مدلسازی خروجی های NOx از دیگهای بخار با سوخت زغال سنگ با استفاده از مطلوبیت رگرسیون بردار پشتیبانی با بهینه سازی کلونی مورچه ها
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|7773||2012||12 صفحه PDF||سفارش دهید||محاسبه نشده|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Engineering Applications of Artificial Intelligence, Volume 25, Issue 1, February 2012, Pages 147–158
Modeling NOx emissions from coal fired utility boiler is critical to develop a predictive emissions monitoring system (PEMS) and to implement combustion optimization software package for low NOx combustion. This paper presents an efficient NOx emissions model based on support vector regression (SVR), and compares its performance with traditional modeling techniques, i.e., back propagation (BPNN) and generalized regression (GRNN) neural networks. A large number of NOx emissions data from an actual power plant, was employed to train and validate the SVR model as well as two neural networks models. Moreover, an ant colony optimization (ACO) based technique was proposed to select the generalization parameter C and Gaussian kernel parameter γ. The focus is on the predictive accuracy and time response characteristics of the SVR model. Results show that ACO optimization algorithm can automatically obtain the optimal parameters, C and γ, of the SVR model with very high predictive accuracy. The predicted NOx emissions from the SVR model, by comparing with the BPNN model, were in good agreement with those measured, and were comparable to those estimated from the GRNN model. Time response of establishing the optimum SVR model was in scale of minutes, which is suitable for on-line and real-time modeling NOx emissions from coal-fired utility boilers.
Nitric oxide (NO) and nitrogen dioxide (NO2) are collectively known as NOx, they have a number of negative effects on air quality: they contribute to photochemical smog, visibility reduction, acid rain and also have a negative impact on human health (Muzio and Quartucy, 1997). Coal is the most popular fuel used in power plants due to its low cost and availability. However, the emission of the nitrogen oxides during coal combustion is a significant pollutant source in the environment (Hill and Smoot, 2000). The main driver for investigations into modeling NOx emissions from coal fired utility boilers is two-fold. On the one hand, with the increasing demand for electricity especially in developing countries, many power plants would need to be built. To satisfy the strict environmental regulation, these power plants are faced with one of the most critical issues for regulations conformity concerning the availability of suitable measurements to monitor the NOx emissions (Tronci et al., 2002). Commonly, continuous emission monitoring systems (CEMSs) are used to measure the emissions. Though this hardware-based CEMS is capable of measuring stack gases with high credibility, they are still relatively expensive to purchase, install and maintain. Besides, due to the harsh environment the analyzer is frequently off-line for maintenance and some redundancy to improve system reliability is also highly desirable. A potentially attractive alternative to installing a CEMS is the use of a PEMS, which estimates the NOx emissions on the basis of their dependence on other relevant system variables using suitable algorithms. That means stack gases from the combustion chamber are possible to be predicted indirectly. At present, the majority of industrial facilities are equipped with distributed control systems (DCSs), which supply a great deal of information on what is happening in the plant (Copado and Rodriguez, 2002). Once a suitable model between the NOx emissions and the various process parameters of the boiler is determined, information downloaded from DCS can be used to infer the emissions at the stack. PEMS offers a number of advantages over expensive CEMS such as in parallel with hardware sensors (Yang et al., 2000), easily implemented on existing hardware (Graziani et al., 2004), real time estimation of NOx emissions (Matsumura et al., 1998), the advantages of lower cost, lower maintenance and higher reliability than more traditional hardware CEMS (Baines, 1999 and Tronci et al., 2002). PEMS has been successfully applied to many different combustion processes including boilers, furnaces and turbines (Kamas and Keeler, 1995). On the other hand, combustion optimization (Radl, 2000 and Zhou et al., 2005) has been proved to be an effective way to realize low NOx combustion in coal fired utility boilers, in which low NOx emission is achieved by carefully setting operational parameter of the boiler using artificial intelligence such as neural network, expert system, fuzzy logic and genetic algorithms. Core of the combustion optimization system will be a NOx emissions model of the boiler incorporated in the software. In other words, the relation between the NOx emission and various operational parameters such as coal quality, load, primary and secondary air velocity, speed of mills and others must be well known. Therefore, in order to reduce NOx emissions, a model predicting NOx emissions from various parameters of the boiler must be established at