Bidirectional Long Short-Term Memory Method for Weather Prediction in Terms of Evacuation of River Drifting Victims

Indonesia, as a strategic archipelago, is vulnerable to natural disasters such as floods and heavy rains. SAR team, Indonesia’s Search and Rescue (SAR) team, is a disaster management agency tasked with evacuating disaster victims, requiring considerable human resources and time to ensure smooth operations. Often, the SAR team must stop the evacuation process abruptly due to extreme weather conditions. To avoid such unwanted actions, a weather prediction model is needed that can help the evacuation process of victims drifting in the river using the Bidirectional Long Short-Term Memory (Bi-LSTM) method. In this research, we present a model that can predict weather conditions in river drifting disaster areas based on a dataset containing weather environment parameters (rainfall, temperature, humidity, air pressure, wind direction, and wind speed). Predictions were made by training hourly data for 3 years (2020-2022) totaling 26304 data points from BMKG Perak Meteorological Station. After partitioning, 21043 data points were used as training data, while 5260 data points were used as test data. The optimal model with the scenario of Bi-LSTM units on layer 1 is 128 with a batch size of 32 and a learning rate of 0.0001, according to the outcomes of the model training test. The model produces an R-squared (R
²
) score of 92% and an RMSE value of 0.19 by using a time span of 3 hours to predict weather conditions up to 8 hours in advance.