SYSTEM AND METHOD FOR MONITORING OF EMISSIONS FROM GAS TURBINES

Abstract

Embodiments of the present disclosure relate to a system (102) and a method (1000) for monitoring of emissions from gas turbines. The system (102) includes a memory with processor-executable instructions, which on execution, causes a processor to collect a dataset of operational and environmental data from a plurality of sensors associated with gas turbines, which act as input attributes to a stacking ensemble model (SEM). Data preprocessing and Pearson correlation analysis are conducted on the collected dataset to train and test the SEM. Deviations between the predicted emissions and actual emissions are analysed to perform anomaly detection and provide recommendations to a plurality of users. Advantageously the present disclosure relates to a system and a method that aims at enhancing emission compliance, reducing environmental impacts, and improving operational efficiency.

Specification

Description:TECHNICAL FIELD
[0001] The present disclosure generally relates to the gas turbine operation and environmental monitoring. More particularly, the present disclosure relates to a system and a method for monitoring of emissions from gas turbines, aimed at enhancing emission compliance, reducing environmental impacts, and improving operational efficiency.

BACKGROUND
[0002] Energy plays a pivotal role in facilitating economic growth and serves as an indispensable component in sustaining the day-to-day operations of the current society. The growing significance of challenges such as energy security, sustainability over the long term, greenhouse gas emissions, and the diminishing availability of fossil fuels has resulted in a heightened focus on clean and renewable energy sources. The issue of providing nations with reliable, adequate, affordable, and environmentally friendly energy has gained essential significance worldwide due to the scarcity of energy resources and the resulting ecological implications.
[0003] As per the published report by the World Health Organization, air pollution can lead to cardiovascular, cerebrovascular (stroke), and respiratory diseases in adults, infants, and other age groups. It is imperative to accurately monitor the emission rate of pollutants from different sources to take the necessary action to reduce environmental emissions. Conventional methods for monitoring pollutants have proven ineffective, which inspired the researcher to move towards artificial intelligence or a machine learning-based monitoring system to measure the emission rate of the pollutants accurately. Based on the literature survey, it is evident that numerous machine-learning techniques, like NNboost, XGB, GA-ANFIS, Fuzzy, GBR, etc., have been developed and applied to predict the emission rates of NOx and CO from various processes. The development of models is crucial and should be carried out on an application basis. It has also been revealed that the emission predictive models based on the stack ensemble technique still need to be explored for CCPP.
[0004] The gas turbine in a combined cyclic power plant (CCPP) produces harmful gases like carbon monoxide (CO) and nitrogen oxide (NOx) into the atmosphere. The rate at which these gases are produced during power generation must be monitored to comply with the industrial standard for emission. Therefore, a system is required to monitor the emission from the CCPP gas turbine continuously.
[0005] To address these limitations, the present invention provides a system and a method for monitoring of emissions from gas turbines that overcomes the shortcomings of the prior art.

OBJECTS OF THE PRESENT DISCLOSURE
[0006] It is a primary object of the present disclosure to provide a predictive system and a method for monitoring of emissions from gas turbines.
[0007] It is another object of the present disclosure to develop a stacking ensemble model using NNR, GAM, BRT (base learners), and GRNN (meta learner) is proposed for the CO and NOx prediction.
[0008] It is yet another object of the present disclosure to provide a system that attains a minimum value of prediction error compared to state-of-the-art methods such as support vector regression (SVR), decision tree (DRT), and linear regression (LIR) using different performance measures in both scenarios.

SUMMARY
[0009] The present disclosure generally relates to the gas turbine operation and environmental monitoring. More particularly, the present disclosure relates to a system and a method for monitoring of emissions from gas turbines, aimed at enhancing emission compliance, reducing environmental impacts, and improving operational efficiency.
[0010] The primary aspect of the present invention is to design a system to monitor the emission from the CCPP gas turbine continuously. The disclosure aims to design a stacked ensemble machine learning (SEM) based predictive model for CO and NOx emission from a CCPP gas turbine. The neural network for regression (NNR), a generalized additive model (GAM), and the bagging of regression trees (BT) act as the base learners. A generalized regression neural network (GRNN) is used as a meta-learner for SEM. The hyperparameters of SEM are optimized using a Bayesian optimization algorithm for CO and NOX prediction.
[0011] In another aspect of the present disclosure, a method for monitoring of emissions from gas turbines is disclosed. The method begins with collecting a dataset of operational and environmental data from a plurality of sensors associated with gas turbines, which act as input attributes to a stacking ensemble model (SEM) and conducting data preprocessing and Pearson correlation analysis on the collected dataset to train and test the SEM. The generated emission predictions are combined from the SEM to generate a final emission forecast with improved emission prediction accuracy. The final emission forecast is compared to predefined regulatory thresholds. An alert is triggered if the predicted emissions exceed the regulatory limits. The med further includes analyzing deviations between the predicted emissions and actual emissions to perform anomaly detection and provide recommendations to a plurality of users.

BRIEF DESCRIPTION OF DRAWINGS
[0012] The accompanying drawings are included to provide a further understanding of the present disclosure and are incorporated in, and constitute a part of this specification. The drawings illustrate exemplary embodiments of the present disclosure, and together with the description, serve to explain the principles of the present disclosure.
[0013] FIG. 1 illustrates an exemplary graphical representation of output variables NOx and CO of the system for monitoring of emissions from gas turbines, in accordance with an embodiment of the prior art disclosure.
[0014] FIG. 2 illustrates an exemplary block diagram of the pearson correlation coefficient matrix, in accordance with an embodiment of the present disclosure.
[0015] FIG. 3 illustrates an exemplary block diagram of the system for monitoring of emissions from gas turbines for CCPP with all the input and output attributes, in accordance with an embodiment of the present disclosure.
[0016] FIG. 4 illustrates an exemplary flow diagram of the overall methodology for the prediction of CO and NOx using SEM, in accordance with an embodiment of the present disclosure.
[0017] FIG. 5 illustrates an exemplary graphical representation of the hydrographs for all the designed ML models for NOx, in accordance with an embodiment of the present disclosure.
[0018] FIG. 6 illustrates an exemplary graphical representation of the Hydrographs for all the designed ML models, in accordance with an embodiment of the present disclosure.
[0019] FIG. 7 illustrates an exemplary graphical representation of the scatter plots for SEM predictive model, in accordance with an embodiment of the present disclosure.
[0020] FIG. 8 illustrates an exemplary graphical representation of the sensitivity analysis of SEM for NOx and CO, in accordance with an embodiment of the present disclosure.
[0021] FIG. 9 illustrates an exemplary graphical representation of the error deviation comparison of all the predictive model for NOx and CO, in accordance with an embodiment of the present disclosure.
[0022] FIG. 10 illustrates an exemplary view of a flow diagram of the proposed method for monitoring of emissions from gas turbines, in accordance with an embodiment of the present disclosure.

DETAILED DESCRIPTION
[0023] The present disclosure generally relates to the gas turbine operation and environmental monitoring. More particularly, the present disclosure relates to a system and a method for monitoring of emissions from gas turbines, aimed at enhancing emission compliance, reducing environmental impacts, and improving operational efficiency.
[0024] A system is designed to continuously monitor CCPP gas turbine's emission. The disclosure aims to design a stacked ensemble machine learning (SEM) based predictive model for CO and NOx emission from a CCPP gas turbine. The neural network for regression (NNR), a generalized additive model (GAM), and the bagging of regression trees (BT) act as the base learners. A generalized regression neural network (GRNN) is used as a meta-learner for SEM. The hyperparameters of SEM are optimized using a Bayesian optimization algorithm for CO and NOX prediction. In addition to this, the performance of SEM is compared with support vector regression (SVR), decision tree (DRT), and linear regression (LIR). Simulation results demonstrate that SEM can reduce the RMSE 5.7-93.8% for NOx and 1%-41.5% for CO compared to other ML techniques. Finally, comparing the results with ML techniques existing in the literature shows the higher predictive accuracy of the proposed SEM.
[0025] The system monitors emissions from gas turbines to at least one user to access a computing device. The system may include a network, one or more computing devices, one or more users, a processing engine and a centralized server. The system may include a processor and a memory. The memory may include a set of instructions, which when executed, causes the processor to diagnose cardiac disorders of a plurality of users.
[0026] In an embodiment, the processing engine(s) may be implemented as a combination of hardware and programming to implement one or more functionalities of the processing engine(s). A database includes data that is either stored or generated as a result of functionalities implemented by any of the components of the processing engine(s).
[0027] In an embodiment, the processing engine(s) may include a training engine, a data collection engine, an application engine, a combining engine, a comparing engine, a triggering engine, an analysing and other module(s), but not limited to the likes. The other module(s) implements functionalities that supplement applications or functions performed by the system or the processing engine(s). The database serves, amongst other things, as a repository for storing data processed, received, and generated by one or more of the modules.
