Illumination control of LED systems based on neural network model and energy optimization algorithm

Energy and Buildings 62 (2013) 514–521

Contents lists available at SciVerse ScienceDirect

Energy and Buildings journal homepage: www.elsevier.com/locate/enbuild

Illumination control of LED systems based on neural network model and energy optimization algorithm Zizhen Wang, Yen Kheng Tan ∗ Energy Research Institute @ Nanyang Technological University, Nanyang Technological University, Research Techno Plaza, X-Frontier, Level 5, 50 Nanyang Drive, Singapore 637553, Singapore

a r t i c l e

i n f o

Article history: Received 10 August 2012 Received in revised form 15 March 2013 Accepted 18 March 2013 Keywords: Intelligent lighting control Neural network Constrained nonlinear optimization LED system

a b s t r a c t Lighting constitutes a large proportion of the main energy consumption loads of a building; energyefficient lighting control is an important topic to be addressed in achieving green building requirement. Within a building, huge amount of lights are being deployed in a distributed manner which poses great challenge in achieving energy saving and personalized lighting control. In this paper, the objective is to satisfy table illumination preference of each office user while minimize energy consumption of the overall lighting system by optimizing the illumination levels of the distributed luminaires. A holistic and scalable neural network model is developed to represent the complex relationship between dimming levels of luminaires and measured illuminance on the table. Based on the developed model, a lighting energy optimization algorithm is proposed to achieve energy saving while having personalized lighting control. The proposed model can serve as a base model for the improved artificial light and even daylight control system in the future study. © 2013 Elsevier B.V. All rights reserved.

1. Introduction With rapid urbanization trend in all cities in the world, there are many challenges to ensure that each city continues to be a livable place. In any building environment, there exists plenty of opportunity to obtain energy conservation. Lighting constitutes 20–38% of energy consumption in commercial buildings in the United States [1]. It is important to explore lighting approaches and technologies to save energy. Solid lighting technology like Lighting Emitting Diode (LED) is cost-competitive and energy-efficient and has large potential in replacing the traditional lightings such as fluorescent and hydrogen lamps [2]. Since dimming levels of a luminaire are related to the power it consumes, dimmable LED luminaires have great advantages in the energy efficient lighting system. Some studies [3–5] have shown that energy consumed by lighting can be conserved by designing an indoor lighting control system with dimming capability. In [5], Users were able to control the lights and adjust the illuminance of their tables, but the ability to control the lighting did not put users in a more positive mood or affect the performance on the task. Lighting simulation model had been used to simulate the indoor illuminance in [6–8]. In [9], simulation and neural networks were used to facilitate a

∗ Corresponding author. Tel.: +65 97922444. E-mail address: [email protected] (Y.K. Tan). 0378-7788/$ – see front matter © 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.enbuild.2013.03.029

responsive lighting control strategy. The objective of [10] was to investigate the accuracy of simulating the illuminance distribution from lighting and the electrical lighting consumption of an exist atrium building. Researchers of [11] compared the energy savings and effectiveness of combinations of occupant detection, light dimming and switching techniques in private offices. An Excel application in [12] was developed to calculate illumination levels and energy impacts of a building. The approaches for rendering illumination have been discussed in several studies [13–16]. In [17], researchers used Genetic Algorithm to real-time predict natural light levels at chosen points within a room. A fuzzy controller was developed in [18] to maintain the lighting illuminance level suitable for robotic manipulation in dynamic environments. In [19], the control of direct sun block to meet the workplane illuminance level was implemented using custom designed blind. In [20], a method was proposed to estimate and disaggregate illumination contributions of incoming light and the different LED sources at the workspace plane. In [21], lighting control was formulated as a linear programming problem to minimize energy usage and meet occupants’ lighting preferences at the same time. Energy-efficient lighting control can contribute great energy saving in a building context, and especially, large common office spaces have great potential. Typically, in a lighting control system, dimmable luminaires are installed and light sensors and occupancy sensors are placed to measure the illuminance and user occupancy. Some researches use simulation software to carry out the lighting

