# A hybrid-model approach for reducing the performance gap in building energy forecasting

**Authors:** Xia Chen, Tong Guo, Martin Kriegel, Philipp Geyer  
**arXiv:** [2206.00460](https://arxiv.org/abs/2206.00460)  
**Date:** 2022-05-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2206.00460v1/

## Key points

- Hybrid-model approach aims to minimize performance gap between predicted and actual energy consumption in the building domain
- Combines first-principles model and machine learning (ML) model
- First-principles method used as encoding tool to convert building static features and predictable patterns
- Results combined with historical records in ML model for capturing implicit performance differences and calibrating output
- Introduction of Level-of-Information (LOI) concept to balance information detail investment and accuracy boost
- Testing done using data from operating commercial building in Shanghai over three-year period
- Higher prediction accuracy and better interpretability compared to traditional methods
- Reduces modeling workload and computational resources required for refining simulations
- Time-series decomposition used to identify systematic and unsystematic patterns in building energy consumption data
- STL decomposition method employed to separate time-series into trend, seasonal, and residual components
- Trend and periodic sub-series suitable for knowledge-based simulation, random sub-series represent stochastic factors that cannot be directly modeled
- Discussion on universality of variables in Building Performance Simulation (BPS) community, higher universality indicates better acceptance, more methodologies/validation, and better data availability.
- Integration of domain knowledge via building simulation with data-driven methods improves prediction accuracy.
- Source code and data available on GitHub for reference purposes.
- Future work could focus on addressing limitations identified during case study and further refining hybrid model approach.

## Summary

The existing summary discusses a hybrid-model approach that aims to minimize the performance gap between predicted and actual energy consumption in the building domain. This gap exists in both the first-principles model and the machine learning (ML) model. The proposed approach combines these two methods by using the first-principles method as an encoding tool to convert building static features and predictable patterns, and then combining the results with historical records in the ML model to capture implicit performance differences and calibrate the output. To further analyze the balance between information detail investment and accuracy boost, a new concept called Level-of-Information (LOI) is introduced. LOI helps determine how much simulation modeling detail is needed for accurate predictions without overwhelming practitioners with excessive workload or computational resources. The approach is tested using data from an operating commercial building in Shanghai over a three-year period. The results show higher prediction accuracy and better interpretability compared to traditional methods. Additionally, it reduces modeling workload and computational resources required for refining simulations. In terms of methodology, time-series decomposition is used to identify systematic and unsystematic patterns in building energy consumption data. The STL decomposition method is employed to separate time-series into trend, seasonal, and residual components. Trend and periodic sub-series are related to systematic physical patterns suitable for knowledge-based simulation, while random sub-series represent stochastic factors that cannot be directly modeled due to implicit factors or lack of information. The concept of universality is also discussed which reflects the degree of usage frequency of different variables in the Building Performance Simulation (BPS) community. Variables with higher universality have better acceptance, more methodologies or validation, as well as better data availability. Overall this hybrid model approach provides a way to integrate domain knowledge via building simulation with data driven methods leading to improved prediction accuracy. The source code and data used in this study are available on GitHub for reference purposes. Moving forward future work could focus on addressing limitations identified during case study as well as further refining hybrid model approach.