A Multivariate Bayesian Inference Model for Streamflow Prediction and Interpretation under Uncertainty

Abstract:

Machine learning models have been gaining increasing attentions in understanding and predicting the response of hydrological process. In this study, a tree-based multivariate Bayesian inference model is proposed to simultaneously improve the model predictive accuracy and interpretability. Within the regression tree ensemble (RTE) framework, the proposed model addresses the autocorrelative effects of streamflows through jointly aggregating the univariant and multivariant regression trees using an advanced Bayesian tree-ensemble approach. By doing so, we found that the univariant model can focus on more uncertainties and stochasticity than multivariant models owing to its more sophisticated decision trees. While in some circumstances, those decision trees may overfits the observations due to the autocorrelation structural of streamflow. The multivariant models on the other hand, can preserve such autocorrelative effect and thereby, enhance the overall performance of prediction. The Bayesian tree-ensemble approach also improves the model interpretability upon the well-known permutation feature importance method by taking the consideration of posterior probabilities of every weak leaner and at various flow quantiles. This interpretation strategy thus enables the model to identify the most informative features under various flood magnitudes, which therefore can help practitioners gaining more insight towards the varying roles those features could play in different hydrological processes. The proposed model is applied to three interconnected irrigation drainage areas located in the floodplain of the Yellow river basin. The knowledge of the proposed model can be transferred to other statistical hydrological models for improving predictive accuracy and bridging the information gap between hydrological processes and relevant statistics.