Balancing accuracy versus precision: Enhancing the usability of sub-seasonal forecasts

Subseasonal forecasting, which bridges the gap between short-term weather forecasts and seasonal outlooks, covers lead times of 2–6 weeks. Traditionally, these forecasts are presented as anomalies over broad areas due to decreasing skill with lead time. However, this approach significantly limits their usability for laypeople, who are used to highly localized forecasts and often struggle with interpreting anomalies, especially without knowledge of the area's climatology for that time of year. We therefore want to share the development and implementation process of a localized subseasonal (3-week) forecast, presenting actual weather variables, integrated into the existing Yr weather service. Yr is a weather app and website (www.yr.no) and is a collaboration between the Norwegian Meteorological Institute and the Norwegian Broadcasting Corporation (NRK). The subseasonal forecast has been operational since January 2024 and is available for the Nordic countries and the Baltic states. During the development process, we prioritized usability over maximizing forecast skill by engaging in co-production with existing Yr users through diary studies, user tests, and feedback forms. This approach provided valuable insights into presenting subseasonal forecasts that users find understandable and actionable. User feedback suggests that despite the inevitable decline in skill by week 3, localized forecasts presenting actual weather variable values—rather than aggregated anomalies—still provide significant value, as long as forecast uncertainty is clearly communicated. Creating a new forecast product through co-production is time-intensive and highly interactive, involving multiple iterations to test different methods for data dissemination. This process requires continual adjustments to both the design and forecast parameters. We will share our lessons learned by presenting not only the final forecast product but also the intermediate versions that were tested and deemed unfit based on user feedback. Lastly, we will discuss the integration of the subseasonal forecast into the Yr weather service, with a primary focus on obtaining consistency with existing forecast products (e.g. the 10-day forecast). Ensuring coherence across products is essential, as user feedback highlighted that most users base their decisions on combining information from different forecast products on Yr.

Recent drought events like in the 2011 Horn of Africa and the ongoing drought in California have an enormous impact on nature and society. Reliable seasonal weather outlooks are critical for drought management and other applications like, crop modelling, flood forecasting and planning of reservoir operation, and would help reduce the potential economic damage from extremes as well as help optimize crop yields during more normal weather years from improved agricultural management. However, most seasonal forecasts are limited by low spatial and temporal model resolutions. The newly released North American Multi-Model Ensemble phase 2 (NMME-2), provides subseasonal forecast that increase the temporal resolution from monthly to daily, enabling subseasonal forecasting for end-user applications that rely on a daily temporal resolution. In this study we give an overview of the current status of the NMME subseasonal forecasts ensembles, their skill over the African continent and the forecast skill of the ensembles for applications related to agriculture and hydrology. We show that the NMME-2 subseasonal forecasts are significantly skilful for both precipitation and 2 m air temperature for large parts of Africa. The precipitation forecasts are skilful up to a lead time of two months, while temperature anomaly forecasts have a significant skill beyond the three months lead for most of Africa. Potential applications that would benefit from the new NMME-2 ensemble were studied in more detail for West Africa. We show that the models have a significant skill in forecasting the onset of the annual rain season in West Africa, and thereby the start of the growing season. Additionally, the models have a significant forecast skill to predict the onset and peak of the high flow season for most parts of West Africa. The low uncertainty in the forecasts compared to the observed anomalies indicates that local stakeholders will benefit from the high temporal resolution that the NMME-2 provides. Results encourage future research into the potential of the new subseasonal NMME-2 forecast ensemble to forecast more specific end-user applications and climate services, which require skilful high temporal resolution forecasts.

Sub-seasonal forecasting (SSF) is the prediction of key climate variables such as temperature and precipitation on the 2-week to 2-month time horizon. Skillful SSF would have substantial societal value in areas such as agricultural productivity, hydrology and water resource management, and emergency planning for extreme events such as droughts and wildfires. Despite its societal importance, SSF has stayed a challenging problem compared to both short-term weather forecasting and long-term seasonal forecasting. Recent studies have shown the potential of machine learning (ML) models to advance SSF. In this paper, for the first time, we perform a fine-grained comparison of a suite of modern ML models with start-of-the-art physics-based dynamical models from the Subseasonal Experiment (SubX) project for SSF in the western contiguous United States. Additionally, we explore mechanisms to enhance the ML models by using forecasts from dynamical models. Empirical results illustrate that, on average, ML models outperform dynamical models while the ML models tend to generate forecasts with conservative magnitude compared to the SubX models. Further, we illustrate that ML models make forecasting errors under extreme weather conditions, e.g., cold waves due to the polar vortex, highlighting the need for separate models for extreme events. Finally, we show that suitably incorporating dynamical model forecasts as inputs to ML models can substantially improve the forecasting performance of the ML models. The SSF dataset constructed for the work and code for the ML models are released along with the paper for the benefit of the artificial intelligence community.

