Edge AI System for EV Trip Analytics: On-Device WLTP Class 3B Processing and Battery Temperature Prediction

Abstract

EV analytics are currently reliant on cloud computing; this introduces latency and does not allow drivers to receive real-time feedback. Current systems can hardly estimate interpretable 1 Hz measurements or do short-term thermal predictions on small embedded environments. This study aims to demonstrate that these tasks can be completely performed on-device. In this study, we describe a low-cost edge platform, based on transforming EV telemetry into real-time analytics on an ESP32-S3 microcontroller. The device is capable of calculating energy consumption, coasting performance, regenerative efficiency, battery-flow efficiency, and pack temperature at 1 Hz. It also executes a TinyML model prediction algorithm of the battery temperature 120 s in advance. To our best knowledge, this is one of the first demonstration of 1 Hz thermal forecasting based on a 1 Hz TinyML model running on the resource-constrained microcontroller. It provides a fully edge native system, and sends reduced size summaries over MQTT to a Node-RED dashboard. WLTP Class 3B cycle experiments indicate that it has stable 1 Hz throughput (1801/1801 messages), fast inference (6 ms), low memory consumption (1 MB flash), and highly accurate thermal prediction (MAE of 0.09 C). These findings demonstrate that real-time EV analytics and compact horizon thermal prediction is possible without cloud computing, which provides a viable alternative to standard EV monitoring architectures.