Multi-Decoder Transformer for Charging Load Prediction of Electric Vehicle Charging Stations

The penetration rate of electric vehicles (EVs) has seen a rapid increase in recent years. Electric vehicle charging loads impose significant demands on the power grid. Particularly, the prediction of charging loads at EV charging stations is of paramount importance for the secure operation of the electrical grid. Currently, deep learning methods based on the powerful Transformer architecture have proven effective in handling long-term dependencies within charging load data and have achieved state-of-the-art predict performance. However, due to the extremely limited samples collected from the real world, traditional deep learning methods are prone to issues of model overfitting, and single-branch decoder often exhibit substantial biases in their predictions. To address these issues, we first employ a diffusion model to synthesize additional charging load data as a complement to the real-world dataset, thereby mitigating the problem of model overfitting. Additionally, we leverage the Transformer-based structure and propose a multi-decoder approach for making predictions from the encoded signals. During the training phase, multiple parallel homogeneous decoders are utilized, and during the testing phase, these multiple decoders are employed for ensembled predictions of charging load, mitigating model biases. The experimental results on real-world datasets consistently demonstrate that our approach consistently outperforms baseline methods.