Probabilistic ultra-short-term solar photovoltaic power forecasting using natural gradient boosting with attention-enhanced neural networks

Abstract

Probabilistic forecasting provides insights in estimating the uncertainty of photovoltaic (PV) power forecasts. In this study, an innovative probabilistic ultra-short-term PV power forecasting framework that integrates natural gradient boosting (NGBoost) and deep neural networks is developed. Specifically, an attention-enhanced neural network combining convolutional neural networks (CNN) and bidirectional long short-term memory (BiLSTM) networks is employed for feature engineering to extract abstract features from time-series data. The extracted features are then fed into an optimized NGBoost model to yield probabilistic forecasts. In comparison to the benchmark models, i.e., the recently reported quantile regression (QR)-based deep learning methods and NGBoost, the proposed model demonstrates an enhanced ability to capture variation patterns in PV power output, further improving the forecast skill score by approximately 15–60 % in deterministic forecasting. In terms of probabilistic forecasting, the proposed model shows superior forecast reliability and sharpness compared to all benchmark methods. Its continuous ranked probability score (CRPS) ranges from 0.0710 kW to 0.0898 kW, achieving reductions of approximately 21–43 % over QR-based models and 29–40 % over NGBoost. Furthermore, within confidence intervals of 10–90 %, the proposed model consistently maintains higher coverage probabilities along with narrower average forecast intervals, as evidenced by a lower Winkler score (WS) than the benchmark models. The findings of this study provide insightful references for probabilistic PV power forecasting research, contributing to efficient solar power management and dispatch.