Time Series Modelling For Weather Forecasting Using The Long Short Term Memory Method (Case Study: Lampung Climatology Station)

Abstract

Weather significantly influences human activities, including environmental man-agement, health, agriculture, and urban planning. Accurate temperature prediction is essential for managing energy consumption, public health, and agriculture. This study employs Long Short Term Memory (LSTM) networks for weather fore-casting at the Lampung Climatology Station, focusing on minimum, maximum, and average temperatures. LSTM networks, capable of learning long-term de-pendencies in data, have proven effective in various forecasting applications. The research aims to identify the optimal LSTM model for temperature forecasting us-ing time series data from the Lampung Climatology Station. Different LSTM model parameters, such as hidden neurons, batch size, and epochs, were tested to find the best configuration with the smallest Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE) values. The results indicate that the optimal LSTM model for forecasting minimum air temperature (X1) includes 100 hidden neurons, a batch size of 4, and 50 epochs, achieving an RMSE of 0.59 and a MAPE of 1.90%. For maximum air temperature (X2), the best model uses 5 hidden neurons, a batch size of 4, and 50 epochs, with an RMSE of 1.22 and a MAPE of 3.15%. The optimal model for average air temperature (X3) comprises 5 hidden neurons, a batch size of 4, and 150 epochs, achieving an RMSE of 0.73 and a MAPE of 2.10%. These findings demonstrate the effective-ness of LSTM models in accurately forecasting air temperatures, providing valu-able insights for applications ranging from agricultural planning to disaster pre-paredness.