Distributed Learning for Probabilistic Forecasting in Energy Systems

The urgent push for decarbonisation in the energy sector has increased the uncertainty surrounding demand and supply, as well as related energy system variables, such as short-circuit level, inertia and voltage support. Moreover, commercial stakeholders and end-users are starting to participate actively to the secure and cost-effective operation of power systems, adding their behavioural uncertainty to the system operation. Leading system operators worldwide have indicated the need for accurate forecast models to enable the integration of the increased number of market actors as well as the massive integration of variable renewable energy sources, such as wind and solar, needed to achieve the climate targets. These forecast models should accurately quantify the uncertainty on the predicted variables and optimally use the available information, including information of privacy-conscious energy system actors. Privacy-preserving forecasting techniques are essential to convince these actors to contribute to the forecast model development and forecast procedure.
The aim of this project is to develop forecasts of energy system variables that accurately quantify the uncertainty on the predicted variables and are privacy-preserving in nature. Distributed learning, which involves the implementation of machine learning algorithms in a decentralised manner, is one way of creating privacy-preserving models. When this technique is used, data does not need to be shared with a central entity; instead a machine learning algorithm is run locally by any actors who would typically contribute data, and the results from this are then combined by a central entity. Existing work has focused on probabilistic privacy-preserving forecasting of wind power production. However, energy system variables, such as inertia and short-circuit level, depend upon multiple uncertain variables, such as renewable energy production and demand. During this project, different methods for distributed learning will be applied to produce probabilistic forecast models of a variety of energy system variables and the performance of the methods will be evaluated in these different application contexts. During the initial stages of this project, these methods will be applied to the forecasting of synchronous generator system inertia. This is an increasingly essential area of research as the decarbonisation of energy systems has led to a reduction in the total amount of system inertia available. Renewable energy sources (RES) which are replacing their fossil-fuel counterparts are also highly unpredictable which makes it hard for the ESO to predict how much of an impact they will have on the system on any given day. Moreover, consideration will need to be given to the design of incentives to convince actors to participate in this decentralized framework. Additionally, the outcome of models created as part of this project will be further improved by unifying and simultaneously training any forecasting and decision-making models. Finally, the use of explainable artificial intelligence techniques (explainable AI) may also be considered in this context as energy system operators (ESOs) must make high-stakes decisions which they should be able to justify; this is not always possible if the model they are using is a 'black-box'.
Overall, this project will help to facilitate the decarbonisation of the energy sector by providing accurate, privacy-preserving forecasting models over a range of areas, starting with synchronous generator system inertia.