An integrated simulation and optimization approach for managing human health risks of atmospheric pollutants by coal-fired power plants

Abstract

This research developed a simulation-aided nonlinear programming model (SNPM). This model incorporated the consideration of pollutant dispersion modeling, and the management of coal blending and the related human health risks within a general modeling framework. In SNPM, the simulation effort (i.e., California puff [CALPUFF]) was used to forecast the fate of air pollutants for quantifying the health risk under various conditions, while the optimization studies were to identify the optimal coal blending strategies from a number of alternatives. To solve the model, a surrogate-based indirect search approach was proposed, where the support vector regression (SVR) was used to create a set of easy-to-use and rapid-response surrogates for identifying the function relationships between coal-blending operating conditions and health risks. Through replacing the CALPUFF and the corresponding hazard quotient equation with the surrogates, the computation efficiency could be improved. The developed SNPM was applied to minimize the human health risk associated with air pollutants discharged from Gaojing and Shijingshan power plants in the west of Beijing. Solution results indicated that it could be used for reducing the health risk of the public in the vicinity of the two power plants, identifying desired coal blending strategies for decision makers, and considering a proper balance between coal purchase cost and human health risk.

Implications:

A simulation-aided nonlinear programming model (SNPM) is developed. It integrates the advantages of CALPUFF and nonlinear programming model. To solve the model, a surrogate-based indirect search approach based on the combination of support vector regression and genetic algorithm is proposed. SNPM is applied to reduce the health risk caused by air pollutants discharged from Gaojing and Shijingshan power plants in the west of Beijing. Solution results indicate that it is useful for generating coal blending schemes, reducing the health risk of the public, reflecting the trade-off between coal purchase cost and health risk.

Introduction

Stack emissions of coal-fired power plants, such as sulfur dioxide (SO₂), nitrogen oxides (NO_x), and particulate matter that are less than 10 μm in diameter (PM₁₀), have a number of adverse impacts on public health due to their correlation with respiratory illness, chronic bronchitis and premature death (Cheng et al., 2007; Schuhmacher et al., 2004). Concerns over alleviating such emissions have thus been increasing in the past decades. Many technologies have been advanced, such as coal blending. As an effective clean-coal combustion technology, coal blending can be used not only for adjusting coal features for specific power plant boilers, but also for alleviating atmospheric pollutant emissions and reducing the resulting health risks of the public in the vicinity of power plants (Cai et al., 2009; Cui, 2008; Liu et al., 2009). Identification of optimal coal blending schemes is desired to achieve multiple targets such as maximization of economic benefits, minimization of pollution emission, and alleviation of the associated health risks. However, such a process is fraught with complexities due to the diversities of coal features, as well as the trade-offs among power generation, pollution emissions, and the resulting human risks. This leads to many challenges for decision makers and managers of many power plants. Therefore, it is necessary to propose an integrated approach through incorporating pollutant dispersion modeling, and the management of coal blending and the related human health risks within a general modeling framework.

Previously, many types of dispersion models were developed to investigate transport and distributions of atmospheric pollutants, including the box model (such as PBM and AURORA) (Jin and Demerjian, 1993; Mensink et al., 2001), the Gaussian model (such as ADMS, AERMOD, and CALPUFF) (Abdul-Wahab et al., 2011; Carruthers et al., 1997; Cimorelli et al., 2005; Vieira de Melo et al., 2012), the Lagrangian/Eulerian model (such as GRAL and GATOR) (Knoth and Wolke, 1998; Oettl et al., 2005), and the computational fluid dynamics model (such as MISKAM and MICRO-CALGRID) (Olesen et al., 2009; Stern and Yamartino, 2004). Among them, California puff (CALPUFF) is a multilayer, multispecies, non-steady-state, Gaussian puff dispersion model that can simulate the effects of time- and space-varying meteorological conditions on air pollution transport, transformation, and removal (Alessandrini et al., 2010). One of the main applications of CALPUFF simulation systems is the prediction of the spatial distribution for air pollutants, which is crucial for quantifying health risks at different exposure locations (Qin et al., 2008). Such a simulation system was recommended by the U.S. Environmental Protection Agency (EPA) for long-range transport regulatory modeling and, on a case-by-case basis, for near-field applications (EPA, 2008). Recently, CALPUFF has been widely applied to air quality management and health risk assessments for cities and countries such as Istanbul (Elbir et al., 2010), Beijing (Hao et al., 2007; Song et al., 2006; Zhou et al., 2003), and Thailand (Sakulniyomporn et al., 2011).

These works revealed that CALPUFF could provide reasonably predictions for the patterns of long-term air pollutant deposition in near fields that were caused be emissions from a discrete source in complex terrain. However, it is computationally expensive, especially when a large number of runs is required for conducting further analysis. One concern of computational efficiency motivated researchers to seek advanced methods for alleviating the huge computational efforts (He et al., 2008a). One solution approach is to create effective surrogates to replace the complex CALPUFF simulation system. Peralta et al. (1991) and Aly and Peralta (1999) employed a first-order approximation to replace simulation models formulating a linear optimization problem. However, because such linear formulations were incapable of reflecting the complex nonlinear relationships, further studies on exploring surrogates with high approximation accuracy were developed For example, Huang et al. (2003) introduced an integrated simulation–optimization approach using a dual-response surface method for solving relevant problems in a simpler and quicker manner. Lall et al. (2006) used a locally weighted polynomial regression surrogate to approximate the arbitrary nonlinear relationship between inputs and outputs. He et al. (2008b) employed an artificial neural network to link the time-consuming fuzzy-simulation process with an optimization model. These research efforts indicated that the surrogates not only could improve computational efficiency, but also could be practical in simulation and optimization processes without introducing significant errors. Among these surrogates, support vector regression (SVR) was regarded as one of the perfect candidates for universal regression purposes (Thissen et al., 2004). Compared to artificial neural networks, SVR has the advantage of producing a global optimization model that is capable of efficiently dealing with high-dimensional input vectors (Vapnik, 1999). Moreover, it can adhere to the principle of structural risk minimization seeking to minimize the upper bound of the generalization error, rather than to minimize the training error (Smola and Schölkopf, 2004). Recently, the SVR model has been frequently used in the environmental field (Dai et al., 2011; Lin et al., 2009; Lu and Wang, 2005; Noori et al., 2009a; Noori et al., 2009b).

