A Deep Convolutional Neural Network Model for Improving WRF Simulations

Advancements in numerical weather prediction (NWP) models have accelerated, fostering a more comprehensive understanding of physical phenomena pertaining to the dynamics of weather and related computing resources. Despite these advancements, these models contain inherent biases due to parameterization of the physical processes and discretization of the differential equations that reduce simulation accuracy. In this work, we investigate the use of a computationally efficient deep learning (DL) method, the convolutional neural network (CNN), as a postprocessing technique that improves mesoscale Weather Research and Forecasting (WRF) one-day simulation (with a 1-h temporal resolution) outputs. Using the CNN architecture, we bias-correct several meteorological parameters calculated by the WRF model for all of 2018. We train the CNN model with a four-year history (2014-2017) to investigate the patterns in WRF biases and then reduce these biases in simulations for surface wind speed and direction, precipitation, relative humidity, surface pressure, dewpoint temperature, and surface temperature. The WRF data, with a spatial resolution of 27 km, cover South Korea. We obtain ground observations from the Korean Meteorological Administration station network for 93 weather station locations. The results indicate a noticeable improvement in WRF simulations in all station locations. The average of annual index of agreement for surface wind, precipitation, surface pressure, temperature, dewpoint temperature, and relative humidity of all stations is 0.85 (WRF:0.67), 0.62 (WRF:0.56), 0.91 (WRF:0.69), 0.99 (WRF:0.98), 0.98 (WRF:0.98), and 0.92 (WRF:0.87), respectively. While this study focuses on South Korea, the proposed approach can be applied for any measured weather parameters at any location.