A Neural Network Ensemble Incorporated with Dynamic Variable Selection for Rainfall Forecast

Abstract

his paper presents a novel ensemble model of artificial neural networks for rainfall forecast incorporating dynamic variable selection. In the first phase of the model, meteorological variables optimal to the response (here rainfall) are selected with the optimal lag value of the response variable. A dynamic variable selection method named, time series least angle regression (TS-LARS) is applied in this phase. In the second phase, an ensemble comprising artificial neural network (ANN) is constructed. The number of hidden neurons in each ANN are selected randomly to speed up the training of the ensemble. The optimization of each ANN is done by Levenberg Marquart Gradient Descent method. In the third phase of the ensemble, the component ANN models are ranked based on mutual information (MI) between the outputs of the base models and the original output. Before applying MI, we have used independent component analysis (ICA) to extract the base models which are independent with each other. Finally the highest ranked base models are combined to construct the ensemble model. A real world case study has been setup in Fukuoka city, Japan. Daily rainfall data from 1990 to 2010 with relevant meteorological variables are extracted to construct the data. The empirical results reveal that, the use of TS-LARS to select most relevant dynamic variables increase the efficiency of the ensemble model, where as the ICA-MI method reduce the number of base models hence reduce the complexity of the ensemble.