Climate CH2025 documents and explains past, present, and future climate change in Switzerland using the latest climate model data, providing the scientific basis for updating the National Adaptation Strategy after 2025. The Climate CH2025 scenarios use climate models from the Coupled Model Intercomparison Project (CMIP), integrating CMIP5-era Regional Climate Models (hereafter called RCMs) and CMIP6 General Circulation Models (GCMs) through both established and newly developed approaches based on Global Warming Levels (GWLs). Observations show that climate response in Switzerland has been particularly pronounced in comparison to other global land regions with mean near-surface air temperatures in 2024 exceeding the preindustrial reference period by 2.9 °C. This is a warming rate about two times faster than on global average. Most models simulate a substantially lower warming trend over this period. The recent warming was likely substantially enhanced by internal variability and by a decline of atmospheric aerosol loads since the 1980s. Regardless, a mismatch identified between RCMs and GCMs, where western Europe and Switzerland warm consistently more in GCMs than RCMs, in particular in spring and summer, limits confidence in the RCMs. This warming mismatch presents the main methodological challenge for Climate CH2025.Several methodological choices were made in Climate CH2025 to reduce the influence of the RCM-GCM warming mismatch on Swiss climate change projections. The first was to set the “present day” base period to 1991-2020, consistent with the current norm period of the World Meteorological Organization. The observed global warming from the preindustrial period to the present day was used to calculate when each GCM reaches a given GWL, defined as a 30-year mean relative to preindustrial conditions. CMIP6 GCMs were brought in to incorporate the latest regional warming estimates, which were used in a regional time adjustment step that ensured RCMs and GCMs warmed the same amount regionally at each GWL. Once regional warming was aligned, local climate responses at 1.5 °C, 2 °C, and 3 °C of global warming could be reported. This method we call the “Block-Time-Shift" (BTS) approach. An advantage of using GWLs is that they relate warming on the global scale to Swiss warming, without relying on specific details in socioeconomic emissions scenarios. A disadvantage is that BTS cannot provide fully transient timeseries. Here we show how the BTS approach shaped results in Climate CH2025, particularly in comparison to earlier Swiss climate scenarios. We report on user feedback on GWLs from communication and technical standpoints and provide guidance for updating workflows from change at fixed time points to change at fixed points in global temperature.

National climate scenarios reflecting the current scientific state of knowledge are an indispensable basis for public and private sectors to plan and design adaptation and mitigation measures. Regional or even local assessments of future climate change are therefore an important climate service. The next generation of climate scenarios for Switzerland is currently under development in the project Klima CH2025. Similar to previous climate scenario generations, the new project is a joint effort involving the Federal Office of Meteorology and Climatology MeteoSwiss, ETH Zurich, C2SM and further partners from academia and administration. The main goal of Klima CH2025 is to develop, update and provide the physical basis of climate change in Switzerland and related products. Two main scientific questions will be addressed: 1) How can we better merge observations and model-based climate scenarios in order to provide consistent and temporally seamless information to best serve user needs?; 2) What is the projected evolution of impact-relevant climate extremes in Switzerland and what are their underlying processes? Guided by these two questions, we will develop a range of new products, engage with stakeholders, and plan active communication and dissemination.We build our scenarios upon the existing CMIP5-based EURO-CORDEX simulations but combine them with CMIP6 GCM information using a variant of a pattern scaling approach. Our approach bridges the gap between different CMIP and CORDEX generations and at the same time merges models and observations to provide consistent information on climate change for the past, the present and the future. In this presentation, the general approach of the Klima CH2025 framework will be presented together with the method applied to bridge models and observations and to integrate CMIP6 evidence into CMIP5-based EURO-CORDEX simulations. In addition, first results of the project and planned products will be presented.

