Prediction of cost and emission from Indian coal-fired power plants with CO₂ capture and storage using artificial intelligence techniques

Abstract

Coal-fired power plants are one of the most important targets with respect to reduction of CO₂ emissions. The reasons for this are that coal-fired power plants offer localized large point sources (LPS) of CO₂ and that the Indian power sector contributes to roughly half of all-India CO₂ emissions. CO₂ capture and storage (CCS) can be implemented in these power plants for long-term decarbonisation of the Indian economy. In this paper, two artificial intelligence (AI) techniques—adaptive network based fuzzy inference system (ANFIS) and multi gene genetic programming (MGGP) are used to model Indian coal-fired power plants with CO₂ capture. The data set of 75 power plants take the plant size, the capture type, the load and the CO₂ emission as the input and the COE and annual CO₂ emissions as the output. It is found that MGGP is more suited to these applications with an R² value of more than 99% between the predicted and actual values, as against the ~96% correlation for the ANFIS approach. MGGP also gives the traditionally expected results in sensitivity analysis, which ANFIS fails to give. Several other parameters in the base plant and CO₂ capture unit may be included in similar studies to give a more accurate result. This is because MGGP gives a better perspective toward qualitative data, such as capture type, as compared to ANFIS.

Keywords

carbon capture and storage / power plants / artificial intelligence / genetic programming / neuro fuzzy

Introduction

India is a rising developing economy with a gross domestic product (GDP) of more than US$ 2 trillion and a population above 1.29 billion. India’s GDP is expected to increase to US$ 12 to 80 trillion by the end of this century, as predicted by various models. Besides, it is likely for India to become the most populous nation by 2050 [¹]. Therefore, it is expected that India will play a major role in the global decarbonisation process, i.e., in the reduction of the CO₂ emission intensity. The Government of India has expressed its commitment for reduction of the GDP intensity by 33% to 35% by 2035, with respect to the 2005 levels. For a reduction in the greenhouse gas (GHG) intensity levels by such a magnitude, the power sector, which contributes to around 50% of all-India CO₂ emissions [²] will have to play a major role. In addition, the reduction of CO₂ emissions can happen in an easily manageable way in the power sector, which has large point sources (LPSs), in contrast to other sectors with dispersed emission sources, e.g., transport or agriculture sectors.

Recently, steps have been taken to reduce the CO₂ emission from Indian power plants. The newer units that are being commissioned in the country, including those in the ultra-mega power plants of 4000 MW capacity, are all of super-critical type. Steps are also being taken toward power generation through the gasification route [³]. The use of imported/blended coal, if done in a strategic manner, may also contribute to a reduction in the emission (kg-CO₂/kWh_e) from Indian power plants. However, for more radical drops in emissions, the CO₂ capture and storage (CCS) technology needs to be utilized. This technology involves the capture of CO₂ from LPSs, its compression and subsequence transport using pipelines to geological storage reservoirs, like saline aquifers, basalt formations, unmineable coal seams or depleted hydrocarbon reserves. Readers who wish to have a preliminary understanding of the underlying techniques, present status and economic implications of CCS may refer to references [⁴–⁷].

Shukla et al. [⁸] have projected that in the light of decarbonisation policies being put into place, the CO₂ intensity of electricity is expected to decline by more than 10 times between 2010 and 2050. They have assumed two scenarios for decarbonisation, viz., the conventional pathway (having dominance of CCS and nuclear energy) and sustainable pathway (having predominance of renewable energy). They have projected that the amount of CO₂ mitigated by 2050 through CCS will be around 10000 Mt in the conventional pathway and around 7000 Mt in the sustainable pathway. Thus, CCS will play an important role in the meeting of India’s GHG reduction targets, with the predicted total annual emissions of CO₂ of just over 2100 Mt in the sustainable scenario and over 3100 Mt in the conventional scenario in 2050.

