Determination of High Temperature Corrosion Rates of Steam Boiler Evaporators Using Continuous Measurements of Flue Gas Composition and Neural Networks

: The use of low-emission combustion techniques in pulverized coal-ﬁred (PC) boilers are usually associated with the formation of a reduced-gas atmosphere near evaporator walls. This increases the risk of high temperature (low oxygen) corrosion processes in coal-ﬁred boilers. The identiﬁcation of the dynamics and the locations of these processes, and minimizing negative consequences are essential for power plant operation. This paper presents the diagnostic system for determining corrosion risks, based on continuous measurements of ﬂue gas composition in the boundary layer of the combustion chamber, and artiﬁcial intelligence techniques. Experience from the implementation of these measurements on the OP-230 hard coal-ﬁred boiler, to identify the corrosion hazard of one of the evaporator walls, has been thoroughly described. The results obtained indicate that the continuous controlling of the concentrations of O 2 and CO near the water wall, in combination with the use of neural networks, allows for the forecasting of the corrosion rate of the evaporator. The correlation between ﬂue gas composition and corrosion rate has been demonstrated. At the same time, the analysis of the possibilities of signiﬁcantly simplifying the measurement system by using neural networks was carried out. have agreed the the


Introduction
Very strict NO x emission regulations, rising prices in the European Union emissions trading system (EU ETS) and in general zero emission EU policy in energy production forces fossil power plants to adapt markedly the technology and asks for the increased flexibility of large utility units. In the case of solid pulverized fuel combustion, the reduction in NO x emissions with a simultaneous increase in economic efficiency forces different low-emission combustion techniques to be implemented. The latter requirement, in correlation with different power demands changing during the day, intensifies the high-temperature corrosion processes of evaporator walls. The issue stated above also greatly negatively affects the reliability of the boiler and the operating availability of the power unit.
The high-temperature corrosion process in steam boilers is a very complex phenomenon, briefly described in [1][2][3]. The main key factors in low oxygen corrosion in pulverized fired utility boilers combustion conditions by monitoring gas concentration (e.g., O 2 , CO, and other components) [22,23]. To reduce the costs of measurements covering all of the furnace walls, it is worth taking additional measures to reduce the number of measuring points, while maintaining the possibility of the correct interpretation of the results.
For the last two decades, artificial neural networks (ANNs) have been successfully used to represent many processes with complex physics and chemistry. In particular, their fault tolerance for noisy data (typical of boiler data), their abilities to learn from examples, and their generalization properties, have provided great potential in many application areas related to complex combustion processes [24]. They have been used successfully in diverse applications, ranging from modelling coal/biomass devolatilization [25] or rate of char oxidation [26], through predicting quality of fuel blends [27] to predicting the emissions of dioxins from waste power plants [28], identifying slagging in the vicinity of burners [29], or soot-blowing optimization [30]. ANNs were also used to predict the corrosion behavior of metal alloys [31], corrosion fatigue crack growth rates [32], or the atmospheric corrosion behavior of steel and zinc [33]. However, they are mostly focused on particular process phenomena rather than evaluating the corrosion danger, especially low oxygen corrosion in utility boilers. There are also some known works related to ANN utilization for recognizing or finding the compositions of gas mixtures [34,35]. However, none of them (except preliminary authors' papers [36]) were used for predicting O 2 and CO concentrations in the area near the evaporator walls of a steam boiler to forecast the threat of high-temperature corrosion.

OP230 Boiler as an Object of Research
The high-temperature corrosion of evaporators in utility pulverized fired boilers is the adverse phenomenon caused by low emission technologies. It is very important to diagnose corrosion risks in the boiler, since it may cause unplanned and long outages of the boilers. The method of the online monitoring of high-temperature corrosion risks in steam boilers is elaborated, based on the continuous measurement of O 2 and CO concentration in flue gases in the boundary layer of the evaporator [19,37]. The main idea of this method is to determine the high reductive zones with a low level of O 2 and a high level of CO concentration, which can be presented as (current or averaged in the given period) maps of corrosion risk [38]. This information can be directly used by operators or DCS systems to optimize air or fuel flows to the boiler to avoid such zones [39].
The tested OP230 boiler is a tangentially fired hard-coal boiler with natural circulation and a nominal live steam capacity of 230 Mg/h under pressure 13.6 MPa, as shown in Table 1, and it is combined with a 50 MW turbine. Three ball mills supply the fuel to three levels of jet burners where the highest level connects to mill no. 1, while mill no. 3 connects to lowest level of burners. At low capacity, the boiler is powered by two mills operating in variable configurations. Currently, the acceptable level of NOx emissions is maintained through the use of low-emission combustion, mainly the advanced OFA (over-fire air), ROFA (rotating opposed fired air system), air staging, and SNCR (selective non-catalytic reduction) methods. The main role in reducing NO x emissions is played by the primary methods that keep NO x emissions as low as possible. In the Energies 2020, 13, 3134 4 of 17 main burner zone, the average air excess coefficient is maintained at the level of 0.90-1.05. The upper level of the burners is at 15.9 m, while ROFA air is fed at 20 m and 23 m (on the right and left wall). Reducing the airflow in the combustion zone results in the formation of a reducing atmosphere (with substoichiometric air) in a substantial part of the combustion chamber, which is also affected by the configuration of the working burners and boiler capacity. Besides, the frequent changes in boiler efficiency and changes in the configuration of the working mills (and the related work of the burners), contribute to frequent changes in the composition of flue gas near the water walls. Locally, high corrosion rates in this boiler have been observed in recent years that have caused evaporator failure. Such failures always bring measurable financial losses due to the lack of energy production and the need for quick repairs. Therefore, it is advisable to take action aimed at limiting the risk of failure.
It is obvious that many factors, such as fuel composition (including Cl, S) [1,3,40], stoichiometry, flue gas composition (e.g., CO, H 2 S, HCl, SO 2 ) [41,42], and the chemical composition of ash deposits on the tube surfaces [1,3,[42][43][44][45], play a significant role in the high-temperature corrosion process. Moreover, the corrosion history from the period of long term boiler operation may impact the present corrosion rate [46]. The experiences of the authors and other researchers show that the measurement of O 2 and CO near the water wall can be a good indicator of corrosion risk assessment for a given boiler and its fuel [19,23]. Therefore, a lot of research was carried out regarding the implementation of the monitoring of the gas atmosphere near the combustion chamber walls. One such system has been developed and tested in the OP-230 boiler [38,47]. This system uses a continuous on-site measurement of the O 2 and CO concentrations in the layer near the wall of the boiler chamber walls, and enables the identification of a reductive atmosphere and may be useful for preventing intensified corrosion.
The use of the continuous monitoring of the flue gas composition in the area near the evaporator walls for indicating the presence of a reductive atmosphere seems to be an easy and cost-effective solution. With such system boiler can be operated with continuous optimization of the combustion process by making adequate adjustments in air or fuel stream distribution. A decrease in reducing zones number or level of CO and thus the diminishment of the corrosion rate can be achieved. Figure 1 shows the location of the measuring points (no.  for the periodic measurements of the exhaust gas composition using portable gas analyzers (Testo 350), while the red circles indicate the points where the continuous measurements were carried out (S1-S5). The flue gas was sampled for 5 min (with stream approximately 1 L/min) from a layer close to the surface of the evaporator wall, and the measured O 2 and CO values were averaged. Then, the measuring tracks were cyclically blown with compressed air for 10 min (due to the need to clean them). Therefore, the results were recorded in the system with a frequency of 15 min. For this reason, the measurement can be called quasi-continuous.
The electronic transducers and software used in the measuring system have been specially developed and built for this purpose, while the other components are available on the market, such as O 2 zirconia sensors (Bosch LSM11), CO sensors (CF40000 Membrapor/CO-GP-40-5 GRI Instrument), and H 2 S sensors (3HLM CiTicel). The oxygen concentration was measured directly on the boiler wall at points S1-S5, while the CO and H 2 S concentrations were measured in the control cabinet (serving the entire wall) to which the exhaust gas from the S1-S5 connectors was sucked, as shown in Figure 2. The concentration of H 2 S was measured on-site at only one point to test for a correlation with O 2 and CO. Examples of continuous measurements of the O 2 and CO concentrations in the boundary layer at the two selected points on the font wall of the OP-230 boiler-nos. S2 and S3 in Figure 1-are shown in Figure 3.
gas composition using portable gas analyzers (Testo 350), while the red circles indicate the points where the continuous measurements were carried out (S1-S5). The flue gas was sampled for 5 min (with stream approximately 1 L/min) from a layer close to the surface of the evaporator wall, and the measured O2 and CO values were averaged. Then, the measuring tracks were cyclically blown with compressed air for 10 min (due to the need to clean them). Therefore, the results were recorded in the system with a frequency of 15 min. For this reason, the measurement can be called quasi-continuous.   The electronic transducers and software used in the measuring system have been specially developed and built for this purpose, while the other components are available on the market, such as O2 zirconia sensors (Bosch LSM11), CO sensors (CF40000 Membrapor/CO-GP-40-5 GRI Instrument), and H2S sensors (3HLM CiTicel). The oxygen concentration was measured directly on the boiler wall at points S1-S5, while the CO and H2S concentrations were measured in the control cabinet (serving the entire wall) to which the exhaust gas from the S1-S5 connectors was sucked, as shown in Figure 2. The concentration of H2S was measured on-site at only one point to test for a correlation with O2 and CO. Examples of continuous measurements of the O2 and CO concentrations in the boundary layer at the two selected points on the font wall of the OP-230 boiler-nos. S2 and S3 in Figure 1-are shown in Figure 3.   The electronic transducers and software used in the measuring system have been specially developed and built for this purpose, while the other components are available on the market, such as O2 zirconia sensors (Bosch LSM11), CO sensors (CF40000 Membrapor/CO-GP-40-5 GRI Instrument), and H2S sensors (3HLM CiTicel). The oxygen concentration was measured directly on the boiler wall at points S1-S5, while the CO and H2S concentrations were measured in the control cabinet (serving the entire wall) to which the exhaust gas from the S1-S5 connectors was sucked, as shown in Figure 2. The concentration of H2S was measured on-site at only one point to test for a correlation with O2 and CO. Examples of continuous measurements of the O2 and CO concentrations in the boundary layer at the two selected points on the font wall of the OP-230 boiler-nos. S2 and S3 in Figure 1-are shown in Figure 3.

Application of Neural Networks to Estimate the Rate of Corrosion at OP-230 Boiler
There are many different types of artificial neural network (ANN). Typically, an ANN is made up of some number of interconnected processing elements, called neurons, linked together by connections with weights. An ANN is a "black box" with inputs and outputs. The most important part of the ANN performance is the learning phase, which can be understood as the process of updating the connection weights and, sometimes, network structure as a response to a set of training data, obtained usually from experimental measurements. The overall behavior and functionality of an ANN depend on the network architecture, activation function, learning methods, and training set quality. No high computing power is needed for the operation of the neural network used. Among the different types of ANNs, the Feed-Forward Multilayer Perceptron (MLP) network, presented in Figure 4a, is the most popular supervised-learning network and has found application across many disciplines, also in the presented work. It is particularly suited for use in tasks involving the prediction and classification of complex data. A MLP network is typically organized in a series of layers made up of a number of interconnected neurons, each of which contains a transfer function f, as shown in Figure 4b. Input data are first presented to the network via the input layer, which then communicates these data via a system of weighted connections to one or more hidden layers, where the actual neuron processing is carried out. The response of the hidden layers is then transferred to an output layer where the network outputs are computed. During the training phase, as shown in Figure 4, these outputs are compared with the known desired values of the given process. Arising errors from the difference between the predicted and known values are back-propagated through the network and used to update the weights and biases of all the connections. The process is successively repeated until the errors fall below the given threshold. At this stage, the network training is complete. However, it should be appreciated that training is not necessarily a one-off procedure, since the network can learn and be updated on-line if the system or process changes. The most widely used transfer functions for the neurons are tanh, sigmoid, rectified linear unit, or linear function, and are shown in Figure 5. In current work, the tanh function was used (where tanh denotes hyperbolic tangent).

Application of Neural Networks to Estimate the Rate of Corrosion at OP-230 Boiler
There are many different types of artificial neural network (ANN). Typically, an ANN is made up of some number of interconnected processing elements, called neurons, linked together by connections with weights. An ANN is a "black box" with inputs and outputs. The most important part of the ANN performance is the learning phase, which can be understood as the process of updating the connection weights and, sometimes, network structure as a response to a set of training data, obtained usually from experimental measurements. The overall behavior and functionality of an ANN depend on the network architecture, activation function, learning methods, and training set quality. No high computing power is needed for the operation of the neural network used. Among the different types of ANNs, the Feed-Forward Multilayer Perceptron (MLP) network, presented in Figure 4a, is the most popular supervised-learning network and has found application across many disciplines, also in the presented work. It is particularly suited for use in tasks involving the prediction and classification of complex data. A MLP network is typically organized in a series of layers made up of a number of interconnected neurons, each of which contains a transfer function f, as shown in Figure 4b. Input data are first presented to the network via the input layer, which then communicates these data via a system of weighted connections to one or more hidden layers, where the actual neuron processing is carried out. The response of the hidden layers is then transferred to an output layer where the network outputs are computed. During the training phase, as shown in Figure 4, these outputs are compared with the known desired values of the given process. Arising errors from the difference between the predicted and known values are back-propagated through the network and used to update the weights and biases of all the connections. The process is successively repeated until the errors fall below the given threshold. At this stage, the network training is complete. However, it should be appreciated that training is not necessarily a one-off procedure, since the network can learn and be updated on-line if the system or process changes. The most widely used transfer functions for the neurons are tanh, sigmoid, rectified linear unit, or linear function, and are shown in Figure 5. In current work, the tanh function was used (where tanh denotes hyperbolic tangent).  For error estimation, the mean square percentage error (MSPE) was used: where ddi is normalized expected value (real value), dti is normalized estimated by ANN value, n is the number of samples. As the inputs to ANN all measured the O2 concentrations through O2 sensors, they were used -5 inputs. The outputs were O2 and CO concentrations in other measuring points where developed O2 sensors were no placed (for clarification see Figure 1). The value of the threshold in training the back-propagation algorithm was varied depending on the selected point, usually between 0.1 and 0.01. Figure 6 shows the results of the ANN training and testing for O2 and CO predictions. In general, the results of the ANN performance are, in most cases acceptable. However, there are some higher deviations for several points, in particular, where high standard deviations during measurements of O2 and CO occurred (training data), which also caused relatively high MSPE error. However, training was carried out for only eight cases, and testing for only two. Thus, the obtained ANN performances should be assumed as a success. Better predictions were found by authors in other own similar investigations performed on other boilers [36,48], where more training and testing cases were available.
(a) (b)  For error estimation, the mean square percentage error (MSPE) was used: where ddi is normalized expected value (real value), dti is normalized estimated by ANN value, n is the number of samples. As the inputs to ANN all measured the O2 concentrations through O2 sensors, they were used -5 inputs. The outputs were O2 and CO concentrations in other measuring points where developed O2 sensors were no placed (for clarification see Figure 1). The value of the threshold in training the back-propagation algorithm was varied depending on the selected point, usually between 0.1 and 0.01. Figure 6 shows the results of the ANN training and testing for O2 and CO predictions. In general, the results of the ANN performance are, in most cases acceptable. However, there are some higher deviations for several points, in particular, where high standard deviations during measurements of O2 and CO occurred (training data), which also caused relatively high MSPE error. However, training was carried out for only eight cases, and testing for only two. Thus, the obtained ANN performances should be assumed as a success. Better predictions were found by authors in other own similar investigations performed on other boilers [36,48], where more training and testing cases were available.
(a) (b) For error estimation, the mean square percentage error (MSPE) was used: where dd i is normalized expected value (real value), dt i is normalized estimated by ANN value, n is the number of samples. As the inputs to ANN all measured O 2 concentrations through O 2 sensors, were used −5 inputs. The outputs were O 2 and CO concentrations in other measuring points where developed O 2 sensors were no placed (for clarification see Figure 1). The value of the threshold in training the back-propagation algorithm was varied depending on the selected point, usually between 0.1 and 0.01. Figure 6 shows the results of the ANN training and testing for O 2 and CO predictions. In general, the results of the ANN performance are, in most cases acceptable. However, there are some higher deviations for several points, in particular, where high standard deviations during measurements of O 2 and CO occurred (training data), which also caused relatively high MSPE error. However, training was carried out for only eight cases, and testing for only two. Thus, the obtained ANN performances should be assumed as a success. Better predictions were found by authors in other own similar investigations performed on other boilers [36,48], where more training and testing cases were available.   For error estimation, the mean square percentage error (MSPE) was used: where ddi is normalized expected value (real value), dti is normalized estimated by ANN value, n is the number of samples. As the inputs to ANN all measured the O2 concentrations through O2 sensors, they were used -5 inputs. The outputs were O2 and CO concentrations in other measuring points where developed O2 sensors were no placed (for clarification see Figure 1). The value of the threshold in training the back-propagation algorithm was varied depending on the selected point, usually between 0.1 and 0.01. Figure 6 shows the results of the ANN training and testing for O2 and CO predictions. In general, the results of the ANN performance are, in most cases acceptable. However, there are some higher deviations for several points, in particular, where high standard deviations during measurements of O2 and CO occurred (training data), which also caused relatively high MSPE error. However, training was carried out for only eight cases, and testing for only two. Thus, the obtained ANN performances should be assumed as a success. Better predictions were found by authors in other own similar investigations performed on other boilers [36,48], where more training and testing cases were available.
(a) (b)  The criteria for the degree of corrosion risk were selected based on literature data and the authors' own experiences. In Table 2, the degree of corrosion risk (from none to very high) is marked with appropriate colors that correspond to the colors of the maps showing the degree of corrosion hazard generated in the Rachel software, [36,38].

Results and Discussions
To elaborate on this corrosion rate, the data from the high-temperature corrosion monitoring system, operated in the OP-230 hard coal-fired boiler were used against typical ultrasonic thickness measurements of the evaporator wall tubes. The concentration values of O 2 and CO in the boundary layer were collected for six months of boiler operation, whilst wall thickness measurements were carried out in 2014 and 2017 (in 280 and 315 points on the wall, respectively). Moreover, artificial neural networks were used to determine the O 2 and CO gas species concentrations in the boundary layer of the whole wall of the boiler evaporator, based on five O 2 sensor indications. Typical feed-forward neural networks with one hidden layer with 8-10 neurons and tanh activation functions were used with a back-propagation training algorithm.

Determination of Corrosion Rate in the Tested Boilers
The map of change of evaporator thickness over time was drawn up based on measurements made by ultrasonic technique during the boiler outages in April 2014 and in February 2017. Over almost three years boiler operation (including downtime caused by breakdowns and renovation) loss of tube thickness on the front wall reached over 2 mm at selected points, which gives the even rate of 0.6-0.7 mm per year, as shown in Figure 7. The water wall tube thickness was compared at 280 points on the front wall between levels 11.2 m and 21.2 m.
The power unit worked 17,566 h during this period. For 40% of this period (of working hours), it ran with electric power in the range of 42-52 MW, but the average power of the unit in the analyzed time period was about 39 MW. Figure 8 presents changes in the electric power of the unit (which corresponds to boiler capacity) during eighteen months of operation.
Frequent changes in capacity, the configuration of operating burner systems, and frequent boiler outages mean that the average values of the oxygen concentration in the boundary layer recorded online are slightly higher than during the short and periodic measurements carried out using portable flue gas analyzers under steady conditions. The continuous measurements also included short-term boiler operation periods, with higher air excess in the burner zone during the configuration changes of working mills.  Frequent changes in capacity, the configuration of operating burner systems, and frequent boiler outages mean that the average values of the oxygen concentration in the boundary layer recorded online are slightly higher than during the short and periodic measurements carried out using portable flue gas analyzers under steady conditions. The continuous measurements also included short-term boiler operation periods, with higher air excess in the burner zone during the configuration changes of working mills.

Periodic Measurements of the Flue Gas Composition near the Wall of the Combustion Chamber
Measurements of the exhaust gas composition in the boundary layer on the front wall of the boiler were made for several variants of the boiler with a working system of secondary air ROFA.  Frequent changes in capacity, the configuration of operating burner systems, and frequent boiler outages mean that the average values of the oxygen concentration in the boundary layer recorded online are slightly higher than during the short and periodic measurements carried out using portable flue gas analyzers under steady conditions. The continuous measurements also included short-term boiler operation periods, with higher air excess in the burner zone during the configuration changes of working mills.

Periodic Measurements of the Flue Gas Composition near the Wall of the Combustion Chamber
Measurements of the exhaust gas composition in the boundary layer on the front wall of the boiler were made for several variants of the boiler with a working system of secondary air ROFA.

Periodic Measurements of the Flue Gas Composition near the Wall of the Combustion Chamber
Measurements of the exhaust gas composition in the boundary layer on the front wall of the boiler were made for several variants of the boiler with a working system of secondary air ROFA. Variants of the boiler differed in power, the configuration of the working mills, and air discharge. The number of measurement variants was limited since the work of the unit was not tampered with, and measurements were performed only for the conditions existing on a given day.
At 26 available measuring points on the front wall, the concentrations of O 2 and CO were measured. These data were used to study the neural networks in Rachel software. For some boiler operational parameter variants, additional components of the exhaust gas were measured, mainly to determine their correlation with the concentrations of CO and O 2 . Such measurements have to answer to whether it is worth taking other exhaust components, apart from or together with CO as reliable indicators of the occurrence of a reducing atmosphere. For example, Figure 9 shows the difference in O 2 and CO concentrations in the layer near the front water wall at different boiler operating parameters. As can be seen, the area in which the reduction zone occurs promotes faster corrosion changes as the boiler operating conditions change. In the initial testing period, other exhaust gas components, such as H 2 S, were also measured, but due to the clear correlation between H 2 S and CO, as shown in Figure 10, it was abandoned to simplify the measurement system and make it cheaper [38]. operational parameter variants, additional components of the exhaust gas were measured, mainly to determine their correlation with the concentrations of CO and O2. Such measurements have to answer to whether it is worth taking other exhaust components, apart from or together with CO as reliable indicators of the occurrence of a reducing atmosphere. For example, Figure 9 shows the difference in O2 and CO concentrations in the layer near the front water wall at different boiler operating parameters. As can be seen, the area in which the reduction zone occurs promotes faster corrosion changes as the boiler operating conditions change. In the initial testing period, other exhaust gas components, such as H2S, were also measured, but due to the clear correlation between H2S and CO, as shown in Figure 10, it was abandoned to simplify the measurement system and make it cheaper [38]. The presented example, shown in Figure 9, shows the differences in the maps of O2 and CO concentrations depending on the boiler operating parameters (on February 18th, the block power was 44 MW and mills 2 and 3 worked; on February 23rd, the block power was 52.6 MW and mills 1, 2, and 3 worked). In the case of this boiler, this is very important, because its operation is characterized by a large variability of performance. For this purpose, oxygen and carbon monoxide concentration measurements are required for many different boiler operating conditions (load, activity of mills, burner configurations, air distribution, etc.). Periodical measurements of the composition of the exhaust gas near the evaporator wall of the boiler were made for several variants of the boiler operating conditions (differing in load, working mills and burners, and air distribution), shown in Table 3. The number of measurement variants was limited due to the fact that, during the heating season, in the combined heat and power (CHP) plant, the authors had no influence on the possibility of changing the boiler capacity. The average results of six of these variants (variants no. 3-8, which were tested in 2016), are presented in Figure 10     As can be seen from the data shown in Figure 11, the relationship between the instantaneous flue gas composition (corresponding to constant boiler operation parameters), and the average corrosion rate, is characterized by a large spread of results, and the final result will be greatly influenced by the history of the boiler operation parameters. For oxygen concentration, the average values from long-term tests are usually significantly higher than the instantaneous values measured at full boiler capacity. It should be noted that the presented individual variants of the boiler's operation relate to several hours of measurements with constant boiler parameters, while the corrosion rate is, in this case, calculated in nm/hour for 17,566 working hours in the analyzed period (including planned and emergency boiler outages in this period). The presented example, shown in Figure 9, shows the differences in the maps of O 2 and CO concentrations depending on the boiler operating parameters (on February 18th, the block power was 44 MW and mills 2 and 3 worked; on February 23rd, the block power was 52.6 MW and mills 1, 2, and 3 worked). In the case of this boiler, this is very important, because its operation is characterized by a large variability of performance. For this purpose, oxygen and carbon monoxide concentration measurements are required for many different boiler operating conditions (load, activity of mills, burner configurations, air distribution, etc.). Periodical measurements of the composition of the exhaust gas near the evaporator wall of the boiler were made for several variants of the boiler operating conditions (differing in load, working mills and burners, and air distribution), shown in Table 3. The number of measurement variants was limited due to the fact that, during the heating season, in the combined heat and power (CHP) plant, the authors had no influence on the possibility of changing the boiler capacity. The average results of six of these variants (variants no. 3-8, which were tested in 2016), are presented in Figure 10 as a comparison of the O 2 concentration with the average corrosion rate. As can be seen from the data shown in Figure 11, the relationship between the instantaneous flue gas composition (corresponding to constant boiler operation parameters), and the average corrosion rate, is characterized by a large spread of results, and the final result will be greatly influenced by the history of the boiler operation parameters. For oxygen concentration, the average values from long-term tests are usually significantly higher than the instantaneous values measured at full boiler capacity. It should be noted that the presented individual variants of the boiler's operation relate to several hours of measurements with constant boiler parameters, while the corrosion rate is, in this case, calculated in nm/hour for 17,566 working hours in the analyzed period (including planned and emergency boiler outages in this period).  As can be seen from the data shown in Figure 11, the relationship between the instantaneous flue gas composition (corresponding to constant boiler operation parameters), and the average corrosion rate, is characterized by a large spread of results, and the final result will be greatly influenced by the history of the boiler operation parameters. For oxygen concentration, the average values from long-term tests are usually significantly higher than the instantaneous values measured at full boiler capacity. It should be noted that the presented individual variants of the boiler's operation relate to several hours of measurements with constant boiler parameters, while the corrosion rate is, in this case, calculated in nm/hour for 17,566 working hours in the analyzed period (including planned and emergency boiler outages in this period).

Continuous Measurements of O2 and CO Concentration near the Combustion Chamber Wall
In order to achieve a good compromise between complexity and accuracy, it was decided to place five zirconium probes in preselected places on the boiler wall, shown in Figure 1. These locations were selected because they guaranteed the changeable level of oxygen concentration when

Continuous Measurements of O 2 and CO Concentration near the Combustion Chamber Wall
In order to achieve a good compromise between complexity and accuracy, it was decided to place five zirconium probes in preselected places on the boiler wall, shown in Figure 1. These locations were selected because they guaranteed the changeable level of oxygen concentration when boiler performance is changed (capacity, air flows, mills configuration). The CO concentration was also measured at these points. Based on the measurement results, the boiler operator can then decide to change the operating conditions while maintaining low NO x emissions. The monitoring system may appear useful when making decisions regarding protective air installations, investments in anti-corrosive coatings, or combustion process optimization activities.
The dedicated software (Rachel) allows us to average the measurement results from selected periods of the boiler operation time. Assuming the appropriate corrosion hazard criteria, depending on the concentrations of O 2 and CO in the exhaust gas, the degree of corrosion risk of the water wall in the form of maps can be presented, both for instantaneous values and averaged values over a selected period. Figure 12 shows maps with average concentrations of O 2 and CO from the six months of boiler operation and the corresponding corrosion risk map.
Energies 2020, 13, x FOR PEER REVIEW 13 of 18 Figure 12. The maps of O2, CO concentration, and corrosion risk, generated by Rachel software from half year OP-230 boiler operation data [38]. The scale of the corrosion risk in this figure is the same as in Table 2-from none (blue color) up to very high (red color).

Determination of Corrosion Rate Based on Measurements of O2 and CO Concentrations in the Boundary Layer of the Combustion Chamber
Data representing the loss of water wall tube thickness in the long-term operation of the boiler were used to determine the corrosion rate (nm/h), taking into account the boiler's operating hours during that period. To define the correlation between the corrosion rate and O2/CO concentrations, the averages over six months of operation maps of O2 and CO concentrations, generated by artificial neural networks and a map of the wall thickness loss of the evaporator tubes based on the ultrasonic technique, were analyzed. The test procedure is shown schematically in Figure 13 and the results of the analysis are shown in Figures 14-16. The comparison was made for almost 200 points, where O2 and CO concentrations at these points were calculated using neuronal networks. It can be seen that spots of high corrosion rates correspond with low O2 levels and high CO levels.  Figure 12. The maps of O 2 , CO concentration, and corrosion risk, generated by Rachel software from half year OP-230 boiler operation data [38]. The scale of the corrosion risk in this figure is the same as in Table 2-from none (blue color) up to very high (red color).
The corrosion hazard level, shown in Figure 12, indicates good compliance with the corrosion rate map developed based on ultrasonic measurements, as shown in Figure 7. The area of the corrosion risk map that was compared with the corrosion rate map is marked with a dotted yellow line in Figure 12. It should be noted, however, that the rate of corrosion was determined after a much longer time of boiler operation than the period of averaging the results of continuous measurements of exhaust gas composition. The monitoring of gas composition was carried out for a longer period than one year, but in the meantime, the measuring installation was modernized, and it was decided to take into account only the time during which the system worked without breaks and failures. In contrast, the determination of the average corrosion rate, based on the ultrasonic measurements in this boiler, was possible only during outages of the boiler, planned earlier by the operator.

Determination of Corrosion Rate Based on Measurements of O 2 and CO Concentrations in the Boundary Layer of the Combustion Chamber
Data representing the loss of water wall tube thickness in the long-term operation of the boiler were used to determine the corrosion rate (nm/h), taking into account the boiler's operating hours during that period. To define the correlation between the corrosion rate and O 2 /CO concentrations, the averages over six months of operation maps of O 2 and CO concentrations, generated by artificial neural networks and a map of the wall thickness loss of the evaporator tubes based on the ultrasonic technique, were analyzed. The test procedure is shown schematically in Figure 13 and the results of the analysis are shown in Figures 14-16. The comparison was made for almost 200 points, where O 2 and CO concentrations at these points were calculated using neuronal networks. It can be seen that spots of high corrosion rates correspond with low O 2 levels and high CO levels. The graphs visibly show the increase in the corrosive dynamic with decreasing oxygen content and increasing CO content taken from historical data. The corrosion rate can be interrelated with oxygen concentration using the following formula: and with carbon monoxide concentration with the formula: wcor = 14.0·CO + 18.5 (nm/y) (3) Figure 13. The diagram of the testing procedure to determine the correlation between flue gas composition and the corrosion rate of the evaporator. The graphs visibly show the increase in the corrosive dynamic with decreasing oxygen content and increasing CO content taken from historical data. The corrosion rate can be interrelated with oxygen concentration using the following formula: Taking into account the above results, the corrosion rate can be interrelated with oxygen and CO concentrations using the following formula: wcor = −0.04· (O2) 2 − 3.94 O2 + 0.28· CO 2 + 8.21· CO + 43.17 (nm/h) The coefficients of Equations (2)-(4) were determined using the least-squares method. For this purpose, only data from the boiler area located between the measurements points No. 1-26 were used (i.e., data from the inside the convex polygon whose vertices were the following measuring locations: 1, 2, 6, 13,18,26,25,24,23,22,17,9,3), as shown in Figure 6. This is because, inside this area, neural networks are particularly good at predicting O2 and CO concentrations. The O2 and CO concentration values were averaged over time. Linear interpolation was used to determine the average O2 and CO concentrations in the same locations where corrosion rate was measured.
Assuming the above Equation (4), the corrosion rate was calculated based on the average (semiannual) values of O2 and CO concentration. In this way, the relationship between the value calculated from Equation (4) and the measured corrosion rate was determined. A considerable dispersion of values is visible on the graph. However, the proposed trend line indicates quite good agreement with the measurement results, shown in Figure 16. Analyzing the results presented in Figure 16, it should be noted that the online registration period of the flue gas composition in the boundary layer of the combustion chamber, for which the calculated corrosion rate was determined, does not coincide exactly with the period for which the measured corrosion rate was determined. In the analyzed case, the recording time of the flue gas  Taking into account the above results, the corrosion rate can be interrelated with oxygen and CO concentrations using the following formula: wcor = −0.04·(O2) 2 − 3.94 O2 + 0.28·CO 2 + 8.21·CO + 43.17 (nm/h) (4) The coefficients of Equations (2)-(4) were determined using the least-squares method. For this purpose, only data from the boiler area located between the measurements points No. 1-26 were used (i.e., data from the inside the convex polygon whose vertices were the following measuring locations: 1,2,6,13,18,26,25,24,23,22,17,9,3), as shown in Figure 6. This is because, inside this area, neural networks are particularly good at predicting O2 and CO concentrations. The O2 and CO concentration values were averaged over time. Linear interpolation was used to determine the average O2 and CO concentrations in the same locations where corrosion rate was measured.
Assuming the above Equation (4), the corrosion rate was calculated based on the average (semiannual) values of O2 and CO concentration. In this way, the relationship between the value calculated from Equation (4) and the measured corrosion rate was determined. A considerable dispersion of values is visible on the graph. However, the proposed trend line indicates quite good agreement with the measurement results, shown in Figure 16. Analyzing the results presented in Figure 16, it should be noted that the online registration period of the flue gas composition in the boundary layer of the combustion chamber, for which the calculated corrosion rate was determined, does not coincide exactly with the period for which the measured corrosion rate was determined. In the analyzed case, the recording time of the flue gas The graphs visibly show the increase in the corrosive dynamic with decreasing oxygen content and increasing CO content taken from historical data. The corrosion rate can be interrelated with oxygen concentration using the following formula: w cor = 1.1·(O 2 ) 2 − 19.3·O 2 + 93.8 (nm/h) (2) and with carbon monoxide concentration with the formula: w cor = 14.0·CO + 18.5 (nm/y) By comparing both of the above characteristics, the effect of the O 2 and CO levels on the corrosion rate can be presented simultaneously, as in Figure 15.
Taking into account the above results, the corrosion rate can be interrelated with oxygen and CO concentrations using the following formula: The coefficients of Equations (2)-(4) were determined using the least-squares method. For this purpose, only data from the boiler area located between the measurements points No. 1-26 were used (i.e., data from the inside the convex polygon whose vertices were the following measuring locations: 1,2,6,13,18,26,25,24,23,22,17,9,3), as shown in Figure 6. This is because, inside this area, neural networks are particularly good at predicting O 2 and CO concentrations. The O 2 and CO concentration values were averaged over time. Linear interpolation was used to determine the average O 2 and CO concentrations in the same locations where corrosion rate was measured.
Assuming the above Equation (4), the corrosion rate was calculated based on the average (semi-annual) values of O 2 and CO concentration. In this way, the relationship between the value calculated from Equation (4) and the measured corrosion rate was determined. A considerable dispersion of values is visible on the graph. However, the proposed trend line indicates quite good agreement with the measurement results, shown in Figure 16.
Analyzing the results presented in Figure 16, it should be noted that the online registration period of the flue gas composition in the boundary layer of the combustion chamber, for which the calculated corrosion rate was determined, does not coincide exactly with the period for which the measured corrosion rate was determined. In the analyzed case, the recording time of the flue gas composition on-site measurement covered only about 20% of the boiler operation time, which may be the reason for additional uncertainty related to the history of the boiler operating conditions.

Conclusions
A simple tool for the continuous control of the corrosion hazard in PC boilers has been developed and tested. One of the most important advantages of the proposed monitoring system is the simplicity of the measuring system. The simplification of this system was possible due to the use of neural networks. This allowed for simplified measurements of the exhaust gas composition in the boundary layer of the combustion chamber at only a few selected points, and then the determination of results on the entire wall. By carrying out continuous measurements of O 2 and CO, using historical data on the operating conditions of the boiler, and formulating equations for the corrosion rate for the tested boiler, it is possible to make a quick corrosion risk assessment and forecast future wear on evaporator tubes. This, in turn, is very important information for boiler service teams, which may plan outage times, avoiding unexpected and costly failures. It can also be used for periodic boiler operation optimization.