Load Forecasting for Commercial Buildings Using BiLSTM–Transformer Network and Cyber–Physical Cognitive Control Systems

Abstract

With the widespread adoption of electric vehicles (EVs), their charging and discharging schedules pose new challenges for real-time load forecasting in commercial buildings. This study proposes a prediction model based on the integration of bidirectional long short-term memory (BiLSTM) networks and Transformer architecture, along with the introduction of a cognitive control system and cyber–physical systems (CPS) to address issues such as data loss and excessive computation time during the forecasting process. The BiLSTM–Transformer model significantly improves load-forecasting accuracy and real-time performance by combining time-series modeling with global feature extraction capabilities. Additionally, the cognitive control system includes user-aware cognitive control (UACC) and Microgrid Control Center Cognitive Control (MACC). UACC quantifies information gaps in real time and adaptively adjusts strategies during communication instability, while MACC employs Q-learning algorithms to evaluate the impact of data loss on scheduling and optimize power resource allocation. The synergy between these mechanisms ensures system stability and predictive performance in scenarios involving data loss or communication disruptions. Experimental results demonstrate that the model achieves outstanding predictive accuracy under complete data conditions and significantly reduces errors in scenarios with data loss, validating its superior accuracy and robustness. This provides reliable support for load forecasting in commercial buildings.

1. Introduction

In modern energy management and smart city construction, accurate load forecasting is essential for the efficient allocation of power resources and the stable operation of the system. Especially in commercial buildings, load forecasting has become one of the core technologies of building energy management due to complex electricity demand and dynamic load changes. In recent years, with the popularization of electric vehicles (EVs) in commercial buildings, the impact of their charging and discharging behavior on the overall electrical load has become more and more significant [1,2,3,4,5,6]. The complexity of charge peak stacking and discharge feedback creates new challenges for traditional power systems and further highlights the need for load forecasting. At the same time, the rapid development of smart grids and cyber–physical systems (CPS) provides technical support to meet these challenges. Smart grids can optimize power distribution strategies and improve energy efficiency by monitoring power demand in real time, combined with dynamic data analysis and forecasting. In the commercial building environment, the charging and discharging needs of electric vehicles often lead to load uncertainty and volatility, and smart grids can effectively mitigate the impact of load peaks on grid stability by using advanced prediction models and dynamic pricing mechanisms.

As an integrated technology, cyber–physical system further improves the flexibility and intelligence of the energy system by integrating physical systems, computing technology, and communication control. In the application of load forecasting, CPS can not only collect and analyze complex data in real time but also make responsive adjustments to dynamic changes, providing strong technical support for load forecasting. In this context, the energy management of commercial buildings is facing a transformation from traditional single-load forecasting to complex multi-scenario load forecasting. The proliferation of electric vehicles has brought new challenges to building-energy management while also creating opportunities to explore more efficient and intelligent load forecasting methods.

In modern buildings, the combination of building-energy management systems (BEMS) and CPS opens up new possibilities for efficient energy dispatch and management. BEMS is able to monitor the energy usage inside the building in real time, ensuring the efficient operation of all equipment. At the same time, the CPS enables seamless connection between physical devices and digital systems through the integration of sensors, data processing, and feedback mechanisms. Specifically, the CPS provides a real-time data stream for the BEMS, enabling the system to respond quickly in a rapidly changing environment. For example, when the charging and discharging behavior of an electric vehicle causes sudden load fluctuations, BEMS can obtain real-time data through the CPS, analyze and quickly adjust the energy dispatch strategy, thereby reducing the risk of load imbalance. In addition, the CPS is able to support BEMS to make autonomous decisions in the face of unstable communication conditions by quantifying information gaps and uncertainties in data transmission. With this integration, BEMS is not only able to optimize energy use but also to enhance its adaptability to dynamic load changes, enabling accurate load forecasting and resource allocation in the event of incomplete information. This integration effectively improves the efficiency and sustainability of building-energy management and meets the increasingly complex energy needs of modern buildings. In commercial buildings, EV-charging and -discharging scheduling in parking lots has a significant impact on the overall load demand of the building, so accurate load prediction after dispatch is crucial. Accurate load forecasting not only enables smart grids to allocate resources in advance and optimize power supply strategies but also mitigates grid overload pressures that can be triggered by centralized charging and discharging peaks. With accurate forecasting, smart grids can more efficiently distribute power and respond to demand, ensuring stable power consumption in commercial buildings while meeting the demand for EV charging. In addition, load forecasting provides a scientific basis for building energy management, making energy use more intelligent and flexible, reducing unnecessary waste, and enhancing the sustainability and resilience of the power system.

In the load forecasting of commercial building clusters, traditional algorithms are difficult to meet the needs of real-time prediction in dynamic environments due to long computing time and slow response speed. With the increase in the demand for charging and discharging of electric vehicles, the frequency and amplitude of load fluctuations have increased significantly, which puts forward higher requirements for the real-time performance and flexibility of load prediction models. However, traditional load-forecasting methods often rely on large amounts of data processing and complex calculations, which make it difficult to respond in a timely manner due to prediction delays in rapidly changing scenarios. This increased computation time affects the timeliness of forecasts, limiting the ability of smart grids to allocate resources and optimize power supply. This limitation can lead to under- or over-allocation of resources, which can affect grid efficiency and the power stability of building clusters. In addition, the complexity of traditional algorithms leads to prediction delays, making it difficult to quickly respond to sudden load changes, hindering refined dynamic load management. Therefore, in complex and dynamic commercial building power scenarios, the low computational efficiency of traditional algorithms makes it unable to meet the needs of real-time load forecasting. In particular, with the increase in the scheduling frequency of electric vehicles, the fluctuation of the overall load of the shopping mall is more significant, which puts forward higher requirements for the real-time performance and computing power of the prediction model. In contrast, machine learning methods, especially deep learning models, are able to automatically learn features from large-scale data, significantly reducing computation time and better meeting the real-time needs of dynamic load forecasting.

However, existing basic machine learning models still have significant limitations when dealing with complex workload data. Some common network architectures perform poorly in dealing with high-frequency nonlinear load fluctuations and long-term and short-term dependencies, and it is difficult to fully meet the dynamic needs of load forecasting in shopping malls. For example, the traditional one-way LSTM model can only make one-way predictions based on historical data, which is difficult to effectively capture the two-way information dependence in the load fluctuation caused by the charging and discharging of electric vehicles. This limitation reduces the flexibility of the model in a rapidly changing load environment, making it difficult to accurately predict future load fluctuations. In addition, the LSTM model also has difficulties in capturing long-term dependencies and nonlinear features, which limits its performance in complex data scenarios. Similarly, while convolutional neural networks (CNNs) have advantages in feature extraction and are suitable for working with static data such as images, they do not perform well when dealing with dynamic, nonlinear load data. CNN models are usually good at extracting local features, and the time dependence and global characteristics of EV charging and discharging load requirements pose a challenge for CNNs to capture these sequence data features. With the continuous fluctuation of charging demand, it is difficult for CNNs to deal with long-term complex dependencies, and the accuracy and timeliness provided in high-dimensional dynamic load data scenarios are insufficient. In addition, although CNN provides a certain amount of computational efficiency in parallel processing of large-scale data, there are still problems of high computational cost and insufficient response speed in high-dimensional dynamic load data scenarios [7].

Although machine learning methods have accelerated the load prediction process to a certain extent, the basic network is still limited in capturing complex nonlinear features, long-term and short-term dependencies, and dynamic changes in load data, which limits its practical application. In order to achieve efficient and accurate prediction of the load after EV-charging and -discharging scheduling, it is necessary to introduce more advanced deep learning models to deal with the dynamic dependence and nonlinear characteristics in the load demand, so as to improve the real time and accuracy of load forecasting.

The BiLSTM–Transformer model combines the time-processing capabilities of BiLSTM and the global modeling capabilities of Transformer, which can effectively capture the complex nonlinear and long-term dependencies in charging demand and significantly improve the timeliness and accuracy of load forecasting. BiLSTM is able to process past and future information in time series data, improving forecast accuracy and better capturing the dynamic changes in EV-charging demand compared to traditional one-way LSTMs. At the same time, the self-attention mechanism of Transformer is extremely efficient in processing large-scale parallel data and can capture global dependencies without relying on fixed sequences, which enhances the robustness of the model in complex charging data scenarios [8,9]. Therefore, compared with other networks, the BiLSTM–Transformer model not only reduces the interference of abnormal data but also significantly improves the accuracy and computational efficiency of prediction, providing a real-time and stable solution for grid load management.

However, predictive models alone are still not enough to fully solve the problem of sudden load fluctuations and data loss caused by EV charging and discharging. During transmission, unstable communication or data loss can directly affect the real-time monitoring and scheduling efficiency of the building-energy management system (BEMS), resulting in load imbalances, scheduling deviations, and potential energy loss. Therefore, a cognitive control model is introduced in this study, which combines the prediction results of BiLSTM–Transformer with a real-time adjustment mechanism to cope with the uncertainty caused by data loss and transmission anomalies.

In this process, the cyber–physical system (CPS) monitors and analyzes data changes in real time by supporting the cognitive control system to predict the load after the mall is dispatched. Through the integrated information transmission, data processing, and feedback mechanism, the CPS enables the cognitive control system to detect and solve information loss or delay problems in a timely manner, ensuring the accuracy of load prediction even in the case of incomplete data. Specifically, the user-side cognitive control (UACC) in the cognitive control system aims to enhance the resilience of the BEMS in the event of unstable wireless communication or data loss. Each charging station acts as an independent UACC agent and can autonomously adjust the charging and discharging operations according to its own historical data and current communication conditions. UACC quantifies the information gap and makes decisions through the instantaneous reward function and adjusts the charging and discharging strategy according to the weight of the information gap and the communication retransmission cost when the information gap exceeds the set threshold. On the other hand, the microgrid control center-side cognitive control (MACC) solves the problem of data loss through a Q-learning algorithm. MACC evaluates the impact of information gaps on dispatching decisions and, when the information gaps exceed the allowable range, re-estimates the gap between user demand and actual demand and ensures reasonable power resource allocation even when the data are incomplete.