Forecasting winds at the local scale can be challenging due to the highly variable and complex nature of wind patterns, particularly in the case of complex terrain. In such cases, the accuracy of numerical weather prediction models (NWPs) is often limited by the quality of their initial conditions and their grid resolution. This is where the use of observational data through statistical postprocessing techniques can help to improve the quality of forecasts. Statistical postprocessing is nowadays an established component in operational weather forecasting that is used to improve the accuracy, resolution, and calibration of NWP ensemble forecasts with historical observations. In recent years, machine learning techniques have shown great potential in the field of postprocessing, thanks to their ability to deal with increasingly large volumes of data, and the capacity to capture complex relationships between forecasts and observations that are not explicitly represented in traditional postprocessing methods. To capitalize on machine learning for weather applications, and for it to gain acceptance and become a reliable technology for operational use, it is also crucial to consider the technical and engineering challenges that arise when implementing machine learning in a productive environment. MLOps, or Machine Learning Operations, is a set of practices that are used to manage and streamline the deployment, monitoring, and maintenance of machine learning models in production. We will present our recent experience with the development and operationalization of a statistical postprocessing system based on the use of neural networks to predict the probability distribution of forecasts of surface winds. Following MLOps best practices, our framework aims to improve the reproducibility and automation of most common tasks in a machine learning-based system, such as efficient data loading and manipulation, the monitoring and visualization of prediction quality, and the automation of model training and deployment pipelines.

On a daily basis, MeteoSwiss provides a wide range of automatic weather forecasts to the general public, the aviation and to private customers. These data are provided by an ensemble of heterogeneous individual system wich forces the end-user to choose between different sources and sometimes to combine them despite a limited knowledge of their quality and shortcomings. In addition, the forecasts of the different systems are provided in different formats and through different channels, making combined use even more difficult. Furthermore, from a scientific point of view, combining nowcasting and post-processing approaches in a single step using statistical and/or machine learning methods has been shown to give the best forecast performance for twelve hour precipitation forecasts Deep learning for twelve-hour precipitation forecasts. Nat Commun 13, 5145 (2022)). In such an approach, there is no need for a separate system for the nowcasting range, and therefore no need to combine different forecasts a posteriori, which requires further assumptions and introduces further source of errors and inconsistencies. MeteoSwiss has therefore initiated a project to build a system that will integrate data from different weather observation and weather forecasts data (ECMWF, MeteoSwiss regional ICON implementation) in order to provide a consolidated and easily accessible weather forecast dataset. We aim as well for probabilistic gridded forecasts that are seamless in space and time which can be used by a wide range of application, like hydrological models, automatic generation of warning proposals for forecasters, probabilistic animation of precipitation for the MeteoSwiss App. This project is very much oritented toward end-users and a large effort will be made to assess and meet their needs.

<div> <p><span>Correctly representing surface wind is critical for applications such as renewable energy, snow modelling or warning systems. However, numerical weather prediction models with their limited resolution cannot fully represent the strong variability due to complex topography. Downscaling techniques – functionally equivalent to postprocessing when the ground truth is given by observational data - can achieve remarkable results in reducing systematic biases of raw models and can be calibrated to yield accurate probabilistic information at any point in space.</span><span> </span></p> </div><div> <p><span>These techniques can be further improved at analysis time by including real-time measurements, allowing to produce a probabilistic sub-grid resolution analysis of surface wind. Such a product would enable other interesting applications, such as detailed climatologies or nowcasting, and could serve as a ground truth for training deep learning-based postprocessing models with generative approaches, allowing to model spatially and temporally consistent ensembles. </span><span> </span></p> </div><div> <p><span>The first important challenge is to integrate measurements in a statistically optimized and efficient way. Here, we share our ongoing work and preliminary results in a comparative analysis of different approaches, from naïve interpolations to geostatistical techniques or novel approaches based on neural networks. The analysis is based on a multi-year archive of hourly wind observations and NWP analyses from the operational COSMO-1E model over Switzerland.</span><span> </span></p> </div>

