Contextual retrieval-augmented tree ensemble boosting for multi-horizon solar PV power forecasting

Probabilistic forecasting of photovoltaic (PV) power generation is essential for reliable grid operation under high renewable-energy penetration. Existing approaches either separate point prediction from uncertainty estimation or use historical analogues only as external pre- or post-hoc augmentation, applying static similarity weights that fail to adapt to meteorological variability. This paper proposes Contextual retrieval-augmented tree ensemble Boosting (CrateBoost), a probabilistic forecasting framework whose central mechanism is the injection of regime-matched historical analogues directly into the natural-gradient updates of a joint mean–variance booster at every boosting iteration, so that retrieval shapes the learning trajectory itself rather than correcting it externally. The framework is organized as a multi-stage pipeline comprising: (i) data pre-processing with outlier treatment and physics-informed feature engineering; (ii) a precomputed target-supervised PowerUMAP embedding, constructed from irradiance, thermal, and electrical features and indexed with Facebook AI Similarity Search (FAISS), which provides the similarity manifold over which retrieval operates and ensures that retrieved neighbours correspond to comparable PV-conversion regimes rather than superficially similar raw weather; (iii) the core boosting loop, which jointly optimizes mean and variance heads under a Gaussian negative log-likelihood objective with causality-preserving neighbour selection, blends retrieval-augmented gradients at every iteration, and modulates analogue influence according to local clear-sky-ratio dispersion so that the model leans on parametric estimates under stable conditions and on retrieved analogues under broken-cloud regimes; and (iv) a post-hoc Ensemble Model Output Statistics (EMOS) recalibration applied on a held-out validation set to deliver calibrated marginal predictive distributions. The framework is evaluated on six years (2017–2022) of operational data from the Yulara Solar System (Australia) under a ten-fold walk-forward cross-validation protocol at 1-, 3-, and 6-h horizons. Benchmarked against eight state-of-the-art methods spanning traditional machine learning (LightGBM, XGBoost, CatBoost, and NGBoost) and deep learning architectures (TFT, LSTM, Autoformer, and N-BEATS), the proposed method achieves consistent gains across horizons, outperforming the strongest baseline at each horizon by 4 to 14% in normalized mean absolute error and by 1.5 to 15% in the continuous ranked probability score.