[0028] In an embodiment, the system may be configured to collect a dataset of operational and environmental data from a plurality of sensors associated with gas turbines, which act as input attributes to a stacking ensemble model (SEM) via the data collection engine. The data collection engine captures operational data from gas turbine sensors, including combustion temperature, fuel flow rate, gas flow rate, turbine speed, pressure levels, and pollutant concentrations such as NOx, CO, and unburned hydrocarbons. The operational and environmental data are selected from a group comprising ambient temperature (AT), ambient pressure (AP), ambient humidity (AH), air filter difference pressure (ADP), compressor discharge pressure (CDP), gas turbine exhaust pressure (GEP), turbine after temperature (ATT), carbon monoxide (CO), nitrogen oxide (NOx), turbine inlet temperature (ITT) and any combination thereof. The emission of CO increases with a decrease in the ITT and the CDP. The sensors are configured to monitor parameters including but not limited to combustion temperature, fuel flow rate, gas flow rate, turbine speed, pressure levels, and pollutant concentrations such as NOx, CO, unburned hydrocarbons and any combination thereof.
[0029] In an embodiment, the system may be configured to conduct data preprocessing and Pearson correlation analysis on the collected dataset to train and test the SEM via the training engine. The SEM is selected from Boosted Regression Trees (BRT), Generalized Additive Models (GAM), Neural Network Regression (NNR), Generalized Regression Neural Network (GRNN), and Structural Equation Modeling (SEM). An ensemble learning mechanism configured to combine predictions from the plurality of machine learning models to improve emission prediction accuracy. The ensemble learning mechanism is configured to use a combination of bagging, boosting, or stacking techniques to aggregate predictions from BRT, GAM, NNR, GRNN, and SEM models.
[0030] In an embodiment, the system may be configured to apply the SEM to the collected data to generate individual emission predictions via the application engine. The Structural Equation Modeling (SEM) model is configured to establish relationships between latent variables influencing gas turbine emissions, such as operational parameters, fuel type, and ambient conditions. The Boosted Regression Trees (BRT) model uses iterative resampling of the training dataset to improve the prediction accuracy of emissions under varying operating conditions of the gas turbine. The Generalized Additive Models (GAM) model is configured to model non-linear relationships between gas turbine operating parameters and emission outputs, allowing flexibility in capturing complex emission behavior. The Neural Network Regression (NNR) model and Generalized Regression Neural Network (GRNN) are employed to handle high-dimensional sensor data and capture intricate patterns in the operational data to predict future emission levels. The ensemble learning mechanism assigns different weights to the output of each machine learning model based on its predictive accuracy for different operational conditions. The ensemble learning mechanism continuously updates the weights of the BRT, GAM, NNR, GRNN, and SEM models based on performance metrics, improving the accuracy of emission predictions over time.
[0031] In an embodiment, the system may be configured to combine the generated emission predictions from the SEM to generate a final emission forecast with improved emission prediction accuracy via the combining engine. A predictive analytics module processes the data through the ensemble of models to generate emission forecasts.
[0032] In an embodiment, the system may be configured to compare the final emission forecast to predefined regulatory thresholds via the comparing engine.
[0033] In an embodiment, the system may be configured to trigger an alert if the predicted emissions exceed the regulatory limits via the triggering engine.
[0034] In an embodiment, the system may be configured to analyze deviations between the predicted emissions and actual emissions to perform anomaly detection and provide recommendations to a plurality of users via the analysing engine. The recommendations pertain to adjusting operations of the gas turbine to reduce emissions below the predefined thresholds.
[0035] FIG. 1 illustrates an exemplary graphical representation of output variables NOx and CO of the system for monitoring of emissions from gas turbines, in accordance with an embodiment of the prior art disclosure.
[0036] In an exemplary embodiment, information on pollutants from the gas turbine of CCPP was obtained from an undisclosed power plant in Turkey. It consists of sensor data aggregated every hour, covering 2011 to 2015. This dataset is made accessible open source through the UCI repository. The collection comprises a total of 36,733 records, each representing hourly data. The provided data set lacks continuity as it solely encompasses instances when the plant was functioning within load factors ranging from 75% to 100%. A graphical representation of the output attributes is given in Figure 1. The performance of GT is impacted by several environmental conditions like the temperature of the air compressor, humidity, exhaust temperature, input and output pressure ratio of the compressor, fuel type used for the combustion, losses due to discharges, turbine temperature (inlet and outlet), etc. CO and NOx are the primary pollutants produced by GT during energy generation. However, the generation of sulphur oxides and other pollutants depends on the type of fuel used. Furthermore, if the process operates at maximum load capacity, it will produce fewer pollutants than the process operating on partial load. Table 1 shows the statistical analysis of the two output attributes (CO and NOx) and the eight Inputs (AT, AP, AH, ADP, GEP, ITT, ATT, and CDP). The prominent parameters AT, AP, and AH are related to the air conditioning system. Meanwhile, for combined cyclic gas turbine (CCGT), the ADP, GEP, ITT, ATT, and CDP are the influential parameters.

