Research On Inner Anomaly Detection And Temperature Prediction Of Thermal Boilers Based On Multivariate Time Series

The thermal power still occupies a very important position in the industry.The inner and temperature anomaly of a thermal boiler,which is one of the most important equipments in a thermal power plant,have a great impact on the efficiency and safety of power generation.Focusing on the anomaly and temperature prediction of thermal boilers,the research content of the thesis is as follows.(1)The thermal boiler is a complex system composed of combustion subsystem,water circulation subsystem,and flue gas circulation subsystem.Its working parameters can be represented as multivariate time series.So,the inner anomaly caused by component damage and system disturbance will appear as abnormal modes in multivariate time series.Therefore,a model for anomaly detection of multivariate time series named GAT-AD is proposed in the thesis.The model can extract multi-scale features of time series based on one-dimensional convolution.Moreover,Graph Attention Network is introduced to spatio-temporal feature extraction of multivariate time series,and Bidirectional Gate Recurrent Unit is used to improve prediction accuracy and model interpretability.Compared with existing methods,GAT-AD averagely improves F1 Score by 9.07% and 9.09% on public datasets SWaT and MSL respectively,and by 34.7% on boiler anomaly datasets.(2)The heating surfaces of a boiler superheater may contain hundreds of tubes.Owing to the different impacts of desuperheating water and flue gas on tubes with various conditions,the temperature of tubes may be very different.Therefore,a model,WGCNGRU,to predict the temperature of the heating surface is proposed.The boiler heating surface is divided into local ones based on clustering and the Davies-Bouldin index,and their potential relationship in temperature changes is represented as local heating surface graphs.The weighted graph convolutional neural network is then introduced to extract the spatial features among local heating surfaces,and Gate Recurrent Unit is used to predict the temperature of each local heating surface finally.Experiments show that WGCN-GRU can maintain the mean absolute error of prediction below 0.5℃.Compared with existing models,WGCN-GRU can averagely reduce the mean absolute error by11.4%,and also shows great advantages in root mean square error and Coefficient of Determination.