Research On Reservoir Health Data Augmentation And Diagnosis Based On Generative Adversarial Networks

The health of reservoirs,as hydraulic engineering structures for flood storage and water flow regulation,is crucial to promoting sustainable development.Reservoir health is not only about good natural ecological conditions,but also about sustainable social service functions.With the development of automation technology,reservoir health monitoring data is more accurate and comprehensive,and more up-to-date.Therefore,it is of greater engineering significance to make full use of the large amount of operational data obtained during reservoir operation to achieve reservoir health diagnosis and prediction,in order to maintain healthy reservoir operation and make early warning and countermeasures against risks in advance.As reservoirs are in a healthy state most of the time during actual operation,samples of lesion type are difficult to collect,which seriously affects the accuracy of reservoir health diagnosis and prediction.Therefore,effective data augmentation and balanced distribution of the acquired reservoir operation data are indispensable for reservoir health diagnosis and prediction.This paper investigates the reservoir health data augmentation method based on Wasserstein Auxiliary Classifier Generative Adversarial Network(WACGAN)model improved by Variational Auto-Encoder(VAE)model,and the reservoir health diagnosis and prediction method based on the fusion model of Convolutional Neural Network(CNN)and Long Short Term Memory(LSTM)network,and relies on the reservoir health big data intelligent management platform developed by the group.The proposed method was validated by using the multi-source information from the reservoirs of Lake Leize and Dingdong in the platform developed by the group.The main research findings are as follows:(1)Collate,analyse and process reservoir health multi-source dataA total of 282 sets of experimental data were obtained based on the operational data of Lake Leize Reservoir and Dingdong Reservoir from 2018 to 2022 in the reservoir health data management platform,and four reservoir health classes,namely,“healthy”,“sub-healthy”,“diseased” and “critical”,were identified as prediction categories.Thirty-nine different health indicators,including engineering geology and hydrogeology,dam settlement,biodiversity,water use and reservoir sediment changes,were used as feature vectors,and the selected experimental data were pre-processed.(2)Building a VAE improved WACGAN based reservoir health data augmentation modelBased on the feature extraction capability of VAE and the data generation capability of WACGAN,a VAE-WACGAN reservoir health data augmentation model was built to address the problem of uneven distribution of reservoir health data,and the parameters of the model were optimised in several experiments.The activation function is Leaky Re LU except for the last layer,and the Tanh function is used for the last layer.The Dropout retention rate is 0.5.(3)Building a reservoir health diagnosis prediction model based on the fusion of CNN and LSTM networksBased on the internal implicit feature extraction capability of CNN and the temporal feature extraction capability of LSTM network,a CNN-LSTM reservoir health diagnosis and prediction model was built and the parameters of the model were optimised through several experiments,in which the number of layers of CNN was set to 3 and the number of layers of LSTM network was set to 2.The final number of layers of CNN was set to 3 and the number of layers of LSTM network was set to 2.(4)Engineering example analysis relying on the operational data of Lake Leize Reservoir and Dingdong Reservoir in the intelligent management platformThe GAN,DCGAN,ACGAN and WACGAN models were constructed and compared with the generation effect of the VAE-WACGAN model.Using the evaluation method of correlation calculation and comparing the accuracy obtained by the classification algorithms,the experimental results show that the VAE-WACGAN model has the best generation effect and its training process is stable.LSTM and CNN models were constructed to compare the diagnostic prediction results with the CNN-LSTM model.Using the evaluation metrics of ACC,MCC,Precision,Recall and F1 scores,the experimental results show that the lower the degree of imbalance of the data is,the higher the diagnostic prediction accuracy of the model will be,and the diagnostic prediction accuracy and generalization ability of the CNN-LSTM model are higher than the other models.