<p>It is generally accepted that weather forecasts contain errors due to the chaotic nature of the atmosphere. Regression models, such as neural networks, are traditionally trained to minimize the pixel-wise difference between their predictions and ground truth. The major shortcoming of these models is that they express uncertainty about prediction with blurring, especially for longer prediction lead times. One way to tackle this issue is to use a generative adversarial network, which learns what real precipitation should look like during training. Coupled with a loss, such as Mean Squared or Mean Absolute Error, these networks can produce highly accurate and realistic nowcasts. As there is an inherent randomness in those networks, they allow to be sampled from, just like ensemble models, and various probabilistic metrics can be calculated from the samples. In this work, we have designed a physically-constrained generative adversarial network for radar reflectivity prediction. We compare this network to one without physical restraints and show that it predicts events with higher accuracy and shows much less variance among its samples. Furthermore, we explore fine-tuning the network to the prediction of severe weather events, as an accurate prediction of these benefits both automated warning systems and forecasters.</p>

<p>The standard for weather radar nowcasting in the Central Europe region is the COTREC extrapolation method. We propose a recurrent neural network based on the PredRNN architecture, which outperforms the COTREC 60 minutes predictions by a significant margin.</p><p>Nowcasting, as a complement to numerical weather predictions, is a well-known concept. However, the increasing speed of information flow in our society today creates an opportunity for its effective implementation. Methods currently used for these predictions are primarily based on the optical flow and are struggling in the prediction of the development of the echo shape and intensity.</p><p>In this work, we are benefiting from a data-driven approach and building on the advances in the capabilities of neural networks for computer vision. We define the prediction task as an extrapolation of sequences of the latest weather radar echo measurements. To capture the spatiotemporal behaviour of rainfall and storms correctly, we propose the use of a recurrent neural network using a combination of long short term memory (LSTM) techniques with convolutional neural networks (CNN). Our approach is applicable to any geographical area, radar network resolution and refresh rate.</p><p>We conducted the experiments comparing predictions for 10 to 60 minutes into the future with the Critical Success Index, which evaluates the spatial accuracy of the predicted echo and Mean Squared Error. Our neural network model has been trained with three years of rainfall data captured by weather radars over the Czech Republic. Results for our bordered testing domain show that our method achieves comparable or better scores than both COTREC and optical flow extrapolation methods available in the open-source pySTEPS and rainymotion libraries.</p><p>With our work, we aim to contribute to the nowcasting research in general and create another source of short-time predictions for both experts and the general public.</p>

<p>During last year's summer storm season, we have introduced a precipitation nowcasting neural network MWNet and deployed it to operational use. The network tackles the nowcasting problem as a sequence to sequence prediction of radar echo, emphasizing high resolution and accuracy. We have conducted two quantitative experiments comparing MWNet 60 min forecasts to other available precipitation nowcasting models, using the metrics CSI and MSE. Both evaluations, over the domain of Denmark for years 2018 - 2020 and over the Czech Republic for the summer storm season of 2021, concluded in favor of our approach. However, we aim to improve MWNet capabilities further by focusing on severe weather nowcasting, the physical soundness of the predictions, and lead times longer than 60 min. Building on the advances in deep learning and its use in spatio-temporal forecasting, MWNet is based on the idea of disentangling physical dynamics from the residual factors. In this contribution, we consider improvements to the physical part of the network, its incorporation into the whole model, and the loss function used during training. Mainly, we are exploring the effect of implementing non-linear partial differential equations into the physical part, with various levels of hand-engineering equation terms. We analyze the impact on the dynamics learned by each part of the network and prediction quality for each setting. MWNet v1.2, based on the proposed architecture, will be operationally used and evaluated by meteorologists in Meteopress during the summer of 2022. This work aims to contribute to bridging the gap between machine learning and physical modeling in weather forecasting, alongside improving precipitation prediction.</p>