H141-0021
Precipitation forecasting using machine-learning-based ensemble aggregation with Wasserstein-guided weighting
Precipitation forecasting using machine-learning-based ensemble aggregation with Wasserstein-guided weighting
Monday, 14 December 2020
Poster
Abstract:
Precipitation forecasting is one of the most complex modeling tasks, requiring the resolution of numerous spatial and temporal patterns that are sensitive to the accurate representation of many secondary variables (precipitable water column, air humidity, pressure, etc.). Ensemble forecasting approaches are commonly used to better explore the full dynamics of the solution space and reduce the forecast accuracy requirements -- these forecasts are then disseminated as probabilistic products or ensemble averages. Emanating from this, there is a large amount of literature related to ensemble averaging approaches for atmospheric modeling. These can range from simple arithmetic averaging to various model-weighting schemes based on heuristics, prior knowledge, or deviation from observations based on a moving temporal period of varying length.
In this study we implemented a learning-aggregation technique that used past observations and past model forecasts to calculate a weight for each model. We explored a number of schemes to quantify model misfit against observations for the weight computation; namely, a standard L2 norm that considers point-wise comparisons, and a Wasserstein metric that encompasses spatial variation. In its essence, the Wasserstein metric captures spatial dependency in the model data, allowing the misfit computation to pick up on subtle spatial differences, where, for example, the magnitude of the prediction may be correct but the location is wrong. The approach was implemented for an ensemble of up to five regional weather models (e.g. Global Forecast System, North American Mesoscale Forecast System, ...) and at least 177 observation stations covering the entire state of Texas. The aggregated forecasts were compared to observation data to quantify the performance of the model ensemble and aggregation techniques. We demonstrated that while relatively simple algorithms based on the standard L2 norm outperformed the best individual model, spatial complexity must be incorporated to generate a high-performing aggregation framework for precipitation forecasting.
In this study we implemented a learning-aggregation technique that used past observations and past model forecasts to calculate a weight for each model. We explored a number of schemes to quantify model misfit against observations for the weight computation; namely, a standard L2 norm that considers point-wise comparisons, and a Wasserstein metric that encompasses spatial variation. In its essence, the Wasserstein metric captures spatial dependency in the model data, allowing the misfit computation to pick up on subtle spatial differences, where, for example, the magnitude of the prediction may be correct but the location is wrong. The approach was implemented for an ensemble of up to five regional weather models (e.g. Global Forecast System, North American Mesoscale Forecast System, ...) and at least 177 observation stations covering the entire state of Texas. The aggregated forecasts were compared to observation data to quantify the performance of the model ensemble and aggregation techniques. We demonstrated that while relatively simple algorithms based on the standard L2 norm outperformed the best individual model, spatial complexity must be incorporated to generate a high-performing aggregation framework for precipitation forecasting.