Real-Time Prediction for the Temperature Field in Energy Vehicle Charging Piles Based on Autoencoder Dimensionality Reduction

Abstract This study develops a real-time 3D temperature field prediction method for charging piles using an autoencoder (AE) and backpropagation neural network (BPNN) to address thermal management challenges during fast charging. Unlike conventional sensor-based methods (limited to local measurements) or numerical simulations (computationally expensive), the approach combines finite element modeling (FEM), design of experiments (DOEs), and deep learning for efficient global prediction. A validated FEM generates 1000 temperature samples via DOE, which are compressed into latent features by AE and predicted via BPNN. Decoding these features reconstructs the 3D temperature field in seconds. Compared to proper orthogonal decomposition (POD), AE's nonlinear dimensionality reduction improves accuracy by capturing complex thermal patterns, achieving less than 1% relative error. This real-time prediction method provides an effective solution for thermal monitoring in charging piles, supporting operational safety and longevity enhancement.