Thursday, 31 August 2023: 2:45 PM
Great Lakes A (Hyatt Regency Minneapolis)
Since 2021, Latent Heat Nudging (LHN) has been used operationally in Canada’s High Resolution (2.5km) Deterministic Prediction System (HRDPS). Research on this system has demonstrated that a key component of the LHN implementation is the use of pre-defined “idealized” latent heating profiles to warm (and moisten) the modeled atmosphere in locations where precipitation is observed but not simulated. This work investigates whether better forecasts can be obtained using “flow dependent” latent heating profiles inferred from radar observations using machine learning.
In models using explicit microphysics, it is not straightforward to infer latent heating using only information from observed precipitation. Typically, latent heat release is strongest somewhere above and “ahead” of precipitation such that the relationship between latent heating and precipitation rates displays a large amount of scatter.
In this work, a deep RESNET model is trained to infer latent heating profiles using only knowledge of precipitation rates at the surface. Because of limited storage and computing power, it is currently not possible to directly infer 3D fields of latent heating based on 2D precipitation. A number of strategies are thus put forward to make the problem tractable:
First, precipitation rates during a 30 minutes interval are considered to capture how the precipitation evolves and help the inference of latent heating. Second, the HRDPS domain is decomposed in small local neighborhoods surrounding precipitation. This greatly reduces the dimensionality of the problem and allows for large training datasets. Third, nowcasting techniques are used to compute the motion vectors of precipitation. The local neighborhoods are then rotated in the direction of these these motion vectors so that the machine learning algorithm always “sees” precipitation evolving in the same direction.
Training data comes from two months of forecasts initialized every 12h over a North-American domain. More than twelve millions latent heating profiles are thus available. Verification against a validation dataset shows that the latent heating profiles obtained with machine learning greatly outperform the pre-defined profiles currently used in operations.
At the moment of performing assimilation, latent heating profiles are inferred from radar observations of precipitation. The results of assimilation experiments performed using these profiles will be presented.
In models using explicit microphysics, it is not straightforward to infer latent heating using only information from observed precipitation. Typically, latent heat release is strongest somewhere above and “ahead” of precipitation such that the relationship between latent heating and precipitation rates displays a large amount of scatter.
In this work, a deep RESNET model is trained to infer latent heating profiles using only knowledge of precipitation rates at the surface. Because of limited storage and computing power, it is currently not possible to directly infer 3D fields of latent heating based on 2D precipitation. A number of strategies are thus put forward to make the problem tractable:
First, precipitation rates during a 30 minutes interval are considered to capture how the precipitation evolves and help the inference of latent heating. Second, the HRDPS domain is decomposed in small local neighborhoods surrounding precipitation. This greatly reduces the dimensionality of the problem and allows for large training datasets. Third, nowcasting techniques are used to compute the motion vectors of precipitation. The local neighborhoods are then rotated in the direction of these these motion vectors so that the machine learning algorithm always “sees” precipitation evolving in the same direction.
Training data comes from two months of forecasts initialized every 12h over a North-American domain. More than twelve millions latent heating profiles are thus available. Verification against a validation dataset shows that the latent heating profiles obtained with machine learning greatly outperform the pre-defined profiles currently used in operations.
At the moment of performing assimilation, latent heating profiles are inferred from radar observations of precipitation. The results of assimilation experiments performed using these profiles will be presented.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner