An energy-efficient predictive control for HVAC systems applied to

Energy and Buildings 152 (2017) 409–417

Contents lists available at ScienceDirect

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

An energy-efficient predictive control for HVAC systems applied to tertiary buildings based on regression techniques Diana Manjarres ∗ , Ana Mera, Eugenio Perea, Adelaida Lejarazu, Sergio Gil-Lopez Tecnalia Research and Innovation, Parque Tecnológico de Bizkaia, Derio, Spain

a r t i c l e

i n f o

Article history: Received 12 April 2017 Received in revised form 4 July 2017 Accepted 17 July 2017 Available online 24 July 2017 Keywords: Heating ventilation and air conditioning (HVAC) Building energy management system (BEMS) Energy efficiency Optimisation algorithms Building automation Thermal comfort Supply air temperature optimal control

a b s t r a c t Heating ventilation and air conditioning (HVAC) systems represent an important amount of the total energy use in office buildings, accounting for near 30%. Moreover, in countries affected by extreme climates HVAC systems’ contribution to energy demand increases up to 50%. Therefore, the automation of energy efficient strategies that act on the Building Energy Management System (BEMS) in order to improve building energy use becomes increasingly relevant. This paper delves into the devising of a novel HVAC optimization framework, coined as Next24h-Energy, which consists on a two-way communication system, an enhanced database management system and a set of machine learning algorithms based on random forest (RF) regression techniques mainly focused on providing an energy-efficient predictive control of the HVAC system. Therefore, the proposed framework achieves optimal HVAC ON/OFF and mechanical ventilation (MV) schedule operation that minimizes the energy consumption while keeps the building between a predefined indoor temperature margins. Simulation results assess the performance of the proposed Next 24 h-Energy framework at a real office building named Mikeletegi 1 (M1) in Donostia-San Sebastian (Spain) yielding to excellent results and significant energy savings by virtue of its capability of adapting the parameters that control the HVAC schedule in a daily basis without affecting user comfort conditions. Specifically, the energy reduction for the test period is estimated in 48% for the heating and 39% for the cooling consumption. © 2017 Elsevier B.V. All rights reserved.

1. Introduction As mentioned by the Independent Statistics & Analysis of the U. S. Energy Information Administration in 2015, office buildings represent nearly one-fifth of all delivered energy consumed by commercial buildings, and are therefore an important focus for energy efficiency improvements. In fact, median savings from the existing building commissioning was 15% of the whole building energy use [1]. These energy improvements usually depend on the optimal operation strategy implemented at the building energy management system (BEMS), which is the agent responsible for reducing the energy consumption while ensuring user comfort conditions. Nowadays there is a lack of approaches unifying both developments, i.e. the design of optimal energy efficient strategies and its integration for interacting with BEMS. In this context, authors in [2] state that there is already a gap between sensor deployment infrastructures and facility manager’s real actuations.

∗ Corresponding author at: Parque Científico y Tecnológico de Bizkaia C/Geldo, Edificio 700, E-48160 Derio, Bizkaia. E-mail address: [email protected] (D. Manjarres). http://dx.doi.org/10.1016/j.enbuild.2017.07.056 0378-7788/© 2017 Elsevier B.V. All rights reserved.

As widely known, energy demand in office buildings has a large variability depending on the time and the day of the week which severely affects the energy consumption. Consequently, in order to maximize energy savings in an office building’s domain, the automation of optimization procedures which dynamically adapt the HVAC operation mode to the indoor and outdoor conditions becomes essential. Otherwise energy improvements highly depend on an important human component, i.e. the facility manager of the building, which should act on the BEMS at specific times for ensuring energy efficiency. In this context, previous studies [3–10] deal with the dynamic optimization problem for building heating and cooling systems based on three main approaches: (1) physical based models; (2) grey-box models; and (3) black box models. From the first to the latest the knowledge of the structural building characteristics, the computational complexity and the effort to be implemented is decreased while generalization capabilities are gradually increased. The first two approaches involve static conditions based on complex differential equations, heat and mass balance equations and a deep knowledge of the thermal characteristics of the building. This results in a large engineering knowledge, intensive computer simulations and a need of considerable simulation time to