Heatwaves pose a range of severe impacts on human health, including an increase in premature mortality. The summers of 2018 and 2022 are two examples with record-breaking temperatures leading to thousands of heat-related excess deaths in Europe. Some of the extreme temperatures experienced during these summers were predictable several weeks in advance by subseasonal forecasts. Subseasonal forecasts provide weather predictions from 2weeks to 2months ahead, offering advance planning capabilities. Nevertheless, there is only limited assessment of the potential for heat-health warning systems at a regional level on subseasonal timescales. Here we combine methods of climate epidemiology and subseasonal forecasts to retrospectively predict the 2018 and 2022 heat-related mortality for the cantons of Zurich and Geneva in Switzerland. The temperature-mortality association for these cantons is estimated using observed daily temperature and mortality during summers between 1990 and 2017. The temperature-mortality association is subsequently combined with bias-corrected subseasonal forecasts at a spatial resolution of 2-km to predict the daily heat-related mortality counts of 2018 and 2022. The mortality predictions are compared against the daily heat-related mortality estimated based on observed temperature during these two summers. Heat-related mortality peaks occurring for a few days can be accurately predicted up to 2weeks ahead, while longer periods of heat-related mortality lasting a few weeks can be anticipated 3 to even 4weeks ahead. Our findings demonstrate that subseasonal forecasts are a valuable-but yet untapped-tool for potentially issuing warnings for the excess health burden observed during central European summers.

Under the climatic variability and climate change, skilful weather forecast in different spatial and temporal scale encourages the farmers to organize and activate their own resources in the best possible way to increase the crop production. Though medium range weather forecast is used extensively in operational agromet advisory services, sub-seasonal forecast provides additional decision-relevant information to support the timing of crop planting, irrigation scheduling, and harvesting, particularly in water-stressed regions. In view of that, dynamical and statistical and sub-seasonal seasonal forecast is generated and delivered to the farmers as climate information services in number of countries in the world. Under the Gramin Krishi Mausam Sewa (GKMS) Project, Agricultural Meteorology Division, India Meteorological Department (IMD) in collaboration with Indian Institute of Tropical Meteorology (IITM), Pune, All India Coordinated Research Project on Agrometeorology (AICRPAM), CRIDA, Indian Council of Agricultural Research, Hyderabad prepared Agromet Advisory fortnightly taking into consideration realized rainfall during previous fortnight and extended range rainfall forecast for next fortnight and crop information i.e., state and stage of the crops. In the present article agricultural applications of sub-seasonal forecasts on agricultural management in India has been explored. It has been showed how the extended range weather forecast i.e., sub-seasonal forecast has been developed and translated into agromet advisories for the farming communities to increase crop production in India and whether the present state of accuracy could be used for generating advisory under contingent crop planning conditions and other advisories by citing different case studies and ultimately helping the farming communities to improve their economic conditions. It has been demonstrated here that sub-seasonal forecasts are increasingly being used across agriculture in the country. The sub-seasonal forecasting time scale is therefore a new concept for many users. Because of the additional value of sub-seasonal forecasts for decision-making, it is increasingly gaining interest among users. Present case studies clearly suggest the forecast at sub-seasonal time scale is need of the hour.

<p>Increasingly, operational forecasting centres are producing sub-seasonal forecasts, targeted at lead times of 3-6 weeks. These aim to fill the gap between conventional 2-week weather forecasts and longer term seasonal outlooks. However it is often difficult for end-users to know how these sub-seasonal forecasts can be best utilised, and how skilful they are for predicting variables of real world interest.</p><p>Much prior work on sub-seasonal forecasts has focused on assessing skill scores for large-scale smooth fields of mid- or upper-tropospheric variables, or else has looked at heavily time-averaged quantities. How to extend the lessons of these studies to user applications is not always obvious.</p><p>We take a more applied approach, focused on the chaotic and variable weather of Western Europe. We use sub-seasonal temperature forecasts alongside real-world French energy price and demand data in order to directly calculate the financial value of subseasonal forecasts to users in the energy sector. Using this new, real-world framework we make an estimate of cost-loss ratios and so can compare to the results of a simpler potential economic value model.</p><p> </p>

