Rainfall Forecasting Utilizing Statistical Downscaling Of Global Forecast System (GFS) Data Using Support Vector Regression (SVR) And Long Short-Term Memory (LSTM) As Top Management Decision Support

The Brantas River Basin (Brantas RB) often experiences flooding events due to high rainfall, so accurate weather predictions are needed to support management decisions in flood mitigation. Global Forecast System (GFS) data is used in predicting weather, especially rainfall prediction, but its low resolution is not sufficient for the local scale in the Brantas RB. The problem faced is how to improve the accuracy of rainfall prediction by integrating low-resolution GFS data with local rainfall data. Based on this, this research aims to improve the accuracy of rainfall prediction in the Brantas RB by utilizing statistical downscaling techniques using the Support Vector Regression (SVR) method which is known to be good at handling non-linear data and Long Short-Term Memory (LSTM), a type of artificial neural network that can capture temporal relationships in time series data. The method developed from this research is expected to provide more accurate rainfall predictions to support better decision making in flood mitigation. The data required are rainfall predictions from GFS and local rainfall data in the Brantas RB. And the result of this research, it is found that for the Mean Squared Error (MSE) and Root Mean Squared Error (RMSE) values of GFS modeling prediction rainfall data obtained are 21.662 and 4.274, respectively, the MSE and RMSE values of GFS are higher than the MSE and RMSE results of predictions based on Machine Learning SVR and LSTM modeling where for SVR the MSE value is 0.774 and RMSE is 0.817 while LSTM MSE value is 0.582 and RMSE is 0.743. It is found that the results of SVR and LSTM learning are more accurate than the GFS rainfall prediction data. The results of the accuracy evaluation between SVR and LSTM learning in relation to rainfall prediction in the Brantas River Basin have almost the same degree, where LSTM is slightly superior to SVR.