A Novel, Physics-Based Data Analytics Framework for Reducing Systematic Model Errors

Abstract:

Most climate and weather models exhibit systematic biases, such as under predicted diurnal temperatures in the WRF (Weather Research and Forecasting) model. General approaches to alleviate the systematic biases include improving model physics and numerics, improving data assimilation, and bias correction through post-processing. In this study, we developed a novel, physics-based data analytics framework in post processing by taking advantage of ever-growing high-resolution (spatial and temporal) observational and modeling data. In the framework, a spatiotemporal PCA (Principal Component Analysis) is first applied on the observational data to filter out noise and information on scales that a model may not be able to resolve. The filtered observations are then used to establish regression relationships with archived model forecasts in the same spatiotemporal domain. The regressions along with the model forecasts predict the projected observations in the forecasting period. The pre-regression PCA procedure strengthens regressions, and enhances predictive skills. We then combine the projected observations with the past observations to apply PCA iteratively to derive the final forecasts. This post-regression PCA reconstructs variances and scales of information that are lost in the regression. The framework was examined and validated with 24 days of 5-minute observational data and archives from the WRF model at 27 stations near Dugway Proving Ground, Utah. The validation shows significant bias reduction in the diurnal cycle of predicted surface air temperature compared to the direct output from the WRF model. Additionally, unlike other post-processing bias correction schemes, the data analytics framework does not require long-term historic data and model archives. A week or two of the data is enough to take into account changes in weather regimes. The program, written in python, is also computationally efficient.