A Data-Driven Method to Monitor Carbon Dioxide Emissions of Coal-Fired Power Plants

Abstract

Reducing CO${}_2$ emissions from coal-fired power plants is an urgent global issue. Effective and precise monitoring of CO${}_2$ emissions is a prerequisite for optimizing electricity production processes and achieving such reductions. To obtain the high temporal resolution emissions status of power plants, a lot of research has been done. Currently, typical solutions are utilizing Continuous Emission Monitoring System (CEMS) to measure CO${}_2$ emissions. However, these methods are too expensive and complicated because they require the installation of a large number of devices and require periodic maintenance to obtain accurate measurements. According to this limitation, this paper attempts to provide a novel data-driven method using net power generation to achieve near-real-time monitoring. First, we study the key elements of CO${}_2$ emissions from coal-fired power plants (CFPPs) in depth and design a regression and physical variable model-based emission simulator. We then present Emission Estimation Network (EEN), a heterogeneous network-based deep learning model, to estimate CO${}_2$ emissions from CFPPs in near-real-time. We use artificial data generated by the simulator to train it and apply a few real-world datasets to complete the adaptation. The experimental results show that our proposal is a competitive approach that not only has accurate measurements but is also easy to implement.

1. Introduction

Global climate change is mainly caused by the emissions of greenhouse gases from anthropogenic activities [1]. In recent years, CO${}_2$ has become the most important greenhouse gas in the world [2]. Specifically, coal-fired power plants (CFPPs), the primary CO${}_2$ emission source, contributed roughly 40% of energy-related emissions from 2010 [3,4]. In response to global climate change, a number of groups and world governments have developed detailed plans and programs and taken action to reduce emissions. In particular, the Chinese government has solemnly committed to peak its carbon emissions by 2030 and achieve carbon neutrality by 2060 at the General Debate of the General Assembly’s seventy-fifth session [5]. However, in the coming decades, coal-fired power plants are still expected to be the dominant contributors of CO${}_2$ emissions [6]. In China, due to rich coal, poor oil, and less gas energy resource endowment, coal accounts for more than half of the country’s primary energy sources [5]. It is reported that China’s coal consumption will reach  2.8 billion tons of standard coal in 2025, accounting for about 50% of the total energy consumption [7]. Therefore, reducing CO${}_2$ emissions from coal-fired power plants is an important and urgent task.

With the improvement of coal quality in recent years, the coal consumption for power generation is declining year by year [8]. Nevertheless, the percentage of direct CO${}_2$ emissions from CFPPs still reached 90% [9,10]. To reduce CO${}_2$ emissions, many new technologies have been developed and utilized, such as combining coal-fired power plants with solar energy to reduce total emissions [11], or Co-Firing Carbonized Wood Pellets [12], or installing suites of Ultra-Low Emission (ULE) technologies [13], or implementing Carbon Capture and Storage (CCS) to capture the CO${}_2$ emitted by CFPPs [14,15], and so on. Numerous studies have demonstrated that these technologies result in significant reductions in CO${}_2$ emissions. At the same time, all of these studies emphasize that it is necessary to enhance CO${}_2$ emissions monitoring throughout the life cycle. It is becoming an important question of how to effectively and in near-real-time monitor CO${}_2$ emissions from CFPPs. Firstly, it helps determine emission source characteristics and supports air quality modeling, allows a better optimized electricity production process, and achieves such reduction [16]. Second, it is helpful to determine the CO${}_2$ emission reduction effect of ULE, CCS, or similar devices, and it also poses an auxiliary method to optimize these devices’ configuration to achieve the best state [13]. Finally, it would support policy-making to reduce emission levels [17,18], and to be the basis of other related industrial policies [19].

1.1. Early Work

To quantify CO${}_2$ emissions from coal-fired power plants, a lot of research has been done.

According to the Guidelines for National Greenhouse Gas Inventories produced by the Intergovernmental Panel on Climate Change [20], CO${}_2$ emissions from power plants are typically quantified based on bottom-up approaches using fuel consumption and fuel quality. Hence, the Emission-Factor approach is the most important element. Besides the series of default emission factors provided by IPCC, a lot of improved emission factors were proposed [21,22,23,24,25,26,27]. However, most of these emission factors have been developed on long-term emissions, even some studies based on the life cycle model [8,13]. Especially, Alfredo et al. [28] developed a novel bottom-up approach to estimate and forecast emissions, which implemented a series of elements, such as demand, fuel price, etc. However, their temporal resolution is difficult to meet the requirements of CO${}_2$ emission reduction techniques. Although some inventories provide hourly profiles, these profiles do not change from one day to another or from one region to another [29,30]. Liu et al. [13], Chen et al. [31], and Akpan et al. [32] attempted to construct an hourly emission factor, while Liu et al. and Chen et al. did not consider CO${}_2$ emissions [13,31] and Akpan’s approach implemented a power plant-related model to determine the CO${}_2$ emission factor [32]. Therefore, current emission factor inventories are not suitable for monitoring short-term emissions.

To enhance the capability of CO${}_2$ emission monitoring, the U.S. EPA (Environmental Protection Agency) requires CFPPs to equip continuous emission monitoring system (CEMS) and to send the data back continuously [33]. Obviously, the data have a high temporal resolution. However, just a few countries make policies to ask CFPPs to install CEMS [4] because it is expensive to install and maintain [19]. Furthermore, not all CEMS systems are capable of monitoring CO${}_2$. For example, most CEMS equipped with CFPPs in China do not support measuring CO${}_2$ [13], so that many related studies have to avoid discussing CO${}_2$ emissions based on CEMS in China [13,31].

Unlike CEMS, satellite-based CO${}_2$ emission quantification is a novel and convenient method. In recent decades, several satellites have been equipped with special sensors to monitor atmospheric pollution and greenhouse gas, such as SCIAMACHY [34], GOSAT [35], and OCO-2 [36]. Based on these projects, many related works have been done [37,38,39,40]. Because of the narrow scanning area, past studies could not focus on individual point CO${}_2$ emission sources. According to this question, Liu et al. utilized the Ozone Monitoring Instrument to estimate NOx emissions and the CEMS NOx/CO${}_2$ emissions ratio to infer CO${}_2$ emissions [4]. Hu et al. [41] provided a similar resolution, but the difference is that they implemented emission inventories as auxiliary data to estimate the CO${}_2$ emissions of CFPPs. Further, Hu et al. [41] interpolated the resolution of OCO-2 observations by using wind data with higher temporal and spatial resolutions to improve the estimation. However, due to the half-monthly scanning cycle of the satellite, they can hardly observe CO${}_2$ emissions continuously and timely. According to this restriction, Alnaim et al. [19] proposed a near-real-time method to monitor greenhouse gas emissions; however, it could not quantify CO${}_2$ emissions.

