© 2026 EPFL
In collaboration with SBB, EPFL researchers have developed an AI model that improves next-day electricity-demand forecasts across Switzerland's rail network, reducing major prediction errors by up to 80%.
Every day, Switzerland's rail network must anticipate how much electricity it will need to keep trains running. Yet accurate forecasts are challenging as demand depends on many interacting factors, including fluctuating passenger flows, weather conditions, and constantly changing operations. Large prediction errors can lead to significant operational risks and costs. Some studies suggest that even a 1% reduction in forecasting error could translate into annual savings of around one million Swiss francs.
To tackle this problem, a team of researchers at EPFL, led by tenure-track assistant professor Olga Fink and PhD student Raffael Pascal Theiler, developed a new AI forecasting model in collaboration with SBB and partners at Empa, ETH Zurich, and MIT. Their approach combines historical railway data with contextual information such as train timetables, operational planning data, and weather forecasts.
The results, published in Energy Reports, show that incorporating this additional information reduces the average prediction error by 26.6% compared to the model without this information. The improvement is even more striking on unusual days, when train operations or passenger flows differ significantly from the past. In this case, large forecasting errors are reduced by about 80%. "Because the model has information about planned future operations, it is better prepared for unusual days," says Fink. The approach could help make Swiss rail operations more efficient by improving energy management, while reducing costs and environmental impact.
The Swiss rail network: an ideal testing ground
The project initially focused on developing anomaly detection methods for hydropower generation in a single plant. Fink and Theiler soon realized that, by leveraging contextual information from railway operations and the power grid, their model could help address the challenges of forecasting Swiss-wide railway energy demand. While this contextual information is typically fragmented across separate organizations, SBB collects operational data extensively, including data known in advance, such as train schedules. This unique dataset, combining energy measurements with forward-looking operational information, offered an opportunity to investigate how knowledge of future operations can improve forecasting performance.
Outperforming classical algorithms
In systems as complex as national rail networks, many factors influence electricity demand, such as weather conditions and passenger affluence. Classical models that rely solely on historical data often struggle to capture these dynamics. "I'm not really a strong believer in forecasting based on just past information. In many cases, information on the operator's plans can guide us significantly more than what we can infer from what happened in the past," says Theiler.
By combining historical energy-consumption data with "expected future", that is, information that describes how the system is projected to operate, such as railway timetables, planned train services, or scheduled activities, the researchers significantly improved forecasting accuracy compared with conventional approaches. "Large and complex systems can operate efficiently only when many actors coordinate around a shared plan of how operations are expected to unfold. When this plan is documented, it becomes a valuable source of information for forecasting," explains Theiler.
Applications beyond the rails
Railway operation is just one example of how future contextual information can help improve energy demand forecasting. The researchers have already demonstrated the approach in building energy systems, where occupancy schedules and planned activities improve forecasts of energy demand. But the same approach can be applied to other areas such as manufacturing, logistics, and supply-chain management, where production plans and schedules contain valuable information about future activity. "I hope our study motivates people from industry to study their data and use a system like the one we propose," concludes Theiler.
