Optimization of LSTM Parameters for Flash Flood Forecasting Using Genetic Algorithm

Accurate flood forecasts provide a critical time for authorities and the public to enact flood response measures and initiate evacuations. Long short-term memory (LSTM) is widely used in flood forecasting to ensure sufficient response time. The lag length of rainfall (L_R), the number of hidden layers (N_L), and the number of neurons (N_N) are key parameters of the LSTM model. However, flood forecasting research seldom explores their optimization via the Genetic Algorithm (GA). This study introduces a novel LSTM-GA model, which integrates LSTM with GA to optimize the L_R, N_L, and N_N for flash flood forecasting. The case study pertains to the water level forecasting of the Wu River in Taiwan. To assess the improvement brought by the proposed model, a standard LSTM model was utilized as a benchmark. This model accurately forecasted floods in the next 1 to 6 h, achieving a Nash–Sutcliffe efficiency coefficient (NSE) score ranging from 0.896 to 0.906. It also exhibited strong flood peak forecast performance. The integration of GA enhanced the LSTM’s forecasting accuracy, with NSE scores rising to between 0.917 and 0.931. Notably, a shorter forecast lead time augmented the degree of improvement. In the LSTM model, L_R was set as the river’s concentration time, and N_L represented the water storage function of the watershed. For short lead time forecasting, surface runoff was the dominant factor, leading to smaller optimized values for L_R and N_L. Conversely, long lead time forecasting needed to consider the impact of subsurface and groundwater runoff, resulting in larger optimized values for L_R and N_L. In conclusion, the parameters optimized through GA consider the watershed’s characteristics.