Optics-guided and physics-guided multisource deep learning framework for short-term photovoltaic power forecasting

Accurate short-term photovoltaic (PV) power forecasting is essential for secure grid operation and economical dispatch under high renewable penetration. However, PV output exhibits strong nonlinear and nonstationary behavior driven by rapidly changing meteorological conditions, the evolution of cloud optical properties, and complex system operating states governed by solar radiative transfer. To address these challenges, this paper proposes an optics- and physics-guided multisource deep learning framework for multi-horizon PV power forecasting. The proposed model integrates historical PV power, ground-based meteorological measurements, numerical weather prediction (NWP) grids, and plant operation signals into a unified spatio-temporal representation that explicitly reflects the optical modulation of solar irradiance by clouds and aerosols. A convolutional encoder first extracts compact spatial features from NWP irradiance and cloud-cover maps, implicitly capturing the optical evolution of cloud fields, while a temporal fusion module based on a Transformer architecture models long-range dependencies across heterogeneous time series. To ensure physically and optically consistent predictions, domain knowledge is incorporated through optics- and physics-guided loss functions that penalize violations of plant capacity constraints, enforce near-zero nighttime power according to solar elevation, and regularize excessive ramping between consecutive predictions that would be incompatible with realistic cloud-shadow dynamics. In addition, a quantile regression head is employed to generate predictive intervals, providing uncertainty information that is crucial for operational decision making. Comprehensive experiments on a synthetic yet physically plausible PV plant dataset demonstrate that the proposed framework achieves lower RMSE and MAE than persistence, gradient boosting, LSTM, and vanilla Transformer baselines across multiple forecasting horizons. Ablation studies further confirm the effectiveness of the spatial (optical) encoder and optics-/physics-guided constraints in improving both accuracy and robustness under rapidly changing weather and cloud-optics conditions.