A Study On Radial Basis Function Neural Networks And Its Application In Direct-firing Medium Speed Mill Pulverizing System

Boiler Pulverizing System is an important part of the whole coal-fired power plant, and its safety dependability and economics plays a direct part in the boiler's safety dependability and economics. With the improvement of unit load and parameters of the power plant, coal-fired power plant uses medium speed mill direct-firing pulverizing system in common. Automatic control of thermal process is an indispensable technical step and measure to guarantee the security and economical running of thermal equipments. However, establish object's exact mathematical model is also a precondition of automatic control. So, it is necessary to research on the dynamic characteristics of medium speed mill thoroughly.Medium-speed mill is a highly relevant, lag and MIMO nonlinear system and its dynamic characteristics change widely with circumstance. Traditional identifications usually have a disadvantage of more calculation, worse anti-disturbing and lower precision. However, radial basis function (RBF) neural networks have an ability to model arbitrary nonlinear mapping, simple network structure. Based on existing learning algorithms for RBF neural networks, an RBF training method combining immune clustering and immune evolutionary programming is proposed in this paper, and play offline identification on medium speed mill pulverizing system with proposed algorithm. In order to satisfy the real time demand of the thermal automatic control system,"winner neuron"strategy based on minimal resource allocating networks algorithm is proposed in this paper. Proposed strategy simplify adjustment process on network parameters and only those parameters that are related to the input are updated by the EKF algorithm, and thus greatly reduce the training time of network. The pruning strategy of MRAN algorithm is improved, not only to delete the hidden neurons continuous contributing little to the network output, but also to combine the similar hidden neurons and thus to implement a more compact network structure.The two proposed RBF neural networks training methods are all used to identifying the MPS medium speed mill in a power plant. Matlab simulation study indicates both the two training methods have higher precision. Offline modeling has the highest precision but long calculation time, and fit to the control system that requires the highest precision. Although online modeling has lower precision, the identification time is far less than sample time. So the online modeling based on MRAN accords with the real time identification and can be used in model based control algorithm directly.