Forecasting water levels at the Yangtze River with neural networks

Abstract

In the last ten years, the application of neural network models has become an emerging field of research in the field of hydrology. In the present study, three different neural network models, namely the Multilayer Perceptron (MLP), the Jordan net, and the Elman net were used for forecasting water levels at Cuntan station, located at the Yangtze River’s upper reaches. The performances of the neural network models were compared with each other and with the results of a multiple linear regression (MLR) model. As input variables for the models, not only were precipitation data and antecedent water levels implemented, but also two climatic variables which are usually left out in the field of neural network modeling: evaporation and snow data. Before the models were adopted, the optimal lead time between the input variables and the model output was determined by means of a cross-correlation analysis. The highly significant correlation between the model input and output already indicated a highly linear relationship. Accordingly, the MLR model showed the best performance, even though the results of the other models are only slightly worse. The good capability of the Jordan net in forecasting high water levels should be investigated further. In predicting water levels in general, the integrated snow data improved the performance of the different models only marginally. However, the integration of evaporation data definitely improved the modeling results.