smart street lighting management

ORIGINAL ARTICLE

Smart street lighting management

S. Pizzuti & M. Annunziato & F. Moretti

Received: 2 July 2012 /Accepted: 23 January 2013 /Published online: 8 February 2013# Springer Science+Business Media Dordrecht 2013

Abstract In this work, we propose a new street light-ing energy management system in order to reduceenergy consumption. The key idea we want to accom-plish is that of “energy on demand” meaning thatenergy, in this case light, is provided only when need-ed. In order to achieve this goal, it is critical to have areliable demand model, which in the case of streetlighting turns out to be a traffic flow rate forecastingmodel. In order to achieve this goal, several methodson the 1-h prediction have been compared and the oneproviding the best results is based on artificial neuralnetworks. Moreover, several control strategies havebeen tested and the one which gave the best en-ergy savings is the adaptive one we carried out.Experimentation has been carried out on real dataand the study shows that with the proposed ap-proach, it is possible to save up to 50 % of energycompared to no regulation systems.

Keywords Lighting efficiency . Energy managementsystems . Adaptive control . Neural network models

Introduction

Since the first international recommendations for thelighting of roads (CIE 1995), power consumption andenvironmental aspects have become more and moreimportant, and at the same time, the improved perfor-mance of luminaires and lamps, and especially theintroduction of electronic control gears, has made itpossible to introduce adaptive lighting for motorizedroads and pedestrians areas.

A structured model has been developed for theselection of the appropriate lighting classes (CIE 1152010) (M, C, or P), based on the luminance concept,taking into account the different parameters relevantfor the given visual tasks. Applying for example time-dependent variables like traffic volume or weatherconditions, the model offers the possibility to useadaptive lighting systems with remarkable energy con-sumption savings and therefore high financial benefitsfor those municipalities (The Highland Council 2008)where street lighting is a high percentage of the elec-trical bill.

Today, lighting control approaches range from sim-ple on/off to regulation systems. On/off systems in-clude timers, twilight, and astronomical clocks. Thefirst one is a static system which turns on and off streetlights always according to fixed times. The second onehas light-sensitive photocells to turn them on at duskand off at dawn (Delft University of Technology2011). The third ones are GPS-based street light con-trollers which operate the on/off of the street lightaccording to the location features (longitude, latitude,sunrise, sunset times).

Energy Efficiency (2013) 6:607–616DOI 10.1007/s12053-013-9195-9

S. Pizzuti (*) :M. Annunziato : F. MorettiEnergy New Technologies and Sustainable EconomicDevelopment Agency (ENEA), Rome, Italye-mail: [email protected]

M. Annunziatoe-mail: [email protected]

F. Morettie-mail: [email protected]

S. Pizzuti : F. MorettiAutomation and Computer Science Department, University“Roma Tre”, Rome, Italy

Regulation systems are based on dimmable LED orhigh-pressure sodium vapor lights (Rea et al. 2009) andallow to schedule lights on or off and set dimming levelsof individual or groups of lights. All these systems haveone common feature: they do not care about the real on-line demand and this is a source of high inefficiency.

Thus, in order to overcome the main lack of thecurrent regulation systems, it has recently started thenew intelligent street lighting (ISL) approach whichlooks very promising (Delft University of Technology2011; Fundazioa 2011). Therefore, here we propose anISL approach (smart adaptive control) based on theconcept of “energy on demand,” whose goal is to dy-namically set the light intensity as function of the fore-seen demand, namely the traffic flow rate 1 h forecast.

Thus, in such context, the demand model has acritical role and its accuracy strongly affects the per-formance of the regulation system.

