A novel intelligent interval prediction framework for daily generation of hydropower reservoir

Accurate interval prediction of hydropower generation is crucial for optimizing reservoir operation and enhancing the resilience of clean energy systems. To address the challenges of data quality, model integration, and error accumulation in existing methods, this study proposes an intelligent interval prediction framework for cascade hydropower systems. First, a data preprocessing strategy is developed to systematically perform outlier detection, missing value identification, and imputation on the original "meteorological-hydrological-hydropower" time series of the hydropower system, providing high-quality data for subsequent modeling. A Gated Recurrent Unit (GRU) is then integrated with the Deep Autoregressive Recurrent (DeepAR) probabilistic forecasting model to achieve unified probabilistic prediction under non-negativity constraints of hydropower generation. Furthermore, an Error Correction (EC) method is introduced to ensure accurate, reliable, and stable interval predictions. Engineering case study results from cascade hydropower reservoirs in the Jinxi River Basin, Fujian, China, demonstrate that the proposed method significantly outperforms other comparative models. As an example in Chitan reservoir, for an 8-day forecast period, the proposed method achieves a Nash-Sutcliffe Efficiency coefficient of 0.8584 and a Prediction Interval Normalized Average Width of 946.8693, which represent a 16.06% improvement and a 12.26% reduction, respectively, compared to the DeepAR method's 0.7396 and 1079.193, indicating the higher accuracy of the proposed method. Evaluations under different hydrological conditions and varying confidence intervals consistently confirm the superiority and adaptability of the approach. This research provides an effective solution for daily hydropower generation interval prediction, supporting uncertainty quantification and sustainable water resource management.