<p>Improving our ability to forecast the weather and climate is of interest to all sectors of the economy and government agencies from the local to the national level. In fact, weather forecasts 0-10 days ahead and climate forecasts seasons to decades ahead are currently used operationally in decision-making, and the accuracy and reliability of these forecasts has improved consistently in recent decades. However, many critical applications require subseasonal forecasts with lead times in between these two timescales. Subseasonal forecasting—predicting temperature and precipitation 2-6 weeks ahead—is indeed critical for effective water allocation, wildfire management, and drought and flood mitigation. Yet, accurate forecasts for the subseasonal regime are still lacking due to the chaotic nature of weather.</p><p>While short-term forecasting accuracy is largely sustained by physics-based dynamical models, these deterministic methods have limited subseasonal accuracy due to chaos. Indeed, subseasonal forecasting has long been considered a “predictability desert” due to its complex dependence on both local weather and global climate variables. Nevertheless, recent large-scale research efforts have advanced the subseasonal capabilities of operational physics-based models, while parallel efforts have demonstrated the value of machine learning and deep learning methods in improving subseasonal forecasting.</p><p>To counter the systematic errors of dynamical models at longer lead times, we introduce an <em>adaptive bias correction</em> (ABC) method that combines state-of-the-art dynamical forecasts with observations using machine learning. We evaluate our adaptive bias correction method in the contiguous U.S. over the years 2011-2020 and demonstrate consistent improvement over standard meteorological baselines, state-of-the-art learning models, and the leading subseasonal dynamical models, as measured by root mean squared error and uncentered anomaly correlation skill. When applied to the United States’ operational climate forecast system (CFSv2), ABC improves temperature forecasting skill by 20-47% and precipitation forecasting skill by 200-350%. When applied to the leading subseasonal model from the European Centre for Medium-Range Weather Forecasts (ECMWF), ABC improves temperature forecasting skill by 8-38% and precipitation forecasting skill by 40-80%.</p><p>Overall, we find that de-biasing dynamical forecasts with our learned adaptive bias correction method yields an effective and computationally inexpensive strategy for generating improved subseasonal forecasts and building the next generation of subseasonal forecasting benchmarks. To facilitate future subseasonal benchmarking and development, we release our model code through the subseasonal_toolkit Python package and our routinely updated SubseasonalClimateUSA dataset through the subseasonal_data Python package.</p>

Few hundred million to billion parameters of autoregressive transformer-based weather foundation models (FMs) have demonstrated generalizabilities for various downstream applications such as regional forecasting to downscaling. They occasionally outperform traditional physics-based models for medium-range forecasting skills as well as enable significantly faster execution speeds. These ML approaches are designed for timeframes ranging from hourly to up to 10 days, and sub-seasonal forecasting, defined as a range spanning two weeks to two months, often receives less attention for their downstream tasks due to the inherent challenges in predicting the chaotic nature of atmospheric systems. However, the sub-seasonal to seasonal forecast has socio-economic impacts influencing actions from seasonal extreme weather events and economic activities. While community standards for benchmarking studies have been conducted for the medium-range forecasts, the benchmarking of sub-seasonal forecasts still needs further efforts. In this study, we are aiming to fine-tune foundation models to predict sub-seasonal forecasts for various variables to conduct comprehensive benchmarking for weather foundation models. Particularly, to reduce the complexity of tasks, our fine-tune task forecasts two-week averaged atmospheric variables with a forecasting lead-time of two weeks. For this task, we resample the community standard dataset, WeatherBench, for the two-week averaged dataset. We primarily work with the Oak Ridge Base Foundation Model for Earth System Predictability (ORBIT), and extend the benchmarking to other FMs across Aurora, ClimaX, and Prithvi WxC models. Our initial fine-tuning task uses a 100 million parameters ORBIT model to predict geopotential height at 200 hPa with two-week lead time, a key indicator for extreme precipitation in Central Southeast Asia. The preliminary results demonstrate that the fine-tuned ORBIT predicts realistic spatial distributions achieving an MSE of 24.32 m when evaluated against the 2018 data. The comprehensive sub-seasonal forecasting benchmarking can highlight the potential of weather FMs whether they capture underlying principles of atmospheric dynamics, thereby enabling their performance to be extended to longer forecast lead-times. 