In the last decade, one of the most widely usedmethods in order to solve modeling problems is that ofartificial neural networks (ANNs) (Arbib 1995; Haykin1999). In particular, traffic flow forecasting has beenfaced since the 1990s (Canca et al. 1997; Dougherty andCobbett 1994; Ishak et al. 2003; Park et al. 1999; Taylorand Meldrum 1995; van Lint et al. 2003; Zheng et al.2006) up today (Pamuła 2011; Bucur et al. 2010;Çetiner et al. 2010; Jawanjal and Bajaj 2010) withANN. As example, among the most recent work(Jawanjal and Bajaj 2010) focuses on traffic flow fore-casting approach based on particle swarm optimizationwith wavelet network model. Pamuła (2011) reviewsneural network applications in urban traffic manage-ment systems and presents a method of traffic flowprediction based on neural networks. Bucur et al.(2010) propose the use of a self-adaptive fuzzy neuralnetwork for traffic prediction suggesting an architecturewhich tracks probability distribution drifts due to weath-er conditions, season, or other factors.

All the mentioned applications have one featurein common: they use one single global model inorder to perform the prediction. A novel approachis to use not only one model, but an ensemble ofmodels.

Modeling methods

In this paragraph, we shortly describe the modeling andcontrol techniques we compared in the experimentation.

Statistical modeling

One of the simplest and most widely used models is tobuild an average weekly distribution of the traffic flowrate sampled hourly. Thus, from the data, we computefor each day the average traffic flow rate hour by hourin such a way that we get an average distribution madeof 24×7=168 points.

Artificial neural networks

ANNs are computational models which try to sim-ulate some properties of biological neural networksin order to solve complex modeling problems ofnonlinear systems. An ANN is an interconnectedgroup of artificial neurons (called also nodes) thatuse a mathematical or computational model forinformation processing based on a connectionisticapproach to computation. In more practical terms,ANNs are nonlinear data modeling or decision-making tools which can be used to model complexrelationships between inputs and outputs or to findpatterns in data. ANNs are referred also as blackbox or data-driven models and they are mainly

Fig. 1 Artificial neuron model

Fig. 2 Feed-forward neural network topology

608 Energy Efficiency (2013) 6:607–616

used when analytical or transparent models cannotbe applied. Building such models needs severalstages as input analysis and training through algo-rithms which minimize the error between the realvalues to be modeled and the ANN output. ANNdemonstrated their effectiveness in modeling manyreal-world applications.

Once modeling an ANN model, we must takeinto account three basic components. First, thesynapses of the biological neuron are modeled asweights. Let us remember that the synapse of thebiological neuron is the one which interconnectsthe neural network and gives the strength of theconnection. For an artificial neuron, the weight isa number and represents the synapse. A negativeweight reflects an inhibitory connection, whilepositive values designate excitatory connections.The following components of the model representthe actual activity of the neuron cell. All inputsare summed altogether and modified by theweights. This activity is referred as a linear com-bination. Finally, an activation function controlsthe amplitude of the output. Mathematically, thisprocess is described in Fig. 1. From this model,

the activity of the neuron can be shown to be:

y ¼ faX

WiXi � θ� �

ð1Þ

where θ is a threshold called basic input activationsystem which identifies the sensitivity of the neu-ron to respond to the external inputs. The mostcommon function used to model fa are the hyper-bolic tangent, the sigmoid, and the linear function.

Therefore, each unit performs a relatively simplejob: receive input from neighbors or external sourcesand use this to compute an output signal which ispropagated to other units. Apart from this processing,a second task is the adjustment of the weights. Thesystem is inherently parallel in the sense that manyunits can carry out their computations at the sametime. Within neural systems, it is useful to distinguishthree types of units: input units which receive datafrom outside the neural network, output units whichsend data out of the neural network, and hidden unitswhose input and output signals remain within thenetwork.

The way units are connected defines the networktopology or architecture. In the past years, many ofthem have been studied and the most widely used andis the feed-forward one. In this network structure,neurons are grouped into layers. There exist at leasttwo layers, the input and the output one, which arethose gathering the corresponding input and outputvariables. This basic structure is also known as per-ceptron (Fig. 2) (Rosenblatt 1957).

