MTML-EV: Multi-target meta-learning for cross-zone urban electric vehicle charging infrastructure management

Rapid electric vehicle (EV) adoption is reshaping urban mobility and energy systems, intensifying challenges for charging infrastructure management, including congestion, uneven spatial demand, and grid stress. Addressing these challenges requires reliable forecasting of multiple charging-system states across both near-real-time operation and short-term planning horizons, particularly for newly deployed or evolving zones with limited historical data. Existing urban EV charging forecasting approaches often focus on individual charging-system variables and exhibit limited adaptability to previously unseen zones. To address this limitation, this paper presents MTML-EV, a multi-target meta-learning framework that jointly predicts charging occupancy, energy demand, and electricity price across 1-, 6-, and 12-hour forecasting horizons. MTML-EV formulates each traffic zone as a learning task and leverages a BiGRU-based temporal encoder optimized with first-order model-agnostic meta-learning (FOMAML) to learn transferable temporal representations that enable rapid adaptation to previously unseen zones. Extensive experiments on the UrbanEV benchmark evaluate MTML-EV across multiple forecasting horizons and unseen-zone adaptation scenarios against deep learning, hybrid, and Transformer-based baselines. For 1-hour ahead prediction on unseen zones, MTML-EV reduces RMSE by 7.5% for energy demand and 59% for electricity price compared with the strongest Transformer baseline while maintaining high occupancy forecasting accuracy. Additional robustness analyses across multiple unseen-zone partitions demonstrate that these performance gains remain stable under varying zone configurations, while interpretability analysis confirms that the model relies on meaningful temporal, charging-related, and contextual factors. Overall, the results demonstrate the effectiveness of MTML-EV for multi-target cross-zone forecasting under heterogeneous urban conditions, highlighting the value of transferable temporal representations for adaptive EV charging infrastructure management.