In the framework of the renewal of the warning system at MeteoSwiss, we use the possibility to redesign the software as well as the scientific approach of generating warnings. The production chain is designed such that the data is guided and refined along the pathway. Here, we present the operational production process currently in development for the next generation warning system. The first step is the combination of all available data into one data stream (seamless weather). The second stage prepares warning proposals for the forecasters. The only human interaction with the system (third step) is the one of the forecaster in case of extreme weather. However, the automatically generated warning proposals aim to minimize the time needed by the forecaster to issue a warning. In a last step at the end of the chain, the warning products are customized and distributed to our customers. Furthermore, the warnings will also be verified automatically to monitor the quality of the system. Along the chain, we run into a number of challenges, which ask for clever solutions: How do we group grid cells with similar extreme weather information into meaningful warning polygons? How do we facilitate the interaction of the system with the forecaster? Which aspects of the warning do we verify? How do we judge the quality of the warning? Finally yet importantly, what does the warning that we issue actually mean? If we do not have a common understanding about what we are expecting to happen within a warning polygon a clear communication of uncertainties and measuring the quality of our warnings is impossible. Hence, the system involves new applications based on specific developments aiming at generating the greatest value for public warnings. Most of the warning chain is purely machine driven; nonetheless, the human interaction remains a key aspect of the new warning system.

Under hot conditions the human body is able to regulate its core temperature via sweat evaporation, but this ability is reduced when air humidity is high. These conditions of high temperature and high humidity invoke heat stress which is a major problem for humans, in particular for vulnerable groups of the population and people under physical stress (e.g. heavy duty work without appropriate cooling systems). It is generally expected that the frequency, duration and magnitude of such unfavorable conditions will increase with further climate warming. In this respect, climate services play a crucial role by putting together climatological information and adaptation solutions to reduce future heat stress. We here assess the recently developed CH2018 scenarios for Switzerland (https://www.climate-scenarios.ch) in terms of heat stress conditions including their future projections. For this purpose, we characterize future extreme heat conditions with the use of climate analogs. By doing so, we attempt to produce more accessible climate information which might foster the use and understanding of regional-scale climate scenarios. Here heat stress is expressed through the Wet Bulb Temperature (TW), which is a relatively simple proxy for heat stress on the human body and which depends non-linearly on temperature and humidity. It is assessed in terms of single-day events and heat stress spells. Projections based on the CH2018 scenarios indicate increasing heat stress over Switzerland, which is accentuated towards the end of the century. High heat stress conditions might be about 3-5 times more frequent for an emission scenario without mitigation (RCP 8.5) than for the mitigation scenario (RCP 2.6) by the end of the 21st century. The projected increase of heat stress results in more and longer heat stress spells, thus highlighting the importance of timely and precise prevention strategies in the context of heat-health action plans. Spatial climate analogs based on heat stress spells in Switzerland greatly vary depending on the emission scenario and are found in Central Europe under a mitigation scenario and in southern Europe under unmitigated warming.

Weather forecasting centers currently rely on statistical postprocessing methods to minimize forecast error. This improves skill but can lead to predictions that violate physical principles or disregard dependencies between variables, which can be problematic for downstream applications and for the trustworthiness of postprocessing models, especially when they are based on new machine learning approaches. Building on recent advances in physics-informed machine learning, we propose to achieve physical consistency in deep learning-based postprocessing models by integrating meteorological expertise in the form of analytic equations. Applied to the post-processing of surface weather in Switzerland, we find that constraining a neural network to enforce thermodynamic state equations yields physically-consistent predictions of temperature and humidity without compromising performance. Our approach is especially advantageous when data is scarce, and our findings suggest that incorporating domain expertise into postprocessing models allows to optimize weather forecast information while satisfying application-specific requirements.