Z. Wang, Y.K. Tan / Energy and Buildings 62 (2013) 514–521

configuration in the test bed to establish the relationship between lights and table illuminance [10]. The drawback of this approach is that the performance of the simulation software could not imitate the real environment totally. For some researches, they implemented a few luminaires and sensors that researchers do not need to divide into groups. When considering a large office space with multiple lights and sensors, researchers also could put lights and sensors into different groups by placing partitions among them [22]. However, in many common offices, there are usually multiple lights and no higher partitions. The illuminance of a table in a test bed can be easily affected by several luminaires and the number of luminaires and tables increases as the test bed becomes bigger, it is difficult to group the luminaires with sensors and to get the optimal control results. In this situation, more energy can be conserved if the solution can be improved. When using illuminance sensors to assist in the feedback lighting control, the brightness of lights is changing, sometimes tremendously, before the lights become stable and it will affect users’ eyesight and working efficiency. And if illuminance sensors are placed on the office tables, they will affect the working users and vice visa, users may affect the function of sensors accidentally. In this paper, dimmable LED luminaires are implemented in our test bed. We use neural network technique to map the relationship among dimming levels of luminaires and table illuminance. Our goal is to minimize lighting energy consumption by adjusting the luminaire dimming levels since dimming level of a LED luminaire is proportional to energy it consumes. We also satisfy users’ various illumination preference and consider the non-office hour scenario when some tables are not occupied and more energy can be conserved. We use a nonlinear optimization method to obtain the optimal lighting performance and the minimum energy consumption. The neural network model we built is based on the mapping relationship between luminaire dimming levels and table illuminance. Thus, this model is currently applicable for artificial lighting control at night. In the future, we will establish a second model to map the relationship between day light and table illuminance and further combine it with the current model into a complete model for lighting control in the day time. In this paper, we only introduce this base model which presents the relationship between luminaire dimming levels and table illuminance. This paper is organized as follows. In Section 2, we introduce the test bed for implementing the dimmable LED luminaires. The neural network method used for mapping the relationship among luminaires and table illuminance is presented in Section 3. The lighting optimization solution with numerical results is shown in Section 4. Conclusions are drawn in Section 5.

Fig. 1. The LED lighting control test.

will calculate based on the measured illuminance information and determine the final dimming setting for all luminaires. Since the system uses overall control and optimization, no grouping among luminaires and tables is needed and minimized energy consumption will be achieved given that illuminance reference of each table is met. Our lighting control test bed is shown in Fig. 1. To show the illuminance distribution and isoline, a simulated test bed is created by using DIALux lighting simulation software, shown in Fig. 2. And the lighting layout is shown in Fig. 3. 3. System identification using neural network Since luminaires and tables are hard to group in a large, open, complex lighting system, it is difficult to make control decisions for all luminaires when facing different illuminance preferences, especially during the non-office hour at night. The illuminance of one table is normally affected by more than one luminaire and traditional control methods are unable to handle it well. Before the lighting control, we need to get familiar with the test bed and figure out the relationship among luminaire dimming values and table illuminance by using the logged data from a portable lux meter inputted into the developed Dialux simulation model. The distribution of the lighting flux in the room can then be visualized and understood. If we treat the test bed as the plant to be controlled, luminaire dimming values are seen as inputs of the plant

2. LED lighting control test bed Our test bed is an 8.5 m × 8.2 m office with installation of 9 * 54 W LED luminaires and 5 * 19 W LED luminaires (Philips BBS360 1×LED3500 NW (special: DC-DC) Luminaire efficacy: 65 lum/W) on the ceiling. These luminaires are all DALI controlled. There are 12 tables in the office. The numbering for the tables is shown in Fig. 15. Since there are no high partitions between tables, luminaires can affect the illuminance of more than one table. The office lighting illuminance followed in this setup is to be between 320 and 500 lux. All the LED luminaires used are dimmable and they can be controlled by the central controlling computer. The dimming level of a luminaire is proportional to its power consumption. The goals of this lighting control system are to satisfy different table users’ preference of illuminance and to minimize the overall power consumption of lighting during the office hour as well as non-office hour. During non-office hour, only the illuminance reference of the occupied tables will be considered and the optimization algorithm

515

Fig. 2. A simulated test bed created using DIALux.

