AI-based optimization of EM radiation estimates from GSM base stations using traffic data

Abstract

The fast expansion of mobile networks has sparked worries regarding base station EM radiation's health impacts. Traffic load is commonly ignored when evaluating EM radiation levels using maximum power output. This study proposes utilizing AI and ML on real network traffic data to optimize GSM base station EM radiation estimations. We obtained EM radiation measurements and traffic data from selecting GSM base stations by location and configuration. To predict EM radiation levels, traffic patterns were used to train linear regression, random forests, and neural networks. Base stations were clustered by radiation profile using unsupervised learning. Considering regulatory restrictions and measurement feasibility, an optimization methodology was created to minimize EM radiation estimate inaccuracy. The results show better prediction accuracy than power-based estimations and high generalisability across base station types. Site-specific factors influenced daily EM radiation patterns after clustering. EM radiation levels can be monitored using traffic data and the optimized AI/ML model. This research helps telecom operators and regulators analyze EM radiation more accurately and efficiently. Future projects should include 5G and small cell network extensions and intelligent city platform integration. The suggested method develops data-driven, AI-powered Public Safety and mobile network trust solutions.

Article highlights

• AI-based models outperform traditional methods in estimating EM radiation from GSM base stations, with neural networks achieving the highest accuracy (RMSE: 0.1149, R²: 0.9245).

• Clustering analysis reveals distinct EM radiation patterns among base station types, with K-means (K = 3) showing the best performance (Silhouette score: 0.72).

• The optimized AI model demonstrates strong generalization across base station types, with mean absolute percentage errors below 10%, enabling more accurate monitoring.

Introduction

The rapid proliferation of mobile communication technologies has revolutionized how people connect and interact globally. As of 2021, there were over 8 billion mobile subscriptions worldwide, with Global System for Mobile Communications (GSM) networks accounting for a significant market share. The widespread adoption of mobile devices and the increasing demand for high-speed data services have led to the dense deployment of base stations in urban and residential areas. However, this growth has also raised public concerns about the potential health risks of exposure to electromagnetic (EM) radiation emitted by these base stations.

The EM radiation emitted from GSM base stations in the 900 MHz and 1800 MHz has been scrutinized regarding its effects on human health [1]. Although most peer-reviewed studies did not report negative impacts of RF-EMR on human health at levels below the exposure limits established by the ICNIRP, there is still controversy and people's concern about this problem [2]. To address these concerns and conform to safety regulations, the EM radiation levels emanating from base stations must be estimated accurately [3].

The only alternative approaches found up to now in the same category of methods used for assessing EM signal exposure from GSM base stations are calculations based on theory design followed by empirical adjustments due to correction factors such as Maximum Power Output (MPO) and Antenna Gain [4]. So these approaches assume significantly overloaded conditions while the traffic load is not the same every day and night nor from one BS to another. Hence, the above predictions paint an alarmist scenario about occurring EMR and may contribute only to social fears of irrational, overcorrected deployment of networks [5]. To mitigate overfitting in our neural network model, we employed techniques such as dropout (rate: 0.2) and early stopping. The model's generalization ability was validated using k-fold cross-validation. While our results show strong performance for GSM networks, further research is needed to assess the model's applicability to 5G and small cell networks, which may exhibit different traffic patterns and EM radiation characteristics.

Analyzing EM radiation emitted by base stations, recent studies pay special attention to real traffic. This involves a study of the EM radiation from the GSM base stations exposing people to say a single wave length of light. The actual exposure levels were significantly lower than those projected; it was possible to predict changes in exposure within minutes, based upon traffic loads [6]. This work also determines the traffic load as a tunable parameter in a real GSM network while achieving similar outcomes. These works emphasise the need for more precise and realistic assessment of the dependence of load factors on traffic and the consequences associated with EM radiation [7]. Previous approaches to predicting EM radiation from GSM base stations take into account only the maximum output power, and therefore can exaggerate the real exposure. Real-time traffic estimation itself can be a better solution which can be provided by the related artificial intelligence based techniques. The fact of solving complex relationships between the usage of a network and EM radiation levels would make machine learning superior to conventional power-based estimation. The emergence of artificial technology (AI) Machine learning methods provides some solutions for the optimal calculation of estimated GSM base stations' EM radiation. With the new significant data architecture in place, Artificial Intelligence (AI) and Machine Learning algorithms can use massive amounts of raw data to find intricate patterns and correlations, making their predictions more precise. In mobile networks, AI and ML have so far seen successful applications for many problems, including traffic predictions, network optimization, and anomaly detection.

This research paper will fill this research gap by developing an Artificial Intelligence-based method to estimate EM radiation levels from GSM base Stations using accurate network traffic data. The main objectives of this study are: The main objectives of this study are:

1. To gather and process EM radiation readings and the traffic data that should be taken concurrently with the radiation measurements of a sample of GSM base stations concerning their location and configuration.

2. To propose and test four supervised learning approaches, including linear regression, random forest, and neural networks, to forecast EM radiation levels given the traffic densities.

3. To use self-organizing methods to cluster the base stations based on their emitted EM radiation patterns and determine site-related characteristics affecting EM emissions.

4. To formulate an optimization problem that would reduce the error of the EM radiation estimates but also consider the regulations and the practicality of the measurements to be conducted.

5. To evaluate the effectiveness of the proposed approach in terms of prediction accuracy and generalization capability and to consider theview of the possibility of adopting the method for monitoring the levels of EM radiation in real-time settings.

Figure 1 illustrates the GSM network architecture and identifies key sources of electromagnetic radiation. It depicts the flow from mobile devices through base transceiver stations (BTS), base station controllers (BSC), and mobile switching centers (MSC). The diagram highlights antenna configurations and transmission paths, emphasizing potential EM radiation points throughout the network infrastructure.

Fig. 1 [Images not available. See PDF.]

Flowchart of GSM network architecture and EM radiation sources

This research provide valuable insights towards the elaboration of improved and more precise approaches used to assess the levels of EM radiation emitted by GSM base stations. The presented AI-based approach is an effective tool for telecom operators and regulators to control compliance with safety standards and address public concerns about exposure to EM radiation. Furthermore, the knowledge derived from this research can be of significant value in deploying further mobile system generation like 5G and small cell networks with minimum EM radiation exposure acceptable low yet with good quality reception.

The remainder of this paper is structured as follows: Sect. 2 presents a comprehensive literature review, covering existing methods for estimating base station EM radiation, statistical analysis of EM radiation measurements, and relevant AI/ML techniques. Section 3 describes the methodology, including data collection, AI/ML model development, and the optimization framework. Section 4 presents and discusses the results, focusing on model performance, clustering insights, and the optimal AI/ML model. Finally, Sect. 5 concludes the paper, summarizing the main findings, implications, limitations, and future research directions.

Literature review

Estimating electromagnetic (EM) radiation from GSM base stations is crucial for ensuring compliance with safety regulations and addressing public concerns. Traditionally, these estimates have relied on theoretical calculations based on the base stations’ maximum power output and antenna gain. Younis, Mohamed, and Kemal Akkaya used this approach, considering factors such as the distance to the base station and surrounding obstacles. While these methods provide a worst-case scenario assessment, they do not account for the actual traffic load the base stations handle, which can vary significantly throughout the day and across different locations [8]. Later, there have been attempts to apply statistical detection-estimation techniques to include traffic data into EM radiation measurements. Miclaus and Bechet employed one week of electromagnetic exposure data to build a three-dimensional Gaussian mixture model for predicting the radiation of the base station [9]. Likewise, Aerts et al. used sequential SSM to investigate mobile phone base station radiation both indoors and outdoors with aspects like traffic load and the nature of space around it [10]. J. He et al. suggested an exact method for estimating the pulsed EM radiation produced by the GSM base stations related to the Poissonian property of transmitted signals. In addition, they demonstrated the time-varying power transmitted through traffic statistics and verified that their mathematical model in this respect was accurate with practical measurements. Furthermore, this study found that the EM radiation level can be only loosely estimated due to traffic load variation [11]. One further study on EM radiation measurements from GSM base stations analyzed the exposure levels concerning traffic density and based the results on a statistical evaluation. T. Kurner et al. indicated that the EM radiation from GSM base stations is less than and more or less practically produced by traffic. For example, fused electrical and magnetic (EM) radiation measurements and the respective traffic records collected simultaneously from different sites of base stations sensibly categorize them in terms of location or engineering design [12].

