Deep generative modeling is able to generate highly realistic atmospheric fields, one prominent example being precipitation. So far, almost all studies have used generative adversarial networks (GANs) for this purpose, but recent progress in machine learning research has had a new class of methods called diffusion models replace GANs in many applications. Diffusion models have been often shown to be able to generate a wider variety of samples than GANs, suggesting that they might be able to better capture uncertainty in applications such as weather and climate where quantifying it is important.

In this presentation, we describe our research on using diffusion models for short-term prediction (nowcasting) of precipitation fields. We adapt the latent diffusion model used by Stable Diffusion (Rombach et al. 2022) to the this problem, predicting precipitation up to 3 hours ahead to the future at 5-min temporal resolution and 1-km horizontal resolution. Predictions can be produced as an ensemble where each member represent a possible future evolution of the precipitation field.

We show that our model:

• generates highly realistic precipitation fields that are consistent with the past precipitation used as input.
• outperforms the state-of-the-art GAN-based Deep Generative Models of Rainfall (DGMR) model by most relevant metrics.
• performs particularly well at representing the uncertainty of its own predictions, as shown by uncertainty quantification methods developed for ensemble forecast verification.

Therefore, it appears that diffusion models are indeed suitable for generative modeling of precipitation fields with highly realistic representation of uncertainty. Our model architecture also permits multiple inputs data sources to be combined, in particular allowing seamless generative predictions to be made by exploiting observations and numerical weather predictions.

How to cite: Leinonen, J., Hamann, U., and Germann, U.: Latent diffusion models for generative nowcasting and uncertainty quantification of precipitation fields, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-9531, https://doi.org/10.5194/egusphere-egu23-9531, 2023.

This work introduces a novel deep-learning method for generating realistic ensembles nowcast of radar-based precipitation at a five-minute time resolution for the next 60 minutes and longer.

The proposed method is composed of a combination of two models: the first model is trained to compress and decompress the spatial domain into and from a discrete representation (tokens), while the second model evolves the compressed representation over time. Specifically, the compression and decompression model is based on a combination of a Quantized Variational Autoencoder with a Generative Adversarial Network, while the prediction over time leverages a Generative Pretrained Transformer (GPT) architecture.

This separation of concerns (discretized spatial compression/decompression and temporal extrapolation) adds several desirable features not present in more commonly used deep learning methods based on recurrent/convolutional deep learning architectures: 

• transformer output probabilities can be leveraged to generate ensemble/probabilistic forecasts (without the need of injecting noise)
• the discretized spatial representation can be used to characterize each token, adding interpretability and explainability to the model
• the combination of transformer probabilities and token characterization can be used at inference time for forecasts conditioning based on external factors (e.g. NWP forecast output)

The presented architecture is trained and tested on a 7-year radar dataset of reflectivity composites of the Emilia-Romagna Region, Italy. The method is then applied at two different scales: regional, over Emilia-Romagna, and national, on the entire Italian domain, showing the adaptability of the approach to multiple spatial domains. We will present the performance of this model for both deterministic and ensemble settings by comparing it with respect to other commonly used extrapolation and deep learning methods.

How to cite: Franch, G., Tomasi, E., Poli, V., Cardinali, C., Cristoforetti, M., and Alberoni, P. P.: Ensemble precipitation nowcasting by combination of generative and transformer deep learning models, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-15153, https://doi.org/10.5194/egusphere-egu23-15153, 2023.

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.

How to cite: Choma, M., Murín, M., Bartel, J., Troller, M., and Najman, M.: Probabilistic Precipitation Nowcasting with Physically-Constrained GANs, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-15514, https://doi.org/10.5194/egusphere-egu23-15514, 2023.

