Improving the prediction accuracy of monthly streamflow using a data-driven model based on a double-processing strategy

Streamflow forecasting has great significance in water resource management, particularly for reservoir operation. However, accurately predicting streamflow is challenging due to the non-stationary characteristics of hydrologic processes and the effects of noise. To improve monthly streamflow forecasting, this study proposes a data-driven model based on a double-processing strategy, which combines singular spectrum analysis (SSA), improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and extreme learning machine (ELM) approaches. In the proposed double-processing model, called SSA-ICEEMDAN-ELM, the original streamflow series are first processed via SSA for denoising; then, the processed series are reprocessed via ICEEMDAN to decompose them into relatively stationary sub-series; finally, these sub-series are modelled using ELM. The performance of the proposed model is tested for one-month-ahead prediction using streamflow data from the Caojiahu and Shibalipu reservoirs in the Gulang River Basin. In addition, the proposed double-processing model is compared with four single-processing models, namely, empirical mode decomposition (EMD)-ELM, ensemble EMD (EEMD)-ELM, ICEEMDAN-ELM and SSA-ELM, and two single models without any processing, namely, autoregressive integrated moving average (ARIMA) and ELM. The results show that: (a) the four single-processing models have higher prediction accuracy than the single models, and the performance of the SSA-ELM model is the best of these single-processing models, implying that noise in hydrological series cannot be ignored; (b) the proposed SSA-ICEEMDAN-ELM model is superior to the single-processing models and single models, demonstrating that the double-processing approach can further improve streamflow prediction accuracy. Thus, the proposed model, which is a promising method that is expected to benefit reservoir management, can better reduce the influence of noise and capture the dynamic characteristics of hydrological series.