MeteoSwiss is currently implementing a new NWP postprocessing suite for providing automated local weather forecasts to the general public. As these forecasts are nowadays mainly accessed via smartphone app, we aimed at global postprocessing approaches, that is, optimizing forecasts not only at observation sites but at any location in Switzerland. The system takes advantage of both regional area and global NWP ensemble models with different forecast horizons for providing seamless probabilistic predictions over two weeks leadtime. Finally, the postprocessing suite also considers operational aspects such as robustness towards missing or delayed input data or the ability to cope with limited reforecasts records for training.Both ensemble model output statistics (EMOS) and machine learning (ML) methods are applied for postprocessing the target parameters temperature, precipitation, wind and cloud cover. Forecast skill in terms of CRPS improves by up to 30% compared to direct model output, with largest benefits for temperature and wind in areas of complex orography and only marginal gains for precipitation during seasons with a high fraction of convective situations. The postprocessing of multiple NWP sources not only allows seamless forecasts, but also proved more skillful than single-model postprocessing. EMOS postprocessing performed well even in case only short reforecast records were available, but was outperformed by ML approaches given sufficient training data. While this general-purpose postprocessing suite improves forecasts overall, it showed weaknesses in some warning-relevant weather situations. Future developments will aim at extending its applicability to these less frequent situations, target further parameters, and extending the use of ML methods.

To make sound decisions in the face of climate change, government agencies, policymakers and private stakeholders require suitable climate information on local to regional scales. In Switzerland, the development of climate change scenarios is strongly linked to the climate adaptation strategy of the Confederation. The current climate scenarios for Switzerland CH2018 - released in form of six user-oriented products - were the result of an intensive collaboration between academia and administration under the umbrella of the National Centre for Climate Services (NCCS), accounting for user needs and stakeholder dialogues from the beginning. A rigorous scientific concept ensured consistency throughout the various analysis steps of the EURO-CORDEX projections and a common procedure on how to extract robust results and deal with associated uncertainties. The main results show that Switzerland’s climate will face dry summers, heavy precipitation, more hot days and snow-scarce winters. Approximately half of these changes could be alleviated by mid-century through strong global mitigation efforts. A comprehensive communication concept ensured that the results were rolled out and distilled in specific user-oriented communication measures to increase their uptake and to make them actionable. A narrative approach with four fictitious persons was used to communicate the key messages to the general public. Three years after the release, the climate scenarios have proven to be an indispensable information basis for users in climate adaptation and for downstream applications. Potential for extensions and updates has been identified since then and will shape the concept and planning of the next scenario generation in Switzerland.

Automated forecasting provides the basis for everyday forecast products used by a wide range of users. Continued progress in numerical weather prediction allows to produce local forecasts with considerable accuracy. To further reduce systematic errors and thereby render such local forecasts more beneficial to users, statistical postprocessing can be employed. While statistical postprocessing can readily be optimized for specific applications, optimization is less straight forward for general-purpose forecasts that are used across a diversity of applications and decisions. This issue is illustrated with ensemble postprocessing for automated precipitation forecasts. While medium-range precipitation forecasts are often communicated with hourly granularity, beyond the nowcasting range most applications are likely less affected by the precise timing of precipitation. In contrast, (sub-)daily aggregated precipitation may be a more relevant quantity. In addition, predictability of hourly precipitation is generally very limited days in advance and statistical postprocessing for hourly precipitation forecasts will therefore be strongly affected by the regression-to-the-mean problem (i.e. statistical postprocessing resorts to issuing climatological forecasts in the absence of predictability). To overcome the above issues, we propose a combined postprocessing approach operating on daily aggregated precipitation and hourly fractions of daily precipitation. We present results for a simple disaggregation according to the NWP precipitation and disaggregation according to a separate postprocessing of the hourly fraction of daily totals. The latter approach allows us to correct systematic biases in the diurnal cycle of precipitation occurrence particularly relevant for convective situations in complex topography as is the case in Switzerland. The same approach can be used to extend daily precipitation forecasts into the sub-seasonal to seasonal range.

