Non-Intrusive Load Monitoring and Controllability Evaluation of Electric Vehicle Charging Stations Based on K-Means Clustering Optimization Deep Learning

Abstract

Electric vehicles have the advantages of zero emissions and high energy efficiency. They have a broad potential in today’s social life, especially in China where they have been widely used. In the current situation, whereby the storage capacity of electric vehicles is continually increasing and the requirements for grid stability are getting higher and higher, V2G technology emerges to keep up with the times. Since the electric vehicle charging station is a large-scale electric vehicle cluster charging terminal, it is necessary to pay attention to the status and controllability of each charging pile. In view of the lack of attention to the actual operation of the electric vehicle charging station in the existing vehicle–network interaction mode, the charging state of the current electric vehicle charging station is fixed. In this paper, deep learning is used to establish a load perception model for electric vehicle charging stations, and K-means clustering is used to optimize the load perception model to realize random load perception and non-intrusive load monitoring stations for electric vehicle charging. The calculation example results show that the proposed method has good performance in the load perception and controllability evaluation of electric vehicle charging stations, and it provides a feasible solution for the practical realization of electric vehicle auxiliary response.

1. Introduction

In the context of the 2030 carbon-peaking action plan proposed by the Chinese government in 2021 [1], in order to rationally regulate oil and gas consumption, gradually reduce the dependence of transportation on fossil energy, and guide the public to choose a low-carbon/carbon green travel mode, the energy industry has high expectations, and the low-carbon transformation of transportation needs to be accelerated [2]. The power system is optimized to make the power system safer and more efficient. The design of the power system to the smart power system will increase its transparency. This transparency refers to the small particle size running data that contains various equipment status—including the type of electrical appliances, running time, and the operating habits of users—which has great value. By understanding the operating information of a single electrical appliance, that is, the small particle size operation data of the electrical appliances mentioned above, these data can intuitively reflect the operation of the electrical appliances to energy users. This can guide users’ energy-saving behavior. As of October 2020, China has 666,500 public charging piles and 43,300 charging stations, and has built a charging facility system with the largest number of charging facilities, the largest radiation area, and the most complete service vehicles in the world [3]. However, when a large number of electric vehicles enter the grid for charging, the stability of the grid will be greatly challenged. At the same time, with the rapid development of battery technology and vehicle-to-grid (V2G) control technology, the storage capacity of the battery has further increased, and the battery control foundation of electric vehicles has become more mature. Assisting the operation of new energy power systems has become a hot research topic [4]. At the same time, the electric vehicle charging state needs to be obtained before the implementation of the V2G algorithm, so how to perceive the electric vehicle load becomes the key to the operation of the cluster electric vehicle auxiliary power system.

There are two different ways to implement load monitoring, namely Intrusive Load Monitoring (ILM) and Non-Intrusive Load Monitoring (NILM). Intrusive load monitoring monitors each device individually by installing sensors on each device to capture the power consumption information of each device. This monitoring solution focuses mainly on the development of hardware and how to extract the above information quickly, efficiently, and accurately while reducing the cost of hardware equipment. Non-invasive load monitoring focuses mainly on software algorithms, especially how to use the power data of the electric vehicle charging station terminal to analyze the charging behavior of each charging pile. Although the ILM scheme based on sensors has high precision advantages, it also has disadvantages such as high prices and huge waste. Therefore, despite the limited accuracy of NILM, due to its low price, it is very popular among energy users. As a result, it has come to the forefront of research in the field of academic and industrial.

NILM was proposed by Professor Hart G. W of MIT in the 1990s [5]. Due to the high computational complexity and low accuracy of the original NILM, it did not attract much attention at the time. In recent years, due to the rapid development of computer science and artificial intelligence, especially in the extensive application of in-depth learning in model recognition, NILM has now attracted the attention of scholars and has become a hot spot in smart grid research. At present, NILM is classified into two categories: low frequency and high frequency, according to different sampling rates. Due to the inherently simple structure of the data, the analysis results of low-frequency data may lead to performance degradation. Since high-frequency data usually contains more information, such as current harmonics, voltage–current traces, and high-frequency transient waveforms [6], the results of high-frequency data analysis are usually more accurate. However, compared with low-frequency data analysis, high-frequency data’s high requirements for monitoring equipment show a cost disadvantage. From a practical application point of view and considering the industrial aspect, solutions based on low-frequency data are becoming increasingly more attractive due to stringent cost requirements and acceptance of loss of accuracy, thus becoming the focus of our work.

Various NILM solutions have been proposed. First, some mathematical models are explored [7,8,9]. The fuzzy model obtains the objective function of membership of each sample signal to all home appliance centers through optimization, so as to determine the appliance category of the sample signal and to achieve the purpose of automatically identifying the sample data [8]. Graph signal processing first simply builds an undirected graph based on the signal, then groups the on/off events of the load, and finally defines an optimization problem to find the signal with the smallest change [7]. The linear programming model will be an optimization of the NILM problem to form a multi-feature objective function, so as to realize the load decomposition and identification of different characteristics [9]. Although the current research based on mathematical models has achieved some results, it still has some limitations. The most prominent problem is that most of these explorations are based on optimized NILM schemes, resulting in low scalability of the formula and high algorithm dependence; that is, when the complexity of the problem increases or the model has structural changes, such methods will have large deviation. Therefore, the decomposition performance of these schemes is highly dependent on the scene, and the model needs to be adjusted for the actual application scene at the same time, which reduces the operability caused by this requirement. The rise of deep learning algorithms offers some effective ways to overcome these obstacles. Hidden Markov Models (HMMs) are a typical class of these studies, in which a bi-stochastic process combining electrical states and an explicit stochastic function is used to build the sequence of operations [10]. In addition, neural networks (NNs) can also be combined with HMMs, where the firing probability of the HMM is modeled by a Gaussian distribution representing the state of a single device and a DNN representing the state of the aggregated signal [11]. Through the above studies, the great potential of deep learning methods in NILM problems is demonstrated. The optimized deep learning algorithm has high accuracy for non-intrusive load monitoring and high adaptability to changing models, so it has better prospects than simply using mathematical model solutions for non-intrusive load monitoring.

