Research On Operation Diagnostic And Optimization System Of Coal-fired Units Based On Neural Network

Nowadays coal-fires units take more and more peak-shaving mission.The focus of energy conservation switches from designed load mode to whole range load mode,which brings new requirement and challenge to the optimization method.Current research approach focuses on the coupling between different devices,the thermal performance in wide range of load mode and the mutual influence of power unit and environment.With such specialization,it is of great meaning to utilize machine learning method into the modeling of coal-fired power unit.In the meanwhile,the parameters of model can be analyzed to achieve the operation optimization.A stable data determining method based on second order smoothing window is proposed.After analysis on the history data,a conclusion is made that there are lots of redundant and high-correlated data in the original data set.Then,K-means is used to perform clustering on the data to get scientific working load segmentation.Redundant date parameters are combined and Pearson correlation index are calculated.The high-correlated is removed and combined to get an initial dimension-reduced data set.The rough set theory is introduced with several dimension reduction methods based on rough set.In the process of building decision table,the economic index of power unit,standard coal rate,g/kwh,is chosen as the decision variable.By building decision table of history data,a further dimension-reduced data is stored.This accelerates the process of machine learning and avoids dimension disaster.The concept of hidden operating state is proposed in order to solve problems in the benchmark determining process.By viewing the operation of unit as a Markov process,a hidden state transform matrix and operating transform matrix is built.History operating data is analyzed to train these two matrices.After analyzing the hidden state transform probability,the trend of unit operating can be predicted of whether the performance will improve or deteriorate,and this trend can be used to determine whether to store this operating status into benchmark database.The operating mechanism of unit is analyzed and,combined with neural network method,a numerical model of unit is constructed.Directed by the mechanism of power unit,the connection method(partial connection and full connection)is made in order to build a feature extraction layer and feature comparison layer.BP algorithm is used here to train the network within an acceptable time.After training the network,it is used in online energy conservation diagnostic.Devices with large deviation from benchmark will be recognized and the contribution of each parameter to the decision variable is calculated.By decomposing the error of coal rate down to error in each parameter,guide of operation adjust can be given from the perspective of data method.