Moving Average for Optimizing Flood Risk Detection and Rainfall Forecasting with PSO-LSTM

Abstract

Monthly forecasting is important in flood risk 
detection because it can help plan for its prevention in the long 
term. However, long-term forecasting in flood risk detection still 
has limited performance. This paper aims to create flood 
forecasting using long short-term memory (LSTM) optimized 
by moving average smoothing, particle swarm optimization 
(PSO), and residual noise reapplication. In this study, the 
dataset used contains information about rainfall in millimeters 
(Rainfall - MM), which is then normalized to improve the 
prediction model's performance. We perform a moving average 
analysis on the performance of the model prediction. Then, PSO 
improves the prediction ability of LSTM by finding the 
optimum hyperparameters. Finally, we perform residual noise 
reapplication to obtain important outliers and indicate whether 
or not a flood occurs. The test results show that PSO gets the 
optimum R-squared at the fourth iteration, with the optimum 
hyperparameters being learning rate, LSTM units, epoch, and 
batch size of 0.01, 200, 93, and 64. Moving average is proven to 
improve the performance of LSTM and PSO+LSTM, where the 
MAPE decreases from 2.74 and 1.30 to 0.32 and 0.31, 
respectively. Finally, residual noise reapplication affects flood 
risk detection, where the detection accuracy increases from 
0.985 to 0.996.