記憶に基づく推論による天気予測

抄録

We have implemented a weather prediction system, WINDOM, which uses memory-based reasoning (MBR). The observation data from Japan Meteorological Agency (JMA)'s networks, AMeDAS and the surface observation centers, are used directly as a case database. WINDOM runs on the parallel computer AP1000 and predicts the weather around Tokyo several hours ahead by matching the past data in the database with the present observational data. In MBR, the weighting methods for features and its metric definition are important to get high accuracy. In the experiments, four weighting methods based on conditional probability and three calculation methods of similarity are tested, and the best pair of methods is selected empirically. The relation between accuracy and parameters such as the area of the observational data, the quantity of data, and prediction hour are shown. By using 9 years of observational data in the middle and the west part of Japan to predict whether the weather of Tokyo 6 hours ahead would be rainy or not, an accuracy of 87.2% is achieved. The average accuracy of prediction in Kanto-area is slightly worse than that of JMA, but in some prefectures our results are comparable with their average results.