Moreover, in order to let the model cope withnonlinear problems, it is possible to add one or more

Fig. 3 Ensembling

Fig. 4 Static control profile(central Italy, spring season)

Energy Efficiency (2013) 6:607–616 609

intermediate layers, known as hidden layers. Thesemodels are also known as multilayer perceptrons(MLP) (Rosenblatt 1961).

The flow of data from input to output units isstrictly in one direction (forward). The data processingcan extend over multiple (layers of) units, but nofeedback connections are present, that is, connectionsextending from outputs of units to inputs of units inthe same layer or previous layers.

Ensembling

The term “ensemble” describes a group of learningmachines that work together on the same task; in thecase of ANN, they are trained on some data, runtogether, and their outputs are combined as a single

one. The goal is obtain better predictive performancethan could be obtained from any of the constituentmodels (Fig. 3).

In the last years, several ensembling methods havebeen carried out (Krogh and Vedelsby 1995; Liu andYao 1999; Breiman 1999). The first one, also known asbasic ensemble method (BEM), is the simplest way tocombine M neural networks as an arithmetic mean oftheir outputs yi. This method can improve the globalperformance (Perrone and Cooper 1993; Bishop 1995)although it does not take into account that some modelscan be more accurate than others. This method has theadvantage to be very easy to apply.

A direct BEM extension is the generalized ensem-ble method (Perrone and Cooper 1993; Bishop 1995)in which the outputs of the single models are

Fig. 5 Smart adaptive control

Fig. 6 Adaptivecontrol profile


combined in a weighted average where the weightshave to be properly set, sometimes after an expensivetuning process. Other methods are bootstrap aggregat-ing (Kohavi 1999) and Adaboost (Drucker 1997;Avnimelech and Intrator 1999).

Control approaches

In this section, we describe the two control approacheswe compared. The first is a classical one (static con-trol), and the second is a novel one (smart adaptivecontrol).

Static control

The easiest and one of the most widely imple-mented regulation strategies is the static control(StaC). This simply sets the dimming levels ofthe street lights always according to fixed timesand a typical schedule is the one which keeps thepower level at 100 % for half night and at 50 %for the remaining half.

The strategy control profile depends on the locallongitudinal position and on the season of the year. InFig. 4, the profile applied in our experimentation (cen-tral Italy during the spring season) is shown as anexample.

Smart adaptive control

Smart adaptive control (SmAC) is the novel regulationsystem we have carried out based on energy on de-mand approach (Fig. 5). The basic idea is to set thepower level of the following hour as a function of theANN ensemble forecast.

Ptþ1 ¼ f ftþ1

� � ð2Þwhere Pt+1 is the power level normalized in [0,1] to beset for the next hour, and ϕt+1 is the traffic flow rate

neural forecast which is

ftþ1 ¼ anne ft; ft�1; . . . ; ft�nð Þ ð3Þwhere anne is the ANN ensemble result, and ϕt−i is themeasured traffic flow rate at time t−i.

For street lighting applications, the function f inEq. 2 can be shaped in different ways. Among these,we applied a linear profile although internationalstandards (CIE 115: 2010) suggest a nonlinear onethat we will apply in future work.

If fk < 0:25; then f ¼ 0:5If fk > 0:5; then f ¼ 1Else ¼ 2fk

ð4Þ

where ϕk is the predicted traffic flow rate at time k(Eq. 3) normalized in [0,1] (Fig. 6).