Throughout the above studies of accounting and monitoring of CO${}_2$ emissions, there are not only advantages but also various shortcomings. The lack of absolutely precise CO${}_2$ emission data makes it difficult to assess the accuracy of each method. In addition, there are significant differences between the CO${}_2$ emission data obtained by different techniques. For example, Ackerman et al. compared the CO${}_2$ emission data obtained by the U.S. EIA’s and EPA’s databases, which were obtained by different techniques, and found that the difference between the two databases was more than 20% [42]. Therefore, some researchers believe that emission estimates of multiple independent and different technical sources should be a desirable complement for validating and improving the current CO${}_2$ emission inventories [4]. Therefore, this paper attempts to present a novel method for estimating CO${}_2$ emissions from CFPPs using net power generation and other auxiliary emission information.

Machine learning (ML) is a sort of artificial intelligence where classification, prediction, regression, and decision processes can be applied in various fields by creating a mathematical model. Currently, Deep learning (DL) is becoming the most important branch of the ML technique in solving complex problems. Deep neural network (DNN) simulates the brain’s structure with multiple layers of neurons, fitting complex functions, and characterizing the input data distribution, and has demonstrated excellent capacity of automatically learning features. It is widely adopted in computer vision [43], speech-recognition [44], natural language processing [45], etc. Due to the huge success achieved by machine learning and deep learning, a lot of related methods have been proposed to forecast CO${}_2$ emissions at the country level [46,47,48,49]. Beyond that, there is some research focus on special zones or application areas, such as Huang et al. provided a support vector regression (SVR) and long short-term memory network (LSTM) based joint model to predict CO${}_2$ emissions of the Yangtze River Economic Zone [50]. In addition, Alfaseeh et al. applied LSTM to predict greenhouse gas emissions of a road network [51]. Singh et al. proposed an LSTM-based model which uses vehicle On-Board Diagnostics (OBD-II) port data to infer CO${}_2$ emissions of the vehicle [52]. Since the above studies gained amazing performance with implementing ML or DL on estimating CO${}_2$ emissions, this paper attempts to provide a DL model to discover the relationship between power generation and CO${}_2$ emissions, and use electricity data to infer CO${}_2$ emissions from CFPPs.

1.2. Challenges

Using electricity data to infer CO${}_2$ emissions is a cheap and convenient way to monitor CO${}_2$ emissions from CFPPs. Especially for China, it is an important method to support reducing CO${}_2$ emissions and achieve the national vision of carbon peaking by 2030 and carbon neutrality by 2060. Referring to the above related work, a purely data-driven approach of estimating CO${}_2$ emissions would meet a series of challenges as follows:

• How to design a suitable deep learning model to estimate short-term CO${}_2$ emissions from CFPPs? Most current research is around long-term and total CO${}_2$ emissions. They belong to the question of point prediction, while estimating short-term CO${}_2$ emissions is a question of sequence prediction. That is to say, the requirement of short-term estimation not only focuses on the accuracy of total prediction but also the progress information within a certain time range. Hence, designing a model focus on the macroscopic and microcosmic CO${}_2$ emissions from CFPPs is a challenge to us.
• How to obtain enough data to train the DL-based model? The larger the dataset in DL, the better the learning process takes place and the better results are obtained. However, in the real world, high temporal resolution CO${}_2$ emission information of CFPPs is extremely limited. Hence, to prepare enough training data is a precondition to build a DL-based estimator.
• How does the CO${}_2$ emission estimation model adapt to different CFPPs? It is well known that different CFPPs own different generating units in different techniques and capacities, and use different types of coal. Finally, all of these result in different CO${}_2$ emission inventories, especially the short-term CO${}_2$ emitting behavior. Obviously, the unified CO${}_2$ emission model is impossible and unacceptable. How to make a well-trained model adapt to different CFPPs is the primary constraint to implement.

1.3. Purpose of This Study

According to the above challenges, the primary work and contributions of this paper are as follows:

• This paper analyzes the importance of various inherent factors contributing to the CO${}_2$ emissions from CFPPs.
• This paper provides a novel and purely data-driven approach to estimate near-real-time CO${}_2$ emissions from CFPPs through a heterogeneous neural network by utilizing hourly readings of a power meter at the link between CFPPs and the smart grid.
• A coal-fired unit-oriented CO${}_2$ emission simulator is proposed and generates a large number of training data to allow heterogeneous neural networks to be well trained.
• This paper proposes an adaptation method to help the model trained by simulating data to adapt any given CFPP and gain satisfied performance.
• Comparison of the results with other baselines, and demonstration of the strengths and limitations associated with each.

This work is organized as follows. Section 2 analyzes the inherent factors contributing to the CO${}_2$ emissions from CFPPs. The methodology related to the utilized DL model is in Section 3, which includes the simulation and estimation model about CO${}_2$ emissions. Results and discussion are in Section 4. Finally, concluding remarks and future outlook are presented in Section 5.

2. Study on the Mechanism of CO₂ Emissions from CFPPs

2.1. Accounting Boundary

The boundary of CO${}_2$ emissions from CFPPs throughout the life cycle includes 5 stages: coal production, processing, transportation, coal-fired power generation, and resource utilization [8]. This paper primarily investigates the CO${}_2$ emissions among the stage of coal-fired power generation. Hence, the calculation was provided at the facility level of individual power plants. Considering the wide applicability and foundation of facility estimating, the boundary of CO${}_2$ emissions from CFPPs is described in Figure 1.

In the process of coal-fired power generation, coal consumption is the most important energy source of power plants, but also the largest direct CO${}_2$ emission source. The emissions mainly depend on the type of coal, combustion technology, operation conditions, maintenance quality, and the service life of the coal-fired unit. Furthermore, most of CFPPs are also equipped with flue gas desulfurization (FGD), selective catalytic reduction (SCR), and bag-hose precipitation (BHP) devices. Referring to these processes, they also produce direct CO${}_2$ emissions.

