Model Time Series Dengan Exogenous Variable Menggunakan Long Short Term Memory (LSTM) Untuk Prediksi Curah Hujan Di Kabupaten Sidoarjod

Abstract

Sidoarjo Regency is one of the areas that has about 90% of technically irrigated rice fields with low-lying conditions, so that when rainfall is high, some rice fields experience flooding and can result in decreased crop productivity. Therefore, prediction of rainfall is needed to overcome this. In addition, Sidoarjo district has a weather station close to the airport, making it possible to obtain more complete data as an exogenous variable in predicting rainfall. Besides that, there are several difficulties when including exogenous variables, namely having to ensure that exogenous variables are relevant to the forecasted time series data and possibly missing values or experiencing measurement errors in the data. The method used to predict rainfall is using Long Short Term Memory (LSTM) with the advantage of having a memory block that is able to find the output value most relevant to the given input, and using the backward method to overcome missing values. The results of the study using daily period rainfall data obtained an RMSE value of 13.68 mm when using yesterday's timestep data (t-1) with the proportion of training and testing data of 90%:10% and the best parameters were batch size 32, epoch 100, and neuron 10, and overcome the missing value by using the backward method. While the results of the study using monthly rainfall data obtained an RMSE value of 2.11 mm when using yesterday's timestep data (t-1) with the proportion of training and testing data of 90%:10% and the best parameters were batch size 32, epoch 100, and 30 neurons.