AI Chiller: An Open IoT Cloud Based Machine Learning Framework for the Energy Saving of Building HVAC System via Big Data Analytics on the Fusion of BMS and Environmental Data
AAI C
HILLER : A N O PEN I O T C
LOUD B ASED M ACHINE L EARNING F RAMEWORK FOR THE E NERGY S AVING OF B UILDING
HVAC S
YSTEM VIA B IG D ATA A NALYTICS ON THE F USION OF
BMS
AND E NVIRONMENTAL D ATA
A P
REPRINT
Yong Yu ∗ Siemens Advanta SolutionsNew Territories, Hong Kong [email protected] † November 3, 2020 A BSTRACT
Energy saving and carbon emission reduction in buildings is one of the key measures in combatingclimate change. Heating, Ventilation, and Air Conditioning (HVAC) system account for the majorityof the energy consumption in the built environment, and among which, the chiller plant constitutesthe top portion. The optimization of chiller system power consumption had been extensively studiedin the mechanical engineering and building service domains. Many works employ physical modelsfrom the domain knowledge. With the advance of big data and AI, the adoption of machine learninginto the optimization problems becomes popular. Although many research works and projects turn tothis direction for energy saving, the application into the optimization problem remains a challengingtask. This work is targeted to outline a framework for such problems on how the energy saving shouldbe benchmarked, if holistic or individually modeling should be used, how the optimization is to beconducted, why data pattern augmentation at the initial deployment is a must, why the graduallyincreasing changes strategy must be used. Results of analysis on historical data and empiricalexperiment on live data are presented.
Keywords AI · HVAC System · Machine Learning · Energy Saving · Cloud Computing · Big Data · BMS
To combat climate change, the Paris Agreement aims to strengthen the global response by keeping global temperaturerise this century well below 2 degrees Celsius above pre-industrial levels [1]. The United Nations Environment Programestimates that 40% of global energy is used in buildings(Figure 1) [2]. The residential and commercial buildingsconsume approximately 60% of the world’s electricity [3–5]. Because of the high energy consumption, the buildingsector has also been shown to provide the greatest potential for delivering significant cuts in emissions and costglobally [2], and such potential has been projected to increase in the future [5,6]. For metropolis like Hong Kong, 66.5%of greenhouse gas emissions came from electricity in 2016, and Hong Kong committed to reduce its absolute carbonemissions by 26-36 per cent by 2030 from 2005 levels [7]. Centralized HVAC plants are widely used in commercialbuildings, and many are using water-cooled chiller systems during the last decades in substitution of the air-cooledchiller systems, which had increased the cooling efficiency significantly, and consumes up to 20% less electricity [8].When new installation to new buildings or major retrofit to existing buildings to the water-cooled chiller systems ∗ Dr. Yong Yu is a Data Scientist at Siemens Ltd. Hong Kong. This work was done during his employment with Siemens. † Ms. Jimalyn Yao is a Product Manager, Digital Solutions Manager at Siemens Ltd. Hong Kong, please contract her for anyenquiries regarding this paper via [email protected] . a r X i v : . [ c s . OH ] O c t I Chiller: An Open IoT Cloud Based Machine Learning Framework for the Energy Saving of Building HVAC Systemvia Big Data Analytics on the Fusion of BMS and Environmental Data
A P
REPRINT
Industry24% Buildings48%Transportation28% Industry77% Buildings15%Transportation8%Buildings(Domestic andTertiary)37%Transportation33%IndustrialProcess21%IndustrialBuildings7% Agriculture2%(a) (b) (c)
Figure 1: The energy consumption by sector for (a) the United States, (b) the European Union, and (c) China [3]exhibited remarkable savings during the last two decades, more potentials of savings are necessary to strive for morepower savings.When hardware installation and upgrade have been widely adopted, the industry turns to the optimization on theoperations of the HVAC systems in search of more power savings. In deed there had been many research works on theoptimization techniques of HVAC systems, especially chiller plants [9–11]. In the fields of mechanical engineering andbuilding service engineering, the chiller plant optimization methodologies usually rely on the physical models of theHVAC systems, and such models would have to be based on many assumptions on the