Radar Forecast Model – Deep Learning

Long Short-term Memory

LSTM is good at processing time series data.

The model structure can be decomposed by two networks, one is encoder, and the other is predictor.

Model configurations:

Model Name	hidden layers	loss function	# training samples	tsize	# epochs	training process	Date
newest-5_8	16	built-in RSME	400	10	500	Fig.1	2019.5.8
newest-5-13	16	RMSE+FAR	400	10	500	Fig.2	2019.5.13

Fig.1 Training loss on 5.8
Fig.2 Comparison of predction and observation

Wind Effect

To add wind effect in, one way is to customize the loss function by comparing LSTM modeled image at t and predicted image with Lagrangian plus wind at t. In this way, we drive the LSTM to the correct advection direction.

GAN

GRU

We select one event from our radar data shown below

Kalman Filter

Semi-Lagrangian movement

we create several wind field to test the robustness of this algorithm with radar movement, the first goes with uniform speed gradient wind vector.

The basic idea of semi-lagrangian goes here ... the stability analysis of semi-lagrangian algorithm goes here...

unsolved problem:

1. when cloud converges due to conteract wind vector, it will collapse somehow.

Optical Flow

The results are like this ...

Left panel is the ground truth, while the right is prediction.

This is pretty messy because the optical flow assumes that all pixels have the same intensity which is not the case in radar images because localized growth and decays may be generated.

Here is one implementation with optical flow Motion Vectors in Weather Radars

FlowNet

Gaussian Mixture Models

A full description and simple code about how Gaussian Mixture Model (GMM) works can be found at Jupyter Notebook

Kalman Filter

To-do List

• Test on circular wind vector (Semi-lagrangian)