The number of issues, stakeholders needs and regulations that a policy must consider in the current world is so high that addressing them only with common-sense is unthinkable. Policy makers have to consider disparate issues, as diverse as economic development, environmental aspects, as well as the social acceptance of the policy. A single person cannot be expert in all these subjects. Thus, to obtain a well assessed policy in the current complex world one can adopt decision support systems featuring optimization components. Leveraging on previous work on Strategic Environmental Assessment, we developed a fully-fledged system that is able to provide optimal plans with respect to a given objective, to perform multi-objective optimization and provide sets of Pareto optimal plans, and to visually compare them. Each plan is environmentally assessed and its environmental footprint is provided in terms of emissions, global warming e↵ect, human toxicity, and acidification. The heart of the system is an application developed in a popular Constraint Logic Programming system on the Reals sort. It has been equipped with a web service module that can be queried through standard interfaces. An intuitive graphic user interface has been added to provide easy access to the web service and the CLP application.
Policy making, in the current connected world, has to consider such a number
of issues that a single person cannot possibly consider without introducing vast
approximations. For example European regions should provide Regional Energy
Plans to define strategic objectives and political actions for the energy sector.
These policies must take into consideration
? This paper is an extended version of [
The policy contains strategic objectives on the energy share and energy efficiency, measures and activities to cope with the increased energy needs, new regulations, etc. In the case of regional planning, the plan is typically very highlevel: it includes activities such as building new power plants for some total output power, the share of each fuel type (nuclear, fossil fuels, biomasses, photovoltaic, windmills, etc.) and the type of produced energy (electric or thermal); but it lacks information about, for example, the actual placement of the plants in the region, or the size of each of the plants. More detailed plans will be done at lower scale, like the province or municipality levels.
By EU directives, regional policies on the energy sector should also include an environmental assessment of the plan. Being the plan so high-level, usually the assessment is done only in a qualitative way.
In a previous work [
In a following work [
In this work, we show how the first prototype of the planner was extended to a fully-fledged application. The current version of the software supports – plans that consider decommissioning obsolete power plants; – computation of emissions of the power plants for various types of pollutants, in a quantitative way; – quantitative assessment of the e↵ect of the plan on human health, global warming, and acidification potential; – multi-criteria optimization considering a variety of objective functions based on qualitative and quantitative information; – computation of the Pareto front, for two or more objective functions; – a web interface, that can be accessed both as a Graphical User Interface (GUI) and as a web service.
This work is one of the components of the EU ePolicy project6. The final application will use the optimal planner as the center of a large application, that will include an opinion mining component, to assess the acceptance of the policies from the public considering information coming from blogs and social networks; a social simulator component, that will simulate how the population will react to the policies adopted by the Region; a mechanism design component, that will include information from game theory to provide the best allocation schemes of regional subsidies to the stakeholders in the Region; and an integrated visualization component.
The rest of the paper is organized as follows. We first explain the planning and environmental assessment as they are currently done by experts in the Emilia-Romagna region of Italy, and recap the basic CLP program of the first prototype (Section 2). In Section 3, extend it for the new features. We show the design and features of the web service and GUI in Section 4. Finally, we show an experimental evaluation in Section 5, and we conclude in Section 6. 2
Problem considered and CLP solution
The strategic environmental assessment, in the Emilia-Romagna region of Italy,
is currently performed by considering two matrices, called coaxial matrices [
The first matrix, M, considers the activities that can be undertaken in a plan, and links them with the so-called environmental pressures. Environmental pressures can be positive or negative, and they account for the impact on the environment of human activities. Each element mij of the matrix M can take values {high, medium, low, null}, and defines a qualitative dependency between the activity i and the negative or positive pressure j.
The second matrix, N , relates the pressures with the environmental receptors, that register the e↵ect of the pressures on the environment. For example, the activity “coal-fueled power plant” generates the pressure “emission of pollutants in the atmosphere” on the environment; on its turn, this influences the receptor “air quality” (as well as other receptors, like “human wellbeing” or “wildlife 6 http://www.epolicy-project.eu wellbeing”). In the same way, the pressure “emission of greenhouse gases” influences receptor “global warming”; while the pressure “emission of pollutants in the water” influences the “quality of river waters”. Again, each element nij of the matrix can take the same qualitative values: high, medium, low or null.