410

D. Manjarres et al. / Energy and Buildings 152 (2017) 409–417

model the building in order to provide accurate predictions. In this sense, authors in [11] propose optimization techniques based on dynamic programming providing an ON/OFF schedule and temperature set-points for optimizing the cost function and avoiding peak electrical demand. Nevertheless, in this case the authors assume non-realistic conditions, such as a perfect weather forecast, and also a deep knowledge of thermal building characteristics. In order to characterize the thermal building performance, authors consider the heat balance equations whose parameters are determined by means of employing the building heat transfer simulation program HASP/ACLD/8501. Authors in [12] use the TRNSYS dynamic model for the evaluation of energy saving techniques in HVAC systems, such as heat pumps with heat recovery. In the same line of research, DOE-2 is another building energy analysis program especially designed for allowing engineers to perform design studies of building energy use under certain weather conditions [13]. [14] in his thesis employs DOE-2 as a fitness function providing building energy and cost calculation to be hybridized with a Genetic Algorithm as an optimization procedure. The proposed grey-based algorithm optimizes building control parameters such as: indoor temperatures, shade position, artificial light power, and outdoor air ventilation rates for an entire building achieving savings between 10 and 30% in a typical building in Montreal. With the aim at reducing the computational complexity of the hybrid schemes, the authors in [15] propose a Neural Network (NN) approach as a multiparametric non-linear forecast procedure. In this case DOE-2 is employed as a tool for generating the training and testing data to be used in a mixed expert system based on a fuzzy-neural assistant prediction, which has shown significant advantages for optimizing building energy use as a dynamic method. Another highly employed solution is to call EnergyPlus module as an hourly energy building simulation. EnergyPlus is a modular loop based method which estimates thermal zone loads based on heat balance equations [17]. Dealing with the numerous drawbacks of physical based approaches, authors in [18] explain the causes of the cost function’s discontinuities present in the above-mentioned building simulation programs and evaluate a set of algorithms for such optimization problems with different smoothness levels. The authors show that due to the computational complexity of these methodologies, stochastic optimization algorithms are employed using a reduced set of simulations which conditioned the probability to be stacked at a local minimum, reducing their accuracy. As mentioned in [19] practical applications demonstrate that there exist large discrepancies in results from different models using distinct Building Energy Modeling programs. Therefore, the authors in this work summarize methodologies, processes and the main assumptions of three widely used simulation tools which could be responsible for those discrepancies. Due to the above-mentioned limitations of both physical and grey box approaches, in the last decade an upsurge of black box models based on machine learning procedures [20–24] has emerged. More specifically, in [23] the authors present a Support Vector Machine black box model to learn the features of the thermal model of a room. More recent studies propose the integration of these approaches with the existing BEMS using a specialized web based approach [24]. The main advantage of analytical approaches based on machine learning procedures is that they operate as black box models, learning the relationship between inputs (environmental magnitudes) and outputs (control parameters) based only on the study of previously recorded data. Consequently, if it is assumed that the dynamic relationship between the weather and occupancy magnitudes which affect building energy consumption is recorded in the historical data, the use of black box models provides higher generalization capabilities without the necessity to spend a huge amount of time and effort developing physical or grey-box based

models. Another advantage of black box models is that they can be adapted to dynamic responses of the building due to possible buildings’ structure changes, such as windows replacement or others affecting building energy consumption, just selecting the adequate training data set which is affected by these structural changes. This paper advances over the state of the art in building energy management optimization by presenting a novel framework, hereafter coined as Next 24 h-Energy, for providing an optimal energy-efficient predictive control of the HVAC system in terms of two main actuators: (1) HVAC ON/OFF and (2) Mechanical Ventilation (MV) schedule operation optimization. The main core of the Next 24 h-Energy framework is based on an internal temperature estimation, based on Random Forest (RF) techniques, with utilizes a minimal set of measured magnitudes trying to provide its maximum generalization capability to other locations or building characteristics. The automation of this optimization procedure integrated with the BEMS is implemented in a real office-building in Mikeletegi 1 (M1) in Donostia-San Sebastian (Spain) demonstrating its energy savings’ capabilities. In order to perform the simulations, specific indoor conditions of office buildings in Spain based on the Royal Decree 1826/2009 in [25] are considered. In summary, it defines among other conditions indoor temperatures between 21 and 26 ◦ C (Celsius degrees). The paper is organized as follows: Section 2 presents the proposed methodology: M1 office building characteristics, the HVAC description and the test deployment architecture and the proposed Next 24 h-Energy optimizer. Section 3 presents the simulation and tests results of the proposed approach and outlines the energy savings results. Finally, Section 4 ends the manuscript by presenting the conclusions extracted from this work and by outlining future research lines.

