Spatio-temporal Forecasting of Schedulable Capacity of Shared Electric Vehicles

Aiming at the schedulable capacity forecasting of shared electric vehicles participating in power system reserve service through demand response (DR), based on the historical trajectory data, this paper proposes a schedulable capacity evaluation model combining with the model- agnostic meta-learning, the convolutional neural network, the long- and short-term neural network and the attention mechanism (MAML-CNN-LSTM-Attention). Specifically, the LSTM is used to model and learn the dynamic changes of the effective feature vectors extracted by the CNN from the historical data, and the MAML is adopted to train the initialization parameters of the CNN-LSTM network. While solving that it is difficult for the traditional neural networks to effectively extract the potential high-dimensional features in the historical sequences and the important information is apt to be lost when the time series is too long, the meta-prediction network is fine-tuned by multi-task training to adapt to the new prediction tasks quickly, in order to improve the prediction accuracy and generalization ability of the model. The attention mechanism is added to highlight the timing sequence information which plays a key role in the results to further improve the forecasting accuracy. Simulation results show that the model is able to effectively forecast the schedulable capacity of the shared electric vehicles in different date types and functional areas, and provides a reference for the risk assessment study of the shared electric vehicles participating in the reserve services through DR. © 2023 Power System Technology Press. All rights reserved.