Deep learning has been rapidly adopted in short-term precipitation prediction, such as simulating precipitation movement and predicting extreme weather events. Recently, generative adversarial neural networks (GANs) have been shown to be effective at dealing with field smoothing with increasing lead time. Several studies (Jing et al., 2019; Ravuri et al., 2021) demonstrated the potential of GAN by solving spatial smoothing problems and demonstrating reliable predictive performance. However, despite promising results from GANs, unbalanced datasets and human annotations can limit the predictive ability of deep learning and induce biased results. In addition, precipitation is a complex process that depends on various factors. Thus, approximating the model into a single latent space is a challenge, and furthermore, there is a risk of mode collapse. This study introduces an algorithm for predicting precipitation by clustering precipitation types using self-supervised learning (SSL) and estimating rainfall distribution according to precipitation types. First, we derive precipitation-type labels by self-clustering a generator that is a multi-layer ConvGRU. And then, we predict six-hour precipitation based on the gaussian distribution of each type. SSL improves the performance of precipitation forecasting based on type-specific representation learning through adaptive sampling in latent space. The proposed methodology was verified using hybrid surface rainfall (HSR) dataset at a spatial resolution of 500m with a resolution of 2,305 (longitude) × 2,881 (latitude) and a temporal resolution of 5 min. The images consist of 256×256 pixels from scaling down to a resolution of 4 km and are extracted at 30-minute intervals. Experimental results show that our method outperforms a state-of-the-art method on a six-hour prediction basis with a mean squared error and critical success index on unseen datasets. Also, the proposed algorithm can predict various precipitation types without spatial smoothing.

How to cite: An, S., Oh, T.-J., Na, I., Jang, J., Park, W., and Kim, J.: GAN-based forecasting model via self-adaptive clustering approach, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-4735, https://doi.org/10.5194/egusphere-egu23-4735, 2023.

    Due to the threat of extreme rainfall associated with mesoscale convective systems and summer afternoon thunderstorms, very short-term quantitative precipitation forecasting during 0−3 h is critical in Taiwan. In this study, deep learning models are developed for high-resolution quantitative precipitation nowcasting in Taiwan up to 3 h ahead. The baseline model based on the convolutional recurrent neural network is trained with a dataset containing radar reflectivity and rain rates at a granularity of 10 min. As previous works tend to produce overprediction in low-rainfall regions, the currently proposed model is improved and further driven by highly related heterogeneous weather data, including visible channel satellite observation, environmental winds, and environmental thermo-dynamical profiles. Note that an innovative “PONI module” is added to the deep learning model to integrate a variety of heterogeneous data with various spatial and temporal characteristics. Moreover, model performance is evaluated from statistical and spatial rescaling perspectives represented by R =  R_avg + R'_, where R denotes original rainfall, R_avg and R' are spatial moving averages and the values deviated from R_avg, respectively. Statistical verification shows that the R_avg of the new model outperforms the previous model, while the performance of R' is comparable. The new model integrated with heterogeneous data selected upon domain knowledge can restrain the nowcasts that overestimate in low-rainfall regions. Last but not least, quasi-operational verifications against other state-of-the-art techniques in Taiwan Central Weather Bureau are presented as follows: (1) the CSI of the first-hour prediction from the deep learning model is comparable with QPESUMS-QPF and better than RWRF and iTeen. (2) 3h ahead prediction CSI of RWRF and iTeen are inferior to the performance of deep learning model owing to their misprediction of rainfall regions. The deep learning model can accurately predict medium and extreme amounts of precipitation at a fraction of the computational cost.

How to cite: Chen, D.-Y., Chang, C.-T., and Chen, B.-F.: Precipitation Nowcasting Based on an Optimized Deep Learning Model Trained with Heterogeneous Weather Data, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11043, https://doi.org/10.5194/egusphere-egu23-11043, 2023.

Quantitative precipitation nowcasting (QPN) is crucial for forecasting precipitation within the next several hours (generally up to 6) to prevent substantial socioeconomic damage. In general, ground radar data has been widely employed in QPN due to its high spatial-temporal resolution and more precise precipitation estimation than satellite. With the remarkable success of deep learning (DL), recent QPN studies have actively adopted DL using radar data. Although these studies yielded high skill scores in forecasting precipitation areas with a weak intensity (about 1 mm/h), they failed to effectively simulate the horizontal movement of precipitation areas and showed poor ability in forecasting precipitation with stronger intensities. In addition, despite the fact that the skill score is highly dependent on the characteristics of each precipitation event, there was a lack of evaluation over various precipitation cases. From the motivation that there can be room for improving QPN using the advanced DL model in video prediction, this study suggests the QPN model based on simple yet better video prediction (SimVP), which is a state-of-the-art DL model. We trained the SimVP model using radar data in South Korea from June to September (JJAS) for the period of 2019-2022, which includes the summer and early fall. In terms of the critical score index (CSI) with a lead time of 120 minutes (0.46, 0.23, and 0.09 for 1, 5, and 10 mm/h thresholds, respectively), the proposed model showed significant improvement over the existing DL models based on an evaluation from JJAS 2022. Considering different precipitation conditions, three case studies were conducted for heavy rainfall, typhoons, and fast-moving narrow convection events. The suggested model showed comparable or the highest CSI in 120 min with a 1 mm/h threshold in all cases, demonstrating robust performance (0.49, 0.69, and 0.29 for heavy rainfall, typhoon, and narrow convection, respectively). Qualitative evaluation of the proposed model also showed better results in terms of horizontal displacement movement and less underestimation than the other models. In addition, we further explored the possibility of real-time learning (RTL) with newly added radar data. By repeatedly optimizing DL model for currently facing precipitation events, RTL contributed to deep learning models predicting results more similar to actual radar patterns. It is expected that the proposed SimVP and RTL would serve as a new baseline for DL-based QPN due to their ease of implementation and enhanced performance. 