Skilful subseasonal forecasts are crucial for issuing early warnings of extreme weather events, such as heatwaves and floods. Operational subseasonal climate forecasts are often produced by global climate models not dissimilar to seasonal forecast models, which typically fail to reproduce observed temperature trends. In this study, we identify that the same issue exists in the subseasonal forecasting system. Subsequently, we adapt a trend‐aware forecast postprocessing method, previously developed for seasonal forecasts, to calibrate and correct the trend in subseasonal forecasts. We modify the method to embed 30‐year climate trends into the calibrated forecasts even when the available hindcast period is shorter. The use of 30‐year trends is to robustly represent long‐term climate changes and overcome the problem that trends inferred from a shorter period may be subject to large sampling variability. Calibration is applied to 20‐year ECMWF subseasonal forecasts and AWAP observations of Australian minimum and maximum temperatures with forecast horizons of up to 4 weeks. Relative to day‐of‐year climatology, raw week‐1 forecasts reproduce temperature trends of the 20‐year observations in many regions while raw week‐4 forecasts do not exhibit the 20‐year observed trends. After trend‐aware postprocessing, the behaviour of forecast trends is related to raw forecast skill regarding accuracy. Calibrated week‐1 forecasts show apparent trends consistent with the 20‐year observations, as the calibration transfers forecast skill and embeds the 20‐year observed trends into the forecasts when raw forecasts are inherently skilful. In contrast, calibrated week‐4 forecasts exhibit the 30‐year observed trends, as the calibration reverts the forecasts to the 30‐year observed climatology with trends when raw forecasts have little skill. For both weeks, the trend‐aware calibrated forecasts are more reliable, and as skilful as or more skilful than raw forecasts. The extended trend‐aware method can be applied to deliver high‐quality subseasonal forecasts and support decision‐making in a changing climate.

Application of real time S2S forecasts over Eastern Africa in the co-production of climate services

<p>Climate variations have the potential to strongly affect aquaculture production. By having access to reliable predictions at extended and long-range lead times, aquaculture can take preventative measures. For instance, variability in water temperature influences the growth and mortality rates of farmed fish. Fish farmers can, if they have reliable forecasts, take action against unfavorable changes in water temperature by moving the sea cages and alter feeding schemes and slaughter times accordingly. In this way, one can minimize production loss, and production can become more sustainable. We present how sub-seasonal forecasts  from ECMWF can be used to provide skilful forecasts at lead times of two to four weeks at various fish farm locations in Norway by including post-processing methods that use on-sight observations as a predictor. Sub-seasonal forecasts are expected to capture grid scale variations and larger-scale phenomena in sea temperature. However, fish farms often lie in complex coastal areas and are therefore prone to local effects like river runoff and smaller scale currents, which are not adequately represented in the sub-seasonal forecast models.  First, we assess the forecast skill for all seasons for the fish farms along the Norwegian coast. The Norwegian fish farms are located in various regions, from off-shore to practically closed-off fjord environments. It is clear that forecast skill is reduced the further in the fjords  the fish farms are located. Post-processing the forecasts by including information on the persistence of water temperatures improves the skill in the fjords, compared to using the ECMWF sub-seasonal forecasts alone. The post-processing model is simple to implement and may enhance water temperature forecast skill in regions that are influenced by local processes. Moreover, this overview of forecast skill may guide forecasters and fish farmers on when, and where, to trust the sub-seasonal forecasts, which is crucial for decision making and can be beneficial for the economy and the industry’s  environmental sustainability.</p>

Sub-seasonal forecasting (SSF) focuses on predicting key variables such as temperature and precipitation on the 2-week to 2-month time scale. Skillful SSF would have immense societal value in such areas as agricultural productivity, water resource management, and emergency planning for extreme weather events. However, SSF is considered more challenging than either weather prediction or even seasonal prediction, and is still a largely understudied problem. In this paper, we carefully investigate 10 Machine Learning (ML) approaches to sub-seasonal temperature forecasting over the contiguous U.S. on the SSF dataset we collect, including a variety of climate variables from the atmosphere, ocean, and land. Because of the complicated atmosphere-land-ocean couplings and the limited amount of good quality observational data, SSF imposes a great challenge for ML despite the recent advances in various domains. Our results indicate that suitable ML models, e.g., XGBoost, to some extent, capture the predictability on sub-seasonal time scales and can outperform the climatological baselines, while Deep Learning (DL) models barely manage to match the best results with carefully designed architecture. Besides, our analysis and exploration provide insights on important aspects to improve the quality of sub-seasonal forecasts, e.g., feature representation and model architecture. The SSF dataset and code are released with this paper for use by the broader research community.

