An integrated assessment tool to define effective air quality policies at regional scale

at SciVerse ScienceDirect

Environmental Modelling & Software 38 (2012) 306e315

Contents lists available

Environmental Modelling & Software

journal homepage: www.elsevier .com/locate/envsoft

An integrated assessment tool to define effective air quality policiesat regional scale

Claudio Carnevale a, Giovanna Finzi a, Enrico Pisoni a,*, Marialuisa Volta a, Giorgio Guariso b,Roberta Gianfreda c, Giuseppe Maffeis c, Philippe Thunis d, Les White e, Giuseppe Triacchini f

aDepartment of Information Engineering, University of Brescia, Via Branze 38, 25123 Brescia, ItalybDepartment of Electronics and Information, Politecnico di Milano, Piazza Leonardo Di Vinci 32, 20133 Milan, Italyc TerrAria s.r.l., via Melchiorre Gioia 132, Milan, Italyd European Commission-DG Joint Research Centre, Institute for Environment and Sustainability, I-21020 Ispra, Italye Les White Associates, 42 Blunts Wood Road, Haywards Heath, West Sussex RH16 1NB, United KingdomfDietary and Chemical Monitoring Unit, EFSA e European Food Safety Authority, Via Carlo Magno 1A, I e 43126 Parma, Italy

a r t i c l e i n f o

Article history:Received 22 July 2011Received in revised form15 May 2012Accepted 11 July 2012Available online 4 August 2012

Keywords:Integrated assessment modelingModel reductionAir quality modelingMulti-objective optimizationDecision support

* Corresponding author. Tel.: þ39 030 3715520; faxE-mail address: [email protected] (E. Piso

1364-8152/$ e see front matter � 2012 Elsevier Ltd.http://dx.doi.org/10.1016/j.envsoft.2012.07.004

a b s t r a c t

In this paper, the Integrated Assessment of air quality is dealt with at regional scale. First the paperdescribes the main challenges to tackle current air pollution control, including economic aspects. Then itproposes a novel approach to manage the problem, presenting its mathematical formalization anddescribing its practical implementation into the Regional Integrated Assessment Tool (RIAT). The mainfeatures of the software system are described and some preliminary results on a domain in NorthernItaly are illustrated. The novel features in RIAT are then compared to the state-of-the-art in integratedassessment of air quality, for example the ability to handle nonlinearities (instead of the usual linearapproach) and the multi-objective framework (alternative to cost-effectiveness and scenario analysis).Then the lessons learned during the RIAT implementation are discussed, focusing on the locality, flexi-bility and openness of the tool. Finally the areas for further development of air quality integratedassessment are highlighted, with a focus on sensitivity analysis, structural and non technical measures,and the application of parallel computing concepts.

� 2012 Elsevier Ltd. All rights reserved.

1. Introduction

Managing environmental quality at any level is always a difficulttask since it results from human actions on one side and from thenatural, unmanageable aspects, on the other side. The idea of‘quality’ itself, though intuitively clear, can be formulated in manydifferent ways when dealing with a pollution level which contin-uously varies in time and space. Possibly none of the currentapproaches fully addresses the needs of the many decisions makersworking in this policy area. The complexity of the problem (Janssenet al., 2009) further increases when dealing with secondarypollutants, that are formed in the atmosphere from complexchemical reactions and physical processes involving precursoremissions, namely VOC, NOx, NH3, primary PM and SO2. Their highconcentrations, that represent one of the most significant airpollution problems in many developed countries, cannot be

: þ39 030 380014.ni).

All rights reserved.

regulated directly, but must be tackled indirectly by definingreduction plans of precursor emissions (as required, for instance, byEU directive 2008/50) taking into account the cost of implementingthese policies (Reis et al., 2005).

Air pollution Integrated Assessment (IA) models are tools thatbring together data on pollutant sources (emission inventories),their contribution to atmospheric concentrations and humanexposure, with information on potential emission reductionmeasures and their respective implementation costs. At the EUscale, IA models have been developed in the recent years to providea technical base for intergovernmental negotiations in a structuredway. In the context of UNECE Convention on Long Range Trans-boundary Air Pollution (CLRTAP), the integrated assessment modelRAINS/GAINS (Wagner et al., 2007), developed by IIASA, has beenextensively used to determine cost-efficient policies to reduceemissions and achieve EU-wide target for various air quality indi-cators (e.g. acidification, eutrophication, tropospheric ozone,primary and secondary particulate).

Furthermore, these European scale models have been adaptedto the national scale (RAINS-Italy as in D’Elia et al. (2009), RAINS-

C. Carnevale et al. / Environmental Modelling & Software 38 (2012) 306e315 307

Netherlands as in Aben et al. (2003), FRES-Finland as in Syri et al.(2002), UK-IAM as in Oxley and ApSimon (2007), Belgium-IAM asin Deutsch et al. (2008)). Based on a similar approach to the onedeveloped at the European scale, these models can then be used tooptimize emission reductions within a given country.

At the local/urban scale few IA models (IAMs) have beendeveloped and applied (Mediavilla-Sahagun and ApSimon, 2003).However these IAMs have generally been used for non-reactivespecies, and so their application to suggest optimal policies toreduce secondary pollutants has some relevant limitations.

