Water level forecasting using a hybrid algorithm of artificial neural networks and local Kalman filtering

Abstract

The dynamic processes in the tidal reaches of the Yangtze River lead to the complexity of short-term water level forecasting. Historical data of daily water level are obtained for the lower reaches (Anqing–Wuhu–Nanjing) of the Yangtze River. Stationary time series of water level is derived by making the first-order difference with the raw datasets. An artificial neural network–Kalman hybrid model is proposed for water level forecasting, in which the Kalman filtering is introduced for partial data reconstruction. The model is calibrated with the hydrologic daily water level data of years 2014–2016 for MaAnshan station. Comparing with the traditional artificial neural network model, daily water level predictions are improved by the hybrid algorithm. Discrepancies appear under the circumstance of sharp variations of water level observations. Moreover, the implementation strategy of Kalman filtering algorithm is explored, which indicates the superiority of local Kalman filtering.