A hybrid deep learning model for load forecasting of electric vehicle charging stations using time series decomposition

The precise load forecasting is one of critical factors for the safe operation of electric vehicle charging stations (EVCSs), and it can also support planning decisions for expanding charging infrastructures. Due to the uncertain charging demands and external influences, current EVCS load forecasting models generally face the challenges of strong nonlinearity and instability. In this research, a novel hybrid deep learning model that combines the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and TimesNet is proposed for the load forecasting of EVCSs, which effectively integrates the time-domain decomposition and the frequency-domain modeling. The ICEEMDAN decomposes the raw load series into multi-scale components, enabling the noise suppression and finer feature separation without manual feature engineering. The TimesNet then models these components in the frequency domain to capture complex temporal patterns across multiple scales. The proposed ICEEMDAN-TimesNet forecasting model is analyzed and evaluated under different scenarios, including the multi-step-ahead forecasting with varying time window and changes in the input sequence length. Results demonstrate that the proposed ICEEMDAN-TimesNet model consistently outperforms other state-of-the-art benchmark models, demonstrating superior accuracy, robustness, and generalization ability under all different scenarios.