How to cite: Han, D., Choo, M., and Im, J.: A data-driven precipitation nowcasting framework using advanced deep learning model for video prediction and real-time learning approaches, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-13644, https://doi.org/10.5194/egusphere-egu23-13644, 2023.

On 26 November 2022, an extreme rainfall event occurred over Ischia Island (Italy). It triggered a mudflow that swept over Casamicciola Terme town and caused 12 victims. Based on available rainfall data from 4 rain-gauge stations over the island, the precipitation values registered during the event were higher than the annual maxima values of the previous 15 years. With regards to 1 and 24 hours, the rain-gauge stations measured values between 40.6 and 57.6 mm, and between 145.4 and 176.8 mm, respectively. Since one of the main challenges during these phenomena is predicting rainfall sufficiently in advance in order to allow water managers to take action (issue warnings or real-time control), this study investigates how much time before the peak - or threshold exceedance - a machine learning model is able to capture the peak - or threshold exceedance. A model that predicts rainfall intervals and the corresponding probability of occurrence for lead times from 10 minutes to 6 hours is proposed. The model employs cumulative rainfall depths from recording stations in an area of 50 km radius from the Ischia Island as inputs for a Feed Forward Neural Network to nowcast rainfall in the 4 rain-gauges over the study area. Based on almost 400 rain events observed during years 2009-2022, 24 machine learning models were independently trained for each rain-gauge and each of the 6 lead-times - 10, 30, 60, 120, 180 and 360 minutes. The performance of each model was evaluated and compared using different metrics, both continuous (RMSE and MAE) and categorical (POD and FAR). In addition, the Eulerian Persistence (EP) was considered as a benchmark model. The rainfall nowcasts showed encouraging results. Even though for convective rain events the potential lead-time is short, the models produced consistent nowcasts for lead-times up to 2 hours. With probabilities of almost 90%, the thresholds exceedance was forecasted up to 1 hour before. As expected, predictive accuracy and probabilities gradually decreased as the lead-time increased, according to physically based models. Moreover, the proposed models outperformed the benchmark EP for all the lead-times and performance criteria. Results confirmed that the use of cumulative rainfall depths for precipitation nowcasting made this approach a promising tool for nowcasting purposes, and his flexibility and conceptual simplicity resulted in a rapid, easily replicable and convenient nowcasting approach. To conclude, the proposed models enhanced a first identification of critical thresholds, which should be further analysed in order to achieve a better, complementary understanding of the occurring phenomenon. 

Keywords: Precipitation nowcasting; Multi-step predictions; Rain-gauge measurements; Pattern recognition; Feed forward neural networks; Cumulative rainfall fields.

How to cite: Pirone, D., Del Giudice, G., and Pianese, D.: Machine Learning models for probabilistic rainfall nowcasting applied to a case study in Italy: the extreme rainfall event on 26 November 2022 over Casamicciola town, Ischia Island., EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-14545, https://doi.org/10.5194/egusphere-egu23-14545, 2023.