By integrating the real-time monitoring and feedback mechanism of CPS, the cognitive control model plays an important role in quantifying information gaps and effectively assessing data uncertainty caused by communication instability. This integration not only enhances the accuracy and flexibility of load forecasting after mall scheduling but also enhances the system’s ability to handle uncertainty in a dynamic environment. On this basis, the innovative algorithm based on BiLSTM–Transformer proposed in this paper combines BiLSTM and Transformer architecture, focusing on the charging and discharging management of a large number of electric vehicles in the parking lot of commercial buildings, so as to achieve accurate prediction of load consumption after shopping mall scheduling. The model not only overcomes the limitations of real-time response and accuracy in traditional methods but also enhances the robustness of the system in an incomplete data environment through cognitive control, providing efficient and reliable support for complex and dynamic shopping mall load management [10].

The main contributions of this paper are as follows:

1. A shopping mall load prediction framework based on BiLSTM–Transformer was constructed. In this framework, the training and prediction processes of the BiLSTM–Transformer network are separated. Offline training enhances the generalization ability of the model, enabling it to quickly generate optimized shopping mall load consumption predictions in real-time applications. The separation of offline training and online prediction ensures real-time performance while enhancing the adaptability and stability of the model.

2. A BEMS model based on BiLSTM–Transformer was proposed. The model combines the powerful feature extraction capabilities of LSTM and the global modeling capabilities of Transformer, enabling the system to efficiently and accurately capture complex patterns and structures in BEMS input time series. This approach significantly improves the computation speed, especially when processing large-scale data, and greatly improves the accuracy and responsiveness of load forecasting.

3. A cognitive control method is introduced to enhance the adaptability and robustness of the system. Cognitive control simulates the learning and adaptive capacity of the human brain, enabling the system to adjust control strategies based on real-time feedback. By combining the long-term predictive power of BiLSTM with the adaptability of cognitive control, the system maintains prediction stability in the event of data loss or unstable communication and makes optimization decisions in a dynamic and uncertain environment, thereby improving the performance of BEMS.

2. Problem Description

The current building-energy management system (BEMS) in shopping malls faces multiple challenges in load management, especially in the context of increasing demand for electric vehicle (EV) charging and discharging. The main task of the BEMS is to collect, predict, and optimize loads through systematic steps to achieve more efficient and economical energy management. This process aims to cope with increasingly complex fluctuations in load demand while maximizing the utilization of electricity resources.

Firstly, the BEMS collects real-time data on the overall load of the shopping mall and EV charging and discharging through a data acquisition module. These real-time data are the foundation for subsequent load forecasting and can provide key basis for the system to understand and analyze future changes in electricity demand. With the increasing impact of EV charging and discharging behavior on the load demand of shopping malls, it has become particularly important for the BEMS to continuously and accurately collect these data in dynamic environments. Only when the immediacy and accuracy of data are guaranteed can the rationality of system forecasting and scheduling be ensured and scheduling errors caused by information loss or delay be avoided.

After completing data collection, the core task of the BEMS is to perform load forecasting. Based on the collected EV charging and discharging data and the basic needs of the shopping mall, the BEMS can predict future load changes. Accurate load forecasting enables the system to have stronger foresight and flexibility in power dispatch decisions. For example, through load forecasting, the BEMS can prioritize EV charging during periods of low electricity prices, avoiding high electricity consumption during periods of high electricity prices. During periods of high electricity prices, the BEMS can reasonably schedule the discharge of EVs and use the stored electricity as an auxiliary power source for shopping malls, thereby reducing dependence on the grid, saving electricity costs, and improving electricity efficiency.

The scenario is illustrated in Figure 1.

The load P_grid of the mall is determined by the sum of the building’s basic demand P_demand and the charging and discharging of the electric vehicle P_EVi, and the Equation is as follows:

(1)$P_{\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{d}}=P_{\mathrm{d}\mathrm{e}\mathrm{m}\mathrm{a}\mathrm{n}\mathrm{d}}+{\sum }_{i=1}^N{P}_{\mathrm{EVi}}$

The primary purpose of the cognitive control system designed within BEMS is to ensure the integrity and reliability of data transmission, particularly in cases where communication instability, data loss, or interference may occur during data collection and transmission. In this context, cyber–physical systems (CPSs) act as a critical infrastructure, providing real-time data acquisition, processing, and feedback mechanisms that further enhance BEMS’s data management capabilities. Since BEMS heavily relies on precise real-time data for load forecasting and scheduling optimization, any data loss or incompleteness can directly affect the system’s forecasting accuracy and scheduling performance. Therefore, the integration of CPS with the cognitive control system is essential for the efficient operation of the BEMS.

Specifically, the CPS can monitor changes in EV charging and discharging as well as the load of shopping centers in real-time, collecting data through its distributed sensor network and transmitting it promptly to the BEMS. On this basis, the CPS continuously monitors the status of data transmission, detecting and transmitting issues such as packet loss, transmission delays, or interference to the cognitive control system in BEMS for resolution. The cognitive control system mitigates the impact of data loss on forecasting and decision-making through methods like retransmission, selective compensation, or inferred completion based on historical data. Additionally, the cognitive control system can adaptively adjust transmission paths or priorities to minimize the impact of network interference on data quality. Through these measures, the cognitive control system ensures that BEMS can receive complete and reliable data support in complex network environments, enabling accurate load forecasting and optimized scheduling while maintaining robustness and efficiency in dynamic and uncertain conditions.

However, challenges such as communication instability, data loss, and transmission interference during data collection and transmission pose significant threats to the BEMS’s load forecasting and scheduling effectiveness. To address these issues, the BEMS integrates a cognitive control system to enhance data transmission integrity and reliability. The core task of the cognitive control system is to monitor data transmission status in real time, promptly identifying potential issues like packet loss, transmission delay, or interference. By doing so, the cognitive control system effectively reduces the negative impact of data loss on system forecasting and decision-making. It can also adaptively adjust transmission paths and priorities, further enhancing data transmission stability by minimizing the effects of network interference.

The integration of the cognitive control system not only improves data transmission integrity and reliability but also strengthens the BEMS’s adaptability and robustness in uncertain environments. By flexibly responding to data loss and transmission delays, the system ensures that the BEMS can still receive complete and reliable data support even in complex network environments, thus achieving more accurate load forecasting and optimized scheduling. With the cognitive control system, BEMS can maintain stable operation even in scenarios with significant fluctuations in EV demand, allowing the system to handle load changes more flexibly and efficiently.

3. Dataset Processing Based on Simulated Annealing Algorithm

In this section, we use the traditional Simulated Annealing (SA) algorithm to construct an optimization model for a building energy management system (BEMS). This model aims to optimize power costs and improve energy efficiency by reasonably scheduling the charging and discharging behaviors of electric vehicles (EVs) and managing power exchanges with the grid. Additionally, the dataset generated by this model will serve as training samples and reference data for subsequent machine learning methods, ensuring that the machine learning model can make forecasts and decisions based on reliable historical data.

The Simulated Annealing algorithm is a heuristic global optimization method inspired by the physical annealing process of solids. By controlling the simulated “temperature”, the algorithm initially accepts suboptimal solutions to avoid getting trapped in local optima. As the “temperature” gradually decreases, the algorithm reduces the acceptance probability for suboptimal solutions, ultimately converging toward a global optimum in [11,12,13,14,15].

Furthermore, this model introduces an “attack system”, which randomly omits portions of the mall’s load data at each time step. This setup simulates the impact of data loss or network attacks on system data transmission in real-world scenarios, increasing the model’s robustness requirements. By handling these incomplete datasets, the model can maintain effective scheduling even in the presence of missing information, further enhancing its practical applicability.

Equation (2) represents the objective function of this model, which is the minimization of the total electricity cost for each vehicle. In this Equation, C_goal represents the total target price; C(t) represents the real-time electricity price; P_EVi represents the power of charge and discharge; R(t) stands for reactive power expense rate; Q_EVi stands for reactive power. Reactive power refers to the power component generated due to the phase difference between voltage and current during the charging and discharging processes of electric vehicles. It is primarily used to maintain the stability of the electromagnetic field in the power grid. The consumption of reactive power reflects the energy exchange characteristics between electric vehicles and the grid. Excessive reactive power consumption can reduce the efficiency of the power grid and increase the difficulty of dispatching. Therefore, it is necessary to incorporate reactive power calculation and optimization into the model. However, reactive power does not directly calculate electricity bills, so the value of R(t) in this formula is 0

(2)$C_{\mathrm{g}\mathrm{o}\mathrm{a}\mathrm{l}}=\mathrm{m}\mathrm{i}\mathrm{n}{\sum }_{i=1}^N{\sum }_{t=1}^{t\mathrm{m}\mathrm{a}\mathrm{x}}(C\left(t\right)P_{\mathrm{EVi}}\left(t\right)+R(t\left)Q_{\mathrm{EVi}}\right(t\left)\right)+C_{\mathrm{PF}}+C_{\mathrm{THD}}$

Equation (3) indicates the power factor fine, and if the power factor PF is too low, the grid will impose a fine. And Equation (4) indicates the harmonic pollution fine, and if the total harmonic distortion THD exceeds the standard, a fine will be imposed.

(3)$C_{\mathrm{PF}}=\alpha ·(0.95-\mathrm{PF})·P_{\mathrm{m}\mathrm{a}\mathrm{x}}$

(4)$C_{\mathrm{T}\mathrm{H}\mathrm{D}}=\beta ·({\mathrm{T}\mathrm{H}\mathrm{D}}_{\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{u}\mathrm{a}\mathrm{l}}-{\mathrm{THD}}_{\mathrm{t}\mathrm{h}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{h}\mathrm{o}\mathrm{l}\mathrm{d}})$

Equation (5) is the time-of-use electricity price formula, which shows the difference in electricity charges at different times, where C₁ represents 0.8 ($), C₂ represents 0.75 ($), C₃ represents 0.9 ($), and C₄ represents 1.05 ($). T₁ is 0:00; T₂ is 10:00; T₃ at 16:00; 19:00 on T₄; T₅ is 24:00.