Pragmatic planning and understanding of the effects of CCS on Indian coal-fired power plants is of utmost importance for overall planning of the energy sector. This paper attempts to understand the effect of retrofitting CO₂ capture systems on the current fleet of Indian coal-based power plants, using AI techniques.

In this paper, an attempt has been made to estimate the cost of electricity (COE) and overall emissions of the current fleet of Indian coal-fired power plants using two AI techniques, viz., adaptive network based fuzzy inference system (ANFIS) and multi gene genetic programming (MGGP). The parameters that have been studied are the plant size, the utilization factor of the plant, the current emission factor, and the type of capture mechanism used. These parameters have been selected as they have been known to have significant impact on the capture dynamics of the plant, while other parameters also exist, which are generally studied in for sensitivity studies [⁹–¹¹]. A larger plant size means larger CO₂ emissions and, therefore, a significant LPS for CO₂ abatement. The plant load or utilization factor has a significant impact on capture results. A higher load results in a lower COE but a higher emission due to more operation of the plant. The effect of load on COE is discussed in detail in Section 3. Capture technologies also affect the CO₂ avoidance cost (as also discussed in Section 3). Two types of capture technologies, advanced amine based capture and membrane based CO₂ capture techniques, have been studied. Finally, a higher CO₂ emission results in more emissions and a lower cost for a CCS system, both with undesirable outcomes. Thus, the selected parameters result in a complex outcome for the plant and an optimized solution needs to be found out for individual plants. In other words, there is no such solution as “one size fits all.” Therefore, the modeling using AI techniques may be utilized. The better modeling technique for application in these systems is identified and, thereby, sensitivity analysis has been done to identify the dominant parameter for cost and emission in CCS based power plants.

It should be noted here that this is a first of a kind analysis for a major chunk of Indian power plants. However, many more parameters have an important effect on the overall results of costs and emissions for such power plants. Thus, while more than 70 power plants have been analyzed, to get an overall view on the feasibility of CCS adoption in these plants, such a study can be further refined in future works, to study individual plants, to understand the effects of more operational and financial parameters toward CCS deployment. It is hoped that this paper will prove to be useful to policy makers and the industry for better planning toward CCS in India.

AI methods

ANFIS

ANFIS is a tool for constructing a set of fuzzy if-then rules using suitable membership functions to generate the specified input-output pairs. The technique combines the features of adaptive neural networks (ANN) and fuzzy inference systems (FIS) and it belongs to the family of adaptive networks which are functionally corresponding to the fuzzy inference system [¹²]. The method creates a FIS where the membership function parameters are adjusted using suitable algorithms. This incorporates within itself both the learning abilities of the ANN and the decision making abilities of the FIS to improve the system performance and avoid the need for manual optimization of the fuzzy logic based parameters with the help of the tuning up characteristic of the ANN. The logical pattern of the prediction in the results is achieved by creating a set of fuzzy conditional statements (if-then rules) backed up with the membership functions to generate the desired results.

Due to the simplicity of the requirement of lesser rules and parameter estimation from the data set using optimization techniques, the rule-based model is often chosen as the Takagi-Sugeno-Kang (TSK) type [¹³,¹⁴]. The output is generated as a zero or first degree polynomial as a linear combination of the input variables. The ANFIS architecture can be illustrated in Fig. 1, by taking into consideration a FIS model having two inputs (x and y) and one output (f). For a first order TSK model fuzzy inference system consisting of two if-then rules [¹⁵],

\mathrm{Rule}\hspace{0.17em}\mathrm{1}:\mathrm{If}\hspace{0.17em}x\hspace{0.17em}\mathrm{is}{{\displaystyle A}}_{\mathrm{1}}\hspace{0.17em}\mathrm{and}\hspace{0.17em}y\hspace{0.17em}\mathrm{is}{{\displaystyle B}}_{\mathrm{1}},\mathrm{then}{{\displaystyle f}}_{\mathrm{1}}=p_{\mathrm{1}}x+q_{\mathrm{1}}y+{{\displaystyle r}}_{\mathrm{1}}.