At present, deep learning algorithms in non-intrusive home appliance monitoring learning algorithms can be roughly divided into three categories, namely supervised algorithms [12], unsupervised algorithms [13], and semi-supervised algorithms [14]. The supervised algorithm learns to build a model from the training data, infers new instances according to the model, and realizes load perception and prediction. It has the advantages of simplicity, convenience, speed, accuracy, and small storage space [12]. The supervision algorithm has simple requirements for data structure. After the neural network training is completed, the model can be used directly for analysis. At the same time, the model decomposition speed is fast and the required storage space is not large. However, when the spatial characteristics of the data are large, the logistic regression performance of the algorithm is poor, and the neural network also has shortcomings, such as underfitting or overfitting and poor self-learning ability. When the structure of the system changes and adjusts, the neural network fails. It is more likely that the model must be established and trained again. Unsupervised algorithm refers to a data processing method that classifies samples by analyzing many samples of the research object without category information. It has strong self-learning ability, and new data can be directly learned without retraining and added to the dataset; however, also has the disadvantage of low accuracy of the analysis results [15]. Research on semi-supervised regression problems is relatively limited.

Among all the deep learning methods for supervised learning that have been studied, neural networks have attracted much attention due to their excellent performance. In addition to the literature [16] on the use of neural networks to improve NILM performance, there are many research publications discussing the advantages of neural networks in enhancing NILM. Reference [17] is from the early literature on using NN to solve NILM, and it was the first to propose the research of neural network NILM and inspired related research. Reference [16] proposed a non-intrusive harmonic source identification method based on neural network. In order to be suitable for specific problems, some NN methods are modified for different NILM application problems. Reference [18] proposed an additional optimization to be embedded into the NILM formulation to form a denoising autoencoder method. Reference [19] proposed a multi-label learning method based on convolutional neural network (CNN). Reference [20] proposed a general electrical appliance recognition model based on convolutional neural network. Reference [21] studied the problem of identifying electrical loads connected to a house with a CNN-based NILM method and proposed a system capable of extracting the energy demand of each individual device. At the same time, recurrent neural network (RNN) has also received extensive attention from scholars in NILM research. Reference [22] used the RNN model to extract the appliance features of the steady-state part as model input for recognition after an event is detected. To overcome the difficulty of RNNs in learning long-term dependencies, Reference [23] proposed a 1DCNN-based method and solved the NILM problem by considering the current signal and using a long short-term memory (LSTM) neural network. Reference [24] highlighted neural NILM by building a multilayer perceptron with multiple hidden layers. Reference [25] used the simplest neural network, the Back Propagation Neural Network (BPNN), using the abrupt values of active power and the corresponding odd harmonics to achieve load identification. It can be seen from a large number of current literature publications and research that neural networks will have a high research interest for a long time in the future.

Unsupervised learning is roughly divided into eight categories, including hierarchical clustering such as K-means clustering and other clustering algorithms, Principal Component Analysis (PCA), Probabilistic Latent Semantic Analysis (PLSA), Markov Chain Monte Carlo (MCMC), Latent Semantic Analysis (LSA), Singular Value Decomposition (SVD), Latent Dirichlet Analysis (LDA), and PageRank Algorithms. Reference [26] used a hierarchical approach to cluster data objects in the form of trees called hierarchies. The results of Reference [27] showed that the GMM method outperformed other methods to a certain extent. Reference [28] described a method for computing the partial singular value decomposition of a matrix, which is suitable for problems where the matrix is known to have low rank and only the dominant singular vectors are of interest. Reference [29] proposed a rotation-invariant principal component analysis method based on the maximum correlation entropy criterion (MCC). A semi-quadratic optimization algorithm is used to calculate the relevant entropy objective. Reference [30] described that probabilistic latent semantic analysis (PLSA) can effectively capture semantic and statistical data for modeling. The K-means clustering method has the advantage of simple model among many methods, so it has been widely used. Reference [31] utilized the strong adaptability of unsupervised learning to optimize a non-intrusive load-sensing model of household electrical load using neural networks. Reference [32] subdivided data user power consumption patterns. On the basis of clustering analysis of neighbor propagation for loads according to working and rest days, the rationality judgment of equipment working state is added, and the load and its working state are decomposed from the total power combined with genetic optimization.

Aimed at the characteristics of similar load characteristics and a large number of random loads in electric vehicle charging stations, a new non-intrusive load monitoring method for electric vehicle charging stations is proposed. In the case of using the neural network to establish the load perception model of the electric vehicle charging station, the K-means algorithm is used to optimize the neural network to realize the perception of random loads. Finally, the controllability evaluation of electric vehicle charging station is realized. In other words, the proposed NILM method can adapt to a random electric vehicle charging station and can even solve random electric vehicle loads without sufficient information. This contribution fills the research gap in related fields. The detailed technical contribution can be summarized as:

• A practical and adaptive NILM model is established based on BP neural network.

• Non-supervised learning optimization is based on the proposed NILM model to improve the scalability and robustness of the method.

• The proposed NILM solution will combine the neural network model based on unsupervised learning optimization and supervision, which can determine the load of random electric vehicle while filling an important research gap.

2. Modeling of Non-Intrusive Load Monitoring Algorithms

2.1. K-Means Clustering Algorithm

The K-average algorithm is a cluster algorithm that can summarize the data type according to the two data features. The corresponding scheme is as follows. The first step of the K-average algorithm is to randomly determine the k-initial center and calculate the Euclidean distance of each power point to the initial center. The Euclidean distance referred to here is the closest distance between the two power points in the PQ two-dimensional space. Assign each PQ power point to each cluster center point class according to the calculated Euclidean distance and then calculate the average distance. This process is repeated until the squared error standard function meets the requirements [33]. The squared error is defined as:

(1)$\min E=\min {\displaystyle \sum _{i=1}^K{\displaystyle \sum _{j=1}^{t_i}\Vert x_j-m_i{\Vert }^2}}$

K is the number of cluster centers. K-means clustering is a clustering algorithm based on data spacing, operating on the principle that the closer the distance between two data points, the higher the similarity. Below is the formula $dist(m,x)$, which calculates the Euclidean distance between points $m$ and $x$:

(2)$\min {\displaystyle \sum _{i=1}^K{\displaystyle \sum _{x\in m_i}^{}dist{(m_i,x)}^2}}$

In this example, we replace the XY axis with the PQ axis, where P represents the active power of each charging pile in the electric vehicle charging station when charging, and Q represents the reactive power of each charging pile during operation. Although there are real-time changes of the electric vehicles connected to each charging pile, the operating characteristics within a certain period of time can be clustered due to the limited vehicle types.

