Study On Thermal Load Forecasting Of Coal-fired Industrial Boiler

Industrial boilers are widely used in China, the designed thermal balance efficiency of them is always about 80%, but the actual operating efficiency of most boilers is about 60%. It is the running conditions that decide the thermal efficiency in practice. The real operation is effected by many factors including weather, load change, operation management, working environment and overall operating efficiency. The main reason leads to the low efficiency is that the change in range and frequency of load fluctuation are both strong, but the hysteresis of boiler system make response in time impossible. This paper is based on the load forecasting system which is suitable for industrial boilers. Combining dynamic response characteristics of industrial boilers with mechanism analysis of stoker-firing chain grate boiler, this paper offers instructions to the optimization of actual operation.To inllustrate the influence of load waving on operating parameters, selected one 20t/h boiler whose load change frequently as object, we have researched the variation of grate speed , excessive air coefficient, air distribution and thickness of coalbed and atmosphere of burning bed surface under different loads.Support vector machine is selected as load forecasting method for it is global optimal, easily extended and it has a simple structure, and the work is based on the analysis of research on short-term load forecasting. The input sample has a great influence on the prediction results when using SVM for load forecasting. It is necessary to do cluster analysis on the samples to find the proper input. The advantage and disadvantage of SOM neural network and K-MEANS are compared. SOM network can do cluster analysis unsupervised, but the convergence time is long and can only get approximate clustering results when girds do not converge. K-MEANS get different clustering results with different initial clustering centers, but it is high-precision when clustering center and number are known. This paper introduces a cluster method that combines SOM neural network with K-MEANS. SOM is used first to get the clustering center and number, and K-MEANS then used to re-cluster the samples and get detailed cluster information. Use the similar days got from this combination as the input of SVM and thus improve the prediction accuracy of SVM.