Keywords: Real-Time Range Prediction, Long Short-Term Memory (LSTM), Time Series Modelling, Deep Learning, Vehicular Telemetry, CAN Data, Range Anxiety
Accurate and real-time range estimation is a critical enabler for efficient energy management and operational planning in Commercial Electric Vehicles (CVEVs). This project presents a robust deep learning-driven framework for vehicle range prediction, leveraging a Long Short-Term Memory (LSTM) network trained on high-frequency powertrain CAN data obtained from CVEVs operating across diverse driving environments.
The dataset comprises key dynamic and environmental parameters—including State of Charge (SOC), vehicle speed, acceleration, and temperature—captured during real-world operations. Extensive data pre-processing techniques were applied to ensure signal integrity, including high-resolution timestamp synchronization, noise filtering, and imputation of missing values, thereby establishing a reliable foundation for time-series modelling.
The LSTM network architecture was meticulously designed to capture the temporal dependencies and nonlinear interdependencies inherent in vehicular telemetry. Trained on a large-scale telematics dataset, the model demonstrated superior predictive performance, evaluated through key statistical metrics such as Mean Squared Error (MSE) and the Coefficient of Determination (R²), outperforming traditional range estimation algorithms.
To evaluate its practical deployability, the trained LSTM model was optimized and deployed on an automotive-grade embedded processor. Functional validation was carried out on a test bench using live CAN feed replay from recorded vehicle logs. The model maintained high predictive accuracy under real-time constraints, demonstrating its resilience and robustness in an embedded automotive environment.
This intelligent range estimation solution offers significant advantages to fleet operators by enabling dynamic, context-aware range predictions. It supports optimized route planning, minimizes operational downtime, and mitigates range anxiety—contributing to enhanced fleet efficiency and energy utilization. From an engineering perspective, this work substantiates the viability of deploying advanced deep learning architectures on resource-constrained embedded platforms for real-time inference in electric mobility applications.