At the same time, many methods were developed for supporting the management of coal blending, atmospheric pollutant emission reduction, and the corresponding health risks (Chakraborty and Chakraborty, 2012; Chen and Liao, 2006; Guo et al., 2009; Lewis et al., 2012; Liao and Ling, 2003; Liao et al., 2012; Liu, 2008; Liu and Sherali, 2000; Sakulniyomporn et al., 2011; Schuhmacher et al., 2004; Senior et al., 2013; Tan et al., 2010; Tian et al., 2012). For example, Zemba et al. (1996) introduced a multiple-pathway health risk assessment method to evaluate the possibilities of adverse effects to human health that may result from contaminant releases from air pollution. Schuhmacher et al. (2001) used a Monte Carlo simulation technique to obtain variability and uncertainty propagation in the health risks due to air pollutants for the residents living in the surroundings of a municipal solid waste incinerator. Erarslan et al. (2001) developed a linear programming model to identify optimum coal blending patterns in terms of both quality and quantity. Schuhmacher et al. (2004) investigated the health risks due to combustor emissions in the cement manufacturing for the population living in the neighborhood of a cement kiln in Catalonia, Spain, where the ISC3-ST model was applied to estimate air dispersion of the pollutants emitted by the cement plant. Liao et al. (2005) utilized the genetic algorithm (GA) to optimize coal blending schemes in a power plant. Liu (2008) proposed an integer programming model to identify optimal blending and distribution decisions for coals in many domestic power plants. Based on GA, Chakraborty and Chakraborty (2012) developed a multiple-objective programming model to identify the blending plans of multiple-grade coals to satisfy the requirements of the end users with desired specifications.

Though many simulation and optimization modeling efforts have been undertaken for dealing with issues of coal blending, atmospheric pollution dispersion, and health risk management, several shortcomings need to be remedied. One of the major shortcomings remaining for these studies is that no inherent linkages exist among them. This leads to several problems. First, in terms of the simulation modeling, they focused on analysis and evaluation of pollutant concentration without any mitigation actions for remedying the emissions of pollution when necessary. At the same time, those studies focusing on optimization of coal-blending schemes could not evaluate the effectiveness of the generated decision alternatives within a real-time context. Moreover, such simulation and optimization efforts were scarcely connected with risk management actions and thus could not successfully reflect trade-offs among coal-clean technologies, power generation benefits, and human health risks. Decision makers or managers thus could not adjust their strategies to maintain a balance between economic benefits and human health risks.

Therefore, the objective of this research is to develop a simulation-aided nonlinear programming model (SNPM). The simulation effort (i.e., CALPUFF) is to forecast the fate of air pollutants for quantifying the health risk under various conditions, while the optimization studies are to identify the optimal coal blending strategies from a number of alternatives. Due to the difficulty in searching for the optimal solutions of SNPM, an indirect search approach will be proposed. In the approach, the complex CALPUFF simulation system will then be replaced with a rapid-response SVR simulator, and subsequently the optimal solutions will be found via GA. The objective entails the following tasks: (i) carrying out the CALPUFF simulation system for generating a number of statistical samples, (ii) using SVR to establish a set of surrogates for providing a bridge between coal blending schemes and the associated health risk, (iii) advancing a nonlinear optimization method through incorporating the surrogates into the optimization framework, and (iv) applying the SNPM to manage health risks and coal blending in the west of Beijing for demonstration.

Model Development

Figure 1 presents the framework of the study method. Specifically, the proposed SNPM contains three modules: simulation and risk assessment, statistical analysis, and risk management and optimization. First, simulation (i.e., through CALPUFF) and health risk assessment (i.e., through eq 4a) is performed to predict concentrations of air pollutants and the resulting noncarcinogenic risk levels, respectively. Subsequent, numerical experiments are conducted to select and identify statistical samples comprising explanatory and response variables representing coal blending schemes and hazard quotient of air pollutants, respectively. The statistical samples are obtained through computer-assisted randomly sampling. Then, a set of surrogates is created through adopting SVR to capture the relations between coal blending schemes and health risk levels. With the obtained surrogates, health risks are estimated in terms of the outputs of the surrogates, and then the SNPM will be solved through the genetic algorithm.

Figure 1. Framework of the study method.

Numerical simulation

The CALPUFF modeling system contains a meteorological processor (CALMET), a dispersion model (CALPUFF), and a postprocessor (CALPOST). Among them, CALMET is a diagnostic wind field model with multiple micro-meteorological modules for overwater and overland boundary layers (Scire et al., 2000a). Comparatively, CALPOST is a postprocessing program with options for computation of time-averaged concentrations and deposition fluxes predicted by CALPUFF. Generally, CALPUFF can simulate continuous plumes from a source as a series of discrete puffs (i.e., packets of pollutants) that are transported and dispersed through a three-dimensional wind and micro-meteorological field. Through the adoption of CALPUFF, a snapshot approach is used to evaluate the contribution of a puff to the total pollutant-concentration at a receptor. At the particular time intervals (sampling steps), the concentration of each puff is calculated. The puff can then move and change in size until the next sampling step (Ranzato et al., 2012). The total concentration at a receptor is the sum of the contributions of all nearby puffs averaged for all sampling steps within the basic time step (Scire et al., 2000b). In total, the basic equation for the contribution of a puff at a receptor can be presented as follows (Scire et al., 2000b; Vieira de Melo et al., 2012):

1\rm a

1\rm b

where C is the ground-level concentration (g/m³); Q_p is the pollutant mass in the puff (g); σ_x is the standard deviation of the Gaussian distribution in the along-wind direction (m); σ_y is the standard deviation of the Gaussian distribution in the cross-wind direction (m); σ_z is the standard deviation of the Gaussian distribution in the vertical direction (m); d_a is the distance from the puff center to the receptor in the along-wind direction (m); d_c is the distance from the puff center to the receptor in the cross-wind direction (m); g is the vertical term of the Gaussian equation (m); H is the effective height above the ground of the puff center(m); and h is the mixed-layer height (m). The summation in the vertical term (i.e., the term g) accounts for multiple reflections off the mixing lid and the ground. It reduces to the uniformly mixed limit of 1/h for σ_z > 1.6 h. Puffs within the convective boundary layer meet this criterion within several hours after release. For a horizontally symmetric puff, with σ_x = σ_y, the equation can be reformulated as follows:

2

where R is the distance from the center of the puff to the receptor (m) and s is the distance travelled by the puff (m). The total concentration at a receptor is the sum of the contributions of all nearby puffs averaged for all sampling steps within the basic time step, which can be shown as follows:

3

Details of CALPUFF were introduced by Scire et al. (2000a) and Scire et al. (2000b).

