Collaborative modeling and scheduling for generator and energy storage hybrid power systems using multitask prediction and reinforcement learning

Abstract To improve modeling accuracy and operational response in generator–storage hybrid power systems, this paper proposes an AI-enhanced collaborative modeling method integrating a multi-task attention neural predictor and a policy-gradient reinforcement learning scheduler. The predictor adopts an encoder–attention–decoder structure, and the scheduler employs the Proximal Policy Optimization (PPO) algorithm for adaptive decision-making under dynamic disturbances. Experiments were conducted in a MATLAB/Simulink–Python co-simulation environment using 2.4 × 10⁵ mixed samples from public and laboratory datasets. Compared with MPC, Heuristic-EMS, and NN-Predict baselines, the proposed method achieved a MAPE of 4.2%, an energy utilization efficiency of 92.4%, a SOC-RMSE of 3.6%, and an average response time of 36 ms. Repeated experiments over ten independent random seeds further showed that the proposed method maintained stable performance under load fluctuation, renewable-output disturbance, and temporary generator-failure scenarios, confirming its robustness and engineering adaptability. The results verify that the method effectively enhances the intelligent operation and optimization capability of generator–storage integrated systems.