Study And Application On Mass Data From Thermal Process Via Data Mining Approaches

Power plant safe and economic operation is the foundation to ensure the steady development of the national economy.Therefore,the performance monitoring and operation optimization technology for large thermal power units is of great significance.Due to the complex structure,strong nonlinear,changeable running environment of the thermal process,it is difficult to establish accurate applicable mathematical model.Thus,it is necessary to use data mining methods to analyze complex data in thermal processes for unit operation knowledge and rules.In this way,s thermal process monitoring and optimization can be achieved.This paper takes massive historical operation data and real-time data from real-time data acquisition/storage system in the thermal process as the research object.After the preprocessing of thermal process data,attribute reduction and clustering analysis methods are taken to acquire target condition library and monitor the running state of the power plant unit.The main contents of this paper can be described as follows:1?Data preprocessing for thermal process has been carried out.Information entropy theory with basic statistical method is proposed as a means of signal transformation in different transform space.This method can realize the quantitative characterization of the signal in different levels.Multiple substitution nonlinear regression models with time delay samples related to dependent variable is employed to conduct fault detection.Sliding sample entropy is used to deal with the normalized sample data.With steady-state threshold value,steady state factor can be obtained for stability determination to thermal process data.ECNN algorithm is used to compress the thermal process data by reducing threshold weight for low steady data.This method can keep samples with higher degree of stability in the compressed data set.2?In view of the defects of the traditional discrete algorithm for the discrete number should be set in advance,a kind of entropy based clustering(E_Cluster)method has been proposed for continuous attributes discretization without presetting the initial clustering parameters.On the basis of the complementary condition entropy of rough set,the improved rough set complementary condition algorithm(D_RED)is proposed for incremental data.E_Cluster and D RED algorithms are used for boiler controllable operation parameters discretization and attribute reduction.Moreover,operation parameters importance degree to boiler combustion efficiency and concentration of NOx is analyzed under different unit loads.3?Based on the numerical data clustering analysis,a new clustering evaluation index(Vnew)is proposed and utilized to participate in the improvement of Kmeans algorithm by comparing each partition clustering validity index values to determine the optimal clustering number.The improved Kmeans algorithm is then utilized to analyze the relationship between furnace pressure signal characteristic value and the unit load.4?To deal with the mixed type data existing in the thermal process,an improved K-prototypes method has been proposed.This method is combined with tabu particle swarm optimization(TS_PSO)algorithm to optimize the clustering objective dissimilarity function,D(x,y).Furthermore,with self-adapted K-prototypes based on TS_PSO clustering algorithm as the main technical method,the target operating condition library for best efficiency of boiler and furnace outlet concentration of NOx can be achieved under different loads.5?EKFCM algorithm is proposed to evolve the shortcomings of fuzzy C-means(FCM)by no necessity to set up the initial value of clustering number.The EKFCM algorithm,combining with the information entropy theory and Kmeans,FCM clustering algorithm,optimal clustering number is considered as the classification number when the difference of the entropy value get the minimum.The optimal clustering number is then taken as the initial parameters of the FCM method.When dealing with the data of continuous operation system,a new sample online incremental learning strategy is proposed.First Out by Time Decay for Subset(Sub-TDFO)strategy,which is based on the First In First Out(FIFO),takes time decay into consideration.The proposed EKFCM incremental clustering algorithm is then applied to the real-time analyzing of the thermal process data to monitor blocking in air preheater operation and steam turbine flow passage fouling.6?A complete development program of the KDTPS is designed which includes system theory,architecture design and system function.The integration KDTPS into SIS is also introduced.