Simulation-based optimization for health risk management

Consider a health-risk-based coal-blending management system that is comprised of three components, including (i) a coal blending sector that needs to mix low- and high-grade coal to meet fuel requirements and reduce pollution emissions and economic costs, (ii) a power generation sector that burns the blending coal to generate electricity and discharge atmospheric pollution, and (iii) a receptor sector where the discharged air pollutants diffuse and deposit into the receptor zones, resulting in the potential health risk of relevant residents. A decision maker is responsible for reducing the ground-level concentration of air pollution and protecting human health through identifying coal blending options and thus puts restrictions on the corresponding risk. Therefore, in this research, the total hazard quotient is used as the objective function of the optimization system for facilitating the health risk management.

At the same time, the focus is put on noncancerous aspects of health impacts that the air contaminated by power plants poses and/or may pose in the future to human beings. Previously, hazard quotient (HQ) was widely used to estimate the noncancerous health risk (Li et al., 2007). For the individual living in the lth exposure location, HQ caused by the kth air pollutant (i.e., ) can be calculated as follows (Liao and Ling, 2003):

4\rm a

where i is the name of component coals, i = 1, 2, …, I; j is the name of power plants, j = 1, 2, …, J; k is the species of air pollutants, k = 1, 2, …, K; l is the name of exposure locations, l = 1, 2, …, L; is the hazard quotient of air pollutant k at exposure location l; is the inhalation rate (m³/day); is exposure frequency (days/yr); is the average exposure duration (years); is a reference dose of air pollutant k (mg/kg-day); is a body weight index at exposure location l; is average body weight of the people at exposure location l (kg); is averaging time for noncarcinogens (days); is the daily amount of component coal i burned by power plant j (tonnes/day); is the ground-level concentration of air pollutant k at exposure location l (mg/m³), which can be computed through the CALPUFF modeling system; denotes the functional relationships between and , which is because (i) the emission rate of air pollutant k in power plant j (i.e., ) is the input parameter of CALPUFF modeling system, which can influence the calculation results of ground-level concentration, and (ii) is computed by the equation (i.e., , where is the emission factor of component coal i for air pollutant k [kg/tonne], and is the removal rate of air pollutant k in power plant j by using pollution-control technologies [%]), and the value of can affect the value of ; is used to compute emission rate Q_jk, which is one of the CALPUFF inputs; and is the CALPUFF output. Thus, the functional relationships between and (i.e., ) can be simulated by the CALPUFF modeling system.

In eq 4a, HQ mainly considers the individual health risk but cannot reflect the population factor of exposure locations. In fact, the larger the population there is in an exposure location, the greater will be the health risk generated by air pollutants. Therefore, for estimating the population health risk of the study system, a population-weighted hazard quotient (PHQ) caused by kth air pollutants (i.e., ) can be formulated as follows:

4\rm b

where is the population-weighted hazard quotient for air pollutant k; is the population of exposure location l; and is the total population of all the exposure locations. Based on the assumption of dose addition (EPA, 1986), a total hazard quotient of a chemical mixture (i.e., ) can be estimated by the following equation:

4\rm c

The objective function should be minimized and subject to a number of environmental, economic, coal quality, and technical constraints. The environmental constraints require that the concentration of air pollutants emitted from power plants should be less than or equal to a regulated environmental target. For the economic constraints, the revenues from power selling should be more than or equal to the costs, including coal purchasing costs and emission reduction costs of pollution-control technologies. The coal quality constraints require that the specific features (i.e., volatile matter content, heat rate, ash content, moisture content, and sulfur content) of the blended coals should be limited within a regulated range for meeting the quality specification of coal used by boilers. In the coal quality features, volatile matter content refers to the components of coal, exclusive of moisture, that are given off by a material as gas or vapor at high temperature in the absence of air. In general, coal with a higher volatile matter content may be more flammable. Additionally, the amount of the blended coals transported into power plants should be equal to its demands. This constraint is proposed to maintain a supply–demand balance and meet local demand for electricity. Therefore, the simulation-aided nonlinear programming model (SNPM) can be formulated as follows:

5\rm a

subject to

5\rm b

5\rm c

5\rm d

5\rm e

5\rm f

5\rm g

5\rm h

5\rm i

where is the decision variable, which means the daily amount of component coal i fired by power plant j (tonnes/day); is the emission factor of component coal i to air pollutant k (kg/tonne); is the removal rate of air pollutant k in the power plant j by taking pollutant-control measures (%); is the emission allowance of pollutant k (kg/day); is the emission reduction cost of pollution-control measures for reducing pollutant k in the power plant j (RMB¥/kg); is the coefficient of component coal i to produce electricity (kWh/tonne); is the unit price of electricity (RMB¥/kWh); is the procurement cost of component coal i (RMB¥/tonne); , , , , and are the lower bound of coal quality (i.e., volatile matter content (%), heat rate (GJ/tonne), ash content (%), moisture content (%), and sulfur content (%), respectively) for meeting the operation allowance of the jth power plant; , , , , and are the upper bound of coal quality (i.e., volatile matter content (%), heat rate (GJ/tonne), ash content (%), moisture content (%), and sulfur content (%), respectively) for meeting the operation allowance of the jth power plant; , , , , and are the ith component coal’s quality (i.e., volatile matter content (%), heat rate (GJ/tonne), ash content (%), moisture content (%), and sulfur content (%), respectively); is the number of days in a year (days/year); and is the annual coal demand of the jth power plant (tonnes/year).

Solution method

An indirect search method is proposed to solve the SNPM (i.e., model 5). First, the Latin hypercube sampling approach (Huntington and Lyrintzis, 1998; Iman et al., 1981) is implemented to produce a number of training and testing samples, with each one having the inputs of the amount of component coal fired (i.e., ) and the outputs of the population hazard quotient (i.e., ). Then a surrogate simulator is developed via SVR model to reproduce the relationships between the inputs and outputs. GA is finally employed to seek the optimal solutions of SNPM. Figure 2 shows the flowchart of the procedures for solving SNPM. The specific procedures can be described as follows:

Figure 2. Flowchart of the solution method.

• Step 1: Input the geophysical data (e.g., terrain elevations and land cover) and routinely available meteorological data (e.g., hourly surface observations) into the CALMET, and run CALMET.