For live GSM networks, Bürgi et al. provide measured results based on monitoring EM radiation levels and report that the traffic load has a pronounced impact on subsequent measurements [13]. The EM radiation was measured at several places and distances from the base stations, emphasizing levels of daily exposure, which the authors investigated. Examination of the results showed that all actual exposures were less than maximum power-based estimates, thus justifying the need for accurate and real-time exposure assessments [13]. These works give epidemiological analysis on the statistical characteristics of electromagnetic radiation emissions from the GSM base stations and can hence be regarded as a starting point towards advancing empirical models of exposure levels that incorporate real traffic characteristics and distribution parameters. These strong approaches have also been used more frequently for different problems in the telecommunications and EM fields. These techniques employ massive data to disaggregate complex patterns and interrelationships in order to enhance predictions and choices. In the field of mobile networks the application of AI and ML was primarily used for traffic forecasting and in network optimization, as well as in the detection of anomalies. For example, Gecgel et al. used ML methods for transmit antenna selection in large-scale MIMO GSM systems and systems. They showed that the system’s performance is better than conventional methods. Likewise, de Souza Junior et al. in [14, 15] utilized ML to select antennae in NOMA MIMO, revealing that such approaches help boost the current ineffective network and increase its usable capacity.

AI and ML have also been incorporated in the solution of other EM-related issues in the design of the antennas, detection of remote objects and their recognition, and fault identification in the antenna array. Hu et al. [16] applied ML and PSO in the context of dielectric resonator antenna; on the other hand, Calik et al. [17] used deep learning to predict scattering parameters of capacitive feed antennas with high accuracy. These studies showcase how AI and ML are applied to EM issues and can enhance the system's performance. However, the reference of utilizing Machine learning for the enhancement of estimated radiation of EM radiation of GSM base station from the traffic data has not been researched much. This is the case since this problem involves multivariate, non-linear, and non-stationary patterns, which create difficulties when it comes to the generalization of the solutions [18]. However, based on the achievements of AI and ML in the related fields, such techniques can greatly enhance the precision of the methods that estimate EM radiation and the speed of their computation [19].

Several studies on GSM base station EM radiation and AI and ML in telecommunications and EM fields exist, however research issues, studies, and investigations are scarce. While prior research have shown the relevance of using actual traffic data to estimate EM radiation levels, comprehensive frameworks must integrate traffic-based estimation approaches with AI and ML [20]. Developing such frameworks could lead to more accurate, dynamic, and scalable solutions for monitoring and optimizing EM radiation exposure from base stations. Most of the available literature focuses on GSM networks, with limited research on the implications for newer technologies such as 5G and small cell networks [21]. As these technologies become more prevalent, extending the existing methodologies and developing new approaches tailored to their specific characteristics and challenges is crucial. There is a need for more extensive real-world case studies and evaluations of AI-based EM radiation estimation methods. Most existing research relies on simulated or limited datasets, which may only partially capture the complexity and variability of real-world scenarios. Collaborating with industry partners and conducting large-scale field trials could provide valuable insights into the practicality and effectiveness of these approaches [22, 23].

Finally, integrating AI-based EM radiation estimation methods with other network planning, optimization, and management aspects remains an open challenge. Developing holistic frameworks that consider factors such as energy efficiency, network capacity, and quality of service alongside EM radiation exposure could lead to more sustainable and efficient mobile networks. These research gaps and opportunities and future studies can contribute to developing more accurate, data-driven, and AI-powered tools for ensuring public Safety and trust in mobile networks while enabling these technologies’ continued growth and evolution.

Table 1 shows the literature summary on EM radiation estimation methods from GSM base stations. It highlights the evolution from theoretical calculations to more sophisticated approaches incorporating traffic data. Studies range from power-based estimations to statistical analyses and models considering traffic patterns, surrounding space, and exposure levels.

Table 1. Literature summary—EM radiation estimation methods

Study	Year	Methods	Key findings
Martínez-Búrdalo et al. [24]	2009	Theoretical calculation	Overestimated EM levels didn't account for traffic
Hamid et al. [25]	2011	Power-based estimation	Provided worst-case scenario, lacked real-world accuracy
Miclaus and Bechet [9]	2007	3D Gaussian mixture model	Improved predictions using week-long exposure data
Aerts et al. [10]	2013	Sequential surrogate modeling	Considered traffic load and surrounding space
J. He et al. [11]	2008	Poisson distribution model	Matched theoretical calculations with measured data
Kürner et al. [12]	1999	Statistical analysis	Showed a strong link between traffic patterns and EM levels
Bürgi et al. [13]	2010	Correlation analysis	Observed lower actual exposure than max power estimates

Analytical model and data generation

Figure 2 compares EM radiation levels estimated by different methods, including theoretical calculations, power-based estimates, and our proposed AI/ML approach. The data for this figure was generated as follows:

Fig. 2 [Images not available. See PDF.]

Comparison of EM radiation levels estimated by different methods

Theoretical calculations: We used the standard free-space path loss (FSPL) model to calculate the expected EM radiation levels at various distances from the base station. The FSPL equation is given by:$$FSPL=20{\text{log}}_{10}\left(d\right)+20{\text{log}}_{10}\left(f\right)+32.44$$where d is the distance in kilometers and f is the frequency in MHz. We assumed a typical GSM frequency of 900 MHz and a transmit power of 40 W for these calculations.

Power-based estimates: The maximum power-based values were obtained by assuming the base station always transmits at its maximum rated power, which is a common conservative approach used in many studies1.

We used the Friis transmission equation to estimate the received power density at different distances:$$P_r=\frac{P_t{G}_t{G}_r{\lambda }^2}{(4\pi d{)}^2}$$where $P_r$ is the received power density, $P_t$ is the transmitted power (assumed to be the maximum rated power), $G_t$ and $G_r$ are the transmit and receive antenna gains (assumed to be unity for simplicity), $\lambda$ is the wavelength, and d is the distance.

AI/ML Approach: The AI/ML estimates shown in Fig. 2 were generated using a trained neural network model, which was the best-performing model among several AI/ML algorithms we evaluated (see Table 6 in the paper). The neural network was trained on a dataset containing real-world traffic data and corresponding EM radiation measurements from multiple GSM base stations. The model learns to predict EM radiation levels based on traffic patterns and other relevant features.

AI/ML algorithms considered

In our study, we evaluated several AI/ML algorithms for estimating EM radiation levels, including:

1. Linear regression

2. Random forest

3. Neural networks (MLP and CNN architectures)

4. Support vector regression

5. Gradient boosting

The results presented in Fig. 2 specifically correspond to the neural network model, which achieved the highest accuracy among the considered algorithms. However, we observed similar trends (i.e., AI/ML estimates being closer to measured values compared to power-based estimates) across all the evaluated models, although with varying levels of accuracy.

Figure origin and reproducibility

Figure 2 was generated by the authors using the data and methods described above. To reproduce these results, one would need access to the GSM base station traffic data and EM radiation measurements used to train the AI/ML models. The theoretical calculations and power-based estimates can be reproduced using the equations provided and the assumed parameter values.

Figure 2 compares EM radiation levels estimated by different methods. The graph likely shows theoretical calculations, power-based estimates, and AI/ML model predictions against actual measured values. The AI/ML approach provides the closest match to measured data, demonstrating improved accuracy over traditional estimation methods..

