More examples

Forecasting to build a sustainable and safe society

The UK is the world’s first major economy to set a legally binding target of being net zero by 2050. With renewable energy already part of our overall energy supply, achieving a balance between the greenhouse gases emitted and removed from the atmosphere will require a transition to a zero-carbon electricity system. According to the UK National Grid, the biggest proportion of electricity entering the national grid is generated by natural gas, which is largely from imported fossil fuel that emits harmful greenhouse gases into the atmosphere when burnt to produce energy. To meet the UK’s demand while reducing carbon emissions, considerable efforts are being made by the local and national authorities to increase the amount of energy from renewable and low-carbon technologies. In 1991, renewable energy accounted for just 2% of the UK’s electricity generation, which in 2020 increased to 43%. For the first time in UK’s history, in 2021 electricity came mainly from renewable energy: a mix of wind, solar, bioenergy and hydroelectric sources, overtaking the use of fossil fuels consumption. The UK’s endeavour to establish a carbon-neutral infrastructure for energy supply has been greatly assisted by data science and AI technologies.

In 2017, ASG researchers collaborated with the UK National Grid Electricity System Operator (ESO) to improve their work in ‘balancing’ the electricity demand with electricity being produced while accounting for intermittent generation such as wind and solar power. While there is data on the location of UK government-supported solar farms, there is less information on smaller domestic installations of photovoltaic (PV) solar panels. In the absence of accurate forecasting of solar energy input from solar PV installed capacity, electricity produced by natural gases can unnecessarily enter the energy circulation (due fossil fuelled generators running in the background) adding to the carbon emission. Attempting to improve the forecast accuracy for ESO and run the National Grid more economically, researchers and the Turing’s doctoral students at a Data Study Group in 2017 contributed to the adoption of PV solar panel forecasting models for the National Grid. The initial model was further validated, tested and implemented by ESO, alongside a range of other new forecasting approaches, achieving 33% more accurate day-ahead forecasts. These more accurate forecasts will help balance energy supply and demand, lowering costs for consumers and ultimately meeting the ‘net zero’ goals globally ([Stowell et al., 2020]](https://www.nature.com/articles/s41597-020-00739-0)). Using data from citizen science and crowdsourcing efforts, researchers have created a fully open geographic data source for solar PV within the UK, suitable for reuse in different geographical data and other applications such as machine vision and short-term solar forecasting.

Another example comes from a project dedicated to quantifying the effects of climate change on extreme weather events using the ‘distributional downscaling’ approach, an approach to infer high-resolution information from low-resolution data. Extreme precipitation events and floods caused by climate change expose over 1.8 billion people to significant flood risk worldwide, causing nearly $520 billion in economic losses including well-being costs per year (Rentschler et al. 2022). These events disproportionately affect low-income countries lacking infrastructure investments. Climate models are used for forecasting and predicting the risk of floods and heavy precipitation events, acting as an important source of information for policy-makers. To improve the prediction of high-resolution precipitation, researchers in this project apply a deep learning approach using model fields (weather variables) that are more predictable and generalisable than local precipitation (Adewoyin et al. 2021). To this end, they present TRU-NET (Temporal Recurrent U-Net), an open source Python-based deep learning framework to effectively model multi-scale spatiotemporal weather processes and provide high-resolution predictions of rainfall. Funded by Alan Turing Institute under Climate Action Pilot Projects, TRU-NET uses HPC Facilities at the University of Warwick. The project was benchmarked against existing tools and its suitability was assessed through rigorous experimentation predicting seasonal performance metrics for each model tested on the whole country and predictive errors on 5 specific cities across the range of precipitation profiles. Based on their analysis, the TRU-NET team concluded that the available data is sufficient for applying a deep learning approach to produce robust, high-quality predictions for rainfall across seasons and varying regions. Climate prediction methods such as TRU-NET are important for guiding the decision of policy-makers reducing the financial and societal risk posed by flooding (Shukla et al. 2019).