This research project aims to develop machine learning systems that keep their uncertainty estimates reliable as conditions change. It uses electricity markets as a test case. Deadline for Application & Eligibility The deadline to submit your first-stage application form is 11 February 2027, 23:59.You must follow the CDT MLSystems application process as described in those webpages. Please contact the project's PI ahead of submitting your application to first check suitability and interest.Eligibility: this project is open for applications to Home students only. Co-funding CompanySSE plcSSE plc is one of the UK’s leading renewable energy generators and electricity market participants. It operates a large portfolio of renewable generation assets across Great Britain and Europe, and builds forecasting and optimisation systems for dispatch, portfolio managementand risk management. For an operator of physical assets, forecast uncertainty matters as much as the forecast itself, because prediction intervals inform the positions the business takes. Intervals that are right on average but too narrow in particular conditions cause problems exactly when uncertainty is highest.https://www.sse.com/Supervisory teamUniversity of Edinburgh PI: Amanda Lenzi, Lecturer in Statistics, School of Mathematics, Amanda.Lenzi@ed.ac.ukPersonal website: Amanda Lenzi - University of Edinburgh Research ExplorerUniversity of Edinburgh Co-supervisor: Amos Storkey, Professor of Machine Learning and Artificial Intelligence, School of InformaticsCompany supervisor: Pietro Colombo, Quantitative Analyst, SSE Energy MarketsAbstractForecasts used to manage renewable energy portfolios often come with prediction intervals, but their reliability is typically assessed across all operating conditions. This can mask important failures: intervals that achieve their target coverage overall may be far too narrow under particular conditions, such as periods of high wind and transmission congestion. These failures are especially consequential when forecasts inform operational and trading decisions.This project develops machine learning systems that keep their uncertainty estimates reliable as conditions change. Instead of treating calibration as a fixed post-processing step, we treat it as a sequential decision problem. An adaptive agent monitors the forecasting environment, infers which operating regime it is in, decides which parts of history are still relevant, and adjusts its calibration strategy, while seeking robustness to errors in regime identification. Electricity markets are the test case, but the methods apply to any deployed ML system whose behaviour switches between regimes. You will start on open data and work towards a real-time demonstrator with SSE.Project BackgroundMost forecasting pipelines train a model, apply a fixed calibration procedure, and output uncertainty estimates. This assumes future forecast errors resemble those in the calibration set. In operational systems that assumption fails, and it fails unevenly: some conditions are well represented in recent data and others are not. The usual fix is a rolling window of recent observations. That uses time as a proxy for "conditions like today’s", and assumes recency implies similarity. When operating regimes recur, a rolling window discards exactly the history that is most informative when a regime returns. We want methods that use similar past periods as context alongside recent ones. An electricity market is an abstraction of a physical system made of interconnected sub-systems, each with its own behaviour. Looking only at the top-level price hides which sub-system is driving a shift, so it becomes essential to identify, categorise and separate the sources of change. The methods developed here should deliver their largest improvements in the edge and complex cases of such a system, where standard calibration is least reliable. Regimes of interest include:An interconnector or network boundary that is congested versus uncongested;high wind coinciding with low demand;cold, still periods with tight supply;structural change, such as new market rules or growth in renewable capacity.The central question is:How should a machine learning system determine which past observations remain informative for uncertainty estimation under changing conditions, and what can it guarantee when that determination is wrong?This is a statistical problem, because coverage guarantees must survive an imperfectly inferred regime. It is also a systems problem, because the answer has to be computed online, inside a deployed pipeline, under a daily deadline.We seek to investigate novel methods that will afford greater ability to adjust for this known weakness in relying on recency, by expanding to include similar reference periods as an additional source of context for the next period, alongside recency.Project AimsThe project has a defined starting point and a core question. Beyond that, you will shape which directions to pursue in depth with the supervisory team.Year 1: show the problem and build the benchmark. Using public data only (energy flows and capacities, GB and continental day-ahead prices, open ensemble weather forecasts), you will test across several realistic pipeline designs whether regime-conditional miscalibration is systematic and whether average-coverage validation misses it. Calibration will be measured by coverage and interval scores within each regime. Outputs: an open benchmark and evaluation toolkit, and a first paper.Core: regime-aware calibration with guarantees. You will learn