To address this issue, many air quality managers rely on scenarioanalysis utilizing complex multiphase air pollution models,1 tosimulate the effect of emission reduction scenarios on pollutionconcentration (Finzi et al., 2000; Cuvelier et al., 2002; Sokhi et al.,2006; Carnevale et al., 2008a). These multiphase models,provided sufficient data are available on the local meteorology andemissions, can indeed be used to estimate air quality at each pointin time and space, but they cannot solve more policy orientedproblems such as where to invest more efficiently, or to estimatethe cost required to achieve a given level of air quality. This is whythere is a need for novel methodologies and tools that can supportsecondary pollution control decisions at sub-national scale.

The aim of this paper is to present and discuss the RegionalIntegrated Assessment Tool (RIAT), tuned to local and regional scaleapplications, and to assess its potential as a decision-making aidshowing some preliminary results for a specific Italian domain. Themain goal of this tool is to identify the most efficient mix of localpolicies required to reduce tropospheric ozone and particulatematter, in compliance with National and International air qualityregulations (e.g. EU directives), while accounting for local pecu-liarities in terms of emissions, meteorology and technological,financial and social constraints. From the analysis of RIAT featuresand performances, the paper attempts to identify the most inter-esting directions for further development of integrated assessmentmodeling in the field of air quality.

2. The challenges of the air quality decision problem

As common tomany environmental fields, the definition itself ofthe decision problem is not an easy task: neither the objective(s),nor the actions (decisions), nor the constraints have a definite andgeneral form when dealing with regional air quality problems.

As to the objectives, it is clear that any action aimed atimproving air quality has a certain cost, that we want to minimize.There is extensive experience in such ‘cost optimized’ approaches(see for instance the results of GAINS project, as in Wagner et al.(2007)). However, as already noted, air quality varies in time andspace, as the result of the combined action of different pollutants,and ultimately impacts on humans, vegetation, ecosystems, andconstruction materials. Thus it is almost impossible to condense allthe information relevant for decision making into one or even fewvariables. Existing legislation currently defines a certain number ofrelevant air quality indicators (the average pollutant concentration,the maximum over 8 h, the sum of values exceeding a giventhreshold, etc) mainly filtering the temporal variations. This isconsistent with the idea that outdoor air quality (at presentsvalues) plays a significant role mainly in the long term, i.e. on thehorizon of the reduction plans, which is of the order of some years.It also means that the spatial distribution of air quality cannot beeasily summarized and the differences between different areas of

1 Multiphase models simulate the physicalechemical processes involvingsecondary pollutants (as ozone and PM) in the atmosphere, implementing gas-phase chemical and aerosol modules.

the same region (typically urban and country parts) are relevant forthe decision maker and should be clearly separated.

This concept is inextricably linked to the second item that mustbedefined:which actions, i.e. decisions, have tobe considered. Sinceair quality, particularly for secondary pollutants, is a regionalphenomenon, actions taken on a certain domain may impact airquality conditions tens of kilometers away. So a decision maker(DM) may be interested in checking the consequences over an areawhich is outside his/her control. Another important aspect is whatthe DM can actually decide. Most of the previous work in integratedair quality modeling has assumed that the overall emission reduc-tionof a certainmacrosector (say for instance, domestic heating) canbe reduced bya certain percentage (Pisoni et al., 2009), using the so-called ‘lumped’ approach. This approach tends to over-simplify theproblem (there is only one decision to be taken for eachmacrosectorand each pollutant), and also makes it difficult to estimate how toimplement in practice the requested reduction, i.e. which is theoptimal (least cost) mix of measures that can achieve the results.This said, in case of scarce data availability, the ‘lumped’ approachcan be a helpful way to assess a priority list of macrosector emissionreductions. However, defining the adoption of a specific reductiontechnology for each specific activity is amuchmore explicit decision(the decision maker can then impose the use of that technology fora specific process/emission) and facilitates the calculation of costs,with the clear disadvantage of increasing the number of decisionvariables by at least two orders of magnitude.

Adopting the last definition of decision variables, constraints areeasy to be defined. Clearly, costs should not exceed the budget, anylegal air quality standard must be satisfied and the technology canbe adopted at a percentage between a minimum (possibly, zero)and a maximum value (possibly, 100%) of the emission due to therelated activity. It is interesting to note in this respect that theminimum adoption (or penetration) value is set by the foreseentechnology application in a given year under Current LEgislation(CLE), and thus normally varies with time. For instance, the EURO Vstandard for light-duty vehicle emissions (operating since 2009)had a low penetration level in 2010 (one year after its introductionthe new standard will have influenced only a small part of thefleet), but will have a higher CLE in 2020 (because after 11 years,a significant amount of circulating cars will have been replaced).This means that the decision about what is the optimal rate ofadoption of EURO V can span a larger range for 2010 and a muchsmaller one in 2020. Additionally, due to this time varying low limitof the decision ranges, the decision maker needs to understand ifa certain solution that was optimal for 2010 conditions will still bethe best (or at least will not be too far from the optimum) also in thefollowing years. When this is not the case, some of the technologiespreviously adopted should be substituted to take into account thenew context. Some others on the contrary, may be too expensive toscrap. The RIAT system described in the following chaptersattempts to tackle the above problems in an effective way.

3. The RIAT approach overview

RIAT is an integrated modeling environment strongly linkingtabular and geographic data, simulation and optimization models,graphical and geographical user interface, focused on the local andregional spatial scales (see Fig. 1 for a schematic description of theRIAT system). It incorporates explicitly the specific features of thearea, i.e. the local meteorology, the detailed pattern of the emissionprecursor sources and the prevailing chemical regimes. Its partic-ularities are:

(1) A multi-objective (air quality vs emission reduction measurescosts) optimization problem (Carnevale et al., 2008b; Pisoni

Fig. 1. RIAT system simplified scheme.

