Full Program »
Panel 3: Data science and energy systems -- industry challenges and case studies
Friday, 21 August 2020
09:00 - 11:00
Chair: Chris Dent, University of Edinburgh and Alan Turing Institute
Data Science and Energy Systems: Industry Challenges and Case Studies
Chair: Dr. Chris Dent (University of Edinburgh and Alan Turing Institute
There is a need for more effective use of data in power system planning and operation, due to the integration of ever higher variable generation penetrations, to increasingly active coordination of demand, and the ongoing desire to run systems economically and with an appropriate level of security of supply. This panel will demonstrate the value of bringing together the professional data science and power system analysis communities to solve practical industry challenges – in recent years such collaboration has happened less broadly with the data science community than for instance with control and optimization. There will be four talks from experts in industry and research organisations, demonstrating where innovative data science methods have brought benefits to real world questions, and discussing the challenges (both technical and organisational) associated with making wider use of advanced data science methods in practical situations.
Hui Yan (EDF)
Probabilistic forecasting methods for energy management systems
Peak shaving is a critical task in the energy sector arising from several motivations, including grid safety and energy cost optimization. Various methodologies and tools can be used to perform probabilistic forecasting or quantile regression. Here, we present a use case where we use a distribution-based approach – GAMLSS (Generalized Additive Model for Location, Scale and Shape) – to provide a peak prediction service for a battery storage management system. New ideas also appear such as a multi-resolution framework using information from different temporal scales efficiently in the context of daily demand peak forecasting or multi-resolution neural network.
Hui Yan is a research engineer in the load forecasting group at EDF Lab Saclay, in France. His team focuses on developing statistical/machine learning models for load forecasting problems which differ in time frame, aggregation level and client type. His main personal research interests are probabilistic forecasting for energy management systems, and automatic calibration of energy forecasting models at scale. He holds a French engineer’s degree in Applied Mathematics from INSA Rouen, and a MSc in Stochastic Modeling from University of Paris Diderot.
Henrik Madsen (Danish Technical University)
Probabilistic Forecasting for System Operators in a Low-Carbon Society
The green transition implies that the power system will undergo a fundamental change from a system where centralized power production follows demand to a system where the demand follows the renewable power production at all scales of the future power system. We are facing a disruption of our energy system, driven by decarbonisation, technology, and changing consumer behavior. This implies that energy forecasting will play an important role for system operators in the future low-carbon society, and for system operators it is crucial to have access to the best possible tools.
In this talk also state-of-the-art methodologies for probabilistic energy forecasting will be described. In particular we will focus on methods for multivariate probabilistic forecasting of load and renewable power generation. It will be argued that tools for integrated forecasting across domains (wind, solar, load, prices, …) will become essential, and replace more silo-oriented and independent tools for individual areas like wind power. It will be demonstrated that such probabilistic forecasts will become essential to obtaining reliability and profitability in the operation of the future low-carbon energy systems. Examples will be taken eg. from the Danish power system where on average more that 50 pct of the load now is covered by wind and solar power.
Martin Barons (Warwick University)
Structured expert judgement elicitation approach to estimating the probability of a major power system unreliability event
Preventing, containing and recovering quickly and safely from disturbances to the power system, is essential. Security standards and defence plans are in place worldwide but how often might a system collapse and black start be required? Major events remain very rare and have shown a number of common phenomena, but the precise pathway is different each time and depends on complex and uncertain system behaviour and characteristics. The use of modelling to estimate the probability of system shutdown is therefore extremely difficult.
We describe a Structured Expert Judgement (SEJ) approach. SEJ is used where major risks need to be understood in respect of phenomena that are not readily amenable to modelling. We develop a conceptual, high-level model of the power system and use it as the basis for questions to elicit the judgements of relevant experts. Facilitated discussion tests experts’ buy-in to the model, understanding of the questions, and highlights the range of scenarios included. SEJ allows not only the elicitation of each expert’s own judgement, but also the assessment of the experts’ abilities to estimate a probability and an uncertainty range within which they believe the true answer lies. This is used in weighting experts’ answers to the main questions of interest. Experts are given the opportunity to share rationales and discuss the questions in two rounds to enable enhanced understanding of the situation and to adjust assessments based on these insights.
Martine Barons is the Director of the Applied Statistics & Risk Unit at the University of Warwick. Martine started her career in Accountancy where she gained an appreciation of multiple types of business, government and third sector organisations, their imperatives and modes or working. Martine re-trained in Mathematics, gaining her PhD in 2013. Her research interests are in decision support, with a particular interest in graphical models. Through AS&RU she undertakes and manages research projects with industry partners in a large variety of domains.
Jean-Paul Watson (Lawrence Livermore National Laboratory)
Data-Driven Models for Decision Making Under Uncertainty
Explicit representations of uncertainty are starting to appear in various power systems operations and planning contexts, in support of both reliability and resilience objectives. Examples include scenario trees for stochastic programming models and uncertainty set descriptions for robust optimization models. Accurate estimation of key uncertain parameters requires significant historical data, which is often not archived and therefore generally unavailable. We describe primary challenges associated with and methods for developing uncertainty representations for key power systems operations and planning problems, with illustrative applications.
Dr. Jean-Paul Watson is a Senior Research Scientist at Lawrence Livermore National Laboratory in Livermore, California, USA. Dr. Watson leads numerous research efforts in the areas of power systems operations, planning, and resilience; sponsoring agencies include DOE/ARPA-E, DOE/OE, and DOE/EERE. He has nearly 17 years of experience applying and analyzing algorithms for solving difficult combinatorial optimization and informatics problems, in fields ranging from logistics and infrastructure security to power systems and computational chemistry. He is a co-developer of the Pyomo (www.pyomo.org) open-source software package for modeling and solving optimization problems, and has published over 65 journal articles in the areas of optimization algorithms and their application.