EPFL, Empa and SBB Develop AI Model to Improve Electricity Demand Forecasting for Swiss Rail

Forecasting errors regarding SBB’s electricity demand can lead to operational risks and high costs. Image: Adobe Stock

(IN BRIEF) Researchers from EPFL and Empa, in collaboration with SBB, have developed an AI model that improves next-day electricity demand forecasting across Switzerland’s rail network. The model combines historical energy-consumption data with contextual information such as train timetables, operational planning data and weather forecasts. The results, published in Energy Reports, show a 26.6% reduction in average prediction error and an approximately 80% reduction in major errors on unusual operating days. The approach could help SBB improve energy management, reduce operational risk, lower costs and decrease environmental impact. The researchers also noted that similar methods could benefit other sectors such as building energy systems, manufacturing, logistics and supply-chain management.

(PRESS RELEASE) DÜBENDORF, 20-Jul-2026 — /EuropaWire/ — Researchers from EPFL and Empa, working in collaboration with SBB, have developed an artificial intelligence model that improves next-day electricity demand forecasts across Switzerland’s railway network.

The model reduces average prediction errors by 26.6% compared with approaches that do not use additional contextual information. On unusual operating days, when railway activity or passenger flows differ significantly from previous patterns, the model can reduce major forecasting errors by around 80%.

Accurate forecasting is critical for Switzerland’s rail system, which must anticipate daily electricity needs to keep trains operating reliably. Demand can vary significantly depending on passenger flows, weather, timetable changes and operational conditions.

Large errors in electricity demand forecasts can create operational risks and additional costs. Previous studies suggest that even a 1% reduction in forecasting error could result in annual savings of around one million Swiss francs.

The research team was led by Olga Fink, tenure-track assistant professor at EPFL, and PhD student Raffael Pascal Theiler, with partners from SBB, Empa, ETH Zurich and MIT.

Their model combines historical railway energy-consumption data with contextual information, including train timetables, operational planning data and weather forecasts.

The results, published in Energy Reports, show that information about planned future operations can significantly improve forecasting accuracy, especially on days when the rail system operates differently from historical patterns.

Fink said the model performs better on unusual days because it has access to information about planned future operations, allowing it to prepare for changes before they occur.

The approach could help improve Swiss rail energy management, reduce costs and lower environmental impact by enabling more efficient use of electricity across the network.

The project initially focused on anomaly detection for hydropower generation at a single plant. However, the researchers later identified a wider application in forecasting electricity demand across the Swiss railway system by using contextual information from railway operations and the power grid.

Such information is often fragmented across different organisations. SBB, however, collects extensive operational data, including train schedules and other information available in advance. This allowed the researchers to investigate how knowledge of future operations can improve forecasting performance.

Theiler said classical forecasting models that rely only on past data often struggle with complex systems where demand depends on many interacting factors.

He said information about an operator’s plans can often be more useful than relying only on what happened in the past.

By combining historical electricity consumption with expected future activity, such as timetables, planned train services and scheduled operations, the researchers were able to outperform conventional forecasting approaches.

Theiler explained that large and complex systems depend on many actors coordinating around a shared operational plan. When such plans are documented, they become a valuable source of forecasting information.

The researchers said the approach is not limited to railways. Similar methods have already been demonstrated in building energy systems, where occupancy schedules and planned activities can improve energy demand forecasts.

The same principle could also apply to manufacturing, logistics and supply-chain management, where production plans and schedules provide useful signals about future activity.

Switzerland’s railway power system includes more than 1,800 kilometres of transmission lines, around 70 substations and 13 power plants and converter stations.

SBB operates a dedicated power grid separate from the conventional electricity network. Because train movements create highly dynamic electricity demand, the company uses pumped-storage hydropower facilities in part to store energy and adjust generation to changing operating conditions.

The researchers said the new AI model demonstrates how future operational context can support better energy forecasting and help complex infrastructure systems operate more efficiently.

The Swiss railway power system

The Swiss railway power system comprises more than 1,800 kilometers of transmission lines, around 70 substations, and 13 power plants and converter stations.  SBB operates a dedicated power grid, separated from the conventional electricity network. Because train movements create highly dynamic electricity demand, SBB relies in part on pumped-storage hydropower facilities to store energy and adjust generation to changing operating conditions.

Media Contacts:

Leandro Von Krannichfeldt
Urban Energy Systems
Phone +41 58 765 6051
leandro.vonkrannichfeldt@empa.ch

Dr. Andrea Six
Communications
Phone +41 58 765 6133
redaktion@empa.ch

SOURCE: Empa