Currently, the matrices relate 115 activities with 29 negative pressures, 19 positive pressures and 23 receptors. They can be used to assess a variety of regional plans, including Agriculture plans, Forest, Fishing, Energy, Industrial, Transport, Waste, Water, Telecommunications, Tourism, Urban plans. The environmental assessment is usually done using a spreadsheet and eliminating (by hand) those activities that do not belong to the given type of plan; then pressures that are not influenced by remaining activities are removed, and again receptors that are not influenced by the remaining pressures are removed as well. Finally, these reduced matrices are evaluated by environmental experts, that state which parts are most important, mainly considering clusters of High values.
Clearly, this process is very slow, experts might overlook combinations of medium or low values that might produce a significant e↵ect, and, most importantly, it can be done only after the plan has been provided by the policy maker. At this stage, usually only minor modifications can be backpropagated to the plan, and it is practically impossible to compare the e↵ect of the plan with other alternative plans, because this would need to start again another planning phase. As a matter of fact, although the evaluation of alternative plans is obligatory by EU regulations, this is usually not done, because planning and environmental assessment are done in a strictly sequential way.
To overcome the limitations and improve on current practices, we devised an
expert system able to automatically assess regional plans [
A first CLP solution We model the planning problem in CLP on the Reals sort (CLP(R)). In CLP, one can define a problem through a set of variables, ranging on given domains; the possible assignments are restricted through a set of constraints; a solution is an assignment of values to variables such that all the constraints are satisfied. In many cases, solutions are not all equivalent, but there is an objective function to be maximized or minimized.
Given a number Na of activities, we consider a vector A = (a1, . . . , aNa ) in which we associate to each activity a variable ai that defines its magnitude. The domain of ai depends on the availability of the resource on the given Region; for example some regions are very windy, while others can exploit better biomasses or solar panels.
We distinguish primary from secondary activities: primary activities are of primary importance for the given type of plan, while secondary activities are those supporting the primary activities by providing the needed infrastructures. 8 j 2 AS aj =
X i2 AP
dij ai Na X ai ci i=1
BP lan Na X ai outi i=1
outP lan.
Moreover, given an expected outcome outP lan of the plan, we have a constraint ensuring to reach the outcome: For example, in an energy plan the outcome can be to have more energy available in the region, so outP lan could be the increased availability of electrical power (e.g., in kilo-TOE, Tonnes of Oil Equivalent). In such a case, outi will be the production in kTOE for each unit of activity ai.
Concerning the impacts of the regional plan, an environmental expert suggested to convert the qualitative values in the matrices into the following numeric coecients: high=0.75, medium=0.5, low=0.25 and null=0. We sum up the contributions of all the activities and obtain the estimate of the impact on each environmental pressure:
E.g, in an Energy plan, the activities that produce energy (e.g., power plants) are of primary importance; such activities require other activities (e.g., power lines, waste stocking, streets, etc.) to be performed, and they also should be considered in the environmental assessment. Let AP be the set of indexes of primary activities and AS that of secondary activities. The dependencies between primary and secondary activities are considered by the constraint: Each activity ai has a cost ci; given a budget BP lan available for a given plan, we have a constraint limiting the overall plan cost: (1) (2) (3) (4) (5) In the same way, given the vector of environmental pressures P = (p1, . . . , pNp ), one can estimate their influence on the environmental receptor ri by means of the matrix N , that relates pressures with receptors:
Possible objective functions include maximizing the produced energy, minimizing the cost, or maximizing one of the receptors (e.g., maximizing the “air quality”), or a linear combination of the above. 3
Extended solution The CLP(R) program described in Section 2.1 was practically used in the development of the 2011-13 Regional Energy plan of the Emilia-Romagna region of Italy. The main objective of the plan was to increase significantly the share of renewable energy in the energy mix, to fulfill the 20-20-20 directive. Indeed, during the years from 2011 to 2013, a large number of new renewable power plants was installed in the region, thanks to subsidies, that make them appealing from the market viewpoint. For the next years, technicians in the region foresee that old power plants fueled by fossil fuels will become obsolete, and they will have to be shut down completely, or, possibly, used only when renewable energy is unavailable or in peak hours. They asked us to extend the DSS to allow for possible closing of power plants, which means that some of the activities could have a negative magnitude: the magnitude, in MW, of oil-based power plants could be reduced with respect to the previous years.