representations of forecaststate similarity that identify the operating regime and hence which historical examples remain relevant. These will be anchored in observable physical diagnostics, so they stay interpretable and can be checked against how the system works. The central theoretical result is a bound on how within-regime coverage degrades when the regime is misidentified. Baselines include group-conditional, adaptive and similarity-weighted conformal methods.Directions to choose from:Calibration agents. Agents that choose calibration strategies, observation weights, adaptation rates and when to widen or narrow intervals. They optimise long-term reliability using a proper interval score with a penalty for coverage shortfall within each regime, and the coverage bound keeps their decisions within safe limits.Continual adaptation in a live pipeline. Forgetting indexed on operating regime instead of time. A key question is whether calibration parameters drift smoothly or jump when the regime changes. You will also characterise the trade-off between compute spent on regime inference, calibration-set size and achieved coverage under a daily deadline.Tails and joint targets. For a renewable generator the critical tail is high wind coinciding with collapsed or negative prices, which drives capture price and curtailment exposure. Capture price depends on the joint distribution of wind output and price, and their dependence switches with congestion. This direction extends regime-conditional calibration to extreme quantiles and joint quantities using multivariate extreme value and copula methods.Beyond energy. The same ideas apply to other ML systems whose behaviour switches between unobserved regimes.Expected Outcome and TimescaleYear Focus Main outputs 1Benchmark on public data; first regimerepresentationsOpen benchmark and evaluation toolkit;diagnostic paper2Regime-aware calibration andmisclassification theory; first agentsMain methods paper3Chosen directions; continual adaptation;optional SSE placementOpen-source calibration framework;real-time demonstrator 4Tails and joint targets, or further depth ina chosen direction; integrationFurther papers; thesisTarget venues span machine learning (NeurIPS, ICML, AISTATS), systems (MLSys), statistics (JRSS-B, JASA) and forecasting (International Journal of Forecasting). Applications include renewable generation and electricity price forecasting, capture price and curtailment risk, financial risk forecasting, and any ML system operating under non-stationarity. Overall objective is a general framework for probabilistic forecasting systems that recognise changing environments, adapt their calibration strategy and keep uncertainty estimates reliable within each operating regime.Students Gains and RequirementsYou will be supervised across statistics, machine learning and industry. You will build expertise in distribution-free uncertainty quantification, online and continual learning, and running ML reliably in a deployed pipeline, and you will learn how electricity markets work from people who operate in them. These skills transfer directly to research and industry roles in ML, forecasting, energy and quantitative finance.The ideal candidate will have:First Class Honours degree or equivalent in Computer Science, Mathematics, Statistics, Data Science, Machine Learning or a related discipline.Strong programming skills, and an interest in building software.Interest in machine learning systems, uncertainty quantification and statistical methods.Desirable but not essential: Experience with forecasting, time series, reinforcement learning, sequential decision-making or probabilistic modelling. Prior energy-market experience.More on the CDT MLSystems requirements for candidatesWorking with SSEYou will have an industry supervisor at SSE Energy Markets in Edinburgh throughout the project, with regular contact with the team that builds and uses these forecasts. An on-site placement at SSE Energy Markets in Edinburgh is available and encouraged. It is not required to complete the project. The first two years run on public data, so progress does not depend on access to company data. Later work with SSE can use synthetic or obfuscated data that SSE generates and provides, so the project does not rely on private, licensed or proprietary data. Where access to SSE systems is useful, SSE can provide a laptop with secure access and specific permissions, protecting both SSE and the University. The aim is a real-time calibration demonstrator alongside SSE’s forecasting pipeline, which may run on synthetic or obfuscated data.Further ReadingAngelopoulos, A. N. and Bates, S. (2023). Conformal Prediction: A Gentle Introduction. Foundations and Trends in Machine Learning, 16, 494–591.Gibbs, I. and Candès, E. (2021). Adaptive Conformal Inference Under Distribution Shift. NeurIPS 34.Zaffran, M., Féron, O., Goude, Y., Josse, J. and Dieuleveut, A. (2022). Adaptive Conformal Predictions for Time Series. ICML, PMLR 162, 25834–25866.Gibbs, I., Cherian, J. J. and Candès, E. (2025). Conformal Prediction with Conditional Guarantees. Journal of the Royal Statistical Society: Series B, 87(4), 1100–1126.Sun, S. and Yu, R. (2025). Conformal Prediction for Time-series Forecasting with Change Points. NeurIPS 38.Stocker, Małgorzewicz, Fontana and Ben Taieb. A Gentle Introduction to Conformal Time Series Forecasting.Stocker, M., Małgorzewicz, W., Fontana, M. and Ben Taieb, S. (2025). A Gentle Introduction to Conformal Time Series Forecasting. arXiv:2511.13608. This article was published on Tuesday 4 November 2025