Emission control
As environmental regulations get tighter, it becomes more important to have better control on sulphur dioxide emissions from waste incineration plants, as well as to operate the flue gas desulphurization (FGD) efficiently. Nonlinear modelling can determine the effects of significant variables and help us achieve this... by Abhay Bulsari and Jouni Perttilä Ekokem Oy is the leading hazardous and municipal waste treatment company in Finland. It has three rotary kilns for hazardous waste incineration and one grate-fired boiler for domestic and industrial waste. For these four lines, there are three flue gas cleaning systems, of which two are wet scrubbing systems and one is a humidified dry system called NID™ by Alstom. The NID process is connected to rotary drums and boilers of line 2 and medium temperature kiln. The gas cleaning system was constructed in 2007. It uses unslaked lime and activated carbon as reagents. Unslaked lime is charged from a silo to the hydrator, where a small amount of water is sprayed on the lime, which partly hydrates. It overflows to the next vessel, called a mixer, where further slaking takes place with three water spraying nozzles. Slaked lime, which is still freely fluent, flows into a vertical, rectangular flue gas duct, which is the reactor. Activated carbon is also fed into the reactor. This dust mixture is entrained with flue gas flow up to the fabric filter. During the 20 metre flow in the reactor duct and on the filter surface cake, chemicals react with the acidic gases and harmful pollutants in the flue gas. Dust mixture is mainly recirculated through the mixer to the reactor and partly discharged to the ash silo. The main acidic gases SO2 and HCl form dry salts CaSO3, CaSO4 and CaCl2. A part of the cleaned gas is recirculated to the reactor, if it is needed for increasing the flue gas velocity. The process flowsheet is shown in Figure 1. Incineration line 2 is meant for hazardous waste incineration, but it is used also for various kinds of solid and liquid wastes. Middle temperature kiln is mainly used for solvent and oil contaminated soil remediation, where in-situ methods are not effective enough. The raw concentration of SO2 is in the range of 400 to 900 mg/Nm3 while HCl is in the range of 300 to 1400 mg/Nm3. Due to hydroscopic property of CaCl2 a certain amount of HCl is desirable for moistening reactants and improving desulphurization reactions. Concentrations in the stack have to be under the EU waste incineration directive limits of 50 mg/Nm3 for SO2 and 10 mg/Nm3 for HCl. Variations in the incoming gas are quite large, so some buffer removal capacity has to remain. Additional chemical expenditure increases costs by two ways: procurement and disposal. Operating the process closer to optimal conditions also reduces disturbances and increases the capacity. It is therefore quite important to operate the flue gas desulphurization at a good efficiency. Mathematical modelling Mathematical models represent knowledge of quantitative effects of relevant variables in a concise and precise form. They can be used instead of experimentation if they are reliable enough. Mathematical models also permit the user to carry out various kinds of calculations, like optimisation, which can be used to improve the efficiency of processes. Mathematical modelling is performed in many ways, and different ways are suitable in different situations. Physical or phenomenological modelling is not particularly effective for predicting the results of heterogeneous reactions. Physical modelling for desulphurization requires quantitative knowledge of heat transfer, mass transfer and fluid dynamics coupled with multiple chemical reactions. In addition, the varying surface area of the slaked lime particles as they react with hydrochloric acid and sulphur dioxide is nearly impossible to estimate. This requires plenty of assumptions and simplifications. The reactions taking place at different temperatures are poorly known, let aside their kinetics. Even if possible, the solution of such partial differential equations tends to be very slow, making it unsuitable for on-line process guidance. On the other hand, empirical and semi-empirical modelling does not need any major assumptions or simplifications. Empirical models simply describe the observed behaviour of a process. Conventional techniques of empirical modelling are linear statistical techniques, which tend to have limitations because nothing in nature is very linear. It therefore makes sense to use better techniques which take nonlinearities into account. Nonlinear modelling Nonlinear modelling is empirical or semi-empirical modelling which takes at least some nonlinearities into account. The older techniques include polynomial regression, linear regression with nonlinear terms and nonlinear regression. These techniques have several disadvantages compared to the new techniques of nonlinear modelling based on free-form nonlinearities. Nonlinear modelling can also be performed with feed-forward neural networks, series of basis functions, multivariate splines, kernel regression and other techniques. Among these new techniques, feed-forward neural networks have turned out to be particularly valuable in process modelling. Feed-forward neural networks have several features which make them efficient tools for nonlinear empirical modelling. Besides the universal approximation capability, it is usually possible to produce nonlinear models with some extrapolation capabilities with feed-forward neural networks. They consist of neurons or calculation units in layers directionally connected to others in the adjacent layers. Nonlinear modelling of industrial processes Besides the power generation sector, nonlinear modelling has successfully been used in several other industrial sectors including plastics, metals, concrete, glass, pharmaceuticals, medicine, biotechnology, mineral wools, semiconductors and food. It has been successfully utilized for a variety of purposes including quality control, product development, process guidance, software sensors and fault detection. Some things are common to process modelling of various kinds of processes. One would like to determine the best values of the feed characteristics effecting to process variables such as product properties, emissions and other consequences of the process are within desired limits. Within these constraints preferably a production economic variable, e.g. production rate, raw material consumption, energy efficiency, purity, number of defects, emissions or other important variable, is maximized or minimized. The problem looks somewhat similar from the process modelling point of view for a wide variety of processes. From the process modelling point of view, emissions, efficiency, product properties, corrosion, wear, production rate, operating costs, etc. are consequences of feed characteristics and process variables. The modelling procedure Most plants today collect large amounts of data which is normally left unused even though it can contain valuable information. Data is normally collected