First of all, one should notice that negative activities introduce nonlinearities. For example, if building a new power plant i has a cost ci in Euros per MW, decommissioning it will not give a profit of ci e/M W .
Our implementation is based on the ECLiPSe CLP language [
IsP osi = ⇢ 1 if ai 0 0 if ai < 0
P osi = ⇢ ai if ai 0 ; 0 if ai < 0 this can be obtained by imposing the following linear constraints: P osi ai P osi 0 P osi ai + (1 IsP osi) · M P osi IsP osi · M
Na X P osi ci i=1
BP lan. (where M is a suciently large positive number), and with the further integrality constraint IsP osi 2 { 0, 1}. The cost constraint (2) is now rewritten as
Similarly, we do not want that secondary activities are decommissioned automatically when decommissioning primary activities; so we impose their relationship only with the positive part of primary activities.
Concerning the environmental assessment, as a first attempt we left the original linear constraints of equations (4-5), but the results were not considered satisfactory by the environmental expert. In fact, a new activity has a number of impacts, some for the construction of the activity (e.g., land use for building a coal power plant, pollution due to the construction site, etc.), and others due to running the activity (e.g., air pollution for burning fuel, water for cooling the plant, etc.). If we assume that the same coecients in equation (4) can be used also for negative activities, we would correctly account for the second type of impacts, but we would wrongly assume that decommissioning a power plant means restoring the construction site as it was before. Moreover, we would not consider the end-life of the power plants, which can be significant (for example, consider nuclear power plants).
To account correctly for these cases, the environmental expert added new activities on the co-axial matrices (e.g., “Reduced use of fossil fueled power plants”), together with their impacts on environmental pressures. All the pressures are now computed only on positive activities, i.e., Equation (4) is substituted with Then, we considered the new activities as a new type of secondary activities: the “Reduced use of fossil fueled power plants” is a secondary activity that becomes positive only when one of the activities “Coal-based power plant”, “oil-based power plant”, etc., has a negative value. More precisely, we have secondary activities that are linked to the decommissioning of other activities: e.g., activity “Reduction of fossil fuel power plants” is a secondary activity that is positive if one of the fossil fueled power plants has a negative value. Associated to activities we now have two matrices of dependencies between activities. In particular we have a Na ⇥ Na square matrix D+ where each element di+j represents the magnitude of activity j per unit of activity i, and another Na ⇥ Na square matrix D where each element dij represents the magnitude of activity j per unit of reduction of activity i.
The dependency primary-secondary activities in Equation (1) is now
substituted with
(7)
(8)
where
The base CLP program in Section 2.1 was able to provide the environmental
assessment only in terms of qualitative information. We extended it to consider
also quantitative information, in particular the emission of pollutants in the
air for each power plant type. We rely on the data provided by two databases:
INEMAR [
While ISPRA provides the average emission for each type of plant (biomasses, oil, coal, etc.), INEMAR provides fine grained information, in which emissions depend also on the type of boiler and the size of the plant (in MW).
We relate the power produced by plants with that of each boiler type. Let NB the number of boiler types, we have a vector of constrained variables B = (b1, . . . , bNB ) where bi is the total output power of the plants using boiler type i. Let O be the matrix that relates power plants and the di↵erent kinds of boiler: each element oij of the matrix is set to 1 if the boiler bj 2 B can be used for the power plant ai 2 A, and zero otherwise. We impose that the output power of each plant type is the sum of the power of its boilers: (9) (10)
Let E = (e1, . . . , eNe ) be the vector of emissions and T the matrix that
relates them with the boilers. An element tij of the matrix represents the grams
of pollutant ei 2 E emitted when 1GJ of fuel is provided to the boiler bj 2 B.
