WHEN IS A BUILDING SMART
This is a paper submitted and accepted to the CIBSE Technical Symposium 2019. A PDF can be downloaded here.
Abstract
We are facing a societal shift towards significantly increased use of automation across industries. Artificial Intelligence (AI) represents the next wave of this shift and this is being embraced at government and policy level (1). This opinion paper discusses the current status of artificial intelligence and raises some of the opportunities – and challenges – of using AI within building control systems.
To start with a tangible example, the world’s largest internet provider recently reduced the energy cost associated with cooling it’s data centres by 40%. It achieved this by deploying machine learning algorithms, developed by its sister company DeepMind that optimised the operation of system components (2). This is startling given these were already highly engineered facilities, where energy costs are a significant overhead. This is notable as an example of how digital technologies will provide opportunities for building services design to reduce energy use and maximise efficiency.
There is an oft used term that “people use energy, not buildings” alluding to the fact that behaviour within buildings are responsible for a significant amount of the energy use. This leads to a drive for the controls to be intuitive, and users to be educated in how to operate them. The counterpoint to this would be to allow AI to take control of building systems. Machine learning would learn occupant behaviours and then optimise the operation of heating and ventilation systems to maximise comfort and, in places of work, enhance productivity, whilst minimising energy consumption. Such handing over of control to digital systems has significant potential to save energy. It also brings risks.
This opinion paper has been prepared to postulate a future where we move away from procedural based control systems to those which use AI, specifically machine learning, to accommodate change. Machine learning is capable of continually optimising a system to maximise comfort whilst minimising energy consumption.
The paper is a practitioner’s view and is focused predominantly on the mechanical systems in commercial office spaces. However, the principals could equally be applied to other systems and building types.
Context
In June 2018, the UK Committee on Climate Change released an annual report on the progress in the reduction of carbon emissions in the UK (3). This report illustrates a dramatic improvement in the carbon emissions associated with the electrical network. However, progress on reducing the energy consumption from buildings is much less significant. This is illustrated in Figure 1 (left). Data from this report indicates that buildings account for 66% of electrical consumption in the UK, which is equivalent to emission of 48MtCO2 (Million tonnes of CO2). Direct emissions of CO2 is an additional 83MtCO2 which is largely associated with heating, of which dwellings account for 77%. If the decarbonisation of heat involves significant electrical based heating, then a similarly significant proportion of the 83MtCO2 will transfer to the electrical network which will need to increase in capacity.
It has been well documented that the measured energy consumption of buildings when in-use is generally much higher than that predicted by using thermal modelling software at design stage to demonstrate compliance with the building regulations Part L. This, so-called, performance gap means that the measured energy consumption in buildings is often in the region of 1.5-2.5 times greater than that which was predicted at design stage from modelling (4).
In order to address the performance gap in the early stages of a project, the Chartered Institution of Building Services Engineers (CIBSE) produced Technical Memorandum 54 (TM54), titled, Evaluating operational energy performance of buildings at the design stage (4).
Figure 1 (right) is taken from TM54 and illustrates the difference between the energy consumption compliance model for an office project and the actual energy in use. This shows that mechanical systems including cooling, fans and pumps contribute a significant part of this discrepancy and are estimated to contribute around 46% of the total building energy consumption.
Figure 1 – Left - Trends in UK carbon emissions, (Source - Committee on Climate Change, Reducing UK Emissions, 2018), Right – Case study illustrating energy by end use and the performance gap (Source – CIBSE, TM54)
Following the methodology set out by TM54 does not necessarily provide greater insight into the operation of the building or indeed lead to reduced energy consumption. It instead offers the design team and client a better estimate of what might be the case. This then raises questions around the best approach to addressing the performance gap. We should also not lose sight of the fact that the key issue is to ultimately realise a reduction in carbon emissions from building systems.
What makes a building smart?
The term ‘smart building’ is a popular term in the building industry. What qualifies a building as smart is not precisely defined and the term appears to be used interchangeably with ‘automated buildings’ and ‘intelligent buildings’.
A report published by the Royal Academy of Engineering (5) indicated that a key element of a smart building is that networked systems such as security, building management systems, lighting control and any associations are enabled as elements in the Internet of Things.
The term Internet of Things (IoT) is now widely applied and refers to internet enabled devices which can provide a network of sensors all reporting information to a web-based platform (6). There is significant interest in the application of IoT within the built environment. One possible avenue for this technology is how it can allow building environments to be more responsive to occupants.
In 2018 the British Council for Offices (BCO) released a report on digital technology and how It could improve the responsiveness of buildings (7). This highlighted how the various data-based systems in a building generally operate within silos. In recent times we’ve seen the emergence of ‘converged networks’ where the buildings data network is used for communication within the systems shown in Figure 2. However, this often only means convergence at a hardware level in the sense of shared structured cabling and there is still no connectivity at the software level. This siloed effect is illustrated in Figure 2.