When we design our model based on N types of vehicles for clustering, we consider how many different combinations of cluster centers can be obtained from the switching actions on M charging piles. At this time, the total number of cluster centers is:

(3)$1+N+{\displaystyle \sum _{i=1}^M{C}_M^i}$

The number of cluster centers calculated by the above formula is used as K for clustering, and the obtained cluster centers can decompose the charging characteristics of each electric vehicle. When we obtain these cluster centers, we can compare them with the characteristics of the cluster centers that were run in the previous period, identify the changing load, and finally complete the update of the training network to handle the randomness of the charging users of the electric vehicle charging station.

2.2. Neural Network Algorithms

The artificial neuron is a basic element of the neural network. The M-P model (McCulloch–Pitts model) is a threshold weighted sum model and a processing unit in the neural network. The basic principle is shown in Figure 1, also known as an artificial neuron.

As shown in the figure above, $x_1\sim x_i$ represents all the inputs of the artificial neuron by other neurons or systems, and these inputs represent the characteristics transmitted by the previous signal transmission source. Each input $x_i$ has a corresponding weight $w_i$, and the weight corresponding to each input reflects the connection between the internal output of the neuron and the external input. Neuron learning, as this is called, establishes the input from it, starting with the output link. It can be seen that the learning direction of the neural network is the direction of $x$ transmission, and $x$ is the content transmitted by the previous neuron or system to the next neuron. Then, its input matrix X can be described as:

(4)$X=[x_0,x_1,x_2,\dots ,x_n]$

The $w_i$ weight represented by each $x_i$ mentioned above is the key to the self-learning of the neural network. As we train the neural network, the weights are continually adjusted to reflect the connections among the variables. Finally, after multiple trainings, a layer-by-layer close connection is established between the input and the output. In fact, our training of neural networks is essentially multiple adjustments to the weights. The expression of the weight vector W is:

(5)$W=[w_{i0},w_{i1},w_{i2},\dots ,w_{in}]$

After we add the product of the weight and the input, the neuron also has its own threshold. When the threshold of the neuron is ${\theta }_j$, the threshold is usually called the bias. The connection between the input and output obtained by the artificial neuron is as follows:

(6)$ne{t}_i={\displaystyle \sum _{j=1}^n{w}_{ij}{x}_j-{\theta }_j}$

where $ne{t}_i$ acts as the net activation of neuron $i$ in the neural network. The output result of the weighted summation of the inputs of $i$ and the subtraction of the threshold. It can be simplified using a vector as:

(7)$ne{t}_i=XW$

By applying an activation function $f$ to the final output. The function of the activation function or transfer function mentioned here is actually to convert the net activation obtained by our previous calculation. This net activation is a numerical value obtained by calculation. Using the activation function, the numerical result can be converted into a classification result. The main categories of activation functions are sigmoid function, softmax function, tanh hyperbolic tangent function, and relu function. The sigmoid function is used mostly in pattern recognition problems in machine learning. The softmax function is used mostly in multi-classification problems. The tanh yperbolic tangent function is an odd function that ultimately converts the net activation to a value between −1 and 1. The relu function can mainly improve the running speed of the network. The final output $y_i$ can be simplified as.

(8)$y_i=f(ne{t}_i)=f(XW)$

If a neuron is said to be active, then its network activation network is positive. If the neuron is in an inhibitory state, the neuron’s net activation is negative. Many of the above neurons can form a neural network. During training, the feedforward neural network will calculate the activation value of each layer and then propagate forward to calculate the activation value of the second layer one by one, and so on, until reaching the output layer. For a three-layer feedforward neural network, X is used as the input vector of the neural network, W is used as the connection weight vector of each layer, and F is used as the activation function. Then, the expression of the output $Y_1$, $Y_2$, $Y_3$ of each layer is:

(9)$Y_1=F_1(X{W}_1)$

(10)$Y_2=F_2(F_1(X{W}_1)W_2)$

(11)$Y_3=F_3(F_2(F_1(X{W}_1)W_2)W_3)$

2.3. Neural Network Modeling for NILM

NILM problems involving electric vehicle charging stations are high-order nonlinear functions. Therefore, the activation transfer function of linear BP neural network cannot meet the requirements of non-intrusive load monitoring problem. Therefore, an appropriate nonlinear function should be chosen as the activation function. A two-layer neural network with a nonlinear function as the activation function can approximate any nonlinear function. Thus, we choose a differentiable activation function as shown in the following formula:

(12)$f(x)=\frac{1-e^{-x}}{1+e^{-x}}$

When performing load sensing and controllability assessments of electric vehicle charging stations, first, the short-term charging power data of the charging station should be obtained through the energy information terminal in the electric vehicle charging station, including the total active power, power factor, and time stamp. After the initial processing of the obtained data, the active and reactive power data are obtained, and then K-means clustering is performed to obtain the power characteristics of each cluster center. Then, the load is decomposed according to the power characteristics of each cluster center, and the decomposed load power data is normalized, and its expression is:

(13)$P_{norm}=\frac{P-P_{\min }}{P_{\max }-P_{\min }}$

where $P$ is the matrix of all measured powers, $P_{norm}$ is the normalized power matrix of power, and $P_{norm}$ and $P_{\max }$ are the minimum and maximum power in the power matrix, respectively.

