Soft Dynamic Time Warping Neural Networks for Electric Vehicle Charging Station Load Forecasting

The article focuses on studying the time-series characteristics of Electric Vehicle Charging Station (EVCS) loads. It highlights the superior performance of Soft Dynamic Time Warping loss-based Neural Networks (SDTW-NNs). First, we conduct an exploratory analysis of EVCS loads. By examining the spatial distribution map of EVCSs, we can understand the spatial distribution of EVCSs. We then reveal the changing patterns, volatility, and potential usage trends of EVCS loads through descriptive statistics, rolling analysis, and violin analysis. Next, we delve into the changing patterns, anomaly detection, and pattern extraction of EVCS loads using methods such as piecewise and symbolic aggregate approximation and matrix profiles. Finally, we apply the SDTW-NNs for predicting EVCS loads and validate with real data that neural networks based on the SDTW loss can significantly enhance forecasting performance in time-series analysis.