[M-GI26] Data-driven approaches for weather and hydrological predictions

[MGI26-01] Machine learning enables real-time proactive quality control: A proof-of-concept study

★Invited Papers

*Takumi Honda¹, Akira Yamazaki² (1.Faculty of Science, Hokkaido University, 2.Application Laboratory, JAMSTEC)

Keywords:Proactive quality control, Machine learning, Numerical weather prediction, Ensemble forecast sensitivity to observations, Reservoir computing

To improve the forecast accuracy of numerical weather prediction (NWP), it is essential to estimate accurate initial conditions by assimilating available observations. It has been known that some observations could degrade the forecast accuracy. Such detrimental observations can be detected by ensemble forecast sensitivity to observations (EFSO). Denying detrimental observations detected by EFSO as proactive quality control (PQC) has been shown to be effective for improving the forecast accuracy of NWP. However, EFSO requires future observations to evaluate current observations’ impacts on forecasts, so that PQC cannot be real-time in general. This study proposes using machine learning (ML) predictions in place of future observations in EFSO. By doing so, EFSO and PQC do not need to wait for future observations and could be performed in real-time. This study conducts proof-of-concept with the 40-variable Lorenz model and reservoir computing as ML. The results indicate that observation impacts on forecasts are generally consistent between ML-based EFSO and conventional EFSO. Furthermore, our ML-based PQC successfully improves the forecast accuracy although it does not require future observations.