Robust prediction of lithium-ion battery temperature using wavelet scattering and deep learning

Lithium-ion batteries are essential for electric vehicles and energy storage systems; nevertheless, inconsistent temperature distributions jeopardize their safety and reliability, potentially resulting in thermal runaway. Thus, precise temperature forecasting models are crucial for enhancing performance and guaranteeing operational safety. In this research, we present a hybrid framework that combines the Wavelet Scattering Transform (WST) with Long Short-Term Memory (LSTM) neural networks to predict battery temperature distributions. The methodology was validated using datasets from the NASA Ames Prognostics Center of Excellence (PCoE), which included authentic cycle data for voltage, current, and temperature. To guarantee reliability, model predictions were calibrated and assessed using three methods: RAW outputs, bias correction, and Poly2. The proposed WST-LSTM framework showed improved stability and consistency relative to the raw-signal LSTM baseline, supporting its suitability for thermal-safety evaluation and battery-management applications. The proposed framework was implemented in MATLAB/Simulink for an electric-vehicle battery model, thereby validating its applicability to real-world automotive platforms. This study emphasizes forecasting temperature at a specific sensor point rather than reconstructing the temperature field. The suggested WST-LSTM model predicted the temperature trajectory at this location utilizing electrical and environmental signals.