Jack Kelly

Using computers to fix climate change. Currently focused on forecasting electricity demand.

Greater London, England, United Kingdom

About

I'm terrified by climate change and determined to do everything in my power to reduce emissions as quickly as possible. In early-2019, I co-founded Open Climate Fix, a non-profit entirely focused on using open-science to urgently reduce emissions. Prior to Open Climate Fix, I was a research engineer at DeepMind where I applied machine learning to predict wind power. Prior to DeepMind, I did an MSc, PhD and postdoc at Imperial College London on electricity disaggregation. During my academic work, I collected and released a large dataset of domestic appliance-level electricity consumption, and applied deep learning to disaggregation, and was a lead developer on an open-source framework for electricity disaggregation.

Experience

  • Co-Founder at Open Climate Fix
    Jan 2019 - Present · 7 yrs 6 mos

    Non-profit focused entirely on helping the energy community to reduce greenhouse gas emissions at scale. Very open and collaborative. Everything will be open-source. Avoid all business norms which get in the way of reducing greenhouse gas emissions ASAP.

  • Concept Team Member at International Centre for AI, Energy, & Climate
    Feb 2019 - Present · 7 yrs 5 mos

  • Research Engineer at DeepMind
    Jun 2017 - Dec 2018 · 1 yr 7 mos

    I was a research engineer on the team who used machine learning to predict wind power

  • Computer Science Post-Doc research associate at Imperial College London
    Sep 2016 - Jan 2017 · 5 mos

    Has disaggregation accuracy improved since the 1980s? Which algorithms are most accurate for a given use-case? Which (if any) use-cases are well served by disaggregation already? It's pretty much impossible to answer any of these questions with confidence. We can't directly compare published results across papers and companies because people use different datasets, different metrics, different pre-processing, etc. This means that we can't measure progress over time. Nor can we decide which NILM algorithms are most promising and which might be dead-ends. These are bad problems to have. This postdoc aims to outline some potential solutions to these problems.

  • PhD placement at IBM
    Jul 2014 - Oct 2014 · 4 mos

    Commercial buildings have management systems consisting of many thousands or even tens of thousands of sensors. Detecting faults in this mass of data is not trivial. Existing approaches use rule-based systems where a large set of `if, then' rules are hand-coded. These rule-based systems are fragile, difficult to set up, inaccurate and require user input to modify the rules whenever a change is made to the building. We set out to build an automated anomaly detection system which can passively learn the normal behaviour of a building and then detect when the building deviates from that norm. I used a type of recurrent neural network called a `long short-term memory' (LSTM) network to learn the dominant patterns in the data and then to detect anomalies and this worked well. I also wrote code to pull data from a BACnet building management system (this code was a mixture of C, Cython and Python).