Figure 2 above is a three-layer BP neural network model. When the input and output are not equal, the output error E is expressed as:

(14)$E=\frac{1}{2}{\displaystyle \sum _{k=1}^{\mathcal{l}}}{(d_k-o_k)}^2$

When the output error is transferred to the hidden layer, the error expression is:

(15)$E=\frac{1}{2}{\displaystyle \sum _{k=1}^{\mathcal{l}}[d_k-f({{\displaystyle \sum _{j=0}^n{w}_{jk}{y}_j)]}}^2}$

If it is further transferred to the input layer, it becomes an input error, and its expression is:

(16)$E=\frac{1}{2}{\displaystyle \sum _{k=1}^{\mathcal{l}}{d}_k-f[{{\displaystyle \sum _{j=0}^n{w}_{jk}f({\displaystyle \sum _{j=0}^m{v}_{ij}{x}_i})]}}^2}$

Expression (16) shows that the input error is a function related to the sum of the weights of each layer, so adjusting the weights of each layer will affect the change of the function error result E. The error is obtained by using the final output result of forward propagation, and the partial derivative of the error is calculated. This partial derivative is used to perform a weighted summation with the previous layer. After reaching the input layer, the partial derivative result calculated by each node is used to update the weights. Obviously, the principle of weight adjustment is to gradually reduce the error, so that the weight value is proportional to the gradient descent of the error. Its expression is:

(17)$\Delta w_{jk}=-\eta \frac{\partial E}{\partial w_{jk}}j=0,1,2,\dots ,n;k=1,2,\dots ,\mathcal{l}$

(18)$\Delta v_{ij}=-\eta \frac{\partial E}{\partial w_{jk}}j=0,1,2,\dots ,n;i=0,1,2,\dots ,n$

It is analogized layer by layer according to the weight adjustment formula. The weight adjustment is determined by the error signal X of the output signal of this layer, the input signal Y of this layer, and the learning rate. The difference between the expected output of the network and the actual output is expressed as the output error and feedback in the output layer.

2.4. NILM Model of Electric Vehicle Charging Station Based on Deep Learning

First, use the intelligent terminal of the electric vehicle charging station to collect the charging operation data of the electric vehicle charging station and perform further data processing on the initial data. We will use a portion of the initial data to train the deep neural network, multiply this portion of the data, and use multi-layer training to increase the dimension of the neural network to improve the depth of the neural network. Then, each charging vehicle is decomposed so as to perceive the electricity consumption laws and habits of regular electricity consumers among the charging users and then evaluate the controllability of the electric vehicle charging station. According to the weighted sum of the charging power of each user at different times and their possible electricity consumption probability, the adjustable capacity evaluation of the entire electric vehicle charging station can be obtained. When random vehicle charging occurs in the electric vehicle charging station, such as adding a vehicle or removing a vehicle, our monitoring results will have errors. Then, the charging data clustering center of each electric vehicle will be obtained by K-means clustering of its active and reactive power characteristics. After comparing the cluster center at this time with the previous cluster center, different cluster centers are found, and these cluster centers are stored as new electric vehicle data. Then, according to the characteristics of each electric vehicle charging data clustering center, we decompose each electric vehicle charging data collected from the smart terminal of the electric vehicle charging station. This decomposition scheme is to use the characteristics of the cluster centers to set an appropriate threshold, which is approximated by the gradient descent method to obtain an appropriate threshold. Finally, we filter and decompose the total data of electric vehicle charging stations according to the set threshold so as to obtain the data cluster of the newly added load. We then update the training data of deep learning so as to retrain to update the neural network model and complete the perception of the randomness of electric vehicle charging. The flow of this process is shown in Figure 3.

3. Case Analysis

3.1. Electric Vehicle Charging Station Power Feature Dataset

In real life, the active power of the DC charging pile of the electric vehicle charging station is relatively fixed. There are different powers such as 60 KW, 150 KW, and 300 KW, and when the electric vehicle is not fully charged, the shape of the power–time curve of the electric vehicle charger during charging is roughly a rectangular wave. There will be a stepped descent when the electric car is fully charged, and each brand of electric vehicle charging reactive power has its own characteristics and differences in power factor. In the analysis after this article, the charging power of most of the small car charging station charging piles on the market is kept mainly within 60–90 KW, so we chose 60 KW as the analysis object of the charging power of the charging station in this article. In addition, four charging vehicles with different reactive power characteristics are selected for follow-up research.

To discuss the effectiveness of the program, we first present a case study based on a public dataset generated from electric vehicle charging characteristics [34]. In order to avoid the excessive noise interference of the real data of electric vehicle charging stations and the safety protection of the information of domestic electric vehicle charging stations at present, we will generate charging data for the EV charging stations used in our experiments based on real EV charging characteristics. It is currently known that the frequency of the low-frequency dataset of the target electric vehicle charging station is 1 Hz.

From the above conditions, we have generated four electric vehicle charging characteristics, code-named EV1, EV2, EV3, and EV_new; there is a slight difference in the active power and reactive power among them, so as to simulate the current electric vehicles of various brands. There are differences in charging characteristics, but it is set so that the three electric vehicles have charging power fluctuations of no more than 10 W to simulate the influence of the power noise of the electric vehicle charging station on our modeling. At the same time, the charging habits of the three electric vehicle users are very different, and the random charging load existing in the electric vehicle charging station is set at the same time. The data for ten electric vehicle charging stations are set by the above rules, and the power–time curve of the electric vehicle charging station composed of them is shown in Figure 4.

In addition, in order to simulate the randomness of electric vehicles in the electric vehicle charging station, we will add new car data; the impact on the electric vehicle charging station after adding is shown in Figure 5. The random load added to the trained neural network is a strange load and can be used to detect the effect of the optimized algorithm in this article.

