Dynamic Neural Networks for Nonstationary Hydrological Time Series Modeling

Evidence of nonstationary trends in hydrological time series, which result from natural and/or anthropogenic climatic variability and change, has raised a number of questions as to the adequacy of conventional statistical methods for long-term (seasonal to annual) hydrologic time series forecasting. Most conventional statistical methods that are used in hydrology will suffer from severe limitations as they assume a stationary time series. Advances in the application of artificial neural networks in hydrology provide new alternative methods for complex, nonstationary hydrological time series modeling and forecasting. An ensemble approach of competitive recurrent neural networks (RNN) is proposed for complex time-varying hydrologic system modeling. The proposed method automatically selects the most optimally trained RNN in each case. The model performance is evaluated on three well-known nonstationary hydrological time series, namely the historical storage volumes of the Great Salt Lake in Utah, the Saint-Lawrence River flows at Cornwall, and the Nile River flows at Aswan. In each modeling experiment, forecasting is performed up to 12 months ahead. The forecast performance of the proposed competitive RNN model is compared with the results obtained from optimal multivariate adaptive regression spline (MARS) models. It is shown that the competitive RNN model can be a good alternative for modeling the complex dynamics of a hydrological system, performing better than the MARS model, on the three selected hydrological time series data sets.