Short-Term Forecasting of Rainfall Using Sequentially Deep LSTM Networks: A Case Study on a Semi-Arid Region

Abstract

Weather prediction is a key aspect of today’s society and its activities. Accurate predictions are crucial for efficiently organizing human activities, particularly in the agricultural, transportation and energy sectors. In this paper, two deep neural networks, designed based on a long short-term memory architecture, are developed to predict the occurrence of rainfall events and the respective amount of rainfall on the island of Nisyros in the south Aegean Sea. Two deep neural networks are developed, serving two different learning tasks. The first network acts as a classifier that assesses whether it is going to rain or not and, sequentially, the second network performs a regression task, quantifying the anticipated amount of rainfall. The performance of such prediction models is highly dependent on input sequences. Among others, the lookback time window shapes those input sequences by determining the number of past time steps to be taken into account. The ideal time window for the classifier involves 24 time steps (i.e., 4 h), resulting in increased accuracy levels exceeding 96.45%. The predictions of the regression model, which has the same lookback time window, feature low errors, measured as 6.635 and 1.411 mm using the mean-square-error and mean-absolute-error indices, respectively.

Keywords

• Deep Neural Networks
• short-term forecasting
• LSTM
• rainfall