In order to comprehensively evaluate the properties of the research such as accuracy, this paper selects the four most important indicators in the non-intrusive load monitoring and evaluation system: recall rate, F1 score, accuracy rate, and mean absolute error as evaluation indicators. The calculation formula of each index is as follows [31]:

(19)$PRE=\frac{TP}{TP+FP}$

(20)$REC=\frac{TP}{TP+FN}$

(21)$F1=2\times \frac{PRE\times REC}{PRE+REC}$

(22)$MAE=\frac{1}{T_1-T_0}{\displaystyle \sum _{t=T_0}^{T_1}abs(\frac{{\tilde{y}}_t-y_t}{y_t}})$

TP indicates that the actual work and state of the load are consistent with the work state of the NILM analysis results and both are data sequences that are working. FP and FN represent the data sequence in which the actual work and state of the load are inconsistent with the work state of the NILM analysis result; however, in FP, the load is actually working, but the NILM analysis result is not working. In FN, the load is not actually working, but the NILM analysis results are working [31]. $y_t$ is the actual power of the EV charging station at time t, $\stackrel{\sim }{y_t}$ is the decomposition power at time t, and MAE is the mean absolute error from T₀ to T₁ [31]. PRE, REC, and F1 scores are the basic indicators of NILM and can be used to reflect the correctness of the analysis results judged by the NILM model. MAE reflects the accuracy of results across time periods. The lower the resulting value of MAE, the higher the precision of its decomposed value.

3.2. Load Perception of Electric Vehicle Charging Station Based on Deep Learning

First, after obtaining the experimental data, the depth of the neural network can be extended without multi-layering the neural network. Instead, the normalized data is directly used as the neural network training data. The resulting neural network training results are poor, as shown in Figure 6.

To sum up, the direct training of the neural network on the electric vehicle charging station data without multi-layer processing will lead to insufficient network depth. The neural network cannot get enough neuron connections and parameters. As a result, the neural network cannot perceive the operating characteristics of the electric vehicle well and cannot effectively decompose it. Therefore, it is necessary to perform multi-layer processing on the input data and expand the original two-column array into a data matrix so that the user’s electricity consumption habits are repeatedly connected to each other and increase the depth of the neural network.

After many trials, when the number of hidden layers is set to 16, the maximum number of failures is set to 10, and the learning rate is set to 0.01, the two requirements of speed and accuracy can be met. Figure 7 shows the training results of matlab’s BP neural network.

The decomposition result of EV1 is basically in line with the real operating data after filtering, indicating that the neural network can effectively detect the user’s electricity consumption habits and the charging characteristics of the vehicle. The decomposition result is shown in Figure 8.

As can be seen from the above figure, although there are still some fluctuations in the decomposition of the power data of the electric vehicle charging station by the deep neural network, its waveform can already fit the power curve in the real situation. The data of the first four days is used for training, and the data of the last four days is used for testing, which proves that the deepening of the neural network in this scheme can better realize the load perception of the electric vehicle charging station. At the same time, we use the decomposed data to calculate the precision rate, recall rate, and F1 score of each load data by using the various indicators in the evaluation system of the non-intrusive load monitoring algorithm mentioned above. The analysis results and calculation indicators are shown in the Table 1.

3.3. Electric Vehicle Charging Station Load Perception Based on Optimized Deep Learning for Newly Added Electric Vehicle Loads

The vehicles in the electric vehicle charging station are mainly divided into two categories: habitual long-term users and random short-term users. The main focus of our monitoring is habitual long-term users. For the monitoring of short-term users, we can sacrifice a little accuracy. However, if an undetected electric vehicle is added to the currently established electric vehicle charging station load model, the current model will not be able to perceive this vehicle. If the car is a random load, it will not have very much influence on the evaluation of the load perception and controllability of the electric vehicle charging station. However, if it is a habitual long-term user, our algorithm will have a big loophole. How can the parameters of the current deep learning neural network be updated for a long time to keep the model with high accuracy? Therefore, it is proposed to continually update the training set of the deep learning neural network, so that the neural network can always update the parameters in it to complete the perception of the newly added electric vehicle load.

Thus, we added a new EV to the bus based on the 10 EVs we generated earlier. The modified electric vehicle data is shown in Figure 9.

After increasing the load of EV11, in addition to the inability to perceive the current model of EV11, there will be more or less deviations in the perception of other vehicles. For example, in the perception of EV6, due to the influence of EV11, more fluctuations on the decomposed power curve may affect subsequent identification difficulties, as shown in Figure 10.

When EV11 data is added to the current EV charging station model, it will have a greater impact on the analytical capabilities of deep neural networks. Therefore, the training set is constantly updated to update the parameters in the deep neural network. First, we input the new load EV charging station bus data into the algorithm to get the initial K value in the K-means clustering algorithm. Because the active power of each charging pile in the electric vehicle charging station is the same, we can get the size of K by comparing the active and reactive power. Then, we perform K-means clustering on EV charging stations without EV11 added. Figure 11 shows the clustering results.

After the above cluster centers are obtained, K calculation and K-means clustering are also performed on the data of the electric vehicle charging station with the newly added load. The clustering results are shown in Figure 12.

The cluster center characteristics of newly added and unadded loads can be seen in the Table 2.

It can be seen from the above table that a new cluster center appears in the current electric vehicle charging station bus data after the new load is added. The appearance of this new cluster center is considered to be a new feature generated by the overlap of the newly added EV11 load with other loads during this time. Therefore, we use time comparison and power data difference to generate EV11 operating data. Figure 13 is a comparison diagram of EV11 data supplemented by this scheme and its real data.

After using the clustering and decomposition algorithm to obtain the EV11 running curve, we then use it together with the original running data as training data, update the deep neural network training set, and retrain the neural network. The number of hidden layers is set to 16, the maximum number of failures is 10, and the initial learning rate is set to 0.01. The training results are shown in Figure 14.

Then, we use the updated data neural network to analyze the newly added load EV11 and decompose the running power–time curve of EV11. Figure 15 is a comparison graph with the real power–time curve.

Although for the exploration of new loads, the accuracy of the neural network’s analysis after updating the parameters is not very high, it still distinguishes the running time of each EV11. This scheme is still very effective and can handle the randomness of users of electric vehicle charging stations. The Table 3 shows the indicators obtained by running the non-intrusive load monitoring algorithm for EV charging stations after the addition of EV11.