516

Z. Wang, Y.K. Tan / Energy and Buildings 62 (2013) 514–521

with input neuron j, then the weight vector W1 in the Hidden layer can be defined as:

⎡ w1

1,1

1 w1,2

···

1 wL,1

1 wL,2

···

1 w1,M

⎤

⎢ w1 w1 · · · w1 ⎥ ⎢ 2,1 2,2 2,M ⎥ ⎥ W =⎢ ⎢ . . .. ⎥ . .. .. ⎣ .. . ⎦ 1

1 wL,M

2 as the weight that conAnd with the same way, we denote wi,j nects hidden neuron i with input neuron j, then the weight vector W2 in the output layer can be defined as:

⎡ w2

1,1

2 w1,2

···

2 wN,1

2 wN,2

···

2 w1,L

⎤

⎢ w2 2 2 ⎥ ⎢ 2,1 w2,2 · · · w2,L ⎥ ⎢ ⎥ W =⎢ .. .. ⎥ .. ⎣ ... . . . ⎦ 2

2 wN,L

The bias vectors in the Hidden layer and Output layer are denoted as b1 and b2 , respectively. Each of hidden neurons and output neurons has a bias variable. Fig. 3. The lighting layout of the test bed.

and tables’ illuminance can be seen as output of the plant. Luminaire dimming setting is the variable that can be controlled. Some researches [20,22] have presented the approach to establish the linear relationship between the dimming level of each luminaire and the illuminance of all the tables and thus incorporate them to map the whole relationship. The disadvantage of this approach is that with increased number of luminaires and tables, the complexity of the calculation is larger. In this paper, we used neural network developed in the MATLAB Simulink environment to identify our test bed and recognize the relationship between luminaires and table illuminance. Artificial neural networks are composed of individual elements and connections between elements. The connections between elements with different transfer functions largely determine the network function. A neural network can be trained to perform a particular function with the adjustment of the values of the connections. Typically, pairs of inputs and targets are put into the neural network training. The network is adjusted based on a comparison of the output and the target. Neural networks have been trained to perform complex functions in various fields, including pattern recognition, identification, classification, vision, and control systems. We collect the mapping data at night to avoid the human interference. Without knowing the relationship among luminaire dimming values and table illuminance, the plant looks like a black box. The neural network can solve this black box problem by using function approximation. A two-layer feed-forward network is used for mapping between a data set of numeric inputs, dimming levels of 14 luminaires and a set of numeric targets, the corresponding illuminance of 12 table given the dimming setting. Thus, the network is with 14 inputs and 12 outputs, shown in Fig. 4. Thus, the network is with M inputs, L neurons in the Hidden Layer and N outputs. The input, hidden neuron and output vectors are respectively: x = [x1 , x2 ,. . ., xM ]T h = [h1 , h2 ,. . ., hL ]T o = [o1 , o2 ,. . ., oN ]T where the input, hidden neuron and output of the network are M-dimensional, L-dimensional and N-dimensional vectors respec1 as the weight that connects hidden neuron i tively. If we denote wi,j

b1 = [b11 , b12 , ..., b1L ]

T

b2 = [b21 , b22 , ..., b2N ]

T

The sum of the weighted inputs and the bias b1 forms the input to the transfer function tansig. The function tansig generates outputs between −1 and 1 as the neuron’s net input goes from negative to positive infinity. And the output of the hidden layer and the bias b2 in the output layer forms the input of the second layer to the transfer function purelin that is a linear transfer function. Multiple layers of neurons with nonlinear transfer functions allow the network to learn nonlinear relationships between input and output vectors. Thus, the output o can be represented as: h = tansig(W 1 x + b1 ) =

o = W 2 h + b2 =

2 −1 1 + exp[−2(W 1 x + b1 )]