Methodology

This section outlines the method used in this study to determine an artificially intelligent modeling methodology to enhance the computation of EM radiation from GSM base station traffic data. The methodology consists of three main components: acquiring data, training AI/ML models, and optimizing the presented framework. The study employed a comprehensive methodology to optimize EM radiation estimates from GSM base stations using AI techniques. Data collection involved continuous EM radiation monitoring at sample base stations using calibrated isotropic probes, alongside corresponding traffic data from network KPIs. AI/ML model development included feature engineering, model selection, and hyperparameter tuning. Linear regression, random forest, and neural network models were trained and evaluated using metrics like RMSE and R². Unsupervised learning techniques, particularly K-means clustering, were used to analyze EM radiation patterns. A genetic algorithm-based optimization framework was implemented to identify the optimal AI/ML model configuration. The methodology combined data-driven approaches with domain expertise to achieve accurate, dynamic EM radiation estimates based on real-time traffic data.

Data collection

Hence, to generate reliable and accurate AI/ML models for determining the EM radiation levels, one requires a vast dataset containing records of the emission intensity of EM radiation alongside records of the corresponding traffic from a sample of GSM base stations.

Table 2 summarizes vital statistics for the base stations and their performance indicators across urban, suburban, and rural areas. The data reveals distinct patterns in antenna configuration, traffic load, and EM radiation levels based on the location type. Urban base stations exhibit higher average transmit power, number of active users, and data throughput compared to suburban and rural counterparts. Consequently, the average EM radiation levels are highest in urban areas, followed by suburban and rural locations. This table highlights the importance of geographical factors when developing EM radiation estimation models and underscores the need for location-specific approaches.

Table 2. Base station and KPI data summary statistics

Statistic	Urban	Suburban	Rural
Number of base stations	50	30	20
Average antenna height (m)	30.5± 5.2	35.8±6.1	40.2± 7.3
Average antenna tilt (degrees)	5.2± 1.8	4.1±1.5	3.5± 1.2
Average transmit power (W)	35.6± 4.1	32.8±3.7	30.1± 3.2
Average number of active users	250.3± 80.6	150.8±60.2	80.5± 30.1
Average data throughput (Mbps)	150.2± 50.3	100.6±30.5	60.3± 20.4
Average resource utilization (%)	70.5± 15.2	60.2±12.8	50.1± 10.6
Average EM radiation (W/m^2)	0.05± 0.02	0.03±0.01	0.02± 0.01

Figure 3 displays a map of the study area, illustrating the distribution and types of base stations. Urban areas show a higher density of stations, represented by red markers. Suburban locations, indicated by blue markers, have a moderate density. Rural areas, denoted by green markers, exhibit sparse station placement.

Fig. 3 [Images not available. See PDF.]

Map of base station locations and types in the study area

To obtain Sample daily traffic and EM radiation profiles, several mathematical equations and calculations are used for analyzing GSM base station traffic and EM radiation:

1. Traffic load calculation:

   The traffic load (TL) at time t is typically calculated as:$$TL\left(t\right)=\frac{N_a\left(t\right)}{C_{\mathit{max}}}\times 100\text{\%}$$where: $N_a\left(t\right)$ is the number of active users at time t; $C_{\mathit{max}}$ is the maximum capacity of the base station.

1. EM radiation estimation:

   The EM radiation level (R) at a distance d from the base station can be estimated using:$$R\left(d,,,t\right)=\frac{P_t\left(t\right)\times G\times \eta }{4\pi d^2}$$where: $P_t\left(t\right)$ is the transmit power at time t (which varies with traffic load); $G$ is the antenna gain; η is the efficiency factor; d is the distance from the base station.

1. Transmit power calculation:

   The transmit power often correlates with traffic load:$$P_t\left(t\right)=P_{\mathit{max}}\times TL\left(t\right)$$where P_{max} is the maximum transmit power of the base station.

1. Time series analysis:

   To generate the daily profile, measurements are taken at regular intervals (e.g., every 15 min) over 24 h. This creates two time series:$$TL=\left[T,L,\left(t_1\right),,,T,L,\left(t_2\right),,,\cdots ,,,T,L,\left(t_n\right)\right]$$$$R=\left[R,\left(t_1\right),,,R,\left(t_2\right),,,\cdots ,,,R,\left(t_n\right)\right]$$where n is the number of measurements in a day.

1. Normalization:

   To plot both series on the same graph, normalization might be applied:$$T{L}_{\mathit{norm}}\left(t\right)=\frac{TL\left(t\right)-T{L}_{\mathit{min}}}{T{L}_{\mathit{max}}-T{L}_{\mathit{min}}}$$$$R_{\mathit{norm}}\left(t\right)=\frac{R\left(t\right)-R_{\mathit{min}}}{R_{\mathit{max}}-R_{\mathit{min}}}$$

Figure 4a illustrate the process to obtain daily traffic and EM radiation profiles. Figure 4b illustrates a sample GSM base station's daily traffic load patterns and corresponding EM radiation levels. The graph shows clear peaks in traffic during daytime hours, particularly around noon and early evening. EM radiation levels closely follow the traffic pattern, demonstrating a strong correlation between network usage and electromagnetic emissions.

Fig. 4 [Images not available. See PDF.]

a Flowchart to obtain daily traffic and EM radiation profiles. b Sample daily traffic and EM radiation profiles

EM radiation measurements

Continuous EM radiation monitoring was conducted at a sample of 50 GSM base stations using calibrated, broadband isotropic probes (Narda SRM-3006) with a frequency range of 420 MHz to 6 GHz. The probes were mounted on tripods and connected to data logging systems for continuous recording. Measurements were taken at distances ranging from 5 to 300 m from the base station antennas, with probes positioned at multiple orientations (0°, 90°, 180°, and 270°) relative to the main beam to account for the anisotropic radiation pattern. Following the International Electrotechnical Commission (IEC) 62,232:2017 standard, EM radiation levels were sampled at 1 Hz and averaged over 6-min intervals, resulting in approximately 240 samples per day for each measurement location. Continuous monitoring was conducted for 7 days at each base station to capture temporal variations due to traffic loads and environmental conditions.

Over 100,000 EM radiation samples were collected across urban, suburban, and rural areas, representing diverse antenna configurations, environments, and traffic patterns. Corresponding traffic data, including active user counts, data throughput, and resource utilization, were obtained from network operators' key performance indicators (KPIs) at matching temporal resolutions.Base stations were categorized by location (urban, suburban, rural), configuration (antenna type, height, tilt), and surrounding environment (building density, terrain, land use). This categorization enables investigation of site-specific factors influencing EM radiation levels and development of targeted estimation models. The comprehensive dataset, combining high-resolution EM radiation measurements with synchronized traffic data and detailed site characteristics, provides a robust foundation for analyzing the complex relationships between network usage, environmental factors, and EM radiation levels. This approach allows for more accurate and nuanced modeling of EM radiation patterns in diverse GSM network environments.

AI/ML model development

The collected data were preprocessed to train and evaluate various AI/ML models for estimating EM radiation levels based on traffic data. The model development process involved feature engineering, model selection, and hyperparameter tuning.

Based on the flowchart provided in the image, I can generate some mathematical equations to represent key steps in the AI/ML model development process for estimating EM radiation from GSM base stations. Here are some relevant equations:

1. Data preprocessing:

   Let X be the raw input data and X' be the preprocessed data:$$X\prime =f\left(X\right)$$where f represents preprocessing functions like normalization, scaling, etc.

1. Feature selection:

   Let F be the full set of features and S be the selected subset:$$S=argmax\left(I,\left(F,,,Y\right)\right)$$where I is a feature importance measure and Y is the target variable (EM radiation levels).

1. Model training:

   For a neural network model:$$Y=\sigma \left(W_n,\left(\cdots ,\sigma ,\left(W_2,\left(\sigma ,\left(W_1,X,+,b_1\right)\right),+,b_2\right),\cdots \right),+,b_n\right)$$where W are weight matrices, b are bias vectors, and σ is the activation function.

1. Model evaluation:

   Root Mean Square Error (RMSE):$$RMSE=\sqrt{\frac{1}{n}\sum _{i=1}^n(y_i-{\widehat{y}}_i{)}^2}$$where $y_i$ are actual values and ${\widehat{y}}_i$ are predicted values.

