Summary of the different tasks performed in ISSUL:
1. Two targets of different complexity (see black curves in the attached figure), which describe regional summer temperature conditions in Iberia, were computed. This was done to test the sensitivity of the machine learning based forecast model to the complexity of the targets.
2. A pool of predictors containing 121 potential drivers was created. This pool of predictor is composed of 120 PCs (20 leading PCs of the 6 predictor variables) plus one more predictor: the day of the summer that we aim to predict. The predictor variables are: the sea surface temperature (SST), the 2m-temperature over Europe (T2M), the soil moisture (SM) over Europe, the outgoing-longwave radiation (OLR), the sea ice cover (SIC) in the Northern Hemisphere and the snow cover (SC) in the Northern Hemisphere.
3. The optimization algorithm was coupled to a regression algorithm. This machine learning model was trained then trained using the training data. Finally, a prediction of the test set part of the target was performed. These tasks are part of Objective 2
4. Once a satisfying skill was obtained for the targets, the dynamical link between the set of optimal predictors selected by the optimization algorithm and the target was analysed. This task is part of Objective 1.
Main results achieved in this project:
1. A machine learning prediction model based on the coupling of an optimization algorithm with a regression algorithm skilfully predicts regional monthly mean temperature conditions over southern Europe with one month lead time. This is the main scientific result of the project and its main technological achievement. This work further demonstrated that a machine learning prediction model of Iberian monthly 2m maximum temperature outperforms a simple statistical model based on the climatology for a lead time of at least one month (see figure).
2. Regional monthly heatwave intensity in southern Europe is more difficult to predict one month in advance than the mean temperature conditions. The model did not achieve a skilful prediction for the regional heatwave intensity index. Although the result can be considered negative, it also illustrates that extreme events are more difficult to predict. This outcome was somewhat anticipated, given that extreme events are inherently rare, resulting in an unbalanced training set (fewer days with a heatwave than non-heatwave days) to train the model.
3. Confidence in the model: dynamical connections between recurrent predictors and the target. ISSUL demonstrates that there are recurrent sets of optimal predictors (i.e. preferred combinations of skilful predictors’ states), and hence windows of opportunity to further improve the predictability of summer high temperatures and extremes. Second, the most commonly selected predictors have impacts on regional temperature in the next month, indicating that the model is not an artefact and that the drivers have true predictive signals. More specifically, the calendar day and the first PC of the SIC (or SST) allows the model to learn the climatology and trend in the data. The fourth PC of the OLR appears to be associated with cooler 2m maximum temperatures via a wave train developing in the tropical Atlantic during its positive phase and a weaker anomaly dipole during its negative phase related to summer NAO-like negative phase. All these results contribute to uncover new predictors of monthly 2m maximum temperatures over southern Europe and to understand the sources of its skill, thus enhancing trust in our model's performance.
The results of the project were disseminated in three presentations in international conferences and one scientific paper (in preparation). No website has been developed for the project.
