How can battery storage systems make smarter decisions about when to charge and discharge – even when electricity tariffs change?
A new scientific paper involving researchers from the Technical University of Munich (TUM), a partner in the INTELLIGENT project, explores an innovative machine-learning approach to make battery control more adaptable, efficient and scalable.
The paper, “Tariff-Aware Imitation Learning for Transferable Battery Storage Control,” was authored by Manuel Katholnigg (TUM), Sheng Yin (Kempten University of Applied Sciences), Elgin Kollnig (TUM), and Christoph Goebel (TUM). It was published in the proceedings of E-Energy ’26 – The 17th ACM International Conference on Future and Sustainable Energy Systems, held from 22–25 June 2026 in Banff, Alberta, Canada.
For commercial buildings, electricity costs are not determined only by how much energy is consumed. In many tariff systems, businesses also pay significant charges based on their highest electricity demand during specific periods. This means that a short peak in electricity use can have a major impact on the monthly bill.
Battery storage can help by storing energy and releasing it when demand is high – a strategy known as peak shaving. But deciding exactly when a battery should charge or discharge is complex. Existing optimisation-based approaches can work well, but often require significant computing resources and need to be specifically adapted when electricity tariffs or operating conditions change.
The researchers therefore asked a practical question: Can a battery controller learn the logic behind good energy-management decisions and apply it when tariff schedules change?
The researchers developed a method called Tariff-Aware Imitation Learning.
Put simply, an advanced optimisation model first acts as a “teacher”, demonstrating how the battery should be operated. A much lighter machine-learning model then learns from these examples.
The important difference is that the new model does not simply memorise at what time the battery was previously charged or discharged. Instead, it receives information about the electricity tariff – for example, whether a peak-pricing period is currently active, when the next pricing period begins or ends, and what the current energy price is. This allows the system to learn the economic logic behind battery operation, rather than simply following a fixed schedule.
This is particularly important in real-world applications, where tariff structures can vary between locations and change over time.
The approach was evaluated using data representing 25 commercial buildings, including offices, retail buildings, warehouses, restaurants and schools, and across three different types of electricity tariff structures.
The results are promising: the tariff-aware controller was able to adapt to tariff schedules it had not seen during training without being retrained, reducing electricity bills by approximately 12–15% across the three tariff designs tested.
It also substantially outperformed a simpler rule-based controller given the same tariff information. Another interesting result emerged when uncertainty was introduced into electricity-demand forecasts: with forecast errors of 5% or more, the learned controller performed better than the optimisation-based approach because it reacted to actual observed conditions rather than relying on inaccurate forecasts.
The approach is also computationally lightweight. In the experiments, the learned controller made decisions around 400 times faster than the optimisation model used as its teacher, highlighting its potential for deployment on lower-cost hardware.
The research points towards a future in which intelligent battery controllers could be deployed across different buildings and tariff environments without requiring extensive site-specific engineering or retraining.
By enabling batteries to respond not only to energy demand but also to the economic signals embedded in electricity tariffs, approaches like this could make battery storage easier to deploy at scale and help businesses use stored energy more effectively.
For the broader energy transition, this is an important step towards more flexible and intelligent energy management systems – systems capable of adapting to changing conditions while helping users reduce peak demand and electricity costs.
Interested in the full research? Read the scientific paper here:
Tariff-Aware Imitation Learning for Transferable Battery Storage Control
