Research On Dam Deformation Prediction Method Based On STL-CS-LSTM Combination Model

By 2018,China has built more than 98000 dams,which is the country had the largest number of dams in the world.Dam can not only regulate the spatial and temporal distribution of water resources,but also is one of the most important engineering measures for rational allocation of water resources.Dam has flood control,drought resistance,power generation,shipping,irrigation,breeding,tourism and other functions.Therefore,the dam plays a vital role in China’s national economic construction.However,dam-break often causes huge loss of people’s lives and property around the country.The dam damage is generally from quantitative change to qualitative change,this process can not be found by human intuition,so it is necessary to carry out perfect and reliable deformation monitoring.Therefore,effective monitoring of dam deformation and accurate prediction of deformation monitoring data are very necessary to ensure the safe operation of the dam.Dam deformation is affected by many factors,such as water pressure,temperature,geological conditions and so on.Limited by the management cost and level,it is difficult to obtain the above monitoring data for small reservoir dams.Therefore,for small reservoirs,it is a cost-effective and effective method to use only historical dam deformation monitoring data for prediction.In this paper,the horizontal displacement data of 5 points obtained from Lishan reservoir dam,a small reservoir,are used for the experiment,and the data are measured from December 16,2018 to January 12,2019.Five groups of horizontal displacement data with 672 pieces of data were obtained by eliminating the outliers of five groups of data and interpolating the missing values.For horizontal displacement data,the first 500 data are selected as the training set,and the last 172 data are selected as the test set.Aiming at the problem that the dam deformation monitoring data have the characteristics of nonlinearity,trend,periodicity and randomness at the same time,a combined model STL-CS-LSTM is constructed by STL、CS and LSTM in this paper.Firstly,the STL method is used to decompose the time series into trend component,periodic component and residual component,and the optimized LSTM model based on cuckoo search algorithm is used to predict the trend component and residual component;The data used in this paper is hourly,and the period is set to 24,that is,the sequence value at time t is equal to the sequence value at time(t+24).Finally,the predicted values of the three components are added to get the final predicted result.After preprocessing the data,the prediction results of the combined model STL-CS-LSTM are compared with the predictions of LSTM,FNN,SVR,XGBoost and GRU models,respectively The accuracy of the 6 models can be described as follows:STL-CS-LSTM model,STL-CS-LSTM model.LSTM model,FNN model,SVR model,GRU model and XGBoost model.The RMSE of STL-CS-LSTM model with the best prediction performance is less than 0.028mm,MAE is less than 0.025mm,and R~2 is more than 0.9.The prediction accuracy can meet the actual needs of the project.