| With the continuous development of global economy,the consumption of energy grows accordingly.Meanwhile,environmental problems,especially air pollution are getting more and more serious.Since the dust is one of the main components of atmospheric pollutants,it is meaningful for controlling air pollution to reduce dust emissions of human activities effectively.As an efficient,low-cost and maintenance friendly method,the electrostatic precipitation technology has been widely developed and applied among various dust removal technologies.Nowadays,the environmental standards for dust removal efficiency has been more and more stringent,which brings both new chances and challenges to the electrostatic precipitation technology.Particularly,outstanding models and model-training methods of the dust removal efficiency are very significant,including the fault prediction and optimal predictive controlling of electrostatic precipitators.In this paper,our research focuses on the relation of electrical control variables such as working voltage and current in electrostatic precipitation systems,and the dust concentration in outputs of systems.At the same time,the generation load of the power plant was included in the modeling procedure as an uncontrollable input variable.Recurrent neural network(RNN)and its variant,Long Short-Term Memory network(LSTM)were applied to the intelligent prediction modeling of the electrostatic precipitation process.Moreover,novel algorithms of training models were proposed,based on the combination of cybernetics and model-training process.Research contents in this paper can be summarized as follows:(1)Equivalent control process of neural network optimization was proposed from the perspective of cybernetics.The process of neural network optimization was regarded as a nonlinear MIMO control system.Particularly,the gradient descent method was pointed out to be completed equivalent as the integral control algorithm.(2)Based on the Model-Free Adaptive Control(MFAC)algorithm,further researches were conducted by the proposed equivalent method.A novel optimization algorithm named CFGD(Compact Form Gradient Descent)was proposed,based on the control algorithm of the Compact Form Dynamic Linearization(CFDL)MFAC.Furthermore,CFGD algorithm was proved to be the general form of the traditional gradient descent method,with the additional ability to adaptively adj ust its learning rate.Effectiveness of the proposed CFGD algorithm was demonstrated on benchmark datasets,and CFGD was proved to be more efficient and stable than the traditional gradient descent method.(3)A novel optimization algorithm named PFGD(Partial Form Gradient Descent)was proposed,based on the control algorithm of the Partial Form Dynamic Linearization(PFDL)MFAC.Furthermore,PFGD algorithm was proved to be a general form of the traditional Momentum method,with the additional ability to adaptively adjust both its learning rate and momentum.Effectiveness of the proposed PFGD algorithm was demonstrated on benchmark datasets,and PFGD was proved to be even more efficient than CFGD.(4)Intelligent prediction models of the dust removal efficiency in the industrial electrostatic precipitation process were built based on the traditional RNN and its variant LSTM.With the comparison test to traditional modeling algorithms,RNN was verified to be effective in the modeling of the electrostatic precipitation process.On the experimental dataset of modeling the efficiency of the electrostatic precipitation,the RMSEP of LSTM model is 0.2207,and the lowest RMSEP of traditional regression models is 0.23 16.Moreover,the proposed optimization algorithms based on cybernetics were applied to optimize the LSTM model.As a result,a lower RMSEP(0.2097)was achieved with a higher training efficiency.This value is 9%lower than the lowest value of traditional regressive models,and 5%lower than that of LSTM model optimized by traditional gradient descent method,which demonstrated the proposed optimization methods to be effective and practical again. |
PDF Full Text Request