The Spring Festival is approaching, and many people must have embarked on the journey home, but due to the sudden cooling in the past two days, many friends have encountered extreme weather, and even delays or cancellations of planes and trains. This sudden onset of extreme weather has turned weather forecasting into a toy.
But now, that may be about to change. At present, there are many large language models that already have the ability to achieve high accuracy weather**. Compared to traditional weather** techniques, these are called:Large AI meteorological models(Large AI Weather Forecast Model, LWMS) has better results in extreme weather**.
Pangu-weatherPangu-Weather is a weather** system launched by Huawei that uses deep learning technology to improve the accuracy of weather forecasting.
There are two key strategies for Pangea Weather to improve accuracy, including a custom 3D Earth-Specific Transformer (3DEST) architecture that formats altitude information as cube dataIn addition, the research team also designed a hierarchical time aggregation algorithm to mitigate the cumulative forecast error. Studies have shown that Pangea weather shows great advantages in short-to-medium-range forecasts (i.e., forecasts from one hour to one week). Moreover, the system also supports extreme weather forecasting and multiple ensemble forecasts.
In terms of forecasting extreme weather events, Pangu Meteorology has shown significant advantages. On the tracking of 88 named tropical cyclones tested in 2018, the 3-day and 5-day mean direct position errors were lower than those of ECMWF-HRES, which were 120., respectively29 km and 19565 km, better than the latter's 16228 km and 27210 km.
However, there are still some limitations and room for improvement in Pangu meteorology. Since the training data is based on ERA5 reanalysis data, there is a large bias, especially in terms of intensity**. Secondly, the Pangu meteorological training process requires a lot of computing resources, and the cost of training and regular maintenance is extremely high. Moreover, the generalization ability of Pangea meteorology in unseen data or different climatic conditions needs to be verified.
graphcastGraphcast is a graph neural network-based weather model developed by Google DeepMind that is capable of handling complex spatial dependencies and providing accurate weather forecasts on a global scale.
Graphcast takes advantage of GNNS in dealing with complex spatial dependencies to improve the accuracy of air forecasting. In the first process, Graphcast will perform data preprocessing in Xi'an, convert the meteorological data into a map structure, each node represents the geographical location, and uses edges to represent the spatial relationship; After that, GraphCast will extract features for each node in the graph, such as temperature, humidity, wind speed, etc.; Then, the graph neural network modeling is used to update the content of each node by aggregating the information of neighbor nodes, and the complex spatial dependencies are learned. Then, the time series**, the meteorological data variables of each node in the future period; Finally, the results are processed and evaluated to improve the readability and accuracy of the forecast.
There are similarly problematic graphcasts. The performance of Graphcast relies heavily on high-quality and comprehensive meteorological data, and missing or inaccurate data may affect the model's best results. In addition, despite Graphcast's ability to handle complex spatial dependencies, there may still be challenges in terms of real-time updates and rapid response to weather changes.
FengwuFengwu is a medium-range weather forecasting system for the world, which is released by the Shanghai Artificial Intelligence Laboratory, the University of Science and Technology of China, Shanghai Jiao Tong University, Nanjing University of Information Science and Technology, the Institute of Atmospheric Physics of the Chinese Academy of Sciences, and the Shanghai Central Meteorological Observatory. It employs a deep learning architecture of multimodal and multi-task learning, including model-specific encoders-decoders and cross-modal fusion transformers. These components learn under the supervision of uncertainty loss, balancing the optimization of different ** devices in a regionally adaptive manner. Fengwu also introduced a replay buffer mechanism that improved long-term performance by storing the results of previous optimization iterations and using them as input to the current model.
The biggest feature of Fengwu is that it uses multi-modal and multi-task learning capabilities to improve the effective forecast of global weather to 1075 days. It has been trained for 39 years on ERA5 reanalysis data and is able to accurately simulate atmospheric dynamics and future land and atmospheric states at 37 vertical levels.
Performance evaluations have shown that the wind crow outperforms Graphcast on most targets, such as reducing the root mean square error (RMSE) of the 10-day global lead of the Z500** from 733 to 651 ms. It costs just 100 milliseconds for inference on NVIDIA Tesla A600 hardware, which is less computationally expensive in terms of training and inference compared to Graphcast.
Like other similar models, the wind crow faces data and cost challenges, and as climate patterns change, it needs to be updated regularly to maintain accuracy, which requires ongoing resource investment and expertise.
climatenetClimateNet is a model that uses convolutional neural networks to identify climate features from satellite imagery and aims to solve the challenges of identifying, detecting, and locating extreme weather events in the field of weather and climate science. ClimateNet creates datasets through the ClimateContours tool, which allows experts to label climate events and optimize based on LabelMe. Its research team has carried out several annotation activities, resulting in hundreds of expert-annotated snapshots of climate data, forming the ClimateNet dataset, and conducting strict quality control.
ClimateNet uses the DeepLabV3+ architecture to train a deep learning model to achieve pixel-level segmentation of climate data. The model was trained and tested in two different climate model scenarios and demonstrated better performance than the heuristic-based model. In addition, the use case of ClimateNet shows how its segmentation results can be used for conditional precipitation analysis, which can help to understand the impact of climate change on extreme weather events.
Although ClimateNet has made significant progress, it is still in the research phase and has not yet been put into use due to limited training data and the need for continuous updating and maintenance.
FuxiFuxi is a machine learning-based weather forecasting system capable of providing a 15-day global weather forecast with a 6-hour temporal resolution and 0Spatial resolution of 25°. The system was developed based on the ERA5 reanalysis dataset of the European Centre for Medium-Range Weather Forecasts (ECMWF) and trained on data equivalent to 39 years.
Fuxi adopts a novel cascaded machine learning model architecture, which consists of three pre-trained Fuxi models, which are optimized for the forecast time windows of 0-5 days, 5-10 days, and 10-15 days, respectively. These models generate a complete 15-day forecast in a cascading fashion. The performance of Fuxi in the 15-day forecast is comparable to the average of the ECMWF ensemble, which significantly reduces the cumulative error and improves the accuracy of the long-term forecast. To deal with the uncertainty of weather forecasting, Fuxi also provides a collective forecast with 50 members.
The biggest problem with Fuxi's current prediction time is that its performance will decline after 9 days, and Fuxi's best game relies on ECMWF's ERA5 reanalysis dataset, which is completely independent of traditional numerical weather prediction models, which makes it difficult to put into practical applications at present.
In addition to these large, mature AI weather** models, a number of models are being trained to reduce the impact of extreme weather. Of course, due to problems such as cost, maintenance, and accuracy, it is still difficult to achieve comprehensive accuracy in the current weather model, but I believe that with the development of technology, more and more AI models will be put into use in the future, so that the road home for the New Year will be "unimpeded".