Enhancing robustness of monthly streamflow forecasting model using gated recurrent unit based on improved grey wolf optimizer

Abstract Accurate and reliable mid- to long-term streamflow prediction is essential for water resources management. However, streamflow series exhibits strong non-stationary and non-linear; thus the traditional single model may fail to capture the characteristic of hydrological series and hard to maintain the robustness of the model. To overcome this problem, the application of empirical wavelet transform combined with improved complete ensemble empirical mode decomposition with adaptive noise (ICEEWT) is firstly to eliminate the redundant noise and decompose original series into several components as inputs. Secondly, gated recurrent unit (GRU) is employed to deep learn the relationship between historical streamflow series and those of the future. Finally, balancing underfitting and overfitting the improved grey wolf optimizer (IGWO) is adopted with the hybrid model to identify the approximate parameter combination. The case study is conducted by monthly streamflow data at Shangjingyou station and Fenhe reservoir station in the upper reaches of Fenhe River. In comparison with parallel experiments using extreme learning machine (ELM) and least squares support vector machine (LSSVM), the validation results demonstrate that the GRU model can exhibit satisfactory performance in monthly streamflow forecasting. Moreover, the standard GRU model is further improved by IGWO, which consistently remains the performance of in-sample and out-sample. The proposed model ICEEWT-IGWO-GRU is superior to the single GRU model average reducing MAE and RMSE by 50% and 52% for two stations, which indicates its superiority with a more robust prediction.