VMDI-LSTM-ED: A novel enhanced decomposition ensemble model incorporating data integration for accurate non-stationary daily streamflow forecasting

Abstract

Accurate daily streamflow forecasting is crucial for effective flood control and water management. However, the non-stationary nonlinearity in actual streamflow poses a challenge to accurate forecasting. While decomposition ensemble models can address non-stationary nonlinear streamflow, they still suffer from low forecast accuracy when dealing with highly non-stationary streamflow. Recent studies have shown that incorporating lagged streamflow into long short-term memory (LSTM) networks, known as data integration (DI), represents an effective approach for streamflow forecasting. Nevertheless, existing decomposition ensemble models do not fully leverage the benefits of recent observations. To enhance the precision of non-stationary streamflow forecasting, we propose an improved decomposition ensemble model based on DI called VMDI-LSTM-ED, which utilizes recent observations to improve prediction while processing the subsignals of Variational mode decomposition (VMD) decomposition using LSTM with Encoder-Decoder framework (LSTM-ED). In order to evaluate the reliability and applicability of VMDI-LSTM-ED and demonstrate its superiority, we conducted model tests in six different basins in the United States and compared VMDI-LSTM-ED with VMD-LSTM, Transformer, and LSTM-ED. The results indicate that VMDI-LSTM-ED yields the best streamflow forecast result across all tested basins, with an average Nash-Sutcliffe Efficiency (NSE) of 0.880 for 1-day ahead forecasts over the six basins; whereas NSE values for VMD-LSTM, Transformer, and LSTM-ED are only 0.687, 0.556, and 0.368 respectively. In addition, VMDI-LSTM-ED is good not only for high-streamflow areas but also for low-streamflow areas, and the prediction effect of peak streamflow is the best.