Cause-driven Streamflow Forecasting Framework Based on Linear Correlation Reconstruction and Long Short-term Memory

Abstract

Reasonable runoff forecasting is the foundation of water resource management. However, the impact of environmental change on streamflow was not fully revealed due to the lack of enough streamflow features in many previous studies. In contrast, too many features also could lead cause undesired problems, including unstable model, interpretation difficulty, overfitting, high computational complexity, and high memory complexity. To address the above problems, this study proposes a cause-driven runoff forecasting framework based on linear-correlated reconstruction and machine learning model and refers to this framework as CSLM. We use variance inflation factor (VIF), pairwise linear correlation (PLC) reconstruction, and long short-term memory (LSTM) to realize this framework, referred to as VIF-PLC-LSTM. Four experiments were conducted to demonstrate the accuracy and efficiency of the proposed framework and its VIF-PLC-LSTM realization. Four experiments compare 1) different filter thresholds of driving factors, 2) different combination prediction features, 3) different reconstruction methods of linear-correlated features, and 4) different CSLM models. Experimental results on daily streamflow data from the Tangnaihai station at the Yellow River source and the Yangxian station at the Han River show that 1) data filtering has the risk of feature information loss, 2) when the streamflow, ERA5L, and meteorology data are used as inputs at the same time, the performance of the model is superior to the combination of other prediction features; the prediction effect of different prediction features, 3) the reconstruction of linear-correlated features is not only better than dimension reduction but also can improve the forecasting performance for streamflow prediction, and 4) among different CSLM models, LSTM is superior to other models.