C. Carnevale et al. / Environmental Modelling & Software 38 (2012) 306e315308

et al., 2009) is solved to select and present to the user the entireset of efficient abatement solutions. The decision variables ofsuch a problem are the emission abatement measures.

(2) A full 3D deterministic multi-phase modeling system(describing the non linear dynamics linking precursor emis-sions to air pollutant concentrations) cannot be embedded andrun in real time within the RIAT optimization procedurebecause of its computational requirements (Carnevale et al.,2009). Thus, a series of long term simulations is performedbeforehand with such a model. The results of the simulationsare then used to derive simpler relationships (Ratto et al., 2011;Castelletti et al., 2012) between emission sources and airquality indicators at given receptor sites (S/R models) whichcan then be used directly in the optimization algorithm. Arti-ficial Neural Networks (ANN) are used in RIAT to derive thesesourceereceptor relationships (Carnevale et al., 2012). Thisapproach has the following advantages:� Compared to the traditional approach used in many IAmodels based on linear regression analysis, ANNs capture thenon-linearities (Gabusi et al., 2008; Carnevale et al., 2010) inthe relationships between emissions and concentrations.

� ANNs represent area to grid source relationships, whereasother approaches mostly deliver country (or region) to gridrelationships.

� ANNs requirements in terms of input (i.e. the long-term 3Dsimulations performed beforehand) are significantlyreduced compared to the traditional linear regressionapproach.

(3) The potency of an emission reduction (change of concentrationper unit emission reduction) depends on the height at whichthis emission reduction occurs. Emission reductions made at ornear the surface (e.g. traffic) are generally more potent (healthimpact/unit of emission) than emission reductions made athigher release heights (e.g. power sector). This is especiallytrue in urban environments where low level emissions aregenerally located. When impacts on population exposure are

considered, these potency differences tend to increase (Thuniset al., 2010). RIAT facilitates the management of both low andhigh emission reduction separately, so correctly accounting forthe potency differences.

(4) Other particular features of the system include the possibilityfor the user to:� Define different air quality indexes;� Constrain the overall expenditure to a specific value;� Spatially visualize the improvement in air quality index andhence show their distribution on the territory;

� Adjust automatically to the foreseen development of legis-lation and technologies in time, adopting a specific classifi-cation of technologies to distinguish between those that canor cannot be replaced with time;

� Facilitate the exploration of the ‘policy envelope’ throughthe use of the Pareto curve providing a detailed view on allpossible efficient solutions ranked by costs.

The next section describes the RIAT formalization approach inmore details.

4. The RIAT approach formalization

Even if this paper is mainly devoted to present the RIAT meth-odology, from this Section onward, the approach is described,formalized and exemplified considering the case of Lombardyregion in Northern Italy. The geographical domain was discretizedwith a 6 � 6 km2 grid and comprises roughly 6000 cells (see Fig. 2).The area where decisions are assumed to be implemented (purplecells in Fig. 2) coincides with the areawhere air quality is evaluated,though RIAT may be used also when the two differ. Some prelim-inary results related to this test case will be shown and commentedin Section 6.

RIAT implements and solves a multi-objective problem, tosupport the selection of effective policies to control population

Fig. 2. Lombardy region domain: cells in purple are the cells considered in the test application presented in this paper. (For interpretation of the references to colour in this figurelegend, the reader is referred to the web version of this article.)

C. Carnevale et al. / Environmental Modelling & Software 38 (2012) 306e315 309

exposure to primary and secondary pollutants. To do so, the systemaccesses information from various data bases:

(1) Current and prospective emission reduction technologies andrelated costs (in the illustrated application, these are derivedfrom GAINS (Wagner et al., 2007)).

(2) Regional activities and emission data (in the following applica-tion, these are derived from the INEMAR inventory (Dommenet al., 2003), the official emission inventory of Lombardy region).

(3) Source-receptor models, developed for the specific regionalenvironment (a surrogate of the much larger and detailed, butalso time demanding, chemical transport model (Carnevaleet al., 2012)).

These data, together with all the CORINAIR activities that releaseprecursor emissions, classified into macrosector-sector-activitytriples defined from here onward as activity i, constitute theinternal database of the system. Processing this informationaccording to the scheme in Fig. 1, the following user requests can beanswered:

(1) Scenario evaluation approach: the user defines the set ofemission reduction measures to be used.

(2) Multi-objective approach: the user requests the system to solvea multi-objective problem:� Computing the entire Pareto boundary;� Computing a single efficient solution for a specified cost target.

The system reports the optimal (or required) levels of adoptionof each technology, the correspondent emission reductions, the airquality indexes (through S/R models) and the costs (throughabatement measure DB).

4.1. The decision problem

The solutions of the multi-objective problem are the efficientemission control policies in terms of air quality and emissionreduction costs. Since, as already noted, defining largely shared airquality indexes is often difficult, a standard approach is to refer to

legislation in force that usually concerns pollution limits, whichhave to be attained. The problem can be formalized as follows:

min ½P1ðEðXÞÞ.PhðEðXÞÞ.PNðEðXÞÞCðEðXÞÞ� (1)

where E represents the precursor emissions, which are functions ofa set of decision variables (emission control measures) X;Ph(E(X)),h ¼ 1,.,N are air pollution indexes concerning differentpollutants; C(E(X)) represents the implementation costs of pollu-tion reduction measures.

