Long short-term memory integrating moving average method for flood inundation depth forecasting based on observed data in urban area

Since flooding in urban areas is rarely observed using sensors, most researchers use artificial intelligence (AI) models to predict flood hazards based on model simulation data. However, there is still a gap between simulation and real flooding phenomenon due to the limitation of the model. Few studies have reported on the AI model for flood inundation depth forecasting based on observed data. This study presents a novel method integrating long short-term memory (LSTM) with moving average (MA) for flood inundation depth forecasting based on observed data. A flood-prone intersection in Rende District, Tainan, Taiwan, was adopted as the study area. This investigation compared the forecasting performance of the backpropagation neural network (BPNN), recurrent neural network (RNN) and LSTM models. Accumulated rainfall (Ra) and the moving average (MA) method were applied to enhance the LSTM model performance. The model forecast accuracy was evaluated using root mean square error, coefficient of determination (R2) and Nash–Sutcliffe efficiency (NSE). Analytical results indicated that the LSTM had better forecasting ability than the RNN and BPNN, because LSTM had both long-term and short-term memory. Since Ra was an important factor in flooding, adding the Ra to the model input upgraded the LSTM forecasting accuracy for high inundation depths. Because MA reduced the noise of the data, processing the model output using the MA also elevated the forecasting accuracy for high inundation depths. For 3-step-ahead forecasting, the NSE of the model benchmark BPNN was 0.79. Using LSTM, Ra and MA, NSEs gradually increased to 0.83, 0.88 and 0.91, respectively.