Impact of Input Filtering and Architecture Selection Strategies on GRU Runoff Forecasting: A Case Study in the Wei River Basin, Shaanxi, China

Abstract

A gated recurrent unit (GRU) network, which is a kind of artificial neural network (ANN), has been increasingly applied to runoff forecasting. However, knowledge about the impact of different input data filtering strategies and the implications of different architectures on the GRU runoff forecasting model’s performance is still insufficient. This study has selected the daily rainfall and runoff data from 2007 to 2014 in the Wei River basin in Shaanxi, China, and assessed six different scenarios to explore the patterns of that impact. In the scenarios, four manually-selected rainfall or runoff data combinations and principal component analysis (PCA) denoised input have been considered along with single directional and bi-directional GRU network architectures. The performance has been evaluated from the aspect of robustness to 48 various hypermeter combinations, also, optimized accuracy in one-day-ahead (T + 1) and two-day-ahead (T + 2) forecasting for the overall forecasting process and the flood peak forecasts. The results suggest that the rainfall data can enhance the robustness of the model, especially in T + 2 forecasting. Additionally, it slightly introduces noise and affects the optimized prediction accuracy in T + 1 forecasting, but significantly improves the accuracy in T + 2 forecasting. Though with relevance (R = 0.409~0.763, Grey correlation grade >0.99), the runoff data at the adjacent tributary has an adverse effect on the robustness, but can enhance the accuracy of the flood peak forecasts with a short lead time. The models with PCA denoised input has an equivalent, even better performance on the robustness and accuracy compared with the models with the well manually filtered data; though slightly reduces the time-step robustness, the bi-directional architecture can enhance the prediction accuracy. All the scenarios provide acceptable forecasting results (NSE of 0.927~0.951 for T + 1 forecasting and 0.745~0.836 for T + 2 forecasting) when the hyperparameters have already been optimized. Based on the results, recommendations have been provided for the construction of the GRU runoff forecasting model.

Regional characteristics’ impact on the performances of the gated recurrent unit on streamflow forecasting

“…The mean and median values of these metrics also follow the same pattern. This phenomenon has also been reported in relevant studies that employ the LSTM or GRU for streamflow forecasting (Le et al 2019;Kao et al 2020;Wang Q et al 2020). It is caused by the increasing amount of useful information contained in the data at the previous timesteps being excluded with the increased lead time.…”

Improved runoff forecasting based on time-varying model averaging method and deep learning

“…Compared with LSTM, GRU has less gate and less parameters than LSTM, but it can also achieve the same functions as LSTM. The update equations are as shown in Eqs ( 3 )–( 5 ) [ 52 ]: …”

A Hybrid Model Integrating Elman Neural Network with Variational Mode Decomposition and Box–Cox Transformation for Monthly Runoff Time Series Prediction

“…The Wei River Basin (WRB) covers an area of approximately 13.5×10 4 km 2 between 104°00′-110°20′E and 33°50′-37°18′N. Topographically, the elevation of the WRB ranges from 336 m to 3929 m (Wang et al 2020), and a digital elevation model (DEM) of the WRB is presented in Fig. 3.…”