• Step 2: Use Latin hypercube sampling method to generate a set of sample () subject to , where 10 equally probable intervals are selected for each variable . Moreover, according to the equation , the corresponding emission rate of kth air pollutant in jth power plant (i.e., ) can be calculated.

• Step 3: Run CALPUFF dispersion model, where the input data include CALMET outputs and the emission inventory of power plants (e.g., , stack height, stack diameter, stack base elevation, and exit temperature and velocity).

• Step 4: Input the outputs of CALPUFF into CALPOST, and run CALPOST for extracting the simulation results of the ground-level concentration of kth air pollutant at lth exposure location (i.e., ).

• Step 5: Calculate the population hazard quotient of kth air pollutant (i.e., ) according to eq 4b.

• Step 6: Repeat steps 2 to 5 N times, where N is the sum of training and testing samples.

• Step 7: Train and test the accuracy of the surrogate simulator (i.e., SVR) based on the generated N samples, where (explanatory vector) and (response vector) are the inputs and outputs of SVR, respectively. In SVR, the purpose of train and test is to set the regularization constant C, precision parameter ϵ, and the kernel parameters (Cherkassky and Ma, 2004). The optimal values of parameters C and ϵ are determined by employing a grid search in a n-fold cross-validation approach (Kulkarni et al., 2004). Different kernel functions require tuning the different kernel parameters. The most used kernel functions contains linear kernel, polynomial kernel, radial basis function kernel, and multilayer perception kernel (Campbell, 2002; Tzeng, 2002). Selecting the kernel function parameters is usually based on application-domain knowledge and may reflect distribution of input values of the training data (Chapelle and Vapnik, 1999; Vapnik, 1999).Section S1 of the supplementary material provides the methodology of SVR.

• Step 8: GA is employed to seek the global optimization solutions of SNPM based on the created surrogate simulators. GA mainly includes four components: selection, crossover, mutation, and adaptive evaluation. In GA, we should set four main parameters, including the probability mutation and crossover operations, population size, and generation number. For details regarding GA refer to Goldberg (1989) and Davis (1991).

• Step 9: Obtain the optimal coal blending schemes under the objective for minimizing the total human health risk.

Case Study

The modeling domain (39º44^′38^′′–40º06^′31^′′ N, 115º54^′02^′′–116º22^′05^′′ E) is located in the west of Beijing, covering four urban districts (i.e., Haidian, Fengtai, Xicheng, and Shijingshan) and two suburban districts (i.e., Mentougou and Fangsha). Figure 3 shows the overview of the modeling domain, which is arranged as a 40 × 40 grid with a 1-km horizontal resolution. Ten vertical layers were set up below 2,000 m, and the lowest layer was at 10 m. Shuttle radar topography mission (SRTM) 3’s were used to collect terrain data. The northwest of modeling domain is bordered by Yan Mountain, with a maximum elevation of over 200 m above sea level. The main area of the domain is located within the Northern China Plain, with elevations of no more than 60 m. The landcover data were generated by the moderate resolution imaging spectroradiometer (MODIS) on the Terra satellite. National Centers for Environmental Prediction final analysis data (NCEP-FNL) and meteorological observation data from four surface stations located in the modeling domain were also used. The time range of these data covers the full year of 2008 (1 January to 31 December). The roughness lengths were 1.6 m for urban surfaces of Beijing, obtained from recent measurements (Gao and Bian, 2004). Input parameters of CALPUFF, such as chemical parameters for dry deposition of gases and miscellaneous dry deposition parameters, were based on the default values recommended by the EPA.

Figure 3. The modeling domain.

Beijing has implemented a project for upgrading boilers in power plants from coal-based to natural-gas-based ones in order to protect the atmospheric environment since the 2008 Olympic Games. Recently, two coal-fired power plants (i.e., Gaojing and Shijingshan power plants) are considered as the two major point pollution sources that are causing harmful effects on local residents (BMBS, 2012). Table 1 presents the parameters of the two plants (NAAQS, 2013). These parameters include the stack and emission information, fuel quality requirement, emission reduction measure, and the related efficiency and cost. The source of data for related parameters contains Beijing Statistical Yearbook, 2012 annual summary report of Gaojing and Shijingshan power plants, and the investigation of power plant workers. Three pollutants (SO₂, NO_x, and PM₁₀) are harmful for the health of residents. Moreover, the two power plants have taken measures for reducing the emission of pollutants since 2000. For example, the Gaojing power plant has adopted a lime spray dryer system to absorb SO₂ in the scrubber through a fine spray of slaked lime with a 95.9% removal efficiency. Shijingshan power plant has taken a selective catalytic reduction measure for the reduction of NO_x emissions, and this measure utilizes a catalyst to increase the NO_x removal efficiency at low temperatures. The electrostatic precipitator technology employs electrical forces to separate suspended particles from the flue gas stream. Thus, Gaojing and Shijingshan power plants have used it to reduce PM₁₀ emissions. The emission reduction cost in this research did not contain the building cost. It mainly denoted operational and material costs. For example, through using a lime spray dryer system (LDS) to reduce SO₂, we merely considered the operational and limestone purchase costs. In terms of selective catalytic reduction (SCR), the operational and catalyst purchasing costs constituted the expenses for emission reduction. Furthermore, coal quality is essential to support the normal work of the steam boiler of power plants. For instance, the Gaojing power plant required that the volatile matter of fuels must be larger than 6.5% and less than 27%, heat rate must be larger than 26 GJ/tonne, ash content must be less than 20%, moisture content must be less than 5%, and sulfur content must be less than 1.2%. In general, the efficiency of the two power plants in producing electricity was 33.5%.

Table 1. Parameters of power plants

	Gaojing power plant (j = 1)	Shijingshan power plant (j = 2)
Exhaust gas and stack
Stack height (m)	120	60
Stack diameter (m)	6	5
Stack base elevation (m)	108	102
Building downwash flag	No	No
Exit temperature (K)	360	316
Exit velocity (m/sec)	18	12
Fuel quality requirement
Volatile matter (% weight)	> 6.5 and < 27	> 9 and < 40
Heat rate (GJ/tonne)	> 26	> 24
Ash content (% weight)	< 20	< 24
Moisture content (% weight)	< 5	< 5.5
Sulfur content (% weight)	< 1.2	< 1.0
Emission reduction measure
For SO2 (k = 1)	LDS	LDS
For NOx (k = 2)	SCR	SCR
For PM10 (k = 3)	EP and FF	CDE and EP
Emission reduction efficiency
For SO2 (k = 1) (%)	95.9	95.7
For NOx (k = 2) (%)	84.3	85.4
For PM10 (k = 3) (%)	98.5	96.0
Emission reduction cost
For SO2 (k = 1) (RMB¥/kg)	1.8	2.1
For NOx (k = 2) (RMB¥/kg)	14.2	15.3
For PM10 (k = 3) (RMB¥/kg)	3.1	2.1

Notes: LDS, lime spray dryer system; SCR, selective catalytic reduction; EP, electrostatic precipitator; CDE, cyclone dust extractor; FF, fabric filter.

