If we are to successfully mitigate climate change we must reduce global emissions to near-zero by 2050. To this end, three solutions exist:

  1. Generate energy using emission-free technologies;
  2. Capture and bury emissions using carbon capture and storage (CCS); and
  3. Use energy more efficiently.

In the UK, emission reduction efforts via solution 1) have so far proven extremely successful. Emission-free energy technologies generated 45% of UK electricity demand in 2020 which has caused a decrease in grid carbon intensity of 66%. Headway is being made, albeit slowly, toward solution 2), DRAX hope to deploy two Bioenergy CCS plants by 2030 that will remove 0.32 tCO2 per person per year. To date, mitigation efforts via solution 3) have proven less fruitful. UK final energy consumption has not meaningfully decreased in recent years; only 7% on 1990 levels according to the IEA. Despite widespread acknowledgment of the need to use energy more efficiently, over 30 years we’ve made little progress.

Our inefficient use of energy is linked closely with uncertainty. We overuse our resources today because tomorrow’s needs are difficult to predict. Building contractors, for example, pre-purchase more material than is required because they are unsure exactly how much they will use in construction. Similarly, supermarkets buy more food than customers want because they struggle to predict customer demand. And home owners use their central heating for longer than needed because they cannot be sure how long it takes to warm up. Better predictions of our uncertain future allow us to make informed, efficient choices in the present that mitigate wasted resource.

To make such predictions we use probabilistic machine learning. And with our predictions, we can then think about ways of acting optimally in these futures using Reinforcement Learning. For example, using a probabilistic forecast of this afternoon’s cloud cover, we could predict when the solar panels on our roof are likely to produce energy and delay the use of air conditioners until then. Using predictions for control in such a way, we can reduce emissions at near-zero cost, in a way that other mitigation strategies struggle to do.

My work is about making the best use of our predictive abilities to use resources more efficiently and reduce emissions.