2. Methodology 2.1. Site M1 office building characteristics Mikeletegi 1 (M1) is a 8500 m2 office building located in Donostia-San Sebastian, Spain. The building area is divided in three floors that are vertically split into two identical parts (North and South orientation) by the common areas. Fig. 1 depicts the North side of the second floor area under which the simulation tests have been carried out. It can be observed that the main area is an open plan space with some cellular offices and meeting rooms located in the perimeter areas.

2.2. HVAC system description and test deployment architecture The main HVAC (heating ventilation and air conditioning) system is a centralized system consisting of an Air Handling Unit (AHU) which provides a constant tempered fresh air to all the spaces (open plan and perimeter areas). The final elements in the open plan area consist of 4 pipe Fan Coil Units (FCU). The AHU preconditions the outdoor air at a temperature set point by the facility manager and the FCUs then further provide heating or cooling to reach the set point temperature. In the open plan area there are three room thermostats that control three areas. The perimeter offices and meeting rooms also receive fresh air from the same AHU. The heating and cooling required in the AHU and FCU are produced by two reversible air to water heat pumps. A centralised Building Manager System (BMS) controls parameters of the HVAC devices (heat pumps, AHU and fan coils). Table 1 presents in detail the elements that form the HVAC system at M1 building considering the different sections: generation, distribution and control.


411

Fig. 1. Floor plan of the Mikeletegi 1 (M1) case study in Donostia-San Sebastian (Spain).

Table 1 HVAC system elements at Mikeletegi 1 (M1) building. System

Equipment

Generation

Cooling: 1 Heat pump Heating: 1 Heat pump

Distribution

Centralised air and water: Fresh air AHU + 4 pipe FCU

Control

Centralised BMS

Fig. 2 shows the centralised generation of heating and cooling and its distribution inside the building. It can be observed that the

hot and cold water is centrally produced and distributed to the AHU coils and to each level, to feed the 4 pipes FCUs. In order to train the Next 24 h-Energy optimizer and be able to learn the building and the HVAC system operation performance, real measured data is needed in an hourly basis. For that purpose, several activities have been carried out during the test deployment phase prior to the simulation process, such as: the installation of additional sensors and meters in the building, the provision of connectivity with the BEMS, the creation of a Database for storing the measured data, among others. First of all, it is important to make an assessment of the requirements of the HVAC system, its controls and

Fig. 2. HVAC system schematic diagram of the case study in Mikeletegi 1 (M1), Donostia-San Sebastian (Spain).

412


Fig. 3. Communications architecture for the proposed Next 24 h-Energy framework.

communication protocols. Then, several procedures are identified which are summarized in the following points: • Identification, selection and purchase of additional sensors and meters, such as: relative humidity and temperature sensors for indoor and outdoor environment, electricity and thermal energy meters. • Creation of a Database for registering the new monitored data. • Connection with the BEMS for programming the Next 24 hEnergy optimizer outputs. • Development of a communication architecture for new sensors. • Purchase and installation of new equipment. • A service automatisation for the weather forecast provision. Once the initial assessment of the building is completed and the above-mentioned actions deployed, a set of real measured data coming from different sources, i.e. sensors, meters, etc., is stored at the Database. Therefore, relationships between indoor and outdoor magnitudes (temperature and relative humidity sensors) and electrical and thermal energy consumptions (electrical and thermal meters) for different HVAC system conditions are considered in the proposed Next 24 h-Energy optimizer. More specifically, Fig. 3 depicts the Next 24 h-Energy communications’ architecture implemented in M1 building in order to feed the proposed Next 24 h-Energy Optimizer. As it can be seen, it relies on existing sensors and BEMS, includes new sensors and runs over the existing building infrastructure. It is important to note that the deployed infrastructure does not substitute any existing SCADA, BEMS or local intelligence, but relies on them, and builds on top of them a more robust and smartness system, looking at wider optimization time horizons and making use of inputs like weather forecast, occupancies, or energy prices. In this context, additional sensors and meters in M1 building have been installed in order to develop and train the proposed Next 24 h-Energy optimizer. Although not all sensors are required to train the model, both the investment and the engineering instal-

lation work have been done in order to have a perfect control of the building’s thermodynamics and energy balances during the validation process.

