ATTEMPT OF ONE-MONTH WEATHER FORECAST USING MACHINE LEARNING

Abstract

 It is difficult to make one-month weather forecasts using physics-based numerical models. Therefore, the Meteorological Agency announces probability forecasts for temperature and precipitation in three classes rather than numerical forecasts. In this study, we examine the possibility of numerical prediction of average temperature and precipitation using machine learning. We used LSTM suitable for learning time series data and CNN suitable for image learning in the middle layer of the model using ground meteorological observation data and sea surface temperature data as training, validation and evaluation data. Although the average temperature was predicted to rise and fall, there was also a tendency to underestimate summer temperature and overestimate summer and winter temperatures. In the forecast of precipitation, although the trend of increase and decrease can be generally predicted, it was not possible to predict the sudden increase of precipitation. It is also suggested that the sea surface temperature data may deteriorate the prediction accuracy.