[0037] FIG. 2 illustrates an exemplary block diagram of the pearson correlation coefficient matrix, in accordance with an embodiment of the present disclosure.
[0038] In an exemplary embodiment, understanding the relationship between input and output attributes is vital to the prediction process. The correlation coefficient (CRC) is a statistical measure that may be utilized to examine the interrelationships between attributes. The CRC value indicates the level of interdependence among the attributes. If CRC=0, it indicates that the two variables are independent. A positive CRC indicates that both variables are changing in the same direction, whereas a negative CRC indicates that they are changing in opposite directions. CRC between the attributes is estimated using the Pearson correlation, as shown in Figure 2. It can be seen from Figure 2 that CO shows higher values of CRC for all the attributes except AP. Furthermore, a significant difference in the CRC values of input attributes is seen for CO and NOx. The CO negatively correlates with AT, ADP, GEP, ITT, and CDP, with correlation coefficient values of -0.1141, -0.5172, -0.5959, -0.7232, and -0.6418, respectively. This means that the emission of CO increases with a decrease in the turbine's inlet temperature and the compressor's discharged pressure.
[0039] FIG. 3 illustrates an exemplary block diagram of the system for monitoring of emissions from gas turbines for CCPP with all the input and output attributes, in accordance with an embodiment of the present disclosure.
[0040] The emission of NOx is higher when there is a decrease in AT (-0.5699). During winter, GT is operated at higher temperatures to subsidize the NOx emission. It is also negatively correlated to ADP (-0.1619), GEP (-0.1610), ITT (-0.1179), ATT (-0.02764), and CDP (-0.1273), respectively. Furthermore, CO is positively correlated with AP (0.01403), AH (0.08432), and ATT (0.1703). Similarly, NOx is positively correlated with AP and AH with a coefficient of correlation values of 0.1837 and 0.1605. The schematic diagram for CCPP is shown in FIG. 3.
[0041] FIG. 4 illustrates an exemplary flow diagram of the overall methodology for the prediction of CO and NOx using SEM, in accordance with an embodiment of the present disclosure.
[0042] A bootstrap method aims to develop many identical autonomous forecasters and then average their results to acquire the final prediction. While using the BRT, numerous trees (decision trees) are merged to improve the forecast performance of the model, which is called bagging. Consequently, BRT is used to reduce the variance of regression trees and to alleviate the overfitting problem in a single tree. The initial stage in BRTs is to generate L separate training of the same size l as the real data by randomly selecting l out of l observations and replacing them with data from the real training dataset. Furthermore, train each tree individually using new training sets and take the mean of all predictions to arrive at a final prediction. The final prediction for BRT is given as follows.
(1)
[0043] where gi is the tree model on bootstrap data i.
[0044] Consider different base learners a1 (x), ……, an (x) to apprehend the essential concept of bagging. Assumes that the output function estimated from the set of input attributes with distribution p(x) is O(x). So, the estimated error is given as follows.
(2)
[0045] MSE is given by
(3)
[0046] The mean of the error for regression functions is defined by Equation (4)
(4)
[0047] Further assumes that all the errors are uncorrelated and unbiased.
(5)
[0048] The overall regression function is given as , whereas MSE can be calculated as
(6)
(7)
(8)
(9)
The interpretation problem associated with complicated models can be effectively addressed using a GAM. GAMs can be considered as an extension of Linear Regression models. A standard linear regression model establishes the linear relationship among input and output variables. Assume Z is the output variable with µ mean distribution and β2 variance. The linear association between Z and input xi is as follows.
(10)
where α, τ, and M are the predicted value of Z, the attribute of the predictor, and the number of predictors available, respectively.
Equation (10), can be modified using link function b to relate α to xi as
(11)
[0049] Equation (11) represents a functional expression of generalized linear models (GLMs). The GAM is an expansion of GLMS that incorporates non-linear representations of predictive variables. Non-linear predictors are connected to utilizing a suitable link function, expressing their relationship.