system and equipment runningconditions, which may or may not hold true for the studied target systems. Especially, physical equipment and systemsoften vary due to the different installation conditions, extrinsic parameters, and the operation conditions often deviateacross the lifetime of the equipment.Modern buildings often come with Building Automation Systems (BAS), or Building Control Systems (BCS), typicallyconsist of building (energy) management systems (BMS/BEMS), which control HVAC primary components such as airhandling units (AHUs), chillers, and heating elements. As a major effort for the automation of building operations, suchBMSs collect abundant data from the components they monitor and control, but they stick to the purpose of automation,not much value-added analysis and application on the data collected. For the optimization of chiller systems to minimizethe building energy consumption, a new generation of BMS which is armed with data driven models to achieve theenergy saving by controlling and scheduling the components in the buildings is required.In addition to data from BMS systems, with the advances in the IoT sensor technologies, it becomes convenient andeconomical to install IoT sensors to collect live data and transmit back via wireless channel. This makes it possibleto collect whatever data required for the purpose of chiller plant optimization. On the other hand, the booming ofAI/Machine Learning (ML) power and technologies enables the capability of in-depth analytics on the data collected.Hence advanced and flexible optimization algorithms and systems becomes feasible by applying AI/ML models on thehistorical and live data from BMS and/or IoT sensors.The building HVAC system in itself is a complex system which consists of many interconnected electrical andmechanical machines, and the system is affected by many extrinsic and intrinsic factors that make the system extremelyhard to model. But the prevalent adoption of IoT and big data based infrastructure and the maturity of ML, especiallyDeep Learning (DL) algorithms makes the optimization of such system possible.There are two main directions on the power savings of the HVAC systems, one is to reduce the cooling productiondirectly, the other is to reduce the power consumption when at the same time maintain a given cooling productionprofile. In deed, in either way, one needs to set the appropriate cooling production targets or benchmarks, which couldbe a fixed value, a predefined target, or some other cooling profile calibred/forecasted by historical patterns, real-timeoccupancy, demand etc.. The former direction is to directly lower the production, but in many circumstances, one has toguarantee some service level or fulfil some given demand. Although this is an easy direction to go on the optimizationaspect, effort on identifying the actual demand to be fulfilled must be made. The latter direction is to setup a coolingtarget, or cooling target profile, and to optimize the HVAC system to achieve the optimal power consumptions, whileachieving the target or target profile at the same time. 2I Chiller: An Open IoT Cloud Based Machine Learning Framework for the Energy Saving of Building HVAC Systemvia Big Data Analytics on the Fusion of BMS and Environmental Data
A P
REPRINT
We go for the data-driven optimization approach, in stead of making many assumptions and rely highly on oversimplified theoretical models, we take advantage of the abundance of data and train models based on the big data, whichtake various system extrinsic and intrinsic factors into consideration, hence is highly adaptive to many circumstanceslike aging devices, deteriorating working conditions, etc.. On one hand, we recommend to use AI/ML algorithms,which in nature the appropriate companion to big data problems, in the modelling of big data, and we recommendto model the whole HVAC system as a whole under such conditions. On the other hands, we recognize that in somesituations, when data are limited and time is constrained, traditional algorithms/models combined with domain expertknowledge might be the right approach for the startup of the whole implementation of such optimization problems, andalso sometimes modelling individual components of the system might be the better choice.In this work, we first outline a framework towards the energy savings on the building HVAC system, we then elaborateon the directions and present an empirical experiment on the implementation of the later direction. Finally we draw theconclusion.
