- Email: [email protected]
Advanced street lighting control
Expert Systems with Applications 41 (2014) 999–1005
Contents lists available at SciVerse ScienceDirect
Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa
Advanced street lighting control Igor Wojnicki ⇑, Sebastian Ernst, Leszek Kotulski, Adam Seß dziwy AGH University of Science and Technology, Department of Applied Computer Science, Al. Mickiewicza 30, 30-059 Krakow, Poland
a r t i c l e
i n f o
Keywords: Lighting design Outdoor lighting Control system Graph transformation Artificial intelligence Rule-based
a b s t r a c t Design and control of outdoor lighting systems is a complex task, which is made even more difficult by introducing features like dynamic, sensor-based operation, multiple lighting levels and sophisticated, adjustable luminaires. This paper proposes an integrated approach, based on formal graph-based models and methods, to handle both of these tasks. The introduced formalisms help handle the state-space explosion related to the aforementioned characteristics. Control is performed by means of AI techniques (including rule-based systems and pattern matching), which is applied to the system using graph transformations. An illustrative, simple example is carried out throughout the paper, but the presented methods are highly scalable, which made them applicable to several practical projects of varying scale and characteristics. 2013 Elsevier Ltd. All rights reserved.
1. Introduction Outdoor lighting systems are usually designed and operated with several objectives in mind. These include conformance to lighting standards (i.e. provision of minimum required illuminance in given regions), minimization of power consumption as well as improvement of user comfort (which is partly due to standard conformance) and aesthetics. Recent advancements in lighting systems bring features which may be helpful in achieving these objectives, such as multiple luminaire states (e.g. featuring more than two light levels), complex or reconfigurable geometries, different operating modes (i.e. night or emergency modes) and, last but not least, sensor-based dynamic control. Although helpful, these features dramatically increase the number of possible states of the lighting system, making its design a very complex task, which cannot be controlled by hand. To solve this problem, formal methods such as distributed graph transformations must be employed in the design phase. These should be complemented by appropriate solutions for control, including advanced AI algorithms (such as pattern matching or automated planning) and agent-based systems. This paper proposes a formal basis for integrated and coherent design and control of outdoor lighting systems as well as a simple example to illustrate its concepts. It is organized as follows. Section 2 describes the relationship between design and control and outlines the advantages of this approach. Section 3 provides an insight
⇑ Corresponding author. Tel.: +48 508444450. E-mail addresses: [email protected] (I. Wojnicki), [email protected] (S. Ernst), [email protected] (L. Kotulski), [email protected] (A. Se ß dziwy). 0957-4174/$ - see front matter 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.07.044
into the formal, graph-based approach to lighting design. In Section 4, a supplemented graph structure – the control availability graph (CAG) – is introduced. Section 5 adds an AI approach to lighting control, and Section 6 provides details how the rules can be applied to the CAG by means of graph transformations.
2. Design and control Design and control of lighting systems are tasks which are often handled separately. This is usually due to the fact that tools and formalisms applicable to one of these problems are usually not universal enough to handle the other. The proposed graph-based approach is different in that it provides a formal framework able to encompass information relevant to both the design phase and the exploitation phase, i.e.: spatial characteristics of the area under consideration, requirements regarding the illumination of individual subregions, parameters of the luminaires, all inputs and outputs, including sensors and luminary control, methods for selecting appropriate configurations and determining state transitions. This approach has numerous advantages, first of which is the ability to create designs better suited to real-life exploitation based on the assumed usage scenarios. This allows systems to be controlled more efficiently, and lamp parameters may be selected more appropriately with regard to the required lighting levels. Moreover, analysis of control scenarios may positively influence sensor deployment by helping to identify missing or excess sensors
1000
I. Wojnicki et al. / Expert Systems with Applications 41 (2014) 999–1005
in the system. Finally, an integrated lighting design/control process may yield more accurate logical decomposition of the area under consideration (for instance, with regard to predicted dynamics in various subregions) and trigger redesign when necessary. Related research regarding intelligent outdoor lighting control indicates a common theme: the possibility to save energy. This has been confirmed by experiments. Outdoor lighting optimization regarding a highway tunnel, presented by Fan, Yang, and Wang (2011), has shown significant reduction of energy use. The tunnel is equipped with vehicle and luminance detectors. While minimizing energy use, the system still had to fulfill luminance requirements in order to comply with safety regulations. Guo, Eloholma, and Halonen (2008) confirm that street lighting for highways and intersections which takes both traffic and weather conditions into consideration can lead up to 40.9% energy savings. The aforementioned solutions are tailored and tuned to particular cases. The approach presented in this paper is design-driven: design gives basis for intelligent control using AI techniques. This way, general rules, which are easily transferable to other cases, can be used for control purposes. A proposal of a control system for complex, intelligent street lighting can be found in Wojnicki and Kotulski (2012).
(see Seß dziwy & Kozien´-Woz´niak, 2012), and dimensionless points are added as its supplementary vertices. The crucial feature of the hypergraph model is that it provides a calculation grid (2D/3D) necessary to perform photometric computations. Let us note that the results of those computations may be encapsulated within attributes of nodes and edges/hyperedges. Thus, they are accessible at the abstract layer. Fig. 2(a) shows an extremely simplified scene containing the gas pump, a section of the ground area and the canopy with its supporting pillars. All these objects are assumed to have cuboidal forms. Additionally, two fixtures and a single sensor are present. The graph model of this scene is shown in Fig. 2(b). For better clarity, we collapse hypergraphs representing particular solids and mark them as hexagonal vertices. To achieve a lighting design goal (required light level with minimal energy usage), one has to solve an optimization problem, formulated at the beginning of the section, using the calculation grids provided by the graph model. It should be emphasized that the results of this phase include not only the adjustments of internal fixture parameters (photometric profile, inclination, dimming and so on) but also the distribution of fixtures and, possibly, sensors. For this reason, the design phase strongly impacts the control phase. In particular, a graph structure obtained in the design process creates a framework for the control task.
