Model intraday forecast errors

The difference between a forecast and what happens (the 'outturn') is the forecast error.

Forecast errors are modelled on historical data

We model day-ahead forecast errors on the following variables using historical data.

Replicating observed patterns in forecast values of wind, solar and demand

Examining data from recent history, we find several trends.

The forecast will likely be closer to the outturn value (ie the error is small) on particularly windy or non-windy days. It is those days that sit in the middle where the error is higher. It's easier to say, 'This will be a very windy day' or 'This will be a still day' than 'There will be a small amount of wind in the morning and a bit more in the afternoon'. The same thing applies to solar.

In addition, if a forecast is lower than it should be in the morning, it will likely be low all day. A forecast will rarely bounce from higher than the outturn to lower than the outturn in each subsequent hour. There is a pattern to how a forecast's error behaves across time called 'autocorrelation'.

We want to preserve these effects when we project forward forecast errors between day-ahead and intraday markets.

Matching similar days in an error-weather year to our forecast-weather year

To apply forecast errors to our modeled demand, solar, and wind time series, we match similar days in a 'error-weather' year (2019) to our forecast weather year (2018).

Here's an overview of how it works for solar generation.

1. For each month in the fundamentals model, rank days from highest to lowest load factor. This is 2018, the weather year we used in the forecast. This graph shows January 2018: Jan 25th was the sunniest day, so it is labelled '1'.

2. Repeat the same process for 2019, the year we use for forecast errors. 'Day ahead' is the solar production forecast for the day before delivery, and 'Intraday' is the solar outturn value.

3. Match days in the fundamentals model to days in the forecast error data with the same rank. For example, the an error from the sunniest day in January 2019 is applied to outturn data for the sunniest day in January 2018. In this we use the error profile from the forecast error data set to forecast the solar production generated at day ahead for 2018, as well as the outturn or real) load factor.

Using the errors in the forecast of intraday prices

The load factor of the outturn production is used for the intraday price forecast.

The load factor of the outturn production profile plus the modelled error is used for the day ahead price forecast.