Building Energy Calibration based on Parameter Estimation and Machine Learning
- Golden Analytics
- M.S.Addison and Associates
- SAM Associates; Arizona State Univ., Tempe, AZ (United States)
Buildings rarely perform as designed/simulated and there are numerous tangible benefits if this gap is reconciled. Building energy simulations are essential not only during the design phase but also offer benefits during building operation. Examples are commissioning, fault detection and diagnostics, identifying and evaluating retrofit opportunities, and optimal predictive control especially in the context of electrical grid services meant to assure/enhance reliability and flexibility with increasing renewable penetration. This project develops a scientific yet pragmatic methodology that focusses on parametric estimation rather than a blind forced-fit to energy use data. It starts with rapidly and inexpensively created simulation inputs and reconciles simulated energy and indoor temperatures with actual performance using readily available hourly or sub-hourly performance data from smart meters, building automation systems and local weather stations. Current calibration methods provide little physical insight into actual parameters and often over-promise accuracy. Rather than subjectively adjusting a large number of inputs randomly and doing a large number of simulations, the new method is more scientifically-grounded in that it identifies a small number of parameters, and estimates them and their uncertainty bounds using a multi-stage process involving different types of energy flow balances.. The parameters have direct physical significance. The residual error in energy balance is related to forcing functions using machine learning to enhance the accuracy of calibration. Thus the burden of arriving at a good final model is shared between reasonable initial inputs, a calibration process based on the physics of heat flows in buildings and a machine learning final step to capture unexplained variation. In Phase I, this is demonstrated in a synthetic building and one 75,000 sq. ft. actual building in Pennsylvania. The effect of HVAC and primary systems are also integrated into the methodology. In Phase II, the methodology and tool will be refined and a prototype developed for field deployment. This is expected to create new commercial opportunities for building performance assurance not only for the building energy services community, but also for utility services community involved in dispatching flexibility services to the electrical grid. Phase I was highly successful in meeting the project objectives. In summary: Various sources of zone level heat flows were defined and computed using specially formulated EnergyPlus simulations Least Squares fit to the model residuals of the energy balance was accomplished by introducing and estimating parameters with physical significance and was demonstrated in synthetic and real buildings A complementary method was developed (i) for analyzing the building shell, the HVAC system is replaced by an ideal system; and (ii) for analyzing the HVAC system, the building shell is replaced by a box with only process loads. Improved model accuracy was achieved by implementing a neural net for the residuals remaining after the least squares correction. Inputs to the neural net are computed heat flows that automatically incorporate the history of temperatures, solar radiation, internal gains, etc., thus simplifying a time series problem into one in which each hour is independent as far as the neural net is concerned. A multistage parameter estimation technique can reduce the effect of multi-collinearity and results in less biased physical parameters. This aspect needs further investigation.
- Research Organization:
- Golden Analytics LLC
- Sponsoring Organization:
- USDOE Office of Energy Efficiency and Renewable Energy (EERE); USDOE Office of Science (SC), Engineering & Technology. Office of Small Business Innovation Research (SBIR) and Small Business Technology Transfer (STTR) Programs
- DOE Contract Number:
- SC0018811
- OSTI ID:
- 1515544
- Type / Phase:
- SBIR (Phase I)
- Report Number(s):
- DOE-GA-0018811
- Country of Publication:
- United States
- Language:
- English
Similar Records
LLNL Kimberlina 1.2 NUFT Simulations June 2018 (v2)
Carbon dioxide prediction model for VAV system part-load evaluation