(5)$C\left(t\right)=\left\{\begin{array}{c}{C}_1,t_1\le t<t_2\\ C_2,t_2\le t<t_3\\ C_3,t_3\le t<t_4\\ C_4,t_4\le t<t_5\end{array}\right.$

Equation (6) is the range of the arrival time of each vehicle: 8:00 for T₆; T₇ is 9:00. Tarrival represents the range of the arrival time of the electric vehicle; i indicates the serial number of the vehicle.

(6)${t_6\le T}_{\mathrm{a}\mathrm{r}\mathrm{r}\mathrm{i}\mathrm{v}\mathrm{a}\mathrm{l}}^i\le t_7$

Equation (7) is the interval range for the departure time of each vehicle. T₈ is 21:00; T₉ is 22:00. Tdepeture indicates the departure time of each vehicle; i indicates the serial number of the vehicle.

(7)${t_8\le T}_{\mathrm{d}\mathrm{e}\mathrm{p}\mathrm{a}\mathrm{r}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{e}}^i\le t_9$

Equation (8) shows the value of the initial SOC range.

(8)$0.3\le {\mathrm{SOC}}_{\mathrm{i}\mathrm{n}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}}\le 0.35$

Equation (9) indicates the target value of the SOC, i.e., the lowest value after charging is completed. SOC_required has a value of 0.2.

(9)${\mathrm{S}\mathrm{O}\mathrm{C}}_{\mathrm{f}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{l}\mathrm{l}\mathrm{y}}={\mathrm{S}\mathrm{O}\mathrm{C}}_{\mathrm{i}\mathrm{n}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}}+{\mathrm{S}\mathrm{O}\mathrm{C}}_{\mathrm{r}\mathrm{e}\mathrm{q}\mathrm{u}\mathrm{i}\mathrm{r}\mathrm{e}\mathrm{d}}$

Equation (10) represents the SOC constraints.

(10)${\mathrm{S}\mathrm{O}\mathrm{C}}_{\mathrm{m}\mathrm{i}\mathrm{n}}\le \mathrm{S}\mathrm{O}\mathrm{C}\left(T\right)\le {\mathrm{S}\mathrm{O}\mathrm{C}}_{\mathrm{m}\mathrm{a}\mathrm{x}}$

Equation (11) indicates the change in SOC, where S represents the battery capacity of the electric vehicle, Kop represents the charging power, and Koi represents the discharge power.

(11)$\mathrm{S}\mathrm{O}\mathrm{C}\left(T\right)=\left\{\begin{array}{ll}SOC\left(T-1\right)+{\displaystyle \frac{P_{\mathrm{E}\mathrm{V}\mathrm{i}}\left(T\right)\cdot T\cdot \mathrm{K}\mathrm{o}\mathrm{p}}{S}},& 0<P_{\mathrm{E}\mathrm{V}\mathrm{i}}\left(T\right)\\ SOC\left(T-1\right)+{\displaystyle \frac{P_{\mathrm{E}\mathrm{V}\mathrm{i}}\left(T\right)\cdot T}{S\cdot \mathrm{K}\mathrm{o}\mathrm{i}}},& P_{\mathrm{E}\mathrm{V}\mathrm{i}}\left(T\right)<0\end{array}\right.$

Equation (12) represents the value of P_EVi, where P_charge represents the charging power; P_discharge represents the discharge power; P_nothing means that it is neither charging nor discharging at this time.

(12)$P_{\mathrm{E}\mathrm{V}\mathrm{i}}=\left\{\begin{array}{cc}{P}_{\mathrm{c}\mathrm{h}\mathrm{a}\mathrm{r}\mathrm{g}\mathrm{e}},& 0\le P_{\mathrm{E}\mathrm{V}\mathrm{i}}\\ P_{\mathrm{d}\mathrm{i}\mathrm{s}\mathrm{c}\mathrm{h}\mathrm{a}\mathrm{r}\mathrm{g}\mathrm{e}},& 0\ge P_{\mathrm{E}\mathrm{V}\mathrm{i}}\\ P_{\mathrm{n}\mathrm{o}\mathrm{t}\mathrm{h}\mathrm{i}\mathrm{n}\mathrm{g}},& 0=P_{\mathrm{E}\mathrm{V}\mathrm{i}}\end{array}\right.$

Equation (13) represents the fitness function in the simulated annealing algorithm. The goal of the fitness function is to minimize the cost of electricity and meet the requirements of the eventual SOC increase. Penalties are imposed if the SOC requirements are not met. SOC_final is the ultimate SOC for electric vehicles.

(13)$\mathrm{F}\mathrm{i}\mathrm{t}\mathrm{n}\mathrm{e}\mathrm{s}\mathrm{s}=\left\{\begin{array}{cc}\mathrm{Cost},& if{\mathrm{S}\mathrm{O}\mathrm{C}}_{\mathrm{f}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{l}}\ge {\mathrm{S}\mathrm{O}\mathrm{C}}_{\mathrm{i}\mathrm{n}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}}+{\mathrm{S}\mathrm{O}\mathrm{C}}_{\mathrm{r}\mathrm{e}\mathrm{q}\mathrm{u}\mathrm{i}\mathrm{r}\mathrm{e}\mathrm{d}}\\ \infty ,& if{\mathrm{S}\mathrm{O}\mathrm{C}}_{\mathrm{f}\mathrm{i}\mathrm{n}\mathrm{a}\mathrm{l}}<{\mathrm{S}\mathrm{O}\mathrm{C}}_{\mathrm{i}\mathrm{n}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{a}\mathrm{l}}+{\mathrm{S}\mathrm{O}\mathrm{C}}_{\mathrm{r}\mathrm{e}\mathrm{q}\mathrm{u}\mathrm{i}\mathrm{r}\mathrm{e}\mathrm{d}}\end{array}\right.$

Equation (14) indicates the probability formula for accepting the new solution. where Current Fitness is the fitness value of the current solution; New Fitness is the fitness value of the new solution; Temperature is the current “temperature”.

(14)$P_{\mathrm{a}\mathrm{c}\mathrm{c}\mathrm{e}\mathrm{p}\mathrm{t}}=\exp ({\displaystyle \frac{\mathrm{C}\mathrm{u}\mathrm{r}\mathrm{r}\mathrm{e}\mathrm{n}\mathrm{t} \mathrm{F}\mathrm{i}\mathrm{t}\mathrm{n}\mathrm{e}\mathrm{s}\mathrm{s}-\mathrm{N}\mathrm{e}\mathrm{w} \mathrm{F}\mathrm{i}\mathrm{t}\mathrm{n}\mathrm{e}\mathrm{s}\mathrm{s}}{\mathrm{T}\mathrm{e}\mathrm{m}\mathrm{p}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{e}}})$

Equation (15) indicates the temperature update formula (cooling rate). where Cooling Rate is the cooling rate, which in this model is 0.98.

(15)$\mathrm{T}\mathrm{e}\mathrm{m}\mathrm{p}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{e}=\mathrm{T}\mathrm{e}\mathrm{m}\mathrm{p}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{t}\mathrm{u}\mathrm{r}\mathrm{e}·\mathrm{C}\mathrm{o}\mathrm{o}\mathrm{l}\mathrm{i}\mathrm{n}\mathrm{g} \mathrm{R}\mathrm{a}\mathrm{t}\mathrm{e}$

Equation (16) indicates that the load of the mall is lost at a certain time step in the attack simulation formula, and the load is set to 0 at this time.

(16)$P_{\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{d}}=0$

In summary, this section demonstrates the effectiveness of the BEMS model constructed using the Simulated Annealing algorithm in optimizing EV charging and discharging behaviors and reducing costs, while also introducing attack simulations. Additionally, this model provides a dataset for subsequent machine learning applications.

4. BILSTM–Transformer Based Bems Model

In this Section, a machine learning method based on BILSTM–Transformer is proposed, and firstly, the preliminary content of the method is presented in Section 4.1. Subsequently, in Section 4.2, the system control structure is constructed.

Some details of the web-based training and implementation process are provided in Section 4.3 and Section 4.4, respectively. In Section 4.5, a filtering process is designed to improve the decision-making performance of the network. Finally, Section 4.6 summarizes the overall flow and structure of the proposed algorithm.

4.1. The Preliminary Description of the Method

As described in Section 2, the main task of BEMS is to dispatch the charge and discharge of numerous electric vehicles parked in the parking lots of commercial buildings.

Since the load change in the mall shows a significant time dependence, its forecasting can be regarded as a typical time-series-forecasting problem. In general, the main purpose of a time-series-forecasting problem is to predict a future state or value based on trends, periodicity, and other relevant characteristics in the time series by analyzing the historical data of the system. In this study, the complexity of the historical data of the system is mainly due to the dynamic changes in time of the charging and discharging needs of electric vehicles, the electricity prices of the upper and lower grids, and environmental factors.

In order to deal with the complexity of multimodal historical data in the system, effectively capture the temporal correlation of P_gird, and manage the charging and discharging operations of a large number of electric vehicles in the parking lot of commercial buildings, we chose BiLSTM (bidirectional long short-term memory network) as the infrastructure. Combined with the multi-head self-attention mechanism of Transformer to further improve the modeling ability of remote time dependence, a scheduling method based on BiLSTM–Transformer short-term load-forecasting model was proposed.

When processing time series data, long short-term memory networks (LSTMs) have become a classic choice in time series forecasting due to their powerful time-dependent modeling capabilities in [16,17,18,19]. By introducing the forgetting gate, input gate, and output gate, LSTM effectively solves the gradient vanishing and gradient explosion problems encountered by traditional recurrent neural networks (RNNs) in long-term sequences in [20,21,22]. Its unique cell structure allows it to retain important information in time series data while ignoring extraneous or unnecessary features.

