Periodic Reporting for period 1 - ARTIST (ARTificial Intelligence for Seasonal forecast of Temperature extremes)
Okres sprawozdawczy: 2022-04-01 do 2024-03-31
The project ARTIST had the overarching objective of improving seasonal foreacsts of heat waves in Europe and providing a better understanding of what drives this predictability, making use of artificial intelligence techniques to complement dynamical models. Using a data-driven approach, a prediction of seasonal heat wave intesity is made over ten different European areas. The results are promising, with heat waves in a few regions that demonstrated to be possibly skilfully predicted by a data-driven model, much cheaper than a dynamical system.
The architecture developed is potentially important for the society, since it introduces an innovative and very cheap methodology to perform seasonal forecasting. In fact, once the training is performed and the model has passed a benchmark for the evaluation of its performance, steps already done in the framework of ARTIST, the seasonal forecast for the following season can be performed in a few minutes with a personal laptop, an internet connection and basic knowledge of the software python. Farmers, energy providers, tourist operators are among the potential users of the project results.
The architecture developed is potentially important for the society, since it introduces an innovative and very cheap methodology to perform seasonal forecasting. In fact, once the training is performed and the model has passed a benchmark for the evaluation of its performance, steps already done in the framework of ARTIST, the seasonal forecast for the following season can be performed in a few minutes with a personal laptop, an internet connection and basic knowledge of the software python. Farmers, energy providers, tourist operators are among the potential users of the project results.
The implementation and tuning of a machine learning architecture for the prediction of heat waves was the main achievement of the action.
The model features used as input have been separated between large scale (remote) and local scale predictors. Sea surface temperatures from several global ocean regions, together with sea ice concentration from four Arctic areas were the remote predictors, while soil moisture and temperature, surface fluxes and snow water equivalent were the local drivers. The two main groups differed for the prescribed lag time to which they produce their main effect on the heat waves: one month for the remote drivers, two weeks for the local drivers. In addition, a global predictor was used, namely the global CO2 concentration, as a proxy of the global warming trend that has been demonstrated to be a main driver of predictability, especially for summer weather in Europe.
For each region, a feature selection algorithm creates a subset of features that are ingested by a random forest base learner. This model calculates the root mean squared error (RMSE) between observed heat waves, used for training, and predicted heat waves. The feature subset allowing for the minimum RMSE represent the driver pool that better contribute to the prediction of heat waves in that region.
The goodness of the ML-model’s fit is given by the coefficient of determination. While there is a lot of diversity among regions and seasons, the summer prediction is generally better than the prediction for the other seasons, and this may be due to larger persistence and lesser interannual variability that characterize the summer season in Europe. Many areas are clearly characterized by no or very limited prediction skill; however, a few regions such as the Western Mediterranean, the Middle East, and west and central Europe, show considerable skill.
Heat waves are predicted by an ensemble of best models in every region for the four main seasons. he more represented features are generally those that contribute the most to the predicted European HWP: the relative contribution of each feature is given by the SHAP value. The most represented feature is the CO2 concentration, which appears as a predictor of summer heat waves in all the European regions. Heat waves in other seasons are also associated to CO2 concentration in many areas. CO2 concentration is an obvious proxy of the anthropogenic forcing that kept increasing since the 1950s, with an acceleration in the last 40 years, and has previously been linked to the skill of European seasonal forecasts . Apart from the external forcing, some predictors seem to be important for the heat waves in many regions in specific seasons: SST anomalies in different areas of the Atlantic ocean for the spring predictions and in the Mediterranean Sea for the summer predictions, North Pacific SST anomalies for the autumn, soil moisture anomalies at several layers for the fall and summer, snow water equivalent for the winter.
The western Mediterranean is one of the regions where the model skill is the highest in the summer, with correlation between observed and predicted heat waves higher than 0.6. This value is even higher than that of the dynamical prediction provided by the ECMWF System 5, one of the most widely used prediction systems.
The model features used as input have been separated between large scale (remote) and local scale predictors. Sea surface temperatures from several global ocean regions, together with sea ice concentration from four Arctic areas were the remote predictors, while soil moisture and temperature, surface fluxes and snow water equivalent were the local drivers. The two main groups differed for the prescribed lag time to which they produce their main effect on the heat waves: one month for the remote drivers, two weeks for the local drivers. In addition, a global predictor was used, namely the global CO2 concentration, as a proxy of the global warming trend that has been demonstrated to be a main driver of predictability, especially for summer weather in Europe.
