A neural network-based local rainfall prediction system using meteorological data on the Internet: A case study using data from the Japan Meteorological Agency

Abstract

In this study, we develop and test a local rainfall (precipitation) prediction system based on artificial neural networks (ANNs). Our system can automatically obtain meteorological data used for rainfall prediction from the Internet. Meteorological data from equipment installed at a local point is also shared among users in our system. The final goal of the study was the practical use of “big data” on the Internet as well as the sharing of data among users for accurate rainfall prediction. We predicted local rainfall in regions of Japan using data from the Japan Meteorological Agency (JMA). As neural network (NN) models for the system, we used a multi-layer perceptron (MLP) with a hybrid algorithm composed of back-propagation (BP) and random optimization (RO) methods, and radial basis function network (RBFN) with a least squares method (LSM), and compared the prediction performance of the two models. Precipitation (total amount of rainfall above 0.5mm between 12:00 and 24:00 JST (Japan standard time)) at Matsuyama, Sapporo, and Naha in 2012 was predicted by NNs using meteorological data for each city from 2011. The volume of precipitation was also predicted (total amount above 1.0mm between 17:00 and 24:00 JST) at 16 points in Japan and compared with predictions by the JMA in order to verify the universality of the proposed system. The experimental results showed that precipitation in Japan can be predicted by the proposed method, and that the prediction performance of the MLP model was superior to that of the RBFN model for the rainfall prediction problem. However, the results were not better than those generated by the JMA. Finally, heavy rainfall (above 10mm/h) in summer (Jun.–Sep.) afternoons (12:00–24:00 JST) in Tokyo in 2011 and 2012 was predicted using data for Tokyo between 2000 and 2010. The results showed that the volume of precipitation could be accurately predicted and the caching rate of heavy rainfall was high. This suggests that the proposed system can predict unexpected local heavy rainfalls as “guerrilla rainstorms.”