However, the load variation of the mall is not only affected by the historical state but also by the future state. For example, relying solely on traditional LSTMs for one-way time series modeling has limitations for such complex scheduling problems.

To solve this problem, we introduce a bidirectional long short-term memory network (BiLSTM). Unlike traditional LSTMs, BiLSTMs are able to comprehensively consider past and future state information at each time step of the series by simultaneously processing both forward and backward information of the time series.

At the same time, with the exponential growth of the number of electric vehicles, the generalization time of the traditional model also increases exponentially, and the Transformer network enhances the system’s ability to model the dependence between remote time steps through the global attention mechanism. Especially for high-dimensional time series data, Transformer can process the input of each time step in parallel, thus reducing the time overhead required by BILSTM when processing large-scale data. Through the combination of BiLSTM and Transformer, we can capture both local and global time dependence, so that our method can not only predict the trend in load in shopping malls but also deal with the complexity and uncertainty in large-scale EV scheduling.

The structure of the model consists of an input layer, a convolutional layer, a Transformer encoder layer, a bidirectional LSTM layer, and a fully connected output layer. In the process of model training, we use the convolutional layer, Transformer encoder layer, and BiLSTM layer to extract features according to the historical P_gird state of the mall, and finally output the predicted P_grid value through the fully connected layer. By repeating this forecasting process at each time step, we can generate a P_grid forecasting for the mall in the future to guide its parking-lot charging and discharging scheduling.

4.2. Based on the BiLSTM–Transformer System Architecture

In this study, we propose a system architecture based on BiLSTM–Transformer to optimize the forecasting of power change after charge–discharge scheduling of electric vehicles in parking lots of commercial buildings. In order to achieve fast and efficient forecasting, the system adopts a design that separates training and execution.

The network-training stage is completed at the central management level of the system, which is mainly responsible for learning and modeling the total historical data of the entire electric vehicle charging and discharging scheduling. The BiLSTM–Transformer network is trained on the sum of the time series data of the charging and discharging history of electric vehicles and the historical data of the basic demand power of the shopping mall and has the generalization ability to map the input state to the optimal scheduling decision. After training, the BiLSTM–Transformer model can generate an optimized scheduling strategy and obtain the trained network parameters for the online scheduling stage.

In the execution phase, the charging pile of each electric vehicle can transmit the charging and discharging power to the BEMS, and the system can predict the change in the load of the mall in the next time period according to the real-time status and grid load of each electric vehicle.

According to the system structure, the training part of the BiLSTM–Transformer model can be completed offline, while the online execution part can be run together in the central system and each charging station. This design approach not only allows the central system to continuously optimize the model but also allows each charging station to respond in real time to the dynamic charging and discharging needs of the electric vehicle. Through the separation of offline training and online execution, the system can adapt to the complexity of large-scale electric vehicle charging and discharging scheduling and achieve millisecond-level scheduling response in actual operation.

Different from the traditional methods, the system structure based on BiLSTM–Transformer can make full use of the global modeling ability of Transformer to improve the computational efficiency and response speed of the system in dealing with large-scale electric vehicle charging and discharging scheduling while capturing the long-term and short-term dependencies in the time series. As new data are generated, the central system can periodically retrain the BiLSTM–Transformer model to ensure that the system always makes optimal scheduling decisions based on the latest data.

4.3. Offline Training Based on the BiLSTM–Transformer System Architecture

This Section introduces the details of offline training in practical applications.

4.3.1. Input and Output

As mentioned earlier, we chose BiLSTM–Transformer as the core tool to manage the forecasting of load changes in the mall after the charge/discharge dispatch of an electric vehicle connected to a commercial building parking lot. The output of each BiLSTM–Transformer network is the load of the mall at the next time step t + 1.

The forecasting of P_gird(t) belongs to the time-series-forecasting problem, which is closely related to the previous state. Therefore, the load P_gird(t) of the mall at the current time step t is one of the key inputs to the network. In addition to P_gird(t) itself, the load forecasting also depends on a number of other influencing factors, including the P_demand(t), the sum of the charging and discharging power of electric vehicles P_EVi(t), and the current electricity price C(t) based on the time–characteristic curve. Together, these input variables determine the load P_gird(t) forecasting at the next time step t + 1.

In this study, the input layer consists of five nodes, corresponding to P_gird(t), P_demand(t), P_EVi(t), electricity price C(t), and some temporal characteristics (e.g., diurnal temperature curves). After these inputs are processed at the BiLSTM and Transformer layers, the network generates a node at the output layer, which is P_gird(t + 1).

Due to the powerful bidirectional time correlation capture capability of BiLSTM and the global modeling characteristics of Transformer, the relationship between system input and output can be described as follows: the P_gird(t) forecasting of time step t + 1 not only depends on the current input but also correlates with the state information of the past and future time steps, so as to more comprehensively reflect the change trend in the load of the mall after future charging and discharging scheduling.

Equation (17): Input and output descriptions:

(17)$$\begin{gathered} {\widehat{P}}_{\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{d}}\left(t+1\right)=f[P_{\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{d}}\left(t\right),P_{\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{d}}\left(t-1\right),\dots ,P_{\mathrm{g}\mathrm{r}\mathrm{i}\mathrm{d}}\left(t-l1\right),C(t),C(t-1),\dots ,C(t-l+1) \\ {P}_{\mathrm{d}\mathrm{e}\mathrm{m}\mathrm{a}\mathrm{n}\mathrm{d}}\left(t\right),P_{\mathrm{d}\mathrm{e}\mathrm{m}\mathrm{a}\mathrm{n}\mathrm{d}}\left(t-1\right),\dots ,P_{\mathrm{d}\mathrm{e}\mathrm{m}\mathrm{a}\mathrm{n}\mathrm{d}}\left(t-l+1\right) \\ {\sum }_{i=1}^N{P}_{\mathrm{E}\mathrm{V}\mathrm{i}}(t),{\sum }_{i=1}^N{P}_{\mathrm{E}\mathrm{V}\mathrm{i}}(t-1),\dots ,{\sum }_{i=1}^N{P}_{\mathrm{E}\mathrm{V}\mathrm{i}}(t-l+1)] \end{gathered}$$

4.3.2. Training Dataset

The training dataset consists of the input and output historical data of the system, aiming to provide high-quality training samples and reference data for the BiLSTM–Transformer network. The dataset is generated by the above-mentioned Building Energy Management System (BEMS) model based on simulated annealing algorithm (SA), which reasonably schedules the charging and discharging behavior of electric vehicles, optimizes the power exchange and power cost of the power grid, and simulates the attack and data loss scenarios in the power grid. Each time step in this optimization process provides rich historical data for subsequent machine learning model training.

Specifically, the training dataset contains the sum of the load P_gird(t) of the shopping mall, the P_demand(t), and the charging and discharging power of the electric vehicle P_EVi(t), and the current electricity price C(t). These input features are generated by the BEMS system in different time periods, corresponding to the Pgrid changes in the shopping mall at each time step. Through this process, the training dataset can also reflect the charging and discharging behavior of electric vehicle groups under different electricity prices and power demands.

Considering the large differences in the numerical range of these input features, this study adopts a min–max normalization method to scale all data to the interval [0,1] [0, 1] [0,1]. Since BiLSTM–Transformer is sensitive to the scale of the data, this normalization process helps to improve the training stability and convergence speed of the model.

In addition, the construction of the dataset also takes into account the data loss in the attack simulation, so as to ensure that the system can still make reasonable scheduling decisions in the face of possible data incompleteness in practical applications, and the corresponding network structure (see Section 4.5). By processing these incomplete data, the dataset can provide diverse scene inputs for the BiLSTM–Transformer network, and enhance the generalization ability and robustness of the model.

Since the BiLSTM–Transformer-based EV-dispatching management system is a data-driven technology, in order to ensure the effectiveness of the model in large-scale EV-charging and -discharging scheduling, the training dataset must cover a variety of different dispatching scenarios and power demand situations. In this way, the network can be provided with sufficient samples to have good generalization capabilities to cope with the diverse scheduling requirements in the actual environment.

4.3.3. Data Feature Extraction and Network Parameter Optimization

In general, the data series in the time-series-forecasting problem tend to have significant time-correlation features, so these sequences can be fed into the neural network for modeling. However, in the task of EV-charging and -discharging scheduling and load forecasting in shopping malls, there is the complexity of several data features, which increases the difficulty of model fitting. If these unprocessed data sequences are used directly for training, it can lead to the model overfitting historical data and failing to learn effective forecasting patterns. In this case, even if the model performs well on the training set, it will not perform well in the real world because it cannot handle discrete data with no temporal correlation.

In order to solve the above problems, the steps of data preprocessing and feature recognition are introduced into the network architecture. First, to solve the problem of diversity and discreteness of data, we use a fully connected layer and a one-dimensional convolutional layer to preprocess the original data. This step can extract locally important features by performing convolution operations on the input data, and at the same time filter out possible invalid information or noisy data, ensuring that the model can focus on the key information that really affects the change of P_gird.

Next, the network adopts stacked N attention–BiLSTM modules, and the attention mechanism enhances the model’s ability to pay attention to the global features of the input sequence. The attention mechanism can dynamically assign weights to ensure that the model can capture the important information of P_gird at different time steps when processing time series. Combined with the time series modeling capability of bidirectional long short-term memory (BiLSTM) network, the module further improves the capture of bidirectional time dependencies in the model, so that it can effectively cope with the complex time correlation in the charging and discharging scheduling of electric vehicles.

In particular, the Transformer network is used to extract global data features, and captures long-range dependencies through the multi-head self-attention mechanism to ensure that the system can fully understand the overall structure of the input data, so as to further eliminate the interference of invalid information. The introduction of the Transformer module enables the system to show stronger computational efficiency and global modeling capabilities when processing large-scale data, and avoids the common local optimal problems in traditional time series models.

Finally, after feature extraction by BiLSTM, the network uses a fully connected layer to output the final P_gird forecasts. The fully connected layer aggregates all the information after feature extraction to generate a P_gird forecasting value for the mall at the next time step. This deep model architecture, which combines feature extraction, attention mechanism, and BiLSTM network, ensures the robustness and accuracy of the system in the face of complex charging and discharging scheduling problems.