2W 2 + b2 − 1 1 + exp[−2(W 1 x + b1 )]

Now that the architecture of the multilayer network has been defined, the design process is described as following. Collect the data. Before the start of the network design process, we first collect and prepare sample data. It is generally difficult to incorporate prior knowledge into the network, therefore the network can only be as accurate as the data that are used to train the network. The goal of our network is to obtain the relationship among the entire luminaire dimming levels and the illuminance of all the tables. After obtaining the pairs of inputs and targets of the network, network will be trained and the desired outcome is that the actual output of the network is as accurate as the target. The input for training the network consists of dimming levels of M LED luminaires and the target consists of table illumination values of N tables. We create 600 random sets of input data. The number of output is 12. The range of dimming levels is from 40 to 99%. We use illumination meter to measure the table illuminance when a set of input data is given. The measured value will be one set of target given an input. Therefore, we finally get corresponding 600 sets of target data based on the input data. The accuracy of the measured data is verified by putting a repeated series of dimming levels in the order from 95 to 40% for every individual light, so that for every same set of dimming levels, measured illuminance output should

Z. Wang, Y.K. Tan / Energy and Buildings 62 (2013) 514–521

517

Fig. 4. The two-layer feed-forward network.

MSE =

Best Validation Performance is 20.2558 at epoch 36 5

10

Train Validation Test Best

4

10

Mean Squared Error (mse)

be more or less the same. And based on our assumption, the illuminance is proportional to dimming levels of lights so that we also ensure the measured data we got is accurate. Preprocess and postprocess the data. Neural network training can be made more efficient if we perform a certain preprocessing steps on the network inputs and targets. One step is to remove inputs/targets that are constant. Since random input is used, there may be some repetitive input rows in the data set and so the redundant rows will be removed. In our network, hyperbolic tangent sigmoid transfer function is used in the hidden layer. This function becomes saturated when the net input is greater than three (exp(−3) ∼ = 0.05). If this happens at the beginning of the training process, the gradients will be very small, and the network training will be very slow. It is standard practice to normalize the inputs before applying inputs/targets to the network. Generally, the normalization step is applied to both the input vectors and the target vectors in the data set. In this way, the network output always falls into a normalized range. Dividing the data. When training the network, we first divide the data into three subsets. The first subset is the training set, which is used for computing the gradient and updating the network weights and biases. The second subset is the validation set. The error on the validation set is monitored during the training process. The validation error and the training set error normally decrease during the initial phase of training. However, when the network begins to overfit the data, the error on the validation set typically begins to rise. The third subset is the test set. The test set error is not for training but used to compare other models. The ratios for training, testing and validation are 0.7, 0.15 and 0.15, respectively. Train the network. The training process requires a set of examples of proper network behavior – network inputs x and target outputs t. The process of training a neural network involves tuning the values of the weights W and biases b of the network to optimize network performance. The performance function for feedforward networks is mean square error (MSE) – the average squared error between the network outputs o and the target outputs t. The curve of the actual calculated MSE during the training process is shown in Fig. 5. The MSE is defined as follows:

3

10

2

10

1

10

0

10

0

5

10

15

20

25

30

35

Fig. 5. Mean square error of the network during the training process.

lights so that the lights will not affect other lights. In our test bed, 14 LED luminaires are arranged on the ceiling and 12 office tables are placed. No partitions are placed between tables so the total illuminance of one table achieved from all the LED luminaires is the aggregated superposition of the illuminance from each individual luminaire. The area of the test bed is 8.2 m × 8.5 m. The room is 2.8 m high and the work plane is 0.8 m high. We consider the illuminance on the work plane. Illuminance distribution of the test bed with full brightness of each luminaire is shown in Fig. 6.