Additionally a common approach is to reduce the objectives toonly two in order to ease the visualization and the understanding ofthe results. To this purpose, a linear combination PðxÞ ¼P

hwhPhðXÞ of the pollution indexes is assumed.

4.2. The decision variables

The decision variables are the entities of the abatementmeasures, i.e. the application rates (or penetration levels) of emis-sion reduction technologies, representing the fraction of emission(in each activity i) that can be controlled through the adoption ofsuch technologies. It is important to note that we refer here only totechnical measures, i.e. process modifications or emission filteringthat can be implemented at the end-of-pipe, without modificationof the related activity type and entity. For instance, we consider thepossibility of reducing traffic emissions by adopting an improvedtype of muffler (i.e. a better technology) but not that of reducing cartraffic (by other types of actions, as congestion charges, etc.).

Starting from these assumptions, the total emission ofa pollutant p remaining after the application of a set of technologiescan be calculated as follows:

Ep ¼Xi

�EunAbip � Aifip

Xt˛Tip

hitpXit

�(2)

where:

� EunAbip represent the emission of pollutant p and activity i if notechnologies were applied (unabated emission).

C. Carnevale et al. / Environmental Modelling & Software 38 (2012) 306e315310

� Tip is the set of technologies that can be applied in activity i toreduce pollutant p.

� Ai is the level of activity i.� fip represents the unabated emission factor [kton/Activity_U-nit] for activity i and pollutant p.

� hitp is the removal efficiency associated to technology t. Moreprecisely, it is the fraction (between 0 and 1) of pollutant p thatis removed by a 100% application of a particular technology t inactivity i.

� Xit represents the application rate (between Xit and Xit ,respectively minimum and maximum value) of technology t toactivity i.

Note that a certain technology may have a non-zero reductionefficiency for more than one pollutant (this is the case, for instance,of catalytic mufflers). As an example, in the test case of Lombardyregion, about 1600 technologies have been identified and evaluatedand some 700 were found relevant for the specific domain.

As already anticipated, one characteristic feature of RIAT is thebuilt-in ability of following the development of legislation. The so-called CLE (Y) scenario represents the technologies penetrationlevel foreseen at a particular year Y, under current legislation. It isthus possible for the user to select a given reference year andautomatically embed into the problem definition the technologiesactive at that time; then, a static optimization is performed,referred to the selected year Y. Since all the variables refer to thistemporal index, such a dependency is implied in what follows.

An additional possible specification is related to technologiesthat can be replaced or upgraded a few years after having beeninstalled, according to the legislation in force and technologies thatin practice cannot be reasonably substituted (for instance, whenimplying the complete replacement of an industrial plant withina few years of it being built).

4.3. The objectives

The Air Quality objective can be defined by the user as one thefollowing indexes:

� Annual mean PM10 concentration� Annual mean PM2.5 concentration� SOMO35: maximum daily 8-h running mean ozone concen-tration, accumulated over a threshold of 35 ppb (it is normallyused as an indicator of impacts on human health);

� AOT40: ozone concentrations accumulated over a threshold of40 ppb (an indicator of impacts on vegetation).

All the indexes can be computed over different domains. Therelationship between the decision variables and the indexes ismodeled by Artificial Neural Networks, identified processing a setof twenty long-term simulations of a Chemical Transport Model(Carnevale et al., 2012). Some preliminary results on PM2.5 will beillustrated later for the test case considered in this work.

The Cost Objective is calculated as follows. For each activity i,the cost of applying all technologies is computed as:

CiðXÞ ¼ Ai

Xp

Xt˛Tip

CitXit (3)

where:

� Ci are the abatement costs [Meuro] for activity i.� Cit are the annualized unit costs [Meuro/activity_unit] of theapplication of technology t. More in detail, these data arederived by GAINS database, and have been computed consid-ering investments, fixed and variable operating costs at

production (and not consumer) level (Amann et al., 2011). As anexample, the cost to apply “high efficiency covered outdoorstorage ofmanure” to the activity “dairy cattle-dairy cows slurrysystems” is 25.858 millions euros for 1 million of animals.

So the total costs [Meuro] is:

CðXÞ ¼Xi

CiðXÞ (4)

The two-objective optimization problem is solved following the˛ -Constraint Method (Ehrgott, 2000): the Air Quality objective isminimized while the Cost objective is included in the set ofconstraints with a parametric right-hand side, i.e.:

minX

PðXÞ (5)

CðXÞ � L; 0 � L � L (6)

where L is the cost of a full application of all the available tech-nologies. This is the same form of the standard cost-effectivenessanalysis: a problem that the user may be interested to solve,when the budget L is known.

4.4. The constraints

There are two possible configurations of constraints, in RIAT:

(1) In the case no technological substitution is admitted, thefollowing constraints are defined:� Technology feasibility (the technology penetration level hasto remain between a minimum and maximum applicationrate):

XCLEit � Xit � Xit ; ci; t (7)

� Measures complementarity (the sum of the application ratesof the technologies applied to a particular pollutant has to beless than 1, i.e. two measures cannot be applied to the sameportion of an emission source):X

Xit � 1; ci;p (8)

(2) In the case technological substitution is admitted, the followingconstraints are defined:� Technology feasibility and measures complementarity asabove

� Technology increasing efficacy (optimal removed rates haveto be not lower than those requested by legislation):

X XCLE

t˛Tip

hitpXit �t˛Tip

hitpXit ; ci; p (9)

� Incremental abatement (the share of non controlled emis-sions has to decrease in comparison to CLE values):

Xt˛Tip

Xit �Xt˛Tip

XCLEit ; ci; p (10)

In Section 6 some examples for Lombardy region, consideringboth the case with and without technology substitution, will beshown.

