Structural Optimization Design And Application Of RBF Neural Network With Hybrid Connections

Artificial neural network(ANN)is a mathematical model constructed by early neuroscientists to imitate the human brain nervous system.It carries out parallel information processing by simulating the way the brain nervous system processes information.The model simulates the biological neural network by abstracting the neural network of the human brain and establishing connections between different neurons according to a certain topology.Neural networks have the ability to store and use empirical knowledge,which is acquired and stored through learning and the strength of connections between internal neurons to simulate the brain.Radial basis function(RBF)neural network is a powerful type of localized neural network with its fast-training advantage,which is often used to predict highly complex problems.However,in practical application,RBF neural network design has some limitations.For example,in network model training,the stability of the model is difficult to ensure.On the other hand,the current methods mostly determine the structure through experience and training error,ignoring the physical influence of output on input in sample data.In the actual process,the RBF model structure only adapts to the change of the task by changing the weight parameters.How to construct an RBF neural network to make the structure dynamically adjust the weight parameters at the same time,and closely combine with the characteristics of process data is a problem that should be solved for RBF neural network.In addition,in order to capture the statistical features and long-distance spatial correlation between time series,recurrent neural network(RNN)with time feature memory is considered to be an excellent tool for data modeling.However,the determination of the recursive time and the amount of recursive information in the model is still an open and unsolved problem.Therefore,this thesis proposes a structure optimization design method of RBF neural network with hybrid connections to improve the stability and generalization performance of the network and verify the accurate prediction of effluent quality parameters in the process of sewage treatment industry.Different from the traditional artificial neural network design,this thesis fully and deeply analyzes and studies the operation process of sewage treatment industrial process,and improves the accuracy of the model according to the dynamic characteristics of the process rather than relying only on data;Breaking the traditional idea of training before testing,the parameter training process based on training error and the testing process based on network complexity are integrated into a model framework,and the global linear superposition loss function is established to evolve iterative learning,and the design method with high generalization ability is obtained;On this basis,fully considering the space-time characteristics of network recursive information,a double recursive RBF neural network structure consistent with industrial process is designed,and a complete RBF neural network optimization design algorithm is obtained.The main research work and innovations of this thesis are as follows:1.Parameter optimization method of RBF neural network based on improved Levenberg-Marquardt(LM)Many existing modeling methods aim at accuracy,but ignore the stability of the model.A parameter optimization method of RBF neural network based on improved Levenberg-Marquardt(LM)was proposed to ensure the stability of the network.Firstly,a typical sample mechanism with variance reduction is proposed,which can reduce the error of gradient estimation and use accurate gradient information to guide learning.Secondly,an improved LM optimization algorithm was proposed to update the model parameters,which ensured the stability of the network while improving the convergence speed of the network,and theoretical analysis was carried out.Finally,a multi-step updating rule based on typical sample and single sample is designed,which effectively reduces the sample bias introduced by single sample.To prove the validity of the proposed technique,the model is applied to two benchmark simulation platforms and the prediction of biochemical oxygen demand(BOD)of effluent quality in urban wastewater treatment process.The results indicate that this method has good performance in learning speed and stability.2.Research on self-organizing RBF neural network based on network sensitivityAiming at the problem that it is difficult to guarantee the generalization ability and prediction accuracy of nonlinear process model at the same time,a self-organizing RBF neural network method based on output sensitivity is proposed.Firstly,based on the local characteristics of RBF neural network,the sample mean and variance are introduced to establish a local generalized error bound to obtain a smaller error interval and reduce the error range.Furthermore,a self-organizing structure adjustment method based on error bound and network output sensitivity is established to obtain the appropriate number of neurons and simplify the network structure.In this way,while ensuring the generalization ability and model accuracy,the model complexity is effectively reduced and the structural redundancy is alleviated.In addition,based on Lyapunov stability theory,the convergence of the proposed method is proved under the conditions of fixed structure and dynamic adjustment of structure.Finally,the proposed method is applied to two numerical simulations and a practical wastewater treatment process to testing the validity of the method.3.Design algorithm of bilevel self-organizing RBF neural network based on hybrid superpositionThe RBF neural network learning algorithm is difficult to balance over fitting and under fitting when solving practical problems,resulting in poor generalization ability.This thesis studies a hybrid superposition evolutionary learning method of bilevel self-organizing RBF neural network.Firstly,the training process and testing process are integrated into a unified framework to effectively balance the over fitting and under fitting problems.Secondly,an interactive evolutionary learning algorithm with two layers is proposed.The upper layer adjusts the network structure based on the network complexity and test error,and the lower layer uses the LM algorithm as the optimizer to optimize the connection weight of self-organizing RBF neural network.Then,the final output of the model is generated by using the comprehensive information from multiple sub networks,and the global loss function is set to accelerate the global convergence of the network.Finally,the test experiments are carried out in the benchmark problem of Lorenz chaotic time series prediction and the practical application of total phosphorus(TP)prediction of effluent quality in the process of sewage treatment.The results show that under the test accuracy similar to or even better than the traditional neural network structure,this method can not only achieve faster training convergence,but also produce a more compact RBF neural network model,effectively solve the imbalance between over fitting and under fitting,and has stronger generalization ability.4.Structure design of dual recursive RBF neural networkThe determination of recursive time and recursive information in recursive RBF neural network has always been a difficult problem in this field.To solve this problem,an ammonia nitrogen prediction model based on dual recursive RBF neural network is proposed in this thesis.Firstly,feedforward and feedback neural networks based on plug flow and reflux flow characteristics are designed.Secondly,based on the Muskingum method,a recursive time calculation strategy is proposed to estimate the finite time domain of dynamic feedback in the network.The strategy can adaptively adjust the recursive time and recursive information of dual recursive RBF neural network according to the real-time internal and external reflux flow in the wastewater treatment process.Finally,the attention mechanism is introduced to update the dual recursive RBF neural network structure according to the changes of ammonia nitrogen related characteristic variables.This mechanism makes the dual recursive RBF neural network avoid structural redundancy in the process of ammonia nitrogen prediction.The proposed dual recursive RBF neural network model is applied to predict the ammonia nitrogen concentration of water quality parameters in the process of wastewater treatment,and the excellent performance of this method is verified.