Typhoon forecasting in the Philippines using an optimal Multilayer Feedforward Artificial Neural Network model trained in Resilient Propagation algorithm

Timely and accurate forecasting is indispensable during typhoons, and a rapid, efficient, and inexpensive forecasting method is a positive approach in reducing risks and fatalities in the population. This study developed 16 models for typhoon-forecasting in the Philippines using an artificial neural network (ANN). The models were based on three criteria: number of hidden neurons according to Kolmogorov's theorem; activation function (TANH or SIGMOID); and initial neuron weights (random or calculated). The optimal ANN model has 16 hidden neurons, calculated initial neuron weights, and used sigmoid activation function that obtained an error-rate of 0.070627 and mean prediction accuracy of 97.375 percent. The ANN model was trained using Resilient Propagation (RPROP) algorithm following the Multilayer Feedforward ANN (MLFANN) architecture capable of attaining ideal results with low Root-Mean-Square Errors. Neural network training used Join Typhoon Warning Center (JTWC) typhoon data from 1945 to 1995, and neural network evaluation used data from 1996 to 2013.