Objective forecast verification provides the basis to motivate changes to the forecast system. At MeteoSwiss, we are introducing statistical ensemble postprocessing into our automated forecast production. These automated forecasts are accessed by the Swiss general population through the MeteoSwiss website and mobile app and they will form the basis for a range of derived products. Therefore, it is crucial to evaluate the new forecasts broadly, i.e. on the diverse aspects of forecast quality relevant for the variety of products used across all of Switzerland. The core component of this evaluation system consists of a web portal for interactive visualization of verification measures. This web portal provides the means to compare forecast meteograms at stations with the corresponding observations for a quick visual inspection akin to the information available on the mobile app. In addition, forecast quality of the most recent forecasts is monitored using various verification measures. Finally, re-forecasting with experimental configurations is integrated to produce in-depth reporting on the effect of novel forecasts to support decisions on the development of the forecasting system. To facilitate near real-time analyses, atomic verification scores per meteorological parameter, forecast source, forecast issue time, station, and time are pre-computed and stored in a partitioned database. The partitioning allows for rapid multi-threaded access at analysis time from the interactive web portal. The objective verification is complemented with feedback by forecasters on duty on individual cases during the pre-operational phase of new forecast developments. Here, we will showcase how each of these parts is used to assess release candidates for the automatic forecast production.

Probabilistic predictions of precipitation call for rather sophisticated postprocessing approaches due to its low predictability, high spatio-temporal variability and highly positive skewness. Moreover, the large number of zeros makes the generation of physically realistic postprocessed forecast scenarios using standard approaches like ensemble copula coupling (ECC) rather difficult. In addition to classical statistical approaches, recently, machine learning based methods gained increasing popularity in the field of postprocessing of probabilistic weather forecasts. In this study, we compare conditional generative adversarial network (cGAN) based postprocessing of daily precipitation with a quantile regression based approach. In principle, an appropriately trained cGAN model should be able to generate postprocessed forecast scenarios that improve forecast skill and cannot be distinguished from observed data in terms of spatial structure. While we use ECC to generate physically realistic forecast scenarios from quantile regression, cGAN does not need any additional ECC steps. For training and verification, we use COSMO-E ensemble forecasts with a grid resolution of about 2 km over Switzerland and the corresponding CombiPrecip observations, which are a gridded blend of radar and gauge observations. Preliminary results suggest that it is possible to generate realistic looking forecast scenarios using cGAN, but up to now, we have not been able to increase forecast skill. On the other hand, quantile regression seems to increase forecast skill at the expense of relying on an additional ECC step to generate forecast scenarios.

Verification is a core activity in weather forecasting. Insights from verification are used for monitoring, for reporting, to support and motivate development of the forecasting system, and to allow users to maximize forecast value. Due to the broad range of applications for which verification provides valuable input, the range of questions one would like to answer can be very large. Static analyses and summary verification results are often insufficient to cover this broad range. To this end, we developed an interactive verification platform at MeteoSwiss that allows users to inspect verification results from a wide range of angles to find answers to their specific questions.We present the technical setup to achieve a flexible yet performant interactive platform and two prototype applications: monitoring of direct model output from operational NWP systems and understanding of the capabilities and limitations of our pre-operational postprocessing. We present two innovations that illustrate the user-oriented approach to comparative verification adopted as part of the platform. To facilitate the comparison of a broad range of forecasts issued with varying update frequency, we rely on the concept of time of verification to collocate the most recent available forecasts at the time of day at which the forecasts are used. In addition, we offer a matrix selection to more flexibly select forecast sources and scores for comparison. Doing so, we can for example compare the mean absolute error (MAE) for deterministic forecasts to the MAE and continuous ranked probability scores of probabilistic forecasts to illustrate the benefit of using probabilistic forecasts.

