基于神经网络的小型水库洪水预报研究

Flood Forecasting for Small Reservoirs Based on Neural Networks

针对经验预测方法精度不高,传统水文模型应用至小型水库进行洪水预报工作量大、推广较难的问题,引入具有强大特征学习能力的人工神经网络(ANN)方法,结合遗传算法(GA)寻参,对小型水库进行洪水预报。利用GA实现ANN中时间步长和隐含层神经元节点参数自动寻优,可避免寻参盲目性,针对性地为各小型水库构建个性化洪水预报模型。通过构建反向传播(BP)、长短期记忆(LSTM)、门控循环单元(GRU)神经网络洪水预报模型,对实测洪水过程进行模拟对比试验。结果表明:LSTM模型预报精度高、稳定性良好,能学习并模拟实际洪水过程水位变化规律,预报性能优于BP和GRU模型。

In flood forecasting,empirical prediction methods report low accuracy,and traditional hydrological models face the problems of large workloads and difficult promotion when they are applied to small reservoirs.Hence,an artificial neural network(ANN)method is introduced,which is equipped with powerful feature-learning capability.It is combined with the genetic algorithm(GA)to find the optimal parameters for flood forecasting of small reservoirs as GA can realize automatic optimization of the time step and hidden-layer neuron nodes in ANN.In this way,parameter search can be targeted,and personalized flood forecasting models can be constructed for each small reservoir.In addition,the flood forecasting models based on the back propagation(BP),long short-term memory(LSTM),and gated recurrent unit(GRU)neural networks are built,and comparisons between simulations and measured data are conducted for the flood process.The results show that the LSTM model has high prediction accuracy and good stability and can learn and simulate the water-level change pattern of the actual flood process,demonstrating better prediction performance than BP and GRU models.