Daily Rainfall Forecasting Using Meteorology Data with Long Short-Term Memory (LSTM) Network

Abstract

Rainfall is a natural climatic phenomenon and prediction of its value is crucial for weather forecasting. For time series data forecasting, the Long Short-Term Memory (LSTM) network is shown to be superior as compared to other machine learning algorithms. Therefore, in this research work, a LSTM network is developed to predict daily average rainfall values using meteorological data obtained from the Malaysian Meteorological Department for Kuching, Sarawak, Malaysia. Six daily meteorology data, namely, minimum temperature (°C), maximum temperature (°C), mean temperature (°C), mean wind speed (m/s), mean sea level pressure (hPa) and mean relative humidity (%) from the year 2009 to 2013 were used as the input of the LSTM prediction model. The accuracy of the predicted daily average rainfall was assessed using coefficient determinant (R2) and Root Mean Square Error (RMSE). Contrary to the common practice of dividing the whole available data set into training, validation and testing sub-sets, the developed LSTM model in this study was applied to forecast the daily average rainfall for the month December 2013 while training was done using the data prior of this month. An analysis on the testing data showed that, the data is more spread out in the testing set as compared to the training data. As LSTM requires the right setting of hyper-parameters, an analysis on the effects of the number of maximum epochs and the mini-batch size on the rainfall prediction accuracy were carried out in this study. From the experiments, a five layers LSTM model with number of maximum epoch of 10 and mini-batch size of 100 managed to achieve the best prediction at an average RMSE of 20.67 mm and R2 = 0.82.