Energy Forecasting Innovation Conference - Building capacity from modern statistical methodology

'Energy forecasting innovation conference - Building capacity from modern statistical methodology' will present findings from three major projects funded by UK Research and Innovation with a combined value of approximately £1.5 million. The conference will include:

• Presentations from academic researchers and their industrial partners (NGESO, EDF, SSEN, Shell and others);
• An afternoon workshop to identify future research needs and opportunities, in close engagement with EPSRC-UKRI;
• Two optional training courses in open-source software: ProbCast (Dr Jethro Browell) and Statistical Inference with Max-Stable Processes (Dr Kirstin Strokorb and Dr Marco Oesting).

The three projects are:

• Multivariate Max-stable Processes with Application to the Forecasting of Multiple Hazards (Dr Claudia Neves, EPSRC Innovation Fellowship)
• System-wide Probabilistic Energy Forecasting (Dr Jethro Browell, EPSRC Innovation Fellowship)
• Analytical Middleware for Informed Distribution Networks (Dr Bruce Stephen, Principal Investigator of AMIDiNe)

 

Registration

This is a hybrid event (in person and online), hosted by King's College London. Please click 'Register for this event' to register via Eventbrite. Please ensure that you select the correct ticket type (In-Person Ticket or Online Ticket), and book a ticket for each day you would like to attend. There is no registration fee. Refreshments and lunch will be provided free of charge to those attending in person.

The two training courses will run simultaneously and will take place in person only. Please register for a maximum of one training course. Each of the two courses will be limited to 25 participants, who are requested to bring their own laptop with R-software installed.

After you have registered, you will be emailed a short survey. We kindly ask you to fill this in as part of the registration process.

 

Timetable

Tuesday 24 May - morning

Industry Challenge - Industry speakers including representatives of National Grid ESO, EDF, SSEN and E.ON. Speakers include:

• Daniel Drew (NGESO)
• Maciej Fila (SSEN)
• Matthew Allcock (EDF Energy)

Tuesday 24 May - afternoon

Research - Findings of three UKRI projects presented by their leaders - Claudia Neves, Jethro Browell and Bruce Stephen - plus invited speakers, including:

• Dr Jennifer Wadsworth (Lancaster University)
• Professor Dr Marco Oesting (Universität Stuttgart)
• Gordon McFadzean (TNEI Services)
• Adam Brown (Bellrock Technology)

Wednesday 25 May - morning

Innovation Showcase - Translating research into practice: academic collaborations with Shell, TNEI, Bellrock Technology, Capgemini Engineering, Energy Systems Catapult and others. Speakers include:

• Professor David Brayshaw (University of Reading)
• Théo Moins (Inria Grenoble Rhône-Alpes)
• Danica Greetham (Capgemini Engineering)
• Stephen Haben (Energy Systems Catapult)
• Dr Bruce Stephen (University of Strathclyde)

Wednesday 25 May - afternoon

UKRI Workshop - Interactive workshop to define research needs, spin out partnerships and identify funding opportunities.

Thursday 26 May - morning (in-person only)

Training - Two training courses in the use of research outputs and dedicated software, as outlined below: 

ProbCast: an R package for probabilistic forecasting

Learning objectives

Participants will leave this course with a working R script implementing state-of-the-art methods for electricity demand and wind power forecasting. After this training course, participants will be able to:

• Describe the principles of parametric probabilistic forecasting, and non-parametric (multiple quantile regression) probabilistic forecasting
• Use ProbCast to produce parametric probabilistic forecasts building on the framework of Generalised Additive Models for Location Scale and Shape (GAMLSS)
• Use ProbCast to produce multiple quantile forecasts based on gradient boosting machines (a best-in-class algorithm for renewable energy forecasting)
• Use ProbCast to visualise probabilistic forecasts
• Use ProbCast to evaluate probabilistic forecast in terms of reliability/calibration and Pinball Score

Requirements

Prior knowledge of basic probability and probability distractions, regression models, and programming (ideally but not necessarily R) is essential. Attendees should bring a laptop with R and ProbCast already installed. If you cannot bring your own laptop, please contact us and we will make every effort to arrange sharing.

Instructions for installing ProbCast.

About ProbCast

ProbCast is an R package for producing and working with predictive distributions. It was developed with energy forecasting in mind, but its functionality is general and may be useful for any probabilistic prediction task. ProbCast defines two new data classes, one for parametric probabilistic forecasts, and another for non-paramedic predictions in the form of multiple quantiles. The latter may be augmented with parametric tail distributions. Functions are provided to rapidly visualise and evaluate such forecasts. Furthermore, ProbCast provides wrapper functions for a range of model types, including generic and tree boosting methods (gbm, lightGBM, mboost, rq) for quantile regression, and gamlss-type models for distributional regression (gamlss, gamboostLSS).

ProbCast was initially developed by Jethro Browell and Ciaran Gilbert as part of the EPSRC Fellowship “System-wide probabilistic energy forecasting” (EP/R023484/1). It is maintained on GitHub, where a full list of contributors and acknowledgements can be found.

 

Statistical inference with max-stable processes (using R-software)

This course will complement the talk by Dr Marco Oesting on the afternoon of 24 May.

Learning objectives

• Perform statistical inference with spatial data using max-stable processes
• Simulate max-stable processes using a range of available methods
• Learn how method choices can be adapted to practical needs

Prior knowledge

Basic understanding of probability distributions required. In the training event, basic familiarity with the R language (including usage of data frames) is assumed. Participants are expected to bring their own laptop with R installed.

About the course

Max-stable processes have become popular models for spatial and spatiotemporal extremes, specifically in environmental applications. In statistical practice, they are used to approximate pointwise areal block maxima taken over sufficiently long time periods. This can help to answer practical questions about the distribution of (joint) areal maxima of a quantity of interest (e.g. temperature or wind) or the distribution of simultaneous threshold exceedances in specified areas. This course will guide through the main steps of statistical analysis using max-stable processes. While simulation plays an important role in the analysis, most of the available methods can at first sight seem practically prohibitive due to the high computational cost for each simulation. We will provide guidance how method choices can be adapted to practical needs.

Sponsorship

This event is sponsored by three projects funded by UK Research and Innovation.