Design of a Weather-Normalization Forecasting Model · Problem Statement Climate changes have caused NOVEC to question whether the current weather-normalization methodology can be

Design of a Weather-Normalization Forecasting Model

Final Briefing09 May 2014

Abram GrossJedidiah ShireyYafeng PengOR-699

Sponsor:Northern Virginia Electric Cooperative

Agenda Background

Problem Statement

Sponsor, Purpose, Objective

Key Definitions

Scope

Assumptions

Limitations

Methodology

Excursions

Recommendations

2

Background Northern Virginia Electric Cooperative (NOVEC) is an energy

distributor serving parts of 6 Northern Virginia counties. Mandated to meet all customer energy requests.

Service provided by energy market purchases from regional providers: 1) Bulk purchases via contracts 5 years in advance.

2) Spot purchases up to one day prior to delivery.

NOVEC conducts analysis to inform energy purchases. Forecasts of energy demand over a 30-year horizon.

Decomposes forecasted demand into a base load and seasonal load.

Base-load: average demand from customer base.

Seasonal-load: weather’s impact on base consumption.

Predicted energy consumption determines bulk purchase quantity. Large error leads to increased costs.

Recently requested by Sales Department to improve forecast accuracy.

3

Problem Statement Climate changes have caused NOVEC to question whether the current

weather-normalization methodology can be improved.

NOVEC requests a review of weather-normalization methods that account for changing weather trends:

Evaluate new methodologies to compare to existing procedure.

Improve estimates of customer base trends.

NOVEC needs a model to accurately predict total energy consumption.

Forecast over 30 year horizon; energy demand reported monthly.

Emphasis on first 5 years to align with bulk purchase contracts.

Normalizing for weather to quantify customer growth, including impact of economic factors.

4

Sponsor, Purpose, Objectives Sponsor

NOVEC.

Purpose

Provide a candidate methodology to normalize weather impact on monthly energy purchases.

Objectives:

1) Assess historic relationships between economic, weather, and power data.

2) Develop a forecast model to test normalization methodologies.

3) Test weather-normalization methodologies; recommend one for implementation based on accuracy and robustness.

5

Key Definitions Heating Degree-Day (HDD): Measure used to indicate amount of energy need to heat during cold

weather.

Cooling Degree-Day (CDD): Measure used to indicate amount of energy need to cool during hot weather.

6

T b

30

40

50

60

70

80

Actual_Temperature Upper_Bound Lower_Bound

CDD

HDD30

40

50

60

70

80

Actual_Temperature Upper_Bound Lower_Bound

CDD

HDD

t

T b

Neutral Zone

Scope Data:

NOVEC monthly energy purchases data since 1983.

Dulles Airport weather data since 1963.

Historic economic factors data since 1980s metro D.C.; state and county data not evaluated.

30 years of Moody’s Analytics forecasted economic factors.

Model:

Inputs: historic energy purchases, weather data, economic variables, and customer-base.

Output: Monthly predictions for energy consumption over a 30-year horizon.

Regression: characterize dynamics between parameters.

Weather-normalization: remove seasonal weather impacts on NOVEC’s load.

Forecast: facilitate testing of varied normalization methodologies.

Ensure synergy with NOVEC’s existing models (regression, weather-normalization, forecast).

7

Assumptions Neutral zone between HDD/CDD has insignificant impact

on energy consumption.

55 and 65 degrees are the lower and upper bounds.

Economic variables currently utilized provide proper indicators for gauging future power demand:

Employment: Total Non-Agricultural

Gross Metro Product: Total

Housing Completions: Total

Households

Employment (Household Survey): Total Employed

Employment (Household Survey): Unemployment Rate

Population: Total

8

Limitations Unable to develop deep understanding of NOVEC’s

current forecast model due to complexity and time constraints:

Hinder adopting into existing model.

Skew comparisons of forecast accuracy.

Economic regression model determines customer base; potential for inconsistent forecast comparisons of this and NOVEC’s current model output.

Baseline economic scenario only scenario evaluated.

9

Determine Historic Relationships

Moody’s Economic Report / NOVEC Data Moody’s Forecast

7 Economic Scenarios

Approach

10

2012-13 NOVEC

Data Test

1983 1984 … … 1990 1991 … ... 2011 2012 2013 2014 2015 2016 2017 … … 2040 2041

Model Data

Function of:1) # Customers2) Avg. Demand

f(Economics)

g(Temperature)

Base Load

Seasonal Load

Power Demand

Forecast

HDD & CDD

Holt-Winters;ARIMA;BAT

R Model

Excel Model

Methodology

11

Data Sources Data Validation Conduct Analysis

Moody’s Data

NOVEC Data

TemperatureData

ComputeHDD/CDD

Clean, Compute,

Format Combined Linear

Method

Ratio Method

Forecast

Data processing included linear interpolation for data gaps, disaggregating quarterly economic data, as well as aggregating hourly weather data, up to monthly resolution.

Primary methodologies: Split Regression Model, Combined Regression Model, and Ratio Model

Upper/Lower

Bound Temp

Start/End Year

Economic

Scenario

Start/End Year

Economic

Variables

Economic

Data Region

Split LinearMethod

Combined Linear Regression Model

12

Intuition: Usage should be a function of economic contributions and

weather contributions.

Accuracy of Linear Regression Model Adjusted R-square = 0.925

13

Additional Forecasting Methods Split Regression

total load = residential load + non-residential load

residential load = # of residential customer * avg residential

non-residential load = # of non-residential customer * avg non-residential

Customer Ratio Method

total load = residential load + non-residential load

residential load = # of residential customer * avg residential

non-residential load = residential load * ratio

14

Estimate of Customer Base

15

Customers are categorized as either residential or non-residential.