Figure 2 – Systems in a building have historically been siloed
In order to realise the benefits of the IoT it’s important that these devices can communicate on a common integrated platform as illustrated in Figure 3. This shows how systems will transfer data to the integrated platform which allows it to be analysed and then decisions relayed. This level of integration will ultimately mean that sensors can be common across all systems rather than distinct. For example, a camera could be used to monitor occupancy, measure daylight levels to control lighting and highlight security issues – and feed back to a cloud-based platform.
Figure 3 – Integrated approach to building systems
Implementing IoT enabled systems within buildings will allow data to be collected to show how a building truly operates. This data can then be used to inform decisions around building operations and target savings in energy. However, if ‘smart’ refers to ‘intelligent’ then a building needs not only to acquire knowledge, but also to apply that knowledge. Currently, systems controlling building operations have no ability to learn.
Most construction projects are unique and in a commercial building the mechanical systems which keep a building comfortable with good internal air quality are complex. The control system which integrates the components of the mechanical systems is in turn unique to each project- albeit based on some standard algorithms.
A building control system is procedural which means it can only react to circumstances in a pre-defined way. This means designers and control specialists need to predict every possible scenario the building might experience and design an appropriate response. Given the uniqueness of a project and the complexity of a system the engineering time required for this is considerable. In order to make the process more manageable we tend to reduce the possible scenarios by fixing certain operational parameters.
As an example of this final point, it is common that a chilled water cooling system will have fixed flow temperature and return temperatures of 6° and 12°C respectively, the design for the peak cooling scenario of the building is then formed around these parameters. However, at part load or when the external conditions are below the peak design criteria, it might be more efficient to operate the plant differently.
In the case of a building using fan coil units, at part load the chilled water temperature might be raised, this improves chiller efficiency, Coefficient of Performance (COP), but the air supply temperature which can be achieved is lower so a higher volume flow rate in the fan coil unit needs to be maintained which increases fan energy. Alternatively, the chilled water temperature could stay at 6° and 12°C, and the fan coil unit flow rate be reduced. When other components of the system are introduced the possible options go on; increase the temperature difference in the chilled water to reduce pumping power, however apparatus dew point is increased affecting the fan coil unit. Selecting the most efficient option at any given point in time from a myriad of different operational scenarios across all components is impossible for humans to calculate, let alone programme.
These constraints are therefore generally imposed on the building at design stage and can remain fixed throughout the year, and for many years. This may lead to excessive energy consumption for several reasons:
Operating parameters are not optimal when part-load or varying external conditions are considered
Occupancy/working patterns are different to that expected at design stage and the building is not able to adjust operating parameters accordingly
The parameters assumed at design stage may not be optimal given the way the building is operated in reality
Buildings that learn
Machine learning is a form of computer science where algorithms are designed to learn and improve from experience rather than being explicitly programmed.
In order to learn, large amounts of data is required in order to train the system. It is no coincidence that the surge in interest in AI in recent years follows from huge datasets being generated through the world wide web and smartphones (8).
Through collecting data about how the system operates a machine learning algorithm could establish the optimal route to maintaining a particular internal condition for the least amount of energy.
Figure 4 – Effect of Machine Learning Algorithm on Power Usage Effectiveness in a Google Datacentre (source: Deep Mind)
This type of AI controlled adjustment to the operating systems in a building has been successfully trialled by Google at its data centres. There, the automated adjustments of energy use has saved 40% of the energy consumption for cooling and resulted in a 15% improvement in the Power Usage Effectiveness (PUE) (2). The effect of the machine learning algorithm on PUE is demonstrated in Figure 4. Similar outcomes were also observed in research carried out on data centres in Singapore (9).
This illustrates the potential for machine learning to optimise a complex system such as a buildings mechanical ventilation and air conditioning system.
In the case of data centres, the target for the machine learning system is relatively simple – requiring a specific temperature is maintained in each server rack, whilst consuming the minimal amount of energy. Human occupants are much more complex but in the first instance the operating target of a machine learning algorithm can be to achieve a particular internal condition in specific rooms, in a similar way to the data racks. In time AI systems offer the scope for a buildings control system to respond to much more than the space temperature and humidity.
Implementation
Construction is not known for adopting change rapidly and is one of the least digitised industries (10). As such we need to consider the steps required to best facilitate the deployment of AI within building control systems, including the impact it might have on the way we design systems now.
Data is constantly generated in large quantities by building control and energy monitoring systems. For example, in a work place where staff use security swipe cards, data may be being gathered every day on how many people are in the building, how many people are making use of a meeting room relative to an open plan office space and – correspondingly – the use of energy in those spaces, temperate and air flow to those rooms and so on. Much of this data is periodically overwritten, and never analysed by engineers.