Experimentation

In this paragraph, we test and compare the methodspresented in the previous sections. The test case hasconcerned three different urban streets (Table 1) locat-ed in the town of Terni (about 90 km north of Rome)and the data set is made of 3 months (13 weeks) ofmeasurement corresponding to 2,184 hourly samples.The data set has been partitioned into training/testing

Table 1 Street featuresMaximumtraffic flowrate

Street 1 600

Street 2 800

Street 3 950

Table 2 Window history length (hours) selection

Number of samples Street 1 (%) Street 2 (%) Street 3 (%)

3 5.72 6.88 5.81

5 3.9 5.07 3.99

8 3.29 3.43 3.02

10 3.54 4.12 3.74

Table 3 Model comparison

Statistic(%)

ANN ANN ensembling(%)

Street 1 5.90 3.74 % (±0.10 %) 3.29

Street 2 5.56 3.48 % (±0.09 %) 3.02

Street 3 7.14 4.00 % (±0.10 %) 3.43

Average 6.20 3.74 % (±0.10 %) 3.25


and validation made respectively of 10 and 3 weekseach (Table 1).

Modeling

The ANNs are feed-forward MLP with 10 hiddenneurons and one output (the one hour flow fore-cast) with sigmoid as activation function for all theneurons. The number of inputs is the same of thedynamic window length (Eq. 3) and it has beenchosen with a preliminary analysis by calculatingthe validation prediction error, after the ensemblingstage, for different number of hourly samples(Table 2). By this analysis, it turned out the opti-mal number of input neurons (namely the lengthof the history window) to be 8 h.

Training has been performed through the backpro-pagation algorithm with adaptive learning rate andmomentum stopping after 108 iterations and a “savebest” strategy to avoid overfitting. The reported resultsare averaged over 10 different runs (with standard

deviation in brackets) and the ensemble is thereforemade by the same 10 models.

The reported errors are measured as

e ¼ x� yj j M � mð Þ= ð5Þ

where x is the real value to be predicted, y is the outputmodel, M is the real maximum value, and m is theminimum (Table 2).

At last, the following table shows the compari-son of the models considered in this work in termsof prediction accuracy over the validation set andFigs. 2 and 3 show a graphical comparison(Table 3).

From this analysis, it is clear that in general, theensembling approach outperforms the statistical ap-proach providing a remarkable improvement in predic-tion accuracy. Such level of precision is very importantwhen dealing with applications like traffic and lightingcontrol where the higher themodel accuracy is, the moreeffective the control system is (Fig. 7).

Fig. 7 Model comparison


From this graph, it is clear that the ANN ensemblemodel performs much better than the statistical modelbecause when out of normal conditions, the ANN ensem-bling takes into account the real traffic dynamics (Eq. 3).

Control

In this paragraph, we compare the results of the StaC andthe SmAC introduced in “Control approaches” section.In the experimentation, we calculated the energysaving of the two methods with respect to the noregulation strategy, namely when lights are alwayson at 100 % of their power for the whole night.

The light on demand control assumes dimmablelights (SAP or LED). On the streets where wecarried out this study, there were no such lampsand data about the real consumptions were not

available; therefore, the experimentation has beencarried out off-line by calculating the potentialenergy consumptions in the following way. It hasbeen assumed the maximum hourly nominal powerconsumption to be one, then the following quanti-ties have been calculated:

C100 ¼X

x1i; x1 2 0; 1f g ð6Þ

where x1i is the hourly power level for the ith

sample according to the no control strategy (nightpower level always at 100 %) and therefore C100

is its overall consumption.

CStaC ¼X

x2i; x2 2 0; 0:5; 1f g ð7Þ

where x2i is the hourly power level for the ith

sample according to the StaC strategy (Fig. 4)and therefore CStaC is its overall consumption.

CSmAC ¼X

x3

i; x3 2 0; 1½ � ð8Þ

where x3i is the hourly power level for the ith sample

0

200

400

600

800

1000

1200

1400

1600

1800

15 16 17 18 19 20 21 22 23 0 1 2 3 4 5 6 7

traf

fic

flo

w r

ate

hours

real statistic neural ensembling

Fig. 8 Model comparison (night detail)

Table 4 Control strategies comparison: energy saving

StaC (%) SmAC (%)

Street 1 25 44.5

Street 2 25 47

Street 3 25 37.5

Average 25 43


(Eq. 4) according to the SmAC strategy (Fig. 6) andtherefore CSmAC is its overall consumption.