1 1 (ei )2 = (ti − oi )2 N N N

N

i=1

i=1

Create and initialize the network. A two-layer feed-forward neural network with M neurons in the hidden layer is created. One hidden layer has already produced excellent results, so we do not try two hidden layers. The weights and biases of the network are also initialized. The regression and the best validation performance of the network are shown in Figs. 13 and 14. 4. Lighting control approach For the traditional lighting control solutions, controlled lights have their own controlling references. It is easy to control the lights since the control methods are implemented on the different lights. It is a feasible control solution in the environment where each light is placed far away with others or there are no partitions between

40

42 Epochs

Fig. 6. The illumination isoline with full brightness of luminaires.

518

Z. Wang, Y.K. Tan / Energy and Buildings 62 (2013) 514–521

Function value

Current Function Value: 403.806 800 600 400 200 0

0

10

20

30

40

50

60

Iteration Fig. 7. The final values of dimming levels and table illuminance and optimization process.

From the figure above, we can see that the central area of the work plane has a much higher illuminance than the rest area. If we control the lights one by one or group by group, it is difficult to meet various illumination preferences and maximize the energy conservation. Thus, we consider the system as a whole. We have established the relationship among luminaire dimming levels d = [d1 , d2 ,. . ., dM ], the M × 1 vector, and table illuminance t = [t1 , t2 ,. . ., tN ], the N × 1 vector, so now we consider the lighting control solution so as to achieve illumination preference with minimum power consumption. The traditional single-inputsingle-output control methods are not suitable for our case. Global optimization approach is what we look at. It is able to consider multiple variables and perform the optimal outcome subject to the constraints we set. The objective function is the total energy consumed by the whole test bed which we try to minimize. The total energy consumed by all the luminaires can be defined as, E=

M 

Ei =

1

M 

Pi di

1

where E is the overall energy consumption, Ei is the energy consumed by the ith luminaire, Pi is the rated power of the ith luminaire with full brightness, and di is the current dimming level of the ith luminaire, where 0 ≤ di ≤ 1. The lowest dimming level is 0 and the highest dimming level is 1. M is the number of luminaires in the test bed. Since luminaires are dimmable and dimming levels are proportional to the power consumption of each luminaire, the total power consumption can be minimized by dimming down the luminaires.

Fig. 9. The isoline under the optimal control.

The objective of minimization of energy consumption can be seen as minimization of luminaire dimming levels. Thus, we can determine the optimum dimming vector d that solves E = min

Pi di

1

Table Illuminance

500

Actual Desired

Actual Desired

450 400

400 350

Illuminance/Lux

Illuminance/Lux

M 

s.t. 0≤d≤1 0≤r≤t t = net (d) where r = [r1 , r2 ,. . ., r12 ] is the preferred illuminance vector and the output illuminance vector t = [t1 , t2 ,. . ., t14 ] should be equal to or larger than its preferred one ri . net is the neural network model built earlier to map the relationship among luminaire dimming levels and table illuminance. And M is the number of luminaires. The table illuminance t and luminaire dimming level d has a nonlinear relationship established by the neural network model discussed earlier, the constraints of this problem are nonlinear and

Table Illuminance

300 250 200

350 300 250 200 150

150

100

100

50

50 0

Ei = min

1

500 450

M 

0 1

2

3

4

5

6

7

8

9

10

11

12

Table No. Fig. 8. The actual final table illuminance vs. reference value.

1

2

3

4

5

6

7

8

9

10

11

12

Table No. Fig. 10. The actual final table illuminance vs. reference value in non-office hour scenario.

Z. Wang, Y.K. Tan / Energy and Buildings 62 (2013) 514–521

Fig. 11. The non-office hour scenario created by DIALux. Fig. 12. The isoline of the non-office hour scenario.

Fig. 13. The regression result of the trained network.