Extreme hydrological events, which are highly concerned by local governments, hydraulic units and hazard response centers due to their potential to bring heavy rainfall and cause serious floods, have frequently occurred and impacts on Taiwan urban area in recent years under the circumstance of climate change and global warming. The frequent occurrence of high intense storm always leads to flooding-related disasters within a short period, which makes rainfall monitoring a disaster prevention. Therefore, this study utilizes Long Short-Term Memory Neural Networks (LSTM) and Back Propagation Neural Networks (BPNN) to extract the characteristics of radar observations and forecast rainfall with time 1-step-ahead to 6-step-ahead (T+1~T+6) in Taiwan’s capital, Taipei City. The data collection was included in the Shulin dual-polarization radar (RCSL) observations, such as differential phase shift, specific differential phase, reflectivity and doppler radial wind field, and rain gauge data from May 2021 to November 2021 in the Taipei City. With a view to capturing the movement of hydrometeors continually changes within the time step, an algorithm which can calculate velocity and direction of specific hydrometeors on two-dimensional matrix were developed and applied to simulate location of the specific hydrometeors on n-step-ahead (T+n). Finally, the rainfall forecast can be achieved by using the simulated location of specific hydrometeors and its physical properties from radar observations as input data to fit rainfall from the gauge. This study aims to investigate the relationship between short-duration rainfall and radar observations by artificial neural network (ANN), and forecast the rainfall  within a short period.

Keywords: Dual-Polarization Radar; Rainfall Estimation; Artificial Intelligence (AI), Artificial neural network (ANN); Long Short-Term Memory Neural Networks(LSTM)

How to cite: Lin, J.-L., Hsu, C.-Y., and Chang, L.-C.: Improving Dual-Polarization Radar-based rainfall estimation using Long Short-Term Memory Neural Networks, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-14709, https://doi.org/10.5194/egusphere-egu23-14709, 2023.

Severe convective weather events, such as hail, lightning and heavy rainfall pose a great threat to humans and cause a considerable amount of economic damage. Nowcasting convective storms can provide precise and timely warnings and, thus, mitigate the impact of these storms. Dual-polarization weather radars are a crucial source of information for nowcasting severe convective events. These radars provide important information about the microphysics of the convective systems, on top of the rainfall rate and vertical structure of the reflectivity. Nevertheless, polarimetric variables, which can provide additional information about the size, shape and orientation of particles, are often not considered in nowcasting.

This work presents the importance of polarimetric variables as an additional data source for nowcasting thunderstorm hazards using machine learning, compared to using radar reflectivity alone. We add these data to the neural network architecture of Leinonen et al. 2022 (Seamless lightning nowcasting with recurrent-convolutional deep learning), which uses convolutional and recurrent layers and analyzes inputs from multiple data sources simultaneously. This network has a common framework, which enables nowcasting of hail, lightning and heavy rainfall for lead times up to 60 min with a 5 min resolution. The study area is covered by the Swiss operational radar network, which consists of five operational polarimetric C-band radars. In addition, we analyze the contribution of quality indices as an additional information source, which takes the uncertainty of the radar observations throughout the complex mountainous terrain and scanning strategy in Switzerland into account. Results indicate that including polarimetric variables and quality indices improves the accuracy of nowcasting convective storms.

How to cite: Rombeek, N., Leinonen, J., and Hamann, U.: Exploiting radar polarimetry for nowcasting of convective hazards using deep learning, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-12127, https://doi.org/10.5194/egusphere-egu23-12127, 2023.

One major task of the National Meteorological and Hydrological Services (NMHS) is the provision of consistent and integrated forecasting products from minutes to several days ahead (seamless forecasting). The former EUMETNET (European Meteorological Services’ Network) project ASIST (Application oriented analysis and very short-range forecast environment) which started in 2015 focused on the nowcasting and very short range forecasting. Then, it was extended to the EUMETNET Nowcasting Programme (E-NWC) which started in 2019 and will last until the end of 2023 with focus on nowcasting and also on seamless prediction.

In this presentation, the main objectives of the E-NWC Programme will be introduced. E-NWC supports NMHS in sharing expertise, experiences and best practices for developing and implementing nowcasting, very short-range forecasting and seamless prediction systems. Key activities lie in the exchange of information and experiences with the users during e.g. the every two years European Nowcasting Conference and the strong cooperation with the World Meteorological Organization (WMO) and EUMETSAT, and in summarizing the relevant findings in project reports and joint peer-reviewed papers. Highlights of this contribution comprehend a few results from studies and surveys carried out recently.

How to cite: Schmid, F., Sivle, A., Agersten, S., Simon, A., and Atencia, A.: EUMETNET Nowcasting Programme, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-14544, https://doi.org/10.5194/egusphere-egu23-14544, 2023.

