The fireside corrosion of various components of a coal-fired boiler may be attributed to the following:

- Reducing (sub-stoichiometric) conditions caused by impingement of incompletely combusted coal particles and flames,

- Accelerated oxidation from overheating,

- Molten salt or slag-related attack.

The fireside corrosion is generally localized in regions specifically near the walls of the burners. Reducing atmospheric corrosion can result from direct reaction of the waterwall tubes with a sub-stoichiometric gaseous environment containing sulphur, or with partially combusted char containing FeS. The reducing conditions have two major influences on the corrosion process. First, they tend to lower the melting point of any deposited slag, increasing its ability to dissolve the normal oxide scales, and second, the stable gaseous sulphur compounds include H2S, which is more corrosive than SO2 that prevails under oxidizing conditions [5, 6]. The overheating leads to an accelerated oxidation of both the fireside and the steam side surfaces of the tubes that produces thickened, hard scales. Above 570 °C, a considerable non-protective scale of wustite (FeO) can be formed on iron, which culminates in rapid oxidation [7].

Mechanisms and causes and of Fireside corrosion

Fireside corrosion is often a very complicated process. Generally, the same corrosion types as those from the steam side can be found in the fireside of a tube, but the acid and base corrosion mechanisms are usually defined differently. Corrosion of super-heaters has become the limiting factor in the attempt of increasing the final steam temperature in energy production. The strength of present steels would allow very high steam pressures and temperatures, but the high oxidation rate or superheater corrosion caused by contaminants in fuel limits the maximum allowed temperature, especially in the combustion of biomass and recycled fuels [3,4]. Superheater corrosion can take place inside the tube, as steam side corrosion, or outside the tube, as fireside corrosion. The steam side corrosion is mainly due to faults in the steam quality control, while the fireside corrosion is caused by corrosive components present in the flue gas. The methods of measuring these two main corrosion types differ remarkably. It is seldom possible to measure directly the thickening of the oxide scale inside a tube, while a significant material loss on the outside can be detected with smaller effort. Further, coal exhibits wide variations in many of its properties, including composition. Low-rank coals commonly contain relatively large amounts of organically associated elements such as Na, Mg, Ca, K, and Sr, present as salts of organic acid groups, as well as mineral grains, although they commonly contain less chlorine than high-rank coals. The Indian coals in general are low rank coals having much less chlorine contents [8]. By contrast, high-rank coals commonly contain more iron and sulfur than low-rank coals. When coal particles are fired into the boiler furnace, the moisture and the volatile species are driven off; the fixed carbon in the pulverized particles begins to burn. The contained mineral matter may be melted or vaporized, and is largely oxidized. The sulphur bearing compounds in the coal (such as FeS) are converted to oxides, such as Fe2O3, K2O, Na2O, SO2 and SO3. The relative proportions of SO2 and SO3 in the flame depend on the oxygen availability and the flame temperature. SO2 is thermodynamically favored at higher temperatures; the formation of SO3 can be catalyzed by certain metal oxides. Thus, the gaseous species released, as the coal passes through the flame, contain potential corrodents such as sulphur, vapor of alkali metal salts, and chlorine compounds [9]. For high-temperature corrosion due to sulfur, the presence of alkali pyrosulfates in furnace wall deposits and alkali-iron trisulfates on the leading edge of the final superheater tube surface is the primary cause for tube wastage. The sulfates interact with Fe2O3 in ash and with SO3 present at the tube surface, to form alkali tri-sulfates. The fireside corrosion rate enhances due to formation and de-stabilization of molten complex alkali-iron tri-sulphates, with increasing temperature. Some of the most common reactions of forming alkali-iron tri-sulphates are [10]:

Chlorine (more than 0.2 wt%) has been found to promote the release of both Na and K into the flame [12], and acts as a strong catalyst for the molten tri-sulphate attack. There is also evidence that HCI formed in the flame can destroy the Fe2O3 layer on a steel surface, thereby exposing it to additional oxidative attack [3, 4]. Research on alkali tri-sulfates reveals that high-temperature corrosion of furnace walls and superheater surfaces is minimized by maintaining an oxidizing environment and avoiding flame impingement on furnace walls.