3. Formalizing design The main objective of lighting design is development of a lighting infrastructure which satisfies given criteria, e.g. a suitable illuminance level, low power consumption or other, business-oriented goals (compare with Boyce, Hunter, & Vasconez (2001)). An approach frequently applied to outdoor lighting design addresses two of these criteria, i.e. it provides optimization targeted at minimizing the input power level, while still fulfilling the mandatory lighting standards (see: Illuminating Engineering Society of North America (IESNA), 2000; British Standards Institution (BSI), 2003; Commission Internationale de l‘Eclairage, 2010). Such an approach is assumed for the considered case study. The first step towards creating a lighting design is identifying all usage/lighting scenarios for a given area. They may be related to factors like the presence and behavior of (volatile) objects, e.g. cars or persons, but also to weather conditions, current time and date, the ambient light level, etc. Potentially, each of these scenarios requires different adjustments of particular luminaires (the lumen flux values) and, possibly, different control schemes when transiting from/to other scenarios. In further considerations, we will use the notion of a profile which refers to a set of requirements regarding the illuminance of a segment. In most cases, a profile is ascribed to some subarea containing luminaires, sensors and relevant infrastructure. Thus, fixture adjustments, switched dynamically (triggered by environment state changes) and in conformance with given profiles, define a control scheme. The next phase in the design process is creation of a representation of the considered area (see Fig. 1(left)) by means of a proper formal model. A formal model allows for efficient computations (especially for large-scale problems), thanks to the possible application of computational techniques which are supported by such an abstract layer (Kotulski & Sedziwy, 2012). For lighting design problems, a graph-based formalism will be used. The considered environment representing a gas station1 is given in Fig. 1. All elements of the environment are regarded either as solids (gas pumps, pillars, canopy, kiosk, ground area) or as dimensionless points (fixtures, sensors). A composite solid being a compound of multiple adjacent polyhedrons is represented by a hypergraph 1
http://en.wikipedia.org/wiki/File:Esso_gas_station_finland.png GNU Free Documentation License.
4. Control availability graph The environment regarding a particular area, equipped with sensors and lamps illuminating its segments, is shown in Fig. 1. It is formally represented as a graph (see Fig. 3), called the control availability graph. The following naming convention is assumed. A vertex is attributed with its type, index and any number of optional key-value pairs. Regarding the examples, for simplicity, the mandatory attributes (type, index) are expressed by labeling the vertex with the type name, the index being its subscript. Additional attribution regarding information relevant to dynamically-changing control, such as the current sensor values or indication that a particular lamp configuration is enabled, is also provided. It is presented as a third value in square brackets at selected vertices. Formally, it is an attribute value named according to a particular vertex type (see Table 1). These are used in subsequent sections. A lamp is shown as a vertex labeled L. The configuration of a lamp is represented as C, while a segment, which is an area illuminated by a number of lamps – as S. A single lamp configuration contains parameters for a group of lamps, which achieve the desired light profile on a given segment. The parameters are expressed as labeled edges between C and L, while profiles are edges between S and C. There are the following sensors: P – presence sensor, K – darkness sensor, D – ‘done fueling’ sensor (indicating that the gas needs to be paid for), H – day or night hours sensor (which determines the operating mode of the station, described further in this section). There are relationships between sensors and segments, defined as appropriate edges. Numerical indices indicate particular lamps (L1, L2, etc.), configurations (C1, C2, etc.), segments (S1, S2, etc.), or sensors (P1, P2, D1, D2, R1, etc.). Formally, vertex labeling corresponds to attributes. Each vertex has its type and index attributes which hold the values (e.g.: C2 corresponds to type = C, index = 2). 5. Rule-based approach Analysis of the requirements indicated the following terms and rules. There are three levels of illumination: high, low and off. There are two modes of operation: day hours and night hours. During the
1001
I. Wojnicki et al. / Expert Systems with Applications 41 (2014) 999–1005
Fig. 1. Gas station case: segments defining enclosed areas (S1 . . . S9), lamps (L1 . . . L10), and sensors (detecting presence: P1 . . . P6, or if fueling is complete: D1 . . . D8).
Sensor Fixture 1
Fixture 2 Canopy
Pump Pilar 1
Pilar 2
Ground
(a) Solid-based model of a gas station
(b) Graph representation of the scene shown in (a)
Fig. 2. Gas station and its graph description.
