KISS, or "keep it simple, stupid", refers to a design principle, which tells us to pursue simplicity in design and avoid unnecessary complexity. A similar saying is "the principle of parsimony", which is probably the most frequent term I heard from my PhD advisor during my graduate study. When developing the forecasts, a forecaster does act like a designer, designing a model with scientific tools and artistic creativity.

Let's look at the two figures below. Both figures show 6 years of annual peak loads for a small area. In each case, we fit the 6 observations using a regression model, and then derive a one step ahead forecast. A simple linear regression model is used in Figure 1, which results in an R-square value of 0.7955. A polynomial regression model is used in Figure 2, which results in an R-square value of 1. Although we got a "perfect" R-square in Figure 2 with a 5th ordered polynomial regression model, its forecast does not seem to be as reasonable as the one in Figure 1.

Other than the misleading R-square, what conclusion shall we draw here? Is the simple linear regression model more useful than the polynomial regression models? Is this how the KISS principle works?

Let's look at Figure 3 below, which is a scatter plot of hourly loads and temperatures for a year. Will a simple linear regression model do a better job than a 2nd or 3rd order polynomial model?

I don't think so.

Does it mean that the KISS principle is not applicable here?

You may say, in Figures 1 and 2, the trend looks like linear, while in Figure 3, the relationship seems to be nonlinear.

In simple examples like these two, it may be a good enough explanation. Unfortunately, just "eye-balling" is not enough to guide a rigorous practice or justify a defensible forecast.

In the next talk of Tao's Energy Forecasting Webinar Series, we will discuss the meaning of KISS in load forecasting, how to apply this principle to build elegant models, and some DOs and DONTs. If you are interested, feel free to register through the webinar page.

*What does "KISS" mean in load forecasting?*Let's look at the two figures below. Both figures show 6 years of annual peak loads for a small area. In each case, we fit the 6 observations using a regression model, and then derive a one step ahead forecast. A simple linear regression model is used in Figure 1, which results in an R-square value of 0.7955. A polynomial regression model is used in Figure 2, which results in an R-square value of 1. Although we got a "perfect" R-square in Figure 2 with a 5th ordered polynomial regression model, its forecast does not seem to be as reasonable as the one in Figure 1.

Figure 1. A simple linear regression model for annual peak forecasting.

Figure 2. A polynomial regression model for annual peak forecasting.

Other than the misleading R-square, what conclusion shall we draw here? Is the simple linear regression model more useful than the polynomial regression models? Is this how the KISS principle works?

Let's look at Figure 3 below, which is a scatter plot of hourly loads and temperatures for a year. Will a simple linear regression model do a better job than a 2nd or 3rd order polynomial model?

I don't think so.

Figure 3. A scatter plot for load-temperature relationship.

You may say, in Figures 1 and 2, the trend looks like linear, while in Figure 3, the relationship seems to be nonlinear.

In simple examples like these two, it may be a good enough explanation. Unfortunately, just "eye-balling" is not enough to guide a rigorous practice or justify a defensible forecast.

In the next talk of Tao's Energy Forecasting Webinar Series, we will discuss the meaning of KISS in load forecasting, how to apply this principle to build elegant models, and some DOs and DONTs. If you are interested, feel free to register through the webinar page.

## No comments:

## Post a Comment

Note that you may link to your LinkedIn profile if you choose Name/URL option.