Probabilistic Forecasting of Renewable Energy Output Under Uncertainty Using Deep Quantile Learning: A SolarPV Case Study

Abstract

Renewable generation varies with irradiance, cloud cover, temperature, seasonal effects, ramp events and measurement uncertainty in solar-PV systems, while similar uncertainty arises from wind speed and turbulence in wind generation. A single deterministic forecast is therefore not sufficient for reserve scheduling, storage dispatch, market bidding or real-time balancing. This paper presents an attention-based convolutional neural network-bidirectional long short-term memory (CNN-BiLSTM) quantile-regression model for probabilistic renewable-energy forecasting. The model was trained and evaluated on a curated hourly solar photovoltaic subset derived from the ECMWF Ensemble Prediction System solar forecasting dataset. Lagged PV output, ensemble weather predictors, calendar terms and ramp-related variables were used to estimate the 0.05, 0.10, 0.25, 0.50, 0.75, 0.90 and 0.95 conditional quantiles. The 0.50 quantile served as the median point forecast, while the 0.05 and 0.95 quantiles defined the 90% prediction interval. On the held-out test set, the proposed model achieved MAE of 0.041 p.u., RMSE of 0.060 p.u., MAPE of 5.95%, R² of 0.954, mean pinball loss of 0.038, CRPS of 0.044, PICP of 89.1%, PINAW of 0.252 and Winkler score of 0.128. Compared with the strongest Transformer-quantile baseline, the model reduced mean pinball loss from 0.041 to 0.038 and narrowed the interval width from 0.265 to 0.252 while keeping empirical coverage close to the nominal 90% level. The improvement is modest, but it is consistent across probabilistic loss, empirical coverage and interval-width metrics, which is important for short-term operational use