4.3.4. Network Structure and Hyperparameter Settings

To train the BiLSTM–Transformer model, we used the Adam optimizer with a learning rate set to 0.001, keeping other parameters at their default values. After comparing forecasting accuracy and network complexity, the model’s structure adopts four stacked attention–BiLSTM modules. In the Transformer layer, the number of heads in the multi-head attention mechanism is set to four to fully capture global dependencies across different features, while the Transformer encoder layer is set to one to reduce computational overhead and improve modeling efficiency. The BiLSTM layer contains 64 hidden nodes. Experiments indicate that further increasing the network depth adds complexity with limited improvements in forecasting accuracy and significantly increases training time.

Additionally, the model’s time step is set to five, and the output layer is a fully connected layer. Mean-squared error (MSE) is used as the loss function to measure forecasting error (a regularization term for the microgrid optimization based on cognitive control is discussed in Section 4.5). With these hyperparameter settings, the model achieves optimal forecasting accuracy while maintaining a reasonable training time, avoiding overfitting due to excessive network depth.

4.4. Execution Phase

During the network execution phase, the BiLSTM–Transformer model obtains the power of the currently generated EV through each charging pile. When an EV is connected to a charging station, the charging station first detects the vehicle’s current SOC and updates this value to the model’s input vector. At the same time, the charging station also obtains information about the current state of the grid, including the P_demand(t), the charging and discharging power of the electric vehicle P_EVi(t), and the current electricity price C(t).

After all inputs are successfully updated, the BiLSTM–Transformer model starts predicting the P_gird of the next time step t+1. Because the model combines the time-dependent modeling ability of BiLSTM and the global feature extraction ability of Transformer, it can accurately capture the dynamic change trend of P_gird in the future time step. Based on this forecasting, the system generates a power-scheduling scheme with the current time step to ensure that the charging and discharging operations of each electric vehicle meet the optimization requirements.

The process is repeated at each time step, allowing the model to be continuously adjusted based on real-time data, resulting in efficient power management and system optimization to meet the needs of building operators, power authorities, and consumers.

Refer to Figure 2 for the network structure.

4.5. Microgrid Optimization Method Based on Cognitive Control

After applying the Transformer–BiLSTM network, the building-energy management system (BEMS) can effectively predict the P_gird and control the charging and discharging power in most cases, but due to the increasing demand for EV charging and the increasing complexity of the system, the BEMS faces the problem of unstable communication or data loss in data transmission.

On the one hand, the delay or loss of data in communication will make it impossible for the charging pile to obtain the latest status of the central system and the electric vehicle in time, thus affecting the charging and discharging scheduling decisions of the electric vehicle. This uncertainty can directly affect the energy management of the system, leading to an imbalance in the load on the grid and potentially additional energy losses.

On the other hand, the forecasting results of the Transformer–BiLSTM network built into the charging pile depend on the integrity and stability of the data. In the case of missing data or transmission failure, the model may output inaccurate P_gird forecastings, causing the charging and discharging schedule of electric vehicles to deviate from the optimal scheme and increase the scheduling error of the power grid. In particular, when a large number of electric vehicles are connected to a central system at the same time, the accumulated forecasting error will further amplify the scheduling bias and affect the overall energy efficiency of the system.

Therefore, the communication delay and the acquisition of incomplete information will directly or indirectly affect the real-time monitoring ability of the BEMS system for electric vehicles, resulting in the charging and discharging needs of some electric vehicles cannot be met in time, which further reduces the response speed and scheduling efficiency of the system.

In order to solve the above problems, a cognitive control model was introduced to optimize the performance of BEMS, enhance the robustness of the control system, and reduce the impact of data uncertainty. Cognitive control models are a model-free adaptive learning method that can learn and speculate in response to data loss, latency, or missing information in an environment where wireless communication is unstable. The specific implementation steps are divided into the following three parts: in the first part, user-side cognitive control (UACC) is introduced to help the system autonomously adjust the charging and discharging behavior of electric vehicles in the case of unstable wireless communication or data loss. In the second part, the system is globally scheduled and optimized by the building-energy management center (MCC), which addresses the problem of data loss by estimating deviations in user demand. In the third part, the information gap quantification will be used to measure the data information uncertainty caused by communication instability, the Q learning algorithm will be used as the core learning mechanism to help the system find the optimal scheduling strategy in the case of incomplete data, and finally the joint cost function will be used to help the system achieve a balance between energy management and information uncertainty.

In the BEMS of this study, the implementation of the cognitive control model is mainly achieved through user-side cognitive control (UACC), building-energy management center (MCC)-side cognitive control (MACC), and the quantification of information gaps. Each of these directions takes on different levels of tasks, and their synergy optimizes the performance of BEMS in environments of data loss and unstable communications. The following is a description of the implementation steps and formulas of the model.

4.5.1. User-Side Cognitive Control (UACC)

On the user side of BEMS, the cognitive control system (UACC) leverages CPS’s real-time monitoring and data transmission capabilities to achieve adaptive responses to data loss and communication instability in dynamic environments. CPS’s sensor network collects real-time data on EV charging and discharging status, user demand, and communication channel quality, transmitting this information to UACC. This enables UACC to make adjustment decisions based on information gaps, even in unstable communication environments.

Specifically, UACC quantifies the data gaps transmitted by CPS to balance information gaps, retransmission costs, and channel switching costs, thereby adaptively optimizing charging and discharging strategies. When CPS detects that the information gap H_u(k) exceeds a threshold, UACC can adjust the charging strategy through an instantaneous reward function to reduce scheduling errors caused by incomplete information. Additionally, the real-time data stream from CPS supports UACC in triggering channel switching or increasing the number of retransmissions when channel failures or high bit error rates occur, ensuring the integrity and reliability of information transmission.

(18)$r_u\left(k\right)={\displaystyle \frac{1}{\beta H_u\left(k\right)+C_r{N}_r+C_c\left(a_c\left(k\right)\right)}}$

H_u(k) represents the information gap, reflecting the data uncertainty, β is the weight coefficient of the information gap, 0 ≤ β < 1, C_r is the cost of each retransmission, N_r is the upper limit of the number of retransmissions, C_c(a_c(k)) is the cost of switching the alternate channel, and a_c(k) indicates whether to switch the channel at time k.

In offline training, we will simulate the data loss scenario in a random masking manner to improve the robustness of the model in an uncertain environment. Specifically, the program dynamically generates a masking mask at each time step and applies it to the data to simulate packet loss and guarantee data uncertainty. When the UACC detects that the information gap H_u(k) exceeds the threshold, it will help the UCC agent adjust the charging and discharging strategy according to the information gap weight coefficient β and the communication retransmission cost C_r. The program implementation of the UACC algorithm is shown in Algorithm 1.

Algorithm 1. User-side cognitive control (UACC)

for each user agent u in U do  Calculate information gap H_u(k)  if H_u(k) > threshold then     Adjust charging/discharging strategy based on H_u(k)     r_u(k) = 1/(β * H_u(k)) + C_r * N_r + C_c(a_c(k))     if communication loss occurs then          Use backup channel and apply additional costs C_c(a_c(k))      Update user charging state based on current environment  end if end for

Firstly, when the current channel fails or the bit error rate η_i(k) is high, UACC triggers channel switching based on the switching cost C_c(a_c(k) in the immediate reward function and the long-term expected loss, so as to reduce the information gap caused by communication loss.

Secondly, in the case of incomplete information, UACC will obtain key data by increasing the number of communication retransmissions N_r to improve data integrity. That is, when UACC detects a large information gap H_u(k), it will remind the BEMS to trigger a retransmission, so as to ensure that the required information can arrive in time. The number of retransmissions of the BEMS will also be limited by the cost C_r and the upper limit of the retransmission N_r, which will help the BEMS find a balance between reliability and economics.

Finally, in the case of information loss, the UACC will reduce the charging rate to ensure the smoothness of energy consumption. For example, if the charging rate uncharacteristically exceeds the current power dispatch requirements, the UACC may adjust the charging rate based on information gaps to prevent power waste from the BEMS or overload the grid.

The Function (19) for quantifying the information gap is as follows:

(19)$H_u\left(k\right)={\eta }_i\left(k\right)\left|{\delta }_i^t\left(k-1\right)-{\delta }_i^t\left(k\right)\right|$

where: $\left|{\delta }_i^t\left(k-1\right)-{\delta }_i^t\left(k\right)\right|$ represents the user’s actual power demand in the k iteration, η_i(k) is the frame error rate (FER) of the current channel, which reflects the communication quality.

4.5.2. Microgrid Control Center (MCC)-Side Cognitive Control (MACC)

On the microgrid control center (MCC) side of BEMS, CPS further enhances data transmission and distributed processing, assisting MACC in managing data uncertainties in power scheduling. CPS transmits real-time load demand and communication channel status to MACC, enabling it to quantify information gaps based on these data and balance immediate and long-term rewards through Q-learning. The instant reward Function (20) of MACC is expressed as follows:

(20)$r_m\left(k\right)={\displaystyle \frac{1}{H_m\left(k\right)}}={\displaystyle \frac{1}{{\eta }_i(k)\left|{\widehat{\delta }}_i^t\left(k\right)-{\delta }_i^t\left(k\right)\right|}}$

H_m(k) is the estimated information gap on the MCC side, ${\widehat{\delta }}_i^t\left(k\right)$ is the estimated value of the user’s demand, ${\delta }_i^t\left(k\right)$ is the user’s actual power demand, and η_i(k) is the bit error rate of the communication channel.

In offline training, MACC uses the Function (20) to evaluate the impact of the information gap H_m(k) on scheduling decisions. When H_m(k) exceeds the allowable range, MACC will adjust the power allocation strategy by re-estimating the deviation between user demand ${\widehat{\delta }}_i^t\left(k\right)$ and actual demand ${\delta }_i^t\left(k\right)$ to ensure that the MCC side of the BEMS can still allocate power resources reasonably even if the data are incomplete. The program implementation of the UACC algorithm is shown in Algorithm 2.

