Meta-learning Gibbs sampling combined with the Transformer neural network for EV Charging Charging Forecasting

Electric vehicle (EV) charging event prediction is a key technology for optimizing charging infrastructure and enhancing the stability of smart grids. In recent years, spatio-temporal prediction methods based on deep learning have made significant progress in improving prediction accuracy, but they still face challenges such as small sample learning and spatial prediction ambiguity. To address these challenges, this study optimizes an integrated approach for regional EV charging event prediction, namely meta-learning Gibbs sampling combined with the Transformer neural network. The model consists of two modules: 1) Model-Agnostic Meta-Learning (MAML) training method, which enables the model to solve new learning tasks with only a small number of training samples, aligning with the characteristic of scarce and hard-to-obtain charging data in reality. 2) GTN: Gibbs sampling combine with the neural network with the transformer embedding layer, this module more accurately capture the spatial relationships of charging events. Quantitative experiments demonstrate that relative position encoding reduces spatial prediction ambiguity by 23% measured through JS divergence between location clusters. The evaluation results on the dataset of 8132 electric vehicle charging piles in Shenzhen, China,show that the proposed method can achieve precise prediction under small sample learning and accurately capture spatial relationships. Compared with the original model, the proposed model 1) It only decreased by 14% in the indicator for measuring the prediction accuracy, but the convergence speed increased by 52% under the small sample data;2) avoids overfitting of geographical locations.