These quantities have been calculated over3 months of actual traffic flow rates obtained by onstreet coil sensors. Thus, we computed the consump-tion saving of the StaC and SmAC strategies withrespect to the no control approach in the followingway:

SStaC ¼ 1� CStaC C100= ð9Þ

SSmAC ¼ 1� CSmAC C= 100 ð10ÞIn the following table, we report these values for

the three considered streets (Table 4).Results show that it is possible to save on average

43 % of energy, meaning that lamps will work at 57 %

of their nominal power having as inferior limit 50 %(Eq. 4 and Fig. 6) in order to avoid periods duringnormal operation with almost no light due to lightoutput drop.

From these results, it is clear that the SmAC ap-proach provides a remarkable improvement in termsof energy saving (43 % on average) in particular onstreets with medium–low traffic flow rate.

Moreover, in the following figure, an example ofhow the two strategies work is shown, where on the Yaxis, we report the normalized traffic flow rate valuesand the normalized hourly power consumptions of thedifferent strategies. From the figure, it is possible to seethat the SmAC strategy is capable to follow the realdemand (traffic flow rate) achieving the energy on de-mand concept. In particular, it is interesting to point outthat SmAC improves not only energy efficiency (orange

Fig. 9 Control strategies comparison


dotted area), but also safety (yellow dashed area) be-cause it provides light when actually needed (Fig. 8).

Conclusions

In this work, we proposed a new approach for adaptivestreet lighting control based on the energy on demandidea. In order to achieve this goal, it is critical to have areliable demand model, which in the case of street light-ing turns out to be a traffic flow rate forecasting model.

Thus, we showed a novel modeling approach basedon artificial neural networks ensembling in order toprovide a 1-h forecast of urban traffic flow rates.Experimentation has been carried out on three differ-ent classes of real streets and results showed that theproposed approach clearly outperforms the statisticalmethods (6 % prediction error) achieving a 3 % pre-diction error. The reason for that is that the neuralensembling model is capable to provide more reliableestimations when out of standard conditions because itconsiders the real traffic dynamics.

Moreover, the proposed adaptive control strategyhas been tested and compared to a traditional regula-tion system on the same streets (Fig. 9). Resultsshowed that the adaptive control provides, on average,almost doubled energy savings (43 vs 25 %).

Future work will firstly focus on dimming profilesaccording to international standards and then furthermodeling improvements (using more sophisticatedensembling methods as well as trying to develop hy-brid models) will be investigated, and lastly, the eco-nomic impact of the proposed methodology will becarried out.

References

Arbib, M. A. (1995). The handbook of brain theory and neuralnetworks. Cambridge: The MIT Press.

Avnimelech, R., & Intrator, N. (1999). Boosting regressionestimators. Neural Computation, 11, 491–513.

Bishop, C. M. (1995). Neural networks for pattern recognition(pp. 364–369). London: Oxford University Press.

Breiman, L. (1999). Combining predictors. In A. J. C. Sharkey(Ed.), Combining artificial neural nets—ensemble andmodular multi-net systems (pp. 31–50). Berlin: Springer.

Bucur, L., Florea, A., Petrescu, B.S. (2010). An adaptive fuzzyneural network for traffic prediction, Control & Automation

(MED), 18th Mediterranean Conference on, Marrakech,1092–1096.

Canca D., Larrañeta J., Lozano S. and Onieva L. (1997). Trafficintensity forecast in urban networks using a multilayerperceptron. Joint International Meeting EURO XV -INFORMS XXXIV. Barcelona, Spain.

Çetiner, B. G., Sari, M., & Borat, O. (2010). A neural networkbased traffic-flow prediction model. Mathematical andComputational Applications, 15(2), 269–278.

CIE 115 (1995) Recommendations for the lighting of roads formotor and pedestrian traffic.

