Short Paper: Predicting and Analyzing EV Energy Consumption in Bangladesh : A Machine Learning Approach

The global shift to cleaner energy has accelerated EV adoption, offering a sustainable solution to reduce transportation-related carbon emissions [15], [30]. While EVs contribute to decarbonization, they also place significant strain on power networks, especially in developing countries like Bangladesh, where increased demand for EV charging burdens the national grid during peak hours [6], [22].

The rising energy consumption from EVs risks overloading grid capacity, potentially leading to instability and outages. To address these challenges, effective energy management solutions are critical, particularly through machine learning (ML) [22], [29], [30]. ML algorithms analyze historical charging patterns, energy consumption, and grid conditions to predict future demand [24], enabling strategies like optimizing energy distribution, reducing peak loads, and improving grid capacity [30]. Additionally, understanding driver behavior, including charging times and locations, helps refine ML models to improve grid operations and infrastructure planning [23].

In Bangladesh, limited localized EV data presents a challenge. To overcome this, our study uses California’s EV charging data as a proxy [8], developing predictive models tailored to Bangladesh’s needs. These models provide valuable insights into energy demands, assisting policymakers and grid operators in planning sustainable EV infrastructure while ensuring grid stability.

We do not claim this model as a solution to the energy crisis but as a tool to help anticipate and manage challenges from growing EV adoption. Our goal is to support the sustainable integration of EVs into Bangladesh’s transportation system while ensuring grid resilience.

The rapid growth of electric vehicles (EVs) has prompted significant research into optimizing charging infrastructure and predicting energy consumption patterns. Machine learning (ML) and deep learning techniques have been explored to improve demand forecasting, scheduling, and charging station efficiency [2, 4, 9, 10, 16, 19, 27, 28].

[10] analyzed energy consumption estimation, emphasizing adaptive models due to rapid developments in neural network architectures. The study focused on NeuralPower and SyNERGY for GPU-based estimation and fragmented ML ecosystems. [28] proposed an ML solution to optimize water pumping station operations on remote islands, achieving a 15% reduction in energy consumption and suggesting that integrating renewable energy sources could enhance efficiency further.

[9] used multiple linear regression models to forecast residential energy consumption, improving accuracy by incorporating temperature and solar radiation, with potential for further refinement using smart meters. [27] combined Stationary Wavelet Transform (SWT) with deep transformers for household energy forecasting, achieving a 45% improvement in RMSE. [2] employed LSTM networks to estimate energy usage in educational facilities, achieving 94.31% accuracy. Future studies suggested adding device types to refine predictions. [19] reviewed hybrid and ensemble ML models for energy usage prediction, stressing the need for faster, more accurate models.

[4] explored how ML and AI optimize LoRaWAN performance by adjusting spreading factors and bandwidth, focusing on cross-layer optimization. [16] presented a hybrid model using CatBoost, XGBoost, and a multilayer perceptron for energy forecasting on Jeju Island, considering factors like weather and holidays.

The demand for EV infrastructure has driven research on optimizing charging stations and predicting energy use [1, 5, 11, 12, 13, 14, 17]. [13] presented data from 22,200 charging stations in Germany, revealing trends in utilization and energy transfer. [1] used SARIMA models to estimate power usage at EV stations. [14] proposed a microgrid-based system to reduce operational costs at PV-powered stations.

Studies in [20, 21, 22, 23, 24, 25, 29, 30, 31, 32] investigated smart grids, centralized management, and scheduling algorithms to optimize EV charging, supporting EV infrastructure growth and minimizing grid strain. However, in Bangladesh, research on EV infrastructure remains limited, requiring tailored ML approaches to support sustainable EV growth.

Bangladesh’s energy infrastructure faces mounting pressure from the rising demand for EV charging stations. As EV adoption grows, efficient energy distribution and grid stability become critical. Factors like traffic patterns and peak-hour charging increase the unpredictability of demand, risking grid overload, outages, and inefficiencies. Accurate predictions of EV charging behavior are essential for resource optimization, preventing overloads, and ensuring smooth energy distribution. Without such insights, the infrastructure struggles to meet the EV market’s needs sustainably.

