Mnthly

TIPS FOR WEATHER NORMALIZING AND FORECASTING MONTHLY ENERGY SALES

Monthly Electric or Natural Gas Sales

Software: When a large number of tariffs or jurisdictions is involved, it is most efficient to use a programmable modeling software package so that data preparation and modeling can be automated. We favor SAS for this application.

The Dependent Variable: For residential and small commercial tariffs, we model sales per customer. For the other tariffs we model sales.

Degree Days: Heating and cooling degree days are most often used to explain monthly variations in heating and cooling loads. In the US, the government publishes monthly heating and cooling degree days computed with a base temperature of 65 degrees Fahrenheit. However the calendar month is not aligned with the way in which sales are billed. Sales are typically read from meters on billing cycles throughout the month, one cycle is read on each workday. The sales that are billed during a month will therefore have occurred over a two-month period. Sales for the first cycle read each month will have actually occurred mostly over the previous month. Sales for the last cycle read each month will have occurred mostly during the current month. Therefore heating and cooling degree days must be computed separately for each billing cycle using the dates the meters were read and then weighted by the number of customers or the amount of sales on each cycle. Or alternatively, each billing cycle can be modeled separately. Several decades ago, it was standard practice to use a two month average of calendar month degree days, but this introduces a huge "errors in variables" problem. Degree days can be interacted with space heating and air conditioning saturations if these are available.

Degree Day Base Temperatures: While 65 degrees is used most often in the US, other base temperatures may provide a better model fit, particularly for commercial customers. Normal degree days are published for other base temperatures, but not as frequently or as timely as for 65 degrees.

Number of Weather Stations: More stations are better, everything else being equal. But the quality of data collected by different stations is not equal. In the US, NOAA collects and publishes weather statistics. Some weather stations are operated by NOAA or the National Weather Service whereas others are manned by volunteers. The latter tend to have less accurate and more missing data. With few exceptions, we favor using only "first class" weather stations for jurisdictions in the US. Data from these stations is easier to obtain and is available sooner. It is also edited in various ways.

Billing Errors: Accounting data often includes errors that are reversed in subsequent months. Billing errors for electricity and natural gas occur often, and the reversals often occur in a month after the error was made. If large errors are made, model fit based on this data can become poor and some explanatory variables statistically insignificant. This problem can be corrected by restating the billing statistics to reassign corrections to the month the error occurred so that the error and any corrections cancel out. This task is fairly easy to program in SAS even if there are several million customers and many years of data, but it requires access to the monthly billing records, which can be difficult to obtain. It also requires a field that specifies which month a correction is made for.

Number of Customers: Surprisingly, some billing systems do not provide good customer counts. Some systems will double count a premise if someone moves out of a residence and someone else moves in during the same month. If the billing records are accessible, it is often better to redo the customer counts to be sure a premise ID is counted only once per month.

Trends: We have included time trends as separate explanatory variables and also interacted a trend with heating and cooling degree days. Typically we find that residential gas heating loads per customer per heating degree day are declining over time whereas electric cooling loads per customer per cooling degree days are increasing. However, we do not extrapolate these trends too far into the future by first decreasing then stopping the growth of the trend variable. The point of this is not as much to forecast the trends as to prevent estimation problems from omitted variables.

Energy Prices: Price elasticities are difficult to measure in this type of model because of insufficient price variation within the data sample. Elasticities are most likely to be statistically significant in cross sectional time series samples, for example, one that includes dozens of small cooperatives each with different rates. In a very rich sample, it is possible to estimate price elasticities separately for heating and cooling loads. If appliance saturations are included, the elasticities will be mainly short-run. If not, a long lagged price term should be included such as one with geometrically declining lag coefficients. Appliance saturations are slow to adjust to prices, but utilization, such as adjusting thermostats, occurs quickly.

Model Diagnostics: It is very important to examine the model errors or residuals for patterns over time and outliers. A pattern indicates omitted variables and outliers usually indicate data errors. Even in a highly automated forecasting system, someone needs to inspect these plots.

Weather Trends: While the cause is debatable, there is a global warming trend. Many jurisdictions require that a standard 30 year average be used to normalize heating and cooling loads. For gas utilities, this causes revenues to be over estimated and rates are set too low. More recent averages will tend to catch up to the current trend, but any historical average will lag the trend. The best approach would be to model and forecast the trend.

Normal Degree Days: NOAA recomputes normal degree days at the end of each decade. The current set for 1971-2000 is no longer consistent with the way they compute actual degree days (has to do with rounding the average temperature).

Asymmetric Price Response: Does energy demand adjust more rapidly to rising prices than to falling prices? Some conservation devices are installed when prices rise but are not removed until they decay.