• Extra temperature and relative humidity sensors for monitoring outside and indoor environment. These sensors are connected via a ZigBee to Modbus/IP gateway. • Also a couple of relative and temperature sensors are installed in the entry and in the outlet conduct of the Air Handling Unit to control de air renovation proportion and enthalpy, as well as the supplied air enthalpy of the air mass entering the building. These sensors are connected to the Next 24 h-Energy framework through a M-Bus to BACNet/IP gateway. • Electric energy meters are installed in the heat pumps electric supply cables and in the Air Handling Unit’s fans supply cable, and thermal energy meters are mounted in the thermal main pipes. The connection of those meters to the system is provided through a M-Bus to Modbus/IP gateway. • The M1 building is equipped with thermal meters at each floor for providing the energy consumption of the fan-coils and roof tops. Then, specific M-bus and LonWorks gateways are employed to connect the measurements to the Next 24 h-Energy framework architecture.

Finally, all data measured from the sensors and also the outputs of the Next 24 h-Energy optimizer are stored in a MySQL Database built ad-hoc for the Next 24 h-Energy module. Two REST APIs are developed in order to separate the functionalities (access to database and to the Next 24 h-Energy optimizer implemented in Python) and a main process holds the control to integrate and connect the different software pieces into a single Next 24 h-Energy framework architecture.


413

Table 2 Mikeletegi 1 (M1) HVAC system normal operation schedule. Period

Winter

Summer

Mon Tue–Thur Frid

02:30–17:30 05:30–17:30 05:30–15:30

02:30–15:30 05:30–15:30 05:30–15:30

2.3. HVAC normal operation schedule The AHU in normal operation, i.e. full HVAC mode, consists on providing tempered air at the temperature set point selected by the facility manager. The air is supplied and heated or cooled as required to reach the room set point temperature indicated by the facility manager in each zone. The normal working schedule of the main HVAC system is a Monday to Friday working schedule, shown in Table 2. It can be seen that the facility manager used a regular periodic HVAC time schedule. As can be observed the HVAC system is started earlier on Mondays (2:30 h) than the other days of the week (5:30 h). This is due to the fact that as it is an office building, the HVAC system remains off at weekends, so much time is needed to reach the temperature set point. At summer period the office building is closed at 15:30 h which directly affects to the OFF schedule time of the HVAC system. It is important to remark that this HVAC schedule is fixed for all days and is independent of the contextual conditions (i.e. outdoor temperatures) that affect the indoor temperature of the office building. 2.4. Proposed Next 24 h-Energy optimizer In this subsection the current HVAC operation schedule is first described and then compared to the proposed optimized operation of the Next 24 h-Energy optimizer. The main difference is its capability of achieving the optimal HVAC ON/OFF and Mechanical Ventilation (MV) schedule operation for the following day based on the weather forecast for the next 24 h and not only on an automatically or manually implementation as the one presented in Table 2 for the M1 building. 2.4.1. Current HVAC operation As mentioned in Section 1 the operation of an HVAC system is normally controlled by a facility manager whose expertise in the area makes him/her capable for suggesting the ON/OFF HVAC schedule. Thus, the success of the HVAC performance highly depends on the knowledge and ability of the facility manager for a particular building. This procedure is manually or automatically performed and adapted to each season, for which the facility manager ensures that indoor conditions are between the comfort ranges that specifies the Royal Decree 1826/2009 [25] which stipulates indoor temperatures between 21 ◦ C and 26 ◦ C (Celsius degrees). Nevertheless, the building performance is not customized for each day of the season but a general schedule is usually proposed (see Table 2). 2.4.2. Proposed HVAC system operation In order to overcome this issue, the proposed Next 24 h-Energy optimizer presents a way to obtain an optimal ON/OFF HVAC schedule dynamically adapted to the current building conditions and weather forecast information with the aim at minimizing the HVAC energy consumption. The main core of the proposed Next 24 h-Energy optimizer is based on an intelligent machine learning method based on random forest (RF) [26]. RF technique is a non-linear multiparameter regressor which tries to infer the relationships between internal and external temperatures and the HVAC historical loads. There-

