SDG 7: Affordable and Clean Energy
SDG 9: Industry, Innovation and Infrastructure
SDG 11: Sustainable Cities and Communities
SDG 13: Climate Action
Energy & Natural Resources
2. Project Details
Company or Institution
Open Climate Fix
Solar Photovoltaic Nowcasting
General description of the AI solution
Every moment of every day, the electricity supply must exactly match electricity demand. The electricity system operator is responsible for balancing the grid in real-time. This job is getting harder as more weather-dependent generation is added to the grid. Good forecasts are essential because large power stations take hours to turn on.
Solar power generation contributes increasingly large amounts of uncertainty to electricity forecasts. Existing solar forecasts are simply not good enough.
In the absence of accurate forecasts, grid system operators mitigate against the risk of a cloud wiping out solar generation by keeping large quantities of fossil-fuel-powered 'spinning reserve' online. This is carbon-intensive and expensive.
Deep learning on satellite imagery for this problem has not been explored widely but has shown great promise in similar domains and the technology is relatively quick and cheap to scale globally.
We are working on developing a Deep Learning model which takes recent satellite images and numerical weather predictions as inputs, and outputs probabilistic solar electricity forecasts for each solar system in the country. The forecasts will be calibrated in near-real-time using live solar data. Everything will be open-source. As far as we know, implementing this approach is entirely novel.
It's only in the last few years that such an approach has become technically feasible due to the large data volumes and compute requirements, and the ML is cutting-edge.
Existing solar electricity forecasts provide low spatial-resolution, hourly forecasts, updated only hourly, and with a significant time lag.
Our approach will provide high-accuracy, 5-minutely forecasts, updated every 5 minutes, for every solar system in the country of interest – a step-change improvement on existing methodologies. We'll deliver nowcasts via an API and an innovative UI to help users make effective decisions.
Open Climate Fix
Excellence and Scientific Quality: Please detail the improvements made by the nominee or the nominees’ team or yourself if your applying for the award, and why they have been a success.
Our innovation is to teach cloud dynamics to a deep learning model from satellite data and hence create a step-change improvement in solar forecasting. The ML model will be able to generate forecasts in minutes, surpassing existing approaches.
We use a relatively new type of deep learning architecture called the "Transformer" architecture, first developed by Google Research in 2017. Transformers are based on the beautifully simple idea of "self-attention" whereby the neural network learns to "attend" to different parts of the input on a case-by-case basis. This attention mechanism is surprisingly simple: At its core, for each input feature, the attention mechanism creates an 'importance score' for every other input feature.
Transformers have been at the heart of several recent breakthroughs in machine learning, including OpenAI's GPT-3 (which produces paragraphs of very believable text, given a short amount of input text); and Google DeepMind's AlphaFold 2 (which has learnt to predict the 3D structure of proteins).
Our work is all about adapting these powerful techniques to the world of solar electricity forecasting. This requires novel engineering (processing vast volumes of data) and novel research (developing new ML models for solar forecasting).
Absolutely all of our work is open-source. We do this to maximise our climate impact. We want other forecasting organisations to be able to integrate our technology into their products because that's the fastest way to reduce emissions globally. We openly collaborate and discuss with other researchers. We discuss our results and ideas before publication. We are entirely motivated by climate impact, not who publishes first.
The work is currently a research project, but we already have very promising results. It is at a Technical Readiness Level of around 5 We will build a prototype operational service early in 2022.
Scaling of impact to SDGs: Please detail how many citizens/communities and/or researchers/businesses this has had or can have a positive impact on, including particular groups where applicable and to what extent.
The primary route to impact for the AI innovation is through providing the electricity grid ecosystem, both public and private, with better solar electricity forecasts. This will enable more efficient operation of a grid with significant solar electricity penetration as well as allowing the level of solar power on the grid to be increased at lower risk.
More efficient grid scheduling and operation directly support SDG 13 on climate action through reducing carbon dioxide emissions, and SDG 11 on Sustainable Cities and Communities through reducing air pollution. Our work with National Grid (network operator in the UK) and Google.org Impact Challenge on Climate has allowed us to estimate the total potential global impact as at least 54 million tonnes of CO2 annually could be saved through adopting these techniques.
By making solar energy more predictable, it will lower obstacles for grid operators to have more solar on the grid, and further, it lowers the risk to commercial investors in financing utility-scale solar projects. This will lead to more affordable and Clean Energy (SDG 7) in the countries which use the service. This in turn allows the electricity grids to be more decentralize and manage in a highly renewable world (SDG 9).
Electricity markets vary wildly across the world and we want to open up multiple routes to impact, depending on the stakeholders:
– For key markets (initially the UK, and expanding to Europe and the rest of the world) we will run a service providing high-quality forecasts through an API, lowering the barrier to using the AI innovation, and
– All the IP behind the innovation is open source, and we are happy to work with forecasting providers, grid operators or solar farm operators anywhere around the globe to help them integrate our IP into their daily decision making.
Scaling of AI solution: Please detail what proof of concept or implementations can you show now in terms of its efficacy and how the solution can be scaled to provide a global impact ad how realistic that scaling is.
Proof of Concept
We obtained very promising results through research collaborations (including Stanford, University of Edinburgh, and UCL) and our own internal research. We conducted experiments to confirm the key stages of the intended technical architecture work.
Customers and Users
Over the course of 2020, we conducted a feasibility study on solar nowcasting funded by the European Space Agency. From interviews with over 20 electricity grid stakeholders, the study concluded that solar nowcasting using satellite imagery, powered by deep learning is technically feasible and that a range of users are hungry for better solar nowcasts.
Following our feasibility study, we won grant funding from Google.org as part of their Impact Challenge on Climate, to make PV Nowcasting a reality and help scale it across Europe. We are in contract negotiations with the UK National Grid Electricity System Operator to build a solar electricity nowcasting system for the UK. To scale globally we have talked to numerous international energy forecasting companies, who are all very interested in working with our IP, as well as our own service for core markets.
Also, over 1,000 people have signed up for our volunteer list.
Enabling a vibrant open-source ecosystem
Bridging the gap between research and industry is core to the mission of Open Climate Fix! We passionately want to engage thousands of researchers to help solve climate change and to change the culture of the energy forecasting community to be far more open and sharing. All our work is open-source. We plan to run ML competitions and to share not just our code but also our trained model parameters (using an innovative service called the Hugging Face Hub) to make it as easy as possible for researchers and practitioners to collaborate and build on our work.
Ethical aspect: Please detail the way the solution addresses any of the main ethical aspects, including trustworthiness, bias, gender issues, etc.
We believe that our application of the AI technology is ethical and equitable. While we do initially focus on delivering this work in Europe, we are already working together with philanthropic funding to further scale our work to areas where there are fewer investment sources in place locally.
All three components of the trustworthiness of our AI solution are in place.
Open Climate Fix is a nonprofit and we care deeply about equality and transparency. We are an equal opportunities employer and all our roles have a predefined salary which is published on the job posts. All our work is done in the open and all inventions are shared open-source under open licenses. Our driving force is not monetary gain but rather reducing carbon emissions.
We understand that training Machine Learning models produce carbon emissions though. To make sure that this does not go against our mission, we include all generated emissions in our carbon reduction calculations. We recently received validation from an external party that our generated emissions are orders of magnitude less than the emissions saved from the AI innovation.
We track all emissions that we generate.