Electric Vehicle Load Forecasting Based on Time Series Decomposition and CPIKS Module

As the number of electric vehicles (EVs) increases, accurate EV load forecasting becomes crucial. EV loads require more efficient and precise predictions compared to traditional loads. Multi-variable EV load forecasting demands advanced methods to handle inter-variable relationships and improve prediction accuracy. This paper proposes a hybrid model combining one-dimensional convolution and time series decomposition. To address complex EV load patterns, we decompose the load series into trend and fluctuation components. We use linear regression to predict the trend component, reducing computational complexity. For the fluctuation component, we introduce the Convolutional Pooling Integrated Key Series (CPIKS) module to capture inter-variable channel correlations. CPIKS employs a centralized strategy to enhance channel interaction efficiency by aggregating features from all channel sequences into a key sequence, which is then fused with each individual sequence. We demonstrate the superiority of our model's predictions using EV load datasets from Caltech and office areas.