基于卷积神经网络的东苕溪瓶窑水文站水位预报

Water Level Forecast of Pingyao Hydrological Station in Dongtiaoxi Based on Convolutional Neural Network

东苕溪流域的洪水防御对杭嘉湖平原安危至关重要,而瓶窑水文站水位预报是东苕溪流域洪水调度的决策依据。基于二维卷积神经网络,构建了东苕溪瓶窑水文站水位预报方法,选择起涨水位、下游水位、上游水库下泄流量及区间降水量构建特征集,设计了三层二维卷积神经网络结构,对输入特征集进行有效特征提取,采用两层全连接神经网络建立水位预报模型,并利用粒子群算法优化模型参数,进而得到最优模型。东苕溪2007~2020年14场典型洪水过程水位预报实例分析表明,基于卷积神经网络的瓶窑水位预报模型预报精度能达到甲级,预报精度和可靠性较高。

The flood defense in Dongtiaoxi basin is very important to the safety of Hangjiahu Plain,and the water level forecast of Pingyao Hydrological Station is the basis for decision-making on flood dispatch in Dongtiaoxi basin.Based on the two-dimensional convolutional neural network,the water level forecast method of Dongtiaoxi Pingyao Hydrological Station was constructed.The rising water level,downstream water level,upstream reservoir discharge flow and interval precipitation were selected to construct a feature set,and a three-layer two-dimensional convolution neural network was designed to effectively extract feature of the input set.A two-layer fully connected neural network was used to establish a water level forecast model.The particle swarm algorithm was adopted to optimize the model parameters.The analysis of water level forecasting examples for 14 typical flood processes in Dongtiaoxi from 2007 to 2020 shows that the forecast accuracy of Pingyao water level forecast model based on the convolutional neural network can reach Class A with high forecast accuracy and reliability.