An Improved Genetic Algorithm Coupling a Back-Propagation Neural Network Model (IGA-BPNN) for Water-Level Predictions

Abstract

Accurate water-level prediction is of great significance to flood disaster monitoring. A genetic algorithm coupling a back-propagation neural network (GA-BPNN) has been adopted as a hybrid model to improve forecast performance. However, a traditional genetic algorithm can easily to fall into locally limited optimization and local convergence when facing a complex neural network. To deal with this problem, a novel method called an improved genetic algorithm (IGA) coupling a back-propagation neural network model (IGA-BPNN) is proposed with a variety of genetic strategies. The strategies are to supply a genetic population by a chaotic sequence, multi-type genetic strategies, adaptive dynamic probability adjustment and an attenuated genetic strategy. An experiment was tested to predict the water level in the middle and lower reaches of the Han River, China, with meteorological and hydrological data from 2010 to 2017. In the experiment, the IGA-BPNN, traditional GA-BPNN and an artificial neural network (ANN) were evaluated and compared using the root mean square error (RMSE), Nash–Sutcliffe efficiency (NSE) coefficient and Pearson correlation coefficient (R) as the key indicators. The results showed that IGA-BPNN moderately correlates with the observed water level, outperforming the other two models on three indicators. The IGA-BPNN model can settle problems including the limited optimization effect and local convergence; it also improves the prediction accuracy and the model stability regardless of the scenario, i.e., sudden floods or a period of less rainfall.