Here is a review of the most common statistical post-processing methods used in renewable energy forecasting. The main methods discussed here are: machine learning, model output statistics, Kalman filters, regression models, historical analogs and a process of post-processing depending on weather typing.
I attended EMS2018 last week in Budapest, where I presented a talk on my on-going work with Conor and Emily Gleeson from Met Éireann. The talk was titled “An Evaluation of Integrated Cloud Condensate in the HARMONIE-AROME NWP Model” as part of the session on cloud-aerosol-radiation interactions.
I used MÉRA for the data and examined the vertically integrated cloud water (ICW) and the vertically integrated cloud ice (ICI). I investigated the model spin-up and found that for both variables the model requires time to spin-up and therefore for the remainder of my project I used the forecast with a long lead time for the hours between 01 UTC and 09 UTC.
I also compared the MÉRA data to a satellite dataset – KMNI MSGCPP Condensed Water Path (CWP; CWP=ICW + ICI). There is generally good agreement between the satellite and MÉRA CWP. However, in summer the error is large and MÉRA tends to produce too much CWP.
I also looked at some individual case studies in which initial results suggest the error might be related to cloud type. Initial analysis suggests fronts may overestimate CWP and CWP may be underestimated for convection.
Published in the Environmental Research Letters journal, in 2017.
This article, produced by the UK’s Met Office Hadley Centre, is a follow up on previous recent research conducted by scientists from the same centre. That previous research has been able to reproduce the North Atlantic Oscillation (NAO) teleconnection signal months ahead of Winter using their seasonal forecast system, the Global Seasonal forecast System 5 (GloSea5) – this system incorporates and represents elements and subsystems of the climate system (ocean, stratosphere, sea ice) which have been shown to mediate predictable teleconnections such as the NAO.
With this forecast system shown to be useful in predicting the NAO months ahead; and with previous research pieces showing significant and relevant impacts of the NAO on components of climate over Britain, such as wind speed and temperature (which can be linked to electricity production and demand variability); this article’s objective was to assess how the GloSea5 forecast system can be useful to predict electricity demand over Britain.
So, after validating GloSea5’s forecast hindcasts’ accuracy against “observed” (ERA-interim) conditions in the UK (average), for Sea Level Pressure (SLP), wind speed and air temperature; the authors check the reliability of the GloSea5 wind and temperature hindcasts for most of Europe (Maps of reliability), where they confirm that indeed Britain (and Ireland; and some parts of Western Europe) and surrounding areas fall on the category of reliable or underconfident seasonal forecasts (regarding the usage of the GloSea5 system).
Therefore, the article proceeds to its centre objective, which is testing the forecast system, through it’s NAO predicted index, against electricity demand using real data from Britain. This NAO-electricity demand relationship is first expressed using the real-observed NAO index. We can see both in the figure bellow.
Figure 1: Time series (left) and scatter plot (right) of the standardised observed winter mean NAO index and Britain’s electricity demand. Both use the same real electricity demand data. The top row plots use the real-observed NAO index. The bottom plots use the forecasted NAO index, from the GloSea5 hindcast. Note: low-frequency variability of electricity demand was removed from data, as it is associated with socio-economical change.
As we can see in the Figure, there is a strong negative relationship between Britain’s Winter (DJF) electricity demand and the NAO, expressed using both the real-observed NAO index (top plots) and using the 1 month-ahead predicted NAO index (bottom plots). This is the first time that seasonal forecasts of the weather dependent component of Britain’s Winter electricity demand have been demonstrated, offering multiple potential benefits to the energy industry.
Relevancy for our research:
- Having this electricity demand-Winter NAO relationship demonstrated, one can expect to have similar skill for similar applications in other Northern-Western European countries, specialy Ireland (due to its proximity to the UK), which is of our interest. This relationship has multiple benefits, specialy for the energy industry, and can be very useful if the NAO-associated Winter average conditions (wind, temperature) or the NAO index itself are sucessfully forecasted months ahead, and the relevant information integrated in the energy grid management.
- Should we apply the same approach using all-Ireland aggregated electricity demand? Maybe with more spatially resolved demand and production data, as NAO signal may vary significantly accross Ireland.
- We need to engage with ESIPP researchers working with demand data in Ireland (Dr. Lucy Cradden) to assess how viable it is to perform this analysis for Ireland.
To quantify the skill of forecast systems or reanalysis models, deterministic skill scores are used. This report describes some of the most common deterministic skill scores used in meteorology and energy/power forecasting and under which circumstances they are most suitable.
These include, among others: Mean error, root mean square error (RMSE) and mean absolute error.