The results of electric vehicles after adding EV11 can be obtained from the deep neural network analysis. The advantages of the non-intrusive load monitoring algorithm proposed in this paper over the traditional neural network algorithm to analyze the non-intrusive load monitoring problem can be seen in the Figure 16, which shows the comparison of the parameters finally calculated by the three methods. It can be seen that the deep neural network is the most accurate for the perception of regular load. Although the deep neural network optimized by unsupervised learning is generally accurate for the perception of random loads, it solves the non-intrusive load monitoring algorithm’s ability to perceive random loads, and the error is also within an acceptable range.

3.4. Evaluation of Adjustable Capability of Electric Vehicle Charging Station

To assess the controllability of an EV charging station, one needs to focus on the controllability of each EV. The concept of electric vehicle controllability is proposed. Although electric vehicles are connected to the power grid when charging, and V2G interaction can be carried out at any time when the power grid is insufficient, it is necessary to consider whether the current state of electric vehicles allows V2G interaction. If the remaining power of the electric vehicle cannot meet the travel power demand of the owner, it will not be able to interact with the Internet of Vehicles. Therefore, this paper divides the vehicle into charging, idling, and discharging states according to the charging state of the vehicle entering the grid. After the previous analysis, the charging state must be maintained until the car can enter the idle state before it can be discharged, and each car owner has different requirements for entering the idle state. In addition, when entering the grid for charging, there is also a difference in the remaining power in the car. Therefore, it is necessary to analyze the average daily travel distance of users and determine the time when the electric vehicle can enter the idle state after being connected to the grid by judging the remaining power of the electric vehicle and the travel demand of the vehicle owner.

The average daily travel distance of users obeys the log-normal distribution [34], and its probability density function is:

(23)$f(d)=\frac{1}{\sqrt{2\pi }d{\sigma }_d}{e}^{-\frac{{(\ln d-{\mu }_d)}^2}{2{\sigma }_d^2}}$

where $d$ represents the average daily driving distance, and ${\mu }_d$ and ${\sigma }_d$ represent the expectation and variance of the log-normal distribution, respectively. In this paper, ${\mu }_d$ = 3 and ${\sigma }_d$ = 1.1. From this, the probability density function curve of the average daily travel distance of the user can be obtained.

After obtaining the probability density of the user’s average daily travel distance, the predicted remaining power of the current electric vehicle charging station system when the vehicle enters the power grid is set according to the probability, and we calculate the charging time required for the electric vehicle to enter the idle controllable state:

(24)$t=\frac{\beta \times d}{P}$

In the above formula, $\beta$ refers to the power consumption required to travel one kilometer, and the product of the probability density product of the average daily travel distance of electric vehicle users divided by the time is the minimum daily charging amount of this type of electric vehicle. Only electric vehicles that reach this charge amount can be regarded as a controllable idle state and can be included in the evaluation of the controllability of electric vehicle charging stations.

Using neural network to build a model can aid in analyzing the charging habits of electric vehicle charging station users and realize non-intrusive load monitoring of electric vehicle charging stations. It can break down the time each user plugs in the electric car and their charging time. The adjustable time of each electric vehicle is calculated by means of the adjustable ability model of each electric vehicle so as to realize the evaluation of the real-time power-adjustable ability of the electric vehicle charging station.

According to the average daily mileage of electric vehicle users, the average daily mileage of the car in the electric vehicle charging station was simulated according to the probability model. As a result, the charging time required for each electric vehicle is calculated, and then the power controllability is evaluated. That is, when the power grid requires, it can converge at the regulatory capacity of electric vehicles without affecting the normal use of electric vehicles. At different times per day, the number of vehicles that can be regulated at electric vehicle charging stations may different. The prediction of the control capacity of the electric vehicle charging station is shown in Figure 17.

4. Conclusions

This paper provides an idea of NILM for electric vehicle charging stations that considers the different power consumption characteristics and random loads of electric vehicle users. The neural network is deepened by using the multi-dimensional data, and the load decomposition of the electric vehicle charging station is realized according to the different power consumption characteristics of electric vehicles. When the random load is added, the existing model error is used for judgment. The random load is sensed through unsupervised learning, and the neural network model is retrained using the sensing results to optimize the neural network. Finally, the total load data of the electric vehicle charging station is decomposed, and the evaluation method of the maximum controllable capacity of the electric vehicle charging station is comprehensively designed based on the charging speed of the electric vehicle and the remaining power of the electric vehicle. On the whole, the example analysis shows that the proposed scheme is feasible for random load identification, and more accurately realizes the load perception of electric vehicle charging stations and the evaluation of the maximum controllable capacity. How to improve the efficiency of unsupervised learning needs further research.

Author Contributions

Conceptualization, S.L. and Q.X.; methodology, S.L. and J.W.; software, X.F. and J.W.; validation, G.L., S.L. and Q.X.; formal analysis, S.L.; investigation, X.F.; resources, G.L.; data curation, Q.X.; writing—original draft preparation, J.W.; writing—review and editing, J.W.; visualization, S.L.; supervision, Q.X.; project administration, Q.X.; funding acquisition, Q.X. All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

Not applicable.

Conflicts of Interest

The authors declare no conflict of interest.

Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Figures and Tables

Figure 1. Artificial neuron.

Figure 2. Three-layer BP neural network.

Figure 3. Non-intrusive load monitoring process of electric vehicle charging station based on deep learning.

Figure 4. Electric vehicle charging station power–time curve.

Figure 5. Power–time curve of newly added random load in electric vehicle charging station.

Figure 6. Decomposition result of EV1 of electric vehicle charging station with no new random load by deep neural network.

Figure 7. Deep Neural Network Training Results.

Figure 8. Decomposition result of deep neural network on EV1 of electric vehicle charging station with no new load.

Figure 9. Added load EV11 power time curve.

Figure 10. Decomposition results of deep neural network on EV6 of new load electric vehicle charging station.

Figure 11. K-means clustering results for electric vehicle charging stations without new random loads.

