Developing weather prediction method from data-driven approach
To address the enormous computational demands that will limit control methods targeting weather, this study develops weather prediction methods using data-driven approaches such as deep learning. By implementing weather prediction by data-driven approaches such as deep learning, we can approximate tangent linear models and connect them to control methods like Model Predictive Control. We train surrogate models by using regional weather predictive system and operational prediction data from the Japan Meteorological Agency, enabling us to achieve large ensemble predictions necessary for Model Predictive Control with a computationally less demanding surrogate model.
Item 2-1Weather prediction computation using latent variable model
Principal investigator: Daisuke Matsuoka
Outline
This study develops weather surrogate models using deep learning. We construct surrogate models by using regional weather predictive system (SCALE-LETKF) and operational prediction data from the Japan Meteorological Agency, enabling us to achieve large ensemble predictions necessary for Model Predictive Control with a computationally less demanding surrogate model. Specifically, we construct latent variable models such as variational autoencoders that take current weather information as input to predict future weather conditions. We create an ensemble by perturbing latent variables.
Methods
Based on deep learning architecture such as Vision Transformer, we construct a time series predictive model, where latent variables are treated as probability distributions. We develop techniques to generate large ensembles at low cost by applying appropriate perturbations in latent variable space, particularly during the occurrence of extreme phenomena such as torrential heavy rainfall.
At the start of the research, no specific numerical targets will be established, because the scale of heavy rainfall over the sea and the reduction in rainfall over land for disaster mitigation are expected to vary depending on the specific disaster events. By the FY 2024, we select cases with high potential to generate heavy rainfall over the sea based on the results of Item 5-2. Flood inundation and economic damage estimates for the selected cases will be calculated by Item 8-1 and 8-2 to investigate the extent to which the heavy rainfall generation over the sea and rainfall reduction over land are required. With the target values for rainfall reduction, effective operations for generating heavy rainfall over the sea using numerical models will be identified by Item 5-1 and 5-3. We early set the phenomena, spatiotemporal scale, and spatiotemporal resolutions to be predicted in our study, in collaboration with other studies.
We also collaborate with the research group from the Meteorological Research Institute of the Japan Meteorological Agency as research participants. Specifically, we collaborate with the Fourth Research Division of the Meteorological Observation Research Department at the Meteorological Research Institute to develop latent variable representations of weather information using deep learning. In addition, we receive data on ensemble prediction information output from regional weather predictive models from them, exchanging techniques and insights regarding ensemble generation using deep learning.
Importance
The process-driven predictive approaches typified by numerical weather models generally have high computational demands, requiring enormous computational cost to demonstrate the vast amount of model predictive control aimed at weather control. By substituting weather prediction with a data-driven approach based on deep learning, it becomes possible to reduce the costs of generating large ensembles necessary for Model Predictive Control.
Expected problems and solutions
Generally, the construction of surrogate models (model learning using large datasets) requires the use of large-scale parallel GPU computer for a long time, but now it is not easy to prepare such environments. We proceed to apply for the use of domestic GPU equipped supercomputers and develop low-computational-load surrogate models using techniques such as super-resolution, aiming to address the challenges related to computational environments.
Members
Item 2-2Developing weather predictive learning with attention mechanism
Principal investigator: Hiroshi Kera
Outline
This study develops weather surrogate models using deep learning. We train surrogate models by using regional weather predictive system (SCALE-LETKF) and operational prediction data from the Japan Meteorological Agency, enabling us to achieve large ensemble predictions necessary for Model Predictive Control with a computationally less demanding surrogate model. Specifically, we develop a surrogate model for regional weather predictive computation that can execute rapidly by using deep learning with attention mechanisms, such as Microsoft’s ClimaX. We utilize techniques such as adversarial perturbations to identify key (effective) intervention points that can significantly change weather prediction.
Methods
We investigate and design deep learning models with attention mechanisms, such as Transformer, selecting models and training framework capable of effectively learning from high-dimensional weather data. We start our study using reanalysis data from ECMWF and the Japan Meteorological Agency as training data. Then we expand to various open datasets, such as operational weather prediction data from the Japan Meteorological Agency.
At the start of the research, no specific numerical targets will be established, because the scale of heavy rainfall over the sea and the reduction in rainfall over land for disaster mitigation are expected to vary depending on the specific disaster events. By the FY 2024, we select cases with high potential to generate heavy rainfall over the sea based on the results of Item 5-2. Flood inundation and economic damage estimates for the selected cases will be calculated by Item 8-1 and 8-2 to investigate the extent to which the heavy rainfall generation over the sea and rainfall reduction over land are required. With the target values for rainfall reduction, effective operations for generating heavy rainfall over the sea using numerical models will be identified by Item 5-1 and 5-3. We early set the phenomena, spatiotemporal scale, and spatiotemporal resolutions to be predicted in our study, in collaboration with other studies.
Importance
The surrogate model for weather prediction is essential for computationally demonstrating the project’s goal of ‘reducing heavy rainfall damage through heavy rainfall generation over the sea and rainfall reduction over land. In particular, weather predictive surrogate model using deep learning is expected to be an effective surrogate model due to its numerically less demanding and high performance. Since small changes in input can significantly alter the output of deep learning models and can be computed efficiently, this surrogate model can also be utilized to verify the effects of interventions, rather than just for predictions.
Expected problems and solutions
To effectively learn from large and high-dimensional weather data, it is essential to carefully define the components and design of the deep learning model. While attention mechanisms are very powerful, they also have high learning costs. Therefore, we use and design efficient models that incorporate appropriate sparsification and recurrent structures. The outputs of surrogate models obtained through simple learning tend to deviate from real phenomena. We determine the loss function and regularization thorough experiments to mitigate this deviation. In calculating small interventions that lead to significant output changes, we incorporate insights from research on adversarial perturbations while including regularization or constraints to avoid unrealistic perturbations.
Members