\mathrm{Rule}\hspace{0.17em}\mathrm{2}:\mathrm{If}\hspace{0.17em}x\hspace{0.17em}\mathrm{is}{{\displaystyle A}}_{\mathrm{2}}\hspace{0.17em}\mathrm{and}\hspace{0.17em}y\hspace{0.17em}\mathrm{is}{{\displaystyle B}}_{\mathrm{2}},\mathrm{then}{{\displaystyle f}}_{\mathrm{2}}=p_{\mathrm{2}}x+q_{\mathrm{2}}y+{{\displaystyle r}}_{\mathrm{2}}.

The node functions in a particular layer belong to the function family of the same layer as discussed below.

Layer 1: Every node in this layer is a square node with a node function

{{\displaystyle O}}_i^{\mathrm{1}}={{\displaystyle \mu }}_{Ai}(x),

where x is the input to node i, {{\displaystyle O}}_i^{\mathrm{1}} is a membership grade of a fuzzy set A_i, which specifies the degree to which the input x satisfies the quantifier A_i, and μ_Ai(x) is the Gaussian membership function given by

{{\displaystyle \mu }}_{Ai}(x)=\mathrm{e}\mathrm{x}\mathrm{p}\left[-{{\displaystyle \left({\displaystyle \frac{x-{{\displaystyle c}}_i}{{{\displaystyle a}}_i}}\right)}}^{\mathrm{2}}\right],

where a_i and c_i are referred to as the premise parameter set.

Layer 2: Every node in this layer is a circle node labeled P which generates the output as the product of the incoming signals, representing the firing strength of the rule.

{{\displaystyle w}}_i={{\displaystyle \mu }}_{Ai}(x)\times {{\displaystyle \mu }}_{Bi}(x),\hspace{0.17em}i=\mathrm{1},\mathrm{2.}

Layer 3: The node in this layer is a circle denoted as N. The ith node gives the ratio of the firing strengths of the ith rule to that of the sum of all rules, called normalized firing strengths.

{{\displaystyle \tilde{w}}}_i={\displaystyle \frac{{{\displaystyle w}}_i}{{{\displaystyle w}}_{\mathrm{1}}+{{\displaystyle w}}_{\mathrm{2}}}},\hspace{0.17em}i=\mathrm{1},\mathrm{2.}

Layer 4: The node i is a square node in this layer with a node function of

{{\displaystyle O}}_i^{\mathrm{4}}={{\displaystyle \tilde{w}}}_i{{\displaystyle f}}_i={{\displaystyle \tilde{w}}}_i(p_{i}x+q_{i}y+r_i),

where

{{\displaystyle \tilde{w}}}_i

is the output for Layer 3 and \{p_{i}x+q_{i}y+r_i\} is the consequent parameter set.

Layer 5: This layer contains a single node labeled as S that calculates the summation of the input variables.

{{\displaystyle O}}_i^{\mathrm{5}}=final\hspace{0.17em}output=\underset{i}{{\displaystyle \Sigma }}{{\displaystyle \tilde{w}}}_i{{\displaystyle f}}_i={\displaystyle \frac{\underset{i}{{\displaystyle \Sigma }}{{\displaystyle \tilde{w}}}_i{{\displaystyle f}}_i}{\underset{i}{{\displaystyle \Sigma }}{{\displaystyle \tilde{w}}}_i}}.

The ANFIS structure generates an output f as the linear combination of consequent parameters for a given set of premise parameters. Hence, the back propagation algorithm is often replaced with the hybrid learning algorithm which combines the gradient descent and least squares method. This reduces the complexity and increases the learning efficiency by learning the weights and nonlinear membership functions in two independent stages.