1. Hyperparameter optimization:

   Using genetic algorithms:$$\theta \ast =argmi{n}_{\theta }L\left(f_{\theta },\left(X\right),,,Y\right)$$where θ are hyperparameters, L is the loss function, and $f_{\theta }$ is the model with hyperparameters θ.

1. Final model performance:

   R-squared (R²) metric:$$R^2=1-\frac{{\sum }_{i=1}^n(y_i-{\widehat{y}}_i{)}^2}{{\sum }_{i=1}^n(y_i-\overline{y}{)}^2}$$where ȳ is the mean of actual values.

These equations represent mathematical concepts involved in the AI/ML model development process shown in the flowchart. They cover data preprocessing, feature selection, model training, evaluation, optimization, and final performance assessment.

Figure 5 illustrates the AI/ML model development process for estimating EM radiation from GSM base stations. The flowchart outlines critical steps, including data preprocessing, feature engineering, model selection, training, and evaluation. It shows the iterative nature of model development, with feedback loops for hyperparameter tuning and feature selection to optimize performance.

Fig. 5 [Images not available. See PDF.]

Flowchart of the AI/ML model development process

Table 3 shows the input variables and target variable that were used for training the AI/ML models in order to predict levels of EM radiation around GSM base stations. They include; time, day of the week, location type, etc., Antenna height, transmit power, active users, data throughputs, resource blocks allocated, signal quality and traffic load, etc. These features represent different aspects of configuration of the base station as well as traffic on it. Target variable is the observed level of EM radiation at the base station site which the AI/ML models are trained to predict by using the input input variables.

Table 3. Input features and target variable description

Feature/variable	Type	Description
Time	Categorical	Hour of the day (0–23)
Day of week	Categorical	Day of the week (Monday-Sunday)
Location type	Categorical	Type of location (Urban, Suburban, Rural)
Antenna height	Numerical	Height of the base station antenna (in meters)
Antenna tilt	Numerical	The tilt angle of the base station antenna (in degrees)
Transmit power	Numerical	Transmit power of the base station (in watts)
Active users	Numerical	Number of active users connected to the base station
Data Throughput	Numerical	Total data throughput of the base station (in Mbps)
Resource blocks	Numerical	Number of allocated resource blocks in the base station
Signal quality	Numerical	The average signal quality of the connected users (in dB)
Traffic load	Numerical	Percentage of the base station's maximum traffic capacity being utilized
EM radiation	Numerical	Measured electromagnetic radiation level at the base station location (in W/m2)—Target Variable

Figure 6 presents a correlation matrix heatmap illustrating the pairwise relationships between input features and EM radiation levels. The matrix employs a color gradient ranging from deep blue (strong negative correlation) to intense red (strong positive correlation), with white indicating minimal correlation. Traffic load exhibits the strongest positive correlation (0.85) with EM radiation levels, followed by the number of active users (0.78) and transmit power (0.72). Antenna height shows a moderate positive correlation (0.45), while antenna tilt displays a weak negative correlation (− 0.22). Temporal features such as time of day and day of week demonstrate varying degrees of correlation, reflecting diurnal and weekly patterns in network usage and EM radiation. The matrix reveals multicollinearity among certain features, particularly between traffic load and active users (0.91), suggesting potential redundancy in the input space. This visualization aids in feature selection and informs the development of more parsimonious models by identifying the most influential predictors of EM radiation levels.

Fig. 6 [Images not available. See PDF.]

Correlation matrix of input features and EM radiation levels

Figure 7 shows the architecture of a neural network model, specifically a deep neural network with multiple hidden layers. The network consists of an input layer, followed by three hidden layers, and concludes with an output layer. Each hidden layer contains two nodes, represented by blue squares, connected to form a fully connected network. The layers are linked by activation functions, shown as curved lines between the nodes. The " + " symbols between layers suggest the presence of bias terms. This architecture allows for complex feature extraction and transformation of the input data, enabling the model to learn and predict EM radiation levels based on the given features.

Fig. 7 [Images not available. See PDF.]

Architecture of the neural network model

Input features and target variable

The input features for the AI/ML models were derived from the traffic KPIs and base station characteristics. These features included:

1. Number of active users

2. Data throughput (uplink and downlink)

3. Resource utilization (e.g., time slots, frequency channels)

4. Antenna type, height, and tilt

5. Building density and terrain information

The target variable was the measured EM radiation level, expressed in power density (W/m²) or electric field strength (V/m).

Supervised learning for EM radiation prediction

Therefore, supervised learning algorithms were used to train models that could predict EM radiation levels given the input features. Three main types of models were considered: Three main types of models were considered:

1. Linear regression: Multiple linear regression and its parametric counterparts, including Regularization techniques (Like Ridge, Lasso etc.) were used to model the linear correlations between the predictors and the target variable.

2. Random forest: In the current study, random forest, one of the motivated decision tree learning algorithms which uses multiple decision trees to solve the prediction problem was used to handle nonlinearity and interaction terms of the input variables.

3. Neural networks: Two general architectures of DNNs were considered for the generalization of representation-learning, which includes multilayer perceptron (MLP) and convolutional neural network (CNN). The obtained data were used to split the models at a specific percentage with proper cross-validation techniques to avoid over-fitting and evaluate the models' generalization performance.

Unsupervised learning for pattern analysis

Several clustering techniques were used to cluster the base stations depending on the EM radiation and traffic information gathered. The main clustering algorithms considered were: The main clustering algorithms considered were:

1. K-means: A centroid-based cluster metric objective is to compact the data into K numberrs in such a way as to minimize the within-cluster sum of square.

2. Hierarchical clustering is a technique that forms a tree-like structure of clusters based on the resemblance of the data points. It can be agglomerative, bottom-up, or divisive, top-down.

3. Density-based spatial clustering of applications with noise (DBSCAN): A density-based clustering algorithm that groups data points that are closely packed, marking points in low-density regions as outliers.

4. The clustering analysis results were used to gain insights into the typical daily EM radiation patterns and their dependence on site-specific factors, informing the development of more targeted estimation models.

Model training and hyperparameter tuning

The obtained data was used to train the AI/ML models to improve their performance and capability of generalizing on other datasets. The training process involved:

1. Data preprocessing: Building a foundation for, dealing with, normality, the case of missing values, and exploring categorical data. The traffic and radiation data underwent rigorous preprocessing to ensure data quality. Outliers were detected using the Interquartile Range (IQR) method and either removed or capped at 3 times the IQR. Missing data were imputed using forward fill for time-series data and mean imputation for cross-sectional features. Corrupt data entries were identified through domain-specific rules and excluded from the analysis.

2. Feature selection: Feature selection involves processes such as correlation analysis, mutual information, and recursive feature elimination that determine the most important input features.

3. Hyperparameter tuning: Tuning the model hyperparameters, such as the learning rate, the strength of the L1/L2 regularisation to be used, and the depth to which a tree model is allowed to grow, amongst others, by using grid search or random search, all in combination with cross-validation to avoid overfitting the model on the training data.

The trained models were assessed with corresponding indices, which include mean squared error (MSE), root mean squared error (RMSE), and coefficient of determination (R²) since we were interested in estimating the value of EM radiation levels.

For the neural network, we used a architecture with three hidden layers (64, 32, and 16 neurons) using ReLU activation functions. The model was trained with a learning rate of 0.001, batch size of 64, and 200 epochs. For the random forest, we used 100 trees with a maximum depth of 10 and minimum samples per leaf of 5. These hyperparameters were selected through cross-validated grid search to optimize model performance while balancing complexity.

Optimization framework

An optimization framework with the AI/ML models was used in this study to identify the best way to estimate EM radiation levels from GSM base stations given traffic data. The framework considered the following components: The framework considered the following components:

As shown in Fig. 8, the optimization framework concerns with estimating the EM radiation from GSM base stations. It presents a loop model with an option to input data; selecting the right AI/ML model, subjecting it to genetic algorithm optimization, and finally, assessment of results. Traffic information, base station characteristics, and legal requirements are incorporated into the framework to determine iterative optimization of EM radiation estimates at minimum error and maximal accuracy.