Coal blending is one of several available methods for reducing pollutant emissions by coal-fired power plants. To decide how to blend, it is important to understand the composition of component coals that are to be blended. Table 2 shows the parameters of component coals, indicating combustion properties and behavior of the coals. The data for related parameters were collected form 2012 annual summary report of Gaojing and Shijingshan power plants. Four types of coals (i.e., Jincheng, Shenhua, Datong, and Shenfu coals) were considered as the raw materials for blending. Jincheng coal has the lowest volatile matter (6.17%), a heat rate (26.88 GJ/tonne) at the second lowest level, the highest ash content (20.16%), a moisture content (4.62%) at the second highest level, and sulfur content of 0.61%, which is just below that of the Datong coal. Table 2 also presents the coal consumption rate, procurement cost, and emission factors for different pollutants (NAAQS, 2013). According to the data provided by the power plants, in order to generate a unit (i.e., 1 kWh) of electricity, 0.52 and 0.33 kg of Shenhua and Shenfu coals would be required. Also, emission factors were adopted to reflect the average amount of a specific pollutant discharging into the atmosphere after combustion of the component coals. The emission factor of Shenhua coal was the highest among the four types of coals, which implies that it would generate the most air pollutants if the same amount of coal were burned. The procurement costs of different types of coal were directly proportional to their corresponding heat rates. Specifically, the order of procurement cost was as follows: Shenhua coal < Jincheng coal < Datong coal < Shenfu coal. Comparatively, the order of heat rate was as follows: Shenhua coal < Jincheng coal < Datong coal < Shenfu coal.

Table 2. Parameters of component coals

Name of component coals	Jincheng coal (i = 1)	Shenhua coal (i = 2)	Datong coal (i = 3)	Shenfu coal (i = 4)
Fuel characteristics
Volatile matter (% weight)	6.17	38.9	26.06	6.83
Heat rate (GJ/tonne)	26.88	23.76	28.75	32 .32
Ash content (% weight)	20.16	6.13	14.42	3.96
Moisture content (% weight)	4.62	5.68	2.5	3.1
Sulfur content (% weight)	0.61	0.28	1.35	0.13
Emission factor
SO2 (k = 1) (kg/tonne)	8.46	17.4	5.67	2.7
NOx (k = 2) (kg/tonne)	6.58	9.30	2.96	2.5
PM10 (k = 3) (kg/tonne)	0.87	1.90	0.10	0.08
Others
Coal consumption (kg/kWh)	0.40	0.45	0.37	0.33
Procurement cost (RMB¥/tonne)	700	620	750	840

In total, 73 exposure locations (as shown in Figure 1) were considered in this research, mostly covering the six districts. Table 3 presents the 30 exposure locations in Haidian district (BAD, 2013). The population at the exposure locations varies from 0.87 to 22 × 10³, while the elevations fluctuate between 44 and 70 m. The exposure duration was defined as the exposure frequency of 252 days/year for 30 years, and an averaging time of 10,950 days was used in characterizing noncancerous risk (Chen and Liao, 2006). The body weight index (i.e., parameter ) was normalized to account for extrapolation to a different body weight from the standard of 70 kg (Liao and Chiang, 2006). Let the body weight equals 60 kg and inhalation rate equals 18 m³/day, according to Liao and Ling (2003). Reference dose (RfD) is defined as the maximum acceptable oral dose of a toxic substance (EPA, 1999). According to the EPA Integrated Risk Information System (IRIS) database, the inhalation RfD of SO₂, NO_x, and PM₁₀ should be estimated by the American ambient air quality standard if the study system is located in the United States. Analogically, for calculating the health risk of people living in the west of Beijing, the Chinese ambient air quality standard can be used to estimate the inhalation RfDs. And in detail, referring to Chinese ambient air quality standard (i.e., GB 3095-2012), the 24-hr average concentration limits for SO₂, NO_x, and PM₁₀ are 60, 40, and 70 μg/m³, respectively. Assume the average body weight is 60 kg; thus, the inhalation RfDs of SO₂, NO_x, and PM₁₀ are derived as 1.00, 0.67, and 1.17 μg/(kg-m³). Moreover, the EPA defines acceptable risk levels for noncarcinogen risk as lower than 1.0.

Table 3. Exposure locations in Haidian district

Exposure locations	Population (× 10^3)	Number of communities	X-direction distance (km)	Y-direction distance (km)	Elevation (m)
L1	17.56	36	12.6	−3.2	61
L2	3.92	16	10.4	−1.9	62
L3	10.91	38	14.0	−1.8	59
L4	11.38	21	16.0	−1.2	54
L5	10.34	35	14.1	0.1	60
L6	12.07	23	15.1	1.5	58
L7	12.43	35	17.5	2.9	57
L8	14.3	39	19.7	2.8	58
L9	17.58	29	16.2	5.6	59
L10	20	33	17.2	4.6	58
L11	10.2	33	14.8	5.0	58
L12	6.64	24	14.9	8.1	48
L13	2.79	9	16.8	8.2	54
L14	1.8	7	13.9	7.0	52
L15	0.87	6	7.4	7.2	69
L16	6.77	28	17.5	11.9	47
L17	12.17	36	18.7	5.5	54
L18	6.6	16	12.0	10.7	54
L19	5.82	30	18.3	11.0	45
L20	10.53	28	10.8	−0.1	60
L21	6.28	14	15.2	11.5	51
L22	1.77	6	11.8	4.7	54
L23	1.29	8	18.4	9.0	44
L24	11.46	18	12.9	2.4	55
L25	3.16	17	3.2	11.5	82
L26	7.31	24	9.3	2.8	57
L27	3.4	26	11.8	13.2	48
L28	3.21	23	1.8	16.1	50
L29	2.18	21	7.3	18.5	44
L30	22	54	15.5	15.8	47