CIE 115 (2010) Lighting of roads for motor and pedestrian traffic.Delft University of Technology. (2011). Intelligent street lighting

saves up to 80 % on energy. ScienceDaily, http://www.sciencedaily.com/releases/2011/07/110712093623.htm.Accessed 13 Jul 2011.

Dougherty M.S., and Cobbett M. (1994). Short term inter-urbantraffic forecasts using neural networks. Proc. SecondDRIVE-II Workshop on Short-Term Traffic Forecasting,Delft, The Netherlands, 65–79.

Drucker H. (1997) Improving regressors using boosting techni-ques. In: Fisher DH (ed.) ICML, Morgan Kaufmann, 107–115.

Fundazioa, E. (2011). New system of intelligent management ofstreet lighting enables 80 % savings in energy.ScienceDaily, http://www.sciencedaily.com/releases/2011/11/111111094558.htm, Accessed 16 Nov 2011.

Haykin, S. (1999). Neural networks: a comprehensive founda-tion (2nd edition). New Jersey: Prentice Hall.

Ishak, S., Kotha, P., & Alecsandru, C. (2003). Optimizationof dynamic neural network performance for short-termtraffic prediction. Transportation Research Record,1836, 45–56.

Jawanjal, S.; Bajaj, P. (2010). A design approach to traffic flowforecasting with soft computing tools. Emerging Trends inEngineering and Technology (ICETET), 2010 3rdInternational Conference on, Goa, 81–84.

Kohavi, R., & Bauer, E. (1999). An empirical comparison ofvoting classification algorithms: bagging, boosting andvariants. Machine Learning, 36, 105–142.

Krogh, A., & Vedelsby, J. (1995). Neural network ensembles,cross validation and active learning. In G. Tesauro, D. S.Touretzky, & T. K. Leen (Eds.), Advances in neural infor-mation processing systems (7th ed., pp. 231–238).Cambridge: MIT Press.

Liu, Y., & Yao, X. (1999). Ensemble learning via negativecorrelation. Neural Networks, 12(10), 1399–1404.

Pamuła, T. (2011). Road traffic parameters prediction in urbantraffic management systems using neural networks.Transport Problems, 6(3), 123–128.

Park, D., Rilett, L. R., & Han, G. (1999). Spectral basis neuralnetworks for real-time travel time forecasting. Journal ofTransportation Engineering, 125(6), 515–523.

Perrone, M. P., & Cooper, L. N. (1993). When networksdisagree: ensemble methods for hybrid neural net-works. In R. J. Mammone (Ed.), Neural networks forspeech and image processing. London: Chapman andHall.

Rea, M. S., Bullough, J. D., & Akashi, Y. (2009). Several viewsof metal halide and high pressure sodium lighting for


http://www.sciencedaily.com/releases/2011/07/110712093623.htm




outdoor applications. Lighting Research and Technology,41(4), 297–320.

Rosenblatt, F. (1957) The perceptron—a perceiving and recog-nizing automaton. Report 85-460-1, Cornell AeronauticalLaboratory.

Rosenblatt, F. (1961). Principles of neurodynamics: perceptronsand the theory of brain mechanisms. Washington: Spartan.

Taylor C., and Meldrum D. (1995). Freeway traffic data predic-tion using neural networks. Proc. 6th VNIS Conference,Seattle, WA, 225–230.

The Highland Council (2008). Transport, Environmental andCommunity Services Committee, reducing energy con-sumption—street lighting. Report no. TEC/87/08.

van Lint, J. W. C., Hoogendoorn, S. P., & van Zuylen, H. J.(2003). Freeway travel time prediction with state-spaceneural networks (pp. 30–39). Washington: NationalAcademies Press. Transportation research record 1811.

Zheng, W., Der-Horng, L., & Shi, Q. (2006). Short-term freewaytraffic flow prediction: Bayesian combined neural network ap-proach. Journal of Transportation Engineering, 132, 114–121.


smart street lighting management

Documents