1610-12-edifes-talk.pdf presentation slides
- 1. EDIFES: Energy Diagnostics Investigator for Efficiency Savings Alexis R. Abramson, Ph.D. Professor, Mechanical and Aerospace Engineering and Director, GLEI Roger H. French, Ph.D. Professor, Materials Science and Engineering Chris Littman Operations Manager, SDLE Research Center Mohammad Hossain, Ethan Pickering, Jack Mousseau, Rachel Swanson, Arash Khalilnejad, Preethi Kumar Case School of Engineering students Youngchoon Park, Steve Vitullo, Pravin Duraisingh, Johnson Controls, Inc. 1
- 2. EDIFES (Energy Diagnostics Investigator for Efficiency Savings) The problem: expensive, complex, high-risk, untrustworthy diagnostics and recommendations The market need and EDIFES capabilities: 1.Virtual energy diagnoses 2.Energy efficiency recommendations and ROI 3.Measurement and verification of energy efficiency measures (post-implementation) 4.Continuous commissioning 2
- 3. Data and code management process EDIFES code development: - Core component is the ‘edifes’ R1 package - Private code repository on bitbucket2 - For easy collaboration - Every function is documented - Majority of functions are covered by unit tests - New software versions published on plan Data Management: Reference: 1. https://www.r-project.org/ 2. https://bitbucket.org/ 3
- 4. Data and code management process (visually) 4 Git Repository Code Activity Time-series Movie Of Commits Showing Evolution of the codebase Visualization made using ● Software version control ● Gource
- 5. Data Preprocessing: Anomaly detection Reference:1. https://github.com/twitter/AnomalyDetection 2. http://www.itl.nist.gov/div898/handbook/eda/section3/eda35h3.htm Elec.consumption(KWh) Method: Annual Hybrid ESD1 (ESD= Extreme Studentized Deviate) Steps: ● Analyze data and break out into Trend+Daily Pattern+Residual bands ● Apply ESD on residual values2 to assess anomalies 5
- 6. “Good” Building data: Characteristics 6 Morning rise Evening fall Lower Weekend Consumption Additional Systems Daytime
- 8. Automating “good” building finding - Uses a semi-supervised time series clustering algorithm - Differentiates between “good” and “bad” buildings - m&m analogy Tell the computer what qualifies as a green m&m and as a red m&m (training data) Computer “learns” and places the new m&m into the appropriate color group (clustering) m m m 8
- 9. EDA example: correlations, pair plots 0.60 0.24 -0.41 -0.05 -0.10 0.07 -0.05 -0.51 0.56 0.85 0.04 0.39 0.12 0.81 0.09 ● Energy and Outdoor Temperature is strongly correlated, ● Relative Humidity shows negative correlation with Energy Temp=Temperature Dewp= Dewpoint RH = Relative Humidity PCP = Precipitation Green = Positive cor. Red = Negative cor. Correlation between Energy and Weather related parameters 9 https://cran.r-project.org/web/packages/psych/psych.pdf
- 10. Building markers: Identify Building Characteristics lighting scheduling identified using derivative analysis + standard deviation Light turn on Light turn off 10
- 11. Building operations signature marker Daily Operational Signature ● Captures many features of the building that can be compared with other buildings ● Captures features that correspond to distinct characteristics ● Provides a unique method for clustering buildings 11
- 12. Predictive modeling Traditional Time series forecasting ● Seasonal Arima Models ● suffers from autocorrelation Supervised prediction models: ● Tree based models: ● RandomForest ● gradient boosting ● Support Vector regression ● Artificial neural network ● Deep learning ● Convolutional network Ensemble Modeling: ● Linear weight averaged assembly ● Apply meta-learners 12
- 13. Prediction Example ● Random Forest method exhibits higher accuracy compared to data. ● Challenges: ○ Night time ○ Unusual events ● Next steps: ensemble of multiple algorithms 13
- 14. EDIFES Engine 14 ● Current version v0.4.0 ● Hosted in github for internal use
- 15. Acknowledgments 15 Research Faculty & Associates •Tim Peshek, Laura Bruckman Graduate Students •Yang Hu, Nick Wheeler •Mohammad Hossain, Yingfang Ma, Arash K. •Devin Gordon, Donghui Li, Yu Wang Undergraduates • Jack Mousseau, Rachel Swanson •Justin Fada, Davis Zabiyaka SDLE Staff: Chris Littman, Rich Tomazin
- 16. 16
- 17. Technology-to-Market progress ➔ Completed our first version of the Tech-2-Market Plan ➔ Hired a Commercialization Associate (Preethi Kumar, MEM Student) ➔ Continuing to conduct market research and familiarize ourselves with the competitive landscape ➔ Established Industry advisory board: ◆ Ali Ahmed (Cisco) ◆ Joe Jankowski (CIO, CWRU) ◆ Rem Harris (JumpStart) 17
- 18. Technology-to-Market progress Conversations with potential customers, collaborators, & competitors thus far include: - Accenture - Ecova Retroficiency - Energy Savvy - SmartWatt Energy Inc. - Pecan Street - Forest City Realty Trust - Progressive Insurance (Beta site option) - Wells Fargo - Welltower 18
- 19. Potential Customer Segments Real Estate Managers Chain Businesses Energy Service Companies
- 20. Market Analysis 107 on Department of Energy Qualified List of ESCOs $5 billion investment in energy efficiency annually
- 22. Other accomplishments and next steps Ethan masters degree (August 2016) 2 papers will be submitted in September 2016 Next steps: Q2: M1.2.2. Select subset of 10+ buildings identified for marker validation. Will begin working with JCI to identify buildings for verification. Q2: M2.1.1. Summary of findings using EDA applied to 40+ buildings, 4+ building equipment. - Ongoing will begin examining population studies as well. Q3: M1.2.3. Select 2+ buildings outside portfolio for alpha testing of same climate zone, building type and sq ft within 20%. In discussion with Progressive Insurance for alpha and beta testing. Q3: M3.1.1. Library of 10+ building markers developed in Cycle I. Presented in summary of findings inclusive of plots that demonstrate applicability. - At ~40 markers, but will continue to identify with a focus on disaggregation, prediction and population studies. 22
- 23. Data cleaning and assembly for ingestion Weather (NOAA) - ingested Solar irradiation, weather (GIS) - partially ingested (ingestion of certain zip codes under negotiation) Currently 65 “good” datasets ingested that meet requirements: ● 50,000 sq ft or less ● 6 climate zones ● 2 years, 15 minute interval ● Wide range of facility types such as ○ Schools, labs, and office buildings ● Less than 5% anomalies (Twitter) ● “Good” shape - clustering algorithm ● 10 groups of 2 or more buildings Challenges to finding “good” buildings: ● Different meters record differently - need to understand meter hierarchy, designations ● Meter data problems: inconsistent time intervals, duplicate data, invalid timestamps, > 15 minute intervals Mitigation: ● Improved communication with JCI to gather more information about data ● Created script for automated “good” building finding Location of the ingested buildings 23 Reference: https://github.com/twitter/AnomalyDetection
- 24. Summary of Ingested “Good” Buildings Facilities by state Different types of facilities 24
- 25. Summer Break Winter Break This analysis reveals that the building has a high likelihood of being a school - LATER VALIDATED Example from EDA: Building type identification 25
- 26. Example from EDA: Other data issues make select datasets not appropriate for EDIFES analysis (at this time) Some issues with datasets: - Duplicate data - Missing data - Data recorded past the present date - > 15 minute interval Duplicate Missing Data January 2017! 26