Figure 12. K-means clustering results of electric vehicle charging stations with newly added random loads.

Figure 13. EV11 data identified by clustering.

Figure 14. Deep Neural Network Training Results.

Figure 15. Decomposition results of new random load electric vehicle charging station EV11 by deep neural network.

Figure 16. Comparison of load perception results of electric vehicle charging stations.

Figure 17. Evaluation of power controllability of electric vehicle charging station.

Table 1

Solving results of electric vehicle charging station with no new random load.

Index	PRE	REC	F1	MAE
EV1	0.9991	0.9985	0.9988	0.0332
EV2	0.9986	0.9917	0.9951	0.1595
EV3	1	0.9786	0.9892	0.1295
EV4	0.9971	0.6705	0.8018	0.1693
EV5	0.9965	0.9957	0.9961	0.1073
EV6	0.8947	0.9989	0.9439	0.2069

Table 2

K-means cluster center.

	P (EV11 Not Added)	Q (EV11 Not Added)	P (EV11 Added)	Q (EV11 Added)
1	0	0	0	0
2	180.159	86.097	60.053	19.780
3	300.253	135.004	360.330	180.104
4	60.047	29.097	240.203	115.236
5	120.104	48.862	120.106	57.021
6	240.196	105.906	180.159	86.097
7	240.203	115.236	300.255	135.006
8	60.053	19.780	60.037	37.236
9	120.093	66.293	240.196	105.906
10	60.037	37.236	60.047	29.097
11	120.106	57.021	240.212	123.336
12	240.212	123.336	120.102	48.862
13			120.093	66.293

Table 3

NILM analysis results of EV charging stations after adding EV11.

Index	PRE	REC	F1	MAE
EV1	0.9891	0.9985	0.9988	0.0332
EV2	0.9986	0.9783	0.9890	0.1414
EV3	1	0.9773	0.9885	0.0889
EV4	0.9971	0.6033	0.7526	0.2153
EV5	0.9965	0.9947	0.9953	0.1203
EV6	0.8947	0.9972	0.9986	0.0557
EV11	1	0.5828	0.7364	0.3747

References

1. Wang, J.; Deng, J.; Liu, Y.; Wang, Y. Non-intrusive load perception and flexibility evaluation for electric vehicle charging station: A deep learning based approach. Proceedings of the 2022 IEEE 5th International Electrical and Energy Conference (CIEEC); Nangjing, China, 27–29 May 2022; pp. 2570-2575.

2. Notice of the State Council on Printing and Distributing the Action Plan for Carbon Peaking Before 2030; Gazette of the State Council of the People’s Republic of China: Beijing, China, 2021; pp. 48-58.

3. Notice of the General Office of the State Council on Printing and Distributing the New Energy Vehicle Industry Development Plan (2021–2035); Gazette of the State Council of the People’s Republic of China: Beijing, China, 2020; pp. 16-23.

4. Ebrahimi, M.; Rastegar, M.; Arefi, M.M. Real-Time Estimation Frameworks for Feeder-Level Load Disaggregation and PEVs’ Charging Behavior Characteristics Extraction. IEEE Trans. Ind. Inform.; 2022; 18, pp. 4715-4724. [DOI: https://dx.doi.org/10.1109/TII.2021.3118101]

5. Hart, G.W. Nonintrusive appliance load monitoring. Proc. IEEE; 1992; 80, pp. 1870-1891. [DOI: https://dx.doi.org/10.1109/5.192069]

6. Cox, R.; Leeb, S.B.; Shaw, S.R.; Norford, L.K. Transient event detection for nonintrusive load monitoring and demand side management using voltage distortion. Proceedings of the Twenty-First Annual IEEE Applied Power Electronics Conference and Exposition, APEC ‘06; Dallas, TX, USA, 19–23 March 2006; 7.

7. He, K.; Stankovic, L.; Liao, J.; Stankovic, V. Non-Intrusive Load Disaggregation Using Graph Signal Processing. IEEE Trans. Smart Grid; 2018; 9, pp. 1739-1747. [DOI: https://dx.doi.org/10.1109/TSG.2016.2598872]

8. Lin, Y.; Tsai, M. Non-Intrusive Load Monitoring by Novel Neuro-Fuzzy Classification Considering Uncertainties. IEEE Trans. Smart Grid; 2014; 5, pp. 2376-2384. [DOI: https://dx.doi.org/10.1109/TSG.2014.2314738]

9. Liu, S.; Liu, Y.; Gao, S.; Guo, H.; Song, T.; Jiang, W.; Li, Z.; Wang, J.; Song, Y. Non-invasive load decomposition method based on multi-feature objective function PCA-ILP. Electr. Power Constr.; 2020; 41, pp. 1-8.

10. Kolter, J.Z.; Jaakkola, T.S. Approximate inference in additive factorial HMMs with application to energy disaggregation. Proceedings of the International Conference on Artificial Intelligence and Statistics; La Palma, Spain, 21–23 April 2012.

11. Yan, X.F.; Zhai, S.P.; Wang, Z.H.; Wang, F.; He, G.Y. Application of Deep Neural Networks in Non-Intrusive Load Decomposition. Autom. Electr. Power Syst.; 2019; 43, pp. 126–132+167.

12. Liu, Q.; Kamoto, K.M.; Liu, X.; Sun, M.; Linge, N. Low-Complexity Non-Intrusive Load Monitoring Using Unsupervised Learning and Generalized Appliance Models. IEEE Trans. Consum. Electron.; 2019; 65, pp. 28-37. [DOI: https://dx.doi.org/10.1109/TCE.2019.2891160]

13. Li, D.; Dick, S. Residential Household Non-Intrusive Load Monitoring via Graph-Based Multi-Label Semi-Supervised Learning. IEEE Trans. Smart Grid; 2019; 10, pp. 4615-4627. [DOI: https://dx.doi.org/10.1109/TSG.2018.2865702]

14. Zoha, A.; Gluhak, A.; Imran, M.A.; Rajasegarar, S. Non-Intrusive Load Monitoring Approaches for Disaggregated Energy Sensing: A Survey. Sensors; 2012; 12, pp. 16838-16866. [DOI: https://dx.doi.org/10.3390/s121216838]

15. Kelly, J.; Knottenbelt, W.J. Neural NILM: Deep Neural Networks Applied to Energy Disaggregation. Proceedings of the 2nd ACM International Conference on Embedded Systems for Energy-Efficient Built Environments; Seoul, Korea, 4–5 November 2015.

