Seasonal forecasts: can the prediction ensembles be collapsed into a reliable single-value prediction?
If seasonal forecasts are still feasible is because atmospheric variability on these time scales is modulated by slowly-varying boundary conditions. However, these are not the kind of single-value predictions we might have in mind since predicting the hourly changes beyond a few days is simply not possible (due to the atmosphere’s chaotic behaviour). Thus, in this scenario, it would not make any sense to trust a deterministic forecast.
Instead, we are talking about monthly (or weekly) means predicted some months (weeks) ago and probabilistic in nature through the use of ‘ensembles’ of forecasts. These ’ensembles’ conform a ’ramification’ of predictions conceivable as a predicted probability distribution and they were principally developed so as to estimate forecast uncertainty due to unavoidable errors in the initial conditions.
The observed atmospheric behaviour in the following months should be described by a subgroup of the predicted ensembles. And, actually, this is the academic standard to address this kind of problem.
Nevertheless, these ensembles may offer a spectrum of possibilities too disperse (especially in windows of bad predictability) for the wind industry when it comes to taking any decision or conclusion based on them. This presents the dilemma of seasonal predictions uptake: which subgroup of ensembles should I trust or how can the distribution be interpreted? Can the ensembles be collapsed into a single value? Is it skilful enough to use the ensembles’ mean?
To answer this, a case study was performed in which several machine learning techniques were leveraged to collapse the ensembles into single-value predictions, and compared against simpler techniques (such as the ensembles’ mean) as well as the climatology baseline. Let’s bear in mind this baseline consists of systematically assuming next month’s wind speed will be equal to the monthly mean over the last 30 years).
Specifically, supervised learning was used to solve what can be conceived as a regression problem: seasonal model ensembles (e.g. from ECMWF, NCEP…), at different height levels, were faced as ‘features’ against the ground-truth (the observed wind speed anomaly), our ‘labels’.
This study (presented at WindTech’s 2022 edition) was validated using several sites around the world and resulted in the following conclusions:
– It is possible to find at least one artificial intelligence method to obtain skilful single-value predictions of the wind speed anomaly (for a given site and lead month).
– This method is skilful in the sense that outperforms simpler strategies such as the ensembles’ mean and baselines as the climatology.
– The best method found may depend on the site and lead month but more than half of the times was DNNs (deep learning method: dense neural networks). This yielded a 22.1% of enhancement: meaning the predictions were more skilful than climatology 22.1% more than just using the ensembles’ mean as a prediction system (which already is often slightly better than climatology).
– Other ML methods runner-ups were: Cubist (with 18.94% of enhancement), Random Forest (17.68%), ElasticNet (16.16%), CATBoost (16.9%).
All the above is not but another step towards the incorporation and use of seasonal forecasts for the wind industry. However, and despite there is still a lot of room for improvement, we must always remind these forecasts’ usability and correct guidelines differ from short-term forecasts (now-cast).
After all, next month can bring more surprises than tomorrow!