day hours, the entire interior of the station, including a convenience store and cashiers, is accessible to customers. During the night hours, only the night cashier window is accessible. Lighting should take the ambient light into consideration. The following rules are established to control lighting. 1. If it is not dark, the station’s exterior is not illuminated. 2. If it is dark, the station’s exterior is illuminated at low. 3. If car or pedestrian presence is detected in a given segment and it is dark, the segment is illuminated at high. 4. When the fueling process is complete and it is dark, appropriate entrance is illuminated at high: (a) the main entrance during day hours, (b) the night cashier during night hours. 5. If gas is paid for (fueling is not done) and it is dark, disable high illumination at: (a) the main entrance during day hours, (b) the night cashier during night hours. 6. When the fueling process is completed during day hours, the interior has to be additionally illuminated. 7. If the gas is paid for (the fueling process is not done) during day hours, additional illumination of the interior is disabled. The above rules can be expressed as a decision table (Table 2) for easier comprehension. For more complex applications, other rule representation and synthesis methods can be used, including decision trees and Contextual Networked Decision Tables (Wojnicki, 2011). For high complexity systems, Context-Based Reasoning (Gonzalez, Stensrud, & Barrett, 2008), providing hierarchicallyarranged groups of rules, can also be applied. Rule #2 enables potential customers to locate the gas station when it is dark, simultaneously providing energy savings by setting
lighting levels to low. Rule #3 increases customer convenience by switching lighting level to high, providing good illumination of the pump and its surrounding area. Rule #4 directs customer attention to the proper place where the payment should be deposited. Rule #6 directs both customer’s and staff’s attention to make a payment or serve the customer, respectively. Rule #5 is complementary to #4, while #7 – to #6. Since the considered environment, including sensors and actuators, is modeled as a graph, expressing the above rules in graphrelated terms is proposed. 6. Graph transformations Decisions made by a rule-based system, resulting in the change of attribute values in a control availability graph G (see Fig. 3) may be viewed as graph grammar productions carried out on G. The following definition formally introduces the notion of a production. Definition 1. Let G be a control availability graph. A production performed on G is a quintuple of the form P ¼ ðl; r; p; embed; ev alÞ, where l and r denote respectively the left and right hand side graph of the production P, p is an applicability predicate function, such that pðPÞ 2 ftrue; falseg, indicating whether a production P may be applied, embed is an embedding rule for the graph r; it specifies a way it is reattached to a processed graph, eval is the function assigning attributes to particular nodes and vertices. The definition of a production, presented above, has to be accompanied by the scheme of its application: 1.
pðPÞ is evaluated in order to check the possibility of using P: if false is obtained, then the production P cannot be fired. Otherwise, we proceed with the next steps.
1002
I. Wojnicki et al. / Expert Systems with Applications 41 (2014) 999–1005
Fig. 3. Gas station, control availability.
I. Wojnicki et al. / Expert Systems with Applications 41 (2014) 999–1005 Table 1 Additional vertex attributes. Type
Attribute
Values
Comment
K C H D P
detected engaged detected detected detected
{true, false} {on, off} {night, day} {true, false} {true, false}
Indicates if it is dark Enables particular lamp configuration Indicates day or nigh mode Indicates if fueling is complete Indicates if presence is detected
Table 2 Gas station, decision table. Rule
Condition Dark Presence Done fueling Day hours Decision Exterior low Exterior high Day entrance high Night entrance high Interior high
#1
#2
#3
#4a
#4b
#5a
#5b
F
T
T T
T
T
T
T
T T
T F
F T
F F
Off Off
On Off
#6
#7
T T
F T
On
Off
Off On On
Off On
Off
This demand requires the introduction of an order into a set of profiles. The decision which of those approaches to select depends on the problem. For the presented examples, a decision was made to use conflict set resolution (Nilsson & Małuszyn´ski, 1990) instead of the applicability predicate. The production in Fig. 5 corresponds to rule #11. Other productions (Figs. 6–13) represent appropriate rules respectively. Comparing the rules given in Section 5 to the productions, a semantic difference can be noticed. The rule conditions include sensor input only, while the left-hand sides of productions, which correspond to the conditions, can contain references to any vertices, edges and their attributes. For example, the production in Fig. 5, which corresponds to rule # 1, references the activation of a lamp configuration Ck[on] being part of the pattern to be matched. This indicates that applying productions, which represent graph-based pattern matching, is a powerful and highly expressive tool. Any entity, either a vertex, an edge or an attribution, can represent a condition or decision if needed. This requires efficient AI techniques which incorporate pattern matching (Russell & Norvig, 2009). To better explain the control process, let us perform the following scenario. 1. It is after dark, during day hours, there are no cars at the gas station. 2. A car pulls over, and stops at a pump number 1. 3. The pump number 1 begins fueling. 4. A car pulls over, and stops at a pump number 8. 5. The pump number 8 begins fueling. 6. The pump number 1 completes fueling. 7. The pump number 8 completes fueling. 8. The pump number 1 is ready to begin next fueling; the balance is paid. 9. The car leaves the pump number 1. 10. The pump number 8 is ready to begin next fueling; the balance is paid. 11. The car leaves the pump number 8.
2. We take a subgraph H # G, isomorphic with l, and remove it from G, together with all incident edges. 3. The graph r is added to G H together with additional edges (connecting both). The number of edges and the way they have to be attached is specified unambiguously by the embed rule. 4. Finally, the eval function assigns proper values to relevant edge/ node attributes. For the considered case of a control availability graph for a gas station, one deals with productions for which l r and embed rule specifies that all edges are recovered in the new graph in such a way that two states of the graph, i.e. before and after applying P, are isomorphic. The key information is contained in the values of eval, which carry control data. Example 1. The problem of switching between lighting levels may be expressed by means of a graph production in two ways. In the first method, we define the production P representing a lighting level change, as shown in Fig. 4. The following applicability predicate, p, which is the conjunction of several conditions, is associated with the above production:
pðPÞ ¼
true;
if 8j – i : Dj ¼ false;
false; otherwise:
Please note that parallel application of several productions of the above form may generate conflicts. A conflict occurs when productions are applied to subgraphs containing vertices {Di}i2I and $S"i 2 I : succ(Di) = S, where succ(x) denotes the successor of node x in a digraph G. In other words, they produce conflicting values of the attribute C. An alternative, more practical approach assumes that from conflicting productions, we select the one for which the value of C satisfies the most restrictive profile (i.e. the highest lighting level).