The UK Met Office is developing an open-source probability-based post-processing system called IMPROVER (Integrated Model Post-Processing and Verification) to fully exploit our convection permitting, hourly cycling ensemble forecasts.  Post-processed MOGREPS-UK model forecasts are blended with deterministic UKV model forecasts and data from the coarser resolution global ensemble, MOGREPS-G, to produce seamless probabilistic forecasts from now out to 7 days ahead. For precipitation, an extrapolation nowcast is also blended in at the start.

A majority of the post-processing within IMPROVER is performed on gridded forecasts, with site-specific forecasts extracted as a final step, helping to ensure consistency. IMPROVER delivers a wide range of probabilistic products to both operational meteorologists and as input to automated forecast production. The system achieved operational acceptance in spring 2022 and will be used in operational products from spring 2023.

Weather symbols provide the general public with a simple, pictorial view of the weather for a time of interest and include sun and cloud conditions, mist and fog, hail and lightning, and three phases of precipitation, both as showers or continuous, and light or heavy. This talk describes how a deterministic most-likely weather type code is generated using a decision tree approach from probabilistic multi model IMPROVER data for 1 hour, 3 hour and daytime periods that are consistent with each other. Recent work to make these weather codes representative of a time-window, rather than an instant, will be discussed. We will present some verification, comparing IMPROVER weather symbols and the current operational Met Office symbols with SYNOP present weather reports.

How to cite: Moseley, S., Ayliffe, B., and Evans, G.: Generating weather symbol data in IMPROVER, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-5837, https://doi.org/10.5194/egusphere-egu23-5837, 2023.

So far, plenty of efforts have been pursued on the numerical weather prediction (NWP). However, systematic errors could never be ignored in the output applications. To supply the numerical forecasts with higher accuracies, statistical postprocessing is often expected to correct systemic biases and has been one of the key components of the forecasting suites. Based on the NWP models and taking advantages of the raw stepwise pattern projection method (SPPM), the neighborhood pattern projection method (NPPM) is newly proposed to postprocess the model outputs and to improve forecast skills of daily maximum and minimum temperatures (Tmax and Tmin) over East Asia for short-term timescales, as well as the Kalman filter based pattern projection method (KFPPM) for longer-term forecasts. For the short-term lead times of 1–7 days, the SPPM is slightly inferior to the benchmark of decaying averaging method, while its insufficiency decreases with increasing lead times. The NPPM shows manifest superiority for all lead times, with the mean absolute errors of Tmax and Tmin decreased by ~0.7° and ~0.9°C, respectively. Advantages of the SPPM and NPPM are both mainly concentrated on the high-altitude areas such as the Tibetan Plateau, where the raw model outputs show the most conspicuous biases. As for longer-term forecasts at the subseasonal timescale, the NPPM effectively calibrates the temperature forecasts at the early stage. However, with the growing lead times, it shows speedily decreasing skills and can no longer produce positive adjustments over the areas outside the plateaus. By contrast, the KFPPM consistently outperforms the other calibrations and reduces the forecast errors by almost 1.0°C and 0.5°C for Tmax and Tmin, respectively, both retaining superiorities to the random climatology benchmark till the lead time of 24 days. The optimization of KFPPM maintains throughout the whole range of the subseasonal timescale, showing most conspicuous improvements distributed over the Tibetan Plateau and its surroundings. Case experiments further demonstrate the above-mentioned features and imply the potential capability of the NPPM and KFPPM in improving forecast skills and disaster preventions for extreme temperature events. Besides, compared with the initial SPPM, they not only produces more powerful forecast calibrations, but also provides more pragmatic calculations and greater potential economic benefits in practical applications.

How to cite: Zhu, S., Lyu, Y., and Zhi, X.: Calibrations of Surface Air Temperature Forecasts at Short- and Long-term Timescales Based on Statistical Pattern Projection Methods, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-11970, https://doi.org/10.5194/egusphere-egu23-11970, 2023.

