Research on Coupling Knowledge Embedding and Data-Driven Deep Learning Models for Runoff Prediction

Abstract

Accurate runoff prediction is crucial for watershed water resource management, flood prevention, and hydropower station scheduling. Data-driven models have been increasingly applied to runoff prediction tasks and have achieved impressive results. However, existing data-driven methods may produce unreasonable predictions due to the lack of prior knowledge guidance. This study proposes a multivariate runoff prediction model that couples knowledge embedding with data-driven approaches, integrating information contained in runoff probability distributions as constraints into the data-driven model and optimizing the existing loss function with prior probability density functions (PDFs). Using the main stream in the Yellow River Basin with nine hydrological stations as an example, we selected runoff feature factors using the transfer entropy method, chose a temporal convolutional network (TCN) as the data-driven model, and optimized model parameters with the IPSO algorithm, studying univariate input models (TCN-UID), multivariable input models (TCN-MID), and the coupling model. The results indicate the following: (1) Among numerous influencing factors, precipitation, sunshine duration, and relative humidity are the key feature factors driving runoff occurrence; (2) the coupling model can effectively fit the extremes of runoff sequences, improving prediction accuracy in the training set by 6.9% and 4.7% compared to TCN-UID and TCN-MID, respectively, and by 5.7% and 2.8% in the test set. The coupling model established through knowledge embedding not only retains the advantages of data-driven models but also effectively addresses the poor prediction performance of data-driven models at extremes, thereby enhancing the accuracy of runoff predictions.