Learning based fusion in ensembles for weather forecasting

In this study, a novel ensemble is proposed to forecast monthly rainfall for sugarcane areas in Queensland, Australia. Multiple ensembles of neural networks have been developed to estimate the amount of monthly rainfall for Innisfail, Queensland Australia. Furthermore, four fusion methods such as average fusion, neural network learning fusion, lowest error based neural network fusion and neural network based particle swarm optimization fusion were proposed and evaluated. The obtained models were compared against alternative models and climatology where results revealed higher accuracy with ensemble generated outlooks. Among the ensembles with four fusion methods, an ensemble of feed forward neural networks using resilient backpropagation algorithm and particle swarm optimization produced the highest accuracy with 166.71 mms Root Mean Square Error (RMSE).