A novel ensemble learning for post-processing of NWP Model's next-day maximum air temperature forecast in summer using deep learning and statistical approaches

Abstract

A reliable and accurate extreme air temperature in summer is necessary to prepare for and respond to thermal disasters such as heatstroke and power outages. The numerical weather prediction (NWP) model is commonly used to forecast air temperature using dynamic mechanisms. Because of its high uncertainty from coarse spatial resolution and unstable parameterization, however, it requires post-processing. Recent studies have proposed advanced post-processing methods using machine learning and deep learning techniques. This study compared various individual post-processing models—multi-linear regression (MLR), support vector regression (SVR), gated recurrent units (GRU), and convolutional neural network (CNN). It also proposed a novel multi-model ensemble (MMESS) that aggregates individual post-processing models based on the skill score (SS) for the Local Data Assimilation and Prediction System (LDAPS, a local NWP model over Korea) model's next-day maximum air temperature (Tmax) forecast data in two different domains: South Korea and Seoul. The pressure and surface data of the present-day analysis and next-day forecast fields of LDAPS were used as input variables. As a result of hindcast validation, CNN showed good overall performance (root mean square error (RMSE) of 1.41 °C in South Korea and 1.50 °C in Seoul) among individual models. We found that CNN demonstrated lower RMSE (1.17–1.58 °C) than other post-processing models (1.43–2.17 °C) at stations where the bias of LDAPS changes, using surrounding spatial information. The proposed MMESS exhibited more reliable, robust results than the individual models did. A further comparison to the simple average ensemble and the constrained linear squares-based MMEsupported the proposed MMESS as a more suitable ensemble method for next-day Tmax forecast, considering the relative significance of the individual models.