Evolving Neural Network Using Hybrid Genetic Algorithm and Simulated Annealing for Rainfall-Runoff Forecasting

Accurately rainfall-runoff forecasting modeling is a challenging task. Recent neural network (NN) has provided an alternative approach for developing rainfall-runoff forecasting model, which performed a nonlinear mapping between inputs and outputs. In this paper, an effective hybrid optimization strategy by incorporating the jumping property of simulated annealing (SA) into Genetic Algorithm (GA), namely GASA, is used to train and optimize the network architecture and connection weights of neural networks for rainfall-runoff forecasting in a catchment located Liujiang River, which is a watershed from Guangxi of China. This new algorithm incorporates metropolis acceptance criterion into crossover operator, which could maintain the good characteristics of the previous generation and reduce the disruptive effects of genetic operators. The results indicated that compared with pure NN, the GASA algorithm increased the diversity of the individuals, accelerated the evolution process and avoided sinking into the local optimal solution early. Results obtained were compared with existent bibliography, showing an improvement over the published methods for rainfall-runoff prediction.