Probabilistic short-term power prediction for building-integrated photovoltaic-storage systems in remote Inner Mongolian pastoral buildings using selective residual fusion
Year: 2026
Type: article
Abstract: Reliable short-term power prediction can inform photovoltaic utilisation and reserve planning, yet field evidence from building-integrated photovoltaic (BIPV)-storage systems in dispersed pastoral buildings remains limited. This study examines a roof-integrated BIPV system serving an Inner Mongolian pastoral residence and uses a 53-day early-spring field record at 5-min resolution for 60 min-ahead power prediction. After data-quality auditing, a Bayesian-optimised gradient-boosting residual selective fusion framework (BO-GBRSF) was developed with random forest (RF) as the primary predictor and CatBoost, XGBoost and ExtraTrees as residual-information providers. On the locked day-holdout test set, the combined configuration achieved root mean square error (RMSE) = 0.2317 kW, mean absolute error (MAE) = 0.1666 kW and R 2 = 0.7961; the corresponding 95 % confidence intervals were 0.1905–0.2684 kW, 0.1299–0.2028 kW and 0.6723–0.8682. Paired moving-block bootstrap analysis did not establish a statistically significant RMSE difference from RF or ExtraTrees. The 80 %, 90 % and 95 % prediction intervals achieved coverage probabilities of 0.8356, 0.9623 and 0.9946. In controlled comparisons, BO-GBRSF had lower aggregate errors than six deep-learning baselines, while quantile regression forests (QRF) produced the sharpest probabilistic intervals and the lowest proper-score values. Errors concentrated under high-irradiance and midday conditions. The forecasts may inform future reserve and storage-margin decisions; battery-control benefits were not evaluated.
Institution: Inner Mongolia University of Technology
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,610
OpenAlex ID: W7203681083