Algorithm 2. Microgrid control center (MCC)-side cognitive control (MACC)

for each control center m in MCC do  Estimate user energy demand δ_i^t(k)  Calculate MCC’s information gap H_m(k)  r_m(k) = 1/(H_m(k))  if H_m(k) > acceptable threshold then     Adjust power allocation strategy for all connected users     Use historical data and estimation to minimize H_m(k)     if partial data received then          Use estimation model to fill in missing information  end if end for

4.5.3. Quantification of Information Gap

In the cognitive control model, CPS serves as the core support system for quantifying information gaps. Through its distributed sensors and communication network, CPS not only collects information in real time but also ensures the reliability of data transmission, enabling the system to accurately assess data uncertainties caused by communication instability. The information gap quantification function utilizes CPS-transmitted data to measure the uncertainties the system faces in decision-making and scheduling. When information gaps are large, CPS collaborates with the cognitive control system to minimize these gaps through Q-learning or behavior adjustment mechanisms, thereby enhancing the system’s adaptability in complex environments. The Function (21) for quantifying the information gap is as follows:

(21)$\mathsf{\Delta }I\left(t\right)=I_{\mathrm{e}\mathrm{x}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{e}\mathrm{d}}\left(t\right)-I_{\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{e}\mathrm{i}\mathrm{v}\mathrm{e}\mathrm{d}}\left(t\right)$

where I_expected(t) represents the complete information that the system expects to receive at time step t, and I_received(t) represents the data information that the system actually receives.

In offline training, Equation (21) is used to quantify the information uncertainty caused by incomplete data transmission. When the information gap ΔI(t) is large, the cognitive control system will minimize the gap through Q learning or behavior adjustment mechanisms on the UCC and MCC sides to ensure the accuracy of scheduling decisions and the stability of BEMS. The program implementation of the Q learning algorithm is shown in Algorithm 3.

Algorithm 3. Quantification of information gap and optimization strategy of Q learning

Initialize Q-values Q(s, a) for all states and actions for each time step t do  Observe current state s_t  Choose action a_t based on exploration-exploitation trade-off  Execute action a_t and receive reward r_t  Update information gap ΔI(t) = I_expected(t) − I_received(t)  Update Q-value:     Q(s_t, a_t) = Q(s_t, a_t) + α * [r_t + γ * max_a Q(s_{t+1}, a) − Q(s_t, a_t)]  if information gap persists then     Use estimated values and Q-learning to adjust strategy  end if end for

As the core learning mechanism of cognitive control model, the Q learning algorithm is used to optimize scheduling decisions under uncertainty. Q learning updates the Q value iteratively, gradually learns and finds the optimal scheduling strategy, and the update Function (22) of Q learning is as follows:

(22)$Q\left(s_t,a_t\right)=Q\left(s_t,a_t\right)+\alpha \left[r_t+\gamma \mathrm{m}\mathrm{a}\mathrm{x}Q\left(s_{t+1},a\right)-Q\left(s_t,a_t\right)\right]$

Among them, s_t is the current system state, a_t is the action taken in that state, r_t is the immediate return, which reflects the effect of the current action, α is the learning rate, which determines the update speed of the Q value, and γ is the discount factor, which controls the trade-off between immediate return and long-term return.

In offline training, the Q learning algorithm dynamically adjusts the Q value at each time step based on the actions taken by the current state s_t, and the immediate return r_t. When data are lost or information uncertainty increases, Q learns to evaluate the current and long-term returns by estimating future returns and gradually optimizes the scheduling strategy. Through the feedback mechanism of Q learning, the cognitive control model can find the best balance between long-term benefits and immediate returns and ultimately reduce the negative impact of information gap on BEMS scheduling.

It is worth noting that in order to consider the impact of both energy cost and information gap, the cognitive control model introduces a joint-cost function. By striking a balance between energy management and information uncertainty, this function helps BEMS optimize energy use while minimizing the cost caused by information loss.

(23)$C\left(t\right)=\sum _{i=1}^n\left(C_{\mathrm{e}\mathrm{n}\mathrm{e}\mathrm{r}\mathrm{g}\mathrm{y}}\left(t\right)+C_{\mathrm{g}\mathrm{a}\mathrm{p}}\left(t\right)\right)$

In Function (23), C_energy(t) represents the energy cost at time step t, and C_gap(t) represents the cost due to information gaps.

In offline training, the cognitive control system evaluates the scheduling effect based on this cost function. By minimizing this joint-cost function, the cognitive control system can balance the relationship between data loss, information uncertainty, and actual energy consumption in an uncertain environment, so as to realize the optimal scheduling of BEMS. The program implementation of the joint-cost function is shown in Algorithm 4.

Algorithm 4. Joint-cost function

for each time step t do  Calculate energy cost C_energy(t)  Calculate information gap cost C_gap(t)  Compute total cost:     C(t) = Σ_i = 1^n (C_energy(t) + C_gap(t))  if total cost exceeds threshold then     Recalculate strategy to minimize cost  end if end for

Since the algorithm of this model is mainly composed of simple logical judgments and assignment statements, it can be implemented quickly and efficiently.

4.6. The Overall Implementation Process of BEMS Model Based on Bilstm-Transformer

In the offline phase, BEMS collects historical data of all charging piles, including the charging and discharging behavior of electric vehicles, building power demand, time-of-use electricity prices, and historical optimal scheduling results. The simulated annealing (SA) algorithm is used to calculate and obtain the optimal scheduling scheme under these datasets, and use it as the supervised output in the training dataset. After the data are preprocessed by the fully connected layer and the convolutional layer, the training stage of the BiLSTM–Transformer network is entered. At this stage, the CPS works in tandem with the cognitive control model to quantify the information gap and optimize the retransmission strategy, so as to effectively enhance the robustness of the network. Since the data distribution may change over time, the system re-enters the offline training phase after a certain period, and the latest data provided by CPS is used to update the model to ensure the timeliness of network parameters.

In the online phase, the system predicts changes in the load of the mall and further optimizes these predictions through a filtering process. The main purpose of filtering is to remove noise and outliers from the data, ensuring that the model can make more accurate and stable decisions. This method enables the system to accurately track the real-time trend in power demand in shopping malls, providing a reliable basis for subsequent scheduling.

Once the filtered forecast data are available, the CPS will enable BEMS to guide the decentralized, real-time dispatch of EVs across multiple parking lots. This approach not only balances the load between different parking lots but also flexibly adjusts the charging strategy based on fluctuations in the mall’s load demand, thus avoiding the strain on the mall’s overall power system. This scheduling process continues during mall business hours, ensuring efficient distribution of power resources to meet the needs of malls and electric vehicles at different times of the day.

A large amount of data generated during the scheduling process is collected by the CPS, including load prediction values, filtered data, and the charging status of electric vehicles in the parking lot, etc., which are recorded and fed back to the system for subsequent offline retraining to improve the performance of the model. By learning and optimizing historical data, the system can better adapt to the fluctuating pattern of power demand in shopping malls and gradually improve the overall decision-making ability and accuracy of the BEMS system, so as to achieve more efficient power management. Figure 3 shows the operational process of the entire system.

5. Result Analysis

This Section is divided into three paragraphs. Section 5.1 introduces the training and testing sets of the experiment, as well as different comparison samples. Section 5.2 introduces the learning performance of the BiLSTM–Transformer network. Section 5.3 analyzed the forecasting results of different situations

5.1. Experimental Dataset and Comparative Samples

In order to effectively predict the load of the mall after scheduling, we collected data for 100 days of the year as the original research material. First, we generated the optimized dataset by simulated annealing (SA). The dataset records in detail key data such as the P_gird(t), the P_demand(t), the charging and discharging power of electric vehicles P_EVi(t), and the time-of-use electricity price (C) of the upper and lower grids. The optimized dataset serves as the basis for subsequent model training. The optimised 100-day grid power consumption dataset is shown in Figure 4, where different colours are used to represent the grid power consumption curves for different days, and the daily grid power consumption is divided into 720 steps for subsequent network learning.

In the process of data collection and model evaluation, we selected partial days of data as the key sample source for the test set, which covers the time series changes in all key variables. The selection of testing equipment takes into account the different charging and discharging behaviors of electric vehicles, as well as load of buildings at different time periods. The remaining time is used for training.

In order to verify the forecasting effect of the model and the control ability of the system, we compare five different methods:

1. BiLSTM–Transformer: This is the core model proposed in this paper, which combines bidirectional long short-term memory network (BiLSTM) and Transformer architecture to effectively capture complex patterns and remote dependencies in time series. The model was trained and predicted using a complete and correct dataset.

2. BiLSTM–Transformer + data loss: This model is based on BiLSTM–Transformer for prediction, but data loss is introduced in the test set to evaluate the performance of the model in the environment with incomplete data.

3. BiLSTM–Transformer + cognitive control system + data loss: The model adds a cognitive control system on the basis of BiLSTM–Transformer to deal with incomplete data or transmission failure. Both the training set and the test set contain data loss scenarios to verify the robustness of the cognitive control system in complex environments.

4. LSTM–Transformer: The combination of traditional long short-term memory (LSTM) network and Transformer is used to compare the prediction accuracy with BiLSTM–Transformer. This model provides a benchmark for evaluating the advantages of the new model.

5. CNN-LSTM: The CNN-LSTM model combines convolutional neural networks (CNN) and long short-term memory networks (LSTM), aiming to fully utilize the advantages of both. CNN excels at extracting local features from input data, especially when processing time series data, and can effectively identify and capture short-term changes in signals. LSTM, on the other hand, can handle long-term dependencies and maintain the flow of information through memory and forgetting mechanisms. Therefore, the CNN-LSTM model can simultaneously focus on local features and long-term trends in time series prediction, thereby improving the accuracy and stability of the prediction.

