Based On The Nonlinear Theory Of Steam Coal Blending Model

This thesis focuses on modeling of power coal blending with application of some non-linear theories. It includes application of artificial neural networks to predicting properties of blended coals and that of a hybrid method of Genetic Algorithm and Simulated-annealing Algorithm to optimizing coal blending process.As a kind of clean coal combustion technology suitable for China, power coal blending has gained much attention and is about to be spread widely in China. Blending coals of different qualities and prices is an appropriate approach to make coal qualities well within the ranges specified for power plant boilers, such as heating value, ignition property, burn-out property, sulfur content and so forth. It is also a powerful means to lower pollutants emission, to reduce fuel costs and to reduce fouling and slag, thus improving performance of the plant while maintaining boiler's capacity. However, changes in coal qualities are generally difficult to predict because of insufficient data and limited understanding of the underlying physical and chemical process in coal blending. At present, weighted averaging method is most commonly used to estimate qualities of coals blended from their component coals and linear-programming technique is usually employed to guide coal-blending projects. Nevertheless, more and more experimental and theoretical researches show that most of the qualities of the blend coal can not always be measured as weighted averages of corresponding indexes of its component coals and the simplified linear-programming technique can not often draw the proper blending scheme.In this thesis, back-propagation (BP) neural network is adopted to determine the relations between qualities of the blended coal and its components. Models are established for predicting blended coal's heating value, volatile content, ash content, sulfur content and ash melting point. Experiments show that the BP models can always get much better prediction results than weighted averaging method and empirical formula. In spite of its simplicity and effectiveness, BP neural network is susceptible to quite a few factors. Suitable network structure, adequate learning samples and appropriate learning times are essential for a good prediction result. Valuable suggestions are proposed in this thesis after these important factors are discussed to a large extent.Power coal blending is a process of optimization. To search for a suitable coal-blending scheme fast and accurately is another important task to be accomplished by power coal blending technology. Enumerative Algorithm (EA) and Mixed Discrete-variables Optimization Design (MDOD) are commonly used before. However, EA is not efficient while MDOD can only achieve part optimization and is easily trapped into partially optimized result. In this thesis,another two algorithms-Genetic Algorithm (GA) and Simulated-annealing Algorithm (SA), which are never used in the area of coal blending, are employed successfully into the model of optimization. With strong robustness, GA is good at searching for the final optimized result. And SA is simple and has a great ability of local searching. On the basis of the two algorithms, this thesis has for the first time proposed a hybrid algorithm-GA+SA+EA. The calculation time with this new algorithm is only one hundredth of that with the single Enumerative Algorithm.