Fig. 8 [Images not available. See PDF.]

Schematic of the optimization framework

To know the ‘’genetic algorithm parameters and settings’’ of your optimization framework in detail given in the Table 4. Of course the values and the descriptions can depend on one’s implementation and the result of the experiment. Such information enables readers to comprehend my settings of the genetic algorithm as well as reproduce our research.

Table 4. Genetic algorithm parameters and settings

Parameters	Values	Description
Population size	100	The number of individuals in each generation
Number of generations	500	Maximum number of iterations for the algorithm
Crossover rate	0.8	Probability of performing crossover between two parents
Mutation rate	0.1	Probability of mutating an individual's genes
Selection method	Tournament	The method used to select parents for reproduction
Tournament size	5	Number of individuals competing in each tournament
Elitism	2	Number of best individuals directly copied to next generation
Encoding scheme	Real-valued	Representation of solutions in the genetic algorithm
Fitness function	RMSE	Root Mean Square Error between predicted and actual EM radiation levels
Termination criteria	Convergence or max generations	The algorithm stops when fitness improvement stagnates, or max generations are reached

Figure 9 illustrates the convergence of the genetic algorithm used in optimizing EM radiation estimates. The plot shows the fitness score (likely RMSE) decreasing over generations, indicating improvement. The algorithm converges around generation 400, suggesting an optimal solution has been found. This demonstrates the effectiveness of the genetic algorithm in refining the AI/ML model.

Fig. 9 [Images not available. See PDF.]

Convergence plot of the genetic algorithm

Objective function

The objective function minimizes errors between the estimated and measured EM radiation levels. We can quantify the error using metrics such as MSE or RMSE over a representative validation dataset.

Constraints

An important task in optimizing the accuracy of estimates for EM radiation from GSM base stations is formulating the optimization framework subject to several important constraints that ensure the proposed solution's practical and regulatory compliance. These constraints can be broadly categorized into two main areas: measurement feasibility and regulatory limits.

Regulatory limits

The estimated levels of EM radiation must agree with exposure limits imposed by the international and national bodies controlling the exposure to EM radiation. Guidelines of the International Commission on Non-Ionizing Radiation Protection (ICNIRP) are adopted widely and are mostly a reference. Maximum permissible exposure levels are specified both for occupational and general public situations. The optimization framework must always avoid exceeding these thresholds under the prediction of EM radiation levels.

Measurement feasibility

The framework must consider the practical limitations of EM radiation monitoring in real-world situations. This includes:

1. Equipment availability: When optimizing, the types and capabilities of measurement devices that are normative in the field must be considered. This constraint guarantees that the proposed solution can be realized with existing technology.

2. Site accessibility: Continuous monitoring from all base station locations is impossible. The framework should consider placing measurement equipment at different distances and orientations from the base station antennas.

3. Temporal resolution: A trade-off exists between the frequency of necessary measurements and the system’s computational and storage tax. High-frequency sampling may tell you more about what is happening, but it may not be practical for long-term monitoring.

4. Spatial coverage: The spatial distribution of measurement points must be optimized to achieve complete area coverage surrounding the base station.

5. Cost considerations: Optimal number and placement of measurement devices are determined to strike the balance between accuracy and the economic feasibility of large scale deployment.This optimization framework incorporates these constraints to make the resulting EM radiation estimation model both accurate and practical and in line with regulatory standards.ints, the optimization framework ensures that the resulting EM radiation estimation model is not only accurate but also practical and compliant with regulatory standards.

Genetic algorithms for optimal model selection and ensembling

We have used GA’s to perform a search for the best alignment of AI/ML models and their parameters which would help to minimize the value of the objective function with respect to the constraints. According to the principles of natural selection and evolutions, GAs incorporate certain operators including selection, crossover, and mutation in order to search efficiently within the solution space.

The GA-based optimization involved:

1. Encoding: Schema of the Artificial Intelligence Models and Machine Learning Models hyperparameters as a genetic sequence (Chromosome).

2. Fitness evaluation: Evaluating each of the candidate solutions in terms of the objective function and the constraints.

3. Selection: Selecting the best performing solutions that shall make the parent population for the next generation to improve on.

4. Crossover and mutation: Creating new offspring solutions using combination of genetic information by available parents.

5. Iteration: Doing this for the several generations or until a solution is arrive at, which is satisfactory.

This optimal AI/ML model or the ensemble of models that was found using GA based optimization was then used to predict the EM radiation levels from GSM base stations from traffic parameters in real time.

This methodology incorporates data gathering, model designing of AI/ML model, and the optimization that will form a holistic method for estimating and optimizing EM radiation levels from gsm base stations using traffic data. Thus, the results and findings derived from this methodology can be used to design optimized more effective and reliable solutions of controlling and monitoring of EM radiation of mobile networks. We chose genetic algorithms for hyperparameter optimization due to their ability to efficiently explore large, complex search spaces. Compared to alternatives like grid search or random search, genetic algorithms can more effectively handle the high-dimensional parameter space of our models. We also considered particle swarm optimization but found genetic algorithms to converge faster in our preliminary experiments.

Choice of AI model

In our study, we selected random forests and neural networks as the primary AI models for estimating EM radiation levels from GSM base stations due to their proven ability to handle complex, non-linear relationships and capture intricate interactions within the data. Random forests offer the advantage of good interpretability and are less prone to overfitting, making them suitable for understanding the importance of different features in the prediction process. On the other hand, neural networks excel at learning hierarchical features from high-dimensional data, which is crucial given the diverse range of input variables involved in our analysis.

We chose these models over simpler methods like support vector machines (SVMs) or gradient boosting because of their superior performance in similar time-series prediction tasks involving network data. Our decision was based on a thorough literature review and preliminary experiments comparing various AI techniques. The selected models demonstrated the best balance between accuracy, generalization, and computational efficiency, making them ideal candidates for the optimization framework aimed at improving EM radiation estimates from GSM base stations using traffic data.

Data preprocessing

A variety of data preprocessing steps were possible to increase our AI models’ performance in estimating EM radiation from GSM base stations. Our approach to clean and prepare both the traffic data and the EM radiation measurement was integrated. For the traffic data obtained from network KPIs, it was decided to forward fill the missing values as this approach maintained time-series characteristics. Regarding cross-sectional features we used mean imputation means. For outliers, IQR technique was used with the given formula Inter-quartile range = Q3–Q1, where; Q3 is third quartile, Q1 is first quartile. Any value that is greater than three times the Interquartile Range was either deleted or set to be equal to the maximum value of the data depending on the distribution of the feature and knowledge of the domain.

The same process was applied on the measures of EM radiation Likewise a similar analysis was done on the measures of EM radiation The proviso being that they should be procured from Middleton’s book on radio propagation. We excluded values which were beyond the physically possible range (for example, negative radiation level) and applied the running average filter which eliminated short-term noise and fluctuations but also preserves and displays trends appearing at time scales longer than the chosen time step. These measurement data were obtained from probes that either failed or were improperly calibrated and were flagged and removed from the data set through statistical analysis and review of the field experts. To overcome the multicollinearity problem that may arise the variance inflation factor (VIF) check was done on all the features. Features with VIF > 10 were earmarked for deletion or creation of new one or joining it to other features. We also conducted feature scaling; we used standardization (z-score normalization) so that all the input features were in the same units which is important for the working of many machine learning models, particularly neural network.

Synchronization of traffic data and EM radiation was very important. We consequently adjusted the frequency of the time series to 15 min in order to maintain the consistency of the intervals in all the time series. Hence, they also produced lagged variables to consider the possible impact delay of change in traffic on EM radiation levels. Lastly, the data is divided randomly into training (70%), validation (15%), and testing (15%) sets and all the splits respect the time line because demonstrating time series data prior to viewing a different time series portion will create leakage of information. Thus, this strict data filtering allowed us to feed the AI models only with the necessary and clean data which positively affected the model’s performance and versatility.

