Application of Deep Reinforcement Learning Techniques for Power Generation Forecasting of a PV Power Plant

Deep Reinforcement Learning (DRL) has recently been widely applied as a powerful Artificial Intelligence (AI) technique in electrical engineering tasks. This paper presents the performance of three DRL techniques, namely Advantage Actor-Critic (A2C), Proximal Policy Optimization (PPO), and Twin Delayed DDPG (TD3) for power generation forecasting of a PV power plant. Each DRL technique was set with comparable conditions and was trained on 62 days of historical data and weather conditions of a PV power plant. Three types of error metrics are used to indicate the accuracy of power generation forecasting: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). Based on trained data, TD3 showed better accuracy than A2C and PPO. The study results highlight DRL's potential for power generation forecasting.