HPC-Enabled Deep Learning for Real-Time Energy Forecasting in Buildings
Organisations involved
Robotina d.o.o. is a Slovenian SME delivering IoT-enabled hardware and cloud software for smart buildings, industrial automation and energy management.
Rudolfovo Science and Technology Centre is a public research institute supporting technological development and innovation in South-East Slovenia.
The University of Ljubljana, Faculty of Mechanical Engineering creates and transfers engineering knowledge for industry and research partners.
The challenge
Robotina provides energy-management solutions for offices, hotels, shopping centres, public buildings and education facilities. Its aim is to extend the xEMS platform with a subscription service that uses real-time building data and AI to optimise energy use, reduce costs and support sustainable operation. Customers need better planning as dynamic tariffs, renewables and storage make buildings harder to manage. The service needs accurate 1-3 day forecasts at 15-minute resolution, using data collected every minute from smart meters, building systems, weather and tariff signals.
Earlier work showed that advanced recurrent neural networks, especially bidirectional long short-term memory (BiLSTM) models, could deliver the required accuracy. However, these models are computationally demanding because each building may need its own training, tuning and validation. On Robotina's standard CPU infrastructure, one demanding BiLSTM configuration with long sequences and multi-step outputs could take up to several weeks to explore. That bottleneck slowed development and restricted how many customer scenarios could be tested. HPC was needed to run model variants in parallel, shorten training cycles and make advanced forecasting scalable for commercial deployment.
The solution
The consortium developed an HPC-enabled forecasting workflow linking Robotina's energy management platform with the VEGA supercomputer. Real building, weather, tariff and market data were prepared automatically for AI training, then used to test and compare alternative forecasting models. On VEGA, the team used four GPU nodes running in parallel to make large model searches practical. The workflow was packaged in Singularity containers with Python and TensorFlow, so it can be reproduced and extended. HPC reduced demanding training and model-search runs from about one week to less than one day. The output is a validated pipeline and optimised model architecture for 15-minute, multi-step forecasting, plus a lightweight TensorFlow Lite version for deployment on edge EMS devices.
Business impacts
The experiment strengthens the xEMS platform by adding an AI-based forecasting module that can be offered as a deployable subscription service for commercial and public buildings. This improves Robotina's competitiveness in smart energy management by moving the company from monitoring and control towards predictive optimisation. HPC gives Robotina a repeatable way to train and refine advanced models for many customers, rather than relying on slow, building-by-building development on local CPU infrastructure.
Robotina estimates that each additional subscription enabled by the forecasting capability could bring around €80 per year, creating a potential annual revenue opportunity of up to €800,000. For customers, accurate 1-3 day forecasts support better scheduling of energy-intensive operations, improved use of dynamic tariffs and renewables, and lower energy costs, with potential savings of up to 10%. Improved accuracy also makes storage management commercially viable and creates a longer-term route into energy trading and flexibility markets.
Business benefits
- Forecasting error reduced from 47.88% to 10.49%, materially improving the customer value of xEMS.
- HPC cut demanding model-search and training runs from about one week to under one day, accelerating releases.
- Potential revenue opportunity of up to €800,000 per year from added subscriptions enabled by forecasting.
- Customers can reduce energy costs by up to 10% through better planning, scheduling and tariff use.
- Validated path from VEGA HPC training to TensorFlow Lite edge deployment on xEMS devices and offline operation.
- Improved accuracy enables storage management, energy trading and flexibility-market services.