Result Analysis

In total, 200 iterations were prepared by iteratively running the CALPUFF and calculating the PHQs. Figure S1 of the supplementary material shows the 200 samples of (i = 1, 2; and j = 1, 2, 3, and 4), through adopting the Latin hypercube sampling method. The sample was used to compute emission rate, which is a CALPUFF input. Through running CALPUFF, we could obtain 200 samples of ground-level concentration of air pollutants. Figure 4 presents the statistic results, such as mean and standard deviation, for 200 such samples. The peak-concentration values of SO₂ would range from 0.6 to 3.8 μg/m³ with a mean of 2.3 μg/m³ and standard deviation of 1.5 μg/m³. For NO_x, the peak-concentration values would vary from 1.5 to 8.1 μg/m³ with a mean of 4.8 μg/m³ and standard deviation of 3.1 μg/m³. With a mean (i.e., 0.18 μg/m³) and standard deviation (i.e., 0.10 μg/m³), the peak-concentration values of PM₁₀ would change from 0.06 to 0.30 μg/m³. The results also indicate that the mean level of ground-level concentration of NO_x would be the highest. It is mainly because the removal efficiency of NO_x was the lowest, which may cause the highest emission rate of NO_x in power plants. Figure S2 of the supplementary material presents the outputs of SVR models, which are the values of PHQs calculated by CALPUFF and the corresponding hazard quotient equation. Figure 5 shows statistic results, including the mean and standard deviation of PHQs. The mean value of PHQ generated by NO_x would be 3 times larger than the one by SO₂, and 75 times larger than the one by PM₁₀. It is indicated that NO_x would be the most serious air pollutant harming the public health in the vicinity of the two power plants. Moreover, the standard deviation value of PHQ caused by SO₂ would be 17 times larger than the one by PM₁₀, but approximately one-third of the one caused by NO_x.

Figure 4. Mean and standard deviation for the ground-level concentration of (a) SO₂, (b) NO_x, and (c) PM₁₀ from Latin hypercube sampling (unit: μg/m³).

Figure 5. Mean and standard deviation of PHQs caused by SO₂, NO_x and PM₁₀.

In this research, these 200 samples were utilized to produce 3 surrogate simulators (i.e. SVR models). Specifically, SVR-SO₂, SVR-NO_x, and SVR-PM₁₀ represented the functional relationships between coal blending operating conditions and the value of PHQ caused by SO₂, NO_x and PM₁₀, respectively. Moreover, 150 of 200 samples were applied to train the surrogate simulators, and the additional 50 testing samples were used to verify their prediction performance. To examine the performance of SVR models for both of these training and testing sets, three statistic indexes, fractional bias (FB), correlation coefficient (CC), and Willmott’s index of agreement (WI) were considered as follows:

6\rm a

6\rm b

6\rm c

where is the value of samples prepared by iteratively running the CALPUFF and calculating the PHQs, and is the results of SVR models; the overbar refers to the average over the data set, and σ is the standard deviation over the data set. WI, ranging from 0 to 1, describes how well the SVR models fit the data. For example, a WI value of 1 indicates that all variability in the response variables can be interpreted by the SVR models, while a value of 0.4 indicates that approximately 60% of the variation in the response variables cannot be explained. Moreover, the values of the statistical measures for a perfect model would be 0 for FB and 1 for .

Figure 6 compares the calculation results of SVR with those through CALPUFF and hazard quotient equation, where the data points in each plot assemble in the left and right sides of the diagonal. Specifically, data points on the left of the diagonal indicate that SVR models overestimate the PHQs calculated through CALPUFF and hazard quotient equation; conversely, data points on the right of the diagonal illustrate that SVR models underestimate the PHQs calculated through CALPUFF and hazard quotient equation. In order to evaluate the performance of SVR, we could use a number of quantified statistical indexes, such as FB, CC, and WI. Figures 6a, 6c, and 6e show the training performance of SVR-SO₂, SVR-NO_x, and SVR-PM₁₀, respectively. In the training experiments, the linear kernel function was used to help SVR achieve the best fitting and generalization ability. Moreover, through using the grid-search on regularization constant C and precision parameter ϵ, the two parameters were searched in the range of C = 2p, p =- 10, -9, …, 0, …, 9, 10 and ϵ = 2q, q = -10, -9, ..., 0, …, 9, 10. All of the pairs (C, ϵ) were tried and the one with the best cross-validation accuracy could be selected. The FB, CC, and WI values of training SVR-SO₂ model would be 0.02, 0.86, and 0.87, with parameters C and ϵ being 2⁴ and 2¹, respectively. The FB, CC, and WI values of training SVR-NO_x model would be 0.01, 0.85, and 0.88, with parameters C and ϵ being 2² and 2⁸, respectively. The FB, CC, and WI values of training SVR-PM₁₀ model would be 0.03, 0.94, and 0.95, with parameters C and ϵ being 2⁻¹ and 2⁹, respectively. Because the values of CC and WI were greater than 0.85 and the value of FB was lower than 0.03, the three SVRs were considered strong and robust training in performance. Through verifying the testing sets, Figures 6b, 6d, and 6f show the predicting performance of SVR-SO₂, SVR-NO_x, and SVR-PM₁₀, respectively. The FB, CC, and WI values of testing SVR-SO₂ model would be 0.03, 0.80, and 0.82, respectively. The FB, CC, and WI values of testing SVR-NO_x model would be 0.03, 0.78, and 0.75, respectively. The FB, CC, and WI values of testing SVR-PM₁₀ model would be 0.05, 0.87, and 0.87, respectively. The results indicate that the training performance would be better than predicting performance. For example, for SVR-PM₁₀ model, the FB, CC, and WI values of predicting performance were 0.05 (>0.03, FB of training performance), 0.87 (<0.94, CC of training performance), and 0.87 (<0.95, WI of training performance). The results also indicate that the predicting accuracy of the three surrogate simulator should be acceptable since the values of CC and WI were greater than 0.7, and the value of FB was lower than 0.1. In order to examine the performance of SVRs for each sample, the relative errors of 200 samples are presented in Figure 4S of the supplementary material. The results indicate that the relative errors for most of the samples would be limited between −30% and 30%. For example, the relative errors for 96% of SVR-SO₂ training samples would vary between −30% and 30%. Results also indicated that the obtained SVR models would have satisfactory fitting and prediction performance.

Figure 6. Training and testing performance of SVRs.