In this section, we review the previous works on the energy saving and optimization problems on the HVAC systems.HVAC systems had been extensively studied in the mechanical engineering and building service domains, and theoptimization of the systems to achieve energy consumption/cost savings had been one of the key concentrations. In [12],Weaver presented several approaches to the energy savings in the HVAC area with the use of closed-loop computercontrol and optimization. In [13], Doukas et al. studied a decision support model using rule sets based on a typicalbuilding energy management system, which incorporated the requirements in both guaranteeing of the desirable levelsof living quality in all building’s rooms and the necessity for energy savings. In [14], a mathematical programmingapproach for determining which available chiller plant equipment to use to meet a cooling load as well as the bestoperating temperatures for the water flows throughout the system was presented. There was work specifically for theoptimizing variable-speed pumps of indirect water-cooled chilling systems [15], they use online control strategy andfind the optimal control by adjusting the pressure set-point according to the estimated derivative of the total powerwith respect to pressure. Ben-Nakhi and Mahmoud [16] designed a General regression neural networks (GRNN)algorithm to predict cooling load for buildings, and further used the algorithm to optimize the HVAC system in officebuildings. Their results showed a properly designed NN is a powerful instrument for optimizing thermal energy storagein buildings based only on external temperature records. Wei et al. [10] deemed that the chiller plant model derived fromdata-driven approach is a nonlinear and non-convex optimization problem, and derived a two-level intelligent algorithmto solve the model aiming at minimizing the total cost of the chilled water plant, a simulation case was conducted andthe corresponding results were discussed.With the popularity of big data and deep learning (DL) during the 2010s [17–19], DL had been used in much broaderscenarios, and applying DL in the data centers to saving energy and money was one of the natural expansions. E.g.,DeepMind [20, 21] optimized the power for cooling service in Google’s data centers, based on the patterns of energyconsumption linked to running status of the machines in the data center. Siemens [22, 23] also offered an AI-poweredthermal optimization service for data centers. Inspired by these applications, more use cases on the power savings of thecommercial buildings were witnessed [24–27]. In [28] reinforcement learning, Li et al. proposed optimizing the datacenter cooling control via the emerging deep reinforcement learning (DRL) framework, which provided an end-to-endcooling control algorithm.
Many of the previous works focused on the optimization part of the HVAC power saving problems, but put not mucheffort on the measurement of the savings. Some of them use averaged previous power consumptions as the benchmark,others use averaged same period (e.g. same month) last year(s). Either way, these approaches introduced significantbias. Our analysis reveals that the extrinsic weather conditions plays the most important role affecting the buildingHVAC demands, and hence the power consumptions. Therefore using just past power consumptions as the benchmarkwithout considering the corresponding weather conditions is subject to inaccuracy. Accordingly, in our framework, wepropose to setup the appropriate benchmark via building a dedicated model, which forecast the cooling load demandand the corresponding power consumption, as shown in Figure 2.We then outline that the first choice for the optimization is to reduce the cooling demand directly, or change the coolingprofile such as pre-cooling. The other choice is to keep the original, or previous cooling load (corresponding to theweather conditions), which search the possible way of saving power via changing the operational patterns. Model(s) is3I Chiller: An Open IoT Cloud Based Machine Learning Framework for the Energy Saving of Building HVAC Systemvia Big Data Analytics on the Fusion of BMS and Environmental Data
A P
REPRINT
Optimization via changing operational patterns Optimization via direct demand cut For measuring saving amount
HVAC Energy Optimization Framework
Extrinsic Parameters Intrinsic + Extrinsic Parameters
Forecasting Model Cooling Load Approximation Model To be reduced Real-time demand detection/forecasting Power to be consumed Cooling Load to be Provided Power Consumption To be benchmarks
Figure 2: A Framework for the Energy Optimization of the HVAC systemsneeded to simulate the performance of the HVAC system, which then can generate searching space with both intrinsicparameters and extrinsic parameters.In the following sections, we introduce an empirical experiment on the optimization of energy user of a Chiller Plant ina office building.