C. Carnevale et al. / Environmental Modelling & Software 38 (2012) 306e315 311

5. Technical implementation of RIAT

The RIAT system (processes and GUI) is developed with a J2SEplatform (java 2 standard edition). Data pre and post processing aremanaged by FORTRAN executables. The optimization module(including S/R and their application) uses MATLAB functions. Thedatabase is developed in Apache Derby, an open source relationaldatabase implemented entirely in Java. A RIAT GIS interface isintegrated in Google Earth.

The RIAT has a user friendly interface that sequentially guidesthe user through modules applying a folder approach. The mainRIAT modules are:

� Project setup� Emission inventory� Gridding data� Measure database� Run setup� Run results

The use of RIAT is simple and intuitive in spite of the complexityof the “back side” of the system. The scheme in Fig.1 shows how thesystem works: the RIAT core system needs a series of input con-cerning the measure database, the emission data and the spatialgrid to start with the optimization process and to produce output interms of tables and maps.

The measure database contains all the information of GAINStechnologies and activities: unabated emission factors for thepollutants (PM10, PM2.5, NOx, SO2, VOC and NH3), activity levels,removal efficiencies (for each pollutant), application rates (CLE andpotential), Unit Costs. This database file could be replaced directlyby the user or modified through the specific RIAT GUI (Measuremanagement GUI); the GAINS DB is integrated in the systemthrough an Excel file. Emission and mapping data are managedthrough the Emission Inventory GUI: emissions can contain areal,point sources and gridded data. Through the Gridding Data GUI, theuser sets grid information and spatial indicator (proxy variables, tosupport the emission spatialization procedure) so that it is possiblefor the system to grid areal emissions.

In the Measure data section, the user should also selecta subgroup of activities for the optimization and decide whethertechnologies can or cannot be substituted. He/she can also decide tomodify activities and to modify or create new measures. Further-more the user has to make the choice of the optimization typebetween Multi-Objective (one or two indicators of air quality) orCost-Effectiveness (one or two indicators): the choice of the latterrequires the setting of the cost budget constraint, while in the caseof two air quality indicators it is compulsory to set their relativeweights. The optimization area (i.e. where the policies are appliedand the optimization is computed) can be chosen among variousalternatives, for instance one or more provinces, critical areas, etc.The choice of S/R models can be made between ANN’s or otheruser-defined models.

When the run is successfully completed, the user can visualizeall the results. Fig. 3 tries to summarize all the types of output thatcan be obtained using RIAT, again using the test case of theLombardy region. If the user selects a Multi-Objective run, forinstance, the Pareto boundary curve (upper left in Fig. 3) is shown,with all the efficient solutions. Then the user can select one ofthe Pareto curve points to visualize the detailed output relative tothat alternative. If the user selects the Cost-Effectiveness modein the run option, only one solution is computed and shown byRIAT.

For both Multi-Objective and Cost-Effectiveness modes, thesystem provides two kinds of detailed output:

� TABLES: they show the optimized solution in a detailed andaggregated way (technology and macrosector details).

� MAPS: a GIS interface based on Google Earth provides the userwith basic and well known GIS navigation functions andcartographic information.

All the data in tables and maps can be exported in ASCII or Excelformat.

Output TABLES show run results in two different ways:

� TECHNOLOGY: contains the optimized measures (technolo-gies), their application rates, the reduced optimized emissionsfor each pollutant and the consequent regional total andincremental cost (see Fig. 3, center bottom).

� MACROSECTOR: contains regional emission reductions for eachpollutant, the consequent total and incremental cost, emissionbefore and after using the optimized measure.

Output MAPS report gridded data that can be visualized andanalyzed as:

� EMISSION. It contains the optimized gridded output annualemissions for the selected optimization year, sector andpollutant.

� AQI. It contains the gridded air quality index obtained by theapplication of chosen S/R on the optimally gridded emissions(see Fig. 3, left).

6. Test application results

Lombardy Region (Fig. 2), located in Northern Italy, has served asa test case for the application of RIAT. The region, despite efforts toreduce precursor emissions, is still facing problems in respectingEU air quality thresholds, and so it is a challenging domain forRegional Authorities in charge of preparing Air Quality plans. Theresults illustrated in the following, and shown in Fig. 3, refer to theoptimization of PM2.5 annual average concentrations and emissionreduction costs in 2010 conditions.

