Adaptive adjustment of photovoltaic power generation prediction model parameters based on reinforcement learning

To address the problem of traditional prediction models experiencing a sharp drop in accuracy and delayed response due to fixed hyperparameters caused by dynamic meteorological conditions in the context of high photovoltaic grid integration, this paper proposes a parameter adaptive adjustment method for photovoltaic power generation prediction models based on proximal policy optimization reinforcement learning. Based on the LSTM (Long Short Term Memory) prediction model, this method constructs a multidimensional state space that integrates prediction error, irradiance change rate, seasonal encoding, and meteorological mutation signatures. A discrete action space is designed, encompassing learning rate, input window length, and number of hidden units. A multi-objective reward function is applied, balancing accuracy, stability, and responsiveness to sudden changes. This drives the agent to dynamically decide on the optimal parameter combination every six hours. Lightweight online fine-tuning is then combined to achieve real-time reconstruction of the model structure. Experiments on a real photovoltaic dataset demonstrate that, compared with fixed-parameter LSTM, Bayesian optimization parameter tuning, and periodic retraining, this method achieves average MAE (Mean Absolute Error) and RMSE (Root Mean Squared Error) of 8.7 kW and 12.33 kW, respectively, with a median response delay of 23.2 min. Furthermore, the proposed method exhibits lower parameter change rates and greater prediction robustness. This research provides a feasible technical path for high-precision and high-robust photovoltaic power prediction in dynamic meteorological environments, which has important engineering significance for supporting safe grid scheduling and high-proportion renewable energy consumption.