Identification of Building Thermal Model - Using Data
summary
Claims a data‑driven pipeline replaces year‑long semi‑physical calibration and the Unscented Kalman Filter by training lightweight models on short windows of measured data and weather forecasts. Uses LightGBM with time‑series split and 6‑hour lag features, sampling every 15 minutes from Z‑Wave smart‑home sensors across 25 zones, with a one‑week training window and 24‑hour rolling forecasts. Implements stepping predictions (injecting prior forecasts as lags) and runs as TinyML on local controllers, yielding about 0.3°C MAE and ~1.6% MAPE (max error ≈1°C).
AI-written summary · about
chapters (20)
- 0:00Tutorial introduction
- 1:05Identify building thermal models
- 3:55Global energy use solutions
- 6:42Traditional modeling challenges
- 8:07Smart home experimental setup
- 10:13Data collection architecture
- 11:40TinyML for smart home control
- 13:47Data preprocessing engineering
- 15:37Merging datasets for modeling
- 21:21Visualizing temperature trends
- 23:54Feature selection for prediction
- 26:51Lagged features for time series
- 29:44LightGBM model setup
- 31:52Time series cross-validation
- 35:02Avoiding data leakage
- 37:53Model performance metrics
- 39:43Testing multiple weeks
- 42:55Analyzing model errors
- 45:01Conclusion and TinyML benefits
- 46:26Acknowledgments
related talks
23mDataEngBytes 2023 - MEL-T3-01 - Rajath Vanikul
35mWorking with real-time data streams in Python
32mPredicting Grid Faults Before They Happen: AI for Weather-Driven Grid Resilience
31mFinding the Needle in the Haystack: Predicting Storage Device Failures in...
28mUse case in demand forecasting
07mSmarter Cities through Geotab
44mDemocratizing Climate Science
28mMastering Demand Forecasting: Lessons from Europe's Largest Retailer