0
$\begingroup$

I am dealing with solar power output forecasting, but I am still new to forecasting. I would like to ask a few questions here:

  1. It seems that most literature forecasts solar irradiance, while solar power is not often directly forecasted. Why solar power times series is not of interest?

  2. Now I have a historical data set that contains the solar power outputs from 2001 to 2012, which covers the sixth day of each month with a time resolution of 5 minutes. In this case, at which "frequency" it should be specified?

  3. I wish to realize very-short-term forecasting. Suppose I would only use the ARIMA method. After training and testing, can I keep the obtained coefficients fixed for a typical day under study? OR they should be changed at each time interval (say, 5 minutes in this case).

$\endgroup$
  • 1
    $\begingroup$ Question 1 may require subject-matter (rather than statistical) expertise. Questions 2 and 3 could be posted separately to make the posts more concrete. $\endgroup$ – Richard Hardy Jun 7 at 17:49
1
$\begingroup$

I will answer your questions as you numbered them:

  1. This question really isn't meant for a statistics forum, but I would say that the power produced by a photovoltaic system is very dependent on the system's efficiency while actual solar irradiance allows you to predict power output given your individual system.
  2. Unless I am mistaken it seems your data was not taken at continuous intervals since it covers five-minute intervals, but not for continuous days (only one a month). For that reason, there is not a particular "frequency." For time series forecasting, however, this is not important. If you string together your data you will obtain artificial patterns but you could still forecast temperatures for every five minutes on the sixth day of any month. If you want to forecast this for any other day, however, you might be in trouble since your sampling is irregular and you could be overlooking cyclical effects (monthly or weekly). It may be ignorable, but I would suggest collecting/finding more data.
  3. For this type of cyclical data and "very-short" forecasting, I would not recommend using ARIMA (as the coeffients would cause problems). If you have a reasonable amount of computing power and want to be really accurate, I would suggest a neural network approach, such as LSTM. Otherwise, approximating what should (contextually) be a smooth curve with a low-number polynomial fit should work just fine.
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.