Our main forecast product is the monthly mean 2-meter temperature for the Nordic countries. For this forecast we take multiple numerical weather prediction models (NWP) and combine them by computing the mean of the models (more precisely, the mean of their ensemble means), in order to achieve more stability in the forecast. Most of these NWPs are available via the Copernicus web site, however, we also add the Norwegian Climate Prediction Model (NorCPM) to the mix, an NWP that is currently developed at the Bjerknes Center and not (yet) available via Copernicus.

The combination assigns the same weight to each model, but is allowed to assign an additional weight to climatology (i.e. the average temperature over the past 25 years for the corresponding month and location) that may differ from the other weights. This climatology-weight is learned from both the past performance of the NWPs and the climatological forecast.

This approach of learning weights for prediction from past data is one of the key elements in machine learning (ML) and AI methods, but the methodology we apply at the moment is pretty down-to-earth and barely qualifies as AI. Our predictions are for now based mostly on monthly means. This simplifies the modelling and is an important intermediate step for understanding dependencies and the skill of the single NWPs. In the long run, however, we hope to improve our seasonal forecast by incorporating daily data which would lead to a finer temporal resolution. Moving to daily data leaves us with vast datasets for training (25 years of daily training data on a spatial grid with 1000 grid points = 9 million data points), and it is usually this big data regime where the (deep) ML and AI methods outshine more classical statistical methods. We therefore believe that ML and AI methods can help us in the future to improve our predictions by allowing us to incorporate more data.

As to how useful the forecasts are: We measure the skill of our methods in terms of the mean square error (MSE) over the training period and compare to the climatological forecast, that always predicts the mean temperature for the corresponding location and month (averaged over more than 20 years). For June, our method improves on climatology by 5% (i.e. -5% MSE). For comparison, the best-performing NWP model available on Copernicus is the model of the UK MO which improves on climatology by 2%, just combining all NWP models on Copernicus (without allowing for a weight on climatology) would improve on climatology by 4%.

Even though 5% does maybe not sound like much, this is still remarkable given that it has been believed until not too long ago that, due to the chaotic behavior of the atmosphere, it is physically impossible to have better predictions than climatology at lead times of over 14 days. Note that our predictions are issued based on data available at the 15th of May, and therefore the June mean temperature forecast has a lead time between 16 and 36 days.

One of the next steps we want to approach within the SFE project is utilizing sudden stratospheric warmings (SSWs) for our seasonal predictions. SSWs are events that are, in particular in winter, often followed by cold spells in Norway that can persist for several weeks. However, this knowledge is difficult to integrate into an operational forecasting system based on NWPs and has therefore not been utilized to a large extend. We plan to investigate to what extend the occurrence of SSWs can be used to improve seasonal predictions, and whether the prediction of SSWs can lead to skillful surface temperature forecasts at longer lead times.

– Claudio Heinrich