Improving Atmospheric Angular Momentum Forecasts by Machine Learning

Abstract

Earth angular momentum forecasts are naturally accompanied by forecast errors that typically grow with increasing forecast length. In contrast to this behavior, we have detected large quasi‐periodic deviations between atmospheric angular momentum wind term forecasts and their subsequently available analysis. The respective errors are not random and have some hard to define yet clearly visible characteristics which may help to separate them from the true forecast information. These kinds of problems, which should be automated but involve some adaptation and decision‐making in the process, are most suitable for machine learning methods. Consequently, we propose and apply a neural network to the task of removing the detected artificial forecast errors. We found that a cascading forward neural network model performed best in this problem. A total error reduction with respect to the unaltered forecasts amounts to about 30% integrated over a 6‐days forecast period. Integrated over the initial 3‐days forecast period, in which the largest artificial errors are present, the improvements amount to about 50%. After the application of the neural network, the remaining error distribution shows the expected growth with forecast length. However, a 24‐hourly modulation and an initial baseline error of 2 × 10−8 became evident that were hidden before under the larger forecast error.

Plain Language Summary: Variations in Earth rotation can be described by changes in Earth angular momentum. Angular momentum functions are calculated from mass redistributions, for example, given by atmospheric models. Typically, atmospheric model forecasts are naturally accompanied by forecast errors that grow with increasing forecast length. In contrast to this behavior, atmospheric angular momentum wind term forecasts show large quasi‐periodic deviations when compared to their subsequently available model analysis data. The detected errors are not random and have some hard to define yet clearly visible characteristics. A postprocessing step using machine learning methods was established to remove the detected artificial forecast errors. A cascading forward neural network approach was able to reduce the forecast error by about 50% for the first forecast days and about 30% for a 6‐day forecast horizon. Moreover, the remaining error distribution shows the expected growth with forecast length. This postprocessing step improves atmospheric angular momentum forecasts without touching the numerical weather prediction model itself. Improved angular momentum forecasts should help to further decrease Earth rotation predictions errors.

Key Points:

Motion terms of atmospheric angular momentum forecasts contain systematic errors.

Machine learning is used to learn and reduce these errors.

Remaining stochastic errors show modulations with a 24‐hr period.