MeteoSwiss is developing and implementing a post-processing suite of multi-model ensemble forecasts to produce seamless probabilistic calibrated forecasts at arbitrary locations in Switzerland (i.e. also for un-observed locations). With the complex topography of Switzerland, the raw output of the numerical model is subject to particular strong biases and conditional errors. Here, we present results for hourly temperature and precipitation predictions. We apply a global ensemble model output statistics (gEMOS) framework. It extends the classical EMOS approach by incorporating static predictor variables describing relevant topographical features and it is trained for all stations together using a 4-year multi model numerical weather prediction (NWP) archive. As NWP sources, we combine data from the COSMO model suites (1.1 and 2.2 km horizontal grid-spacing) and from the ECMWF IFS medium-range forecasting system. Note that the three NWP suites have different forecast horizons. We show that gEMOS is able to improve forecasts for both variables. Depending on selection of predictors, lead-time, hour-of-day and season we find improvements up to 30% in terms of CRPS for both variables with most pronounced improvements in mountainous regions. Particularly for temperature, the multi-model combination further increases the forecast skill compared to postprocessing using high-resolution simulations of COSMO only. While locally optimized approaches show better performance in terms of skill at the observing sites, the advantage of gEMOS lies in the ability to generate calibrated predictions for arbitrary locations in a consistent way. Its computational efficiency makes it a particularly attractive method for operationalization in a realtime context.

Machine Learning has a big potential for various tasks along the whole value chain of a national Met-Service. Indeed, many research groups, private and national weather services have started to explore the possibilities and first real-time operational implementations are in place already. However, the building up of the expertise is difficult, large amounts of data have to be made available in an efficient way and the necessary tools have special and demanding requirements concerning infrastructure and maintenance. Also, the transition from research results towards operational tools being operated in realtime is a particular challenge. Not least, trust from end-users must be built, while trying to avoid falling into the short-term hype trap. In this presentation, we want to present some examples of machine-learning at MeteoSwiss that are in operational use or soon to be. This includes the use in a measurement system to identify pollen species, the quality control of meteorological observations, the postprocessing of numerical weather forecasts and the condensation of weather forecast information for the meteorologists. These examples have different characteristics and cover a wide range of applications, but also share some common properties. We want to juxtapose these properties with the incentives and conditions how machine learning methods are developed and employed in a more research oriented context like in academia. It turns out that an operational setup of machine learning has very different requirements than machine learning in a research context. The identification of these differences, but also the similarities, could help to understand the challenge of bringing research results into operation and how to alleviate this challenge in the future.

Hourly wind forecasts from numerical weather prediction models suffer from a range of systematic and random errors that are to a great extent related to limitations in the model grid resolution. To correct for such biases, statistical postprocessing and downscaling procedures are commonly applied so to leverage the information provided by automatic wind measurements at the surface. More recently, such techniques have been reformulated in a machine learning framework so to profit from the increased availability of data and computational resources. The results reported in the literature are promising and call for a serious evaluation of their potential for operational forecasting. However, there remain several scientific and more applied challenges that need to be addressed before such methods can transition to real-world applications. One such challenge relates to the availability of multiple ensemble forecasts for the same point in time and space, which raises the question of how the information can be efficiently and optimally handled during postprocessing, so to provide added value to the end-user without adding technical debt to the operational system. We propose an approach where a single deep learning model is trained to postprocess a combination of three ensemble forecasting systems, namely the high-resolution regional COSMO model with two configurations, and the ECMWF IFS ENS global ensemble forecasting system. We will show how the training is set up to provide a robust postprocessing model that can account for real time scenarios that include missing data and late model runs, while the quality of the forecasts remains comparable to a single-model approach. We found that the flexibility of the deep learning architecture translates into a robust automatic postprocessing solution that limits the maintenance burden and improves the system’s reliability.