Reliable early forecasting of summer air temperature is important to effectively prepare and mitigate damage such as heat-related mortality and excessive electricity demand caused by heat waves and tropical nights. Numerical weather prediction (NWP) models have been used for operational forecasting of air temperature. However, NWP models have coarse spatial resolution due to massive computational resources arising from complex forecasting systems and unstable parameterization of NWP models, which make the uncertainty of prediction, consisting of systematic and random biases. Therefore, the objective of this study is to develop a novel deep learning-based statistical downscaling approach for the Global Data Assimilation and Prediction System (GDAPS) model’s summer air temperature forecasts over South Korea. This study developed the proposed statistical downscaling model through the decomposition into the temporal dynamics of daily air temperature forecast and spatial fluctuation by pixels. The daily temperature dynamic was estimated using a daily mean GDAPS temperature forecast with simple mean bias correction. The spatial fluctuation by pixels was obtained using the spatial anomaly of downscaled air temperature forecast by the U-Net model. The GDAPS model’s forecast data, present-day high spatial resolution satellite observations, and topography variables were used as input variables for training the U-Net model. The observations at weather stations were spatially interpolated using the regression-kriging, and then we used it as a target image for the U-Net model. The proposed U-net model was compared with the Local Data Assimilation and Prediction System (LDAPS), the dynamically downscaled model of the GDAPS, and the support vector regression (SVR)-based statistical downscaling model. For next-day Tmax and Tmin forecasts, the suggested U-net model showed better performance, having high coefficient of determination (R²) of 0.76 and 0.74 and root mean square error (RMSE) of 2.5 °C and 1.5 °C for next-day Tmax and Tmin forecasts, respectively. When analyzing the skill score (SS) values by stations of the U-Net model, it had remarkably high SS values at stations where the GDAPS had a high absolute value. For Tmax and Tmin forecasts with 1-7 days forecast lead time, the suggested model consistently provided better performance (higher spatial correlation and lower RMSE) than GDAPS and SVR. In addition, the U-net model showed a detailed spatial distribution most similar to that of the observations. These results demonstrated that the suggested model successfully corrected the bias of the GDAPS, improving not only the forecast accuracy but also the ability to capture the spatial distribution of Tmax and Tmin forecasts. Using the deep learning-based suggested model in this study, bias-corrected high spatial resolution air temperature forecasts with a relatively long forecast lead time in summer seasons can be successfully produced.

How to cite: Cho, D., Im, J., and Jung, S.: Deep learning-based statistical downscaling for short-term forecasting of summer air temperatures, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-14894, https://doi.org/10.5194/egusphere-egu23-14894, 2023.

Taiwan is a 35,808-km² island with more than 100 peaks over 3,000 meters. The complex terrain in Taiwan makes forecasters more challenging to forecast rainfall in mesoscale and storm-scale. Besides, the spatial distribution of rainfall stations is quite uneven as well. Moreover, the forecast performance of both the European Centre for Medium-Range Weather Forecasts (ECMWF) and the Global Forecast System (GFS) is limited by Taiwan's complex terrain, having certain systematic deviations in rainfall forecasts. For example, the ECMWF forecast has underestimated heavy rainfall and over-predicted light rain in Taiwan. Consequently, to correct model deviations and provide better rainfall forecast products, advanced statistical or artificial intelligence (AI) methods should be studied.

This research applies the U-net neural network to generate downscaling rainfall prediction. We collected precipitation forecast data from the ECMWF (9 km resolution) and the GFS (22 km resolution) during 2021 as the model input. The Quantitative Precipitation Estimation and Segregation Using Multiple Sensor (QPESUMS) radar data from CWB is used as the label data. QPESUMS data can effectively help describe the complete spatial distribution of rainfall. The testing data is from the 2022 whole year. An innovative design of the proposed model is a geographical attention layer (GAL) in the U-net. The GAL helps to learn the geospatial characteristics from the QPESUMS rainfall observation. Moreover, this study uses a scale-separated loss function for model optimization, for which the rainfall is divided into large-scale smoothing and small-scale disturbance fields.

Results show that this U-net downscaling model successfully learns the feature and corrects the systematic bias in both global models, such as shifts in the rainfall caused by topographical lift and local circulation. Furthermore, based on the overall statistics of 2021, the performance diagram shows that the AI model corrects the over-prediction of light rain, while the critical success index in heavy rain is improved by 25 to 30%. The ongoing work of this research will apply generative adversarial networks to break the limitation of learning wrong features from the original forecast input data.

How to cite: Chang, R.-C., Cheng, Y.-Y., and Chen, B.-F.: An Advanced Deep Learning Rainfall Forecasts Downscaling Method in Taiwan, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-10645, https://doi.org/10.5194/egusphere-egu23-10645, 2023.