(12)
where g is the basis function, and φ is the parameter for the basis function.
To enhance the precision of the traditional GAM, the inclusion of pairwise correlations is proposed. Consequently, equation (12) can be adjusted as follows:
(13)
[0050] The training of the GAM relies on two crucial factors: (a) choosing the shape variable and (b) choosing the learning method employed for training the GAM. In the present investigation, the use of boosted trees and gradient boosting is employed as the shape variable and learning method to train the GAM.
[0051] NNR is the fully connected (FC) feed-forward neural network for regression analysis. The first FC layer connects with all the inputs, and every hidden layer relates to the previous layer. The output of each FC is the product of the input by a weight matrix with the addition of a bias vector. Whereas activation functions are utilized to map the input and output of the FC layers. A limited-memory Broyden-Fletcher-Goldfarb-Shanno quasi-Newton algorithm (LBFGS) is used as a training algorithm.
[0052] The GRNN differs from the backpropagation method because it doesn't require a repeated training strategy. The model can approximate any function that maps input vectors to output vectors by deriving the function estimates using the training data. Furthermore, it is a constant observation that as the size of the training set increases, the prediction error begins to diminish, with only moderate constraints on the function. The regression methodology yields a predicted value of Z that optimizes the mean-squared error (MSE). The GRNN is a computational technique used to estimate the combined probability distribution function (PDF) of variables x and Z. This estimation is achieved solely based on a provided training dataset. The system exhibits a high degree of generality as the PDF is generated only based on the data, without any preconceived notions about its structure. The conditional mean of Z for given x is given by equation (14).
(14)
where, h (x,Z) is the PDF and E is the estimator of the output values. Based on xj and Zj samples of the variables x and Z the probability estimator is defined as
(15). where m is the total number of samples, P is the sample size of x, ρ is the sample probability width, and the probability calculates the sum of sample probabilities. The calculation involves determining the meaning of the recorded values, denoted as Zj. Every recorded value is assigned a weight based on its Euclidean distance from x, with the weights being exponentially distributed. The regression from equation (16) can be immediately used for numerical data applications.
(16)
Where Fj is the scalar function given by equation (17)
(17)
[0053] The SEM is classified as one of the ensembles learning algorithms. In contrast to bagging and boosting algorithms, which involve integrating numerous models, the stacking ensemble model incorporates diverse models. According to the findings of numerical simulations, the stacking strategy delivers superior outcomes in many applications compared to any other base learner. Instead of emphasizing variance or bias, the stacking method significantly improves the model's overall predictable effect. Regarding the effectiveness of machine learning algorithms, hyperparameters are critical because they directly impact the training algorithm's behavior and the model's performance. Hyper-parameter configuration aims to fine-tune the model's precision by determining the best value for every parameter. Therefore, with the relevant hyperparameters, the model can learn the best weights using the available training data and algorithm. Generally, the hyperparameter values are selected based on prior information or using trial and error. The experts have the liberty to change the values at any step. Both methods are time-consuming, and manually selecting the ideal values of hyperparameters will take hours. The issues above can be resolved due to advancements in the optimization field. In recent years, Bayesian optimization (BO) has been widely used to select hyperparameters of ML algorithms. A global optimization algorithm searches the minimum value for the unknown function h(x).
(18)
χ is a subset of Rd, which includes the hyperparameters of the ML algorithm. Whereas h(x) is the objective function that needs to be minimized. Therefore, in this invention, the BO estimates the optimal hyperparameter values of base learners (BRT, NNR, and GAM).
[0054] FIG. 4 illustrates the overall methodology for CO and NOx prediction using SEM. Initially, the primary dataset is partitioned into training and testing sub-datasets. In the next step, the training dataset is given to the first layer of the SEM model to train each base learner using the 5-fold cross-validation method. Each base learner will generate the forecasting outcomes. The anticipated outcomes of each base learner are aggregated into a fresh data set, which is then used as input to train the meta-learner. This process yields the final set of predicted outputs. Once the model is built, it will be tested on the testing subset. Furthermore, the proposed SEM has a complex structure and will take more training and validation time than the single ML-based predictive model. The proposed SEM is computationally costly compared to a single ML technique. However, there is always a trade-off between computational time and accuracy for SEM. The strength of SEM lies in combining the different models, which improves the overall accuracy but increases the computational complexity. The focus of this study is to design an ensemble model that predicts CO and NOx accurately.
