To address the challenges of lacking physical consistency and poor generalization under intense irradiance fluctuations in purely data-driven photovoltaic (PV) forecasting models, existing hybrid methods predominantly employ loose feature-concatenation strategies, which fail to capture deep physical dynamics. This study proposes an architecturally innovative Physics-Embedded Gated LSTM (PhysGated-LSTM) model. Breaking through the bottleneck of shallow interaction, the model utilizes the Crested Porcupine Optimizer (CPO) to identify double-diode parameters and directly maps them into the LSTM gating units, thereby achieving the endogenous modulation of neuronal memory flow via physical mechanisms. Experimental results demonstrate that the model achieves an R 2 of 0.936 for 30-minute forecasting, with RMSE and MAE reduced by 18.18% and 23.81% compared to the baselines, respectively. Crucially, under extreme conditions with rapid irradiance mutations, the model maintains an R 2 of 0.992 (a 13.2% improvement over the baseline of 0.876). This effectively eliminates phase lag and non-physical overshooting, validating its superior physical compliance. • PhysGated Architecture: Embeds DDM parameters into LSTM gates to modulate memory flow via physical laws. • Dynamics Sensing: Uses high-frequency I-V data to decouple microscopic fluctuations from macroscopic factors. • Error Suppression: Achieves 18.18% RMSE reduction using consistency loss and physical residual modulation. • Accuracy Growth: PG-LSTM achieves R 2 of 0.992 under irradiance fluctuations via its “physical brake" effect.
Liu et al. (Thu,) studied this question.