MeteoSwiss has developed and is currently implementing a NWP postprocessing suite for providing automated weather forecasts at any location in Switzerland. The aim is a combined postprocessing of high resolution limited area and global model ensembles with different forecast horizons to enable seamless probabilistic forecasts over two weeks leadtime. Further, the output should be coherent in space and provide predictions at any location of interest, including sites without observations. We use the full archive of MeteoSwiss’ operational local area models (COSMO-1 and COSMO-E) over the past four years and the corresponding IFS-ENS medium range predictions of ECMWF to develop postprocessing routines for temperature, precipitation, cloud cover and wind. Here we present selected key results on the performance of various postprocessing methods we applied but also on practical aspects of their implementation into operational production. Both ensemble model output statistics (EMOS) and machine learning (ML) approaches are able to improve the forecasts in terms of CRPS by up to 30% as compared to the direct output of the local area model. The skill increase obtained by postprocessing varies depending on the parameter, region and season, with best results for temperature and wind in areas of complex orography and only marginal improvements for precipitation during seasons with a high fraction of convective situations. Particularly for temperature, the combined postprocessing of COSMO and IFS-ENS resulted in a skill benefit over postprocessing the COSMO models alone. Locally optimized postprocessing would allow further skill improvements, but only at sites where observations are available. However, the ability of non-local postprocessing approaches to provide calibrated forecast at any point in space is a key advantage for providing automated forecasts to the general public via the internet and smartphone app. Furthermore, the computational efficiency of these non-local approaches makes them attractive for operationalization in a realtime context.

To improve and automate the quality of weather forecasts to the public, MeteoSwiss is redesigning its statistical postprocessing suite. The effort aims at producing calibrated probabilistic predictions to any arbitrary point in space and up to a 15-day lead time, by seamlessly integrating multiple numerical weather prediction models into a unique consensus forecast. For hourly wind forecasts (mean, gust, and direction), the task is formulated as a regression problem in a supervised machine learning framework, where station measurements are used as labels, and co-located NWP forecasts as features. To improve the estimates at ungauged locations, additional static topographical features are derived from a 50m digital elevation model. The probabilistic component is included by training the neural network not to produce a deterministic prediction, but the parameters of a conditional probability function. To this end, the Continuous Ranked Probability Score (CRPS) is used as a loss function. The dataset includes a range of surface parameters at hourly resolution produced by the operational forecasts from three NWP models (the deterministic COSMO-1 model, at 1 km horizontal resolution; the 21-member COSMO-E, 2 km; and the 51-member ECMWF IFS ENS at about 18 km). The data cover the whole of Switzerland over a period spanning more than four years (mid 2016 to end of 2020). Wind measurements from over 500 surface weather stations are included as reference dataset. The study uses a train-validation-test split in both space and time to assess the ability of the postprocessing model to generalize to unseen locations and times. The results indicate that, despite the challenging nature of the problem, the postprocessing model can improve over the baseline NWP forecasts in terms of CRPS on the test set. In particular, the model is effectively correcting for biases relating to altitude error and other misrepresentations in the NWP topography. The results show that it is feasible to downscale numerical predictions to a substantially higher spatial resolution. Moreover, the conditional probabilities shows consistent improvements in terms of calibration, although it remains a significant portions of undetected peak events (positive outliers), possibly to be related to unpredictable phenomena (e.g., thunderstorm gusts). Finally, first results seem to suggest that the gain in prediction skill is mainly driven by a better statistical reliability rather than higher statistical resolution.

Statistical postprocessing is routinely applied to correct systematic errors of numerical weather prediction models (NWP) and to automatically produce calibrated local forecasts for end-users. Postprocessing is particularly relevant in complex terrain, where even state-of-the-art high-resolution NWP systems cannot resolve many of the small-scale processes shaping local weather conditions. In addition, statistical postprocessing can also be used to combine forecasts from multiple NWP systems. Here we assess an ensemble model output statistics (EMOS) approach to produce seamless temperature forecasts based on a combination of short-term ensemble forecasts from a convection-permitting limited-area ensemble and a medium-range global ensemble forecasting model. We quantify the benefit of this approach compared to only processing the high-resolution NWP. We calibrate and combine 2-m air temperature predictions for a large set of Swiss weather stations at the hourly time-scale. The multi-model EMOS approach ('Mixed EMOS') is able to improve forecasts by 30\% with respect to direct model output from the high-resolution NWP. A detailed evaluation of Mixed EMOS reveals that it outperforms either single-model EMOS version by 8-12\%. Valley location profit particularly from the model combination. All forecast variants perform worst in winter (DJF), however calibration and model combination improves forecast quality substantially.