According to the above boundary and bottom-up approach, the CO${}_2$ emissions from CFPPs can be calculated as follow:

$$E_{total}=E_{CF}+E_{FGD}+E_{SCR}+E_{BHP}$$

(1)

where $E_{total}$ is total CO${}_2$ emission from CFPPs, $E_{CF},E_{FGD},E_{SCR},E_{BHP}$ are CO${}_2$ emissions produced in the processes of coal-fired power generating units, flue gas desulfurization, selective catalytic reduction, and bag-hose precipitation, respectively. Referring to the bottom-up approach, most work is around constructing the model about these four parts. However, these works are dominated by the physical model of CFPPs and can hardly adapt to technology upgrading of CFPPs in the future. According to this problem, the following subsections introduce the basis of this research and study the key factors of $E_{total}$.

2.2. Research Basis of This Work

CFPPs emit both direct and indirect CO${}_2$ emissions, where the emission factor of the direct part is closely related to the standard coal consumption for power supply. Referring to the above accounting model (1), it is obvious that coal-fired CO${}_2$ emissions are the primary part. To demonstrate it, this paper collected and extracted statistical data of CFPPs from related research [32,53,54,55] where the capacity of units are between 200 MW and 1000 MW. First, we analyze the emission intensity of varying stages and provide the distribution of it in Figure 2a. It shows that CO${}_2$ emissions of FGD, SCR, and BHP are maintained at a small level, especially, since the SCR and BHP are far smaller than 1 g/kWh. Further, Figure 2b clearly shows that the proportion of CO${}_2$ emissions from coal-fired plants is almost close to 99%. Hence, it is possible to infer the total CO${}_2$ emissions from CFPPs through power generation.

In addition, this paper presents the monthly profiles of power generation and CO${}_2$ emissions in Figure 3, and discovers the phenomenon that their monthly fractions of the annual amount are almost the same. Referring to Liu et al. [30], their study also supports this opinion. It means that it is possible to disaggregate long-term CO${}_2$ emissions (such as annual or monthly) to short-term behavior (such as monthly or weekly). It is necessary to note that the data in Figure 3 only show the monthly decomposed result. Nevertheless, shorter-period emission is easier to be affected by fluctuations in units’ load rates or any other factors. Hence, it is necessary to study all possible elements that might influence the inference near-real-time CO${}_2$ emissions from long-term data and net power generation.

2.3. Key Factors Contributing to the CO₂ Emissions from CFPPs

2.3.1. Boiler Type

According to boiler steam parameters, such as main steam pressure and temperature, boiler types primarily are categorized into subcritical, supercritical, ultra-supercritical, and high-efficiency ultra-supercritical. Figure 4 compares the emission intensity of various boiler types and capacities.

As shown in Figure 4, for 300 MW class units, when the boilers are raised from subcritical to supercritical, the emission intensity could be decreased by ~15%. For 600 MW class units, when the boilers are enhanced from supercritical to ultra-supercritical, the emission intensity might be decreased by ~10%. For the 1000 MW class units, when the boilers are promoted from ultra-supercritical to high-efficiency ultra-supercritical, the emission intensity would be decreased by about ~5%. The primary reason is the improvement of boiler efficiency, and the generating efficiency of the unit has been promoted, too. Hence, as  mentioned in Section 1.2, the model constructed on 300 MW class units might not be suitable for 600 MW or 1000 MW class units.

2.3.2. Capacity of Unit

Generally, the higher the unit capacity of the same type of boiler, the higher its power generation efficiency and the lower the standard coal consumption per unit of power generation. The results in Figure 4 preliminarily show that CO${}_2$ emission intensity decreases with the increase of the unit’s capacity. In addition, this paper further utilizes public and private data to analyze the distribution of emission intensity with various unit capacities. The result is shown in Figure 5.

Referring to the results shown in Figure 5, from 200 MW to 1000 MW, the average emission intensity continues to decline. It also shows that the generalization of the emission model among different CFPPs is important. Further, it also finds that the shape of the distribution of emission intensity is increasingly sharp with increasing unit’s capacity. It indicates that the larger the unit’s capacity, the less the emission intensity is affected by other factors. It is reasonable to believe from the results in Figure 5 that it is more possible to estimate emissions using net generation for larger coal-fired generating units.

2.3.3. Load Rate

The coal consumption of the power plant is the lowest when the power plant operates under the designed load level. Low-load operation will lead to an increase in coal consumption and then increase the emission intensity. Many researchers have attempted to study the relationship between unit load rate and emission intensity. Based on a large number of reliable measurements of actual CFPPs, it is shown that the higher the unit operating load, the smaller the coal consumption, and the higher the unit operating load, the higher the boiler efficiency, the lower the power supply of coal consumption [32,56]. Figure 6 clearly shows that emission intensity decreases rapidly with the increase of unit load rate, regardless of unit load capacity. Therefore, the load rate of the unit must be taken into account when estimating the CO${}_2$ emissions by using net power generation.

2.3.4. Coal Type

Coal is the only fuel for power generation in CFPPs. Different types of coal contain different carbon content, water content, and volatile fraction, etc. It influences boiler thermal efficiency, heat consumption rate, and plant power consumption rate deeply. Hence, the type and quality of coal directly determine the coal consumption and CO${}_2$ emission intensity of power generation units. The IEA estimated the average carbon emission factors for electricity generation to be 860 g/kWh, 870 g/kWh, and 1020 g/kWh for anthracite, bituminous coal, and lignite, respectively [8]. Therefore, it is not appropriate to utilize static models to estimate the CO${}_2$ emissions of different CFPPs.

2.3.5. Condenser Type

Currently, most CFPPs choose an air condenser, which can be divided into direct air-cooled condenser (ACC) and indirect surface condenser (ISC). ACC includes mechanical ventilation and natural ventilation according to the ventilation mode. ISC is divided into the surface condenser and mixed condenser according to the different condensers used. In fact, the type of condenser has no significant influence on the emission intensity of CFPPs, but the power consumption of every cooling equipment varies greatly, which indirectly affects the auxiliary load rate of the CFPPs, thus affecting their emission intensity. It can be found in Figure 7. Obviously, the power consumption of the cooling system is directly related to the load rate of the generation unit [32,57].

2.3.6. Auxiliary Loads

At present, using emission factors to measure the CO${}_2$ emissions of CFPPs is still an important solution. This approach requires, in addition to appropriate emission factors, accurate estimates of the total power generation. The total power generation is composed of net power generation and auxiliary load. Currently, the net power generation can be accurately accounted for, but there is no effective measurement for the auxiliary load. Therefore, estimating the auxiliary load is crucial for accurately measuring the CO${}_2$ emissions of CFPPs [32]. By studying the relationship between auxiliary load and CO${}_2$ emission intensity of CFPPs, Akpan [32] and She [56] found that the fraction of auxiliary load is inversely proportional to net power generation, as shown in Figure 8. Usually, the auxiliary load of a CFPP is composed of four parts: circulating pump, powder making, water pump, and fan. There are some differences between various CFPPs. Therefore, when using net power generation to estimate carbon dioxide emissions, not only the differences in the above factors but also the auxiliary load should be taken into account.

