A data-to-decision adaptive bidding framework for wind-storage systems in coupled markets using off-policy actor–critic learning

This paper proposes a data-to-decision adaptive bidding framework using off-policy actor–critic learning for wind-storage systems in coupled energy and ancillary service markets. To handle the high-dimensional uncertainty in wind power, a Conditional Generative Adversarial Network (CGAN) is first employed to generate forecast-conditioned scenarios. These scenarios are then clustered by a Gaussian Mixture Model (GMM) into representative operational modes, thereby transforming raw data into actionable knowledge. A novel multi-scenario training architecture based on the Twin Delayed Deep Deterministic Policy Gradient (TD3) is subsequently developed. This architecture trains specialized policies for each operational mode as well as a unified robust policy on mixed scenarios, achieving an optimal balance between specialized performance and generalizable robustness. Once trained, the framework maps observed system states directly to coordinated energy and regulation bids in a single step. Simulation results demonstrate that the proposed framework achieves superior performance over baseline DRL algorithms and competitive economic performance compared with classical optimization methods, while enabling sub-second online decision making through policy inference. This favorable balance between economic performance and computational efficiency demonstrates its practical effectiveness for coordinated Wind–ESS market bidding.