It is usually very hard, if not impossible, to directly measure the savings on energy, water or demand [29], since savingsrepresent the absence of energy, water use or demand. Instead, savings often have to be determined by comparingmeasured use or demand before and after the implementation of a program, making suitable adjustments for changesin conditions. One can never obtain both actual measurement at the same. Figure 3 shows the energy-usage historyof a HVAC device before and after the deployment of an energy conservation measure (ECM). At about the time ofinstallation, the demand for the plant production also increased. The energy effect must be separated from that of theincreased demand in order to properly document the savings of the ECM. The “baseline energy” use the patterns beforethe ECM installation to determine the relationship between energy use and the demand on production. After the ECMinstallation, such baseline was used to estimate how much energy the plant would have used each month if there hadbeen no ECM (called the “adjusted-baseline energy”). The saving, or ‘avoided energy use’ is the difference betweenthe adjusted-baseline energy and the energy that was actually metered during the reporting period [29]. Without theadjustment for the change in production, the difference between baseline energy and reporting period energy wouldhave been much lower (or higher depending on the actual conditions). This will introduce significant bias on thecalculation of savings. It is necessary to clearly separate the energy effects of a savings program from the effects ofother simultaneous changes affecting the energy using systems. The comparison of before and after energy use ordemand should be made on a consistent basis. Hence, in [29], an "Adjustments" term is introduced to re-state the use ordemand of the baseline and reporting periods under a common set of conditions. This adjustments term distinguishesproper savings reports from a simple comparison of cost or usage before and after implementation of an ECM. Simplecomparisons of utility costs without such adjustments report only cost changes and fail to report the true performanceof a project. To properly report “savings,” adjustments must account for the differences in conditions between thebaseline and reporting periods. The baseline in an existing facility project is usually the performance of the facility orsystem prior to modification. This baseline physically exists and can be measured before changes are implemented.In new construction, the baseline is usually hypothetical and defined based on code, regulation, common practice ordocumented performance of similar facilities. In either case, the baseline model must be capable of accommodatingchanges in operating parameters and conditions so “adjustments” can be made.Historical data from the weather station and building BMS data are used in the analysis the relationship betweenextrinsic parameters and the power usage of the building chiller plant. Our results reveal that there is strong correlationbetween the weather temperature and the power usage, as shown in Figure 4. A deep learning model, specificallythe RNN-LSTM model, is employed in the forecasting of power usage and cooling load demand. We achieve MeanAbsolute Percentage Error (MAPE) 16.6% ±10% (95% confidence level, same for MAPE hereafter) in the forecastingfor power usage, and MAPE 14.0% ±5.8% for cooling load demand. The results illustrated that there are strongcorrelations between weather conditions and the building cooling load and power usage, and therefore one can forecastthe the load and power from weather conditions quite accurately. Though weather plays the prominent role in affectingthe cooling and power, the MAPEs all greater than 10% also exhibit that there must be other factors which are not as4I Chiller: An Open IoT Cloud Based Machine Learning Framework for the Energy Saving of Building HVAC Systemvia Big Data Analytics on the Fusion of BMS and Environmental Data
A P
REPRINT
Figure 3: Benchmarking of Savings [29]
Month -2.9 4.5% (a) Monthly averaged temperature vs.change of power usage (b) Plot of temperature vs. power Correlation Coefficient of power_total to other factorstemperature 0.81humidity 0.09wetbulb 0.79NonWorkingDay -0.36dayofyear 0.18building_load 0.95power_total 1 (c) Correlation between extrinsic parame-ters and power usage
Figure 4: Benchmarking of Savingsimportant, but still considerable. We regard the actual building occupancy as one of such extrinsic factors, and we leaveit for future study.
There two possible methods in modeling the Chiller Plant system. One is to model the whole system as a whole,the other is to model each of the major devices individually, and combine such sub-models together at the end ?? .The former one comes with simplicity that only the extrinsic parameters and outputs will be considered, and all theintermediate parameters are ignored. The latter, on the contrary, requires intermediate parameters that are related tothose individual devices, hence require more data. Though the latter one had the benefit of easy to start and easy toincorporate domain expert knowledge, the higher data requirement also makes it hard to generalize.5I Chiller: An Open IoT Cloud Based Machine Learning Framework for the Energy Saving of Building HVAC Systemvia Big Data Analytics on the Fusion of BMS and Environmental Data A P
REPRINT
Nov-2018 Nov-2019
Power Change % Saving attributed to our AI
Temperature 22.6 23.4Power (Actual) Power (Predicted by Monthly Linear) Power (Predicted by Deep Learning, 15Minutes Level)
Figure 5: Results on SavingsIn our test case, our BMS data lack of some important intermediate parameters like flow rate of water pipes, hence weadopt the holistic model.
The holistic model of the chiller plant with 5 chillers, 12 pumps and 4 cooling towers consists of almost 80 inputparameters and 2 outputs, such very high dimensional optimization problem is hard to be solved by conventionaloperations research algorithms. We employ the meta-heuristic algorithms in finding the optimal solutions. ParticleSwarm Optimization (PSO) is found to converge faster, but the results are unstable. The Genetic Algorithm (GA) isslower, but it manage to finish the optimization within the time frame of hours, and converge stably. Hence we adoptthe GA as our optimization algorithm.