MGGP

The MGGP algorithm is a modification to the standard genetic programming (GP) algorithm. GP is a biologically inspired machine learning technique that optimizes the structure of the population of computer programs (traditionally referred as tree structures) to perform computational tasks [¹⁷]. This is achieved by randomly generating the trees and reproducing the best performing trees together to create a new population, until a set of programs are obtained that perform the task satisfactorily, as shown in Fig. 2. GP is a specialized version of genetic algorithms (GA), which optimizes the parameters and develops solutions in the form of strings of fixed lengths, as opposed to tree structures of varying sizes in the former [¹⁸]. GP is primarily associated with symbolic regression modeling which automatically evolves the structure and parameters of the model from the data. The primary node of the tree is the operator functions like Boolean operators, arithmetic operators (+ , –, × and ÷ ) and nonlinear operators (also referred to as the function set) and the terminal nodes as operands or simply input process variables (referred to as the terminal set).

MGGP evolves multiple trees (genes) based on multigene symbolic regression. This facilitates the crossover and mutation of different multigene individuals. The number of generations to be run is specified by the user which acts as termination criteria for the process. In addition to this, the performance of the initial population is checked using various fitness functions on the training data. The most commonly used is the root mean square error (RMSE), given by

\mathrm{RMSE=}\sqrt{{\displaystyle \frac{\underset{i=\mathrm{1}}{\overset{N}{{\displaystyle \Sigma }}}{\left|{{\displaystyle G}}_i-{{\displaystyle A}}_i\right|}^{\mathrm{2}}}{N}}},

where G_i is the estimated data for the ith data generated using MGGP, A_i is the actual data from the samples, and N is the number of training samples.

The complexity of the algorithm is measured using node count, constituent trees and total gene count. The selection method used is generally the tournament selection. In this paper, the Pareto tournament selection method is chosen which considers the model complexity along with the fitness over the regular one where only the model fitness is taken into consideration. The crossover operators used are of two types, high level crossover, and GP sub-tree or low level crossover. The low level crossover selects a gene randomly from each parent and the offspring population completely replaces the original parent genes. While, the high level method is selective in the sense that the new population may acquire whole parent genes or have them deleted, providing a chance for exchange of genes [²⁰]. The mutation in the subsequent generation is brought about by sub-tree mutation, use of Gaussian function for constants, substitution and replacement of randomly selected nodes, assigning randomly selected constant to zero, assigning randomly selected constant to one, and substituting randomly selected constant with another randomly generated constant [¹⁸].

The complexity of MGGP involving multiple genes and nonlinearity of the terms may cause over-fitting of the model. Therefore, a restriction on the number of genes and maximum tree depth is imposed to gain control over the model producing compact and accurate results.

Methodology

The methodology for this study involves three steps. Initially, the base data for Indian coal-fired power plants are obtained from the government. Subsequently, additions are made to this data set using the Integrated Environmental Control Model (IECM-cs), developed at the Carnegie Mellon University, USA, to estimate the cost of electricity (COE) in these plants with and without CCS. Lastly, intelligent techniques are applied to the cost and emission data of these plants for prediction. The analytical framework for the study is presented in Fig. 3.

Basic repository of Indian power plants

Initially, the basic data set of selected coal-fired power plants in India have been made with the initial input from the CO₂ Baseline Database for the Indian Power Sector, made available by the CEA [²¹]. This database helps in summarizing the utilization factor and the emission associated with each power plant in the country. Out of this, some power plants which are unsuitable for CCS deployment (owing to low utilization or very high specific emissions) are removed from the database. Thus, 75 power stations across the country are selected, which is listed in Table 1 (source: Ref. [²¹]).

Data set preparation

The database of Indian power plants provided by CEA does not provide any information on the COE or the suitability of CCS in those plants. Due to the inadequacy of the available data to understand the impacts of CCS on power plants, a data set is prepared using the integrated environmental control model version 8.0.2 (IECM-cs), developed at the Carnegie Mellon University, USA [²²]. This model is a multi-parameter, multi-output software framework. The IECM-cs is used to understand the effect of load on COE, and the effect of CO₂ emission on the additional cost incurred and energy penalty (EP) due to the CO₂ capture process. This will help in understanding the implications of CO₂ capture process in this plant.

