Robust Runoff Prediction With Explainable Artificial Intelligence and Meteorological Variables From Deep Learning Ensemble Model

Accurate runoff forecasting plays a vital role in issuing timely flood warnings. Whereas, previous research has primarily focused on historical runoff and precipitation variability while disregarding other factors' influence. Additionally, the prediction process of most machine learning models is opaque, resulting in low interpretability of model predictions. Hence, this study develops an ensemble deep learning model to forecast runoff from three hydrological stations. Initially, time-varying filtered based empirical mode decomposition is employed to decompose the runoff series into several internal mode functions (IMFs). Subsequently, the complexity of each IMF component is evaluated by the multi-scale permutation entropy, and the IMFs are classified into high- and low-frequency portions based on entropy values. Considering the high-frequency portions still exhibit great volatility, robust local mean decomposition is adopted to perform secondary decomposition of the high-frequency portions. Then, the meteorological variables processed by the Relief algorithm and variance inflation factor features are employed as inputs, the individual subsequences of secondary and preliminary decomposition as outputs to the bidirectional gated recurrent unit and extreme learning machine models. Random forests (RF) are introduced to nonlinear ensemble the individual predicted sub-models to obtain the final prediction results. The proposed model outperforms other models in various evaluation metrics. Meanwhile, due to the opaque nature of machine learning models, shapley is employed to assess the contribution of each selected meteorological variable to the long-term trend of runoff. The proposed model could serve as an essential reference for precise flood prediction and timely warning.