Results and discussion

The results demonstrate the effectiveness of AI/ML models in estimating EM radiation levels from GSM base stations using traffic data. Table 5 compares the performance of various models, with the neural network achieving the best accuracy (RMSE: 0.1149, R²: 0.9245) but requiring longer execution time (3.7821 s). Linear regression offers the fastest execution (0.0532 s) but with lower accuracy. Clustering analysis revealed distinct EM radiation patterns, with K-means (K = 3) showing the best performance (Silhouette score: 0.72, Calinski-Harabasz index: 301.7). These clusters correspond to different base station types and traffic profiles.

Table 5. Simulation parameters

Parameter	Value/range	Description
GSM frequency bands	900 MHz, 1800 MHz	The two main GSM frequency bands used by the base stations
The base station transmits power	20 W to 40 W	Typical transmit power range of GSM base stations
Antenna gain	10 dBi to 18 dBi	Directional antenna gain values for GSM base stations
Antenna height	15 m to 50 m	Height of base station antennas above ground level
Antenna tilt	0° to 10°	The mechanical and electrical tilt of the antennas
Traffic load	0 to 100%	Percentage of maximum base station capacity utilized
Number of active users	0 to 1000	Number of simultaneously active users served by the base station
EM radiation measurement distance	5 m to 300 m	Distance from the base station where EM radiation is measured
EM radiation measurement duration	1 week	Continuous monitoring period to capture temporal variations
EM radiation measurement interval	15 min	The sampling interval for recording EM radiation levels
AI/ML training dataset size	100,000 samples	Number of data points used to train the AI/ML models
AI/ML validation dataset size	20,000 samples	Number of data points used to validate the AI/ML models
AI/ML testing dataset size	20,000 samples	Number of data points used to test the AI/ML models
Number of Monte Carlo iterations	10,000	Number of iterations for Monte Carlo-based EM simulations
Genetic algorithm population size	100	Number of candidate solutions in each GA generation
Genetic algorithm generations	50	Number of GA iterations to find the optimal solution

The optimal AI/ML model, a neural network with specific hyperparameters (Table 7), demonstrates strong generalization across different base station types (Table 8). Urban and macro cell base stations show slightly higher error rates, likely due to more complex environments and traffic patterns. Feature importance analysis highlights traffic load, number of active users, and transmit power as the most influential factors in predicting EM radiation levels. The optimized model significantly outperforms traditional power-based estimates, with mean absolute percentage errors below 10% across all base station types. These findings suggest that AI/ML approaches can provide more accurate, dynamic EM radiation estimates, enabling better compliance monitoring and public safety assurance in mobile networks.

Table 5 outlines key simulation parameters for the GSM base station EM radiation study. It includes frequency bands (900 MHz, 1800 MHz), transmit power (20–40 W), antenna characteristics, traffic load (0–100%), and measurement details. The table specifies dataset sizes for AI/ML model training, validation, testing, and genetic algorithm parameters.

Model performance metrics

The neural network model outperformed other AI/ML approaches in estimating EM radiation from GSM base stations. It achieved the lowest MSE (0.0132), RMSE (0.1149), and MAE (0.0912), with the highest R-squared value (0.9245). However, it also had the longest execution time (3.7821 s). Linear regression offered the fastest execution (0.0532 s) but with lower accuracy. Table 5 compares the performance of various AI/ML models for estimating EM radiation from GSM base stations.

Table 6 compares performance metrics for various AI/ML models in estimating EM radiation from GSM base stations. The neural network model achieves the best accuracy with the lowest MSE (0.0132), RMSE (0.1149), and MAE (0.0912) and the highest R-squared (0.9245). However, it has the longest execution time (3.7821 s). Linear regression offers the fastest execution (0.0532 s) but lower accuracy.

Table 6. Performance metrics for different AI/ML models

Models	Mean squared error (MSE)	Root mean squared error (RMSE)	Mean absolute error (MAE)	R-squared (R2)	Execution time (s)
Linear regression	0.0215	0.1466	0.1183	0.8754	0.0532
Random forest	0.0156	0.1249	0.0987	0.9102	1.2456
Neural network	0.0132	0.1149	0.0912	0.9245	3.7821
Support vector regression	0.0178	0.1334	0.1056	0.8967	0.8943
Gradient boosting	0.0143	0.1196	0.0945	0.9187	2.1678

Figure 10 presents a bar graph comparing the performance of various AI/ML models for estimating EM radiation levels from GSM base stations. The graph displays performance metrics such as Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and R-squared (R²) for each model.

Fig. 10 [Images not available. See PDF.]

Bar graph comparing the performance of AI/ML models

Figure 11 presents a scatter plot comparing the predicted EM radiation levels from the optimized AI/ML model against the actual measured values. The plot demonstrates a strong positive correlation between predictions and measurements, with most data points clustered tightly around the diagonal line, indicating high model accuracy. The x-axis represents the measured EM radiation levels in W/m², while the y-axis shows the corresponding model predictions. The plot exhibits a linear trend with minimal dispersion, suggesting the model's robust performance across various radiation intensities. Some minor deviations are observed at higher radiation levels, potentially indicating areas for model refinement. The coefficient of determination (R²) of 0.9245 further corroborates the model's excellent predictive capability. This visualization effectively illustrates the AI/ML model's ability to accurately estimate EM radiation levels from GSM base stations using traffic data, outperforming traditional power-based estimation methods.

Fig. 11 [Images not available. See PDF.]

Scatter plot of predicted vs. measured EM radiation levels

Clustering results

The clustering analysis revealed distinct EM radiation patterns among GSM base stations. K-means clustering with K = 3 demonstrated the best performance, achieving the highest Silhouette Score (0.72) and Calinski-Harabasz Index (301.7). These clusters likely correspond to different base station types and traffic profiles. Urban base stations formed a cluster characterized by higher average EM radiation levels and more variable patterns, reflecting the dynamic nature of urban traffic. Suburban and rural base stations formed clusters with lower average radiation levels and more consistent patterns. The clustering results highlight the importance of location-specific factors in EM radiation estimation and suggest that tailored models for each cluster could further improve prediction accuracy. This insight enables more targeted monitoring and management strategies for different types of base stations.

K-means clustering outperformed hierarchical clustering and DBSCAN in our analysis, achieving a higher Silhouette score (0.72) and Calinski-Harabasz index (301.7). We attribute this superior performance to K-means' ability to handle spherical clusters and its scalability to large datasets. Alternative metrics such as the Davies-Bouldin index were also considered but showed consistent results favoring K-means.

Figure 12 illustrates the clustering results of EM radiation patterns from GSM base stations. The result employs a dimensionality reduction technique, t-SNE to project high-dimensional EM radiation data onto a 2D space. Three distinct clusters are evident, corresponding to the optimal K-means clustering (K = 3) identified in the analysis. Each cluster is represented by a different color, with data points symbolizing individual base stations. Cluster centroids are marked with larger symbols, indicating the average EM radiation pattern for each group. The spatial distribution of points within clusters reflects the similarity of EM radiation profiles, with tighter groupings suggesting more homogeneous patterns. Outliers, if present, are visible as isolated points. This clustering visualization corroborates the quantitative metrics presented in Table 7, providing a qualitative assessment of the clustering effectiveness. The clear separation between clusters underscores the distinct EM radiation characteristics associated with different base station types and environments, supporting the development of tailored estimation models for each cluster.

Fig. 12 [Images not available. See PDF.]

Clustering results—EM radiation patterns

Table 7. Clustering metrics for different algorithms and K values

Algorithm	K value	Silhouette score	Calinski-Harabasz Index	Davies-Bouldin Index	Inertia
K-means	2	0.68	245.3	0.82	156.2
K-means	3	0.72	301.7	0.76	112.5
K-means	4	0.65	278.4	0.89	98.3
DBSCAN	N/A	0.7	267.9	0.79	N/A
Hierarchical	2	0.67	239.1	0.85	162.8
Hierarchical	3	0.69	285.6	0.8	124.1

Table 7 compares clustering metrics for different algorithms and K values. K-Means with K = 3 show the best overall performance, with the highest Silhouette Score (0.72) and Calinski-Harabasz Index (301.7), and the lowest Davies-Bouldin Index (0.76). DBSCAN and Hierarchical clustering also demonstrate competitive results across various metrics.

