In a recent publication in the journal Applied Sciences, researchers reviewed the current landscape of machine learning applications in weather and climate forecasting, a domain of increasing significance spanning agriculture and disaster management.
Weather forecasting and climate prediction play vital roles in society, offering crucial information for various sectors. Moreover, the energy industry relies on climate forecasts to optimize energy distribution, ensuring a stable and efficient energy supply. This review primarily focuses on the application of machine learning in meteorological forecasting, a field that has evolved significantly over time.
Techniques for climate prediction
This article examines different techniques in climate prediction, including statistical and physical approaches. It outlines the historical evolution of predictive models in climate science, divided into three distinct phases: the initial prevalence of statistical methods, the ascendancy of physical models, and the current exploration of machine learning methods.
Statistical Methods: In meteorology, statistical or empirical forecasting techniques boast a rich history. Prior to the advent of computational models, they constituted the primary means of weather prediction. These techniques, which consider both long-term climate shifts and short-term weather patterns, are still essential to data-driven approaches in meteorological forecasting. Notable statistical techniques like Geographically Weighted Regression (GWR) and Spatio-Temporal Kriging (ST-Kriging) are useful tools in this situation.
Physical Models: Meteorological forecasts have traditionally relied on physical models, grounded in fundamental physical principles. The basis of these models lies in the mathematical formulation of governing equations and physical concepts, such as the Navier-Stokes equations in fluid dynamics. Due to their great precision and dependability over a wide range of periods, physical models are essential to climate modeling and numerical weather prediction (NWP).
Machine Learning Methods: As relatively recent entrants in meteorology, machine learning methods have emerged as promising solutions with the potential to enhance predictive outcomes. They offer faster predictions compared to physical models and more accurate results compared to statistical methods. The adaptability of machine learning for error correction and downscaling renders it a valuable addition to weather and climate prediction. Achieving accuracy and efficiency is paramount in weather forecasting and climate prediction, and machine learning demonstrates significant promise in meeting these requirements.
Taxonomy of climate prediction applications
The progression of machine learning applications in climate and weather prediction from before 2010 to recent innovations in 2022/2023 begins with the 1998 Precipitation Neural Network Prediction Model, one of the early Deep Neural Network (DNN) models for short-term precipitation forecasts. Advancements include the use of machine learning techniques such as k-nearest neighbors (KNN) and support vector machines (SVM) for localized precipitation forecasts. In 2009, conditional random fields (CRF) were introduced for medium-to-long-term precipitation prediction.
From 2010 to 2019, machine learning saw notable progress with models such as convolution-based long-short-term memory (ConsvLSTM) for short-term precipitation forecasts and the precipitation convolution prediction model focusing on localized precipitation. In 2020, the CapsNet model employed capsule networks for predicting extreme weather events in North America, and in 2021 brought models such as random forest (RF)-bias correction and sea-ice prediction. These models used random forests and probabilistic deep learning for various forecasts. Recent advancements in 2022 and 2023 introduced more complex architectures, such as Cycle-Consistent Generative Adversarial Networks (CycleGAN), PanGu, FourCastNet, DeepESD-Down-scaling, and Convolution Neural Networks (CNN)-Bias Correction, each leveraging different machine learning techniques for diverse climate and weather predictions.
Short-term weather forecast
To predict atmospheric phenomena within a short time frame, several machine-learning models, including Numerical Weather Prediction (NWP), MetNet, FourCastNet, and PanGu, were employed. These models exhibit unique strengths and advances in weather forecasting, offering a range of capabilities to meet diverse forecasting needs. MetNet balances accuracy and processing speed; FourCastNet excels in fine-scale feature predictions; GraphCast tracks weather patterns effectively; and PanGu introduces a revolutionary three-dimensional approach to meteorological forecasting.
Medium-to-long-term climate prediction
Medium-to-long-term climate predictions span decadal periods and extend beyond immediate meteorological events to encompass broader macroscopic elements, including long-term climate change trends, temperature fluctuations, and variations in precipitation levels. These predictions are important in various sectors, from environmental policy planning to infrastructure development and agricultural projections. Over time, forecasting methods have evolved from traditional climate models to advanced computational techniques. Popular machine learning models in this context are Conditional Generative Forecasting (CGF), CycleGANs, and DeepESD.
CGF outperforms traditional climate models in accuracy, whether using simulated or real data, demonstrating its enhanced predictive capability. CycleGANs improve the precision of long-term climate estimations, showcasing the effectiveness of deep-learning methodologies. DeepESD maintains higher consistency with observed data, indicating its potential for accurate climate predictions.
Machine learning excels at processing extensive datasets quickly and providing accurate short-term weather forecasts. Data scarcity, especially for historical meteorological data around 1979, presents a challenge for applying machine learning to climate prediction. Complex climate change effects influenced by numerous variables make it difficult for machine learning to capture their correlation. Climate systems' non-linear interactions and the sheer number of variables add to the complexity.
In summary, the authors explored the transformative impact of machine learning in meteorological forecasting, uniquely combining short-term weather forecasting with medium- and long-term climate predictions, covering 20 models and introducing eight leading models in depth. The current study illuminates operational mechanisms, serving as a model selection reference. It also highlights challenges, such as limited chronological season data, and suggests future directions, such as data simulation and physics-based constraints.