The three surrogate simulators obtained were linked to the objective function of SNPM (i.e., model 5). The optimal solutions of model 5 were identified through GA. The probability mutation and crossover operations were determined to be 0.15 and 0.80, respectively. The population size and generation number were 40 and 500, respectively. Figure 7 presents the solution results of SNPM. These results illustrate the optimal coal blending schemes for the Gaojing and Shijingshan power plants. The four component coals allocated to the Gaojing power plant would be 852 (Jincheng coal), 1552 (Shenhua coal), 608 (Datong coal), and 827 (Shenfu coal) tonnes/day, respectively. Different from Gaojing, the Shijingshan power plant would require Jincheng coal at 724 tonnes/day, Shenhua coal at 3120 tonnes/day, Datong coal at 986 tonnes/day, and Shenfu coal at 1650 tonnes/day, respectively. Accordingly, the coal purchase cost for the two power plants would be 7.28 × 10⁶ RMB¥/day. Moreover, the solved value of the objective function would be 0.097, which indicates that the health risk caused by power plants could be controlled within a low level. Currently, the Gaojing and Shijingshan power plants, respectively, burn 3840 tonnes of Jincheng coal and 6480 tonnes of Datong coal every day. This coal burning scheme requires a daily coal purchase cost of 7.55 × 10⁶ RMB¥, but it may lead to a higher health risk level (i.e., 0.156). Compared to the current coal scheme, the optimal coal blending system designed can reduce the health risk caused by the two power plants by 38%, and save 270 × 10³ RMB¥/day in the coal purchase cost. Therefore, an appropriate coal blending scheme could not only reduce the human health risk, but also economize the coal procurement cost for power plants.

Figure 7. The optimal coal blending scheme for (a) the Gaojing power plant and (b) the Shijingshan power plant.

Discussion

In SNPM, CALPUFF was used to predict the ground-level concentration of air pollutants for quantifying the human health risks. Unlike the traditional Gaussian plume models, CALPUFF can simulate the continuous plume from a source as a series of discrete puffs (i.e., packets of pollutants), which is transported and dispersed through a three-dimensional wind and micro-meteorological field. Thus, it can provide reasonably accurate predictions of the patterns of long-term air pollutant deposition in the near field associated with emissions from a discrete source in the complex terrain. The developed SNPM was solved through an indirect search approach based on the combination of SVR and GA. In this research, SVR reproduced the functional relationships between coal-blending operating conditions and health risk, and GA could identify the optimal solutions of SNPM. Such an SVR surrogate has three major advantages over the existing ones: (i) it is capable of efficiently dealing with high-dimensional input vectors, and thus the number of explanatory variables is boundless and the time can be greatly saved for selecting the primary contributed explanatory variables; (ii) it adheres to the principle of structural risk minimization, seeking to minimize an upper bound of the generalization error, rather than minimize the training error; and (iii) compared to similar methods such as artificial neural networks, it requires fewer statistical samples for training without sacrifice of accuracy. The statistical samples for SVR training were obtained through a Latin hypercube sampling method, because it can achieve a significantly higher computational efficiency than the Monte Carlo-based sampling technique. The purpose of the indirect search approach was the reduction of the computational efforts in optimization processes. Because the call to the simulation model is the main component of SNPM, the computation time can be analyzed by mainly considering the central processing unit (CPU) time consumed in running the simulation model. In this research, one CALPUFF simulation on average requires 30 min of CPU time to check whether constraints 5(b) to 5(i) are satisfied or not. When assuming that the optimization model needs to be called for 1000 simulations to obtain the optimal solution, the nonsurrogate optimization method (i.e., the CALPUFF simulation model is called by the optimization algorithm) would require approximately 500 hr of CPU time. In contrast, in terms of the developed surrogates, approximately 100 simulations would be accomplished per hour. Thus, the optimal solution would be obtained within 10 hr based on the surrogate-based optimization method (i.e., the surrogates is called by the optimization algorithm). This is many orders of magnitude faster than those nonsurrogate optimization methods. In general, the indirect search approach provides a bridge between simulation and optimization problems. Also, it can pose as a proxy of the nonlinear interactions between many coal-blending operating conditions and the corresponding health risk quotients without too much computational burden in the nonlinear optimization procedure. On the other hand, scenario analysis is an effective tool through simply selecting the best performance of the scenarios that can be directly tested. Compared with SNPM, it has the following three disadvantages: (i) It has to predefine the scenarios for decision variables (i.e., x_ij), but some may be unavailable due to the violation of constraint conditions; (ii) it is a useful for generating the best solution among all the predefined scenarios, while SNPM can get the global optimal solution; and (iii) the computation cost of scenario analysis is unacceptable. In this research, assuming that 10 levels were predefined for each decision variable, 10⁸ simulations would be necessary to test the performances of all the scenarios. This may take approximately 5707 years (i.e., 10⁸ × 30 /(60 × 24 × 365) = 5707).

Due to the complication of health risk management system, there are many reasons for the modeler to be suspicious of model results. Uncertainties would be one of the most important reasons. In general, there are three types of uncertainties in SNPM. The first type is parameter uncertainties, which mean the deviations caused by empirical parameters in model. For example, the dry deposition parameters of gases in CALPUFF were estimated based on the default values recommended by the EPA. However, such parameters cannot completely represent the spatial heterogeneity and temporal variability in the dry deposition process. Thus, this may be one of the contributions to the prediction errors of the ground-level concentration of air pollutants. Accordingly, sensitivity analysis is an approach for analyzing the effects of parameter uncertainties on the simulation and optimization framework. The second type is input uncertainties, which denote the errors caused by the insufficient input data. For instance, NCEP-FNL and meteorological data of surface stations were used to drive CALMET. However, NCEP-FNL has low spatial resolution of 1º × 1º. Also, three of the four surface stations were not located in the complex terrain portion of the grid. Thus, this may generate deviations in simulating the orographic perturbation airflows. Although more than 75% of the exposure locations were in the flat terrain region, uncertainties may exist in forecasting the ground-level concentration of air pollutants at each of these exposure locations. Also, to run CALPUFF, the data on power plant operating characteristics (e.g., exhaust temperature, exhaust flow rate, and exit speed) are essential. The variations of operating characteristics affect the plume rise and pollutant transport, which may cause uncertainties in forecasting the ground-level concentration of air pollutants. To complicate matters, such uncertainties may translate into large uncertainties in the results of the optimization model. Accordingly, a main method to mitigate such input uncertainties is to improve the quality of input data. For instance, with more meteorological data in the complex terrain obtained by using the vehicle-mounted wind profile radar, better prediction accuracy of air quality modeling would be acquired. Also, stochastic analysis (e.g., Monte Carlo) is a useful tool for reflecting the input uncertainties when the probabilistic distribution information of input variables can be identified. Comparatively, if imprecise or vague information for input variables is available, fuzzy sets theory can be used as an alternative tool for handling input uncertainties. The third type is model structure uncertainties, which arise from imperfections in the integrated simulation-aided optimization model. In SNPM, the accuracy of optimization results mainly depends on SVR models that were used to reproduce the relationships between coal-blending operating conditions and health risk. Although SVRs have a good training and prediction performance, they cannot predict the PHQs without deviations. Thus, through solving SNPM based on SVRs, the generated optimal solutions might deviate from the true optima. Even if such prediction errors were minor, they might probably result in unreliable solutions. Accordingly, such model structure uncertainties can be mitigated through enhancing the numbers of the training and testing samples; these massive samples would be easily obtained through the use of parallel machines. Moreover, the prediction accuracy may benefit from improved SVR models, such as Simulated-annealing-based SVR (Lin et al., 2008; Pai and Hong, 2005). In general, uncertainties exist in a number of impact factors and pollution-related processes, such as characteristics of coals, emission conditions, dispersion of air pollutants, and health risk assessment, as well as the solution method. These uncertainties have an adverse impact on the optimal coal blending schemes. Uncertainty techniques, such as fuzzy sets theory, Monte Carlo, and sensitivity analysis, may be used for tackling this issue, while they are not discussed in the research. It is expected that SNPM would be also useful as it can integrate such uncertainty techniques within a general framework. This could be one of the most important topics in future studies based on the previous studies on uncertainty techniques.

