Integrated deep learning framework for electric vehicles' flexibility forecasting and optimal dispatch

Accurate flexibility forecasting of electric vehicle (EV) cluster is essential for charging station operators (CSOs) to participate in energy market arbitrage and grid optimization. This paper proposes an integrated deep learning framework for EVs flexibility forecasting and optimal dispatch. First, a dynamic flexibility model explicitly defines the aggregate operational domain while incorporating energy jump variable induced by EV arrivals and departures. Second, a Multi-Task Learning CNN-GRU (MTL-CNN-GRU) architecture exploits inter-parameter coupling to jointly forecast correlated power and energy operating boundaries. Third, a Quantile Regression CNN-GRU (QR-CNN-GRU) model provides probabilistic interval forecasts of energy jump to quantify operational uncertainty. Fourth, a standalone CNN predicts daily energy sums for global consistency, minimizing computational complexity. Finally, a day-ahead dispatch model integrates these outputs to minimize operating costs. Validation using real-world data demonstrates superior performance: (1) boundary forecasting achieves an R2 of 0.914, reducing the Mean Absolute Error (MAE) by 12.8% compared to single-task baselines, (2) energy jump uncertainty quantification improves by 36.5% in Quantile Score and 29.5% in Winkler Score, and (3) operational implementation achieves cost reductions of 5.0% compared to uncontrolled charging and 12.7% compared to the empirical strategy, maintaining 90.7% of the theoretical optimal economic performance. The framework successfully bridges data-driven forecasting and operational requirements, providing actionable flexibility management solutions for grid-interactive EV cluster.