Figure 13 presents the silhouette plot used to determine the optimal number of clusters for EM radiation patterns from GSM base stations. The plot displays silhouette scores for different cluster configurations, with higher scores indicating better-defined clusters. The x-axis represents the silhouette coefficient, ranging from − 1 to 1, while the y-axis shows the number of clusters evaluated. Each cluster is represented by a colored bar, with the width indicating the number of data points and the height showing the silhouette score distribution within that cluster.

Fig. 13 [Images not available. See PDF.]

Silhouette plot for determining the optimal number of clusters

The plot reveals that a three-cluster configuration (K = 3) yields the highest average silhouette score of 0.72, suggesting well-separated and cohesive clusters. This optimal clustering aligns with the distinct EM radiation patterns observed in urban, suburban, and rural base stations. The clear separation between clusters and minimal overlap support the choice of K = 3 for subsequent analyses, providing a robust foundation for developing location-specific EM radiation estimation models.

Optimal AI/ML model

The optimal AI/ML model for estimating EM radiation from GSM base stations was a neural network with three hidden layers (64, 32, 16 neurons) using ReLU activation. This model achieved the best performance with an RMSE of 0.1149 and R² of 0.9245 on the test set. Key hyperparameters included a learning rate of 0.001, batch size of 64, and 200 training epochs. Dropout (0.2) was employed to prevent overfitting. The model demonstrated strong generalization across different base station types, with rural stations showing the best performance (RMSE: 0.276, R²: 0.958). Feature importance analysis revealed that traffic load, number of active users, and transmit power were the most influential factors in predicting EM radiation levels. The optimized model significantly outperformed traditional power-based estimates, with mean absolute percentage errors below 10% for all base station types.

Table 8 presents the hyperparameters and performance metrics of the optimal AI/ML model for estimating EM radiation levels from GSM base stations. The model is a neural network with a specific architecture of three hidden layers (64, 32, and 16 neurons) using ReLU activation functions. It was trained with a learning rate of 0.001, batch size of 64, and 200 epochs, incorporating a dropout rate of 0.2 to prevent overfitting.

Table 8. Hyperparameters and performance of the optimal AI/ML model

Model type	Hyperparameters	Training RMSE	Validation RMSE	Test RMSE	R2	Execution time (s)
Neural network	Layers: [64, 32, 16] Activation: ReLU Learning rate: 0.001 Batch size: 64 Epochs: 200 Dropout: 0.2	0.0995	0.1024	0.1149	0.9245	3.7821

The model's performance is impressive, achieving a training RMSE of 0.0995, validation RMSE of 0.1024, and test RMSE of 0.1149. The high R² value of 0.9245 indicates that the model explains over 92% of the variance in the EM radiation levels. While the execution time of 3.7821 s is longer than simpler models, it reflects the complexity and depth of the neural network architecture. These results demonstrate the model's strong predictive capability and generalization performance in estimating EM radiation levels based on traffic data.

Figure 14 displays the relative importance of different features in the optimal AI/ML model for predicting EM radiation levels. The bar chart ranks features such as traffic load, active users, and antenna characteristics based on their impact on the model's predictions. This visualization helps identify the most influential factors in estimating EM radiation from GSM base stations.

Fig. 14 [Images not available. See PDF.]

Feature importance plot for the optimal model

Table 9 demonstrates the optimized AI model's generalization performance across different base station types. The model shows consistent accuracy across urban, suburban, and rural environments and macro and microcells. Rural base stations exhibit the best performance, with the lowest RMSE (0.276) and highest R² (0.958), while suburban stations show slightly lower accuracy.

Table 9. Generalization performance of the optimized model by base station type

Base station type	RMSE	MAE	R2	Mean absolute % error
Urban	0.298	0.231	0.945	8.70
Suburban	0.324	0.256	0.932	9.80
Rural	0.276	0.212	0.958	7.90
Macro cell	0.305	0.239	0.94	9.10
Micro cell	0.287	0.223	0.951	8.40

Figure 15 compares the optimized AI/ML model predictions with actual EM radiation measurements. The scatter plot likely shows predicted values on one axis and measured values on the other, with points clustered around a diagonal line indicating good agreement. The plot may include error bars or confidence intervals to illustrate prediction accuracy across different radiation levels.

Fig. 15 [Images not available. See PDF.]

Comparison of optimized model predictions with measurements

Figure 16 illustrates the correlation between predicted and true EM radiation values obtained from the optimized AI/ML model. The scatter plot demonstrates a strong positive linear relationship, with data points closely aligned along the diagonal line, indicating high prediction accuracy. The coefficient of determination (R² = 0.9245) confirms the model's robust performance in explaining the variance in EM radiation levels. Notably, the plot reveals consistent accuracy across the range of radiation values, with minimal heteroscedasticity. Some minor deviations are observed at higher radiation levels, suggesting potential areas for model refinement. The plot's density distribution highlights the model's precision in predicting common radiation levels, while also capturing less frequent, extreme values. This visualization underscores the AI/ML model's capability to generalize across diverse GSM base station configurations and traffic conditions, outperforming traditional power-based estimation methods. The tight clustering of points around the identity line reinforces the model's reliability for real-time EM radiation monitoring and regulatory compliance assessment.

Fig. 16 [Images not available. See PDF.]

Prediction vs. True values

Figure 17 compares the prediction accuracy of the AI-based model with traditional power-based estimates for EM radiation levels from GSM base stations. The graph likely demonstrates that the AI model achieves significantly higher accuracy across different scenarios compared to conventional power-based estimation methods. Result show the superiority of the AI approach in estimating real-world EM radiation levels. The AI model shows consistently lower error rates or higher R-squared values across various base station types, traffic conditions, or measurement distances. In contrast, the power-based estimates exhibit larger errors and more variability, especially in scenarios with dynamic traffic patterns or complex environments.

Fig. 17 [Images not available. See PDF.]

Prediction accuracy vs. power-based estimates

The improved accuracy of the AI model can be attributed to its ability to capture complex relationships between multiple input features and adapt to changing conditions. This comparison underscores the potential of AI-based approaches to provide more reliable and precise EM radiation estimates, which is crucial for regulatory compliance, public safety assurance, and optimizing network planning. The significant performance gap illustrated in this figure supports the adoption of AI techniques for EM radiation monitoring and management in mobile networks.

Conclusion

This research demonstrates the potential of AI-based optimization techniques to significantly enhance the accuracy and efficiency of estimating electromagnetic (EM) radiation levels from GSM base stations. By leveraging real-time traffic data and advanced machine learning algorithms, we developed precise, dynamic, and location-specific models for monitoring and managing EM radiation exposure.

Our results show that the optimized AI model, a neural network with three hidden layers and specific hyperparameters, outperformed traditional power-based estimation methods, achieving an RMSE of 0.1149 and R² of 0.9245 on the test set. The model demonstrated strong generalization capability across different base station types, with mean absolute percentage errors below 10% for all categories. Rural base stations exhibited the best performance (RMSE: 0.276, R²: 0.958), while suburban stations showed slightly lower accuracy.

Clustering analysis using K-means (K = 3) revealed distinct EM radiation patterns among base station types, with the highest Silhouette score (0.72) and Calinski-Harabasz index (301.7). These results highlight the importance of considering location-specific factors in EM radiation estimation and suggest that tailored models for each cluster could further improve prediction accuracy.

The implications of this research are significant for telecom operators, regulatory bodies, and public health officials. Adopting AI-powered estimation methods enables more informed decision-making regarding network planning, optimization, and safety measures.Furthermore, this work lays the foundation for developing intelligent, data-driven systems that can continuously monitor and adapt to changing EM radiation patterns in our increasingly connected world.

