;
;
For many kinds of businesses, California’s regulatory target of net zero energy (ZNE) from commercial construction by 2030 presents real operational challenges. This is particularly true in the energy intensive food retail sector where lighting, refrigeration, air conditioning and other processes consume large amounts of power. With net zero on government agendas worldwide, California’s forward-looking regulations are likely to be a template mimicked elsewhere, and businesses are understandably eager to find adaptations that are both sustainable and profitable.
ZNE requires that any energy consumed must be offset by energy generated by on-site renewables – but plainly this will be difficult for many food/retail outlets, which are energy intensive. The smart approach to ZNE is to reduce energy use as much as possible, and only then meet the remainder with renewables. This is particularly attractive when on-site renewable energy production is constrained, e.g. due to limited roof space. With this in mind, we worked with Whole Foods supermarket to take a different approach, bringing the full power of machine learning to the challenge of energy reduction in one of its stores.
Project Summary
46% energy savings
2,448energy saving options modelled in a week
Finding the best solution in a 27 digit number
Our team combined knowledge of the mechanical engineering of modern buildings and their energy requirements, with the ability to develop creatively powerful machine learning tools. Faced with the need to radically reduce the energy use of a Whole Foods store in San Francisco, we decided to develop a powerful genetic algorithm which discovered the optimum mix of existing, off-the-shelf technologies that could help them meet the ZNE target.
Our process began by working with industry specialists to list out the most effective improvements that could be made to the store’s existing heating, refrigeration, cooking, lighting and air-conditioning systems. Engineering is always about weighing options and working within constraints, so we knew we couldn’t do everything and stay within budget.
Also, some energy improvements enhance each other, while others cancel each other out – it’s not as simple as picking the most effective individual improvements and designing them all together. Working out which exact combination of those 107 energy saving strategies would be optimal (and affordable) would be impossible with the human mind alone – the total number of potential outcomes was a 27 digit number. But we had a hunch that a well-designed machine learning algorithm could overcome the scale of the problem and identify the best combination.
Learning from nature, to achieve energy reduction
Genetic algorithms mimic the selection process we observe in nature, combining outcomes that represent ‘success’ and rejecting those the ‘weaker’ solutions. This has been attempted in research projects but this is the first time it has been used to optimise a supermarket.
Our algorithm duly got to work, finding energy-efficient combinations of the many ideas, building on those, adding and removing elements until it had generated nearly 2,500 costed plans (something that would have taken over 10+ years if done manually). Like genetics in nature, we wrote the algorithm to introduce random mutations, to see if the process could learn from those that helped or hindered, and repeating for 50 iterations until only the most effective combinations of strategies were left. From those solutions, it was a much more human task for our experienced energy engineering team to pick the most workable combination for the client.
The result was an incredible 46% energy reduction for the store, using existing technology. This is a level of calibration and planning that simply wouldn’t be possible any other way. In further evidence that climate adaptation is more than commercial viable, we were able to work closely with the designers and contractors to develop an upgrade plan that could be carried out while the store stayed open and trading.
Making the sustainable, affordable
This project, a collaboration with the California Energy Commission, has been a milestone for us. It proves how machine learning and cloud-computing can be brought to bear on vital public goals, like energy efficiency and the race to net zero, at an affordable cost, in a scalable form. Business is full of patterns, once you know how to find them – and this kind of use of business data has the potential to transform the way we plan, design and operate the built environment, and achieve a sustainably developed world.
Arup was able to identify 15% greater energy savings than would otherwise have been achievable. Following this success, Arup has already used this method to work on larger projects, including identifying over 40% energy savings for the Natural Resources Defense Council’s seven-floor New York HQ, by modelling 38 different energy conservation measures and working with the city of Boston to identify pathways for it becoming carbon neutral by 2050.