6. RNN (Recurrent Neural Network) is a traditional model for processing sequential data, and its structure can effectively capture dependencies in time series. RNN processes each time step in the input sequence through cyclic connections, allowing the model to remember previous input information. However, RNNs may face the problem of vanishing or exploding gradients when dealing with long-term dependencies, which limits their performance in complex time series prediction tasks. Therefore, RNN is often used as a benchmark model to compare with more advanced models such as BiLSTM–Transformer and CNN-LSTM, in order to evaluate the advantages of the new model in load forecasting.

7. True value of annealing algorithm run: The real value generated by the simulated annealing algorithm is used as a benchmark for comparison. The algorithm uses a complete dataset, ensuring accurate raw-state forecasts.

5.2. Learning Performance of BiLSTM Transformer Network

Firstly, the learning performance of the BiLSTM–Transformer network was evaluated. During training, we use the mean-squared error (MSE) as a loss function to measure the deviation between the network output and the target value in the training dataset. The change in the loss function during training is shown in Figure 5. In this study, the BiLSTM–Transformer network was trained on a dataset containing complete information, using high-resolution data of 720 time steps.

In the first 30 iterations of the initial training phase, the loss function drops rapidly, indicating that the network quickly captures the dominant patterns in the data. After 80 iterations, the loss function gradually stabilizes and falls to a level below 10⁻⁴. Although there are occasional slight fluctuations in the loss function in the later stages of training, the overall trend indicates that the algorithm has good convergence.

After the training is completed, the BiLSTM–Transformer network shows strong generalization ability and can effectively cope with the scheduling requirements in the online environment. By combining the temporal feature extraction capability of BiLSTM and the global dependency modeling capability of Transformer, the network can maintain stable forecasting performance in a variety of complex scenarios. The training loss is shown in Figure 5.

5.3. Load-Forecasting Result

This Section will present the experimental results. Section 5.3.1 will show the prediction performance and performance parameters of different network structures (Table 1). Section 5.3.2 will show the prediction results and performance parameters of the cognitive control system with or without data loss (Table 2).

Firstly, from the perspective of symmetry, an ideal forecasting model should be able to make consistent responses to changes in actual data when processing time series and should not exhibit overreaction or lag biased towards a certain direction. For example, if the model can accurately capture changes in actual data during an upward trend, it should also exhibit similar accuracy during the same downward trend. This symmetry is reflected in the balance and consistency of the model, that is, the response to load changes can remain symmetrical in both rising and falling stages. Applied to load forecasting: In load-forecasting tasks, if the actual load rises within a certain period of time, the predicted curve should also rise smoothly, maintaining symmetry with the actual data. If the predicted values exhibit irregular spikes or deviations at certain moments, it indicates that this symmetry has been disrupted and the model’s handling of signal changes at these locations is asymmetric.

5.3.1. Network Performance Comparison

In Figure 6a, The BiLSTM–Transformer (BT) model: It can be seen that the model is almost completely consistent with the overall trend and the actual load change, especially in the key load fluctuation area; the prediction curve can closely follow the actual curve, showing a high degree of synchronization and stability. By combining the bidirectional time series processing capability of BiLSTM and the self-attention mechanism of Transformer, the BT model effectively captures the short-term and long-term dependence of the load, thereby enhancing the accuracy of the model when dealing with complex load changes. As you can see in the graph, the forecast curve almost coincides with the actual load curve with no significant deviations or spikes, indicating that the model performs well in terms of symmetry and continuity. This consistency allows the BT model to maintain a high degree of balance in capturing both upward and downward trends in load.

In Figure 6b, The LSTM–Transformer model: The LSTM–Transformer model combines a long short-term memory network (LSTM), which is used to capture short-term dependencies in time series, and Transformer, which handles global dependencies through a self-attention mechanism and the Transformer architecture. This combination gives the model an advantage in dealing with the complexity of load changes. As can be seen from the Figure, the prediction curve of the LSTM–Transformer model is in good sync with the actual load curve in the general trend, especially in the interval where the load is relatively stable; the model can follow the fluctuation of the actual load more accurately, showing a certain stability. However, in some rapidly changing regions, there is a slight deviation between the forecast curve and the actual load, especially near the peaks and valleys, and there is a certain delay or shift in the model’s predictions. This situation may be due to the sequential processing nature of the LSTM, and although the Transformer part can help alleviate this problem, the response of the model still lags slightly under complex load changes.

In Figure 6c, The RNN model: It can be seen that the RNN model is generally consistent with the actual load curve in terms of overall trend and can capture the basic trend in areas where the load changes are relatively stable. However, RNNs do not perform well in areas with significant load changes. The Figure shows that the predicted curve has deviations in some peak and valley areas, making it difficult to keep up with the fluctuations of actual load. This performance indicates that RNNs may lack the ability to handle long-term dependencies and rapid changes, especially in complex load scenarios where model lag and bias are more pronounced. In addition, the unidirectional structure of RNN limits its ability to capture bidirectional dependency information, resulting in limited performance of the model during severe load fluctuations.

In Figure 6d, The CNN-LSTM model: It can be seen that the predicted curve of the model is generally consistent with the trend in the actual load curve and can closely follow the changes in the actual load in most areas, showing a good fitting effect. Especially in the range of stable load changes, the predicted values of the model are close to the actual values, indicating that the model has strong stability in capturing smooth load trends. However, in some areas, there are slight deviations between the predicted curve and the actual load curve, and there are also slight peaks or fluctuations in some positions. These biases reflect the limitations of the model in dealing with detailed changes in load signals, especially in areas where there are significant changes in load, where the model may experience brief overreaction or lag. This phenomenon indicates that the CNN-LSTM model has limited ability to capture rapid changes in load and is prone to generating certain errors in highly fluctuating intervals.

Mean-squared error (MSE) is a metric that evaluates the accuracy of a prediction model, representing the mean square difference between the predicted value and the actual value. It is calculated by squaring the difference between the predicted and actual values and averaging the differences in squares. The smaller the value of the MSE, the smaller the error of the prediction of the model, and the more accurate the model. Because MSE is sensitive to predicted values with large deviations (because the square increases the weight of large deviations), it can reflect how the model will perform in extreme cases. The root mean-square error (RMSE) is the square root of the MSE and is used to reduce the error to the same units as the original data, making the error more intuitive. RMSE is a common metric to evaluate the standard error of a model’s prediction, and a smaller value indicates that the model has a smaller prediction error in the average sense. Similar to MSE, RMSE is also sensitive to large errors, but its units are consistent with actual data, so it is more intuitive in practical applications.

As can be seen from Table 1, among the load-forecasting performance of the four models, the BiLSTM–Transformer model stands out with the lowest MSE (43.290) and RMSE (6.579), showing extremely high prediction accuracy and being able to effectively track the actual load fluctuation closely. This outstanding performance is attributed to the combination of bidirectional time series processing capabilities of bidirectional LSTM and the attention mechanism of Transformer, which is able to better capture short-term and long-term dependencies. In contrast, although LSTM–Transformer performs well in capturing the overall trend, the one-way LSTM limits its utilization of bidirectional dependent information, resulting in MSE and RMSE of 336.820 and 18.352, respectively. The RNN model has a medium performance, and its simple structure makes it inferior to the LSTM series in long-term dependence processing, and the error is further increased (MSE is 489.220 and RMSE is 22.118). The MSE of the CNN-LSTM model is 505.710 and the RMSE is 22.488, which is the highest error among all models. Overall, the BiLSTM–Transformer model performs best in prediction accuracy and is best suited for complex tasks of load forecasting.

5.3.2. Performance Analysis of BiLSTM–Transformer with Cognitive Control Under Data Loss Conditions

In Figure 7a, BiLSTM Transformer with Data Loss: Although there were some time periods where the predicted load closely matched the actual load, the curve trend remained basically synchronized, indicating that the model symmetrically captured the changing trend in actual load during this period. However, it can be clearly seen that during the time step from 400 to 700, especially around 500, the predicted curve exhibits significant fluctuations and spikes. This phenomenon indicates that the model failed to symmetrically reflect the actual changes in load during this period, resulting in significant differences between forecasts and reality. From a mathematical perspective of symmetry, an ideal forecasting model should capture signal features symmetrically at different time periods along the timeline, i.e., maintain consistent responses in various rising and falling trends. However, in this figure, the predicted curve exhibits irregular spikes at multiple locations, indicating a local asymmetry that suggests the model has failed to maintain stable signal capture capabilities when dealing with data loss. Especially in the high-frequency variation region, the predicted values exhibit overreaction, leading to a significant deviation from the actual load trend. During the scheduling period, there is a significant symmetry imbalance between the predicted load and the actual load of the model, especially with frequent and drastic fluctuations occurring between time steps 400 and 700, resulting in a high mean-squared error.

In Figure 7b, BiLSTM–Transformer combined with cognitive control system and data loss: However, between the 300 and 700 time steps, slight local biases and fluctuations in the forecasting curves can be observed, and although these fluctuations are small, the forecasts are slightly overestimated or underestimated at some peak points compared to the smoothing trend in actual load. This local asymmetry, while not significant, still reflects a subtle inadequacy in the model’s ability to capture complex power changes. The results show that the predicted load of the model generally shows good symmetry during the scheduling period, especially in the middle and early part of the upward trend, the trend in the forecasting curve and the actual curve is basically the same. However, small local fluctuations reflect a slight asymmetry in the model’s response to complex signals, which manifests as a slight deviation from the predicted curve.

In the case of data loss, Table 2 shows the comparison of the performance parameters of the BiLSTM–Transformer model with or without the support of a cognitive control system. When the model does not have a cognitive control system (BT + data loss), the MSE is as high as 37,413.696, and the RMSE is 193.426, showing significant prediction errors, indicating that the model is difficult to accurately capture load fluctuations in the environment with missing data. When the cognitive control system (BT + data loss + cognitive control system) is added, the MSE is significantly reduced to 546.603, and the RMSE is also reduced to 23.380, indicating that the cognitive control system effectively improves the prediction ability of the model in the environment of incomplete data. This significant error reduction shows that the cognitive control system can compensate for the impact of missing data in the case of incomplete information, so as to improve the stability and prediction accuracy of the model and ensure the reliability of grid dispatching.