Fig. 4. Schema of inputs and outputs of the proposed Next 24 h-Energy optimizer.

fore, the building and HVAC system are modelled by means of a black-box model with the aim at inferring the estimated indoor temperature for the next 24 h. Consequently, the Next 24 h-Energy optimizer obtains the ON/OFF and the mechanical ventilation schedule of the HVAC system for the next 24 h with the aim at maximizing energy efficiency and reducing energy consumption, while maintaining the indoor temperature of the building between the predefined comfort margins. Regarding the mechanical ventilation (MV) mode, it has been shown to be extremely useful at night during the summer period when the outdoor temperature is below the indoor temperature of the building and therefore the fresh outside air can be used to cool down the building at reduced energy cost. Likewise, it can also be suitable for spring, autumn or mild winter period in which the outdoor temperature during daytime remains near the air-supply temperature. In fact, the main advantage of this mode is that since heating and cooling coils are switched off, the energy consumption required to operate in the MV mode is the energy to supply and return the air through the AHU and the ductwork reducing to more than half the energy consumption with regard to the full HVAC mode operation. Consequently, the overall expected energy savings by the application of the MV mode are due to a reduction in the number of operating hours of the heat pumps, and therefore in the number of hours of air conditioning (in full HVAC mode). Although an increase in the number of hours of AHU (in MV mode), that is, air pumps, is expected, it is important to remark that energy savings due to the minimization of heat pumps operation hours are much greater than the energy increase of pumping air during more hours. Fig. 4 depicts the schema of inputs and outputs of the machine learning method based on RF used in the proposed Next 24 hEnergy optimizer. In order to obtain the optimal operation schedule of the HVAC system and assess the different HVAC operation mode possibilities, the estimation of the indoor temperature for the next 24 h in terms of the building (indoor temperature, occupancy levels, relative humidity) and also external conditions (outdoor temperature and relative humidity) is computed. Based on that estimation of the indoor temperature for the considered HVAC operation modes and ON/OFF schedules the proposed Next 24 h-Energy approach selects the one which achieves an optimal performance in terms of energy consumption. Algorithm 1.

Proposed Next 24 h-Energy optimizer approach

414

D. Manjarres et al. / Energy and Buildings 152 (2017) 409–417 Table 3 Absolute and relative statistical error results (Min/Mean/Std) for the indoor temperature estimation by the Next 24 h-Energy framework.

Absolute error (◦ C) Relative error (%)

Min

Mean

Std

0.26 1.06

0.43 1.67

0.13 0.52

there are other situations (i.e. extremely cold or hot days) that the HVAC system must be started before the normal operation schedule for maintaining the indoor temperature of the building between the comfort margins. Similarly, the OFF hour is also computed by means of an iterative process assuring the system is switched on in full HVAC mode the minimum number hours while ensuring comfort condition until the working stop hour (IOFF ) of the building. Thus, MV mode is also preferred to full HVAC mode in order to minimize the energy use. 3. Simulation results and discussion 3.1. Test results

More specifically, in Algorithm 1 the steps of the proposed Next 24 h-Energy optimizer are summarized. In general, the Next 24 hEnergy optimizer is executed at night (i.e. at 23:30 h) and the seed corresponds to the measured indoor temperature at that time. The automatization of the weather forecast service allows having the outdoor temperature forecast for the next 24 h, which as shown in Fig. 4, is an input of the Next 24 h-Energy optimizer. In order to obtain the ON/OFF HVAC and MV schedule for the following day, the Next 24 h-Energy approach employs an iterative process in which it is assumed to start the HVAC system and the mechanical ventilation at certain times prior to the working start time (ION ) of the building. Note that it is of utmost importance to ensure comfort conditions in the office building during the working hours. As MV mode reduces the number of operating hours of heat pumps, it is prioritized over the full HVAC mode, i.e. if only with MV the comfort is ensured, the air conditioning in full HVAC mode will be started later. Therefore, the amount of time in full HVAC mode is reduced and the comfort guaranteed with less energy consumption. Nevertheless,

