Parameter Automatic Calibration Approach for Neural-Network-Based Cyclonic Precipitation Forecast Models

Abstract

This paper presents artificial neural network (ANN)-based models for forecasting precipitation, in which the training parameters are adjusted using a parameter automatic calibration (PAC) approach. A classical ANN-based model, the multilayer perceptron (MLP) neural network, was used to verify the utility of the proposed ANN–PAC approach. The MLP-based ANN used the learning rate, momentum, and number of neurons in the hidden layer as its major parameters. The Dawu gauge station in Taitung, Taiwan, was the study site, and observed typhoon characteristics and ground weather data were the study data. The traditional multiple linear regression model was selected as the benchmark for comparing the accuracy of the ANN–PAC model. In addition, two MLP ANN models based on a trial-and-error calibration method, ANN–TRI1 and ANN–TRI2, were realized by manually tuning the parameters. We found the results yielded by the ANN–PAC model were more reliable than those yielded by the ANN–TRI1, ANN–TRI2, and traditional regression models. In addition, the computing efficiency of the ANN–PAC model decreased with an increase in the number of increments within the parameter ranges because of the considerably increased computational time, whereas the prediction errors decreased because of the model’s increased capability of identifying optimal solutions.