When acting as the shared energy storage and participating in electricity market services, the schedulable capacity that shared electric vehicle (shared EV) will provide to the grid needs in future time needs to be predicted accurately. This research first proposes a method to construct a schedulable capacity dataset for shared EVs based on publicly available shared vehicle rental service data. Secondly, a schedulable capacity evaluation model based on model-agnostic meta-learning, convolutional neural network, long short-term neural network and attention mechanism (MAML-CNN-LSTM-Attention) is proposed. Through the model, the aggregated schedulable capacity of shared EVs in different functional communities for the coming 60 minutes is predicted. Model uses MAML to fine-tune the meta-prediction network through multi-task training to quickly adapt to feature changes caused by different travel habits of different functional communities; CNN-LSTM is used to learn spatial features of schedulable capacity and efficiently extract high-dimensional temporal features from historical sequences; Attention mechanism is used to further improve model prediction accuracy. Simulations show that the model proposed in this paper outperforms other existing models and can reliably predict the schedulable capacity for different date types and functional areas, providing useful decision aids for shared EV operators to participate in market services.