As stated in Section 2, the scenario for deploying the tests of the proposed Next 24 h-Energy framework has been the second floor of the M1 office building in Donostia-San Sebastian (Spain). The proposed Next 24 h-Energy optimizer has been trained with real measured data, i.e. indoor, outdoor temperatures and relative humidities and occupancy levels, and different simulations have been employed considering both full HVAC and mechanical ventilation modes. Specifically, a set of 63 days in summer and 46 days in winter have been taken into account for the simulation and the test phase. The accuracy of the proposed Next 24 h-Energy framework is assessed by means of comparing the real measured indoor temperature after the application of the proposed schedule and the indoor temperature estimation obtained by the Next 24 h-Energy optimizer when executed the day before. Table 3 presents the statistical results (Min/Mean/Std) in terms of indoor temperature error estimation of the Next 24 h-Energy approach. As can be observed the absolute and relative errors for the estimation of the indoor temperature are extremely low, being the mean absolute error below 0.5 ◦ C. Moreover, the small value for the error standard deviation reveals that the indoor temperature estimation is accurate and stable enough, i.e. the proposed Next 24 h-Energy approach does not present a high dispersion in the estimation output. Figs. 5–7 depict the boxplot representations of the HVAC electric consumption for different HVAC operation modes. Note that boxes delimit the lower and upper quartiles, the medians are depicted with a solid line and the outliers are marked with asterisks. On the one hand, Fig. 5 presents the historic HVAC consumption values with the normal operation schedule of Table 2. On the other hand, Fig. 6 presents the MV consumption values after the application of the proposed Next 24 h-Energy optimizer. Note that at certain hours, i.e. early in the morning from 5:00 h to 9:00 h and late at the afternoon from 16:00 h to 18:00 h the electric consumption has been considerably reduced by the utilization of the mechanical ventilation mode. Finally, Fig. 7 depicts the HVAC consumption values for the suggested ON/OFF Next 24 h-Energy schedule. While comparing towards the historic HVAC consumption, it is important to note that by means of the proposed Next 24 h-Energy approach, the electrical consumption is minimized at certain hours, i.e. early in the morning from 5:00 h to 8:00 h and late in the afternoon from 16:00 h to 18:00 h. This is due to the possibility of delaying the HVAC system power up and advancing the shutdown early in the afternoon. Thus, the building is capable of keeping the indoor


Fig. 5. Boxplot of the historic HVAC consumption values of the M1 office building in Donostia-San Sebastian (Spain).

Fig. 6. Boxplot of the mechanical ventilation consumption values of the M1 office building in Donostia-San Sebastian (Spain).

415

Fig. 7. Boxplot of the HVAC consumption values for the suggested ON/OFF Next 24 h-Energy schedule of the M1 office building in Donostia-San Sebastian (Spain).

temperature between the predefined comfort margins with an optimized and efficient operation schedule. Finally, it should be emphasized that an optimal performance of the HVAC system is directly correlated with an optimized building energy consumption by virtue of its capability of adjusting the parameters that control the ON/OFF HVAC schedule in a daily basis based on the indoor and outdoor conditions of the building. Moreover, it has been assessed that the full HVAC mode (see Section 2.3) is not necessary the unique mode for operating the AHU (Air Handling Unit), since mechanical ventilation mode (MV) can be used instead in order to supply outdoor air through the ventilation system while heating and cooling coils remain switched off to partially cool down the building. In energy terms, it can be also observed that the normal operation consumed on average 20–30 kWh while the mechanical ventilation reduces the consumption to approximately 7 kWh. This translates into an extra 5 kWh per hour of mechanical ventilation when the HVAC system in normal operation remained off but means an energy reduction of 13–23 kWh per hour of mechanical ventilation instead of full HVAC mode operation.

Fig. 8. Electrical consumption in Mikeletegi 1.

416


Fig. 9. Energy baseline for monthly thermal production, heating heat pump and boiler.