[0055] Furthermore, the performance of the trained SEM is then estimated on the testing subset using performance indices like MSE, root mean squared error (RMSE), mean absolute error (MAE), and correlation coefficient R, respectively. Let us consider the actual output is Yactual and predicted by the machine learning algorithm is YPred; then the performance indices are as follows.
(19)
(20)
(21)
[0056] where N is the total number of samples in the dataset.
[0057] FIG. 5 illustrates an exemplary graphical representation of the hydrographs for all the designed ML models for NOx, in accordance with an embodiment of the present disclosure.
[0058] A gas turbine employed in a CCPP produces toxic pollutants, including carbon CO and NOx, which substantially impact atmospheric contamination. Monitoring the emissions of this pollutant is of utmost importance to ensure adherence to prescribed limits. Moreover, analyzing the trends and patterns of these emissions can provide helpful information that may be utilized to improve operational strategies and eventually mitigate the release of those pollutants. Therefore, a stacked ensemble, constructed using NNR, GAM, and BRT as base learners and GRNN as a meta-learner, is proposed to monitor pollutants. At the same time, Bayesian optimization is used to optimize the hyperparameter values of SEM for NOx and CO emission prediction. In the subsequent sections, the result analysis of the NOx and CO prediction is discussed.
[0059] Table 2 shows the predictive performance of SEM compared with the LIR, DRT, SVR, NNR, BAG, and GAM. The correlation coefficient achieved by SEM is 0.930, which is higher than LIR (0.698), DRT (0.877), SVR (0.914), NNR (0.916), GAM (0.898), and BAG (0.920), respectively. It shows that the predictive values of SEM are highly correlated to the actual values of NOX. It is evident that the SEM approach achieves the lowest mean absolute error (MAE) value of 2.575 and mean squared error (MSE) value of 14.70 compared to the other machine learning algorithms implemented. Additionally, SEM reduces the RMSE (3.834) by 48.14%, 26%, 8.07%, 7.74%, 15.75%, and 5.40% compared to LIR, DRT, SVR, NNR, BAG, and GAM. Figure 5 shows the hydrographs for all the designed ML models. Furthermore, Bayesian optimization is used to estimate the hyperparameter values, i.e., maximum splits =9110, minimum leaf size = 48, neuron in the hidden layer =161, learning rate = 0.0025314, initial learn rate for predictor =0.0084, number of trees for predictor =183, initial learning rate for interaction =0.1515, number of trees for interaction =35, and spread =0.134 of SEM for the prediction of NOx emission from GT.

[0060] FIG. 6 illustrates an exemplary graphical representation of the Hydrographs for all the designed ML models, in accordance with an embodiment of the present disclosure.
[0061] The estimated optimal hyperparameter values are maximum splits =12011, minimum leaf size = 1, a neuron in the hidden layer =219, learning rate = 8.1709e-09, initial learning rate for predictor =0.00167, number of trees for predictor =10, initial learning rate for interaction =0.0849, number of trees for interaction =20, and spread = 0.1295 of SEM for the prediction of CO emission. Figure 6 illustrates the prediction capabilities of all the ML models that have been developed. In the case of CO, the proposed SEM has demonstrated higher accuracy in predicting values that closely align with the actual measurements.
[0062] FIG. 7 illustrates an exemplary graphical representation of the scatter plots for SEM predictive model, in accordance with an embodiment of the present disclosure.
[0063] Furthermore, the quantitative performance of the proposed SEM compared to other ML models is given in Table 3. The RMSE value attained by SEM is 0.614, followed by BAG (0.620), NNR (0.684), GAM (0.714), DRT (0.741), SVR (0.875), and LIR (869), respectively. The SEM model exhibited lower values of MAE (0.417), MSE (0.377), and R (0.880) in comparison to the other machine learning models that were developed. Figure 7 displays scatter plots depicting the relationship between the projected and actual values in both scenarios. These scatter plots indicate the proximity of the predicted values to the proposed structural equation model (SEM).

[0064] FIG. 8 illustrates an exemplary graphical representation of the sensitivity analysis of SEM for NOx and CO, in accordance with an embodiment of the present disclosure.
[0065] In an embodiment, the dependability analysis is also provided to assess the anticipated output's reliance on the input. The mathematical representation denoting the level of dependence of the output variable (yi) on the input variables (xi) is as follows.
(22)
[0066] where, M is the total number of NOx and CO prediction testing samples. Figure 8 shows the graphical representation of the DA values calculated w.r.t individual inputs. It is observed that, AP, AH, ADP, GEP, ITT, ATT, and CDP are the most sensitive input variables for NOx prediction in the same way AP, AH, ITT, and ATT is the most influential input w.r.t CO.