As can be seen in Table 1, two of the parameters available are plant size and annual electricity generation. These are two interlinked parameters, i.e., the annual generation depends upon plant size and plant load. Therefore, the load is calculated from the annual generation and net size using

\mathrm{Load}(\%)={\displaystyle \frac{\mathrm{1000}\times \mathrm{Net}\hspace{0.17em}\mathrm{Generation}\hspace{0.17em}(\mathrm{GWh})}{\mathrm{Hours}\hspace{0.17em}\mathrm{in}\hspace{0.17em}1\hspace{0.17em}\mathrm{year}(\mathrm{8760})\times \mathrm{Plant}\hspace{0.17em}\mathrm{Size}\hspace{0.17em}(\mathrm{MW})}}.

The calculation of the load gives an approximation of the COE generation. As explained in Sub-subsection 3.2.1, a larger load leads to a faster amortization of the initial capital cost invested in the plant. In a grid-connected power supply, it is assumed in this paper that the power cost remains the same irrespective of the plant size. In such a case, load becomes a more appropriate parameter for COE estimation. Further, the current fleet of power plants does not have the CO₂ capture technology, and therefore, the primary data set does not contain any information on the effects of CO₂ capture. Thus, modification and extension of the original data set is necessary for understanding the effects of CCS retrofitting in the existing power plants. Therefore, the effect of load and CO₂ capture type need to be understood to evaluate CCS effects on the plants.

Effect of load

Now, each of the plants is assigned a (COE), which is a result of its utilization factor, for a plant without CCS. A higher utilization means a lower electricity cost and vice-versa. The typical COE for a sub-critical plant operating at a utilization rate of 90% is taken as US$ 55.89/MWh [²³]. The variation of COE with load, following the assumptions made by Singh and Rao [²³], is shown graphically in Fig. 4.

Effect of capture type

Two types of post-combustion capture processes, the amine based one and the membrane based one, are studied. The uncertainty analysis on the variation of the cost of CO₂ avoidance vis-à-vis CO₂ capture efficiency (Fig. 5) shows that the membrane based CO₂ capture is most cost effective in the CO₂ capture efficiency of 50% to 60%, while the amine based capture is cost effective in the CO₂ capture efficiency of 85% to 90%. The latter has also been reported by Rao and Rubin [²⁴]. Thus, for this study, the membrane CO₂ capture efficiency is modeled as 50%, while the efficiency of the amine based capture is adjusted to 90%.

Now, the variation of COE due to the capture process is modeled as the cost multiplier (a). This value is the ratio of the COE of a plant with CCS to that operating without CCS and has been used in a previous study using fuzzy systems in CCS [²⁵]. Further, the EP is also calculated using [¹⁰]

\mathrm{EP}={\displaystyle \frac{{{\displaystyle \eta }}_{\mathrm{ref}}}{{{\displaystyle \eta }}_{\mathrm{CCS}}}}-\mathrm{1},

where h is net plant efficiency while the subscripts ref and CCS refer to the reference plant and the plant with CCS, respectively. Figure 6 illustrates the effect of variation of α and EP, with change in CO₂ emission of the reference plant. It should be noted that the capture efficiency assumed for the membrane based capture is 50%, while that for the amine based capture is 90% and, therefore, the cost multipliers and EPs mentioned cannot be compared directly.

A set of 50 data points from the final database is included as listed in Table 2 where the plant size has been obtained from the CEA database, as explained in Sub-section 3.1. The annual net output of the plant helps in calculation of the plant load. Using the load, the probable COE for the reference plants is estimated. The annual CO₂ emissions in the reference plants are readily available from the CEA database itself. Further, using the procedure explained in Sub-subsection 3.2.2, the COE and the annual CO₂ emissions for the amine and the membrane based capture in the power plants are assessed.