3.2. Energy savings results Fig. 8 depicts the monthly electrical consumption of the heat pumps for heating (dark grey) and cooling (light grey) the M1 office building during more than one year. It can be observed that the building demands simultaneous heating and cooling in both winter and summer periods. This electrical consumption information has been used to create energy baselines of the heating and cooling consumption. The methodology employed is based on the IPMVP protocol [27]. It was observed a linear relationship between the monthly energy consumption and the monthly outdoor conditions, measured as degree days. Degree days have been obtained from [28] for the closest weather station for the building location. Fig. 9 shows the monthly energy baseline for the thermal energy consumption (measured heat pump and boiler heating production) against the monthly heating degree days with base temperature 15 ◦ C (HDD15 ). kWht = 3161.3 + 75.54 · HDD15 month

(1)

Fig. 10. Energy baseline for monthly electrical consumption, cooling heat pump.

Eq. (1) means that the building thermal baseload for heating is around 3150 kWht per month and the external conditions increase the electrical consumption in 755 kWht/month and HDD. Fig. 10 shows the monthly energy baseline for electrical consumption (measured heat pump electrical consumption) against the monthly cooling degree days with baseline temperature of 15 ◦ C (CDD15 ). kWhe = 605.5 + 5.04 · CDD15 month

(2)

Eq. (2) represents that the building electrical baseload for cooling is around 600 kWh per month and the external conditions increase the electrical consumption in 5 kWh/month and CDD. Figs. 11 and 12 depict the comparison between the actual and estimated thermal and electrical consumption employed to cover the heating and cooling demand. Dark grey vertical bars depict the actual thermal and electrical consumption when implementing energy savings strategies, whereas light grey vertical bars represent the estimated consumption that would have been required to cover the building heating and cooling demand when operating the building as usual.

Fig. 11. Energy savings in heating consumption.


Fig. 12. Energy savings in cooling consumption.

As can be observed in Figs. 11 and 12 and according to the energy baseline built before the implementation of the energy saving strategies the energy consumption was reduced in both heating and cooling systems. More specifically, the energy reduction for the test period is estimated in 48% for the heating consumption and 39% for the cooling consumption. 4. Concluding remarks This paper presents a novel HVAC optimization framework, named Next 24 h-Energy, for efficiently controlling HVAC systems in office buildings by an energy minimization perspective. The proposed framework consists of a two-way communication system, an enhanced database management system and a set of machine learning algorithms based on random forest (RF) regression techniques capable of providing an optimal energy-efficient predictive control of the HVAC system by means of a dynamically adjusted HVAC ON/OFF and MV operation schedule based on the weather forecast and an accurate estimated indoor temperature of the building. Note that HVAC systems consumption represents an important amount of the total energy use in tertiary buildings, accounting for near 30%. The proposed Next 24 h-Energy framework has been successfully tested in a real office building (Mikeletegi 1 – M1) in Donostia-San Sebastian (Spain) obtaining significant energy reduction both in heating (48%) and cooling (39%) consumption. Future research will be devoted to the application of the proposed Next 24 h-Energy framework to different buildings typologies in order to assess its generalization capabilities. Acknowledgments This work has been partially supported by the ARTEMIS-2012-1 call, under project number 332987 which corresponds to Arrowhead project. References [1] J. Wall, J.K. Ward, S. West, M.A. Piette, Comfort, Cost and CO, 2008.

