Inconsistent Monthly Runoff Prediction Models Using Mutation Tests and Machine Learning

In a changing environment, the increasing inconsistency of runoff series complicates the development of runoff forecasting models. Mutation tests have frequently been employed to assess runoff inconsistency, yet their integration with runoff forecasting is rare. In this study, we proposed a combination of machine learning models based on mutation tests to predict monthly runoff at the Pingshi Station, located in the Lechang Gorge Reservoir of the Pearl River in Guangdong Province, China. Specifically, the mutation points of the monthly runoff were assessed using the Mann-Kendall test and the Moving T-test (MTT). The development of member models, including Artificial Neural Networks (ANN) and Support Vector Machines (SVM), utilized both the original runoff series and the sub runoff series identified after the first mutation point. Tele-connected factors and historical monthly runoffs served as candidate input variables. These variables were selected through linear and nonlinear filter methods and further refined by a greedy search based on 10-fold cross-validation. The improvement in overall forecasting performance by combining the ANN, SVM, and the Simple Average Method (SAM) was analyzed. The results showed that, for the validation dataset, the root mean square errors of the combined models with MTT decreased by about 14–19% compared to the member models with MTT, and 10–12% compared to the combined models without mutation tests. The corresponding Nash-Sutcliffe efficiency coefficients increased by approximately 35–45% and 12–14%, respectively. This study highlights the effectiveness of integrating mutation tests, machine learning, and combining models to enhance the forecasting performance of inconsistent monthly runoff.