Our whole solution is implemented based on 18 months historical data from Mar 2018 to Aug 2019 and deployed tothe studied building and with our recommendations into actual operations. Results are collected for two months (Oct2019-Nov 2019), and the final results are shown in Figure 5 and Figure 6.
During the early deployment of our solution, we observed forecasting performance drop of the chiller plant simulationmodel. Such phenomena was also witnessed in [27]. When the model is trained on the historical data, which in ourcase, are collected from the BMS that is running under the old operational practices and contain limited and only oldpatterns, it learns only old patterns. When it is deployed, and our recommendations are put in operations, new patternsemerge, the model performance then dropped. Therefore, right after the solution is deployed, a time period for the data(pattern) augmentation is required. The model must be retrained with data colleted from this period. Subsequently, theperformance is only acceptable after the retraining process is complete.
In our solution, we target at providing the "same" cooling profile. Such cooling profile comes from the forecastingmodel which takes the weather temperature and humidity as the input parameters. Our optimization algorithms searchwithin the feature space generated from a second model which simulates the chiller plant performance. Our final resultsof the last month show that 10.8% power savings was achieved. This is obtained from comparing the actual powerusage to the forecasted cooling profile in 15 minute intervals.
As a comparison, we also conducted two linear analyses. One of them used the daily averaged temperature and powerusage to forecast the daily power usage from daily average temperature. The other used the monthly averaged dailytemperature and daily power usage to do the same forecasting. 9% and 6% savings are observed, which cross-validatedpractical savings achieved. 6I Chiller: An Open IoT Cloud Based Machine Learning Framework for the Energy Saving of Building HVAC Systemvia Big Data Analytics on the Fusion of BMS and Environmental Data
A P
REPRINT (a) Daily averaged temperature vs. power usage (b) Monthly averaged temperature vs. power usage
Figure 6: Savings by Linear Models
In this work, we outlined a framework for the energy savings on the building HVAC system, and presented an empiricalexperiment on the chiller plant system, with around 10% savings achieved. As we have also discussed in Section 5.1, itis worth to try power savings on the air-side devices, i.e., the Air Handling Unit (AHU). Both methods can constitute afull picture for the power savings on HVAC system.Another direction is to make a bigger picture of the optimization, which takes the dynamic electricity price intoconsideration. More complex fitness function is required, and the final results will be more actionable recommendations.
Acknowledgements
I wish to acknowledge the help provided by Lei Lu and Felix So on the great R&D works.I would like to express my deep gratitude to Dr. Eric Chong and Mr. Keith Cheng for offering the opportunity and theirsupport on the works. I would also like to thank Mr. Yuelin Liang and Ms. Jimalyn Yao for their advice.I am particularly grateful for the assistance given by Mr. Ricky Liu, Mr. Tim Lo, Ms. Rita Leung and Ms. Hetty Leungon the domain knowledge.I would also like to extend my thanks to my Siemens colleagues, Ms. Christine Yip, Mr. Antrip Ho, and all whocontributed in this work.Assistance provided by the HKSTP colleagues Mr. Oscar Wong, Mr. Thomas Chan and Mr. Charles Leung under theumbrella of the Smart City Digital Hub was greatly appreciated.
References [1] Paris agreement.[2] The United Nations Environment Programme- Sustainable Buildings and Climate Initiative (UNEP-SBCI).Sustainable buildings and climate initiative – promoting policies and practices for the built environment.[3] Milos Manic, Dumidu Wijayasekara, Kasun Amarasinghe, and Juan J. Rodriguez-Andina. Building energymanagement systems: The age of intelligent and adaptive buildings.
IEEE Industrial Electronics Magazine ,10(1):25–39, mar 2016.[4] Jens Laustsen. Energy efficiency requirements in building codes, energy efficiency policies for new buildings. ieainformation paper, Mar 2008.[5] Thomas Weng and Yuvraj Agarwal. From buildings to smart buildings—sensing and actuation to improve energyefficiency.
IEEE Design & Test of Computers , 29(4):36–44, aug 2012.[6] Luis Pérez-Lombard, José Ortiz, and Christine Pout. A review on buildings energy consumption information.