1003
Initial labeling is as follows:
K1[false], H1[day], for each Cj set Cj[off], for each Pk set Pk[false], for each Dl set Dl[false].
Subsequent graph transformations related to the above scenario are presented in Fig. 14; productions are denoted as Un, the initial labeling is indicated as Init. Considering step 1, a new attribution is performed, since it is dark: K1[true], subsequently production 2 (see Fig. 6) matches. As a result, new attribution is performed: for each Cj set Cj[on]. At step 2 P1[true] is attributed, production 3 (see Fig. 7) is matched. New attribution: C2[off] and C1[on]. At step 4, P6[true] is attributed, production 3 is matched again. New attribution: C14[off] and C13[on]. At step 6, D1[true] is attributed, production 4a, 2 and 6 (see Figs. 8, 6 and 12) are matched. New
Fig. 4. Production P.
1004
I. Wojnicki et al. / Expert Systems with Applications 41 (2014) 999–1005
Fig. 5. Production 1 (U1) supporting rule #1.
Fig. 6. Production 2 (U2) supporting rule #2.
Fig. 7. Production 3 (U3) supporting rule #3.
Fig. 8. Production 4a (U4a) supporting rule #4a.
Fig. 9. Production 4b (U4b) supporting rule #4b.
Fig. 10. Production 5a (U5a) supporting rule #5a.
Fig. 11. Production 5b (U5b) supporting rule #5b.
Fig. 12. Production 6 (U6) supporting rule #6.
I. Wojnicki et al. / Expert Systems with Applications 41 (2014) 999–1005
1005
Fig. 13. Production 7 (U7) supporting rule #7.
Fig. 14. Applying productions in consecutive steps of the scenario.
attribution: C16[off], C15[on], C19[on]. At step 7, D8[true] is attributed, no production is matched. At step 8, D1[false] is attributed, production 5a, 2 and 7 (see Figs. 10, 6 and 13) is matched. At step 9, P1[false] is attributed, production 2 (see Fig. 6) is matched. New attribution: C1[off] and C2[on]. At step 10, D8[false] is attributed, production 5a, 2 and 7 are matched. New attribution: C15[off], C16[on], C19[off]. At step 11, P6[false] is attributed, production 2 is matched. New attribution: C13[off] and C14[on]. It needs to be noted that the control process, carried out by a rule-based system with graph-oriented pattern matching and execution, is performed continuously. It reacts for changing environment and takes appropriate actions accordingly, activating and deactivating profiles for given segments, thus constituting reactive behavior (Russell & Norvig, 2009). 7. Summary The approach proposed in this paper aims at solving the complex task of designing and controlling a contemporary outdoor lighting system. Such systems are characterized by features such as dynamic, sensor-based control, multiple luminaire states and complex geometries, which make both the design phase and the control phase a significantly complex task. State-space explosion caused by the aforementioned characteristics makes it necessary to employ formal methods for design – in this case, a graph model. These are accompanied by appropriate means of control, including AI methods (rule-based systems, pattern matching) and distributed graph transformations. For simplicity, the paper uses a rather simple example of a gas station to illustrate these concepts. However, one of the biggest advantages of the proposed approach is its scalability, which makes it applicable to design and control of lighting systems spanning vast areas (Seßdziwy, 2012). Examples of such applications include the Green AGH Campus project (Szmuc, Kotulski, Wojszczyk, & Sedziwy, 2012), dealing with outdoor lighting architectural spaces featuring diverse sizes and characteristics, street lighting for a significant part of the city of Krakow, Poland (approx. 3500 light points) and a KIC InnoEnergy project.
References Boyce, P., Hunter, C., & Vasconez, S. (2001). An evaluation of three types of gas station canopy lighting. Technical report, Lighting Research Center, Rensselaer Polytechnic Institute, Troy, NY 12180-3352.. British Standards Institution (BSI). (2003). Road lighting. Performance requirements. BS EN 13201-2:2003. London: BSI. Commission Internationale de l‘Eclairage. (2010). Lighting of roads for motor and pedestrian traffic. CIE 115:2010. Vienna: CIE. Fan, S., Yang, C., & Wang, Z. (2011). Automatic control system for highway tunnel lighting. In D. Li, Y. Liu, & Y. Chen (Eds.), Computer and computing technologies in agriculture IV. IFIP advances in information and communication technology (Vol. 347, pp. 116–123). Boston: Springer. http://dx.doi.org/10.1007/978-3-64218369-0http://www.springerlink.com/index/N1520PT884727374.pdf. Gonzalez, A. J., Stensrud, B. S., & Barrett, G. C. (2008). Formalizing context-based reasoning: a modeling paradigm for representing tactical human behavior. International Journal of Intelligent Systems, 23, 822–847. Guo, L., Eloholma, M., & Halonen, L. (2008). Intelligent road lighting control systems. Technical report, Helsinki University of Technology, Department of Electronics, Lighting Unit. . Illuminating Engineering Society of North America (IESNA) (2000). American national standard practice for roadway lighting. RP-8-00. New York: IESNA. Kotulski, L., & Sedziwy, A. (2012). On the effective distribution and maintenance of knowledge represented by complementary graphs. Transactions on Computational Collective Intelligence, 6, 105–120. Nilsson, U., & Małuszyn´ski, J. (1990). Logic, programming and prolog. John Wiley & Sons. Russell, S. J., & Norvig, P. (2009). Artificial intelligence: a modern approach. 3rd ed.. Prentice Hall. Seßdziwy, A. (2012). Effective graph representation for agent-based distributed computing. In G. Jezic, M. Kusek, N.-T. Nguyen, R. J. Howlett, & L. C. Jain (Eds.), Agent and multi-agent systems. technologies and applications. Lecture notes in computer science (Vol. 7327, pp. 638–647). Berlin, Heidelberg: Springer. Seßdziwy, A., & Kozien´-Woz´niak, M. (2012). Computational support for optimizing street lighting design. In W. Zamojski et al. (Eds.), Complex systems and dependability. Advances in intelligent and soft computing (Vol. 170, pp. 241–255). Berlin, Heidelberg: Springer. Szmuc, T., Kotulski, L., Wojszczyk, B., & Sedziwy, A. (2012). Green AGH campus. In B. Donnellan, J. A. P. Lopes, J. Martins, & J. Filipe (Eds.), SMARTGREENS (pp. 159–162). SciTePress. Wojnicki, I. (2011). From tabular trees to networked decision tables: an evolution of modularized knowledge-base representations. Pomiary Automatyka Kontrola, 12. Wojnicki, I., & Kotulski, L. (2012). Controlling complex lighting systems. In W. Zamojski, J. Mazurkiewicz, J. Sugier, T. Walkowiak, & J. Kacprzyk (Eds.), Complex systems and dependability. Advances in intelligent and soft computing (Vol. 170, pp. 305–317). Berlin, Heidelberg: Springer. http://dx.doi.org/10.1007/978-3642-30662-4.
