A Unified Prediction-Dispatch Framework for PV Power Scheduling Using Reinforcement Learning

Increased penetration of photovoltaic (PV) generation and electric vehicle (EV) charging stations into the conventional grids introduces significant variability and uncertainty in the system, making the real-time economic dispatch less reliable. This research presents a unified prediction-dispatch framework integrating deep learning and reinforcement learning (RL) to optimize the power generation schedule. For short-term PV power prediction, dynamic clustering of power ramp rate with deep learning-based advanced temporal modeling is used to improve prediction accuracy in highly intermittent scenarios. An EV aggregator model representing a cluster of controllable EVs is used to coordinate with large-scale PV power on the transmission side. The economic dispatch problem is modeled using a reinforcement learning environment for 30-minute intervals that adapts to PV fluctuations and EV charging demands, considering optimized participation factors for conventional generators. Proximal Policy Optimization ensures system security and optimal generator participation while dynamically adjusting PV power generation schedules based on real-time feedback. The RL reward function introduces the cost minimization factor, ramp alignment, and confidence interval penalty factor to optimize the dispatch results. Unlike traditional approaches, this framework integrates prediction and dispatch problems, achieving greater coherence between predictions and actual system behavior. For predictive assessment, an IEEE 30-bus system is considered, and the algorithm shows a 3.65 $/MWh reduction in generation costs compared to conventional methods while enhancing stability, thus providing an economically efficient solution for renewable generation and stochastic load integration.