Artificial Intelligence applications in corrosion modeling

Quantitative determination of fireside corrosion rates, in conjunction with pertinent mechanisms as a function of boiler coal chemistry, fly ash chemistry and operating parameters based on first principle kinetic modeling, has remained a fairly difficult topic. This is due to the phenomenological complexity and sometimes non-linear relationship between the dependent and independent variables of the fireside corrosion phenomenon. Therefore, first principle based kinetic model predictions are not always amenable to realistic plant operating conditions. Therefore, simplified assumptions are often made to overcome a phenomenological complexity, which idealizes realistic problems. Recently, Data driven Artificial Intelligence (AI) or Computational Intelligence (CI) based techniques are increasingly used with to functionally map, in an accurate way, the input-output relationship of complex corrosion processes, although these are quite scanty. In principle, AI techniques (such as ANN, Fuzzy logic and Genetic algorithms are intelligent information-treatment systems with the characteristics of adaptive learning.

Application of an artificial neural network (ANN) model has been reported for data driven modelling for prediction of ash deposition in boiler heat transport system [13]. ANN has also been developed successfully to characterize thermal behavior of boiler tubes in the presence of fouling on the basis of plant data [14], and it has also been reported that such models have been applied to control and minimize the effect of fouling in biomass boilers [15]. Application of an expert system based theoretical approach for boiler fouling assessment has been proposed [16]. A comparative study of Fuzzy logic and ANN has been reported [17] for the prediction of the remaining life of boiler tubes subjected to various damage mechanisms. An adaptive Neuro-Fuzzy technique has been attempted [18] for predicting and characterizing coal slagging in a power plant. However, application of ANN modeling to predict the oxidation scale deposition rate in boiler operations is relatively scanty in the published literature. The objective of the present work is to develop a multi-layer feed forward ANN model to predict explicitly the fireside corrosion rate, as a function of measured plant data (input/output parameters), namely coal ash and sulfur contents, wt% of Na2O and K2O in fly ash, and operating variables such as flue gas temperature and percentage excess air intake for coal combustion of a typical coal fired Indian operating boiler. The proposed ANN model also attempts to characterize effects of some of the operational parameters on the fireside corrosion behavior. In this proposed ANN model, efficient gradient based network training algorithm and optimized neural network architecture have been incorporated to improve the network learning algorithm and to minimize training errors during the network learning process.

Adaptive neural networks modeling of fireside corrosion rate

Mathematically, this can be given as:

where αis neuron j's output from k's layer and βis the bias weight for neuron j in layer k. The model fitting parameters wijk are the connection weights and fk 's are activation functions. Most neural networks will tend to forget old information if we attempt to add new information incrementally. When developing an artificial neural network to perform a particular pattern-classification operation, one typically proceeds by gathering a set of exemplars, or training patterns, then using these exemplars to train the system. During the training, information is encoded in the system by the adjustment of weight values. Once the training is deemed to be adequate, the system is ready to be put into production, and no additional weight modification is permitted. This operational scenario is acceptable, provided the problem domain has well-defined boundaries and is stable. Under such conditions, it is usually possible to define an adequate set of training inputs for whatever problem is being solved.

Optimal network learning algorithm

MLP is usually trained using the error back propagation algorithm. This popular algorithm works by iteratively changing a network's interconnecting weights, such that the overall error (i.e. between observed values and modeled network outputs) is minimized. In principle, network training/ learning uses one of several possible optimization methods to minimize this error term. There are various back propagation (BP) algorithms, such as Scaled Conjugate Gradient (SCG), Levenberg-Marquardt (LM), Gradient Descent with Momentum (GDM), variable learning rate Back propagation (GDA) and Resilient back Propagation (RP) [19]. There is a variety of network optimization techniques that uses gradient of a function to be optimized. One of the most recently developed efficient versions of the quasi-Newton optimization methods is the BFGS algorithm [21, 22], which has largely replaced the classical DFP algorithm.

In the quasi-Newton training method, the weights are updated using the following iterative procedure:

The matrix B here needs not be computed. It is successively estimated employing rank 1 or rank 2 updates, following each line search in a sequence of search directions. This is algorithmically given as follows:

In this iterative algorithm, Bi-1 is the previous value of B. The two important algorithmic relationships to compute ΔBi are as follows [24]:

The above expression pertains to DFP algorithm and the equation given below is the BFGS algorithm:

where,

Thus, BFGS potentially reduces the number of function evaluations [24, 25] required to achieve an optimization procedure which has been successfully applied elsewhere by one of the authors [26].

Network configuration Input-Output variables

The following activation functions, denoted by φ (v), have been used in this neural model. First, there is the Threshold Function which takes on a value of 0, if the summed input is less than a certain threshold value of v, and the value of 1, if the summed input is greater than or equal to the threshold value [27, 28].