519

520

Z. Wang, Y.K. Tan / Energy and Buildings 62 (2013) 514–521

Fig. 15. The numbering for the tables in the testing office. Fig. 14. The best validation performance of the trained network.

a nonlinear constraint optimization approach is used in our study. We used nonlinear programming optimization Algorithm to solve the problem. 5. Experimental results 5.1. Energy saving evaluation of LED lighting system with smart control in office hour The experiments conducted to evaluate the energy saving level of the proposed DC networked lighting with smart control are accomplished in the settings: (a) existing 9 sets of AC lighting with 2 × 28 W 4 T8 lamps and (b) proposed LED lighting controlled by intelligent control algorithm in the PC. Given the reference illuminance value of 350 lux [23] required for each table, the energy consumption of the fluorescent lighting system is 0.52 kWh while the smart LED system is only 0.4 kWh, 22% less energy consumption than its counterpart lighting system; the final energy consumption of the LED lighting system is 404 W, shown in Fig. 7. From the following Fig. 8, we can see that the illuminance reference of all 12 tables have been reached or beyond. Due to the layout of luminaires and tables as well as influence of a luminaire to multiple tables, it is difficult to achieve the exact illuminance target for every table. Thus, some tables will inevitably get higher illuminance. Thus, when there is a conflict that users request for different preferable table illuminance, we ensure the actual lighting users receive is higher than users’ lighting preference. If users require the illuminance level to be lower than some value but have to conflict with others, we can guarantee lighting illuminance levels are higher than preference. In Fig. 9, the isoline, the illuminance distribution of the table under the optimization control is shown. Most areas of this test bed have obtained 320 lux illuminance or higher than 320 lux Since the isoline is the result of the lighting simulation, there is difference between the simulation and the real result in Fig. 8. We want to present an overall idea of the lighting solution, and the isoline graph shows a reasonable outcome. 5.2. Non-office hour solution scenario During the non-office hour period, we assume, for example, only Table 5 is occupied during the non-office hour. The illuminance

preference for Table 5 is set as 500 lux, and those for the rest of tables are 0 lux. The illuminance expectation for Table 5 has been reached after the optimization process, shown in Fig. 10. Since luminaires in the test bed can influence more than one table, tables with the reference of 0 lux also have some illuminance, but very small compared to Table 5. The scenario isoline and illuminance of each table can be seen in Figs. 11 and 12, respectively. In non-office hour scenario, normally only a few users are in the office. The analysis for that is not a must but we would like to show the effect for that scenario. We arbitrarily chose that specific table, and the performance for others is similar. The simulation result is a little lower than the real performance, resulting from the imperfection of simulating the real test bed. 6. Conclusion We developed a solution for overall illumination control of a LED system based on a neural network mapping model. We presented an accurate model for simulating the relationship among luminaire dimming levels and table illuminance of the test bed. Based on this model, the illumination optimization approach adjusted the dimming levels of luminaires so as to achieve energy conservation and satisfy users’ various illuminance preference. We used the simulation just to give readers an overall control effect. We also showed different scenarios to verify the performance of our solution. The advantages of this illumination control system are firstly, it achieves an optimal performance thanks to the constrained nonlinear optimization algorithm. No specific grouping among luminaires and tables is needed. Secondly, after training the neural network and completing the input-output mapping of the test bed, the network model can replace the test bed as the controlled object and no ambient light sensors are needed to measure the illuminance of table and feed the measurement back to the controller. Thus, luminaires will not change dimming values during the controlling process or affect users who are working at the table. The challenges of the system are (1) during the neural network mapping process, the more tables in the test bed, the more sensors needed for the mapping, and the more pairs of data needed to collect for training the network; (2) if the layout of any luminaire or table is changed, the network model can be no longer used. Data mapping and neural network training must be conducted again.

Z. Wang, Y.K. Tan / Energy and Buildings 62 (2013) 514–521

Acknowledgment We would like to thank Huynh Truc Phong for her help in conducting the experiments.