Implementing AI techniques

ANFIS approach

The ANFIS approach is implemented on the data set using the Neuro-Fuzzy toolbox of Matlab 2014b software. Approximately 75% sets of data are selected as the testing data and 25% are selected as the training data. By trial and error, the best results (minimum errors) are obtained by generating FIS with grid partition using input membership function type as Gaussian distribution and output membership function type as linear distribution. The hybrid optimization method is selected. Four membership functions are each selected for all the four input parameters.

MGGP approach

The MGGP approach is implemented using the GPTIPS 2.0 toolbox of Matlab [²⁰]. About one-third of the data set is used as the training data and the rest is used as the testing data. The parameters used for MGGP, based on trial and error, are tabulated in Table 3.

It should be noted that the intelligent techniques stated above take only numerical input and output. Thus, capture type, which is a qualitative parameter, is input as ‘0’ for no capture, ‘1’ for the membrane based capture, and ‘2’ for the amine based capture.

Results and discussion

The performance of the ANFIS and MGGP approach are compared in order to find out a better method for predicting the cost and emission outputs for power plants utilizing the CCS technology.

Cost prediction

For the ANFIS approach, the average training error is obtained as $0.40/MWh for COE and the average testing error as US$ 8.56/MWh. This is a reasonably accurate result as the COE for most power plants in this study is considered to be well above US$ 60/MWh and the average COE of all the plants (with and without capture) is US$ 117.14. Thus, the average testing error amounts to approximately 7.3%, which is acceptable. This is also because many other bias and sensitivities will play a role in further deviation of the electricity costs. For MGGP, the results are somewhat closer to the actual expected results. The results are predicted correctly for a coefficient of determination of more than 99.8%. The average RMSE testing error is only US$ 2.62/MWh, which is significantly less than the results achieved with the ANFIS approach. Figure 7 demonstrates the correlation of the predicted and actual value for the ANFIS and MGGP approaches for COE. The y-axis values (50–250 and 0–350 in Fig. 7(a) and (b) respectively) denote the COE in $/MWh. Figure 8 depicts the surface plot showing the variation in costs, in the backdrop of changing parameters, predicted by neuro-fuzzy logic.

It should be noted that the trends developed using ANFIS have some functional inabilities to predict the COE. This has been discussed in details in Subsection 5.3, where sensitivity analysis has been used to establish the fact that MGGP is a better technique. For instance, Fig. 7 (a) predicts that for no CO₂ capture, the COE increases with increasing plant size. However, this is contrary to common knowledge. Similarly, Fig. 7(b) and 7(c) indicates that the membrane based CO₂ capture results in a lesser COE than a plant with no capture. Again, this is a wrong prediction of the model. Figure 7(f) makes another wrong prediction, i.e., the COE for a utilization factor of 60% is higher than that for 20%, which is incorrect since the COE should decrease monotonically with increasing utilization factor. Thus, while the ANFIS based model does result in a good R² value, Fig. 8 helps in understanding many infirmities of the model, the reasons of which are explained in Subsection 5.3.

Annual CO₂ emission prediction

Similarly, for annual CO₂ emissions of the plant, the predicted values by the ANFIS model lie within 6% of the results in the basic data set, which is accurate enough for prediction purposes. For annual CO₂ emissions as well, the average error is less than 10% of that obtained with the ANFIS approach. Figure 9 displays the correlation of the predicted and actual values for the ANFIS and MGGP approaches, respectively. The surface plots prepared using ANFIS for CO₂ emissions are presented in Fig. 10. For Fig. 10(a), the prediction of increasing CO₂ emissions with plant size is correct. But similar trend is not exhibited by Fig. 10(d). Similarly, the trend of increasing annual emissions with increasing utilization factor is correctly shown in Fig. 10(b), but contradicted in Fig. 10(d). Thus, the surface plots (Figs. 8 and 10) are used here to show the fundamental lack of modeling capability for this application.

