Time-series prediction of 15 min interval data

Question

I have 5 years 15 min interval of electricity demand time series with a datetime index and a target variable. Don't have any other data to use. I'm curious about your experiences. In general how far is it possible to accurately predict the 15 min electricity demand? Our client wants 1 year prediction, but as far as I know it is quite impossible. What is your experience?

Answer 1

It is absolutely possible to forecast out for one year, or for ten years, or for a hundred years. What is not possible is to achieve any desired accuracy. Achievable accuracy in general deteriorates with forecast horizon. If your client's accuracy expectations are realistic, everyone can be happy. If not, well, reality has a way of stubbornly resisting people's expectations.

One key aspect of forecasting electricity on this temporal granularity is the multiple-seasonalities involved. You have intra-daily patterns, but these differ between weekdays, with weekends being quite different from working days. There have been specialized methods proposed to deal with that that typically improve in terms of accuracy on simpler methods - but at great computational cost, which becomes especially relevant if you forecast not a single time series, but thousands.

Electricity demand typically is driven by causal drivers beyond these multiple seasonalities. If you know that the Super Bowl is coming up, you know that during the quarter breaks, electricity consumption will peak, because roughly 100 million Americans simultaneously get another beer from the fridge, causing the compressor to start up to cool the fridge back down. Take a look at my answer to How to know that your machine learning problem is hopeless?, especially the water consumption time series - similar patterns happen to electricity. It's important to understand these drivers. And of course you can predict the temperature (which drives electricity consumption for heating and air conditioning) more easily for tomorrow than for one day from now.

You may want to look at dedicated energy forecasting competitions to get an idea of what tools are effective, like the GEFcoms.

(Blatant self-promotion: I plan on presenting an analysis on multiple seasonalities forecasting at this year's ISF, and one of my datasets is actually about electricity demand. Our workshop on "Forecasting to meet demand" is not explicitly geared towards electricity forecasting, but may still be relevant.)