As mobile technologies continue to evolve, particularly with the deployment of 5G and small cell networks, further research is needed to extend the methodology and assess its applicability to these new scenarios. Exploring dynamic feature selection methods, incorporating additional contextual data, and conducting long-term validation studies are potential avenues for future work to further enhance the model's performance and adaptability.

In conclusion, this research demonstrates the immense potential of AI-based optimization in improving the accuracy, efficiency, and responsiveness of EM radiation monitoring in mobile networks. By providing a data-driven, adaptive approach to estimating EM radiation levels, this work contributes to the sustainable and responsible deployment of mobile network infrastructure while addressing public concerns about potential health impacts.

Implications

• Improved Accuracy with the AI-based approach provides more accurate and dynamic estimates of EM radiation levels compared to conventional methods, enabling better compliance monitoring and public safety assurance.

• The clustering results highlight the importance of location-specific factors, suggesting that tailored models for different base station types could further improve prediction accuracy and enable more targeted monitoring strategies.

• The optimized model's performance indicates its potential for real-time EM radiation estimation using traffic data, which could significantly enhance the efficiency of regulatory compliance and public health protection efforts.

Limitations and future work

• This research focused primarily on GSM networks. Future research should extend the methodology to newer technologies such as 5G and small cell networks, which may exhibit different traffic patterns and EM radiation characteristics.

• The current model relies on a fixed set of input features. Exploring dynamic feature selection methods and incorporating additional contextual data (e.g., weather conditions, urban development) could potentially improve model performance further.

• Long-term studies are needed to validate the model's performance over extended periods and assess its ability to adapt to evolving network conditions and traffic patterns.

Author contributions

S.K. conceptualized the study, developed the methodology, and supervised the research. R.L. and R.K.S. collected and processed the data, implemented the AI/ML models, and performed the analysis. D.K.N. contributed to the literature review and assisted with data interpretation. S.K. wrote the original draft of the manuscript. All authors reviewed, edited, and approved the final version of the manuscript.

Data availability

The datasets used and analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

References

1. Jayaraju, N et al. Mobile phone and base stations radiation and its effects on human health and environment: a review. Sustain Technol Entrep; 2023; 2, 2 100031. [DOI: https://dx.doi.org/10.1016/j.stae.2022.100031]

2. Karipidis, K; Mate, R; Urban, D et al. 5G mobile networks and health—a state-of-the-science review of the research into low-level RF fields above 6 GHz. J Expo Sci Environ Epidemiol; 2021; 31, pp. 585-605. [DOI: https://dx.doi.org/10.1038/s41370-021-00297-6]

3. Hale, A et al. Safety regulation: the lessons of workplace safety rule management for managing the regulatory burden. Saf Sci; 2015; 71, pp. 112-122. [DOI: https://dx.doi.org/10.1016/j.ssci.2013.11.012]

4. Shi, D et al. Electromagnetic radiation estimation at the ground plane near fifth-generation base stations in China by using the machine learning method. IET Microwav Antennas Propag; 2024; [DOI: https://dx.doi.org/10.1049/mia2.12467]

5. Guo, D; Caprani, CC. Traffic load patterning on long span bridges: a rational approach. Struct Saf; 2019; 77, pp. 18-29. [DOI: https://dx.doi.org/10.1016/j.strusafe.2018.11.003]

6. He, QQ et al. Accurate method to estimate EM radiation from a gsm base station. Progr Electromagn Res M; 2014; 34, pp. 19-27. [DOI: https://dx.doi.org/10.2528/pierm13091301]

7. Zheng, J et al. Intelligent cognition of traffic loads on road bridges: from measurement to simulation—a review. Measurement; 2022; 200, 111636. [DOI: https://dx.doi.org/10.1016/j.measurement.2022.111636]

8. Younis, M; Akkaya, K. Strategies and techniques for node placement in wireless sensor networks: a survey. Ad Hoc Netw; 2008; 6, 4 pp. 621-655. [DOI: https://dx.doi.org/10.1016/j.adhoc.2007.05.003]

9. Miclaus, S; Bechet, P. Estimated and measured values of the radiofrequency radiation power density around cellular base stations. Rom J Phys; 2007; 52, 3/4 429.

10. Aerts, S et al. Spatio-temporal exposure assessment of mobile phone base stations in an urban environment. Environ Res; 2013; 126, pp. 15-26.

11. He, J; Versfeld, A; Danikas, L. He, J. Accurate estimation of GSM base station emissions. IEEE international symposium on electromagnetic compatibility; 2008; Detroit, IEEE: pp. 1-6.

12. Kürner, T et al. Kürner, T et al. Measurement and analysis of the electromagnetic fields from GSM base stations. Proceedings of IEEE vehicular technology; 1999; Amsterdam, IEEE: pp. 2600-2603.

13. Bürgi, A; Frei, P; Theis, G; Mohler, E; Braun-Fahrländer, C; Fröhlich, M. A model for radiofrequency electromagnetic field predictions at outdoor and indoor locations in the context of epidemiological research. Bioelectromagnetics; 2010; 31, 3 pp. 226-236. [DOI: https://dx.doi.org/10.1002/bem.20552]

14. Gecgel, C; Goztepe, C; Kurt, GK. Transmit antenna selection for large-scale MIMO GSM systems. IEEE Trans Wirel Commun; 2018; 17, 3 pp. 1536-1547.

15. de Souza Junior, GC; de Figueiredo, FAP; Alves, H; da Costa, EG. "Enhanced NOMA-MIMO networks using deep learning and antenna selection. IEEE Access.; 2020; 8, pp. 211052-211065.

16. Hu, Y; Yang, W; Yi, H; Huang, X; Yang, L. Machine learning and particle swarm optimization for dielectric resonator antenna design. IEEE Antennas Wirel Propag Lett; 2020; 19, 6 pp. 949-953.

17. Calik, N; Belen, MA; Mahouti, P. Deep learning based modified MLP model for precise scattering parameter prediction of capacitive feed antenna. Int J Numer Model; 2020; 33, 5

18. Golichenko, I et al. Extrapolation problem for continuous time periodically correlated isotropic random fields. Bull Math Sci Appl; 2017; 19, pp. 1-23. [DOI: https://dx.doi.org/10.18052/www.scipress.com/bmsa.19.1]

19. Krenn, M; Buffoni, L; Coutinho, B et al. Forecasting the future of artificial intelligence with machine learning-based link prediction in an exponentially growing knowledge network. Nat Mach Intell; 2023; 5, pp. 1326-1335. [DOI: https://dx.doi.org/10.1038/s42256-023-00735-0]

20. Singh, N; Jindal, T. Jindal, T. Electromagnetic field mobile phone radiation toxicity. New frontiers in environmental toxicology; 2022; Cham, Springer: [DOI: https://dx.doi.org/10.1007/978-3-030-72173-2_1]

21. Talaei Khoei, T; Ould Slimane, H; Kaabouch, N. Deep learning: systematic review, models, challenges, and research directions. Neural Comput Appl; 2023; 35, pp. 23103-23124. [DOI: https://dx.doi.org/10.1007/s00521-023-08957-4]

22. Mhlongo, S et al. Challenges, opportunities, and prospects of adopting and using smart digital technologies in learning environments: an iterative review. Heliyon.; 2023; 9, 6 [DOI: https://dx.doi.org/10.1016/j.heliyon.2023.e16348]

23. Jagetia, GC. Genotoxic effects of electromagnetic field radiations from mobile phones. Environ Res; 2022; 212, 113321. [DOI: https://dx.doi.org/10.1016/j.envres.2022.113321]

24. Martínez-Búrdalo, M et al. FDTD assessment of human exposure to electromagnetic fields from WiFi and bluetooth devices in some operating situations. Bioelectromagnetics; 2009; 30, 2 pp. 142-151. [DOI: https://dx.doi.org/10.1002/bem.20455]

25. Hamid S et al. Appropriating online social networking (OSN) activities for higher education: two Malaysian cases. In: Changing demands, changing directions. Proceedings ascilite Hobart (2011). pp. 526–538.