For each region, a feature selection algorithm creates a subset of features that are ingested by a random forest base learner. This model calculates the root mean squared error (RMSE) between observed heat waves, used for training, and predicted heat waves. The feature subset allowing for the minimum RMSE represent the driver pool that better contribute to the prediction of heat waves in that region.
The goodness of the ML-model’s fit is given by the coefficient of determination. While there is a lot of diversity among regions and seasons, the summer prediction is generally better than the prediction for the other seasons, and this may be due to larger persistence and lesser interannual variability that characterize the summer season in Europe. Many areas are clearly characterized by no or very limited prediction skill; however, a few regions such as the Western Mediterranean, the Middle East, and west and central Europe, show considerable skill.
Heat waves are predicted by an ensemble of best models in every region for the four main seasons. he more represented features are generally those that contribute the most to the predicted European HWP: the relative contribution of each feature is given by the SHAP value. The most represented feature is the CO2 concentration, which appears as a predictor of summer heat waves in all the European regions. Heat waves in other seasons are also associated to CO2 concentration in many areas. CO2 concentration is an obvious proxy of the anthropogenic forcing that kept increasing since the 1950s, with an acceleration in the last 40 years, and has previously been linked to the skill of European seasonal forecasts . Apart from the external forcing, some predictors seem to be important for the heat waves in many regions in specific seasons: SST anomalies in different areas of the Atlantic ocean for the spring predictions and in the Mediterranean Sea for the summer predictions, North Pacific SST anomalies for the autumn, soil moisture anomalies at several layers for the fall and summer, snow water equivalent for the winter.
The western Mediterranean is one of the regions where the model skill is the highest in the summer, with correlation between observed and predicted heat waves higher than 0.6. This value is even higher than that of the dynamical prediction provided by the ECMWF System 5, one of the most widely used prediction systems.
Progress in AI-based forecasting on weather timescales, i.e. less than 10 days, has been remarkable in the last few years. In parallel with the rapid rise of AI, forecasting institutes worldwide and Big Tech companies have seized upon the opportunity to improve weather forecasts, gaining skill comparable to that of state-of-the-art dynamical predictions.
Compared to the short timescales, progress on the subseasonal to decadal (S2D) timescale has been less striking. A fundamental challenge is the limited amount of independent training data, roughly one or two orders of magnitude smaller than for weather timescales. In fact, weather predictions may target individual extreme events, while climate prediction can only aim at their aggregation overtime (Meehl et al., 2021), inevitably decreasing the number of observational samples available for the past. This has long hampered the development of long-lead forecasts of extremes like drought and warm spells, which likely have at least some predictability at the S2D scale. However, an increasing number of articles has been published since the “S2S reboot” opinion paper (Cohen et al., 2019), that claimed that ML techniques mostly developed in computer science could be adopted by climate forecasters to increase the accuracy of predictions at subseasonal to seasonal (S2S) scale.
The work developed in the frame of ARTIST is linked to recent studies of similar kind, where long-term predictions of mid-latitude extreme events have been tested using different machine and deep learning algorithms.
Compared to the short timescales, progress on the subseasonal to decadal (S2D) timescale has been less striking. A fundamental challenge is the limited amount of independent training data, roughly one or two orders of magnitude smaller than for weather timescales. In fact, weather predictions may target individual extreme events, while climate prediction can only aim at their aggregation overtime (Meehl et al., 2021), inevitably decreasing the number of observational samples available for the past. This has long hampered the development of long-lead forecasts of extremes like drought and warm spells, which likely have at least some predictability at the S2D scale. However, an increasing number of articles has been published since the “S2S reboot” opinion paper (Cohen et al., 2019), that claimed that ML techniques mostly developed in computer science could be adopted by climate forecasters to increase the accuracy of predictions at subseasonal to seasonal (S2S) scale.
The work developed in the frame of ARTIST is linked to recent studies of similar kind, where long-term predictions of mid-latitude extreme events have been tested using different machine and deep learning algorithms.