Contents lists available at SciVerse ScienceDirect
Expert Systems with Applications journal homepage: www.elsevier.com/locate/eswa
Advanced street lighting control Igor Wojnicki ⇑, Sebastian Ernst, Leszek Kotulski, Adam Seß dziwy AGH University of Science and Technology, Department of Applied Computer Science, Al. Mickiewicza 30, 30-059 Krakow, Poland
a r t i c l e
i n f o
Keywords: Lighting design Outdoor lighting Control system Graph transformation Artificial intelligence Rule-based
a b s t r a c t Design and control of outdoor lighting systems is a complex task, which is made even more difficult by introducing features like dynamic, sensor-based operation, multiple lighting levels and sophisticated, adjustable luminaires. This paper proposes an integrated approach, based on formal graph-based models and methods, to handle both of these tasks. The introduced formalisms help handle the state-space explosion related to the aforementioned characteristics. Control is performed by means of AI techniques (including rule-based systems and pattern matching), which is applied to the system using graph transformations. An illustrative, simple example is carried out throughout the paper, but the presented methods are highly scalable, which made them applicable to several practical projects of varying scale and characteristics. 2013 Elsevier Ltd. All rights reserved.
1. Introduction Outdoor lighting systems are usually designed and operated with several objectives in mind. These include conformance to lighting standards (i.e. provision of minimum required illuminance in given regions), minimization of power consumption as well as improvement of user comfort (which is partly due to standard conformance) and aesthetics. Recent advancements in lighting systems bring features which may be helpful in achieving these objectives, such as multiple luminaire states (e.g. featuring more than two light levels), complex or reconfigurable geometries, different operating modes (i.e. night or emergency modes) and, last but not least, sensor-based dynamic control. Although helpful, these features dramatically increase the number of possible states of the lighting system, making its design a very complex task, which cannot be controlled by hand. To solve this problem, formal methods such as distributed graph transformations must be employed in the design phase. These should be complemented by appropriate solutions for control, including advanced AI algorithms (such as pattern matching or automated planning) and agent-based systems. This paper proposes a formal basis for integrated and coherent design and control of outdoor lighting systems as well as a simple example to illustrate its concepts. It is organized as follows. Section 2 describes the relationship between design and control and outlines the advantages of this approach. Section 3 provides an insight
⇑ Corresponding author. Tel.: +48 508444450. E-mail addresses: [email protected] (I. Wojnicki), [email protected] (S. Ernst), [email protected] (L. Kotulski), [email protected] (A. Se ß dziwy). 0957-4174/$ - see front matter 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.eswa.2013.07.044
into the formal, graph-based approach to lighting design. In Section 4, a supplemented graph structure – the control availability graph (CAG) – is introduced. Section 5 adds an AI approach to lighting control, and Section 6 provides details how the rules can be applied to the CAG by means of graph transformations.
2. Design and control Design and control of lighting systems are tasks which are often handled separately. This is usually due to the fact that tools and formalisms applicable to one of these problems are usually not universal enough to handle the other. The proposed graph-based approach is different in that it provides a formal framework able to encompass information relevant to both the design phase and the exploitation phase, i.e.: spatial characteristics of the area under consideration, requirements regarding the illumination of individual subregions, parameters of the luminaires, all inputs and outputs, including sensors and luminary control, methods for selecting appropriate configurations and determining state transitions. This approach has numerous advantages, first of which is the ability to create designs better suited to real-life exploitation based on the assumed usage scenarios. This allows systems to be controlled more efficiently, and lamp parameters may be selected more appropriately with regard to the required lighting levels. Moreover, analysis of control scenarios may positively influence sensor deployment by helping to identify missing or excess sensors
1000
I. Wojnicki et al. / Expert Systems with Applications 41 (2014) 999–1005
in the system. Finally, an integrated lighting design/control process may yield more accurate logical decomposition of the area under consideration (for instance, with regard to predicted dynamics in various subregions) and trigger redesign when necessary. Related research regarding intelligent outdoor lighting control indicates a common theme: the possibility to save energy. This has been confirmed by experiments. Outdoor lighting optimization regarding a highway tunnel, presented by Fan, Yang, and Wang (2011), has shown significant reduction of energy use. The tunnel is equipped with vehicle and luminance detectors. While minimizing energy use, the system still had to fulfill luminance requirements in order to comply with safety regulations. Guo, Eloholma, and Halonen (2008) confirm that street lighting for highways and intersections which takes both traffic and weather conditions into consideration can lead up to 40.9% energy savings. The aforementioned solutions are tailored and tuned to particular cases. The approach presented in this paper is design-driven: design gives basis for intelligent control using AI techniques. This way, general rules, which are easily transferable to other cases, can be used for control purposes. A proposal of a control system for complex, intelligent street lighting can be found in Wojnicki and Kotulski (2012).
