Improving deep learning precipitation nowcasting by using prior knowledge
This is an incremental improvement for operational meteorology, as it addresses usability issues in precipitation forecasting without enhancing predictive performance.
The paper tackled the problem of deep learning precipitation nowcasting lacking explainability and high-frequency features due to mean error loss optimization, by integrating a hand-engineered advection-diffusion equation into a PhyCell within a PhyDNet model, but found that the model's prediction capabilities remained unchanged.
Deep learning methods dominate short-term high-resolution precipitation nowcasting in terms of prediction error. However, their operational usability is limited by difficulties explaining dynamics behind the predictions, which are smoothed out and missing the high-frequency features due to optimizing for mean error loss functions. We experiment with hand-engineering of the advection-diffusion differential equation into a PhyCell to introduce more accurate physical prior to a PhyDNet model that disentangles physical and residual dynamics. Results indicate that while PhyCell can learn the intended dynamics, training of PhyDNet remains driven by loss optimization, resulting in a model with the same prediction capabilities.