Interpretable hydropower forecasting method using two-stage decomposition and its management applications

Abstract

As a premier renewable energy source for power generation, hydropower boasts substantial construction scale and robust generation capacity. Amid the global energy crisis, precise hydropower forecasting is becoming increasingly crucial, providing governments with a vital foundation for devising energy strategies and supporting the transition to a stable, clean energy future. Although previously applied methods for predicting hydropower generation have achieved commendable results, it is difficult for them to capture the significant seasonality and inherent uncertainty of hydropower generation, and they tend to overlook the importance of internal explainability in forecasting models. An innovative and explainable hydropower generation prediction model consisting of a two-stage decomposition method for input data processing, polar lights optimization (PLO) for hyperparameter optimization, and a temporal fusion transformer (TFT) is employed to address the above limitations. First, a series of subsequences is adaptively obtained using a two-stage decomposition strategy, allowing for a more thorough exploration of the inherent wave characteristics of complex short-term hydroelectric power generation sequences. Subsequently, polar lights optimization was utilized to intelligently and effectively optimize the parameters in the TFT model, guaranteeing the steadfastness and dependability of the prognostic structure. The empirical results indicate that the proposed model exhibits good predictive performance on the evaluated test sets in short-term hydropower forecasting datasets. By exploring the interpretation of hydropower forecasting results, decision-makers can gain a better understanding of future hydropower production capacity, thereby optimizing resource allocation, managing risks, and ensuring the stability of electricity supply.