417

[2] M. Aune, T. Berker, R. Bye, The missing link which was already there: building operators and energy management in non-residential buildings, Facilities 27 (1–2) (2009) 44–55. [3] L.L. Lai, J.T. Ma, J.B. Lee, Application of genetic algorithms to multi-time interval scheduling for daily operation of a cogeneration system, in: Proceedings of the 4th International Conference on Advances in Power Systems Control, Operation and Management, APSCOM-97, Hong Kong, November, 1997, pp. 327–331. [4] R. Kramer, Energy management system and methods for the optimization of Distributed Generation, Patent US 6,757,591 B2 from Nisource Energy technologies, 29th June 2004. [5] A. Collazosa, F. Marechala, c. Gahlerb, Predictive optimal management method for the control of polygeneration systems, Comput. Chem. Eng. 33 (2009) 1584–1592. [6] B. Eisenhower, Z. O’Neill, S. Narayanan, V.A. Fonoberov, I. Mezic, A methodology for meta-model based optimization in building energy models, Energy Build. 47 (2012) 292–301. [7] F. Ascione, N. Bianco, C. De Stasio, G.M. Mauro, G.P. Vanoli, Simulation-based model predictive control by the multi-objective optimization of building energy performance and thermal comfort, Energy Build. 111 (2016) 131–144. [8] S. Ploix, Multi scale optimization of building energy, in: Energy Engineering, Springer, Singapore, 2017, pp. 105–111. [9] I. Mauser, J. Müller, F. Allerding, H. Schmeck, Adaptive building energy management with multiple commodities and flexible evolutionary optimization, Renew. Energy 87 (2016) 911–921. [10] K.A. Pruitt, S. Leyffer, R.J. Braun, A mixed-integer nonlinear program for the optimal design and dispatch of distributed generation systems, Optim. Eng. 15 (March (1)) (2014) 167–197. [11] T. Nagai, Dynamic optimization technique for control of HVAC system utilization building thermal storage, Energy Convers. Manag. (1999) 999–1014. [12] M. Gustafsson, G. Dermentzis, J.A. Myhren, C. Bales, F. Ochs, S. Holmberg, W. Feist, Energy performance comparison of three innovative HVAC systems for renovation through dynamic simulation, Energy Build. 82 (2014) 512–519. [13] F.C. Winkelmann, B.E. Birsdal, W.F. Buhl, K.L. Ellington, A.E. Erdem, J.J. Hirsch, S. Gates, DOE-2 supplement, version 2.1E, Technical Report LBL-34947, Lawrence Berkeley National Laboratory, Berkeley, CA, USA, 1993. [14] A. Hatef, A systematic approach of integrated building control for optimization of energy and cost, 2015 (PhD thesis at Concordia University). [15] S. Kajl, M.A. Roberge, L. Lamarche, P. Malinowski, Evaluation of building energy consumption based on fuzzy logic and neural networks applications, in: Proceedings of CLIMA 2000 Conference, Brussels, Belgium, 1997, p. 264. [17] D.B. Crawley, J.W. Hand, M. Kummert, B.T. Griffith, Contrasting the capabilities of building energy performance simulation programs, Build. Environ. 43 (4) (2008) 661–673. [18] M. Wetter, J. Wright, A comparison of deterministic and probabilistic optimization algorithms for non smooth simulation-based optimization, Build. Environ. 39 (8) (2004) 989–999. [19] X. Zhou, T. Hong, D. Yan, Comparison of HVAC system modelling in EnergyPlus, DeST, and DOE-2.1E, Build. Simul. 7 (1) (2014) 21–33. [20] U. Andreea, N. Ioan, C. Ilinca, T. Sorin, T. George, Adrian, Intelligent control ofHVAC systems. Part II: Perceptron performance analysis, Incas Bull. 5 (3) (2013) 127–135. [21] A. Dounis, P. Tiropanis, A. Argiriou, A. Diamantis, Intelligent control system for reconciliation of the energy savings with comfort in buildings using soft computing techniques, Energy Build. 43 (1) (2011) 66–74. [22] H. Alashaary, B. Moghtaderi, A. Page, H. Sugo, A neuro-fuzzy model for prediction of the indoor temperature in typical Australian residential buildings, Energy Build. 41 (7) (2009) 703–710. [23] G. Platt, J. Wall, J. Ward, S. West, Outside the Box: HVAC Management techniques that consider the World around them, in: Climate 2007 Conference, Helsinki, Finland, 2007. [24] S. Papantonioua, D. Kolokotsaa, K. Kalaitzakis, Building optimization and control algorithms implemented in existing BEMS using a web based energy management and control system, Energy Build. 98 (2015) 45–55. [25] Royal decree 1826/2009. Available from: https://www.boe.es/buscar/doc. php?id=BOE-A-2009-19915. [26] L. Breiman, Random forests, Mach. Learn. 45 (1) (2001) 5–32, http://dx.doi. org/10.1023/A:1010933404324. [27] IPMVP protocol. Available from: http://www.evo-world.org/en/productsservices-mainmenu-en/products-ipmvp-mainmenu-en. [28] Degree days. Available from: http://www.degreedays.net/.