>99% adjusted R square based on the 7 econometric variables from 1990-2011.

Linear regression model provides sufficient accuracy for predicting customer base.

Estimate Average Customer Usage Tested linear regression on 3 similar models

7 Econometric Variables + HDD + CDD

7 Econometric Variables only

HDD + CDD only

16

Customer Usage Ratio Method Ratio of average usage between non-residential vs. residential

Actual Data

Trend

Seasonality

Random Error

17

HDD/CDD Forecasting Methods Holt-Winters Method

ARIMA method

Not good as correlogram violates control limit

BAT Method

Basically a superset of Holt-Winters. No improvement over Holt-Winters Method.

18

Holt-Winter Method for HDD/CDD Forecasting

19

Accuracy of Holt-Winters Method Correlogram Plot Residual Plot

20

Good fit should have

1 or 2 spikes outside of the dotted control other than x=0

Good fit should have

~0 error mean

& almost constant variance

Modeling Excursions Tested sensitivity using different domains of time:

Regression models inform forecast output.

All historic economic variables are actual records:

Same for all scenarios.

Serve as starting point for forecast.

21

0

50

100

150

200

250

300

350

400

450

Feb

-08

Jul-

09

No

v-1

0

Ap

r-1

2

Au

g-1

3

Dec

-14

May

-16

Tota

l Lo

ad (

GW

H)

2000-2008 Model Run - Forecast 2009-2011

Combined Model Split Trends Split Ratio Actual Load

0

50

100

150

200

250

300

350

400

450

Feb

-08

Jul-

09

No

v-1

0

Ap

r-1

2

Au

g-1

3

Dec

-14

May

-16

Tota

l Lo

ad (

GW

H)



Model Selection Merit given to the balance between:

Bulk-energy error for first five years of forecast.

Robust to changes; back-tested on historic data.

22

0

50

100

150

200

250

300

350

400

450

Jan

-06

May

-07

Sep

-08

Feb

-10

Jun

-11

Tota

l Lo

ad (

GW

H)


Combined Model Split Trends Split Ratio Actual Load0

50

100

150

200

250

300

350

400

450

Jan

-06

May

-07

Sep

-08

Feb

-10

Jun

-11

Tota

l Lo

ad (

GW

H)



0

100

200

300

400

500

600

700

800

Jan-

09

Jan-

14

Jan-

19

Jan-

24

Jan-

29

Jan-

34

Tota

l Loa

d (G

WH)

Forecast Results - 30 Year Horizon


0

50

100

150

200

250

300

350

400

450

Jan

-06

May

-07

Sep

-08

Feb

-10

Jun

-11

Tota

l Lo

ad (

GW

H)



CUMULATIVE LOAD (kWH)

ENERGY (kWH) 1 2 3 4 5

Combined Model 3.18E+09 6.57E+09 1.01E+10 1.37E+10 1.75E+10

Split Models 2.89E+09 5.91E+09 9.02E+09 1.22E+10 1.56E+10

Ratio - Split Models 3.09E+09 6.33E+09 9.65E+09 1.31E+10 1.67E+10

Actual 3.00E+09 6.23E+09 9.46E+09 1.27E+10 1.63E+10

ERROR (%) 1 2 3 4 5

Combined Model 6% 5% 6% 8% 8%

Split Models -4% -5% -5% -4% -4%

Ratio - Split Models 3% 2% 2% 3% 2%

FORECAST HORIZON

Using 1990-2005 data for regression modeling:

Informs monthly forecasts for 30 years.

Cumulative relative error assessed.

Robust to changes; back-tested on historic data.

Modeling Excursion - Results

23

Candidate model selection: split-model with seasonal trend forecast.

YEARLY %-ERROR in CUMULATIVE LOAD

Modeling Approach1 2 3 4 5

Combined Model 10% 12% 14% 17% 27%

Split Models -2% 0% 1% 2% 3%


Combined Model 10% 12% 14% 17% 27%

Split Models -2% 0% 1% 2% 3%



Split Models -4% -7% -9% -11% -13%

Ratio - Split Models 4% 1% -1% -3% -6%


Split Models -4% -5% -5% -4% -4%



Split Models 3% 2% 0% 0% 0%


Combined Model 3.7E+09 7.5E+09 1.1E+10 1.4E+10 1.9E+10

Split Models 3.5E+09 7.0E+09 1.1E+10 1.3E+10 1.8E+10

Ratio - Split Models 3.8E+09 7.5E+09 1.1E+10 1.4E+10 1.9E+10

19

90

-20

05

19

90

-20

08

19

90

-20

11

(Dat

a

Do

mai

n)

FORECAST HORIZON

19

90

-19

95

19

90

-20

00

19

90

-19

95

Recommendations We recommend that NOVEC use the “Split-Trend Model”:

Mirrors methodology currently employed.

Potential for improvements with quantifying trends to seasonal loads (Holt-Winters method).

Supplement their current weather-normalization forecasting model.

Suggestions for future analysis: Determine best set of economic variables to predict future customer

base; Population size, and thus tested methodologies, are sensitive to this parameter.

Best subset is likely to change over time.

Determine best set of neutral zone boundaries to compute CDD/HDD.

Perform further analysis on the ratio method and the ARIMA HDD/CDD forecasting with preliminary de-trending.

24

Documents

Design of a Weather-Normalization Forecasting Model · Problem Statement Climate changes have caused NOVEC to question whether the current weather-normalization methodology can be