Track: Operations Research

Abstract

Temperature is a very important weather element whose value can change at any time. Forecasting daily temperature is important information for the society, so forecasting with high accuracy is required. The use of neural network methods in forecasting has become increasingly popular lately. Therefore, this study uses the Long Short Term Memory (LSTM) model to forecast temperatures in Bandung. This method doesn’t require parametric assumptions and can be used for data with long time periods, as there is a cell state to overcome vanishing gradients in the Recurrent Neural Network method. This method is used because the data to be inputed is daily historical data for 5 years. The temperature allows for unexpected fluctuations, so that earlier data is expected to detect motion patterns that weren’t executed by the most recent data. Based on the results of the LSTM analysis, the Mean Absolute Percentage Error (MAPE) for the minimum temperature prediction is 5.61% and MAPE for the maximum temperature prediction is 3.84%. Thus, forecasting temperatures using LSTM can provide high accuracy results for short-term prediction using data with a long period of time.