first. This model can be derived from the physical processes in the boiler. However, consisting of many combustion dynamics, fluid mechanics, heat transfer and nitrogen conversion chemistry, the overall dynamics of the boiler shows the strong non-linear inter-relations and the mutual dependence of various variables; building such an accuracy model for NOx emissions is not a trivial task and sometimes impossible. Because of the complexity of NOx emissions modeling, neural network models represent a valid alternative to this issue in the last ten years. Multi-Layer Perceptron (MLP) models were used to develop PEMS (Baines, 1999 and Graziani et al., 2004), software sensors (Dong et al., 1995, Matsumura et al., 1998 and Tronci et al., 2002) and have been successfully applied to industry (Kamas and Keeler, 1995). The operation experience showed that PEMS predications closely match measured CEMS results with the predicted NOx values typically within 20% of the actual data as measured by hardware CEMS (Kamas and Keeler, 1995). Fuzzy neural network was proposed to model NOx emissions (Ikonen et al., 2000). Liu and Huang (1998) proposed a fuzzy logic model to generate a reliable emissions model and dealt with environmental and economic dispatching when only limited experimental data was available. Various neural networks were also developed for the modeling and control of the nitrogen oxide emissions from coal-fired boilers (Reinschmidt and Ling, 1994, David and Samuelsen, 1995 and Chan and Huang, 2003). A BPNN-based NOx emissions model was incorporated in several combustion optimization software packages and showed good operation experiences (Radl, 2000, Booth and Roland, 1998 and Jia, 2007). A time delay ANN model was designed for the dynamic prediction of nitrogen oxides and carbon monoxide emissions from a fossil fuel power plant (Adali et al., 1999). Neural network-based black box model and a genetic algorithm-based gray-box model were employed to predict the NOx emissions in a 500 MWe coal-fired power plant (Li et al., 2003). A neural network model consisting of 41 input parameters of the boiler, 6–10 neurons in the hidden layer and 1 output neuron was developed to model the NOx emissions and other performances of a 540 MW capacity generator (Frenken et al., 1996). A cascading neural network was used to model the NOx emissions in a dual fired drum type boiler with full load 300 MWe with oil firing or 200 MWe with coal firing (Dong et al., 1995). Later, a genetic algorithm-based neural network was developed for the identical power plant (Li et al., 2002). Estimation of NOx emissions in thermal power plants using a combination of neural network and computational fluid dynamics (CFD) was also performed by Ferretti and Piroddi (2001). NOx emissions from a combined-cycle natural gas power plant have been investigated using artificial neural networks (Azid et al., 2000), and the deviation of the predicted NOx emissions is less than 5% of the measured values by the CEMS system. A MLP was presented to model the gaseous emissions emanating from the combustion of coal on a chain-grate stoker-fired boiler (Chong et al., 2001). Zhou et al., 2001 and Zhou et al., 2004 have proposed an approach to predict the nitrogen oxides (NOx) emissions characteristics of a large capacity pulverized coal fired boiler with artificial neural networks (ANNs). In summary, various variants of artificial neural networks for NOx emissions modeling have attracted much attention in last ten years, as reviewed by Kalogirou (2003). Despite neural networks (NNs) having been used widely in modeling NOx emissions from coal fired utility boiler, the neural network suffers from a number of weaknesses (Vong et al., 2006), which include the need for numerous controlling parameters such as the number of hidden neurons and the learning rate, difficulty in obtaining a stable solution and the danger of over-fitting. Moreover, the selection of network architecture is still problematic and time consuming task when developing a model for practical situation (Niska et al., 2004 and Benardos and Vosniakos, 2007). Some improvements to basic neural network have attracted much attention of many researchers. A parallel genetic algorithm (GA) is proposed to selecting the inputs and designing the high-level architecture of a multi-layer perceptron model (Niska et al., 2004 and Benardos and Vosniakos, 2007). The results show that the GA is a capable tool for tackling the practical problems of neural network design. However, the process remains to be computationally expensive and time demanding (Niska et al., 2004). In recent years, SVR has been successful in mapping the complex and highly nonlinear relationship between system input and output, such as the daily meteorological pollution prediction (Osowski and Garanty, 2007), automatic signature recognition (Frias-Martinez et al., 2006), bearing fault detection (Samanta et al., 2003). It would appear that it also learns the relationship between the NOx emissions and the boiler operational conditions. Compared with the MLP (for BPNN instance) models, the SVR model has certain advantages. Firstly, training for the SVR results in a global minimum (Frias-Martinez