Short-term photovoltaic (PV) power forecasting is crucial for secure and economical grid operation, yet remains challenging under fast and nonstationary irradiance fluctuations. This paper presents a plant-level TA–SH–LSTM framework that integrates temporal attention into an LSTM encoder to highlight informative subsegments for improved ramp tracking and peak localization and applies budget-aware Successive Halving to jointly tune window length and key hyperparameters under a fixed training budget. To enhance PV-engineering interpretability, we establish a first-order thermal inertia surrogate that explicitly links module temperature to ambient temperature and irradiance, and evaluate robustness across irradiance-tercile regimes within the observation window. Experiments on two real PV plants from the Kaggle Solar Power Generation dataset demonstrate consistent gains over a baseline LSTM and an SH-tuned LSTM. On Plant 1, MAE/RMSE decreases from 1141.1/2066.6 kW to 223.4/424.6 kW and R2 increases from 0.932 to 0.997. Without retraining, the model transfers to Plant 2 with 286.1 kW MAE, 477.1 kW RMSE, and R2 = 0.993, confirming strong cross-site generalization and practical utility under varying operating conditions.
Liu et al. (Sat,) studied this question.