<p>The PISSARO project focuses on atmospheric and oceanic forecasting at the subseasonal scale for applications over the South West Indian Ocean basin (SWIO). It is a collaborative academic research project, developed and conducted in partnership with stakeholders from Reunion and Seychelles and a panel of scientific experts in subseasonal forecasting. The aim of this project is to evaluate, improve and valorize subseasonal forecasting data. For this purpose, we mainly use the data archived into the the S2S (Subseasonal-to-Seasonal prediction project) data base in order to 1) evaluate the quality of subseasonal forecasts for tropical cyclones and weather patterns, and 2) develop forecast products suitable for potential users. This project focuses on the SWIO, which has been little studied by the S2S community until now. The different territories of the SWIO are subject to extreme events and a significant cyclonic activity. It is important to take into account the specificities of this region in order to improve their warning systems.</p><p>The ambition to deploy early warning tools cannot be achieved without discussions between potential users and S2S experts. The users specify the characteristics of the products to be developed so that they offer an asset for decision-making, and the experts assess the feasibility of these products. In the presentation, we will first discuss the importance of collaboration between the users and the experts within the project using two concrete actions: the animation of a monthly experimental forecasting briefing with operational forecasters and the participation in conferences in the humanitarian field. Then, we will present the subseasonal forecasting products which are under development for the anticipation of cyclonic risk at monthly scale in the SWIO basin. To address the urgency of the need of the disaster risk reduction, we first made a basic adaptation of already existing tropical cyclone occurrence probability and rainfall forecasting products into products interpretable by non-meteorological users.</p><p>We consider that a crucial information from the S2S data base to provide to users is the level of uncertainty. However, estimating the quality of S2S forecasts is not straighforward. It is actually difficult to match a forecasted cyclone to an actual observed cyclone, let alone detect a false alarm. To this end, we are working on the classification of S2S tropical cyclone trajectories with clustering methods and we will show the first results. We aim to exploit the ensemble character of the subseasonal forecast for the development of future S2S-derived forecast products that would provide probabilities of scenarios of potential trajectories (based on these clusters).</p>

<p>In 2018 the long rains season in Kenya (March-May) was the wettest ever recorded. The country experienced several multi-day heavy rainfall episodes, leading to dam collapse, land and mudslides. 186 people died due to flooding and 300,000 were left displaced. </p><p>The Kenya Meteorological Department issued several advisories during the season that warned of heavy rainfall events a few days before their occurrence. Ahead of this no warnings were given.</p><p>However subseasonal forecasts gave strong indications of the heaviest rainfall episodes, several weeks in advance. With this extra lead time, preparedness actions may have been taken in order to reduce flood risk and save lives. </p><p>To this end, the ForPAc project (Toward Forecast-Based Preparedness Action) has been working in partnerships across Kenya and the UK to evaluate and build trust in subseasonal forecasts, and explore preparedness actions which could be taken in response. Most recently ForPAc has been granted access to real-time subseasonal data as part of phase two of the S2S pilot.</p><p>In this presentation we will first show analysis of the S2S hindcasts over East Africa, demonstrating the relatively high levels of subseasonal forecast skill and linking this to a strong MJO teleconnection that models capture relatively well.</p><p>In the second part we will describe work with stakeholders to co-design forecast products derived from the S2S data, concluding with a report on the forecasts for the ongoing 2020 long rains season and an evaluation of the way in which these have influenced disaster preparedness.</p>

How to improve the subseasonal forecast skills of dynamic models has always been an important issue in atmospheric science and service. This study proposes a new dynamical-statistical forecast method and a stable components dynamic statistical forecast (STsDSF) for subseasonal outgoing long-wave radiation (OLR) over the tropical Pacific region in January-February from 2004 to 2008. Compared with 11 advanced multi-model ensemble (MME) daily forecasts, the STsDSF model was able to capture the change characteristics of OLR better when the lead time was beyond 30 days in 2005 and 2006. The average pattern correlation coefficients (PCC) of STsDSF are 0.24 and 0.16 in 2005 and 2006, while MME is 0.10 and 0.05, respectively. In addition, the average value of PCC of the STsDSF model in five years is higher than MME in 7–11 pentads. Although both the STsDSF model and MME show a similar temporal correlation coefficient (TCC) pattern over the tropical Pacific region, the STsDSF model error grows more slowly than the MME error during 8–12 pentads in January 2005. This phenomenon demonstrates that STsDSF can reduce dynamical model error in some situations. According to the comparison of subseasonal forecasts between STsDSF and MME in five years, STsDSF model skill depends strictly on the predictability of the dynamical model. The STsDSF model shows some advantages when the dynamical model could not forecast well above a certain level. In this study, the STsDSF model can be used as an effective reference for subseasonal forecast and could feasibly be used in real-time forecast business in the future.