EVALUATION OF LONG-SHORT TERM MEMORY (LSTM) AND SEASONAL AUTOREGRESSIVE INTEGRATED MOVING AVERAGE (SARIMA) MODELS FOR RAINFALL PREDICTION IN SUMATRA AREA

Abstract

The demand for weather and climate information remains substantial. Climate change poses complex agricultural issues, underscoring the need for accurate weather and climate forecasts. Achieving highly accurate and consistent weather and climate predictions is a challenge in tropical regions. Numerous weather prediction models have been developed worldwide, yet these models often fail to adequately represent the dynamic parameters specific to Indonesia. Various approaches can be taken to predict upcoming rainfall, such as developing SARIMA and LSTM models. SARIMA enhances the ARIMA model by adding seasonal components, while LSTM has emerged as a competitive model for seasonal data. Rainfall data blending from selected Seasonal Zones (ZOM) in Sumatra is divided into training and test sets. LSTM and SARIMA models are constructed using Python programming language. Model outcomes are verified using metrics like RMSE, Bias, MAPE, and Pearson correlation. Additionally, data is transformed into multi-category contingency tables to verify the calculations for the proportion of correct forecasts and Peirce Skill Score. The SARIMA model produces smaller absolute error statistical values, but exhibits larger bias compared to LSTM. LSTM effectively captures training data patterns. The SARIMA model's outcomes are somewhat flat. Both models exhibit similar levels of accuracy. Peirce Skill Score calculations reveal that the LSTM model outperforms the SARIMA model in terms of the number of ZOMs falling into the "good" category.