Energy and Buildings , 40(3):394–398, jan 2008. 7I Chiller: An Open IoT Cloud Based Machine Learning Framework for the Energy Saving of Building HVAC Systemvia Big Data Analytics on the Fusion of BMS and Environmental Data
A P
REPRINT [7] Shirley Poon, Olivia To, Prashant Vaze, Patrick Yeung, Samantha Lee, Michael Lau, Xianji Wen, Cheryl Lo, andGavin Edwards. Transform hong kong into asia’s most sustainable city.[8] EMSD Hong Kong. Water-cooled air conditioning systems (wacs).[9] J.E. Braun and G.T. Diderrich. Near-optimal control of cooling towers for chilled water systems.[10] Xiupeng Wei, Guanglin Xu, and Andrew Kusiak. Modeling and optimization of a chiller plant.
Energy , 73:898–907, aug 2014.[11] T.T. Chow, G.Q. Zhang, Z. Lin, and C.L. Song. Global optimization of absorption chiller system by geneticalgorithm and neural network.
Energy and Buildings , 34(1):103–109, jan 2002.[12] Terry R. Weaver. Optimum control of heating, ventilating and air conditioning systems. In
UMR-MEC Conferenceon Energy. 96. [13] Haris Doukas, Konstantinos D. Patlitzianas, Konstantinos Iatropoulos, and John Psarras. Intelligent buildingenergy management system using rule sets.
Building and Environment , 42(10):3562–3569, oct 2007.[14] RICK T. OLSON and JUDITH S. LIEBMAN. Optimization of a chilled water plant using sequential quadraticprogramming.
Engineering Optimization , 15(3):171–191, may 1990.[15] Shengwei Wang and John Burnett. Online adaptive control for optimizing variable-speed pumps of indirectwater-cooled chilling systems.
Applied Thermal Engineering , 21(11):1083–1103, aug 2001.[16] Abdullatif E. Ben-Nakhi and Mohamed A. Mahmoud. Cooling load prediction for buildings using generalregression neural networks.
Energy Conversion and Management , 45(13-14):2127–2141, aug 2004.[17] Yoshua Bengio and Yann Lecun.
Scaling learning algorithms towards AI . MIT Press, 2007.[18] Kristian Kersting and Ulrich Meyer. From big data to big artificial intelligence?
KI - Künstliche Intelligenz ,32(1):3–8, Feb 2018.[19] SAS. Machine learning what it is and why it matters, August 2018.[20] Jim Gao. Machine learning applications for data center optimization, 2014.[21] Richard Evans and Jim Gao. Deepmind ai reduces google data centre cooling bill by 40%.[22] Peter Judge. Siemens brings ai-based thermal optimization to europe and asia demand-driven cooling integratingvigilent’s ai-based service now available outside the us. Web, May 2018.[23] Ulrich Kreutzer. The future of energy smart chillers, December 2018.[24] Zimu Zheng, Qiong Chen, Cheng Fan, Nan Guan, Arun Vishwanath, Dan Wang, and Fangming Liu. Data drivenchiller sequencing for reducing HVAC electricity consumption in commercial buildings. In
Proceedings of theNinth International Conference on Future Energy Systems , e-Energy ’18, pages 236–248, New York, NY, USA,2018. ACM.[25] Arun Vishwanath, Yu-Heng Hong, and Charles Blake. Experimental evaluation of a data driven cooling opti-mization framework for HVAC control in commercial buildings. In
Proceedings of the Tenth ACM InternationalConference on Future Energy Systems . ACM, jun 2019.[26] Arun Vishwanath, Vikas Chandan, and Kumar Saurav. An IoT-based data driven precooling solution for electricitycost savings in commercial buildings.
IEEE Internet of Things Journal , 6(5):7337–7347, oct 2019.[27] Hoang Dung Vu, Kok Soon Chai, Bryan Keating, Nurislam Tursynbek, Boyan Xu, Kaige Yang, Xiaoyan Yang,and Zhenjie Zhang. Data driven chiller plant energy optimization with domain knowledge. In
Proceedings of the2017 ACM on Conference on Information and Knowledge Management , CIKM ’17, pages 1309–1317, New York,NY, USA, 2017. ACM.[28] Yuanlong Li, Yonggang Wen, Dacheng Tao, and Kyle Guan. Transforming cooling optimization for green datacenter via deep reinforcement learning.