The analysis shown in Fig. 3 may lead to several interestingconsiderations. For instance, it shows the lowest value that the airpollution index can attain, whatever the amount of money investedin the abatement technologies. Such value is far from zero because,even if the technologies could reduce emissions close to zero, thelocal meteorology and the boundary conditions (due to the emis-sions of neighboring territories) imply anyway a certain level ofpollution. Nevertheless, this is an important value per se, since itprovides the DMwith a definite idea of the margin of improvementthat any decision may have. The form of the Pareto boundary (theexample in Fig. 3 corresponds for instance to the case with nosubstitution of technologies, in 2010) may evidence a strongcurvature in a certain region, thus quantifying the (expected)decreasing effects of investments in abatement technologies. Withan expenditure of about 10% of that corresponding to the fullapplication of abatement technologies, one may already attain80e85% of the possible improvement of air quality. Additionally,the user may compare for instance the effect of allowing or not thesubstitution of CLE technologies at a certain year. Fig. 4 shows fourdifferent Pareto curves related to efficient policies in 2010 and2020, substituting/not substituting technologies already in place inthe corresponding year. In particular, the blue curves are related tooptimal policies at 2010, while the red ones to optimal policies at2020, considering in both cases no technology substitution(continuous lines) or with substitution (dotted lines). The distancebetween the blue and the red lines represents how the situationmay improve thanks to the different technologies that will be

Fig. 3. RIAT: example of results.

C. Carnevale et al. / Environmental Modelling & Software 38 (2012) 306e315312

available in 2020 (in comparison to the 2010 case). In particular it isimportant to note that the Air Quality Index for the CLE case (leftpoint of the Pareto curve) is going to decrease by 5 mg/m3, only dueto the fact that additional measures will be implemented in 2020(the new legislation that will be in force). This difference is visiblealso for the rest of the Pareto curves. The distance between twocurves of the same color highlights the inefficiency of technologiesalready in use and thus provides an evaluation of the advantages ofscrapping them. If the DM limits this analysis to a certain sector (sayfor instance, road traffic), this difference may also be used asa measure of the amount of the incentives that the regionalgovernment may reasonably pay to boost vehicles substitution. Ina similar way, the user may compare the effects of measuresrestricted to a certain portion of the domain (for instance, urbanareas) with those taken on all the region. Or he/she may evaluatethe expected impacts on the population or agriculture, byweighting the air quality in each cell by the corresponding effects(see for instance ExternE coefficients at www.externe.info) andtargets. An additional interesting way of using the results is thedetermination of the effectiveness of new technologies still to beadopted. Looking at the technologies in the efficient solution fora given expenditure, the user understands which are those to beadopted and those to be discarded. To become part of such a solu-tion, a new technology should have an efficiency (cost per unit ofabated pollutant) higher than the less efficient already present.

Fig. 5 shows an example of policies selected by the optimizationprocedure. The case is related to a point of the Pareto curve for2010, obtained when minimizing PM2.5 concentrations andconsidering both the case with and without technologies substi-tution. The Figure is related to the traffic macrosector, and inparticular the “Light duty vehicles: cars and small buses with 4-stroke engines e Liquefied petroleum gas” activity, with theEURO standard technologies to be applied, from Euro I to Euro VI.The histogram represents the Current LEgislation technologypenetration level (CLE, blue), the optimal penetration levels for thecase without substitution (NTS, red) and the optimal penetrationlevels for the case with substitution (TS, green). Without technol-ogies substitution (red bar), it is only possible to increase theapplication rates beyond CLE, and few degrees of freedom are left to

increase, for instance, Euro VI penetration level (the majority of thefleet is in fact already controlled at CLE with some end-of-pipetechnology). The optimal case with technology substitution(green bar) has added degrees of freedom, allowing the scrappageof old technologies (as Euro I to III) and so a larger adoption of morestringent and newest standard.

7. Lesson learned and future perspective

The development of RIAT as well as of many other DSSs (see forinstance Janssen et al. (2009)) has been a recursive process,sometimes appearing as a never-ending one. More and morequestions were posed after the first results had been produced andthese led to a better understanding of the problems and of thepotentialities of the system itself. Recounting the history of itsdevelopment, we think that thewinning features of the system are:

� Locality. The system incorporates the local physical, econom-ical, and social characteristics, with a well identified user/decision maker, taking into account also the specific legislationof the area considered. The spatial distribution of the emissionactivities can be considered in detail, i.e. at the typical level ofmunicipalities of few square kilometers, quite far from thoseused at higher levels where typical cell size is of order ofhundreds square kilometers, and the use of average values risksto lead to poor approximations.

� Flexibility. All the local data are entered by the user, includingthe source-receptor models that incorporate the local meteo-rological and emissive conditions. Even the optimizationmodule can be substituted by a different one. In this way, theapplication of the system to different territorial contexts isstraightforward and is indeed under way. When the context isdefined, the user can still select the type of problem to beaddressed, the areas involved, the emission sector and thespecific technologies to be considered. This allows, for instance,the exclusion of those that for political or social reasons areconsidered outside the power of the DM.

� Openness. Practical experience shows that, despite high flexi-bility, no decision support system can fully represent all the

Fig. 4. RIAT: example of Pareto curves, with (continuous lines) and without (dotted lines) technology substitution, in 2010 (blue lines) and 2020 (red lines) CLE. (For interpretationof the references to colour in this figure legend, the reader is referred to the web version of this article.)

C. Carnevale et al. / Environmental Modelling & Software 38 (2012) 306e315 313

specific features of a certain territory/problem. The use of thesesystems always generates additional questions and it isimpossible to anticipate all the possible statistics, values,indicators and constraints that the user may ask for. It istherefore essential that the system provides an output that canlater be post-processed by a standard spread-sheet or GISsoftware. Another typical output is an optimized emission

Fig. 5. Example of technology selection, for traffic macrosector. Values refer to a specific plevel (CLE, blue), the optimal penetration levels without substitution (NTS, red) and with subthe reader is referred to the web version of this article.)

pattern, to be validated using a standard full 3D multiphasechemical transport model (e.g. MOZART, CAMx, TCAM, .).

