大坝变形GA-LSTM组合预测模型研究

A combined model for predicting dam deformation based on genetic algorithm and long short-term memory neural network

为监测大坝运行过程中的异常状态,防范化解大坝溃坝等重大风险,基于大坝变形大样本、非线性监测数据,引入长短期记忆(Long Short Term Memory, LSTM)神经网络模型对大坝变形趋势进行预测,以测试样本的均方根误差最小为适应度函数,采用遗传算法(Genetic Algorithm, GA)优化LSTM模型参数,建立大坝变形GA-LSTM组合预测模型。以福建水口水电站大坝为例进行验证分析,并与LSTM模型和门控循环神经网络(Gated Recurrent Unit, GRU)模型预测结果进行对比分析。分析结果表明,GA-LSTM模型的预测效果和性能更佳,且相较于LSTM模型和GRU模型各测点预测误差均有减小,平均绝对误差减小量最高达6.92%。

To monitor the abnormal state and prevent the major risks during the operation of the dam,a combined prediction,which is based on a Genetic Algorithm(GA)and Long Short-Term Memory(LSTM)neural network,is proposed to predict the dam deformation trend on the large deformation samples and nonlinear monitoring data.Parameters of the LSTM model are optimized by genetic algorithm with the minimum root mean square error of test samples as the fitness function.Among them,the optimization parameters mainly include the structure of the LSTM model and the number of neurons at each layer.Environmental factors affecting dam deformation and dam deformation effect aretaken as the input samples.Dam deformation monitoring data is taken as the output samples.Then the GA-LSTM deep learning combined prediction model of dam deformatioinsestablishedF.urthermore,ttheShuikou hydropower station dam in Fujian province is taken as an example for comparing the prediction accuracy of the GA-LSTM model,the LSTM model,and Gated Recurrent Unit(GRU)model.The research results show that the proposed GA-LSTM combined model has a better prediction effect and performance.And the fitting effect of the GA-LSTM prediction model is also better.More specifically,compared with the LSTM model and GRU model,the prediction errors of the GA-LSTM model at each measuring point are reduced.The root mean square errors of the GA-LSTM model respectively decrease by 3.23%,2.41%,2.14%,4.13%,2.01%and6.13%,4.94%,6.17%,5.7%,4.02%.The maximum reduction of the root mean square error is 6.17%.The mean absolute errors of the GA-LSTM model respectively decrease by 2.07%,1.61%,1.07%,4.65%,1.08%and 6.53%,5.79%,4.92%,6.92%,4.36%.The maximum reduction of mean absolute error is 6.92%.