Error correction of mesoscale numerical weather forecast by machine learning methods

The metrics of success of short-term numerical weather forecasts for the territory of Belarus based on the mesoscale hydrodynamic model of the atmosphere WRF (Weather Research and Forecasting) with regionally adapted block of underlying surface modelling are analyzed. The efficiency of systematic error correction of numerical forecasts of surface meteorological parameters using the method of exponentially weighted moving average and Kalman filter is evaluated.

A method of increasing the validity of mesoscale numerical forecasts has been proposed, in which corrections to the prognostic fields of meteorological parameters are calculated using machine learning models, and predictors are forecasts of several global and regional hydrodynamic models of the atmosphere. A comparative analysis of the efficiency of correction of numerical forecasts using machine learning models (XGBoost and ElasticNet) and recurrent graph neural network with attention mechanism (TGCN-A) was carried out.

It is shown that due to a more accurate description of the underlying surface in the WRF model and preliminary processing of its output by the exponentially weighted moving average method, the validity of numerical temperature fore- casts for periods up to 48 h increases by 1.0–4.4 %, and the mean square error of wind speed decreases by 0.1–0.2 m/s. Subsequent correction of WRF model forecasts using machine learning and neural network methods additionally increases the validity of numerical temperature forecasts by 3,2 % on average, and also reduces the RMS errors of surface pressure and wind speed forecasts by 0.20–0.74 hPa and 0.32–0.53 m/s, respectively.

1. Harvey A. C. Time Series Models. 2nd edition. New York, Harvester Wheatsheaf Publ., 1993, pp. 179–198.

2. Kalman R. E. A new approach to linear filtering and prediction problems. Journal of Basic Engineering, 1960, vol. 82, pp. 35–45.

3. Alerskans E., Kaas E. Local temperature forecasts based on statistical post-processing of numerical weather prediction data. Meteorol. Appl., 2021, vol. 28, no. 4, 21 p.

4. Bykov F. L. Statisticheskaya korrekciya prognozov pogody po modeli COSMO s pomoshch'yu nejronnyh setej [Statistical correction of weather forecasts based on the COSMO model using neural networks] Meteorologiya I gidrologiya = Meteorology and Hydrology, 2020, vol. 3, pp. 5–20. (in Russian)

5. Vannitsem S., Bremnes J. B., Demaeyer J., Evans G., e.a. Statistical Postprocessing for Weather Forecasts: Review, Challenges, and Avenues in a Big Data World. Bulletin of the American Meteorological Society, 2021, vol. 102, no. 3, pp. E681–E699.

6. Bengtsson L., Ghil M., Kallen E. Dynamic Meteorology: Data Assimilation Methods. New York, Springer-Verlag Publ., 1981, 330 p.

7. Lorenc A. C. Analysis methods for numerical weather prediction. Quart. J. Roy. Meteor. Soc., 1986, vol. 112, pp. 1177–1194.

8. Rozinkina I. A., Rivin G. S., Burak R. N., e.a. Formirovanie produkcii sistem negidrostaticheskogo modelirovaniya atmosfery COSMO-RuBy (Gidrometcentr Rossii) i WRF-ARW (Belgidromet) dlya kratkosrochnogo prognoza pogody [Product development of COSMO-RuBy (Hydrometeorological Center of Russia) and WRF-ARW (Belhydromet) nonhydrostatic atmospheric modeling systems for short-term weather forecast]. Gidrometeorologicheskie issledovaniya i prognozy = Hydrometeorological studies and forecasts, 2021, vol. 4 (382), pp. 6–29. (in Russian)

9. Lysenko S. A., Zaiko P. O. Ocenki vliyaniya podstilayushchej poverhnosti na tochnost' chislennogo prognoza temperatury vozduha na territorii Belarusi s ispol'zovaniem modeli WRF [Estimation of the influence of the underlying surface on the accuracy of the numerical forecast of air temperature in Belarus using the WRF model]. Gidro meteorologicheskie issledovaniya i prognozy = Hydrometeorological studies and forecasts, 2021, vol. 4, pp. 50–68. (in Russian)

10. Lysenko S. A., Zaiko P. O. Geoinformacionnaya sistema povysheniya opravdyvaemosti gidrodinamicheskogo prognoza pogody dlya territorii Belarusi na osnove dannyh distancionnogo zondirovaniya Zemli i ob"ektivnogo analiza meteorologicheskih polej [Geoinformation system for improving the accuracy of hydrodynamic weather forecasting for the territory of Belarus based on remote sensing data and objective analysis of meteorological fields]. Prirodopol'zovanie = Nature Management, 2024, no. 1, pp. 30–40. (in Russian)

11. Kipf T. N., Welling M. Semi-Supervised Classification with Graph Convolutional Networks. ArXiv preprint, 2016, arXiv:1609.02907.

12. Zhao L., Song Y., Zhang C., e.a. T-GCN: A Temporal Graph Convolutional Network for Traffic Prediction. IEEE Transactions on Intelligent Transportation Systems, 2020, vol. 21, no. 9, pp. 3848–3858.

13. Bai J., Zhu J., Song Y., e.a. A3T-GCN: Attention Temporal Graph Convolutional Network for Traffic Forecasting. ISPRS International Journal of Geo-Information, 2021, vol. 10, no. 7, p. 485.