6. Conclusions

Based on the above experimental results, this study draws the following conclusions. First, the BiLSTM–Transformer model demonstrates outstanding performance in load-forecasting tasks, effectively capturing the complex patterns and long-range dependencies of load variations in commercial buildings, achieving extremely high prediction accuracy under complete data conditions. Additionally, although the model’s prediction accuracy decreases in the presence of data loss, the introduction of a cognitive control system and CPS significantly enhances the robustness of the BiLSTM–Transformer model in cases of incomplete data, ensuring more stable prediction performance. CPS plays a critical role in data collection, transmission, and feedback, further improving the system’s responsiveness and adaptability in dynamic environments. Compared with traditional LSTM models, the BiLSTM–Transformer model exhibits greater computational efficiency and accuracy when handling large-scale and dynamic load data, helping to improve the system’s response speed in complex load scenarios. In conclusion, the BiLSTM–Transformer model, combined with a cognitive control system and CPS, provides a practical and reliable solution to address uncertainties and data loss in commercial building load forecasting, offering more accurate data support and decision-making basis for load management and power scheduling in smart grids.

Author Contributions

Conceptualization, X.X., Z.H. and J.S.; Methodology, X.X. and Z.H.; Software, X.X. and Z.H.; Verification, X.X., Z.H. and Y.C.; Formal analysis, X.X.; Investigation, Z.H.; Resources, X.X.; Data management, X.X. and Z.H.; Writing—First draft preparation, X.X. and Z.H.; Writing—Review and Editing, J.S.; Visualization, X.X. and Z.H.; Supervision, J.S.; Project management, X.X. All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

Data are contained within the article.

Conflicts of Interest

The authors declare no conflicts of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. Scenario diagram.

Figure 2. Network structure.

Figure 3. Overall implementation procedure of BiLSTM–Transformer-based BEMS model.

Figure 4. Load include 100 Days.

Figure 5. Training loss of BiLSTM–Transformer.

Figure 6. Prediction performance comparison of neural network models for load forecasting. (a) The BiLSTM–Transformer (BT) model; (b) The LSTM–Transformer model; (c) The RNN model; (d) The CNN-LSTM model.

Figure 6. Prediction performance comparison of neural network models for load forecasting. (a) The BiLSTM–Transformer (BT) model; (b) The LSTM–Transformer model; (c) The RNN model; (d) The CNN-LSTM model.

Figure 7. Comparison of BiLSTM–Transformer with and without cognitive control under data loss. (a) BiLSTM Transformer with Data Loss; (b) BiLSTM–Transformer combined with cognitive control system and data loss.

Table 1

Performance parameters of neural network models.

Network	MSE	RMSE
CNN-LSTM	505.710	22.488
RNN	489.220	22.118
LSTM–Transformer	336.820	18.352
BilSTM–Transformer	43.290	6.579

Table 2

Performance metrics of BiLSTM–Transformer with and without cognitive control under data loss.

Network	MSE	RMSE
BT + data loss	37,413.696	193.426
BT + data loss + cognitive control	546.603	23.380

References

1. Wan, Z.; Zhang, T.; Liu, W.; Wang, M.; Zhu, L. Decentralized privacy-preserving fair exchange scheme for V2G based on blockchain. IEEE Trans. Dependable Secur. Comput.; 2022; 19, pp. 2442-2456. [DOI: https://dx.doi.org/10.1109/TDSC.2021.3059345]

2. Li, Z.; Li, X.; Wang, J. Research progress on thermal runaway warning and protection of lithium-ion battery in energy storage power station. High Volt. Appar.; 2024; 60, pp. 87-99. [DOI: https://dx.doi.org/10.13296/j.1001-1609.hva.2024.01.011]

3. Hu, W.; Su, C.; Chen, Z.; Bak-Jensen, B. Optimal operation of plug-in electric vehicles in power systems with high wind power penetrations. IEEE Trans. Sustain. Energy; 2013; 4, pp. 577-585. [DOI: https://dx.doi.org/10.1109/TSTE.2012.2229304]

4. Guille, C.; Gross, G. A conceptual framework for the vehicle-to-grid (V2G) implementation. Energy Policy; 2009; 37, pp. 4379-4390. [DOI: https://dx.doi.org/10.1016/j.enpol.2009.05.053]

5. Azuatalam, D.; Chapman, A.C.; Verbič, G. Probabilistic assessment of impact of flexible loads under network tariffs in low-voltage distribution networks. J. Mod. Power Syst. Clean Energy; 2021; 9, pp. 951-962. [DOI: https://dx.doi.org/10.35833/MPCE.2019.000136]

6. Liu, Z.; Wu, Q.; Shahidehpour, M.; Li, C.; Huang, S.; Wei, W. Transactive real-time electric vehicle charging management for commercial buildings with PV onsite generation. IEEE Trans. Smart Grid; 2019; 10, pp. 4939-4950. [DOI: https://dx.doi.org/10.1109/TSG.2018.2871171]

7. Xie, W.; Zhao, M.; Liu, Y.; Yang, D.; Huang, K.; Fan, C.; Wang, Z. Recent advances in Transformer technology for agriculture: A comprehensive survey. Eng. Appl. Artif. Intell.; 2024; 138, Pt B, 109412. [DOI: https://dx.doi.org/10.1016/j.engappai.2024.109412]

8. Cui, X.; Zhu, J.; Jia, L.; Wang, J.; Wu, Y. A novel heat load prediction model of district heating system based on hybrid whale optimization algorithm (WOA) and CNN-LSTM with attention mechanism. Energy; 2024; 312, 133536. [DOI: https://dx.doi.org/10.1016/j.energy.2024.133536]

9. Pan, M.Y.; Yang, H.; Fan, X.H. Change detection in remote sensing images based on CNN-transformer structure. J. Liq. Cryst. Disp.; 2024; 39, pp. 1361-1379. (In Chinese)

10. Fang, X.; Han, Y.; Wang, J.; Zhao, Q. A cognitive control approach for microgrid performance optimization in unstable wireless communication. Neurocomputing; 2019; 355, pp. 168-182. [DOI: https://dx.doi.org/10.1016/j.neucom.2019.04.048]

11. Guo, L.; Yu, D.; Li, Y.; Chen, L.; Du, H. Evaluation of spatial straightness error based on simulated annealing algorithm. Tool Technol.; 2024; 58, pp. 151-155. (In Chinese) [DOI: https://dx.doi.org/10.3969/j.issn.1000-7008.2024.05.029]

12. Haode, M.; Zhengtian, W.; Wendi, W.; Shi, K. Multi objective path planning for fire extinguishing vehicles based on improved simulated annealing algorithm. Comput. Digit. Eng.; 2024; 52, pp. 394-398. (In Chinese) [DOI: https://dx.doi.org/10.3969/j.issn.1672-9722.2024.02.017]

13. Luo, Y. Research on Taipei City Road planning construction method based on improved simulated annealing algorithm. J. Spatiotemporal Inf.; 2024; 31, pp. 240-247. (In Chinese) [DOI: https://dx.doi.org/10.20117/j.jsti.202402017]

14. Lu, Y.; Lin, Y.; Peng, Q.; Wang, Y. Overview and parameter exploration of simulated annealing algorithm improvement. Coll. Math.; 2015; 31, pp. 96-103. (In Chinese) [DOI: https://dx.doi.org/10.3969/j.issn.1672-1454.2015.06.020]

15. Yang, H.; Lin, X. A review of the optimization capabilities of genetic algorithms and simulated annealing methods. Mech. Manuf. Autom.; 2010; 39, pp. 73-75. (In Chinese) [DOI: https://dx.doi.org/10.3969/j.issn.1671-5276.2010.02.020]

16. Hu, L.; Jiang, X.; Zhou, J.; Ouyang, F.; Dai, Y.; Zhang, L.; Fu, X. Application of LSTM model considering memory time in runoff forecasting of Ganjiang River Basin. Lake Sci.; 2024; 36, pp. 1241-1251. (In Chinese) [DOI: https://dx.doi.org/10.18307/2024.0454]

17. Li, X.; Su, C.; Yuan, R. Time series reconstruction improves LSTM based dam deformation prediction model. Hydroelectr. Power Gener.; 2024; 50, pp. 67-71. (In Chinese) [DOI: https://dx.doi.org/10.3969/j.issn.0559-9342.2024.06.015]

18. Li, W.; Wang, X. Time series prediction method based on simplified LSTM neural network. J. Beijing Inst. Technol.; 2021; 47, pp. 480-488. (In Chinese) [DOI: https://dx.doi.org/10.11936/bjutxb2020120032]

19. Zhu, Z.; Wu, Y.; Hu, Y.; Jiang, J. A cluster user job execution time prediction model based on LSTM. Comput. Eng. Sci.; 2022; 44, pp. 1331-1341. (In Chinese) [DOI: https://dx.doi.org/10.3969/j.issn.1007-130X.2022.08.001]

20. Yan, P.; Zhang, X.; Shang, S.; Zhang, C. LIF combined with LSTM neural network for mine water source identification. Spectrosc. Spectr. Anal.; 2022; 42, pp. 3091-3096. (In Chinese) [DOI: https://dx.doi.org/10.3964/j.issn.1000-0593(2022)10-3091-06]

21. Ru, H.; Ma, J.; He, H.; An, Y.; Zheng, J. Supermarket vegetable product revenue prediction based on LSTM neural network. Chang. Inf. Commun.; 2024; 37, pp. 69–72, 76. [DOI: https://dx.doi.org/10.20153/j.issn.2096-9759.2024.04.019]

22. Zhong, J.; Wang, S.; Ran, Y.; Ran, X.; Yu, J. Short term user load forecasting based on mutual information and LSTM. Electr. Power Constr.; 2022; 43, pp. 96-102. (In Chinese) [DOI: https://dx.doi.org/10.12204/j.issn.1000-7229.2022.07.011]