(see Seß dziwy & Kozien´-Woz´niak, 2012), and dimensionless points are added as its supplementary vertices. The crucial feature of the hypergraph model is that it provides a calculation grid (2D/3D) necessary to perform photometric computations. Let us note that the results of those computations may be encapsulated within attributes of nodes and edges/hyperedges. Thus, they are accessible at the abstract layer. Fig. 2(a) shows an extremely simplified scene containing the gas pump, a section of the ground area and the canopy with its supporting pillars. All these objects are assumed to have cuboidal forms. Additionally, two fixtures and a single sensor are present. The graph model of this scene is shown in Fig. 2(b). For better clarity, we collapse hypergraphs representing particular solids and mark them as hexagonal vertices. To achieve a lighting design goal (required light level with minimal energy usage), one has to solve an optimization problem, formulated at the beginning of the section, using the calculation grids provided by the graph model. It should be emphasized that the results of this phase include not only the adjustments of internal fixture parameters (photometric profile, inclination, dimming and so on) but also the distribution of fixtures and, possibly, sensors. For this reason, the design phase strongly impacts the control phase. In particular, a graph structure obtained in the design process creates a framework for the control task.
3. Formalizing design The main objective of lighting design is development of a lighting infrastructure which satisfies given criteria, e.g. a suitable illuminance level, low power consumption or other, business-oriented goals (compare with Boyce, Hunter, & Vasconez (2001)). An approach frequently applied to outdoor lighting design addresses two of these criteria, i.e. it provides optimization targeted at minimizing the input power level, while still fulfilling the mandatory lighting standards (see: Illuminating Engineering Society of North America (IESNA), 2000; British Standards Institution (BSI), 2003; Commission Internationale de l‘Eclairage, 2010). Such an approach is assumed for the considered case study. The first step towards creating a lighting design is identifying all usage/lighting scenarios for a given area. They may be related to factors like the presence and behavior of (volatile) objects, e.g. cars or persons, but also to weather conditions, current time and date, the ambient light level, etc. Potentially, each of these scenarios requires different adjustments of particular luminaires (the lumen flux values) and, possibly, different control schemes when transiting from/to other scenarios. In further considerations, we will use the notion of a profile which refers to a set of requirements regarding the illuminance of a segment. In most cases, a profile is ascribed to some subarea containing luminaires, sensors and relevant infrastructure. Thus, fixture adjustments, switched dynamically (triggered by environment state changes) and in conformance with given profiles, define a control scheme. The next phase in the design process is creation of a representation of the considered area (see Fig. 1(left)) by means of a proper formal model. A formal model allows for efficient computations (especially for large-scale problems), thanks to the possible application of computational techniques which are supported by such an abstract layer (Kotulski & Sedziwy, 2012). For lighting design problems, a graph-based formalism will be used. The considered environment representing a gas station1 is given in Fig. 1. All elements of the environment are regarded either as solids (gas pumps, pillars, canopy, kiosk, ground area) or as dimensionless points (fixtures, sensors). A composite solid being a compound of multiple adjacent polyhedrons is represented by a hypergraph 1
http://en.wikipedia.org/wiki/File:Esso_gas_station_finland.png GNU Free Documentation License.
4. Control availability graph The environment regarding a particular area, equipped with sensors and lamps illuminating its segments, is shown in Fig. 1. It is formally represented as a graph (see Fig. 3), called the control availability graph. The following naming convention is assumed. A vertex is attributed with its type, index and any number of optional key-value pairs. Regarding the examples, for simplicity, the mandatory attributes (type, index) are expressed by labeling the vertex with the type name, the index being its subscript. Additional attribution regarding information relevant to dynamically-changing control, such as the current sensor values or indication that a particular lamp configuration is enabled, is also provided. It is presented as a third value in square brackets at selected vertices. Formally, it is an attribute value named according to a particular vertex type (see Table 1). These are used in subsequent sections. A lamp is shown as a vertex labeled L. The configuration of a lamp is represented as C, while a segment, which is an area illuminated by a number of lamps – as S. A single lamp configuration contains parameters for a group of lamps, which achieve the desired light profile on a given segment. The parameters are expressed as labeled edges between C and L, while profiles are edges between S and C. There are the following sensors: P – presence sensor, K – darkness sensor, D – ‘done fueling’ sensor (indicating that the gas needs to be paid for), H – day or night hours sensor (which determines the operating mode of the station, described further in this section). There are relationships between sensors and segments, defined as appropriate edges. Numerical indices indicate particular lamps (L1, L2, etc.), configurations (C1, C2, etc.), segments (S1, S2, etc.), or sensors (P1, P2, D1, D2, R1, etc.). Formally, vertex labeling corresponds to attributes. Each vertex has its type and index attributes which hold the values (e.g.: C2 corresponds to type = C, index = 2). 5. Rule-based approach Analysis of the requirements indicated the following terms and rules. There are three levels of illumination: high, low and off. There are two modes of operation: day hours and night hours. During the
1001
I. Wojnicki et al. / Expert Systems with Applications 41 (2014) 999–1005
Fig. 1. Gas station case: segments defining enclosed areas (S1 . . . S9), lamps (L1 . . . L10), and sensors (detecting presence: P1 . . . P6, or if fueling is complete: D1 . . . D8).