Conclusion

In this research, the California puff (CALPUFF) modeling system was introduced into a nonlinear programming framework for supporting the management of health risks due to pollution emissions of power plants, leading to a simulation-aided nonlinear programming model (SNPM). In this research, CALPUFF was used to forecast the ground-level concentration of air pollutants for quantifying the human health risks. The developed optimization model was to identify the optimal coal blending strategies for reducing the health risk. Four special contributions make SNPM unique compared with the previous models that handled coal blending problems. First, the objective function of the SNPM is the minimization of the health risk quotient value for the public living in the vicinity of power plants, while the conventional coal blending optimization focuses on how to save the system cost. Thus, the proposed SNPM was conducive to protecting human health compared with previous studies. Second, SNPM is formulated by integrating the simulation model and optimization model, which not only can predict the concentration distribution of air pollutants, but also can reflect the interactive characteristics of the coal blending system. Third, SNPM provides a direct and rapid-response linkage between coal blending schemes and health risk through the created surrogates. Lastly, SNPM is solved through an indirect search approach based on the combination of support vector regression (SVR) and the genetic algorithm (GA), which can alleviate the computation cost in searching for optimal solutions.

To demonstrate effectiveness of the developed method, SNPM was applied to minimize the human health risk associated with air pollutants discharged from the Gaojing and Shijingshan power plants in the west of Beijing. In this area, the mean value of the ground-level concentration of NO_x was the highest, and thus NO_x was the most serious air pollutant harming the public health in the vicinity of the two power plants. Moreover, the predicting performance of the three surrogate simulators (i.e., SVRs) was acceptable, where correlation coefficient and Willmott’s index of agreement were not lower than 0.7 and fractional bias statistic was not large than 0.1. Furthermore, the solution results of SNPM illustrated the optimal coal blending schemes for the Gaojing and Shijingshan power plants. The four component coals allocated to the Gaojing power plant would be Jincheng coal at 852, Shenhua coal at 1552, Datong coal at 608, and Shenfu coal at 827 tonnes/day, respectively. Different from Gaojing, the Shijingshan power plant would require Jincheng coal at 724 tonnes/day, Shenhua coal at 3120 tonnes/day, Datong coal at 986 tonnes/day, and Shenfu coal at 1650 tonnes/day, respectively. Compared to the current coal scheme, the optimal coal blending system designed can reduce the health risk caused by the two power plants by 38%, and save 270 × 10³ RMB¥/day in the coal purchase cost.

This research is the first attempt for planning the health-risk-based coal blending management system through integrating the CALPUFF and optimization techniques. The results suggested that the developed method is an effective tool for supporting relevant decision making. However, the proposed SNPM has some limitations. First, to quantify human health risks, relevant pollutants can be characterized as carcinogens and noncarcinogens. The health risk assessment approach contains an excess lifetime cancer risk model for assessing carcinogenic risk, and a hazard quotient model for noncarcinogenic risk. This study focus on the noncarcinogens risk caused by SO₂, NO_x, and PM₁₀, which are the main air pollutants discharged from power plants. Among these, NO_x is not only one of the risk factors for lung disease, but also one of the important predecessors for ozone formation in the air. However, ozone-related health effects are hard to consider due to complex photochemical reactions and the shortage of CALPUFF. Moreover, heavy metals (i.e., arsenic) adsorbing on the surface of PM₁₀ may pose potential health risk. This is also not considered in this research because the emission factor of arsenic is hard to obtain. Second, SNPM is incapable of handling the uncertainties existing in the health risk management system, which may generate an adverse impact on the model results. Therefore, three challenges still need to be overcome in future studies: (i) monitoring more data with high accuracy for alleviating the uncertainties and calculating more comprehensive carcinogens and noncarcinogen risks, (ii) integrating the uncertainty techniques into SNPM for perfecting the modeling framework, and (iii) searching for the model solutions via SVR and GA techniques with high accuracy.

Funding

This research was supported by the National Science Foundation for Innovative Research Group (No. 51121003), and the special fund of State Key Lab of Water Environment Simulation (11Z01ESPCN).

Supplemental Material

Supplemental data for this article can be accessed on the publisher’s website.

Acknowledgment

The authors are grateful to the editors and the anonymous reviewers for their insightful comments and suggestion.

Additional information

Notes on contributors

C. Dai

C. Dai is a doctor and X.H. Cai and H.C. Guo are professors in the College of Environmental Science and Engineering at Peking University, Beijing, China.

Y.P. Cai

Y.P. Cai is a professor in the State Key Laboratory of Water Environment Simulation at Beijing Normal University, Beijing, China.

Q. Tan

Q. Tan and G.H. Huang are professors, and W. Sun is a doctor in the Institute for Energy, Environment and Sustainable Communities at University of Regina, Regina, Saskatchewan, Canada.