RBF Network Optimization Based On Hierarchical Genetic Algorithm And Its Application In Building Soft Sensor For Nitrogen Removal Process In Waste Water Treatment

The online monitoring device for waste water treatment processing can realize the automatic warning of excessive amount of key physical or chemical indexes in waste water, the automatic billing of sewage charges, the automatic generation of data processing workflows and thus composes the basic link of automatic control system for waste water treatment. If some indexes exceed the standards, the environment protection measures can be implemented promptly to stop the expansion of pollution in an early stage, increasing the safety factor of water pollution prevention. It is important for environmental protection and tackling water pollution.Waster water treatment system is a complex nonlinear system containing a huge amount of information, which feartures a large number of variables, uncertainty and strong coupling. Thus, it is hard to build a principle model. Since 1990s, more work has been done in building soft sensor for waste water treatment using Neural Networks. A new soft sensor model using RBF network is studied in this thesis. GA (Genetic Algorithm) as a computational intelligent algorithm with self-organization and self-adaptation ability can solve some complicated nonlinear problem. RBF network can learn fast and avoid local minimum problem caused by BP network, and can better realize online control. The hierarchical GA can be used to optimize the topologic structure and parameters of RBF network simultaneously in wide operating range. It is a new approach to waste water treatment modeling.The soft sensor model is built for TN (Total Nitrogen) estimation in the outlet of a biological Nitrogen removal system in waster water treatment. Since there is a large number of secondary variables in the inlet and reaction tank,PCA (Principal Component Analysis) is used to perform both dimension-reduction and de-correlation in input space in order to simplify the inputs of RBF Network. 13-dimension inputs are reducted to 6-dimension as the statistic contribution rate is set to be no less than 85%. Then topologic structure and parameters of RBF network are optimized in two phases. The hidden layer is optimized by nonlinear method and output layer is optimized by linear method. The structure and parameters of the hidden layer are encoded into one chromosome. The hierarchical GA is further used to optimize the number of hidden layer nodes, width and center of the basis function. LSM(Least Square Method) is then used to define the weights of output layer. The combination of hierarchical GA and RBF network can lead to a non-linear model with self-adaptation, self-learning and error-tolerance ability. The comparison of simulation results of this soft sensor model with those of two other soft sensor models shows that this model may provide higher accuracy and good generalization performance. The proposed approach to the development of soft sensor can realize the online testing of water quality to avoid the testing lag caused by sensor fault or offline analysis in lab, and provides the basis for closed-loop control.