| Artificial intelligence consists of methods inspired by the mechanisms of biologicalevolution and some natural phenomena, which could solve complex systems’ modelingand optimization. Therefore, it could attract more and more attentions of specialist andscholar. For the extremely complex physical and chemical changes of boilers’ combustionprocess in power plant, it is very difficult to set up models of combustion process and toachieve the combustion optimization objective. Therefore, a great deal of research wouldbe done on heuristic optimization methods and neural networks that would be applied tothe combustion process in order to make a boiler combust at high efficiency and lowpollution.Firstly, for the shortage of Artificial Bee Colony (ABC), this paper proposes twoimproved strategies: To improve the convergence precision and stability and reduce theconvergence iterations of ABC, this paper propose two improved methods, an ImprovedArtificial Bee Colony(I-ABC) in which the best-so-far solution, inertia weight andacceleration coefficients are introduced, and an Artificial Bee Colony with the abilities ofPrediction and Selection(PS-ABC) which inherits the bright sides of ABC, I-ABC andGbest-guided ABC(GABC); To reduce the runtime and the convergence time andsimultaneously guarantee or improve the optimization accuracy of PS-ABC, this paperproposes another improved method called PS-ABCⅡwhich has three major differencesfrom PS-ABC:1) population initialization;2) the technique that chooses candidatesolutions;3) the mode that employed bees become scouts. In order to test the validity ofimproved methods, they are applied to23benchmark optimization problems. Experimentsshow they have very good global search ability and convergence speed.Secondly, this paper presents a novel artificial neural network with a fast learningability called Fast Learning Network (FLN). The FLN is a Double Parallel ForwardNeural Network, whose output nodes not only receive the recodification of the externalinformation through the hidden nodes, but also receive the external information itselfdirectly through the input nodes. This paper adopts9classical regression data sets to validate its learning and generalization abilities. In addition, for the shortage of FLN, thispaper also presents two improved learning methods: Optimized Fast Learning Network;Least Square Fast Learning Network (LSFLN), which are employed to set up models ofthe combustion processes of two pulverized coal fired boilers (a300MW and a330MW).Experimental results show that they have very good training precision, stability andgeneralization performances.In order to make FLN have online learning ability, this paper presents two onlinelearning algorithms: Online Fast Learning Network (OFLN) and Online Least Square FastLearning Network (OLSFLN), and adopts them to set up the online model of thecombustion thermal efficiency of the300MW coal-fired boiler. Simulation results showthat OLSFLN could accurately predict the thermal efficiency, and simultaneously stateOLSFLN has very good online learning, generalization abilities, stability and repeatability.Thirdly, this paper presents an offline combination modeling method, which iscomposed of a global model and a compensation model, to model the thermal efficiencyof the300MW pulverized coal boiler. Experimental results show that this combinationmodel could predict the thermal efficiency at a high-precision level, and have very goodstability and generalization performances.Finally, based on the offline and online models of two coal-fired boilers, this paperdetermines their optimization objective functions according to the respectivecharacteristics of the300MW and the330MW coal-fired boilers. Then improved artificialbee colony algorithms are employed to optimize the adjustable operational parameters ofthe two boilers, and search the optimum boiler operational parameter combination.Provided the pulverized coal boilers could combust at the optimum parameters, they couldachieve aims of the combustion optimization. |
PDF Full Text Request