Abstract:

This study investigated the complex nonlinear and nonstationary characteristics of runoff time series in a changing environment. We obtained accurate and reliable runoff prediction results by constructing long-short term memory (LSTM) combination models on the basis of time frequency analysis methods. The combination models were then applied to the Manas River, Xinjiang, China, a typical inland river in an arid area. Empirical mode decomposition (EMD), variational mode decomposition (VMD), and discrete wavelet transform (DWT) were first applied to decompose the original runoff series into several subsequences with different frequencies and characteristics. Second, we used the previous runoff, precipitation, temperature, and atmospheric circulation factors as the input variables of the LSTM model. At the same time, the optimal predictor of each subsequence was determined according to random forest classification and the Pearson correlation coefficient. Finally, the combined models were based on a combination of VMD, EMD, and DWT with LSTM. Accordingly, they were called VMD-LSTM, EMD-LSTM, and DWT-LSTM and are proposed and applied for runoff forecasting. The total output of all submodules was treated as the final forecasting result for the original runoff. A single back propagation neural network, a single extreme learning machine, and a single LSTM were adopted as comparative forecast models. The results indicated that the VMD-LSTM model had the best forecasting performance among all the models in terms of its Nash-Sutcliffe error (NSE=0.930), root mean square error (RMSE=0.385), and coefficient of determination (R²=0.940). The extreme value prediction error for the runoff of the VMD-LSTM model was within 15%, and it had a good effect on the overall trend prediction and extreme value tracking regarding the runoff. This result further verified the accuracy and stability of the VMD-LSTM model. On the basis of the above results, the accuracy and stability of the VMD-LSTM model were further verified. Thus, on the basis of sequence decomposition, considering the influence of predictors on subsequences can help promote accurate and stable prediction results. These research results will provide a reference for river basin water resource planning and dispatching.

Key words: runoff prediction, combination model, time-frequency analysis, long short-term memory