Deep Learning Based Boiler Wall Temperature Prediction And Anomaly Detection

With the rapid development of China’s power industry,the boiler operating system of thermal power plants is more complex,and the operating parameters that need attention have increased,while the safety of boilers in thermal power plants has also put forward higher requirements based on the frequent occurrence of boiler tube burst accidents.Tube burst accidents often occur because the temperature of the metal wall of the boiler is too high for a long time or the temperature fluctuations change too much,in response to this problem,the operator also through the optimization of the temperature measurement point arrangement and the number of measurement points to increase or accurate prediction of wall temperature and abnormal detection.However,with the increase of real-time measurement point data,there is still less predictive analysis of temperature parameters,lack of abnormality detection and location when a fault occurs and experience accumulation and sharing of abnormality diagnosis,therefore,the task of prediction and abnormality detection of multivariate time series data analysis have become more important.For the time being,artificial intelligence leads the technological zeitgeist,and deep learning is one of the powerful tools in artificial intelligence,which has a wide range of applications in production and life.Its application in the direction of power plants often tends to the prediction of parameters such as temperature and load,optimization of the system,etc.With the development of big data and the massive growth of data,power plants have accumulated a large amount of historical data,and it is crucial to analyze the historical data to detect and accurately locate abnormalities in multivariate time series data.In the face of the above research status and background,this thesis selects several models of deep learning for boiler wall temperature prediction and anomaly detection,and the main research work includes.1.To address the problem of boiler wall temperature overtemperature and large fluctuations,the boiler wall temperature is predicted using temporal convolutional networks,and the classical time-series models ARIMA,LSTM and GRU are used as comparison models.The specific prediction method is to compare the predicted value of the model output with the actual value of the temperature sequence,and the three error indicators RMSE,MAE and MAPE are used for evaluation to judge the model performance of the models.Finally,the prediction results of each model are visualized for observation and comparison,and conclusions are drawn.2.Based on the first point,this thesis designs a system for boiler wall temperature prediction based on deep learning.The system contains data storage,data processing,model training and prediction results of the model,which realizes the necessary functions of the model and improves the user experience of human-computer interaction.3.For the anomaly detection problem of boiler wall temperature,this thesis uses anomaly detection method based on Euclidean distance to analyze the boiler working condition,but only one temperature measurement point can be detected,and the detection for multiple measurement points needs to be cycled constantly,and the detection efficiency needs to be improved.To address this problem,a model for anomaly detection and localization of multiple measurement points in power plants based on generative adversarial networks is proposed.First,the multivariate time series of the power plant is divided into sliding windows to obtain the feature matrix and feature images,and then based on the property that generative adversarial networks(GANs)can simulate the distribution of complex high-dimensional images,the goal of generative adversarial networks is to learn to map a series of feature images to the next feature image,and use encoders and decoders to reconstruct the transformed feature images,considering the The comparison between the reconstructed image and the feature image can also detect whether anomalies occur in the data of this time period and can locate the anomalous sequence accurately.The above experimental results show that the prediction accuracy of the temporal convolutional network is better than the rest of the comparison models in terms of wall temperature prediction,and the design and implementation of the system for boiler wall temperature prediction also have a good experience,and finally the newly proposed generative adversarial network-based model can perform simultaneous anomaly detection for multiple measurement points with higher detection efficiency than the traditional Euclidean distancebased anomaly detection method.