Sensor Fixture 1
Fixture 2 Canopy
Pump Pilar 1
Pilar 2
Ground
(a) Solid-based model of a gas station
(b) Graph representation of the scene shown in (a)
Fig. 2. Gas station and its graph description.
day hours, the entire interior of the station, including a convenience store and cashiers, is accessible to customers. During the night hours, only the night cashier window is accessible. Lighting should take the ambient light into consideration. The following rules are established to control lighting. 1. If it is not dark, the station’s exterior is not illuminated. 2. If it is dark, the station’s exterior is illuminated at low. 3. If car or pedestrian presence is detected in a given segment and it is dark, the segment is illuminated at high. 4. When the fueling process is complete and it is dark, appropriate entrance is illuminated at high: (a) the main entrance during day hours, (b) the night cashier during night hours. 5. If gas is paid for (fueling is not done) and it is dark, disable high illumination at: (a) the main entrance during day hours, (b) the night cashier during night hours. 6. When the fueling process is completed during day hours, the interior has to be additionally illuminated. 7. If the gas is paid for (the fueling process is not done) during day hours, additional illumination of the interior is disabled. The above rules can be expressed as a decision table (Table 2) for easier comprehension. For more complex applications, other rule representation and synthesis methods can be used, including decision trees and Contextual Networked Decision Tables (Wojnicki, 2011). For high complexity systems, Context-Based Reasoning (Gonzalez, Stensrud, & Barrett, 2008), providing hierarchicallyarranged groups of rules, can also be applied. Rule #2 enables potential customers to locate the gas station when it is dark, simultaneously providing energy savings by setting
lighting levels to low. Rule #3 increases customer convenience by switching lighting level to high, providing good illumination of the pump and its surrounding area. Rule #4 directs customer attention to the proper place where the payment should be deposited. Rule #6 directs both customer’s and staff’s attention to make a payment or serve the customer, respectively. Rule #5 is complementary to #4, while #7 – to #6. Since the considered environment, including sensors and actuators, is modeled as a graph, expressing the above rules in graphrelated terms is proposed. 6. Graph transformations Decisions made by a rule-based system, resulting in the change of attribute values in a control availability graph G (see Fig. 3) may be viewed as graph grammar productions carried out on G. The following definition formally introduces the notion of a production. Definition 1. Let G be a control availability graph. A production performed on G is a quintuple of the form P ¼ ðl; r; p; embed; ev alÞ, where l and r denote respectively the left and right hand side graph of the production P, p is an applicability predicate function, such that pðPÞ 2 ftrue; falseg, indicating whether a production P may be applied, embed is an embedding rule for the graph r; it specifies a way it is reattached to a processed graph, eval is the function assigning attributes to particular nodes and vertices. The definition of a production, presented above, has to be accompanied by the scheme of its application: 1.
pðPÞ is evaluated in order to check the possibility of using P: if false is obtained, then the production P cannot be fired. Otherwise, we proceed with the next steps.
1002
I. Wojnicki et al. / Expert Systems with Applications 41 (2014) 999–1005
Fig. 3. Gas station, control availability.
I. Wojnicki et al. / Expert Systems with Applications 41 (2014) 999–1005 Table 1 Additional vertex attributes. Type
Attribute
Values
Comment
K C H D P
detected engaged detected detected detected
{true, false} {on, off} {night, day} {true, false} {true, false}
Indicates if it is dark Enables particular lamp configuration Indicates day or nigh mode Indicates if fueling is complete Indicates if presence is detected
Table 2 Gas station, decision table. Rule
Condition Dark Presence Done fueling Day hours Decision Exterior low Exterior high Day entrance high Night entrance high Interior high
#1
#2
#3
#4a
#4b
#5a
#5b
F
T
T T
T
T
T
T
T T
T F
F T
F F
Off Off
On Off
#6
#7
T T
F T
On
Off
Off On On
Off On
Off
This demand requires the introduction of an order into a set of profiles. The decision which of those approaches to select depends on the problem. For the presented examples, a decision was made to use conflict set resolution (Nilsson & Małuszyn´ski, 1990) instead of the applicability predicate. The production in Fig. 5 corresponds to rule #11. Other productions (Figs. 6–13) represent appropriate rules respectively. Comparing the rules given in Section 5 to the productions, a semantic difference can be noticed. The rule conditions include sensor input only, while the left-hand sides of productions, which correspond to the conditions, can contain references to any vertices, edges and their attributes. For example, the production in Fig. 5, which corresponds to rule # 1, references the activation of a lamp configuration Ck[on] being part of the pattern to be matched. This indicates that applying productions, which represent graph-based pattern matching, is a powerful and highly expressive tool. Any entity, either a vertex, an edge or an attribution, can represent a condition or decision if needed. This requires efficient AI techniques which incorporate pattern matching (Russell & Norvig, 2009). To better explain the control process, let us perform the following scenario. 1. It is after dark, during day hours, there are no cars at the gas station. 2. A car pulls over, and stops at a pump number 1. 3. The pump number 1 begins fueling. 4. A car pulls over, and stops at a pump number 8. 5. The pump number 8 begins fueling. 6. The pump number 1 completes fueling. 7. The pump number 8 completes fueling. 8. The pump number 1 is ready to begin next fueling; the balance is paid. 9. The car leaves the pump number 1. 10. The pump number 8 is ready to begin next fueling; the balance is paid. 11. The car leaves the pump number 8.