in terms of selected variables, after deciding the averaging and frequency at which the data should be taken. The data can sometimes contain faulty observations which need to be removed. If the amount of data is too large, a suitable amount needs to be selected such that there is a good variation in the variables of interest. Then the data is analysed in several ways. First of all, it is a good idea to plot the variables against time, and then some variables against other variables. Basic statistics of the data are calculated, followed by correlation matrices. A cluster analysis may also be performed if there is a risk that the data is in the form of only a few clusters. Then the data may be pre-processed. For example, some variables may be combined to define new variables. In this case, the ratio of SO2 after cleaning to the incoming SO2 is a good output variable. After this stage, feed-forward neural networks with various sets of input variables and with different number of nodes in the hidden layer are trained and tested. Changes in the sets of input variables are made based on what is observed in the runs. Transformations of input and output variables are considered. From the better models obtained so far, some parameters are pruned and tested. Input variables are reduced where possible. Finally, a few such models are validated against unseen data. Then the selected models are implemented in software like LUMET system. Normal operation data taken from the plant One minute average data of the process history database was picked in several patches of a couple of days each. Periods were selected according to variations in the operating conditions. Totally 20 different variables, all continuously measured, were examined. For control of the plant the most important input parameter is so called calculated relative humidity of outgoing flue gas. Also incoming SO2 and HCl concentration, temperature and flue gas volume are essential, but difficultly controlled. Only 4000 observations were selected from that data with a simple algorithm for model development purposes. Results Nine most affecting input variables were picked and finalised after several months of developing preliminary nonlinear models. After a large number of attempts of feed-forward neural networks with different configurations, one model with the following characteristics was selected for implementation in the plant. Output variable 1: sulphur dioxide in the cleaned flue gasrms |err| of output variable 1: 32.716mean |err| of output variable 1: 15.654max |err| of output variable 1: 313.477standard deviation of output 1: 125.466Correlation of output variable 1: 0.9320 Output variable 2: temperature of the cleaned flue gasrms |err| of output variable 2: 0.951mean |err| of output variable 2: 0.664max |err| of output variable 2: 10.659Correlation of output variable 2: 0.9156 The rms error, or roughly speaking, the standard deviation of prediction errors of sulphur dioxide concentration in the cleaned gas was 32.72, while that of the temperature of the flue gas was 0.95°C. The correlation coefficients are 93.2% and 91.6% respectively, which are good for this kind of data. The nonlinear models also show effects of the composition variables and process variables as are expected. Figure 2 shows the effect of sulphur dioxide in the incoming gas on sulphur dioxide in the cleaned gas for different water flow rates to the mixer, while keeping other input variables constant. It is natural that if there is more sulphur dioxide in the incoming flue gas, more sulphur dioxide will be in the cleaned gas, if other variables are constant, which is why all the curves are rising. More water in the mixer results in more desulphurization, which is why the curves are lower for higher values of water flow rate to the mixer. The desulphurization efficiency reduces with increasing amounts of sulphur dioxide in the incoming flue gas. Figure 3 shows the effect of temperature of the incoming flue gas on temperature of the cleaned gas for different moisture contents in the incoming flue gas. If the temperature of the incoming gas is higher, the temperature of the cleaned gas will also tend to be higher, which is reason for ascending curves. More moisture leads to less addition of water in the mixer which causes a smaller drop in temperature of the flue gas, explaining the rise in the curves for higher moisture content. Implementation in the control room Engineers and plant operators cannot be expected to be familiar with nonlinear modelling. Therefore, a set of software components has been developed over the years which allow facile use of the nonlinear models without needing to know the details of the embedded mathematics. Depending on the end use, some of these software components are assembled into a software package, referred to as a LUMET system. LUMET systems have many different functionalities, besides simply being able to predict the values of the output variables. For example, effects of input variables or pairs of input variables can be plotted in different ways. Figures 2 and 3 have been plotted using one such system. To see the combined effects of two input variables, one can either plot contours, or surface plots in three dimensions. Besides these, LUMET systems can also have facilities for calculating the best values of process variables in presence of constraints and optimisation objectives. In some cases, a software component allowing for updating the models is also included. Conclusions While environmental regulations get tighter, power plants are also under pressure to improve their production economics. Power plant managers would like to derive the most from their plants without exceeding the emission limits. One alternative is to add new equipment, which requires significant investments. A better alternative is to first utilize the existing equipment more efficiently. Nonlinear modelling is a powerful new way of extracting quantitative knowledge of a process in terms of measured variables from production data. When properly implemented, nonlinear models can be valuable tools for the operators, plant managers as well as R&D engineers. The nonlinear models developed for Ekokem’s hazardous waste incineration plant in Riihimäki have been implemented in suitable software to help improve the operation of the desulphurization process. The nonlinear model is used by operators and production engineers particularly when there are special circumstances or when planning to incinerate high sulphur waste. The LUMET system has been found to be a quick and easy tool to predict how the process will behave under different conditions. Besides helping in making the operation smoother, it is expected to also save on lime consumption. It will allow us to operate the plant better when the incoming sulphur dioxide content in the flue gas is high, thereby increasing the capacity and availability of the waste incineration line. Abhay Bulsari, Nonlinear Solutions Oy, Turku, Finlande-mail: abulsari@abo.fi Jouni Perttilä, Ekokem Oy Ab, Riihimäki, Finlande-mail: Jouni.Perttila@ekokem.fi