TIPS FOR WEATHER NORMALIZING AND FORECASTING HOURLY ELECTRIC LOADS 

• Software: Hourly loads depend on temperatures that have occurred over the last few days. Therefore, lagged temperatures are important explanatory variables and econometric software packages are the most expedient for modeling time series with lagged terms. We also find a programmable (versus interactive) software package to be most useful because the program documents the steps performed, and incremental changes can be quickly applied to the whole modeling process. For this application we favor RATS, Regression Analysis of Time Series, for estimating the model coefficients. The results can be programmed into a variety of software packages to run simulations.
• Model Specification: We always specify a nonlinear equation for hourly loads because a linear equation is too restrictive. We estimate coefficients separately for each hour of the day. Neural networks are useless in these applications.
• Weekday: We estimate separate sets of coefficients for weekdays, Saturdays and Sundays or 72 sets in total (3 day types x 24 hours).
• Daylight: We include a term for daylight intensity based only on the time of year (not using cloud cover).
• Climate Variables: Temperatures are the most important climate variable to explain heating and cooling loads. We have also found wind speed to be useful in explaining heating loads, and humidity useful for explaining cooling loads, although neither of these is essential for a good model fit.
• Climate Variable Frequency: We have tested both hourly data and combinations of the daily high and low temperatures to explain hourly electric loads. If a weighted average of the daily high and low temperature is used, and the weight is a coefficient to be estimated separately for each hour, the model fit using daily data is almost as good as using hourly temperature data. Daily data is easier to collect and maintain.
• Previous Day’s Climate: We have tested both a weighted average of the current and previous day’s temperatures as well as geometrically declining weights over time and the latter performs only slightly better but is more difficult to program. In both cases, the weights were coefficients estimated with nonlinear least squares.
• Model Diagnostics: It is very important to examine the model errors or residuals for patterns over time. Patterns that vary throughout the year can be caused by very large customers that curtail their load for various reasons, or by holiday or school schedules. Binary variables can be used in nonlinear specifications to coincide with these schedules.
• Large Customers: Poor model fit for system hourly loads can be caused by very large customers that curtail their load during certain times of the year. In this case, it is best to obtain the hourly loads for these customers and subtract them from the system load and add them back after weather normalizing the remainder.
• Weather Normalization: Weather normalization of hourly loads is a complex problem because the probability distribution of weather for a whole year is so complex. Model simulations using historical weather data is one solution because the distribution does not need to be determined. Rather than using the actual weekday associated with the historical weather, we vary this by moving the day of week pattern forward one day at a time, seven times total. The resulting simulations are then converted to load duration curves, these curves averaged for each hour, and the resulting curved resorted using the last year’s data. We believe we were the first to use this approach, which is now becoming standard practice.
• Other Issues too complex to summarize here are:
• The advantages of estimating model coefficients with one year vs several years of data (accuracy at weather extremes vs calibration to recent trends).
• Averaging monthly vs annual load duration curves (the highest expected monthly peak is less than expected annual peak).