[0067] FIG. 9 illustrates an exemplary graphical representation of the error deviation comparison of all the predictive model for NOx and CO, in accordance with an embodiment of the present disclosure.
[0068] The error deviation plots are displayed in Figure 9, illustrating the variances between the predicted and actual emission rates of carbon monoxide and nitrogen oxides for each of the predictive machine-learning models constructed. The least and greatest error values that the predictive models achieved are presented in Table 4. The minimum and maximum error values for the proposed SEM are -26.770 and 24.666 for NOx and -3.603 and 4.299 for CO, respectively. These values are for the two different types of emissions. As a result, it is possible to conclude that, compared to other models, the SEM has the lowest error deviation. This conclusion may be reached by visually inspecting Figure 9 and Table 4.

[0069] In an exemplary embodiment, the exploratory data set has 36,733 samples recorded hourly from 2011 to 2015. Using machine learning algorithms, this dataset is utilized to develop an emission prediction system for CO and NOx from the GT of CCPP. Previous studies suggested that there is no thumb rule for how to divide the dataset into training and testing subsets. In this study, the entire sample has been considered, out of which 85% is for training and 5-fold cross-validation, as the proposed ML algorithm will perform both steps simultaneously, and 15% is for testing the developed ML predictive model. Furthermore, a staking ensemble technique (SEM) is designed and implemented using the training and testing subsets. The base learner of the proposed SEM is trained via 5-fold cross-validation. Base learning units will generate a new training dataset identical in size to the original training dataset. Then a meta-learner (GRNN) in the second layer is trained using a new dataset. Eventually, the stacking model generates a prediction for the newly created test dataset.
[0070] SEM leverages various base learners, which helps capture the dataset's different patterns. It will allow the insertion and exclusion of the models depending on the change in the dataset. It will predict the output by giving weights to the predictions made by the models used to construct SEM hierarchically. The other property of SEM is that it reduces the problem of overfitting the data by combining the output of several models. It will also improve the generalization of the overall predictive model compared to single-ML algorithm-based models. Finally, SEM is more robust at handling noisy data and outliers. It can also be seen from the result that in the case of CO and NOx, the prediction accuracy of the SEM is better than other ML-based models. There is an improvement in MAE for NOx (2.575) and CO (0.417) respectively.
[0071] SEM is the best predictive model for CO and NOx, but the problem is still associated with predicting values. Figures 7 (a) and (b) reveal a slight deviation between the predicted and target values from the actual fit line y=x, indicating an overestimation at the lower level and an underestimation at the higher level. FUG. 7 (b) shows that SEM makes significant numbers of outlying predictions. It underestimates values above 8 mg/m3 and overestimates large values below 4 mg/m3. There is scope for improvement in the prediction accuracy of the SEM for CO and NOx, but a detailed analysis of the model is required to find out the reason for distorting the predictive values. It can be ascertained that there must be other influential factors apart from AT, AP, AH, ADP, GEP, ITT, ATT, and CDP. Estimating the other factors from the dataset is impossible as it is taken from an anonymous CCPP plant. Generically, there are certain factors, such as degradation of the compressors and air filters, changes in the inlet air quality, periodic maintenance, and changing the position of the detectors about the turbine exhaust, may distort sample values.
[0072] In an implementation of an embodiment, NOx Result Analysis is performed. Bayesian optimization is used to estimate the hyperparameter values, i.e., maximum splits =9110, minimum leaf size = 48, neuron in the hidden layer =161, learning rate = 0.0025314, initial learn rate for predictor =0.0084, number of tree for predictor =183, initial learning rate for interaction =0.1515, number of trees for interaction =35, and spread =0.134 of SEM for the prediction of NOx emission from GT.
[0073] In an implementation of an embodiment, CO Result Analysis is performed. The estimated optimal hyperparameter values are maximum splits =12011, minimum leaf size = 1, a neuron in the hidden layer =219, learning rate = 8.1709e-09, initial learning rate for predictor =0.00167, number of trees for predictor =10, initial learning rate for interaction =0.0849, number of trees for interaction =20, and spread = 0.1295 of SEM for the prediction of CO emission.
[0074] FIG. 10 illustrates an exemplary view of a flow diagram of the proposed method for monitoring of emissions from gas turbines, in accordance with an embodiment of the present disclosure.