Evaluation of a better technique

The training and testing using ANFIS and MGGP for CO₂ capture evaluation have been performed in Indian coal-fired power plants. However, there is a significant variation in individual parameters of the plants. Also, some other analysis can be conducted. For instance, the COE have been considered to be independent of the plant size. However, whether by adapting to the data, the fact that the systems naturally generate some variation against plant size is an interesting question. For such questions, it will be prudent to perform sensitivity analysis to understand the dominant parameters while accounting for the final costs and emissions. A few cases of plants are analyzed wherein sensitivities of particular parameters have been studied.

As any considerations have not been placed on the variation of reference plant COE with plant size, it comes as a constant line. However, as shown in Fig. 11, the membrane and the amine based capture show arbitrary variation with the variation of plant size. As a result, the ANFIS model does not show any significant relation between plant size and the COE of the plant. Another fact that becomes evident is that the capture technology has a major impact on plant cost, as expected. Figures 12 and 13 show the variation in COE with the change in load factor and specific emissions using the ANFIS model. The COE should decrease with an increase in load factor. This is observed in the case of reference plant as well as a plant with the membrane based capture. However, the amine based capture plant does not show such a result, which shows a weakness of the ANFIS model for such applications. However, Fig. 13, which shows the variation of COE with emission factor, displays the results as expected.

Figures 14–16 show the sensitivity analysis results using the MGGP model. The results show an accurate depiction of the expected trends, as per previous studies. For instance, an increase in plant load factor shows a decrease in COE, regardless of the type of capture. Besides, the increase in plant size leads to a better load factor, which, in turn, results in a lower COE. Thus, it is again found that the MGGP technique is better for modeling CO₂ capture in power plants and estimating national trends, as compared to the neuro-fuzzy technique. The probable reason for this is the lack of understanding of qualitative parameters (capture type) by the ANFIS approach, which is understandable by the MGGP technique. As a result, more appropriate and accurate results are obtained using MGGP in this study.

Conclusions

In this paper, two AI techniques (neuro-fuzzy technique and MGGP) have been applied to modeling of power plants with CO₂ capture. The significant parameters such as plant size, capture type, load and CO₂ emission factor have been used to train and test the data set of 75 existing Indian coal-fired power plants. The plants are initially simulated using the integrated environmental control model (IECM) and then the results are compared to the predicted values using AI techniques. It is found that the predicted values using the ANFIS and MGGP techniques are reasonably close to the values estimated by IECM modeling (with R² values of 96.15% and 99.83% for COE). However, the MGGP model is better in predicting the cost and emission of the resulting plants with CO₂ capture. This is so not only because of the higher degree of correlation but also because the ANFIS model fails to replicate the expected sensitivity analysis results, which MGGP does successfully. A probable reason for this is that ANFIS is not capable enough of handling qualitative parameters such as capture type, which causes inaccurate results. Several other parameters such as uncertainties in the reference plant (boiler efficiency, turbine efficiency, steam cycle variations etc.) as well as in the CO₂ capture unit (capture efficiency, sorbent regeneration heat requirement, compressor use etc.) may be included to further model these plants and predict the values using AI techniques, especially MGGP. Such analysis can prove to be useful over other conventional analyses as they can simultaneously be used to model several plants and units at the regional or national scale. Besides, it can reduce the modeling/simulation time requirement and can be used in cases where some tolerance is acceptable.

References

[1] Sathaye J, Shukla P R. Methods and models for costing carbon mitigation. Annual Review of Environment and Resources, 2013, 38(1): 137–168

[2] Sharma S K, Choudhury A, Sarkar P, Biswas S, Singh A, Dadhich P K, Singh A K, Majumdar S, Bhatia A, Mohini M, Kumar R. Greenhouse gas inventory estimates for India. Current Science, 2011, 101(3): 405–415

[3] Goel M. Recent approaches in CO₂ fixation research in India and future perspectives towards zero emission coal based power generation. Current Science, 2009, 97(11): 1625–1633

[4] IPCC. Special Report on Carbon Dioxide Capture and Storage. Cambridge: Cambridge University Press, 2005,

