Based on the long-term series data of water level, flow, and precipitation, a long short-term memory (LSTM) neural network model using the particle swarm optimization (PSO) algorithm (PSO-LSTM model) for hyperparameter optimization was constructed to provide short-term forecasts with forecast periods of 1 to 3 days for the water level during the flood season at Pingwang Station, the center of the plain river network in the Wujiang section of the Jiangnan Canal in Suzhou. The water level prediction results were compared with those from water level prediction models based on the PSO algorithm, including support vector machine (SVM), random forests (RF), convolutional neural networks (CNN), and gated recurrent units (GRU). The effects of water conservancy projects on the prediction accuracy of water level were investigated. The results show that the PSO-LSTM model has high prediction accuracy for short-term forecasts with forecast periods of 1 to 3 days, but the prediction accuracy gradually decreases with the growth of the forecast period. Compared with the PSO-SVM, PSO-RF, PSO-CNN, and PSO-GRU models, the PSO-LSTM model has a lower mean absolute percentage error and better prediction efficiency. The PSO-LSTM model can efficiently predict water level of the plain river network during the flood season, and the addition of artificial regulatory influences such as water conservancy projects can improve the water level prediction efficiency.