Enhancing robustness of monthly streamflow forecasting model using embedded-feature selection algorithm based on improved gray wolf optimizer

Abstract

Accurate streamflow prediction plays an essential role in guaranteeing the sustainable utilization and management of water resources. In recent years, Artificial Intelligence (AI) models have been widely used for flow prediction. The performance of these models depends on the appropriate calibration of the input features and model parameters. Theoretically, the embedded feature selection method directly takes the final prediction model as a prediction indicator, which has the unique advantage of parallel optimization of feature and prediction model parameters compared with other methods. Despite being widely used in many other fields, its streamflow forecasting abilities are thus far unknown. In this paper, an embedded prediction model (EFS-SVMIGWO) with improved gray wolf optimizer (IGWO) and support vector machine (SVM) is proposed based on the principle of embedded feature selection method and validated with monthly runoff prediction at Kizil reservoir station in Xinjiang, China. The validation results demonstrate that the EFS-SVMIGWO model has consistently better accuracy and stable values than the benchmark methods (Including autoregressive integrated moving average, random forest, neural network and SVM models based on filtered selection methods). Moreover, IGWO is compared to differential evolution (DE), particle swarm optimization (PSO), whale optimization algorithm (WOA), sparrow search algorithm (SSA), and gray wolf optimizer (GWO), and the results show that IGWO has better convergence speed and solution quality in feature and model parameter parallel optimization tasks. Overall research and analysis indicate that the EFS-SVMIGWO model can exhibit convincing performance in monthly streamflow forecasting. Thus, it is of great importance to carefully choose the input variables and parameters to develop more effective models for forecasting monthly streamflow time series.