et al., 2006). On the other hand, the training of BPNN may become trapped at a local minimum (Vong et al., 2006). The second advantage is that the model parameters of SVR are fewer than those of BPNN. With the Gaussian kernel function, there are only two design parameters that need to be tuned, i.e. the generalization parameter C and the width parameter γ in the kernel function. The third advantage is that it is relatively easier to achieve good generalization because structural risk minimization principle is applied by minimizing an upper bound on the expected risk whereas the traditional empirical risk minimization is used in BPNN minimizing the error on the training data (Frias-Martinez et al., 2006). Nevertheless, to the authors' knowledge, SVR has never been applied to model NOx emissions from coal fired utility boiler in the references. It is meaningful to investigate the applicability of SVR for NOx emissions modeling, which will be favorable to develop the on-line and real-time NOx emissions monitoring system and combustion optimization software package. This study will extend the use of SVR in NOx emissions modeling in two ways: (1) unlike former studies that implemented NN model on very small samples (Zhou et al., 2001 and Zhou et al., 2004), this study applied SVR in more realistic conditions characteristic of an actual power plant. Indeed, once NOx emissions model has been built, it must be able to accurately validate a new dataset, which contains in practice a large number of cases. This study contributes to the existing literature using a sufficient sample size for training and validating the SVR models in a NOx emissions framework. (2) Before the SVR can be implemented, two parameters, i.e., the generalization parameters C and Gaussian kernel parameter γ, have to be optimized in order to construct an efficient prediction model. Extracting the optimal parameters is crucial when implementing SVR. Consequently, an ACO-based technique was proposed to perform the selection procedure. ACO has been applied to optimize the weights and bias of BPNN by several researchers ( Liu et al., 2007 and Socha and Blum, 2007). Theoretically, ACO can also be suitable for the selection of SVR model parameters. The technique is also compared to commonly used grid search method (Hsu et al., 2003). The paper is organized as follows. Section 2 presents an introduction of SVR model fundamentals and the model parameters selection technique. ACO-based SVR is also described in this section. Section 3 gives the evaluation measures adopted in this study. Experimental data used to train and validate the SVR model for NOx emissions is given in Section 4. Section 5 presents the implementation process of SVR model, while Section 6 explains the prediction results of the SVR model. The detailed performance comparison between the SVR model and the ANN models is provided in Section 7. Finally, conclusions are drawn in Section 8
نتیجه گیری انگلیسی
In this paper, different models for the estimation of NOx emissions, obtained on the basis of data acquired on the output of an actual power plant, were introduced and their estimation capabilities were described. In particular, the performance (the predictive capability and the time response) of SVR model were compared to those of both BPNN and GRNN models. Because of its universal approximation ability, SVR can be used to model nonlinear processes. Choosing optimal parameters for the models is an important step in the modeling stage. ACO optimization in combination with the holdout method is a reliable way to determine the optimal model parameters, referred to as ACO–SVR. Compared to the widely used grid search method for selection of parameters pair (C, γ), ACO–SVR needs less computing time and shows slightly better predictive accuracy. Additionally, the ACO–SVR can automatically select the optimal model parameters without any manual operation. The model obtained this way was found to have good generalization property. 92% of cases in test data D2 had the relative error smaller than 5%. The predictive accuracy, characterized by mean relative error MRE and correlation coefficient R, and the time response characteristics of the ACO–SVR model are greatly dependent on the product of the population m and the maximum iterations G. The monotonic relations found between them can help us to make the tradeoff. The optimization procedure with the population m=5 and the maximum iterations G=30 took 164 s to obtain a sufficiently accurate SVR model for NOx emissions. The computing time in scale of less than 3 min is suitable for on-line and real-time building of the SVR model. Comparative study shows that the SVR model demonstrated much better performance than the BPNN model in terms of the predictive accuracy and the model robustness. SVR model also presented slightly better predictive accuracy than the GRNN model. The main advantage of the SVR model over the GRNN model is that the former can be incorporated into optimization algorithms such as ACO and GA to regulate the inputs of model to achieve the expected NOx emissions, which is the basic idea of combustion optimization to reduce NOx emissions from coal-fired power plants.