Author(s) :   Govinda Heda¹, Prof. (Dr.) V. B. Singh²

Abstract : The rapid urbanization of Indian cities and the nationwide push towards electric mobility have created an urgent need for localized, data-driven strategies to support electric vehicle (EV) infrastructure. This research focuses on developing a comprehensive predictive analytics framework tailored to forecast EV charging Demand in Nashik, Maharashtra—a tier-2 Indian city witnessing accelerating EV adoption due to increasing environmental awareness, policy incentives, and improved vehicle availability.

To achieve this objective, we leverage a robust multivariate time-series dataset incorporating various influential parameters such as real-time EV charging station logs, historical weather data (temperature, humidity, precipitation), traffic flow metrics (vehicle count and congestion levels), and socio-demographic indicators (population density, income levels, and EV ownership rates). These variables are collected at high granularity to enable precise day-ahead forecasting at 15-minute intervals, essential for dynamic energy distribution and optimal station management.

We employ traditional statistical and modern deep learning techniques for model development. Specifically, the Autoregressive Integrated Moving Average (ARIMA) model is a benchmark due to its effectiveness in linear time-series forecasting. In contrast, the Long Short-Term Memory (LSTM) neural network, a variant of recurrent neural networks (RNN), captures nonlinear and temporal dependencies inherent in high-frequency charging data. The modeling is implemented in IBM SPSS Statistics for data preprocessing, variable selection, and preliminary regression analysis, while the LSTM model is trained using Python with TensorFlow/Keras libraries.

Model performance is evaluated using standard error metrics: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). The LSTM model achieves superior accuracy, with an MAE of 2.51%, RMSE of 3.12%, and MAPE of 3.45%, outperforming the ARIMA model across all benchmarks. This demonstrates the LSTM's capacity to handle complex, nonlinear interactions among multiple variables and make more accurate short-term forecasts.

An SHAP (SHapley Additive exPlanations) analysis was conducted to understand each predictor's relative influence. The analysis reveals that temperature accounts for approximately 30% of the model's predictive power, followed by the day of the week (25%) and traffic volume (20%). These findings are further validated through multivariate regression and hypothesis testing. The results confirm the statistical significance of the identified predictors, with p-values less than 0.01, indicating a strong and non-random association with charging Demand.

This study yields practical and policy-relevant outcomes. Accurate demand forecasting allows for more effective charging station placement, helps determine the number of connectors and battery swapping stations required at each location, and supports grid-level load balancing. For energy utilities and city planners, these insights can be used to reduce peak load issues, manage real-time electricity distribution, and align infrastructure development with demand trends.

Importantly, this research addresses a critical knowledge gap: the absence of localized EV demand forecasting models for mid-sized Indian cities. Unlike metropolitan-focused studies, our approach is scalable and context-sensitive, offering a blueprint for implementing predictive analytics frameworks in similar urban settings across India. The methodology and findings are expected to contribute meaningfully to India’s broader agenda of achieving sustainable and equitable electric mobility through intelligent, anticipatory infrastructure planning.

DOI : 10.61161/ijarcsms.v13i5.3

Pages : 17-31

*Authors are invited to submit papers through E-mail at editor.ijarcsms@gmail.com