The problem
Buildings waste energy because heating, cooling and lighting are planned without a forecast. Deep models (LSTM) predict well but need a lot of history per building; new or renovated buildings have little. Transfer learning reuses what was learned on other buildings, and a light model can run on an edge gateway next to the meters.
Who needs it
Facility management, energy service companies, universities and hospitals with many buildings.
Idea of the approach
- Use public building-energy datasets (hourly meters and weather) plus campus data if available.
- Train an LSTM forecaster on many buildings, then fine-tune it on a target building with only a few weeks of data.
- Compare with simple baselines (same hour last week, linear regression) and train-from-scratch.
- Run the model on a Raspberry Pi gateway and show forecasts on a dashboard.
What you will build and deliver
- Forecasting models, evaluation and an edge dashboard.
- A paper.
Skills needed
Python, pandas, Keras or PyTorch; time-series basics.
Research contribution & where to publish
Contribution: how much history a new building needs when knowledge is transferred, on low-cost edge hardware. Venues: ACM BuildSys, ACM e-Energy, Energy and Buildings, Applied Energy.
Key references — start reading here
- S. Hochreiter, J. Schmidhuber. Long Short-Term Memory. Neural Computation, 1997.
- S. J. Pan, Q. Yang. A Survey on Transfer Learning. IEEE Transactions on Knowledge and Data Engineering, 2010.
- M. Satyanarayanan. The Emergence of Edge Computing. IEEE Computer, 2017.
Interested?
Send the topic title and the courses you have taken.
Contact Dr. Hamza Djigal