Regional Weather Variable Predictions by Machine Learning with Near-Surface Observational and Atmospheric Numerical Data *
Yihe Zhang, Bryce Turney, Purushottam Sigdel, Xu Yuan, Eric Rappin, Adrian L. Lago, Sytske Kimball, Li Chen, Paul Darby, Lu Peng, Sercan Aygun, Yazhou Tu, M. Hassan Najafi, Nian-Feng Tzeng · 2024
Read Regional Weather Variable Predictions by Machine Learning with Near-Surface Observational and Atmospheric Numerical Data * by Yihe Zhang, Bryce Turney, Purushottam Sigdel, Xu Yuan, Eric Rappin, Adrian L. Lago, Sytske Kimball, Li Chen, Paul Darby, Lu Peng, Sercan Aygun, Yazhou Tu, M. Hassan Najafi, Nian-Feng Tzeng (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-10450.txt.
Introduction
Regional weather forecasts drive farm, energy, and safety decisions, but equation-based models can be coarse and slow. Yihe Zhang and thirteen co-authors built a learning model in 2024 that blends ground-station readings with atmospheric model output for finer local forecasts.
The model, MiMa, pairs two encoders, one for five-minute observations from Kentucky weather stations and one for hourly numerical output, with a decoder that forecasts single variables at single stations. A regional variant learns from a few representative stations and then predicts at places with no gauges, using elevation as a tag. The authors argue the micro-macro blend gives sharp short-range forecasts where officials and growers actually need them.
Catalog record
| CREATOR | Yihe Zhang, Bryce Turney, Purushottam Sigdel, Xu Yuan, Eric Rappin, Adrian L. Lago, Sytske Kimball, Li Chen, Paul Darby, Lu Peng, Sercan Aygun, Yazhou Tu, M. Hassan Najafi, Nian-Feng Tzeng |
|---|---|
| YEAR | 2024 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 121 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2412-10450
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Regional Weather Variable Predictions by Machine Learning with Near-Surface Observational and Atmospheric Numerical Data * Yihe Zhang, Bryce Turney, Purushottam Sigdel, Xu Yuan, Eric Rappin, Adrian L. Lago, Sytske Kimball, Li Chen, Paul Darby, Lu Peng, Sercan Aygun, Yazhou Tu, M. Hassan Najafi, Nian-Feng Tzeng Abstract Accurate and timely regional weather prediction is vital for sectors dependent on weather-related decisions. Traditional prediction methods, based on atmospheric equations, often struggle with coarse temporal resolutions and inaccuracies. This paper presents a novel machine learning (ML) model, called MiMa (short for Mi cro- Ma cro), that integrates both near-surface observational data from Kentucky Mesonet stations (collected every five minutes, known as Micro data) and hourly atmospheric numerical outputs (termed as Macro data) for fine-resolution weather forecasting. The MiMa model employs an encoder-decoder transformer structure, with two encoders for processing multivariate data from both datasets and a decoder for forecasting weather variables over short time horizons. Each instance of the MiMa model, called a modelet, predicts the values of a specific weather parameter at an individual Mesonet station. The approach is extended with Re-MiMa modelets, which are designed to predict weather variables at ungauged locations by training on multivariate data from a few representative stations in a region, tagged with their elevations. Re-MiMa (short for Re gional-MiMa) can provide highly accurate predictions across an entire region, even in areas without observational stations. Experimental results show that MiMa significantly outperforms current models, with Re-MiMa offering precise short-term forecasts for ungauged locations, marking a significant advancement in weather forecasting accuracy and applicability. I Introduction Accurate short-term weather predictions with fine temporal resolutions are crucial for sectors that depend on real-time weather-related decision-making, such as transportation, emergency response, solar farm operations, etc. However, current forecasting models, such as the Weather Research and Forecasting (WRF) model with High-Resolution Rapid Refresh (HRRR) [ 2 ] , fall short of meeting these demands due to their coarse hourly outputs and high computational complexity. These models generate around 148 weather parameter values (i.e., variables) per hour over large geo-grids (e.g., 3 km × 3 km), with coarse temporal granularity (hourly forecasts) often deemed insufficient for applications requiring predictions in the interval of 5 or 15 minutes [ 3 ] . Additionally, a lack of near-surface observational data at tactical locations limits their accuracy. On the other hand, regional Mesonet networks, such as the Kentucky Mesonet [ 4 ] , provide real-time, location-specific weather data with fine temporal granularity (e.g., every five minutes). These networks operate under the U.S. National Mesonet Program [ 5 ] and consist of strategically located observational stations. The Kentucky Mesonet, for example, comprises over 70 stations for collecting values