[5] Haszeldine R S. Carbon capture and storage: how green can black be? Science, 2009, 325(5948): 1647–1652

[6] Singh U. Carbon capture and storage: an effective way to mitigate global warming. Current Science, 2013, 105(7): 914–922

[7] Rubin E S, Davison J E, Herzog H J. The cost of CO₂ capture and storage. International Journal of Greenhouse Gas Control, 2015, 40: 378–400

[8] Shukla P, Dhar S, Pathak M, Mahadevia D, Garg A. Pathways to deep decarbonization in India. SDSN-IDDRI, 2015,

[9] Rao A B, Rubin E S. A technical, economic, and environmental assessment of amine-based CO₂ capture technology for power plant greenhouse gas control. Environmental Science & Technology, 2002, 36(20): 4467–4475

[10] Rubin E S, Chen C, Rao A B. Cost and performance of fossil fuel power plants with CO₂ capture and storage. Energy Policy, 2007, 35(9): 4444–4454

[11] Singh U, Rao A B. Estimating the environmental implications of implementing carbon capture and storage in Indian coal power plants. In: Proceedings on Advances in Green Energy (ICAGE).IEEE, 2014, 226–232

[12] Singh R, Kainthola A, Singh T N. Estimation of elastic constant of rocks using an ANFIS approach. Applied Soft Computing, 2012, 12(1): 40–45

[13] Takagi T, Sugeno M. Fuzzy identification of systems and its applications to modeling and control. IEEE Transactions on Systems, Man, and Cybernetics, 1985, SMC-15 (1): 116–132

[14] Sugeno M, Kang G T. Structure identification of fuzzy model. Fuzzy Sets and Systems, 1988, 28(1): 15–33

[15] Jang J S. ANFIS: adaptive-network-based fuzzy inference system. IEEE Transactions on Systems, Man, and Cybernetics, 1993, 23(3): 665–685

[16] Sahu M, Singh P, Mahapatra S S, Khatua K K. Prediction of entrance length for low Reynolds number flow in pipe using neuro-fuzzy inference system. Expert Systems with Applications, 2012, 39(4): 4545–4557

[17] Koza J R. Genetic Programming: on the Programming of Computers by Means of Natural Selection. Boston: MIT Press, 1992

[18] Garg A, Vijayaraghavan V, Mahapatra S S, Tai K, Wong C H. Performance evaluation of microbial fuel cell by artificial intelligence methods. Expert Systems with Applications, 2014, 41(4): 1389–1399

[19] Garg A, Garg A, Tai K. A multi-gene genetic programming model for estimating stress-dependent soil water retention curves. Computational Geosciences, 2014, 18(1): 45–56

[20] Searson D P. GPTIPS 2: An open-source software platform for symbolic data mining. In: Gandomi A.H., Alavi A.H. and Ryan, C. (eds.), Handbook of Genetic Programming Applications. Springer, Switzerland. 2015

[21] CEA. CO₂ baseline database for the Indian power sector, Central Electricity Authority, Ministry of Power, Government of India, New Delhi. 2016,

[22] Berkenpas M B, Frey H C, Fry J J, Kalagnanam J, Rubin E S. Integrated environmental control model technical documentation. Center for Energy and Environmental Studies. Carnegie Mellon University, USA. 1999,

[23] Singh U, Rao A B. Integrating SO₂ and NO_x control systems in Indian coal-fired power plants. Decision, 2015, 42(2): 191–209

[24] Rao A B, Rubin E S. Identifying cost-effective CO₂ control levels for amine-based CO₂ capture systems. Industrial & Engineering Chemistry Research, 2006, 45(8): 2421–2429

[25] Tan R R, Ng D K, Foo D C, Aviso K B. Crisp and fuzzy integer programming models for optimal carbon sequestration retrofit in the power sector. Chemical Engineering Research & Design, 2010, 88(12): 1580–1588

RIGHTS & PERMISSIONS

Higher Education Press and Springer-Verlag Berlin Heidelberg