Precipitation prediction using recurrent neural networks and long short-term memory

Abstract

Prediction of meteorological variables such as precipitation, temperature, wind speed, and solar radiation is beneficial for human life. The variable observations data is available from time to time for more than thirty years, scattered each observation station makes the opportunity to map patterns into predictions. However, the complexity of weather variables is very high, one of which is influenced by Decadal phenomena such as El-Nino Southern Oscillation and IOD. Weather predictions can be reviewed for the duration, prediction variables, and observation stations. This research proposed precipitation prediction using recurrent neural networks and long short-term memory. Experiments were carried out using the prediction duration factor, the period as a feature and the amount of data set used, and the optimization model. The results showed that the time-lapse as a shorter feature gives good accuracy. Also, the duration of weekly predictions provides more accuracy than monthly, which is 85.71% compared to 83.33% of the validation data.