[11]

[12]

References

[13]

[1] U.S. Energy Information Administration, Lighting in commercial buildings. Available: http://www.eia.gov/emeu/cbecs/cbecs2003/lighting/lighting1. html, April 2009. [2] M.S. Shur, A. Zukauskas, Solid-state lighting: toward superior illumination, Proceedings of the IEEE Oct (93) (2005) 1691–1703. [3] J.M. Alonso, J. Ribas, J.J.D. Coz, A.J. Calleja, E.L. Corominas, M. Rico-Secades, Development of a distributive control scheme for fluorescent lighting based on LonWorks technology, IEEE Transactions on Industrial Electronics 47 (2000) 1253–1262. [4] M.R. Atif, A.D. Galasiu, Energy performance of daylight-linked automatic lighting control systems in large atrium spaces: report on two field-monitored case studies, Energy and Buildings 35 (2003) 441–461. [5] P.R. Boyce, N.H. Eklund, S.N. Simpson, Individual lighting control: task performance, mood and illuminance, Journal of the Illuminating Engineering Society 29 (2000) 131–142. [6] D.H.W. Li, E.K.W. Tsang, An analysis of measured and simulated daylight illuminance and lighting savings in a daylit corridor, Building and Environment 40 (2005) 973–982. [7] A. Mahdavi, S. Chang, V. Pal, Exploring model-based reasoning in lighting systems control, Journal of the Illuminating Engineering Society 29 (2000) 34–40. [8] A. Mahdavi, P. Mathew, S. Kumar, V. Hartkopf, V. Loftness, Effects of lighting, zoning, and control strategies on energy use in commercial buildings, Journal of the Illuminating Engineering Society 24 (1995) 25–35. [9] S. Chang, A. Mahdavi, A hybrid system for daylight responsive lighting control, Journal of the Illuminating Engineering Society 31 (2002) 147–157. [10] A.D. Galasiu, M.R. Atif, Applicability of daylighting computer modeling in real case studies: comparison between measured and simulated daylight

[14]

[15]

[16] [17] [18]

[19]

[20]

[21]

[22]

[23]

521

availability and lighting consumption, Building and Environment 37 (2002) 363–377. J.D. Jennings, F.M. Rubinstein, D. DiBartolomeo, S. Blanc, Comparison of control options in private offices in an advanced lighting controls testbed, Journal of the Illuminating Engineering Society 29 (2000) 39. L. Heschong, J. McHugh, Skylights: calculating illumination levels and energy impacts, Journal of the Illuminating Engineering Society 29 (2000) 90–100. D. Li, K.L. Danny, S.L. Cheung, L. Wong, Tony, An analysis of energy-efficient light fittings and lighting controls, Applied Energy 87 (2010) 558–567. F.M. Rubinstein, J. Jennings, D. Avery, S. Blanc, Preliminary results from an advanced lighting controls testbed, Journal of the Illuminating Engineering Society 28 (1999) 130–141. F.M. Rubinstein, J.M. Siminovitch, R.R. Verderber, Fifty percent energy savings with automatic lighting controls, IEEE Transactions on Industry Applications 29 (1993) 768–773. S. Stannard, D. Keith, K. Johnson, Predicting the performance of controlled lighting systems, Journal of the Illuminating Engineering Society 26 (1997) 69–78. D.A. Coley, J.A. Crabb, An artificial intelligence approach to the prediction of natural lighting levels, Building and Environment 32 (1997) 81–85. S.Y. Chen, J.W. Zhang, Intelligent Lighting Control for Vision-Based Robotic Manipulation”, IEEE Transactions on Industrial Electronics 59 (2012) 3254–3263. D.L. DiBartolomeo, E.S. Lee, F.M. Rubinstein, S.E. Selkowitz, Developing a dynamic envelope/lighting control system with field measurements, Journal of the Illuminating Engineering Society 26 (1997) 146–164. A. Pandharipande, D. Caicedo, Daylight integrated illumination control of LED systems based on enhanced presence sensing, Energy and Buildings 43 (2011) 944–950. Y.J. Wen, A.M. Agogino, Wireless networked lighting systems for optimizing energy savings and user satisfaction’, IEEE Wireless Hive Networks Conference (2008) 1–7. Y.J. Wen, A.M. Agogino, Personalized dynamic design of networked lighting for energy-efficiency in open-plan offices, Energy and Buildings 43 (2011) 1919–1924. CIBSE (Chartered institute of building services engineers), Code for Interior Lighting, Lighting Guide 7: Office Lighting, 2005.