2.3.7. Summary

Through the above study, the conclusions can be drawn as follows:

• There is a strong correlation between the CO${}_2$ emission intensity and the load rate of the units. Therefore, it is possible to measure the short-term CO${}_2$ emissions of CFPPs by using the net power generation.
• When using net power generation to measure CO${}_2$ emissions, it is necessary to take into account the differences between the generating units of CFPPs and various operating states.
• The net power generation-based CO${}_2$ emission measurement model must have a strong capability of data fitting and unit adaptation.

Hence, this paper proposes a novel heterogeneous neural network and adaptation algorithm to estimate near-real-time CO${}_2$ emissions from CFPPs through hourly readings of power meters at the entry between CFPPs and the smart grid.

3. Method

In this section, we present our method which infers CO${}_2$ emissions of CFPPs through net power generation. The section introduces the framework of our proposal, the heterogeneous neural network, and its training and adaptation algorithm.

3.1. Architecture and Procedure

Figure 9 illustrates the architecture of this research work. The figure shows that the proposal consists of a deep neural network-based CO${}_2$ emission estimator, a regression-based CO${}_2$ emission simulator, and a physical variable model-based CO${}_2$ emission simulator. Here, the estimator poses a mathematical model between the CO${}_2$ emissions and the net power generation. In this way, the long-term emission can be decomposed into smaller periods. It not only approximates CO${}_2$ emissions by net electricity but also improves the temporal resolution of estimating CO${}_2$ emissions.

Further, to achieve satisfactory estimating performance, this paper designs a two-phase training strategy to eliminate the differences in the production process, unit generating efficiency, coal type, and its quality. Figure 10 is the flowchart of this training strategy. First, we create a regression-based CO${}_2$ emission simulator to generate a great deal of artificial data for pre-training the estimator. Obviously, the current estimator could not work for any real CFPPs, but it is well-trained to be a powerful mathematical model that can map an electricity sequence to an emission sequence. Second, this proposal designs a calibration algorithm to make an estimator adapted to any real CFPPs. In this stage, the pre-trained estimator is fed with few calibration data which contain hourly electricity and vetted daily emission. According to different situations, the vetted emission might be accounted for by bottom-up CO${}_2$ inventories or generated by another physical variable model-based CO${}_2$ emission simulator. After additional training, the estimator can be applied to monitor the CO${}_2$ emissions from CFPPs.

3.2. Emission Simulator

Currently, the CO${}_2$ emission data in high temporal resolution are still limited except for a few developed countries. For example, even though a large number of CFPPs had equipped CEMS in China, it only monitors NO${}_{\mathrm{x}}$, SO${}_2$, and VOCs, not CO${}_2$. On the other side, there are no open datasets available that report CO${}_2$ emissions with high temporal resolution [51], currently. Therefore, an emission simulator is important and urgent to this study. To train and validate our proposed model, this paper provides two simulators with different technical solutions for different scenarios, such as training and validation.

3.2.1. RES: Regression-Based Emission Simulator

Figure 6 shows that the CO${}_2$ emission intensity has a strong relationship with net power generation efficiency. Even though emission intensity is influenced by multiple factors, this paper attempts to fit the correlation between them by a nonlinear model, such as a power model, exponential model, or polynomial model, etc. By comparing their residuals, this paper chooses the cubic polynomial model to fit the relationship between emission intensity and net power generation efficiency. Furthermore, it should be mentioned that the residual of the cubic polynomial model still cannot be ignored because it is affected by many factors, especially, the demand for the power grid. Therefore, we add a normal random factor into the simulation model to ensure that the generated data are closer to the real-world data. The detailed simulation model is shown in Equation (2).

$$E=\left\{\begin{array}{cc}(-2.147R^3+4.611R^2-3.707R+1.99+N)\ast W_G,\hfill & 300\mathrm{MW}\hfill \\ (-2.117R^3+4.449R^2-3.35R+1.767+N)\ast W_G,\hfill & 600\mathrm{MW}\hfill \\ (-0.77R^3+1.762R^2-1.48R+1.246+N)\ast W_G,\hfill & 1000\mathrm{MW}\hfill \end{array}\right.$$

(2)

where E is the CO${}_2$ emission of CFPP in tons, N is the random factor, $W_G$ is the capacity of CFPP in MW, R is the net power generation efficiency calculated by:

$$R=\frac{NP}{W_G}$$

(3)

where $NP$ is hourly net power generation in MWh.

3.2.2. PES: Physical Variable Model-Based Emission Simulator

The physical variable model also belongs to the range of emission factor methods. However, the traditional emission factor methods and material balance methods do not consider the influence of the load rate of the power unit. Hence, they cannot be used to generate near-real-time CO${}_2$ emissions of CFPPs. In this section, another simulator which is based on the variable emission factor model proposed by Akpan [32] is implemented. It takes into account the re-heat architectures of CFPPs, the quality of the coal, the load conditions, and the characteristics of the boiler and cooling equipment. The detailed simulation model is provided as follows:

$$E=\frac{44}{12}\left({\chi }_C-\frac{HHV·Los{s}_{uc}}{33.83}\right)\frac{TCH{R}_{FL}·R^{-\mathbb{H}}}{{\eta }_B{\eta }_G·HHV·(1-Los{s}_{uc})(1+f_{aux})}\ast W_G$$

(4)

where ${\chi }_C$ is a mass fraction of carbon in the coal, $HHV$ is the higher heating value of coal, $Los{s}_{uc}$ is the unburnt carbon loss of a boiler, $TCH{R}_{FL}$ is the turbine cycle heat rate at full load, $\mathbb{H}$ is the exponential coefficient of the variable turbine cycle heat rate. ${\eta }_B$ and ${\eta }_G$ are efficiencies of the boiler and generator, respectively. $f_{aux}$ is a fixed portion of the auxiliary load. $Los{s}_{uc}$ can be computed by:

$$Los{s}_{uc}=-0.0881R+1.2746$$

(5)

Since the fixed auxiliary load fraction is rarely mentioned by other studies, this paper simplifies it by a power function. By extracting the data of several CFPPs drawn from She [56] and Xu [58], this paper provides a simpler fitting model of the $f_{aux}$ as following:

$$f_{aux}=0.311·R^{-0.383}$$

(6)

Different from the regression model, the ${\chi }_C$, $HHV$, $TCH{R}_{FL}$, $\mathbb{H}$ are randomly enough according to characteristics of CFPPs; hence, it is not necessary to add another random factor to this simulation model anymore.

