The vast majority of products and solutions in the current BeMS marketplace, provide hardware and software that are programmable with conditional logic. This allows control assertions to be created such as :
If temperature_sensor < 18 then turn_on the heating
If expected_weather = rain then close the windows
(very basic pseudo logic control)
This approach to control systems is the norm; as with any programmable based system, the success and suitability of the solution comes down to the design, implementation and testing of the logic and control integration. The capability and experience of your technology and engineering team, ultimately defines the success and quality of the delivered solution.
In the near future, AI will be thrust upon the BeMS industry, with product and solution companies proclaiming all sorts of art-of-the-possible. To be clear, in this context when referring to AI, we are discussing principally machine learning technologies.
The assertion, is that AI will allow for more dynamic control, with the building system using vast data sets in order to predict the best course of action in order to achieve the desired result. This requires a shift in mind-set for the technology and engineering teams, focussing on the objective result required and the constraints that must be observed in order to achieve the desired objective, rather than how to achieve it. The solution will then identify patterns within the data from the live environment, to see if based on the training data and model, what the best action could be in order to achieve the desired objective.
This will undoubtedly in time be the basis for many building control platforms, but not for a while; as we have several practicalities that need to be addressed in the short term.
Firstly, in order to train your model in order to identify patterns and derive decision paths, you need real-world data, both positive and negative. The quality and scope of the data used to feed your model will ultimately define how successful it starts out; once unleashed into the real-world. It will ultimately take time and effort in order to construct sufficient relevant data sets allowing for seasonality, geography, usage type and the various physical components that make up the buildings infrastructure.
Secondly, we also have to consider the privacy aspects of some of the data being collected, as well as the potential security impacts. We will want to include a wide range of data from various aspects of a buildings infrastructure and operating environment, including mobile and data usage.
Thirdly, in a real-world example, we need to place constraints on any learning model in order to ensure health and safety as a minimum are covered, let alone any wider policy restrictions that may apply; by either the company owning the property or local, central law makers. It’s one thing to allow a model to play out all sorts of permutations in order to become the worlds best player of ‘Go’. The same approach cannot be used in the physical world when as a result of a decision people and assets could be harmed.
Then, once you finally have all your data sources, cleaned with sufficient Longitudinal coverage; with data providing positive and negative identification; with all the signed off constraints that need to be enforced in order to comply with your overall environmental and corporate policy, then you can start training your models and deploying your AI.
It is not a trivial set of prerequisites in order to get started; ultimately it will be worth it, for those that take a strategic approach; but first, get your data sources ready !