To calculate the emissions, we have to compute the input energy for each boiler
type j, provided the output power bj :
T U is the average running time of a power plant per year (necessary to convert
energy into power, and estimated in 8000 hours by our environmental expert)
and ⌘ is the average eciency (output power/input power) of power plants,
which is prescribed by law as 39% [
The European Commission [
However, the tables provided by the EC do not always have the same
granularity of the information available for emissions. For example, for each plant type
we know the emissions of N OX ; unluckily, in [
Computing the Pareto front In the case of regional planning it is very hard (if not impossible) to devise a unique function that includes all the objectives that are important for the user. The optimization component described in Section 2 can provide optimal solutions with respect to one objective function that can be either the total amount of energy produced (both electrical and thermal), or the total cost, or the values of receptors, emissions and indexes explained in the previous sections. We decided to extend it to support also multi-objective optimization, to let the user compute more than one solution, and compare them.
In a multi-objective optimization problem, a Pareto optimal solution is such that it is not possible to improve the result for one objective function, without worsening at least one other objective function. More precisely, in a multiobjective problem with n functions to minimize, a solution µ⇤ is Pareto-optimal if there does not exist another solution µ such that µj µ⇤j for 1 j n and there exists at least one i, 1 i n such that µi < µ⇤i . The set of Pareto points is distributed on the so-called Pareto frontier. !!"#$%&'()%** !"#$%"&'()" !"#$%"&'()" 01'1$2&133"& *+,$-#.")/%
+,.&/0&(1 (a) The Web service.
rrvee S n o iiltcap p A
Browser User Interface
1.
JEE Servlet 3.
Control er
2.
PDearstiastesnoceurces
5.
JSP Pages View
4.
Java Beans (BOM) Model
Web service
Optimal planning (b) The Web application.
We implemented the normalized normal constraint method [
Graphical User Interface A software for policy making should be usable by non-IT experts, and have an intuitive GUI able to visualize properly the heterogeneous information inherently present in environmental policies. We deployed the CLP planner as a stateless Web service and access it by means of a stateful Web application. This choice is also convenient from the perspective of a possible composition with the other services of the ePolicy application.
The CLP program is embedded inside a Java wrapper (Fig. 1a) that encodes the requests in CLP terms and decodes the results. This component provides a plethora of Java classes that represent the Business Object Model (BOM) of this domain. Any query addressed to this component and all the returned results are expressed in terms of these objects. Then we use the Apache CXF framework to define a Web Service’s Service Endpoint Interface (WS SEI) – an interface containing the signature of the method to call the service – and later to implement such a service taking advantage of the wrapper.
The Web application that stands as a GUI for the Web service is a standard Java servlet (Fig. 1b) following the Model-View-Controller (MVC) pattern: any request made through a browser is intercepted by the servlet which acts as a controller. The requests are forwarded to the BOM objects inside the model ; these objects interact with our Web service and persistence layer to produce results. The controller then uses the JavaServer Pages (JSPs) to generate the view that becomes the response to display in the user’s browser. Both the Web service and the Web application are finally deployed to an application server. The
Web application can be accessed at: http://globalopt.epolicy-project.eu/ Pareto/.
After a welcome page that introduces the software, there are an input page, and a results page. In input, the user can select the language to use (currently, Italian or English), insert bounds (minimum and maximum bounds for each energy source), constraints, and objective functions for the Pareto optimization, as well as the number of Pareto points (s)he wishes to compare. Constraints and objectives can include linear combinations of cost, produced power, receptors, emissions, or indicators. To simplify the input, the user can load the data for the Regional Energy Plan 2011-2013 of the Emilia-Romagna region for Electricity; in the following section we will show the results for this instance. The user can then compute the Pareto optimal plans, and a set of graphs is presented, as described in the next section. 4.1
Interpreting the Results The results page consists of two master-slave panels. The left side-panel is the master and by clicking on any of its entries in one of its three sections, the view in the main panel (the slave) changes accordingly. Each view hosted in the main panel has many tabs and by selecting one of them, an appropriate graph or table is shown.
