Emission monitoring systems using artificial neural networks.

摘要

The problems of global warming and air pollution have led to enforcement of stringent constraints by governments and international community on the release of emissions. These constraints are nowadays among the most important factors impacting plant operations. Inferential sensing techniques have been gaining momentum recently as viable alternatives to hardware sensors (i.e. Continuous Emission Monitoring System) in various situations. The core of inferential sensing is built on modeling and estimation techniques.; This thesis work investigated and developed an inferential sensing technique for process emission monitoring using artificial neural networks. Neural Networks are powerful tool for modeling highly complex non-linear systems, especially when the physics of the systems is not clearly understood or difficult to determine as with emission from processes. Three different neural networks techniques were employed to model emission from the furnace unit of an industrial boiler. Multilayer Perceptron Neural Network was first investigated to construct the desired model; a trade-off was struck between the accuracy and the computational complexity of the model. Due to the drawbacks of multilayer perceptron neural network in getting trapped in local solution and it sensitivities to initial value of weights and biases other approaches were investigated. Two models were developed with two different sets of data using Radial Basis Function Network so that adequate and parsimonious model can be obtained with no local minimum problem. While a third technique was improvised to seize the merits of Neural Networks and Particle Swarm Optimization to deal with approximation embedded in the estimation of the derivative in the multilayer perceptron thereby leading to a more accurate model. On a final note, real plant's data were archived from an industrial process that produces ample amount of emission; the data were used to develop a model for the process using the modeling techniques highlighted above. The results obtained clearly shows the ability of neural network techniques in developing process emission model.