# Support Vector Machine in Prediction of Building Energy Demand Using Pseudo Dynamic Approach

**Authors:** Subodh Paudel, Phuong H. Nguyen, Wil L. Kling, Mohamed Elmitri, Bruno Lacarrière, Olivier Le Corre  
**arXiv:** [1507.05019](https://arxiv.org/abs/1507.05019)  
**Date:** 2015-07-17  
**Summary page:** https://www.summarizepaper.com/en/paper/1507.05019v1/

## Key points

- Growing concern about building energy consumption and efforts to improve energy management
- Prediction of energy consumption allows for optimal operating strategies in a building's energy management system (BEMS)
- Proposal of a prediction model for building energy consumption using support vector machine (SVM)
- Use of relevant days data selection method based on Dynamic Time Warping to train the SVM model
- Incorporation of thermal inertia through a pseudo dynamic model considering energy consumption effects and occupancy profiles
- Application of the proposed methods to case studies at Ecole des Mines de Nantes France Office building
- Support vector machine with relevant data selection achieves high accuracy compared to whole data training
- Significant reduction in computational time with approximately 8-minute training time for relevant data selection compared to 31 hours for weekends and 116 hours for working days in whole data training
- Enables realistic control implementation for online systems
- Contributes to improving building energy management by presenting an effective prediction model that enhances accuracy while reducing computational requirements

## Summary

In recent years, there has been a growing concern about building energy consumption and efforts have been made to improve energy management in buildings. One essential aspect of this is the prediction of energy consumption, which allows operators to develop optimal operating strategies that can be integrated into a building's energy management system (BEMS). This paper proposes a prediction model for building energy consumption using support vector machine (SVM), a data-driven model that is sensitive to the selection of training data. To address this sensitivity, the authors employ a relevant days data selection method based on Dynamic Time Warping to train the SVM model. Additionally, they incorporate thermal inertia by applying a pseudo dynamic model that considers information on the transition of energy consumption effects and occupancy profiles. The proposed relevant days data selection method and whole training data model are applied to case studies conducted at Ecole des Mines de Nantes France Office building. The results demonstrate that the support vector machine based on relevant data selection method achieves high accuracy in predicting building energy consumption compared to using whole data training. Moreover, this approach requires significantly less computational time with an approximately 8-minute training time as opposed to 31 hours for weekends and 116 hours for working days required by whole data training. This also enables realistic control implementation for online systems. Overall, this research contributes to improving building energy management by presenting an effective prediction model for building energy consumption. The incorporation of relevant days data selection and pseudo dynamic modeling enhances accuracy while reducing computational requirements, making it a practical solution for optimizing energy usage in buildings.