Answers are grounded in this paper and cite their sources.
Accurate probabilistic photovoltaic (PV) power forecasting is essential to reliable operation of power systems with high renewable penetration. The rapid changes in the weather lead to strong non-stationarity in PV generation and degrade both prediction accuracy and uncertainty estimation. We propose a weather-aware contrastive state-space learning framework for probabilistic forecasting of PV power. Our framework integrates a selective state-space encoder for efficient long-range temporal modeling, a contrastive learning module that structures representations according to irradiance-ramp-based weather regimes, and a quantile forecasting head for direct uncertainty estimation. This design does not explicitly generate weather scenarios but adapts the learned representations to varying atmospheric conditions. The framework was validated using 118,865 operational measurements collected from a 350 kWp grid-connected photovoltaic system under real operating conditions. The experimental task is formulated as a short-term one-step-ahead PV power forecasting problem using a historical input window of 24 consecutive observations. On the full evaluation set, the proposed model achieved an RMSE of 0.318, an MAE of 0.251, a MAPE of 11.02%, and a CRPS of 0.244 on the original PV power scale after inverse transformation from the normalized training representation, outperforming the persistence, recurrent, Transformer-based, and state-space baselines that were evaluated. Under high irradiance-variability conditions on the evaluated photovoltaic installation, the RMSE was 36% lower than that of LSTM and 17% lower than that of the independent Mamba baseline. Ablation experiments also showed the complementary contributions of selective state-space modeling and weather aware contrastive learning. These results demonstrate the potential of weather-aware state-space representation learning for efficient probabilistic PV forecasting under non-stationary operating conditions.
No takes yet. Share an insight, caveat, or question.
Asiri et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: