Abstract: Accurate prediction of river water levels is of great significance for a high-quality dispatching and management of the water resources in the basin, but the prediction accuracy of a traditional machine learning model is usually difficult to improve further due to the complexity and nonlinear correlation of hydrological data. This paper develops a more accurat4e model of multivariable water level prediction based on a convolution radial basis network. It extracts the spatiotemporal features of hydrological variables fully in parallel, using a multi-layer two-dimensional convolution network; then it achieves high-accuracy predictions of river water levels through a radial basis function network. To verify this model, a numerical experiment is carried out focusing on the predictions of the Qingxi River basin in Sichuan. The results show that compared with four classical models, its root-mean-square error is reduced by 0.039 at least, and the Nash efficiency coefficient increased by 0.056 at least. Compared with the AR-RNN model with the same inputs, its maximum error and root-mean-square error are reduced by 0.348 and 0.017 respectively, verifying its good applicability and effectiveness in basin water level predictions.

Key words: water level prediction, multivariable sequence, convolution network, radial basis function, feature extraction, correlation analysis