The biggest problem with the current method of weather normalization is that it uses a linear modeling approach. In the last post, I talked about the Degree Day method of weather summary, and how it requires a linear model. However, when you use hourly data, linear models are not the only option.
See the graph below, which shows a linear model applied to a graph of temperature and load.
Demand does not respond to weather in a linear fashion, so this approach will never work correctly. A square peg does not fit into a round hole, and a straight line can’t follow a curve. The best you can hope for with a linear model is that the errors will roughly cancel out. In the long run errors will likely balance, but it takes hundreds of observations for it to be achieved.
Running a linear model is essentially like closing your eyes and hoping for the best—not a very analytical approach. A good model will accurately describe weather response across the entire temperature spectrum—even at the extreme ends. See the model below, which was created by Bellweather using the same data as the chart at the top.
The Bellweather model is optimized to work across all temperatures. For the given data set, it would not be possible to have a model with a better fit. However, more can be done to improve accuracy.
In a subsequent post I’ll talk about subsetting data to maximize overall model quality.