of some 22 weather parameters, including temperature, humidity, wind speed, pressure, and precipitation. This high-resolution dataset provides valuable micro-level data that can be leveraged to improve prediction accuracy (see Fig. 1 for station distribution). Fig. 1 : Kentucky Mesonet weather observational stations denoted by yellow circles, with those stations chosen for MiMa model evaluation and pointed by red line segments tagged with their latitudes, longitudes, and elevations. Fig. 2 : Overview of the MiMa (short for Micro-Macro) model inputted with data from both an individual station and WRF-HRRR modeling computation to yield the weather variable predictions. Recent advances in machine learning (ML) technology have propelled weather forecasting into a new era. Numerous studies have explored ML-centric techniques for weather forecasting, yielding promising results. These techniques include the deep neural networks (DNN), convolutional neural networks (CNN), long short-term memory networks (LSTM) [ 6 ] , generative adversarial networks (GAN) [ 7 ] , among others, for predicting values of such parameters as wind speed and direction [ 8 , 9 , 10 , 11 ] , solar radiation [ 12 , 13 ] , precipitation [ 14 , 15 , 16 , 17 , 18 ] , air quality [ 19 ] , weather changes [ 20 , 21 , 22 , 23 ] , etc. However, existing ML forecasting models have not yet achieved accurate predictions of weather variables at fine temporal resolutions (as seen with recent transformer-variant forecasters like [ 24 ] , [ 25 ] , [ 26 ] , [ 27 ] ). This shortcoming is largely due to the lack of geo-location-aligned ground observational data, which is essential for enhancing model training and prediction accuracy. Many models rely solely on satellite or radar images [ 28 ] , limiting their precision. Additionally, most neural network-based models are designed to predict values of specific parameters, lacking a generalizable framework that can be readily adapted to forecast all parameters of interest. To address these shortcomings, this paper introduces the MiMa model, a novel ML-based approach that integrates fine-grained observational data (Micro data) from regional Mesonet stations with larger-scale numerical outputs (macro data) from the WRF-HRRR model [ 3 ] . This integration enables accurate weather predictions over short time horizons (in minutes) for a region of interest. The MiMa model employs an encoder-decoder transformer architecture, where two encoders process multivariate sequences from the micro and the macro datasets, and a decoder forecasts the values of multiple weather parameters across a sequence of time points from t+1 to t+L (see Fig. 2 ). Each instance of the MiMa model is referred to as a modelet, dedicated to predicting a specific weather variable for an individual Mesonet station. To further enhance its utility, the MiMa model is extended to become Regional MiMa (Re-MiMa) modelets, which can predict weather variables at ungauged locations (where no observational stations exist). By using data from several representative stations in a region of interest, typically 3 or 4, tagged with their elevations, Re-MiMa modelets can generalize predictions across an entire region, providing accurate forecasts at both gauged and ungauged locations. This extension addresses a long-standing challenge in meteorological forecasting: accurate predictions at locations without direct observational data. We conducted experiments on weather forecasting at various Kentucky Mesonet stations, focusing on four key weather parameters: air temperature, relative humidity, wind speed, and atmospheric pressure, across eleven stations (Figure 1 ). The MiMa modelets consistently outperform all comparative techniques, achieving the lowest RMSE (Root Mean Squared Error) values in 39 out of 44 forecasting instances (Table VI . Additionally, Re-MiMa modelets, trained using data from three representative stations in eastern Kentucky (BMTN, FARM, and DANV), were tested at eight ungauged stations. The results show that Re-MiMa modelets predicted weather variables with accuracy comparable to, or better than, location-specific MiMa modelets, with 22 out of 32 parameters at those ungauged locations predicted more accurately (see Table X ). This demonstrates the effectiveness of Re-MiMa in avoiding the need for multiple modelets for predicting a given variable while maintaining high accuracy across the entire region. In summary, the MiMa model integrates micro and macro data to deliver precise weather predictions at fine temporal resolutions. The Re-MiMa extension further enhances its regional forecasting capability, making it a versatile tool for applications that require high-accuracy predictions in real-time. The MiMa model code, documentation, and datasets are made publicly available at [ 29 ] for further research and applications to other regions. This paper makes several key contributions: • Novel Weather Prediction Model (MiMa) . We introduce the MiMa model,