Advancing Sub-Seasonal Weather Predictability Through Machine Learning Techniques

点击翻译按钮获取中文摘要

Current operational weather forecasting systems can produce valuable predictions of day-to-day weather, but such predictions are only skillful up to a week or so in advance. On the longer subseasonal timescale, say between two and eight weeks, useful forecasts of mean conditions and the likelihood of extreme events may also be possible. But basic science questions, including the sources and mechanisms of subseasonal variability, their potential predictability, and the essential elements necessary for robust prediction, have not yet been resolved. These questions are of practical as well as scientific interest, as guidance from subseasonal forecasts could have a variety of uses including agricultural planning and emergency management.This project seeks to identify empirical predictive relationships between local climate variability of interest and large-scale patterns in relevant predictors such as sea surface temperature (SST), soil moisture, and atmospheric circulation. Prior work by the Principal Investigator (PI) and colleagues demonstrated such a relationship between heat waves in Texas, including the heat wave associated with the 2011 drought, and an SST pattern covering much of the North Pacific. A key limitation to such prediction methods is that the observed record is too short to identify statistically significant predictive relationships. Thus methods which seem successful when tested on past cases may fail when used for realtime prediction. The PI's strategy for circumventing this limitation is to identify predictive relationships using output from weather and climate models in place of observations, as many thousands of years of simulated weather and climate variability are available from a variety of modeling projects. Model output used here comes from the North American Multimodel Ensemble (NMME), the Subseasonal to Seasonal (S2S) Prediction Project, and the Coupled Model Intercomparison Project (CMIP). A further concern in developing empirical prediction methods is the need for regularization, meaning a way to eliminate spurious small-scale features in predictor patterns which arise due to the large number of data points used to represent the predictor fields. Such features are not usually consistent from model to model or between model output and observations. The PI uses a regularization scheme in which eigenvectors of the Laplacian operator serve to factor out small scales, leading to more robust predictive relationships. To further ensure robust predictions, the PI applies an innovative cross validation technique which bypasses the best statistical model identified in cross validation in favor of the simplest model which is within a standard deviation of the best model.Once robust predictive relationships are identified that hold across different models and in observations, the sources and mechanisms responsible for the relationships will be explored. If the empirical methods can reproduce hindcasts from the models in the NMME archive then the model output can be used to understand the time-evolving dynamical processes linking the predictor pattern to the predicted variability.In addition to the societal benefits of subseasonal forecasts, the project provides support and training to a postdoc, thereby promoting workforce development in this research area. The project also addresses public scientific literacy through public seminars by the PI on sub-seasonal prediction, a topic of interest to the general public.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.