Comparative Analysis of Deep Learning Models for Real-Time Battery State-of-Charge Prediction in Formula E

Accurate real-time State of Charge (SOC) prediction is a critical component of energy management and strategic decision-making in Formula E racing. This thesis presents an end-to-end artificial intelligence framework for real-time SOC estimation based on live telemetry streams and FIA broadcast graphics, designed to operate under realistic race conditions. The proposed framework integrates a telemetry reader, an enhanced SOC extraction module (SOC-TV), and a deep learning predictor (AI Predictor) through a scalable publish–subscribe backbone based on Apache Kafka, which acts as a central event bus to ingest telemetry streams and FIA broadcast graphics, and to deliver time-aligned inputs to the AI Predictor for real-time inference. Special attention is devoted to robust vehicle speed estimation, Pit Boost detection, and real-time inference constraints. Multiple deep learning architectures, including recurrent, convolutional, hybrid, and attention-based models, are systematically evaluated under a unified experimental setup. A comprehensive benchmark is conducted by varying temporal window size, resampling granularity (1 s vs. 5 s), and warm-up strategies. Results demonstrate that temporal context plays a fundamental role in SOC prediction accuracy and that different architectural families exhibit distinct sensitivities to window length. While models trained on 5 s resampled data require explicit warm-up periods to achieve stable predictions, the best overall performance is obtained using 1 s resampling with an appropriately selected window size. Among all evaluated configurations, a fine-tuned LSTM architecture achieves the lowest average error, reaching a Mean Absolute Error of 0.611 kWh (1.58% relative error with respect to the usable battery capacity) across the first two real Formula E races of Season 12. This represents a substantial improvement over previous approaches and highlights the effectiveness of combining high-resolution temporal modeling with deployment-oriented system design.