3.3. Emission Estimation Network (EEN)

In order to obtain the near-real-time CO${}_2$ emissions from CFPPs, we include the hourly CO${}_2$ emission sequence and daily emissions. We proposed a model to transfer the hourly net power generation sequence to the hourly CO${}_2$ emission sequence, and then use the sum of them to be the daily emissions. According to the conclusion presented by [59], a time series has various types of features, and better features can be obtained by ensemble algorithms. Therefore, this paper designed a heterogeneous neural network as shown in Figure 11.

In detail, the EEN consists of AE (Auto-Encoder), MLP (MultiLayer Perceptron), and IncetionTime [60]. In addition, each network could work independently. AE is a widely applied generative model with a strong ability for denoising and data generation. With the help of its characteristics, this paper uses net power generation to generate CO${}_2$ emissions. Due to AE consisting of typical convolutional and transposed convolutional layers, AE is more focused on the local sequence. Hence, an MLP module is brought into EEN and it allows us to learn the global knowledge from the net power generation sequence. In addition, considering that the CO${}_2$ emission behavior of CFPPs has a strong local correlation in time, and the time scale is slightly different with different CFPPs, this paper introduces a revised IncetionTime module into EEN. The IncetionTime module is a ResNet-style convolution network in which every IncetionTime block owns 3 convolution paths with the different receptive field. It allows EEN to learn precise local features to improve the temporal resolution of CO${}_2$ emissions.

Moreover, the EEN designs an ensemble module to integrate their outputs. The following equation explains the ensemble of predictions made by all three modules:

$$\tilde{E}={\omega }_1\ast A\left(R\right)+{\omega }_2\ast M\left(R\right)+{\omega }_3\ast I\left(R\right)$$

(7)

where $\tilde{E}$ is the integrated CO${}_2$ emission sequence, R is the normalized net power generation sequence. All $\omega$s are weighted to ensemble three modules’ estimation and they are all trainable.

3.4. Loss and Training

All modules of the EEN are trained in two phases:

• Independent Phase: In this phase, each neural network runs as an independent multivariate regressor. The EEN updates them to minimize the Mean Square Error (MSE), respectively:

  $$mi{n}_{A,F,I}\mathbb{E}[||E-\tilde{E}{\left|\right|}^2]$$

  (8)

  where $\tilde{E}$ is the estimated emissions by R, and $\left|\right|·{\left|\right|}^2$ is the L2 norm.
• Ensemble Phase: Firstly, EEN collects the outputs of every neural network independently. Then, we ensemble them through trainable weights. At this stage, the above neural networks would be frozen, and just ensemble weights should be updated. This phase can be represented by:

  $$mi{n}_{{\omega }_{1,2,3}}\mathbb{E}[||ES-({\omega }_1\ast {\tilde{E}}_A+{\omega }_2\ast {\tilde{E}}_M+{\omega }_3\ast {\tilde{E}}_I){\left|\right|}^2]$$

  (9)

The EEN must be trained independently with Adam [61]. The pseudocode of the training algorithm with mini-batches is provided by Algorithm 1.

3.5. CFPP Adaptation

Referring to Figure 5 and Figure 6, the CO${}_2$ emission intensity is different for various CFPPs even with the same capacity. It is obvious that the estimator trained with simulation data is bound to lose effect. Therefore, it is necessary to use the data of the target CFPPs to adaptively train the pre-trained model. In the pre-training phase, Algorithm 1 optimizes the EEN based on the sequence data and takes the MSE of the sequence as the loss. Hence, to solve the lack of hourly CO${}_2$ emission data, this paper utilizes daily CO${}_2$ emission data and follows the Algorithm 1 to fine-tune EEN. The detailed steps are as follows:

• Load the pre-trained EEN.
• If there are only monthly emissions, they should be decomposed into daily emissions according to the daily fraction of the monthly net power generation as below:

  $$E_d=\frac{P_d}{P_m}\ast E_m$$

  (10)

  where $P_m$ and $E_m$ are monthly net power generation and CO${}_2$ emissions, respectively. $P_d$ and $E_d$ are the daily net power generation and CO${}_2$ emissions of the $d^{th}$ day.
• To construct the data samples, the normalized hourly net power generation is posed as input and the daily CO${}_2$ emissions is labeled.
• Fine-tune the pre-trained EEN following Algorithm 1 by targeting CFPP’s samples produced by the former step. In addition, the loss function of Algorithm 1 should be changed as below:

  $$mi{n}_{A,M,I}\mathbb{E}[||E_d-\sum _{h=0}^{23}{\tilde{E}}_h{\left|\right|}^2]$$

  (11)

  where ${\tilde{E}}_h$ is the hourly emission, $E_d$ is the corresponding daily emission. ${\tilde{E}}_h$ and $E_d$ are the approximate values of CFPP’s hourly and daily $E_{total}$ in Equation (1). $\mathbb{E}[·]$ means expected value. When fine-tuning the EEN, the ${\omega }_{1,2,3}$ are fixed.

After the above adaptation steps, the EEN can estimate any CFPP’s hourly or daily $E_{total}$ by its net power generation. It should be mentioned that the learning rate of the adaptation stage should be far smaller than Algorithm 1.

