A CNN-LSTM-PINN Photovoltaic Power Forecasting Method Based on Data–Mechanism Fusion
Abstract
With the increasing penetration of photovoltaic power in power systems, accurate photovoltaic power forecasting is important for dispatch optimization, and renewable energy accommodation. To address the problem that data-driven models may ignore photovoltaic generation mechanisms and produce physically inconsistent forecasting results, this paper proposes a CNN-LSTM-PINN photovoltaic power forecasting method based on data–mechanism fusion. The proposed method introduces physics-informed neural networks into a CNN-LSTM temporal forecasting structure, transforms the physical relationship among photovoltaic power, irradiance, and temperature into differentiable constraints, and constructs physical constraint losses based on partial-derivative direction relationships and power boundary conditions, enabling the model to fit historical data while satisfying photovoltaic generation mechanisms. Experimental results show that the proposed method achieves RMSE, MAE, and R2 values of 1.0583, 0.5727, and 0.8912 on the photovoltaic plant dataset, respectively, outperforming PINN CNN-LSTM, LSTM, and traditional machine learning models. Under a 0.30 noise pertur-bation level, its RMSE increases by only 2.86%, demonstrating good stability and anti-disturbance capability.