Reliable time-series forecasting is essential for operating a secure electricity system based on renewable energy. While forecasts at the level of countries or large regions benefit from the aggregation of many individual effects, forecasting becomes much more difficult at the “last mile” of the power grid. At the low-voltage level, where only a small number of homes and businesses may be connected, consumption and generation are more volatile, spiky, and diverse. The growing use of heat pumps, air-conditioning systems, electric vehicles, and decentralised generation makes accurate local forecasts increasingly important.

This workshop explores how time-series foundation models could help address these challenges. Participants will consider how forecasting systems can handle uncertainty, concept drift, and critical peak events while supporting reliable grid operation in a highly electrified energy system.

What participants will do

Participants will begin by exploring the structure of electricity grids and the role of forecasting in balancing generation and demand and preventing grid-asset overloads. They will then work with prepared power-grid time-series data in an online environment such as Google Colab. Using resources from the Netze BW and KIT research context, participants will investigate forecasting at lower grid levels and develop ideas for models, forecast values, or evaluation methods.

A central challenge will be to think beyond average accuracy: participants will consider how a forecast should represent aleatoric uncertainty, adapt to changing data patterns, and perform during high-load or high-generation peaks. They will discuss which metrics or evaluation approaches best reflect these operational requirements.

What to expect

Participants can expect an accessible introduction to the evolution from conventional time-series forecasting to time-series foundation models, followed by a practical exploration of a real energy-sector problem. The workshop will conclude with a discussion of the solutions and ideas developed by the participants, ranging from improved evaluation metrics and forecasting approaches to concrete model outputs.

Participants should bring a laptop and be prepared to work with internet-based tools; the workshop may use Google Colab, which could require a Google account. Background information and data resources are available through the FeederBW dataset and its accompanying code: https://zenodo.org/records/17831177 and https://github.com/KIT-IAI/FeederBW/tree/main