Several companies in the industry do offer systems that continually harvest information from every available point in a building and energy management system (11). Data science is then used to generate visualisations of the building system operation and illustrate when an element is not behaving as intended. For example, ventilation air handling units (AHU) running outside of operating hours when no one is inside the building. If a control system was to be based on learning algorithms, then the intelligence of the system could be utilised to make modifications to a system. In such a situation the problem would need to be corrected by revisiting the underlying programming or set points in the building management system. Currently, this generally requires manual adjustment by a specialist.
Key barriers to the implementation of machine learning based control systems can be summarised as:
1. Limited interoperability between various systems in a building limit the potential for IoT based sensors to interface with building control systems
2. Building control software tends to be closed, this means it is possible to modify pre-defined set points and read points, but it is not possible to change the state of any variable within the system by an external software package. Direct access between a machine learning based control system hosted on a web-based service and the signals which control actuators, fans, pumps, etc. is required
This paper proposes that the industry needs to be proactive in unifying around a common open source software framework to allow the interoperability between various systems. This would allow manufacturers to develop devices which use this framework and ensure compatibility. There are challenges with open source software in terms of where responsibility sits and how security is continually addressed. This means that cooperation between manufacturers is required with a consortium established to oversee the framework.
There are ongoing attempts to develop a suitable framework but a preferred option is not clear currently. There are rightly concerns raised around security and privacy of data with this approach. These issues are beyond the scope of this paper but it is possible to anonymise this data at the point of collection in line with the European General Data Protection Regulation (GPDR) (7) such that systems can harvest general data which is not relatable to specific users.
Continuous improvement
The Soft Landings framework (12) has been established in order to improve the outcomes of building projects, both in terms of user satisfaction and in resource consumption. A significant aspect of this is to continually monitor the building in use and ensure that systems have been commissioned to operate correctly.
The ultimate goal of a smart building should be to optimise this process of commissioning. This will require that there is increased granularity in the level of sensing and also automation within a building.
Several companies have developed IoT based technology that monitor spaces more accurately. Approaches vary and include using cameras to monitor spaces, or asking users to provide feedback on their comfort using their smart phones (7). Using the IoT therefore allows information about spaces to be shared between systems which can contribute to the data processed by learning algorithms.
Increasing the granularity in mechanical systems would mean increasing the amount of automation such that control dampers and valves which may once have been manually adjusted can now be operated by the building control system. This would allow the initial set up of a building to be automated and then also continually refined through learning algorithms in reaction to increased sensor data.
Conclusions
IoT enabled systems will increase the ability to sense occupancy patterns, air quality and thermal comfort at a much higher level of granularity.
Machine Learning has the ability to optimise complex building systems beyond that which is possible by human designers.
Connectivity between the various systems in a building is important to allow data to be harvested and used for algorithmic learning, and improving.
An industry-driven open framework to allow communication between all equipment and systems is required to enable control points to become available to a machine learning based, web based hub.
Increasing the level of automation in mechanical systems and allowing a direct interface between the learning algorithm and control points will allow optimisation of systems at a resolution which matches an enhanced sensing and perception capability.
A machine learning algorithm can then take on the role of aftercare through optimising the building systems to suit the use of spaces and accommodate change.
We need to focus on reducing overall energy consumption in buildings. Embracing AI in building services will enable us to do this, to create spaces that are better tuned to people - thus creating spaces that are efficient, well used and supportive of user’s needs.
References
(1) Hall W and Pesenti J, Growing the Artificial Intelligence Industry in the UK, Department for Business, Energy & Industrial Strategy, 2017.
(2) DeepMind AI Reduces Google Data Centre Cooling Bill by 40%, 12th February 2019 https://deepmind.com/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-40/
(3) Reducing UK emissions 2018 Progress Report to Parliament, 2018, Committee on Climate Change
(4) Technical Memorandum 54 - Evaluating operational energy performance of buildings at the design stage, CIBSE, 2013
(5) SMART BUILDINGS people and performance, Royal Academy of Engineering, 2013
(6) Shen G and Huang X, Advanced Research on Electronic Commerce, Web Application and Communication, Massachusetts Institute of Technology (MIT), 2011, page 392
(7) Fast and Slow Buildings: Responsiveness through Technology and Design, British Council of Offices, 2019
(8) Digitalization and Energy, International Energy Agency (IEA), 2017, page 22
(9) Li, Yuanlong & Wen, Yonggang & Guan, Kyle & Tao, Dacheng, Transforming Cooling Optimization for Green Data Center via Deep Reinforcement Learning, 2017
(10) Imagining construction’s digital future, 12th February 2019, https://www.mckinsey.com/industries/capital-projects-and-infrastructure/our-insights/imagining-constructions-digital-future
(11) Demand Logic / Finding the needle in a haystack of energy data, 12th February 2019, https://www.ashden.org/winners/demand-logic#continue
(12) Agha Hossein M, BG54/2018 - Soft Landings Framework 2018, BSRIA