Setting Energy Efficiency Requirements Using Multivariate Regression

Transcription

1 Setting Energy Efficiency Requirements Using Multivariate Regression Matt Malinowski, ICF, Presenter Dan Baldewicz, ICF EEDAL 2017 Irvine, CA September 13, 2017

2 About ICF ICF (NASDAQ:ICFI) is a global consulting and technology services provider with more than 5,000 professionals focused on making big things possible for our clients. We are business analysts, policy specialists, technologists, researchers, digital strategists, social scientists and creatives. Since 1969, government and commercial clients have worked with ICF to overcome their toughest challenges on issues that matter profoundly to their success. Come engage with us at icf.com. Dan Baldewicz is an Energy Efficiency Analyst for ICF, working primarily on the ENERGY STAR Products Labeling Program. His analysis covers Lighting, Pumps, HVAC, and Consumer Electronics. Dan has a background in computational mathematics, and typically uses the R programming language. Matt Malinowski is a Technical Specialist with over 11 years of experience in engineering and public policy, currently advising government and utility clients on efficiency of electronics. He has advanced efficient products across all parts of the adoption curve (emerging tech, voluntary programs, and codes and standards) and managed projects from scoping to sunset. 2

3 How to Set Efficiency Requirements? 3

4 Requirement Types Energy 4

5 Requirement Types Energy 5

6 Requirement Types Energy 6

7 7

8 Multiple Classes and Slopes 8

9 Base + Adder Approach Can Simplify Base Functionality Top-mount Refrigerator-freezer Bottom-mount Refrigerator-freezer Side-by-side Refrigerator-freezer All-refrigerator Upright Freezer Chest Freezer Compact Refrigerator Compact Freezer Adder Automatic Defrost Through-the-door ice Built-in 9

10 Product Requirement System Summary Simple Requirement, Single Scope All TVs must use < X kwh/year Simple Requirement, Split Scope TVs with Area A must use < X kwh/year; TVs with Area B must use < Y kwh/year Complex Requirements, Split Scope TVs with Area A must use < (6.13*R) + (0.06*A) + 9 kwh/year; TVs with Area B must use < (6.13*R) + (0.21*A) 0.5 kwh/year Base + Adders Plasma TVs < X kwh/year; LED:LCD TVs < Y kwh/year; OLED TVs < Z kwh/year; 4k TVs may have additional T kwh per year; TVs with contrast ratios exceeding 1:100,000 may have additional U kwh/year 10

11 Product Requirement System Pros/Cons Simple Requirement, Single Scope Assigns power based on all product data Pros: Simple Cons Difficult to fine tune Unfairly penalizes largest percentage of products vs other systems Simple Requirement, Split Scope Assigns power based on group(s) of product data Pros: Fair to representative products Cons: Unfairly penalizes non-representative products, such as: value, high end, or specialized Complex Requirements, Split Scope Assigns power based on parameter(s) Pros: Fair to a majority of products and can capture some technology aspects Cons: Difficult to fine tune to optimal requirements, and can miss key technology considerations not covered by parameter(s) Base + Adders Assigns power based on technology Pros: Fair to highest percentage of products vs other systems Cons: Difficult to fine tune to optimal requirements. 11

12 Setting Adders 1. Look to past/similar requirements May not capture rapid changes in technology or nuances between products E.g., computer wireless card different from simpler edge product or router 2. Interviews with manufacturers 3. Testing a. Measuring the energy use of two related product models that differ only in the presence or absence of an adder; b. Measuring the energy use of the same product model with the adder turned on or off; and c. Measuring the energy use of the adder itself, either removed from the product or through dc measurements inside the product. 4. Then repeat across all relevant adders and prod types/sizes 12

13 Process is Time-Intensive Multivariate regression can help! Rather than examining the impacts of features on a single product, multivariate methods examine multiple products with different combination of features simultaneously. 13

14 Linear Regression Dependent variable is a function of Independent Variable Equations fit a line to data Only 1 variable considered at a time YY = ββ 0 + ββ 1 XX + εε Works well in simple cases, but does not account for correlations, or control for each variable Difficult to tie linear regressions on different variables together to get a complete picture 14

15 Multivariate Regression Dependent Variable is a function of Multiple Independent Variables Equations fit a plane (hyperplane) to data YY = ββ 0 + ββ 1 XX 1 + ββ 2 XX ββ PP XX PP + εε Multivariate Regression tools: Automatically account for correlated variables Determine the effect of each independent variable (controlling for the others) 15

16 Feature Selection Power (W) or Energy Consumption (kwh) is Dependent Variable Choose Independent Variables: Subject Matter Knowledge This feature is/is not power intensive Feature A and B are related to each other Modeling Rules Keep 1 of 2 correlated variables, etc. Existing Data/Analysis Feature C already has good power use estimate from other sources 16

17 Reducing number of Model Variables improves accuracy Stepwise process Models which favor more terms tend to perform suboptimally, by incorporating too much noise in the data to predict accurately (overfitting). Akaike Information Criterion (AIC): describes dependent variable variation accounted for by independent variables with an additional penalty per term in the model STB example (On Left, Cable) 17

18 Setting Base and Adders, Set Top Boxes 18

19 Setting Base and Adders, Set Top Boxes The intercept in a regression model is the base value for a product Additional Base types can be included as variables 19

20 Setting Base and Adders, Set Top Boxes The intercept in a regression model is the base value for a product Additional Base types can be included as variables Each adder is estimated in the output 20

21 Setting Base and Adders, Set Top Boxes The intercept in a regression model is the base value for a product Additional Base types can be included as variables Each adder is estimated in the output Overall model quality is included in the output: R 2, Adjusted R 2 21

22 Setting Base and Adders, Set Top Boxes ENERGY STAR V5.0 Set-top Box Additional Functionality TEC Allowances (Excerpt) Additional Functionality Allowance (kwh/year) CableCARD 15 CableCARD Max One Additional 15 Digital Video Recorder (DVR) 35 Multi-room 21 Multi-stream Cable/Satellite 18 Multi-stream IP 18 22

23 Summary of Process Steps 1. Develop model of power or energy as a function of base and adders 2. Remove some adders from consideration through: a) Research (subtract out energy) or b) Step-wise methods (ignore) 3. Multivariate analysis will output mean base and adder allowances 4. After adder allowances finalized, base allowance decreased manually to permit X% models to pass 23

24 Set Top Box Regression Results, IP Products 24

25 Future Work Ongoing ENERGY STAR Specifications: Small Networking Equipment (Routers, Access Points) Imaging Equipment (Multi Function Devices, Printers) Model Improvements (advanced modeling tools that go beyond multivariate regression): Partial Least Squares: Emphasizes connection between Dependent Variable (power) and Independent Variable (adders) LASSO: Automatic feature selection, useful in cases with many variables Additional Application Areas: Adaptable to other sectors/problems with modeling needs: econometrics/energy advisory, industrial energy efficiency, efficiency program evaluation and measure cost effectiveness estimation E.g., Using regression to pick efficiency measures for a particular site or benchmark buildings against similar ones 25

26 Acknowledgements The U.S. Environmental Protection Agency (EPA) funded the work described in this paper. 26

27 Q&A 27

28 Matt Malinowski, ICF Phone: (202) Dan Baldewicz, ICF Phone: (518) Thank You!