Precise Weather Parameter Predictions for Target Regions via Neural Networks

Abstract

Flexible fine-grained weather forecasting is a problem of national importance due to its stark impacts on economic development and human livelihoods. It remains challenging for such forecasting, given the limitation of currently employed statistical models, that usually involve the complex simulation governed by atmosphere physical equations. To address such a challenge, we develop a deep learning-based prediction model, called Micro-Macro, aiming to precisely forecast weather conditions in the fine temporal resolution (i.e., multiple consecutive short time horizons) based on both the atmospheric numerical output of WRF-HRRR (the weather research and forecasting model with high-resolution rapid refresh) and the ground observation of Mesonet stations. It includes: 1) an Encoder which leverages a set of LSTM units to process the past measurements sequentially in the temporal domain, arriving at a final dense vector that can capture the sequential temporal patterns; 2) a Periodical Mapper which is designed to extract the periodical patterns from past measurements; and 3) a Decoder which employs multiple LSTM units sequentially to forecast a set of weather parameters in the next few short time horizons. Our solution permits temporal scaling in weather parameter predictions flexibly, yielding precise weather forecasting in desirable temporal resolutions. It resorts to a number of Micro-Macro model instances, called modelets, one for each weather parameter per Mesonet station site, to collectively predict a target region precisely. Extensive experiments are conducted to forecast four important weather parameters at two Mesonet station sites. The results exhibit that our Micro-Macro model can achieve high prediction accuracy, outperforming almost all compared counterparts on four parameters of interest.