2. We take a subgraph H # G, isomorphic with l, and remove it from G, together with all incident edges. 3. The graph r is added to G H together with additional edges (connecting both). The number of edges and the way they have to be attached is specified unambiguously by the embed rule. 4. Finally, the eval function assigns proper values to relevant edge/ node attributes. For the considered case of a control availability graph for a gas station, one deals with productions for which l r and embed rule specifies that all edges are recovered in the new graph in such a way that two states of the graph, i.e. before and after applying P, are isomorphic. The key information is contained in the values of eval, which carry control data. Example 1. The problem of switching between lighting levels may be expressed by means of a graph production in two ways. In the first method, we define the production P representing a lighting level change, as shown in Fig. 4. The following applicability predicate, p, which is the conjunction of several conditions, is associated with the above production:
pðPÞ ¼
true;
if 8j – i : Dj ¼ false;
false; otherwise:
Please note that parallel application of several productions of the above form may generate conflicts. A conflict occurs when productions are applied to subgraphs containing vertices {Di}i2I and $S"i 2 I : succ(Di) = S, where succ(x) denotes the successor of node x in a digraph G. In other words, they produce conflicting values of the attribute C. An alternative, more practical approach assumes that from conflicting productions, we select the one for which the value of C satisfies the most restrictive profile (i.e. the highest lighting level).
1003
Initial labeling is as follows:
K1[false], H1[day], for each Cj set Cj[off], for each Pk set Pk[false], for each Dl set Dl[false].
Subsequent graph transformations related to the above scenario are presented in Fig. 14; productions are denoted as Un, the initial labeling is indicated as Init. Considering step 1, a new attribution is performed, since it is dark: K1[true], subsequently production 2 (see Fig. 6) matches. As a result, new attribution is performed: for each Cj set Cj[on]. At step 2 P1[true] is attributed, production 3 (see Fig. 7) is matched. New attribution: C2[off] and C1[on]. At step 4, P6[true] is attributed, production 3 is matched again. New attribution: C14[off] and C13[on]. At step 6, D1[true] is attributed, production 4a, 2 and 6 (see Figs. 8, 6 and 12) are matched. New
Fig. 4. Production P.
1004
I. Wojnicki et al. / Expert Systems with Applications 41 (2014) 999–1005
Fig. 5. Production 1 (U1) supporting rule #1.
Fig. 6. Production 2 (U2) supporting rule #2.
Fig. 7. Production 3 (U3) supporting rule #3.
Fig. 8. Production 4a (U4a) supporting rule #4a.
Fig. 9. Production 4b (U4b) supporting rule #4b.
Fig. 10. Production 5a (U5a) supporting rule #5a.
Fig. 11. Production 5b (U5b) supporting rule #5b.
Fig. 12. Production 6 (U6) supporting rule #6.
I. Wojnicki et al. / Expert Systems with Applications 41 (2014) 999–1005
1005
Fig. 13. Production 7 (U7) supporting rule #7.
Fig. 14. Applying productions in consecutive steps of the scenario.
attribution: C16[off], C15[on], C19[on]. At step 7, D8[true] is attributed, no production is matched. At step 8, D1[false] is attributed, production 5a, 2 and 7 (see Figs. 10, 6 and 13) is matched. At step 9, P1[false] is attributed, production 2 (see Fig. 6) is matched. New attribution: C1[off] and C2[on]. At step 10, D8[false] is attributed, production 5a, 2 and 7 are matched. New attribution: C15[off], C16[on], C19[off]. At step 11, P6[false] is attributed, production 2 is matched. New attribution: C13[off] and C14[on]. It needs to be noted that the control process, carried out by a rule-based system with graph-oriented pattern matching and execution, is performed continuously. It reacts for changing environment and takes appropriate actions accordingly, activating and deactivating profiles for given segments, thus constituting reactive behavior (Russell & Norvig, 2009). 7. Summary The approach proposed in this paper aims at solving the complex task of designing and controlling a contemporary outdoor lighting system. Such systems are characterized by features such as dynamic, sensor-based control, multiple luminaire states and complex geometries, which make both the design phase and the control phase a significantly complex task. State-space explosion caused by the aforementioned characteristics makes it necessary to employ formal methods for design – in this case, a graph model. These are accompanied by appropriate means of control, including AI methods (rule-based systems, pattern matching) and distributed graph transformations. For simplicity, the paper uses a rather simple example of a gas station to illustrate these concepts. However, one of the biggest advantages of the proposed approach is its scalability, which makes it applicable to design and control of lighting systems spanning vast areas (Seßdziwy, 2012). Examples of such applications include the Green AGH Campus project (Szmuc, Kotulski, Wojszczyk, & Sedziwy, 2012), dealing with outdoor lighting architectural spaces featuring diverse sizes and characteristics, street lighting for a significant part of the city of Krakow, Poland (approx. 3500 light points) and a KIC InnoEnergy project.
References Boyce, P., Hunter, C., & Vasconez, S. (2001). An evaluation of three types of gas station canopy lighting. Technical report, Lighting Research Center, Rensselaer Polytechnic Institute, Troy, NY 12180-3352.