Research On Time-series Prediction Methods For Monitoring Data Of Water Conservancy Facilities

As a nation with a significant focus on water resources,China possesses an abundance of water conservancy and hydroelectric facilities,along with a vast accumulation of monitoring data.Predicting the monitoring data of these water resources facilities is crucial for understanding their operational patterns and preventing accidents from occurring.The deformation characteristics of the data from a large-scale domestic hydropower station are analyzed in this thesis as an example.A mutation point detection method using CEEMDANBoxplot was proposed to remove sudden changes in the data and complete data preprocessing tasks such as resampling,missing value imputation,and filtering,based on the unique features of the data.With machine learning theory as its foundation,LSTM was selected as the prediction model and an improved Whale Optimization Algorithm was designed for hyperparameter optimization in this thesis.The STL-IWOA-LSTM model,based on the properties of the deformation data,was proposed,which has achieved excellent results in predicting the deformation time series of water conservancy facilities.1)Resampling and imputation of missing values are performed to obtain complete daily average data.In accordance with the characteristics obtained from the point mutation decomposition,the CEEMDAN-Boxplot method was proposed for point mutation detection.By employing CEEMDAN,the data was decomposed into multiple IMF components and a residual term,with the IMF components being classified into high-frequency and lowfrequency components based on the criterion of discrete mean squared error.The Boxplot technique was employed to eliminate outliers in the high-frequency components,and the remaining data was linearly combined with the low-frequency components and residual term to obtain the data after removing the point mutations.The S-G filter was used to filter out noise and residual,slight point mutations to obtain a smooth data curve,which can be inputted into the model for prediction.2)In order to address the difficulties in setting hyperparameters for machine learning methods and the low convergence accuracy and slow convergence speed of existing Whale Optimization Algorithms(WOA),WOA was improved in three aspects: initialization of the population,piecewise non-linear convergence factors,and tail perturbation strategy.The usual random number generation function leads to uneven distribution,causing some individuals to be concentrated during the initialization of the population.The use of piecewise linear chaotic mapping allows for a more uniform distribution of individuals,avoiding resource waste due to concentration.The WOA uses a linearly decreasing convergence factor to control global and local search.However,more global search is needed in the early stages of iteration.A piecewise non-linear convergence factor was designed to address this issue,increasing the proportion of global search in the early stages by approximately 18.4%.To solve the problem of WOA being prone to local optima,a tail perturbation strategy was proposed.When the current iteration’s best solution is the same as the previous iteration’s,individuals with fitness values in the last half are randomly redistributed in the solution space,greatly increasing the probability of escaping local optima.The improved WOA was compared with PSO,SSA,and WOA algorithms for 50 tests on 11 test functions.The average value and standard deviation of convergence speed and accuracy are the best for the improved WOA.3)In order to address the weakening of the trend component in the data by the LSTM model and the poor prediction performance due to the single LSTM model’s parameter settings not being suitable for both the trend and cycle components,the STL-IWOA-LSTM combined model was proposed.This model utilizes the Seasonal and Trend decomposition using Loess(STL)method to decompose the data into trend,cyclical,and residual components,each of which was used as input.The Improved Whale Optimization Algorithm(IWOA)was employed to optimize the hyperparameters of the LSTM model,which was then used to predict each of the trend,cyclical,and residual components separately.The predictions from the three models are then linearly combined to obtain the final prediction.The STL-IWOALSTM combined model was built for the prediction experiment using the preprocessed data,and it achieved favorable results in the prediction of time series deformation data in hydraulic facilities.In comparison experiments with the IWOA-LSTM model and LSTM model,the experimental results demonstrate that the STL-IWOA-LSTM combination model achieved an average reduction of 8.26% and 13.86% in the MAPE indicator,and an average improvement of 6.34% and 13.59% in the NSE indicator.These results indicate that the proposed model in this study exhibits lower prediction errors and higher fitting accuracy.