AI will be incredibly important in the transition to a zero-carbon economy. This is because of its ability to make accurate predictions about the behaviour of complex systems – in this case predictions of demand and supply of energy.

AI: understanding & harnessing the potential

As end-user energy consumption switches from primary sources such as oil and gas to electricity (electric cars, electric heating etc.), there will be a significant rise in demand on electricity generation. Also, as the proportion of electrical energy supplied by renewables grows, supply will become more sensitive to the amount of wind and sunshine on a given day. Shifting some of the peak load to times when renewables are producing electricity, or when demand is low, enables a greater fraction of renewables to be employed in the network without a high redundancy, and helps to avoid burning gas to offset demand surges. Using AI to predict both usage in individual households, and the expected supply and demand across whole power grids, can help reduce and even out the load.

Here we will discuss how predicting consumer behaviour can maintain comfort and convenience for individuals whilst lowering costs and increasing efficiency, and how this requires suppliers to be able to predict supply and demand to inform users about energy availability and price it accordingly. AI can help with these predictions at both ends of the chain of energy supply.

AI in the home

More than half of all energy consumption in the UK is from road vehicles, domestic heating and domestic appliances. Predicting usage behaviour for individual households and automating systems from this prediction can spread and decrease demand. For example, we developed a smart system to autonomously establish the thermal properties of a given building. The system learned the required heating on a per-room basis under a given set of temperature conditions that are necessary to keep the occupants comfortable. This allowed energy used to heat the building as a whole to be minimised without the loss of comfort.

Such systems do not, however, address the problem of extended demand peaks which renewable systems struggle to handle. Anticipating these events in advance enables demand-based pricing, helping to encourage moving energy use to lower-demand times where possible. Using AI, we can predict the energy that will be required for heating a given household, for example, by basing a prediction on the users’ past behaviour and weather forecasts. The energy predicted to be required in the near future can then be stored up at low-demand, low-cost times, for example using storage heaters, or perhaps household battery systems in the future.

Similarly, charging of electric vehicles can be carried out at the lowest-demand, cheapest times, and minimised by taking into account the predicted mileage on the following day. Washing machines, dishwashers and tumble dryers can be automatically run overnight, autonomous vacuum cleaners can be charged at off-peak times during the day. Fridges and freezers can be run colder at off-peak times to ‘store cold’ and allowed to become warmer during expensive times. These few systems together account for the majority of energy use by individuals, and all are compatible with being automatically run more cost effectively during low-demand periods.

AI in the grid

The industry is starting to think about time-dependent pricing based on supply and demand, in as short as 30-minute periods, to encourage shifting some energy use to lower-demand times. In order for this pricing system to work, however, prices need to be set in advance based on expected demand and supply, which requires accurate prediction of both on a network scale. This is necessary for predicting unavoidable demand surges to allow them to be met using short-term renewable methods such as hydroelectric power, rather than resorting to burning gas.

Predicting the output of a complex supply network consisting of a variety of renewable sources is difficult. This potentially includes partly-decentralised smaller suppliers such as household solar arrays, some of whom may be selling directly to local consumers at given periods. Predicting the demand of an equally complex system of millions of consumers and modes of energy use is equally difficult. Both supply and demand have multiple dependencies on weather, time of day, time of year and many other factors. AI’s ability to accurately model such complex systems will therefore make it an invaluable tool in the prediction of supply and demand of energy.

We're involved in a project to improve the prediction of gas demand in the UK, and AI has played an important role in this task. In this case, we have found that using boosted decision tree models have shown good predictive performance in these kinds of scenarios. The automatic learning of complex combinations of decision trees, using the most cutting-edge learning methods, have proven very effective at dealing with the many factors contributing to the demand at a given time. These factors can be  numerical, such as temperature, or non-numerical, such as the day of the week or whether it is raining or not.

Bringing it all together

Once the supply and demand of the network as a whole can be predicted accurately, a price over time can be set and relayed to the customers. Individual’s own systems can then optimise and spread their use based on this information about current supply, demand and cost, and the past behaviour of the individual. In this way, individuals will be able to guarantee the comfort and convenience they want at the best possible price, based on an efficient, load-balanced supply network which can rapidly move towards being entirely carbon free.

This week we're exhibiting at European Utility Week in Vienna. We will be demonstrating an application of AI in a supply network and a low-cost sensor IIoT system that can reduce generator maintenance costs. If you are there please drop by stand Bn38 for a chat with our team.

Author
Sam Connolly
Physicist & Algorithm Developer

Sam is a Astrophysicist by education working in the Algorithms & Analytics group on data analysis and algorithm development. His experience covers a range of fields including statistical analysis, algorithm design, signal processing, computer vision, robotics, machine learning & AI.