The experience developed with RIAT has demonstrated that it ispossible to design and implement a general integrated assessmentsystem for air quality planning at regional level, isolating domainfeatures in the input database. It has also shown that we can go

oint of the Pareto Curve and represent the Current Legislation technology penetrationstitution (TS, green). (For interpretation of the references to colour in this figure legend,

C. Carnevale et al. / Environmental Modelling & Software 38 (2012) 306e315314

beyond the classical approach of using linear S/R models and weexpect the use of non linear models to become a standard in allfuture systems. We also expect more and more systems to bedeveloped using open source software, even if, in the environ-mental sector, this may take place at a reduced speed compared toother application domains: given in fact that most environmentalapplications are critical, managers sometimes require warrantiesthat can be granted only by proprietary software. There are otherimportant directions along which we see the development ofintegrated air quality planning models in the next years.

First, we think that some sort of uncertainty evaluation or sensi-tivity analysis will soon be included in integratedmodels. This aspectis particularly critical for air quality studies since they are based ongiven meteorological conditions that may not represent what willhappenwhen the plan materializes. Unfortunately, in this field thereis no equivalent of the centennial flood or other reference conditionsused in hydrology. It is very hard to define a critical episode since toomany variables have to be considered (e.g. temperature, wind,humidity, pressure in a 3D domain). It would thus be essential topropagate the uncertainty about themeteorological conditions to theoptimal decisions and to the air quality indicator values.

Second, we expect more support for time evolution. Methods tocheck the robustness of any decision in a changing context would bemostly welcome. Some sort of discounting for future effects ofreduction technologies may be attempted. Additionally, somesupport should probably be added to verify the compatibility ofcurrent decisions with future constraints (following the evolution oflegislation) taking also into account the expected useful life of eachtechnology and the problems related to scrapping and substitution.

A third development may concern the possibility of includingsome spatial variable in the decision process. Up to now in fact, thespatial distribution of activities has been considered as fixed. Onrelatively long time horizons, there clearly exists the possibility ofmoving some activities, as, for instance, creating new industrialareas or relocating a large power station. However, this additionopens an enormous number of different alternatives that maymakeagain scenario analysis the only practically viable approach. A gooddeal of conceptual work would be necessary to achieve a correctbalance between these additional optimization options (and notonly the mathematical solvability of the related problem, but alsothe conceptual possibility of grasping its meaning). Related to thisissue, a further important future development of IntegratedAssessment will be in the ability of including non technicalmeasures. These differ from end-of-pipe technologies because theyrelate to the possibility of reducing pollutant emissions acting onbehavioral changes of people (e.g. encouraging the use of bicycleinstead of car, reducing heating and ventilation in houses, etc).Non-technical measures are still not considered within IntegratedAssessment because it is very complex to estimate associated costsand removal efficiencies. This is an open conceptual problem, thatwill certainly need to be dealt in future, given the expectedincreasing role of these type of actions.

A further development will possibly emerge from the paralleli-zation of hardware and software that will lead to almost real-timeexecution of problems like those solved by RIAT. Presently, the timefor a solution even using simple surrogate models is of the order ofseveralminutes, a reasonable time to define an air quality plan lastingfor several years, but quite toomuch to be usedduring ameeting or tosupport a discussion in a public forum. Furthermore these paralleli-zation developments could allow the use of RIAT-like tools for short-term air quality management, and/or the increase of the number oftechnologies/control variables in the optimization problem.

In the future, it is also possible that these systems will allow theuser to define autonomously his/her own objectives by freelycombining model variables in analytical terms. Though this would

be already possible with present computer technologies, the mostdifficult problem to be solved is the education of decision-makers,who should be able to formalize quantitatively the objective(s) theyhave in mind.

Acknowledgment

RIAT project has been funded by JRC-IES (Joint Research CentreeInstitute for Environment and Sustainability), contract number384364. We also acknowledge Lombardia Region, and the MAG-IIASA (Mitigation of Air Pollution & Greenhouse Gases, Interna-tional Institute for Applied Systems Analysis) for data sharing.

References

Aben, J., Hettelingh, J.-P., Schopp, W., 2003. Rains-nl: an integrated assessmentmodel to support Dutch air quality policy making. In: Proc. 19th Intern. Conf.Informatics for Environmental Protection, pp. 513e517.

Amann, M., Bertok, I., Borken-Kleefeld, J., Cofala, J., Heyes, C., Hglund-Isaksson, L.,Klimont, Z., Nguyen, B., Posch, M., Rafaj, P., Sandler, R., Schpp, W., Wagner, F.,Winiwarter, W., 2011. Cost-effective control of air quality and greenhouse gasesin Europe: modeling and policy applications. Environmental Modelling &Software 26 (12), 1489e1501.

Carnevale, C., Decanini, E., Volta, M., 2008a. Design and validation of a multiphase3d model to simulate tropospheric pollution. Science of the Total Environment390, 166e176.

Carnevale, C., Finzi, G., Guariso, G., Pisoni, E., Volta, M., 2012. Surrogate models tocompute optimal air quality planning policies at a regional scale. EnvironmentalModelling and Software 34, 44e50.

Carnevale, C., Finzi, G., Pisoni, E., Volta, M., 2009. Neuro-fuzzy and neural networksystems for air quality control. Atmospheric Environment 43, 4811e4821.

Carnevale, C., Pisoni, E., Volta, M., 2008b. A multi-objective nonlinear optimizationapproach to design effective air quality control policies. Automatica 44, 1632e1641.

Carnevale, C., Pisoni, E., Volta, M., 2010. A non-linear analysis to detect the origin ofPM10 concentrations in Northern Italy. Science of the Total Environment 409,182e191.

Castelletti, A., Galelli, S., Ratto, M., Soncini-Sessa, R., Young, P., 2012. A generalframework for dynamic emulation modelling in environmental problems.Environmental Modelling & Software 34, 5e18.

Cuvelier, C., Thunis, P., Vautard, R., Amann, M., Bessagnet, B., Bedogni, M.,Berkowicz, R., Brandt, J., Brocheton, F., Builtjes, P., Carnavale, C., Denby, B.,Douros, J., Graf, A., Hellmuth, O., Hodzic, A., Honor, C., Jonson, J.,Kerschbaumer, A., de Leeuw, F., Minguzzi, E., Moussiopoulos, N., Pertot, C.,Peuch, V., Pirovano, G., Rouil, L., Sauter, F., Schaap, M., Stern, R., Tarrason, L.,Vignati, E., Volta, M., White, L., Wind, P., Zuber, A., 2002. Citydelta: a modelintercomparison study to explore the impact of emission reductions inEuropean cities in 2010. Atmospheric Environment 41, 189e207.

D’Elia, I., Bencardino, M., Ciancarella, L., Contaldi, M., Vialetto, G., 2009. Technicaland non-technical measures for air pollution emission reduction: the integratedassessment of the regional air quality management plans through the Italiannational model. Atmospheric Environment 43, 6182e6189.

Deutsch, F., Mensink, C., Vankerkom, J., Janssen, L., 2008. Application and validationof a comprehensive model for PM10 and PM2.5 concentrations in Belgium andEurope. Applied Mathematical Modeling 32, 1501e1510.

Dommen, J., Prevot, A., Baertsch-Ritter, N., Maffeis, G., Longoni, M., Gruebler, F.,Thielmann, A., 2003. High-resolution emission inventory of the Lombardyregion: development and comparison with measurements. Atmospheric Envi-ronment 37 (29), 4149e4161.

Ehrgott, M., 2000. Multicriteria Optimization. Springer.Finzi, G., Silibello, C., Volta, M., 2000. Evaluation of urban pollution abatement

strategies by a photochemical dispersion model. International Journal of Envi-ronment and Pollution 14 (1e6), 616e624.

Gabusi, V., Pisoni, E., Volta, M., 2008. Factor separation in air quality simulations.Ecological Modelling 218, 383e392.

Janssen, S., Ewert, F., Li, Hongtao, Athanasiadis, I.N., Wien, J.J.F., Thérond, O.,Knapen, M.J.R., Bezlepkina, I., Alkan-Olsson, J., Rizzoli, A.E., Belhouchette, H.,Svensson, M., Van Ittersum, M.K., 2009. Defining assessment projects andscenarios for policy support: use of ontology in integrated assessment andmodelling. Environmental Modelling & Software 24, 1491e1500.

Mediavilla-Sahagun, A., ApSimon, H., 2003. Urban scale integrated assessment ofoptions to reduce PM10 in London towards attainment of air quality objectives.Atmospheric Environment 37, 4651e4665.

Oxley, T., ApSimon, H., 2007. Space, time and nesting integrated assessment models.Environmental Modelling & Software 22, 1732e1749.

Pisoni, E., Carnevale, C., Volta, M., 2009. Multi-criteria analysis for PM10 planning.Atmospheric Environment 43, 4833e4842.

Ratto, M., Castelletti, A., Pagano, A., 2011. Emulation techniques for the reductionand sensitivity analysis of complex environmental models. EnvironmentalModelling & Software 34, 1e4.

Reis, S., Nitter, S., Friedrich, R., 2005. Innovative approaches in integrated assess-ment modelling of European air pollution control strategies e implications of

C. Carnevale et al. / Environmental Modelling & Software 38 (2012) 306e315 315

dealing with multi-pollutant multi-effect problems. Environmental Modelling& Software 20, 1524e1531.

Sokhi, R., Jose, R.S., Kitwiroon, N., Fragkou, E., Perez, J., Middleton, D., 2006.Prediction of ozone levels in London using the mm5-cmaq modelling system.Environmental Modelling & Software 21, 566e576.

Syri, S., Karvosenoja, N., Lehtil, A., Laurila, T., Lindfors, V., Tuovinen, J.-P., 2002.Modeling the impacts of the Finnish climate strategy on air pollution. Atmo-spheric Environment 36, 3059e3069.

Thunis, P., Cuvelier, C., Roberts, P., White, L., Nyrni, A., Stern, R., Kerschbaumer, A.,Bessagnet, B., Bergstrom, R., Schaap, M., 2010. Evaluation of a Sectoral Approachto Integrated Assessment Modeling e Second Report. Tech. Rep., EURODELTA-.EUR 24474 EN. Publications Office of the European Union; 2010, JRC59501,Luxembourg (Luxembourg).

Wagner, F., Amann, M., Schoepp, W., 2007. The GAINS Optimization Module as of 1February 2007. Tech. Rep., International Institute for Applied Systems Analysis,Laxenburg, (AU).