Evaluating the impact of improved filter-wrapper input variable selection on long-term runoff forecasting using local and global climate information

Abstract

Long-term runoff forecasting (LRF) is great important for water resources management. Employing accurate local and global climate information can significantly enhance the accuracy of LRF. The filter-wrapper input variables selection (FWIVS) method can efficiently select local and global climate indices for LRF. However, binary metaheuristic algorithms (BMAs) in wrapper input variable selection (WIVS), such as golden jackal optimization (GJO) and gray wolf optimizer (GWO), frequently encounter local optima, which significantly impact LRF accuracy and remain unexplored. Additionally, previous FWIVS methods overlooked physical relationship between local and global climate information. Consequently, an improved GJO (IGJO) that considers the complementarity of local weather and global climate indices and prevents local optima was proposed. Four machine learning models and an integrated model (IM) were used as learning models for WIVS. Finally, eight BMAs, including IGJO, improved GWO and so on, were constructed and integrated with five learning models to develop forty wrappers for evaluating impact of various BMAs and models on WIVS and LRF. The Jinsha River was served as a case study. Results demonstrated that the impact of the improved BMAs on wrapper fitness and variables selection was higher than learning model. IGJO retained critical variables for WIVS, especially local weather variables. Moreover, IGJO effectively reduced redundant information and extracted more robust complementary global climate indices for WIVS. In addition, LSTM-IGJO exhibited the greatest enhancement in wrapper fitness and LRF accuracy, achieving a NSE of 0.92 and a MAE of 575.83 m3/s. Specifically, this finding indicated that selecting the key variables and selecting the stronger complementary global climate indices instead the redundant local weather variables for retaining key information, while eliminating redundant information, can efficiently improve the LRF accuracy.