Secondly, there is the Piecewise-Linear function. This function again can take on the values of 0 or 1, but can also take on values between that, depending on the amplification factor in a certain region of linear operation.

Thirdly, there is the sigmoid function. This function can range between 0 and 1, but it is also sometimes useful to use the -1 to 1 range. The sigmoid function is the hyperbolic tangent function given below [29]:

The neural prediction based on two network architectures (MLP 9-4-1 and MLP 9-13-1) is compared with the regression fit between predicted and measured data. It may be observed from the simulation that all these two proposed network architectures have almost similar prediction characteristics.

It may be observed from the figure that the fireside corrosion rate monotonically increases with an increased sulphur content in coal. The corrosion rate varies from 0.1 to 1.0 mm/year (approx.) with respect to sulphur content of coal ranging from 0.3% - 1.0%. There is comparatively less sulfur content present in Indian coal as compared to British coal. However, the sulphur content is sufficient to the extent it ensures that any sodium and potassium compounds released in the combustion process form fusible sulfates. Sulfur typically is found as sodium sulfate in coal ash. At high temperature it dissociates [12] and eventually alters the basicity of the molten ash deposits. Sulfur reacts with sodium in the melt altering the concentration of Na2O, and thereby changing the corrosion rates.

It may be observed from the figure that the corrosion rate linearly increases with increased SOx percentage in flue gas.The predictions conform to the realistic situation [30]. The corrosion rate varies from 0.1 - 0.95 mm/year (approx.) with respect to SOx concentrations of 230 -245 mg/m3 in flue gas. Fireside corrosion in boiler areas is associated with the deposition of alkali sulfates onto the metal surface, their concentration being increased at the metal surface by absorption onto the porous fly ash. The salts formed are usually molten and contain free sulfur trioxide in flue gas, which dissolves the protective oxide film to form iron and chromium based sulfates.

It may be observed from the figure that the corrosion rate enhances with an increase in flue gas temperature. Corrosion reactions get activated with higher activation energy as the flue gas temperature increases [10, 31].

It may be observed from the figure that the corrosion rate increases monotonically with an increased value of excess intake air percentage in the furnace. The sulfur in coal reacts with oxygen in the combustion air forming SO2, and, if the residence time and O2 content are more than sufficient, it forms also SO3. The initial sulfidation reaction seldom continues so as to result in internal sulfidation of the metal.

It may be observed from the figure that the corrosion rate enhances with increased wt% of Na2O and K2O in fly ash. The corrosion rate varies from 0.01 to 0.94 mm/year (approx.) with respect to 0.3-0.4 and 1.0-1.5 wt% of Na2O and K2O in fly ash, respectively. The severe fireside corrosion of tube materials is caused by condensation/accumulation of low melting-point salts from the flue gas onto the tube surface since salts containing chlorides and sulphates of sodium and potassium, easily liquefying at the operating metal temperatures [16, 32]. Reaction of alkali sulphates (Na2SO4, K2SO4) with Feoxides (deriving from oxide scales or ashes) in the presence of SO3 will result in the formation of alkali-iron trisulphates (Na, K)3Fe(SO4)3, held responsible for the degradation of coal-fired plant superheater tubes[15].

It may be observed that the corrosion rate surface in 3-D framework depicting the variation of coal ash and sulphur content has a peak value which corresponds to a higher level of ash and sulphur content of coal.

It may be observed from this figure that the absolute error drops sharply from 0.06 to 0.01 at the very early stage of training (few cycles) and, subsequently, the training and testing errors asymptotically reduces to almost zero, with further increase in the number of cycles. It may be further noted that the error bounds in the measured data and neural computation are quiet small. Although several simulation results have been generated, however, only limited results are discussed in this paper.

Conclusions

The proposed neural network model provides a reasonably accurate predictive framework and compares extremely well with the plant and experimental data. The ANN methodology shows good potential for predictions of the fireside corrosion rate as a function of input variables, namely, coal ash and sulfur contents, wt% of Na2O and K2O in fly ash and operating variables such as flue gas temperature and percentage excess air intake during combustion. This model has a relative advantage over other phenomenological and semi empirical models treating polluted data or data with complex functional dependence. Effects of coal composition and fly ash constituents and process parameters on the fireside corrosion rate have been investigated, and appropriately validated with the measured data. In the numerical domain, it has been found that the efficient gradient based network training algorithm does not require computation of numerically cumbersome Hessian matrix, or calculation of any matrix inverses. This algorithm potentially reduces the number of functions, facilitating faster convergence of training errors within a few cycles, with respect to fireside corrosion rate prediction.

References