Solar energy generation is highly volatile during the day due to the strong dependence on cloud dynamics, which limits its integration into the power grid (Smith et al., 2022). On the other hand, higher utilization of renewable energy is essential to tackle climate change. To increase the share of photovoltaic energy in the grid without jeopardizing grid stability, accurate forecasts are essential to ascertain the balance between energy demand and supply (David et al., 2021).

Photovoltaic energy production mainly depends on downwelling surface solar radiation (). SSR is accurately measured by pyranometers, but their spatial representativeness is limited to a few kilometers. By estimating the SSR from geostationary satellites, we can cover larger areas with high spatial and temporal resolutions, allowing us to track cloud motion.

Previous studies on probabilistic cloud motion focused on optical-flow methods without considering the temporal evolution of clouds as such. We address this issue by presenting a scale-dependent approach to forecast. Our approach is inspired by the works of Bowler et al., 2006 and Pulkkinen et al., 2019 on precipitation nowcasting. The novelty of our study is the utilization of different autoregressive models to forecast the temporal evolution of cloudiness of different spatial scales. Our work is motivated by the scale-dependent predictability of cloud growth and decay. By exploiting more than one autoregressive model, we can predict the noisy evolution of small scales independently of the more deterministic evolution of larger spatial scales.

Our preliminary results over Switzerland indicate that our model outperforms the probabilistic advection model based on Carriere et al., 2021 noise generation by reducing the continuously ranked probability score (CRPS) on the test set by 14%. Moreover, we demonstrate the advantage of cloudiness scale decomposition by comparing our model with the same approach without decomposition. We can reduce the CRPS by 6% and the RMSE by 5% by decomposing the images into multiple cascades

References

Bowler, N., C. Pierce, A. Seed, 2006, “STEPS: A probabilistic precipitation forecasting scheme which merges an extrapolation nowcast with downscaled NWP”, Quarterly Journal of the Royal Meteorological Society, 132, 620, pp. 2127–2155, doi:10.1256/qj.04.100.

Carriere, T., R. Amaro e Silva, F. Zhuang, Y. Saint-Drenan, P. Blanc, 2021, “A New Approach for Satellite-Based Probabilistic Solar Forecasting with Cloud Motion Vectors”, Energies, 14, doi:10.3390/en14164951.

David, M., M. Luis, P. Lauret, 2018, “Comparison of intraday probabilistic forecasting of solar irradiance using only endogenous data”, International Journal of Forecasting, 34, doi:10.1016/j.ijforecast.2018.02.003.

Pulkkinen, S., D. Nerini, A. Perez Hortal, C. Velasco-Forero, A. Seed, U. Germann, L. Foresti, 2019, “Systems: an open-source Python library for probabilistic precipitation nowcasting (v1.0)”, Geoscientific Model Development, 12, 10, pp. 4185–4219, doi:10.5194/gmd-12-4185-2019.

Smith, O., O. Cattell, E. Farcot, R. D. O’Dea, K. I. Hopcraft, 2022, “The effect of renewable energy incorporation on power grid stability and resilience”, Science Advances,  https://www.science.org/doi/abs/10.1126/sciadv.abj6734.

How to cite: Carpentieri, A., Folini, D., Nerini, D., Pulkkinen, S., Wild, M., and Meyer, A.: Short-term probabilistic forecast of cloudiness: a scale-dependent advection approach, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-12924, https://doi.org/10.5194/egusphere-egu23-12924, 2023.

Ensemble forecasts of precipitation with sub-seasonal lead times offer  useful information for decision makers when they sufficiently sample the possible outcomes of trajectories. In this study, we aim to improve  precipitation ensemble forecast systems using a stochastic weather generator (SWG) based on analogs of the atmospheric circulation. This approach is tested for sub-seasonal lead times (from 2 to 4 weeks). The SWG ensemble forecasts  yield promising probabilistic skill scores for lead times of 5-10 days for precipitation (Krouma et al, 2022) and for lead times of 40 days for temperature   (Yiou and Déandréis, 2019) . In this work, we adapt the parameters of the SWG to optimize the simulation of European precipitations from ensemble dynamical reforecasts of ECMWF and CNRM. We present the HC-SWG forecasting tool (HC refers to Hindcast and SWG to the stochastic weather generator) based on a combination of dynamical and stochastic models.

We start by computing analogs of Z500 from the ensemble member reforecast of ECMWF (11 members) and CNRM (10 members). Then, we generate an ensemble of 100 members for precipitation over Europe. We evaluate the ensemble forecast of the HC-SWG using skill scores such as the continuous probabilistic score CRPS and ROC curve.