In particular, the left panel let the user select either a Scenario comparison, to compare all the generated scenarios, or to get detailed information on one scenario. Scenarios are divided into boundary scenarios, that are those that optimize one of the objective functions, and intermediate scenarios, that try to balance the various objectives.
Scenarios comparison. By clicking on General overview on the left panel, the user can compare the scenarios. One comparison is through a spiderweb chart (Fig. 2a) that has an axis for each objective function. Along each axis, the optimal values are far from the origin. Each scenario is represented by a polygon where each vertex is on a di↵erent axis. Generally speaking, a bigger polygon implies a better scenario (note that these solutions are Pareto optimal, so one polygon cannot be completely included into another polygon). Hovering the mouse on the axes, one can obtain the values for each plan. To improve legibility, one can deactivate one (or more) plans by clicking on it on the legend.
The scenarios can also be compared through stacked bar chart, showing, for each scenario, the distribution of costs per energy source (Fig. 2b), or the amount of electric/thermal energy per source. Again, hovering with the mouse over the graphs, more detailed information is provided.
Another scenario comparison is with respect to the values of the Objective functions selected by the users.
Finally, by clicking on Emissions and pollutants, three tabs (Fig. 3) show, in basic column charts, the amount of pollutants divided into the categories Heavy metals, Greenhouse gases, and Other pollutants. With the default data, heavy (a) Scenario comparison. (b) Costs summary. metals are not present, because in the 2011-2013 plan there are no fossil fuels, the only sources emitting metals.
Boundary and Intermediate scenarios. These sections show detailed information for each of the computed scenarios on the Pareto front. Scenarios are divided into boundary scenarios, that are those that optimize one of the objective functions, and intermediate scenarios. For each scenario, the following views are available: – Receptors. This composite view uses 7 VU-meter charts (Fig. 4a). The top part shows the 3 receptors with the best normalised value, while the bottom one the 3 with the worst normalised value. The main chart allows the user to select any receptor and appraise its normalised value. This graph was explicitly requested by an environmental expert to highlight the best and worst receptors. – There are then four interactive tabular views (Fig. 4b) showing respectively, for the chosen scenario, the amount of produced energy per source, the total (a) Receptors.
(b) A tabular view.
cost for each energy source to be spent in primary and secondary activities, the detailed costs for each activity, and the list of emissions. 5
Experiments The software computes the optimal solution for one objective in a very short time; on modern computers it is well below 1 second. The multi-objective version has to compute a number of solutions, that depend on the number of scenarios (points on the Pareto front) requested by the user, so the computing time can grow up to some seconds, to compute around 5-10 scenarios (a number of scenarios that can be visualized and compared visually).
In order to assess the scalability, we performed a series of tests by randomly generating a set of data, including the co-axial matrices, the matrix relating primary and secondary activities, the activity costs, etc. In this way, we were able to stress-test the software with instances containing a number of activities, pressures and receptors larger than those in the actual data provided by ARPA.
The experiments were performed on a laptop computer running Linux with a 8x Intel Core i7-3720QM CPU at 2.60GHz; only one core was used in the experiments. The results are plotted in Figure 5, for a single objective. The xaxis is the size of the instance, i.e., the size of the co-axial matrices (each matrix is N ⇥ N ). The y-axis shows the time required to find the optimal solution.
The computing time, for sizes below 100, is always less than a second. Note
that if N = 100, the matrix that relates activities and pressures has size 100 ⇥
100 = 10, 000, while in the real instance it is just 93 ⇥ 48 = 4, 464.
We presented a DSS with optimization capabilities based on CLP for the regional
planning and with particular emphasis on the environmental aspects of regional
policies. The program was practically used to produce the energy plan 2011-2013
of the Emilia-Romagna region in Italy [