Integrated model and application of non-stationary runoff based on time series decomposition and machine learning

Revealing the fluctuating characteristics of non-stationary runoff series under changing environments can improve the precision of runoff prediction and support water-related project planning. Given the characteristics of non-stationarity, periodicity, and heteroscedasticity of runoff series, the observed runoff data from 2008 to 2018 were collected from Panzhihua, Chenglingji, and Datong stations in the Yangtze River basin, and based on the seasonal-trend decomposition method, the original data was decomposed into periodic sequence, trend sequence, and residual sequence. Combined with the features of each subsequence, an integrated model was applied to obtain the total predicted value of future runoff, and the results were compared with the single model of Prophet, LSTM, and GARCH. The results show that the integrated model combined with time series decomposition and machine learning is superior to the single model in different evaluation indexes, and the simulation accuracy of stations with a strong heteroscedasticity effect is significantly improved. The Nash-Sutcliffe efficiency coefficient of the three stations in the validation period is 0.96, 0.95, and 0.93, respectively, indicating that the model can effectively simulate the runoff fluctuation process in the Yangtze River basin.