[0075] In an embodiment, the proposed method 1000 for monitoring of emissions from gas turbines is disclosed. At step 1002, collecting, by the system 102, a dataset of operational and environmental data from a plurality of sensors associated with gas turbines, which act as input attributes to a stacking ensemble model (SEM).
[0076] At step 1004, conducting, by the system 102, data preprocessing and Pearson correlation analysis on the collected dataset to train and test the SEM.
[0077] At step 1006, combining, by the system 102, the generated emission predictions from the plurality of SEMs to generate a final emission forecast with improved emission prediction accuracy via the combining engine.
[0078] At step 1008, comparing, by the system 102, the final emission forecast to predefined regulatory thresholds via the comparing engine.
[0079] At step 1010, triggering, by the system 102, an alert if the predicted emissions exceed the regulatory limits via the triggering engine.
[0080] At step 1012, analysing, by the system 102, between the predicted emissions and actual emissions to perform anomaly detection and provide recommendations to a plurality of users via the analysing engine.
[0081] While the foregoing describes various embodiments of the invention, other and further embodiments of the invention may be devised without departing from the basic scope thereof. The scope of the invention is determined by the claims that follow. The invention is not limited to the described embodiments, versions or examples, which are comprised to enable a person having ordinary skill in the art to make and use the invention when combined with information and knowledge available to the person having ordinary skill in the art.
, Claims:1. A system (102) for monitoring of emissions from gas turbines, the system (102) comprising:
a processor operatively coupled to a memory, wherein the memory comprises processor-executable instructions, which on execution, causes the processor to:
collect a dataset of operational and environmental data from a plurality of sensors associated with gas turbines, which act as input attributes to a stacking ensemble model (SEM);
conduct data preprocessing and Pearson correlation analysis on the collected dataset to train and test the SEM;
combine the generated emission predictions from the SEM to generate a final emission forecast with improved emission prediction accuracy;
compare the final emission forecast to predefined regulatory thresholds;
trigger an alert if the predicted emissions exceed the regulatory limits; and
analyze deviations between the predicted emissions and actual emissions to perform anomaly detection and provide recommendations to a plurality of users.
2. The system (102) as claimed in claim 1, wherein the operational and environmental data are selected from a group comprising ambient temperature (AT), ambient pressure (AP), ambient humidity (AH), air filter difference pressure (ADP), compressor discharge pressure (CDP), gas turbine exhaust pressure (GEP), turbine after temperature (ATT), carbon monoxide (CO), nitrogen oxide (NOx), turbine inlet temperature (ITT) and any combination thereof.
3. The system (102) as claimed in claim 2, wherein the emission of CO increases with a decrease in the ITT and the CDP.
4. The system (102) as claimed in claim 1, wherein the SEM is selected from a group comprising Bagged Regression Trees (BRT), Generalized Additive Models (GAM), Neural Network for Regression (NNR), Generalized Regression Neural Networks (GRNN) and the like.
5. The system (102) as claimed in claim 4, wherein the NNR, GAM and BRT are configured to act as base learners and the GRNN is configured to act as mesh learner which are proposed for the prediction of CO and NOx.
6. The system (102) as claimed in claim 1, wherein the system (102) is configured to attain a minimum value of prediction error.
7. The system (102) as claimed in claim 1, wherein hyperparameters of SEM are optimized using a Bayesian optimization instruction model for CO and NOX prediction.
8. The system (102) as claimed in claim 1, wherein the sensors are configured to monitor parameters including but not limited to combustion temperature, fuel flow rate, gas flow rate, turbine speed, pressure levels, and pollutant concentrations such as NOx, CO, unburned hydrocarbons and any combination thereof.
9. The system (102) as claimed in claim 1, wherein the recommendations pertain to adjusting operations of the gas turbine to reduce emissions below the predefined thresholds.
10. A method (1000) for monitoring of emissions from gas turbines, the method (1000) comprising steps of:
collecting (1002), by a system (102), a dataset of operational and environmental data from a plurality of sensors associated with gas turbines, which act as input attributes to a stacking ensemble model (SEM);
conducting (1004), by the system (102), data preprocessing and Pearson correlation analysis on the collected dataset to train and test the SEM;
combining (1006), by the system (102), the generated emission predictions from the SEM to generate a final emission forecast with improved emission prediction accuracy;
comparing (1008), by the system (102), the final emission forecast to predefined regulatory thresholds;
triggering (1010), by the system (102), an alert if the predicted emissions exceed the regulatory limits; and
analyzing (1012), by the system (102) deviations between the predicted emissions and actual emissions to perform anomaly detection and provide recommendations to a plurality of users.

Documents