Algorithm 1 Mini-batch pre-training of EEN

Require:  r = a certain training input, the hourly sequence of normalized net power generation efficiency Require:  B = a batch of training inputs, $\left|B\right|$ is the size of batch Require:  e = labels for labeled inputs, the hourly sequence of normalized CO${}_2$ emissions Require:  $A_{\theta }\left(r\right)$ = AE with trainable parameters $\theta$ Require:  $M_{\gamma }\left(r\right)$ = MLP with trainable parameters $\gamma$ Require:  $I_{\varphi }\left(r\right)$ = InceptionTime with trainable parameters $\varphi$ Require:  ${\omega }_{1,2,3}$ = weight of ensemble module     1: fort = 1 to $iterations$ do     2:    Draw a minibatch B from generated samples randomly     3:    ${\tilde{e}}_A^i\leftarrow A_{\theta }(r_i\in B)$     4:    $los{s}_A\leftarrow \frac{1}{\left|B\right|}{\sum }_{i\in B}d(e_i,{\tilde{e}}_A^i)$     5:    update $\theta$ using individual Adam optimizer     6:    ${\tilde{e}}_M^i\leftarrow M_{\gamma }(r_i\in B)$     7:    $los{s}_M\leftarrow \frac{1}{\left|B\right|}{\sum }_{i\in B}d(e_i,{\tilde{e}}_M^i)$     8:    update $\gamma$ using individual Adam optimizer     9:    ${\tilde{e}}_I^i\leftarrow I_{\varphi }(r_i\in B)$   10:    $los{s}_I\leftarrow \frac{1}{\left|B\right|}{\sum }_{i\in B}d(e_i,{\tilde{e}}_I^i)$   11:    update $\varphi$ using individual Adam optimizer   12:    ${\tilde{e}}_i={\omega }_1\ast {\tilde{e}}_A^i+{\omega }_2\ast {\tilde{e}}_F^i+{\omega }_3\ast {\tilde{e}}_I^i$   13:    $los{s}_e\leftarrow \frac{1}{\left|B\right|}{\sum }_{i\in B}d(e_i,{\tilde{e}}_i)$   14:    update ${\omega }_{1,2,3}$ using individual Adam optimizer   15: end for

4. Experiments

In this section, we first introduce the experimental setup, and conduct experiments to answer the following questions.

• Effectiveness: How accurate are the CO${}_2$ emission estimation results computed by EEN and baseline methods on simulation datasets?
• Robustness: How does the robustness of EEN behave with respect to various CFPPs’ capacity levels of the test set?
• Ablation study: Does the proposed EEN contribute to a sufficiently improved estimation performance?
• Sensitivity: How does the hyperparameter of EEN influence the estimation performance?

4.1. Experiment Setting

4.1.1. Experimental Data

All training and validation data are generated by the simulator which is described in Section 3.2. To answer the above questions, this paper implements different simulation models for the training and validation stages. Firstly, this paper selected two CFPPs in each of the 300 MW, 600 MW, and 1000 MW classes for training and testing. At the same time, we collected their hourly net generation data for 1 year from CSG (China Southern Power Grid) to support the simulator to generate training and validation data. Detailed information on experimental data is introduced below:

• Training set: All data samples are generated through the Regression model mentioned in Section 3.2.1. Each daily magnitude would generate 10 samples with different random factors defined in Equation (2).
• Validation set: In order to avoid the overfitting of the EEN, the physical variable model mentioned in Section 3.2.2 was used to generate the validation dataset. For each set of electricity magnitude of CFPP, we choose different parameters randomly. Based on Akpan [32] and extensive research, the important parameters of the physical variable model are provided in

  Table 1. The range of simulation parameters used in the physical variable model.

  Simulated CFPP	Capacity (MW)	${\mathit{TCHR}}_{\mathit{FL}}$	$\mathbb{H}$	Range of ${\mathit{\chi }}_{\mathit{C}}$	Range of $\mathit{HHV}$
  1	300	9016.5	0.1087	[0.56, 0.59]	[21, 22]
  2	300	8623.8	0.1127	[0.56, 0.59]	[21, 22]
  3	600	7962.3	0.0624	[0.58, 0.60]	[22, 23]
  4	600	7548.6	0.0716	[0.58, 0.60]	[22, 23]
  5	1000	7396.4	0.0542	[0.59, 0.61]	[23, 24]
  6	1000	7062.8	0.0602	[0.59, 0.61]	[23, 24]

  Here, ${\chi }_C$ and HHV are selected randomly within the corresponding range, respectively.

  .

Finally, this paper collected about 5000 training samples in total and about 300 validation samples for each simulated CFPP.

It should be mentioned that hourly electricity data are normalized by the unit’s capacity before feeding to the CO${}_2$ emission estimator. In addition, the emission data are also normalized by a preconfigured max value referring to each unit. Hence, the output of the estimator should be inversely normalized.

4.1.2. Baselines

The proposed EEN is compared with the following machine learning methods:

• Support Vector Regression (SVR): SVR is a popular regressor for time-series forecasting. In addition, SVR supports various kernels for different applications. The electricity magnitude and CO${}_2$ emissions are nonlinear. Hence, the kernel is set as Radial Basis Function(RBF) in this section, and the penalty parameter is set to 10.
• Gradient Boost Regression (GBR): GBR is one of the ensemble learning methods and uses gradient boosting with a regularized cost function. According to this study, the number of trees is set as 200, the max depth of each tree is 5, the minimum child weight equals 1, and the learning rate is configured as 0.01.
• Fully Convolution Network (FCN) [62]: FCN is proposed to perform time-series classification by Wang [62] and showed strong competitiveness. In this paper, we borrow FCN’s capability of feature learning and just change the last fully connected layer to a regressor of CO${}_2$ emission sequence. The outputs of the last layer are of the same shape as the input sequence and are activated by sigmoid.
• Temporal Convolutional Network (TCN) [63]: TCN brings about dilated convolutions for time series forecasting and classification and proves that dilated convolutions outperform RNNs in terms of both efficiency and predictive performance. In this paper, the TCN equipped two dilated convolutional blocks. Each block contains two 1D convolutional layers with a dilation parameter of (1,2,4,8) and each convolutional layer consists of 64 filters. The output layer is the same as FCN.
• LSTM: It is a stack of an LSTM layer and a fully connected layer. The hidden size of the LSTM layer is 8, and the output layer is the same as the last layer of FCN.

The above competitor list includes both traditional and deep approaches. In addition, these baseline methods employ different learning strategies. Hence, the above baseline list can well represent the competing performance of this research. The KNN, SVR, and GBR are based on scikit-learn [64]. In addition, all of the deep neural networks are developed on Keras.

4.1.3. Configuration

In this paper, the primary parameters of EEN are shown in Figure 11. Our proposed EEN and all baseline models are optimized with the Adam optimizer, and its learning rate is fixed as $2\times {10}^{-3}$ for pre-train and $2\times {10}^{-4}$ for adaptation. The total number of epochs is 100 with proper early stopping. We set the baseline methods as recommended, and the batch size is 100.

