关键词: SWMM模型, 长短期记忆神经网络, 模式识别法, 遗传算法, Mann-Kendall趋势检验

Abstract: With the rapid development of urbanization, most cities in China are faced with increasingly severe flood problems. In this paper, a storm water management model (SWMM) is constructed on a community compound scale, and its parameters are determined using the genetic algorithm. Then, two types of most representative rainstorms are selected using a fuzzy identification method based on an analysis of precipitation over the study area. Finally, a long short-term memory (LSTM) model is developed for this area based on the dataset of its different rainfall data and SWMM simulations, a relationship between its rainfall and runoff is obtained, and the LSTM model is evaluated against different conditions. The results show that LSTM simulations agree roughly with those of SWMM, but its peak flood fitting is slightly deviated. Compared with SWMM simulations, LSTM simulated peak floods are smaller for small rainfall and greater for greater rainfall, and LSTM performs the best for medium rainfall. The peak flow fitting is better with 50 hidden layer elements, but more elements will produce an even better effect on the peak flow.

Key words: SWMM model, long short-term memory, pattern recognition method, genetic algorithm, Mann-Kendall trend test