基于长短时记忆神经网络的水库洪水预报

Reservoir Flood Forecasting Based on Long-Short-Term Memory Neural Network

精准的洪水预报是做好防洪排涝工作的重要手段之一,而长短时记忆神经网络(long-short-term memory neural network,LSTM)具有很强的时间序列关系拟合能力,非常适用于模拟及预报流域产汇流这一复杂的时间序列过程。为探究LSTM在水库洪水预报领域的适用性,在白盆珠流域针对不同预见期建立LSTM模型,并与新安江模型进行对比。LSTM模型采用流域降雨及水位数据作为输入,不同预见期的水库水位过程作为输出,率定期为5年,验证期为1年。结果表明:LSTM在预见期为1~6 h时都具有较高的预报精度,在预见期为1 h时预报精度最高,达到0.991,随着预见期增长,模型精度逐渐降低,但其预报精度均高于新安江模型。预见期以及隐藏层神经元数量作为神经网络复杂程度的代表既会影响模型预报精度,也会影响模型训练速度。结果证明了基于长短时记忆神经网络模型具有较高的预报精度,对水库洪水预报具有指导意义。

Accurate flood forecasting is one of the main means to well perform flood control and drainage,and the long-short-term memory neural network(LSTM)has a strong ability to fit time series relationships,which thus is very suitable for simulating and forecasting the complex time series process of basin runoff generation and confluence.To explore the applicability of LSTM in the field of reservoir flood forecasting,this paper established an LSTM model according to different forecast periods in the Baipenzhu Basin and compared it with Xinanjiang model.The LSTM model uses the rainfall and water level data in the basin as input and adopts the water levels of the reservoir at different forecast periods as output.The calibration period is five years,and the verification period is one year.The results show that LSTM has high forecast accuracy when the forecast period is 1~6 h,and the forecast accuracy is the highest when the forecast period is 1h,reaching 0.991.As the forecast period increases,the accuracy of the LSTM model gradually decreases,but its forecast accuracy is higher than that of Xinanjiang model.In addition,reflecting the complexity of the neural network,the prediction period and the number of neurons in the hidden layer will affect not only the forecast accuracy but also the training speed of the model.It is proven that the LSTM model has high forecast accuracy and is of guiding significance to reservoir flood forecasting.