Prediction Model And Applied Research, Power Plant Thermal Parameters Based On Real-time Learning Strategy

The power generation efficiency and energy utilization of thermal power generation are relatively low, one of the main influence factor is that lots of important process parameters (such as flue gas oxygen content, ball mill load) and economic indicators are difficult to be real-time measured online, which seriously restricted the process control and optimal operation in thermal power plant. Because the power production process has the characteristic of nonlinear, strong coupling and large operation range, based on the local modeling algorithm of divide-and-conquer principle, the soft-sensing models of flue gas oxygen content and ball mill load are studied in this paper. The main research work and innovative achievements are listed as follows:(1) The improved support vector machine modeling method based on just-in-time learningBecause the boiler combustion process and coal pulverizing system have the characteristic of nonlinear, strong coupling and large operation range, it is difficult to establish an effective global soft-sensor model using global modeling algorithm (such as neural network, support vector machine) based on data-driven. Therefore, by analyzing local modeling theory, an improved support vector machine modeling method based on just-in-time learning is proposed in paper.By analyzing the similarity of historical data sample, a similar sample selection method based on the distance and angle information is proposed. The interested observations in a local neighborhood of the query point are selected by the selection method, and the selected precision can be improved. Because the local neighborhood size of the query point is small, the support vector machine modeling method is selected to establish the local model, and the three key parameters in support vector machine model are optimized by using an improved particle swarm optimization algorithm, so the prediction accuracy can be effectively improved. In order to balance the retrieval accuracy and the retrieval efficiency of the data sample set, a two-step searching strategy based on weighted fuzzy C-means clustering algorithm is offered in this paper. A kind of updating strategy for data sample set is also presented.(2) Application research of the local prediction model in the flue gas oxygen contentThe initial instrumental variables of the flue gas oxygen content prediction model are determined by analyzing the process mechanism of the boiler combustion process. Then based on lots of historical data and data preprocessing (such as 3σabnormal data recognizing and data normalization), the final instrumental variables are obtained by using gray relational analysis method to optimize the initial instrumental variables and the flue gas oxygen content prediction model is established by utilizing the improved SVM modeling method based on just-in-time learning strategy which is proposed in this paper. Because the proposed modeling method has the essential ability of online self-adaptive, so the obtained model can have better adaptation to different operating conditions. Simulation results show that compared with the standard BP neural network and the standard SVM prediction model, although the computing cost is increased, the proposed prediction model has better prediction performance and can satisfy the real-time requirements for the flue gas oxygen content in the boiler combustion process.(3) Application research of the local prediction model in the ball mill loadThe initial instrumental variables of the ball mill load prediction model are determined by analyzing the process mechanism of the coal pulverizing system. Then based on lots of historical data and data preprocessing (such as 3σand wavelet analysis abnormal data recognizing, and data normalization), the final instrumental variables are obtained by using gray relational analysis method to optimize the initial instrumental variables and the ball mill load prediction model is established by utilizing the improved SVM modeling method based on just-in-time learning strategy which is proposed in this paper. Because the proposed modeling method has good online self-adaptive ability, so the obtained model can have better adaptation to different operating conditions. Simulation results show that compared with the standard BP neural network and the standard SVM prediction model, although the computing cost is increased, the proposed prediction model has better prediction performance and can satisfy the real-time requirements for the ball mill load in the coal pulverizing system.