We obtain reasonable forecast skill scores for lead times up to 35 days for different locations in Europe (Madrid, Toulouse, Orly, De Bilt and Berlin). We compare the HC-SWG forecast with other precipitation forecasts to further confirm the benefit of our method. We found that the HC-SWG shows improvement against the ECMWF precipitation forecast until 25 days.

How to cite: Krouma, M., Batté, L., Magnusson, L., Specq, D., Ardilouze, C., and Yiou, P.: Improving the ensemble forecast of precipitation in Europe by combining a stochastic weather generator with dynamical models , EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-1777, https://doi.org/10.5194/egusphere-egu23-1777, 2023.

The accurate forecasting of the intensity of tropical cyclones (TCs) is able to effectively reduce the overall costs of disaster management. In this study, we proposed a deep learning-based model for TC forecasting with the lead time of 24, 48, and72 hours following the event, based on the fusion of geostationary satellite images and numerical forecast model output. A total of 268 TCs which developed in the Northwest Pacific from 2011 to 2019 were used in this study. The Communications system, the Ocean and Meteorological Satellite (COMS) Meteorological Imager (MI) data were used to extract the images of TCs, and the Climate Forecast System version 2 (CFSv2) provided by the National Center of Environmental Prediction (NCEP) was employed to extract atmosphere and ocean forecasting data. In this study, we suggested hybrid convolutional neural network (hybrid-CNN)-based TC forecasting models. It enables to efficiently consider not only the physical but also the spatial characteristics of variables. The Joint Typhoon Warning Center (JTWC) was used for validating the suggested model, and Korea Meteorological Administrator (KMA)-based operational TC predictions were utilized for evaluating the performance of the model. A hybrid-CNN-based prediction model obtained mean absolute errors (MAE) of 13.58, 16.48, and 21.64 kts and skill scores (SS) of 29%, 19%, and 1.6% for 24h, 48h, and 72h forecasts, respectively. Since the rapid intensification (RI) is one of the challenging tasks in the TC intensity prediction, the performance of suggested model for all RIs in 2019 were additionally evaluated. Compared to KMA-based predictions, the suggested models achieved average SS of 66%. Furthermore, using an explainable artificial intelligence (XAI) approach, it is possible to verify how the suggested model works for forecasting TC intensity and propose the feasibility of the suggested model in the meteorology field.

How to cite: Lee, J. and Im, J.: Deep learning-based tropical cyclone intensity prediction through synergistic fusion of geostationary satellite and numerical prediction model, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-14443, https://doi.org/10.5194/egusphere-egu23-14443, 2023.

Although tropical cyclone (TC) forecasts can fairly well capture the TC track and primary rainfall distribution, limited skills are found in forecasting TC structural changes and asymmetric gusty winds. The barrier to further understanding TC structural change is due mainly to the lack of observation, and it is difficult to have systematic 2-D wind analyses. Here, we developed a deep learning model — Deep Learning 2-D Structure Analysis Model for Tropical Cyclones (DSAT-2D) — to produce TC wind analysis in high-temporal-spatial resolutions based on generative adversarial networks (GAN). We use IR1 satellite observation and ERA5 reanalysis data as the model input for the DSAT-2D. The ASCAT surface wind data were collected and used as the label data. Note, however, that the ASACT analysis tends to underestimate winds greater than 15 m/s. Thus, we proposed several methods to fix this issue before training the model. Furthermore, other innovative designs in the DSAT-2D model include: (i) we regrid all data in a polar coordinate to better handle the TC tangential and radial features, and (ii) we also set the target of the DSAT-2D model as the TC radial wind and tangential wind.

Experiment results demonstrate that the DSAT-2D model can capture the TC asymmetric wind structure while possessing the capability of increasing the maximum estimation frequency from approximately 12 hours (e.g., ASCAT data) to less than one hour. The DSAT-2D model may help understand the TC asymmetric wind evolution and improve TC forecasts. Future applications of assimilating this value-added information into the numerical weather prediction model will also be discussed.

How to cite: Cheng, Y.-Y. and Chen, B.-F.: An End-to-end Deep Learning Approach for Analyzing Tropical Cyclone 2-D Surface Winds Utilizing Satellite Data, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-10632, https://doi.org/10.5194/egusphere-egu23-10632, 2023.