4.1.4. Evaluation Metrics and Computing Infrastructure

In order to compare with the above methods, the Mean Absolute Percentage Error (MAPE) can be applied to evaluate the performance of our proposal and baselines. Since the performance of hourly and daily estimations are both important in this paper, the detailed metrics are defined as follows:

$$MAP{E}_H=\frac{1}{n}\sum _{i=1}^n\frac{|e_i-{\widehat{e}}_i|}{e_i}$$

(12)

$$MAP{E}_D=\frac{1}{n}\sum _{i=1}^n\frac{|sum\left(e_i\right)-sum\left({\widehat{e}}_i\right)|}{sum\left(e_i\right)}$$

(13)

where $MAP{E}_H$ and $MAP{E}_D$ are the MAPE of hourly and daily estimation, respectively. $e_i$ and ${\widehat{e}}_i$ are labeled and estimated hourly CO${}_2$ emissions, and $|·|$ is the absolute value function. n is the number of estimations.

4.2. Results

4.2.1. Results and Analysis

Table 2. MAPE of daily CO${}_2$ emissions.

	EEN	KNN	SVR	GBR	TCN	FCN	LSTM
300 MW-1	1.03	2.54	2.23	1.86	1.95	2.02	1.83
300 MW-2	0.98	3.01	2.45	2.25	1.89	1.74	1.72
600 MW-1	1.37	3.42	4.72	2.32	2.47	2.50	2.20
600 MW-2	1.07	3.63	3.63	2.42	2.18	2.61	1.93
1000 MW-1	1.11	2.98	3.12	2.78	2.10	2.73	2.07
1000 MW-2	1.28	3.11	4.35	2.57	2.32	2.81	2.18

and

Table 3. MAPE of hourly CO${}_2$ emissions.

	EEN	KNN	SVR	GBR	TCN	FCN	LSTM
300 MW-1	5.18	8.59	9.05	7.09	6.45	6.95	8.19
300 MW-2	5.33	9.06	9.08	7.57	6.55	6.65	8.58
600 MW-1	5.29	9.37	10.09	7.40	6.74	7.29	8.67
600 MW-2	5.36	9.33	9.75	7.41	6.78	7.49	8.69
1000 MW-1	5.32	8.97	9.16	7.81	6.73	7.48	8.57
1000 MW-2	5.48	8.82	10.18	7.76	6.58	7.34	8.65

summarize the daily/hourly CO${}_2$ emission estimation results of all the methods on simulated CFPPs. Since the KNN, SVR, and GBR could not support daily and hourly estimating simultaneously, this paper collects their daily and hourly estimation results for independent models. Hence, they are not suitable for deep learning-based methods. The best results are highlighted in boldface. It should be mentioned that, unless otherwise noted below, the experimental results are based on 30 adaptation samples.

Daily CO₂ Emission Estimating

Under this setting, KNN, SVR and GBR attain estimations as a single variable over the time series, while the remaining deep learning methods always estimate hourly CO${}_2$ emissions and account them as daily emissions. From Table 2, it can be observed that: (1) The proposed model EEN significantly improves the inference performance across all simulated CFPPs, and their estimate error keeps stable for various CFPPs, which demonstrates the success of EEN in enhancing the estimation capacity in the problem of monitoring CO${}_2$ emissions of CFPPs. (2) Generally, deep learning-based methods are better than traditional machine learning methods, because the deep neural network has a strong data fitting ability. (3) The proposed EEN is outperforms TCN, FCN and LSTM on MAPE by decreasing 40–60% in average. It shows that our method is more stable and powerful.

Hourly CO₂ Emission Estimating

Within this setting, KNN, SVR, and GBR are inappropriate and require to be changed from univariate prediction to multivariate one by redesigning the models. On the contrary, our proposed EEN is easy to adapt to the daily and hourly situation without any adjusting to the model. From Table 3, it can be observed that: (1) the proposed model EEN greatly outperforms other methods, and the findings 1 and 2 in the daily settings still hold for the hourly CO${}_2$ emissions. (2) The EEN model shows better results than baselines, and the MAPE decreases about 20–46% on average. Compared with the daily results, the overwhelming performance is reduced, and such phenomena can be caused by the uncertainty of power demand in a short time period. It is beyond the scope of this paper, and we will explore it in future work.

In addition, this paper uses an exact sample to show that our EEN model is powerful to predict the details of CO${}_2$ emission behavior in Figure 12 and Figure 13. According to this setting, this paper utilizes Akpan [32] as ground truth and compares the effectiveness of all methods on the same example. The most related work shows acceptable results. TCN and TCN have a good performance on daily emissions, but it is easy to miss details when estimating hourly emissions. There are similar problems with other approaches. Our proposed model EEN shows significantly better results than other methods for both daily and hourly CO${}_2$ emission estimating.

4.2.2. Ablation Study

This subsection conducts additional experiments with ablation consideration and demonstrates how well EEN works.

Performance of Ensembling

In this study, we compare the MAPE of our adaptive ensembling and average ensembling in Figure 14. Firstly, due to EEN consisting of heterogeneous neural networks, it allows EEN to capture various features and fit the relationship between net power generation and CO${}_2$ emissions of CFPPs. Comparing the results of Figure 14 and Table 2 and Table 3, average ensembling is still better than a single model. It demonstrates that the EEN is competent for CO${}_2$ emission estimation. Second, due to the ensembling weights of EEN being learnable, EEN could provide better integration of heterogeneous neural networks. Compared to average ensembling, EEN achieved about 20% and 7% improvement on daily and hourly MAPE, respectively.

Performance of Adaptation

Our approach presents a CFPP adaptation mechanism to help EEN which is pre-trained by simulated data to different CFPPs. In this study, we validate the effectiveness of this mechanism. In

Table 4. Comparing the MAPE of EEN before and after adaptation about daily and hourly CO${}_2$ emission estimation.

	${\mathit{MAPE}}_{\mathit{hourly}}$		${\mathit{MAPE}}_{\mathit{daily}}$
	Before Adaptation	After Adaptation	Before Adaptation	After Adaptation
300 MW-1	6.11	5.18	2.55	1.03
300 MW-2	6.09	5.33	2.51	0.98
600 MW-1	8.67	5.29	4.88	1.37
600 MW-2	7.94	5.36	4.71	1.07
1000 MW-1	7.89	5.32	4.25	1.11
1000 MW-2	8.13	5.48	4.72	1.28

, we compare the daily and hourly MAPE before and after adaptation. In this setting, this paper still utilizes 30 samples of target CFPPs to finish adaptation training. Referring to the results of Table 4, it can be observed that the CFPP adaptation mechanism improves the performance of EEN on daily and hourly MAPE about 69% and 27% on average, respectively.

Robustness of EEN

Referring to the limit of effective samples mentioned in Section 3.2, it is necessary to validate the robustness of EEN for fewer samples. In this experiment, we study the performance of EEN for varying adaptation samples. From the results of Figure 15, it can be observed that the EEN could still keep competitive performance even with just 10 adaptation samples. Especially, the difference between daily and hourly MAPE are both limited within 0.4%. However, it could also be found that more adaptation samples would acquire better performance.

4.2.3. Uncertainties

This paper estimated the uncertainty of the fit performance of EEN. We estimated uncertainties of daily and hourly MAPE of EEN for individually simulated CFPP. In this section, the original and unaveraged estimated error and the Absolute Percentage Error (APE) are used to discuss the uncertainty of EEN. In detail, APE is computed as below:

$$AP{E}_i=\frac{|e_i-{\widehat{e}}_i|}{e_i}$$

(14)

where $AP{E}_i$ is the APE of the $i^{th}$ time unit, day or hour. $e_i$ and ${\widehat{e}}_i$ might be the $i^{th}$ labeled and estimated daily or hourly CO${}_2$ emissions.

Uncertainty of Daily CO${}_2$ Emission Estimation

This paper collects all APE of every day and every CFPPs and does a statistics of them, respectively. From Figure 16, it can be found that APE of EEN is mostly located between 0% and 3%. For each CFPP, the standard deviation of daily APE is around 0.9%. Only a few extreme cases are beyond 4%, but still limited within 8%.

Uncertainty of Hourly CO₂ Emission Estimation

In this study, we separately analyzed the distribution of hourly APE across different days. From Figure 17, it can be observed that hourly APE is distributed within a range of $\pm 3\%$ compared to the hourly MAPE. Even though the hourly APE between 0–6 A.M. is higher than that of other time periods, the standard deviation remained stable. Hence, it can be concluded that the uncertainty of hourly CO${}_2$ emission estimation is limited.

Uncertainty of CO₂ Emission Estimation at Different Load Conditions

In this study, we continue to discuss the uncertainty of hourly CO${}_2$ emissions according to different load rates. From Figure 18, the load rate of CFPPs would affect the MAPE of EEN. When the load rate is smaller than 50%, the hourly CO${}_2$ emissions are more easily affected by boiler efficiency and auxiliary loads, and the uncertainty of CO${}_2$ emission estimation is greater. With the increase in load rate, the uncertainty is improved obviously. Especially, the standard deviation of hourly APE would be decreased by about 1% when the load rate is over 60%. According to the results of Figure 18, which has a commonplace with Figure 17, the hourly APE is higher in the period of 0–6 A.M. because the load rate of CFPPs in such time period is lower.

4.3. Summary of Experiments

First, EEN outperforms other baseline machine learning and deep learning algorithms on the same dataset. Not only the MAPE of CO${}_2$ emissions but also the stability and uncertainty, the experimental results all show that EEN has better performance and robustness.

Second, in Table 2, the $MAP{E}_D$ of EEN is only about 1%. It shows that the performance of EEN is close to Akpan [32]. However, besides net power generation and capacity of the unit, Akpan [32] requires a series of parameters, such as the mass fraction of carbon in the coal, heating value of coal, turbine cycle heat rate at full load, exponential coefficient of the variable turbine cycle heat rate, etc. These parameters are not only difficult to collect, but it is also hard to determine their exact values. Hence, our approach is easier to apply.

Finally, this paper utilizes emissions estimated by the Akpan [32] model as the ground truth to analyze the uncertainty of EEN. From the experimental results, the uncertainties are limited to 10%. However, the uncertainties of the methods proposed by Liu [4] and Hu [41] are both more than 20%. Compared to the current EEN, they are more convenient and generalizable because they both use OCO-2 data.

5. Conclusions

Overall, this paper provides the following contributions:

• This paper provides the quantitative relationship between CO${}_2$ emissions from CFPPs and all primary elements. Through analyzing the data extracted from a wide range of related research, we found that there is an inherent mathematical relationship between the net power generation and the emission intensity. Further, the proportion of emissions produced by auxiliary equipment of CFPPs is very small and all of them are related to an auxiliary power load. This conclusion provides the motivation for our research and guides us to design EEN and related algorithms. At last, it could be found that the emission intensity is different from CFPPs.
• This paper provides a novel and purely data-driven approach to estimate near-real-time CO${}_2$ emissions from CFPPs by net power generation. We introduce deep learning into estimating emissions from CFPPs, and our proposed EEN is a deep assembling model based on heterogeneous neural networks. The experimental results show that EEN owns outperforming data fitting and generalization capabilities. Furthermore, it is worth pointing out that our proposed EEN only requires the hourly readings of the power meters on the link between CFPPs and the smart grid as net power generation. Hence, it is an easy solution to implement.
• This paper provides a novel training algorithm so that our proposed EEN could be adapted to any CFPPs by a few realistic datasets. Due to the limitation of real and well-vetted data, we present a two-phase training algorithm: pre-train and adaptation. First, we use a large number of artificial data generated by a regression model-based simulator to pre-train our proposed EEN. Then, we utilize a few datasets collected from the target CFPP to finish the adaptation of EEN. This strategy allows us to utilize less realistic data to adapt to any CFPPs at the stage of application. This further reduces the cost of the application and improves the implementability.

In this study, a novel data-driven framework for monitoring CO${}_2$ emissions from CFPPs in near-real-time is presented to avoid installing expensive and complicated CEMS. Firstly, we investigate all possible elements of CO${}_2$ emissions from CFPPs. We found that the net power generation is strongly related to the emission intensity, and the emission intensity is different from CFPPs. Next, we propose a deep ensembling model based on heterogeneous neural networks with powerful data fitting and generalization capabilities. Then, due to the limitation of real and well-labeled data, we present a two-phase training algorithm: pre-train and adaptation. We use a large number of artificial datasets generated by the simulator to pre-train our proposed EEN. We allow to utilize fewer datasets of the target CFPP to finish calibration at the stage of application. At last, the experimental results show that our scheme outperforms all baselines.

We found that it is feasible to infer CO${}_2$ emissions from the net power generation of CFPPs, but limitations of the current real-world data only allow us to verify our method on the simulator. Looking forward, we anticipate that these limitations will diminish for more CFPPs to place greater emphasis on data collection and analysis so that our proposed EEN could be adapted to more CFPPS. Therefore, future work will be to validate and optimize our method to the real CFPPs equipped with CEMS or to hold perfectly vetted datasets.