DATA CENTRES | ARTIFICIAL INTELLIGENCE Automatic for the people Last year, Googles DeepMind company used machine learning to reduce energy associated with cooling at its data centres by 40%. UWEs Dan Cash considers the potential of artificial intelligence in the industry A DAN CASH is senior lecturer in building services at the University of the West of England (UWE) s you stand at baggage reclaim, you open an app on your phone. This tells you your homes heating is on and will hit your preferred temperature just as you turn the key in the lock. Gone are the days of returning to a cold house; the new generation of home thermostats record your daily routine and preferred temperatures, and learn about your house and its heating system. They predict when to turn on the heating based on your preferences and the forecast external conditions. This intelligent behaviour is calculated using a machine-learning algorithm. Machine learning is a subset of artificial intelligence (AI), with algorithms designed to learn and improve from experience rather than being programmed explicitly, as in traditional systems. Large amounts of data are required to train the algorithm. It is no coincidence that the current surge in interest in AI follows the huge datasets that have been generated through the internet and smartphones. Learning thermostats are effectively Internet of Things (IoT)-enabled devices, which use the internet to upload data from sensors to web-based servers. The collected data is then used to refine operational behaviour. What could be the potential of this idea if scaled up to larger buildings and systems? A well-publicised example is Googles use of machine-learning algorithms through its DeepMind subsidiary. These helped reduce energy associated with its data centre cooling by 40%, leading to an overall energy saving of 15%. This was startling given that these facilities were already highly engineered.1 Key reasons these data centres were good candidates for AI technologies include: The existence of an abundance of training data from historical monitoring from a multitude of sensors The systems controlling the internal climate are complex and there are thousands of ways that individual items of equipment can interact with each other For every permutation of internal server load and external weather data there is an optimum operating scenario of the various components Optimised scenarios cannot be arrived at by human intuition alone. PAPER ACCEPTED Technicalm Symposiu osium .org/symp www.cibse Machine learning as a tool looks for patterns in training data to develop an algorithm that produces a reliable output from given inputs. Once the learning is complete, the algorithm can be used as a prediction tool for optimising outcomes. Initially, when used to optimise the performance of data centres, algorithms made recommendations for human operators to implement. This was to ensure the system did not try to move to a state that would have put the data centre at risk. At the end of last year, DeepMind announced the algorithms had been allowed to make changes to the control system directly, within predefined safe ranges for the equipment. This direct control of the data centre by a machine-learningbased AI agent is exciting for the future of buildings and the built environment as we seek to decarbonise. Non-domestic buildings are generally unique and involve a complex arrangement of services, ranging from chilled water plant through to automated openable windows. Control systems for buildings are, in turn, unique and programmed by hand to predicted operating scenarios developed by the designer, which is inherently limiting as with the operation of data centres. The key to advancing building control systems towards an AI-enabled future is the adoption of communication platforms that can support its deployment. Currently, many companies offer data analytics tools that continuously read the status of points in a building management system (BMS). These can offer insights and allow building managers to make informed decisions. However, the modification of the architecture in a building control system needs human intervention. Moving to a point where its possible to write to the BMS in a meaningful way may take time. In the meantime, when designing buildings, consider the platform for communication-based systems, their interoperability, the ability to collect data, and the security of that data. Data from an access control system could, in future, be as useful to defining the operation of building systems as a temperature sensor. In operation, look after your data, keep BIM models up to date, store data in accordance with approved standards, and consider using a digital twin. References: 1 Safety-first AI for autonomous data centre cooling and industrial control, DeepMind blog http://bit.ly/CJApr19AI il 25-26 Apr 9 1 0 2 54 April 2019 www.cibsejournal.com CIBSE Apr19 pp54-55 AI columns.indd 54 22/03/2019 17:01