16. Tan, P.N.; Steinback, M.; Kumar, V. Introduction to Data Mining. Posts & Telecom Press: Beijing, China.

17. Andrean, V.; Zhao, X.; Teshome, D.F.; Huang, T.; Lian, K. A Hybrid Method of Cascade-Filtering and Committee Decision Mechanism for Non-Intrusive Load Monitoring. IEEE Access; 2018; 6, pp. 41212-41223. [DOI: https://dx.doi.org/10.1109/ACCESS.2018.2856278]

18. Bonfigli, R.; Felicetti, A.; Principi, E.; Fagiani, M.; Squartini, S.; Piazza, F. Denoising autoencoders for Non-Intrusive Load Monitoring: Improvements and comparative evaluation. Energy Build.; 2018; 158, pp. 1461-1474. [DOI: https://dx.doi.org/10.1016/j.enbuild.2017.11.054]

19. Faustine, A.; Pereira, L. Multi-Label Learning for Appliance Recognition in NILM Using Fryze-Current Decomposition and Convolutional Neural Network. Energies; 2020; 13, 4154. [DOI: https://dx.doi.org/10.3390/en13164154]

20. Xi, X.F.; Zhou, G.D. A Survey on Deep Learning for Natural Language Processing. Acta Autom. Sin.; 2016; 42, pp. 1445-1465.

21. Monteiro, R.; Santana, J.; Teixeira, R.; Bretas, A.S.; Aguiar, R.; Poma, C. Non-intrusive load monitoring using artificial intelligence classifiers: Performance analysis of machine learning techniques. Electr. Power Syst. Res.; 2021; 198, 107347. [DOI: https://dx.doi.org/10.1016/j.epsr.2021.107347]

22. Hengyong, L.; Shuaibin, S.; Xuhui, X.U.; Dongguo, Z.; Ruolin, M.; Wenshan, H.U. A Non-Intrusive Load Identification Method Based on Correlation RNN Model. Power Syst. Prot. Control.; 2019; 47, 9.

23. Figueiredo, M.; Ribeiro, B.; Almeida, A.d. Electrical Signal Source Separation Via Nonnegative Tensor Factorization Using On Site Measurements in a Smart Home. IEEE Trans. Instrum. Meas.; 2014; 63, pp. 364-373. [DOI: https://dx.doi.org/10.1109/TIM.2013.2278596]

24. Zhou, M.; Song, X.; Tu, J.; Li, G.; Luan, K. Residential Electricity Consumption Behavior Analysis Based on Non-Intrusive Load Monitoring. Power Grid Technol.; 2018; 42, pp. 1-9.

25. Ciancetta, F.; Bucci, G.; Fiorucci, E.; Mari, S.; Fioravanti, A. A New Convolutional Neural Network-Based System for NILM Applications. IEEE Trans. Instrum. Meas.; 2021; 70, pp. 1-12. [DOI: https://dx.doi.org/10.1109/TIM.2020.3035193]

26. Nisha,; Kaur, P.J. Cluster quality based performance evaluation of hierarchical clustering method. Proceedings of the 2015 1st International Conference on Next Generation Computing Technologies (NGCT); Dehradun, India, 4–5 September 2015.

27. Mirzal, A. Statistical Analysis of Microarray Data Clustering using NMF, Spectral Clustering, Kmeans, and GMM. IEEE/ACM Trans. Comput. Biol. Bioinform.; 2022; 19, pp. 1173-1192. [DOI: https://dx.doi.org/10.1109/TCBB.2020.3025486]

28. Shlien, S. A Method for Computing the Partial Singular Value Decomposition. IEEE Trans. Pattern Anal. Mach. Intell.; 1982; PAMI-4, pp. 671-676. [DOI: https://dx.doi.org/10.1109/TPAMI.1982.4767324]

29. He, R.; Hu, B.G.; Zheng, W.S.; Kong, X.W. Robust Principal Component Analysis Based on Maximum Correntropy Criterion. IEEE Trans. Image Process.; 2011; 20, pp. 1485-1494. [DOI: https://dx.doi.org/10.1109/TIP.2010.2103949]

30. Chien, J.; Wu, M. Adaptive Bayesian Latent Semantic Analysis. IEEE Trans. Audio Speech Lang. Process.; 2008; 16, pp. 198-207. [DOI: https://dx.doi.org/10.1109/TASL.2007.909452]

31. Liu, Y.; Wang, J.; Deng, J.; Sheng, W.; Tan, P. Non-Intrusive Load Monitoring Based on Unsupervised Optimization Enhanced Neural Network Deep Learning. Front. Energy Res.; 2021; 9, 718916. [DOI: https://dx.doi.org/10.3389/fenrg.2021.718916]

32. Xu, Q.S.; Lou, O.D.; Zheng, A.X.; Liu, J.Y. A Non-Intrusive Load Decomposition Method Based on Affinity Propagation and Genetic Algorithm Optimization. Trans. China Electrotech. Soc.; 2018; 33, 11.

33. Wang, K.; Zhong, H.; Yu, N.; Xia, Q. Nonintrusive Load Monitoring based on Sequence-to-sequence Model With Attention Mechanism. Zhongguo Dianji Gongcheng Xuebao/Proc. Chin. Soc. Electr. Eng.; 2019; 39, pp. 75-83.

34. Liang, H.F.; Liu, B.; Zheng, C.; Cao, D.W.; Gao, J.Y. Research on Modeling Method of Residential Electric Vehicle Demand Response Characteristics Based on Load Identification under Smart Grid. Mod. Electr. Power; 2018; 35, pp. 1-9.