This research addresses these challenges by developing machine learning-based predictive models to forecast EV charging behaviors and optimize energy distribution. Leveraging datasets from advanced EV markets, such as California, this study aims to create a scalable, adaptable framework suitable for Bangladesh’s evolving EV infrastructure. These models enable proactive grid management, reduce environmental impacts, and guide sustainable EV infrastructure planning, laying the groundwork for future applications in developing countries.

The methodology for predicting EV charging behavior involves analyzing key features such as session duration, energy consumption [22], gasoline savings, fees, and GHG reductions. Preprocessing steps include data cleaning, outlier detection, feature scaling, and encoding categorical variables. Exploratory Data Analysis (EDA) identifies critical patterns, while feature engineering enhances the dataset for machine learning models.

Three models—Random Forest (RF), Gradient Boosting Regression (GBR), and Decision Tree (DT)—are utilized. RF reduces overfitting by averaging predictions across trees [3], GBR iteratively corrects errors for higher accuracy [7], and DT provides an interpretable structure for regression and classification tasks [3]. These models establish a robust framework for predicting EV charging behavior while effectively handling dataset complexities. At this stage, traditional models like RF and GBR provide a strong foundation, with the potential to adopt custom models or advanced techniques for Bangladesh’s unique requirements.

To refine precision, K-fold cross-validation and hyperparameter optimization using Grid Search CV, Randomized Search CV, and HyperOpt [7] are employed. Performance metrics such as R², RMSE, and MAE ensure accurate energy demand forecasts, helping to manage grid capacity, highlight peak usage, and inform infrastructure planning.

The efficacy of a predictive machine learning model predominantly hinges on the quality of the dataset employed. In this research, the dataset used for modeling and forecasting EV charging behavior derives from the EV Charging Station Usage [8] records of California City. This dataset presents diverse and comprehensive insights, serving as the essential foundation for training and evaluating machine learning models designed to enhance the operational efficiency of EV charging stations.

Table  2 summarizes the hyperparameters and their optimized values obtained via Grid Search CV, Randomized Search CV, and HyperOpt using K-fold validation for Random Forest, Decision Tree, and Gradient Boosting Regression. Gradient Boosting Regression allows greater customization, including learning rate and subsample, while Random Forest and Decision Tree focus on core parameters like tree depth, splitting criteria, and the number of trees or leaf nodes. These parameters are crucial for controlling model complexity, avoiding overfitting, and enhancing predictive accuracy.

Table  3 compares the performance of Random Forest, Decision Tree, and Gradient Boosting with R2, RMSE, and MAE [24] measures. Random Forest obtains an R2 of 0.9999944, while Gradient Boosting has a slightly higher R2 of 0.9999965. The Decision Tree falls behind with an R2 of 0.9999272. All models have remarkable R2 values. Random Forest has the lowest RMSE at 0.003, but a higher MAE of 0.0874, implying worse precision. The decision tree has a higher RMSE of 0.0410 and MAE of 0.0342. In comparison, Gradient Boosting has the lowest MAE (0.0071), making it the most exact model.

This study addresses the critical need to estimate energy consumption at EV charging stations in Bangladesh, where rapid EV adoption increasingly strains the national power grid. Employing machine learning models such as Random Forest, Gradient Boosting Regression, and Decision Tree, the study identifies Gradient Boosting as the most accurate predictor, enabling optimized energy distribution, mitigating grid overload [30], and improving resource utilization. In the absence of localized data, the study utilizes the California City dataset as a proxy [8], offering insights into future energy demands and supporting strategic infrastructure planning. As EV growth intensifies challenges for the already overburdened grid [18], this research provides a foundation for balancing energy supply and demand, advancing grid stability, and fostering long-term sustainability. Addressing a critical gap in the literature, it establishes a proactive framework for policymakers and grid operators, with future efforts focusing on integrating real-time, localized data to further refine predictions and support the transition to cleaner transportation systems