Forecasting tools were developed at five case study sites, with four in Europe and one in South Australia (Fig. 1). All are catchment–lake systems, and lake water level is regulated at all but the Irish site. The main characteristics of the catchments and lakes/reservoirs are given in Table 1, and catchment maps are shown in SI1 in the Supplement (see code and data availability section).

Mount Bold reservoir is the largest in South Australia. Its main water supply is the Onkaparinga River, but during the dry season inflows are supplemented with pumped water from the Murray River via a pipeline. Mount Bold provides water to the Happy Valley reservoir further downstream, which is the drinking water source for the city of Adelaide (around 1.3×106 people). Pumping water through the Murray pipeline is expensive, and operational decisions relating to pumping and release would benefit from advance knowledge of the likely hydrological conditions. In addition, the reservoir is susceptible to phosphorus resuspension, which contributes to algal blooms in the Happy Valley reservoir.

The Wupper reservoir is located near Cologne and is created by a large dam on the Wupper River. Its catchment is heavily built up compared to the other study sites, and the reservoir is an important recreational area. The water level is managed for flood control, maintaining environmental flows, and recreation, all of which are challenging as the water level fluctuations are large. Cyanobacterial blooms are problematic during hot summers and low water levels.

Lake Vansjø is located in one of the most productive agricultural areas of Norway which, combined with the prevalence of phosphorus-rich clay soils in its catchment, means it is prone to poor water quality. The lake has two main basins, an eastern and a western basin, and is a very important recreational area. The eastern basin (Storefjorden) is deeper and provides drinking water to the city of Moss (population of 60 000) and flows in to the shallower western basin (Vanemfjorden). The focus in this study was on the western basin, where algal blooms are particularly problematic. Toxic cyanobacteria blooms led to bathing bans in much of the period from 2000 to 2007, for example. Water level in the lake is regulated for hydropower, flood protection, and recreation.

Sau reservoir is the main water supply source for the Barcelona metropolitan area, with a population of up to 4.4×106 people. It is part of the Ter River catchment, which is the main source of water to the reservoir. The reservoir is vulnerable to both wet and dry seasonal climate events, as high river discharge washes in nutrients from the catchment resulting in poor water quality, whilst dry and warm seasons may result in low water levels, algal bloom development, and anoxia.

Sau, Mount Bold, and Wupper reservoirs are part of a larger chain of reservoirs, and water managers therefore face challenges in developing optimum release and pumping strategies throughout the chain. All the lake and reservoir sites face water quality challenges. High river discharge may wash in excess nutrients and lead to poor water quality, while prolonged dry and warm periods are often associated with algal blooms. The primary management opportunity at these sites is in adjusting water storage, release, and pumping strategies to minimise operational costs, while ensuring drinking water provision, flood protection, recreation, minimum environmental flows, and environmental protection. Advance warning of cyanobacteria bloom risk was of particular interest to the end-users in Lake Vansjø, in order to inform lake monitoring strategies, as was the likelihood of meeting water quality environmental targets. Additional background information on these four lake/reservoir sites is provided in Mercado-Bettín et al. (2021).

In the Burrishoole catchment in northwestern Ireland, the focus was on the timing of diadromous fish migration. This site is an extremely important Atlantic salmon and eel research catchment, with historic data on fish migration going back to the 1950s, together with a comprehensive catchment monitoring programme. The primary end-user interest at this site was the sustainable management of diadromous fish stocks and the development of a prototype seasonal forecasting tool that could be potentially transferred elsewhere in Ireland to inform the timing of eel trap and release schemes, for example, which are carried out in Ireland to enable eels to safely migrate around run-of-river hydropower structures.

The water managers engaged in the project included reservoir operators and water supply companies at the Mount Bold, Wupper and Sau reservoirs, the Morsa River Basin Management Authority in Norway, which is a partnership organisation responsible for implementation of the European Water Framework Directive (WFD; 2000/60/EC) in the catchment, and the Marine Institute in Ireland, who are responsible for diadromous fish stock monitoring and informing fisheries management (Table 1). One or two people were involved in the project from each organisation, including someone in a position to take water management decisions or, in the case of Mount Bold, responsible for research into new operational methods. A technical staff member (i.e. with responsibility for running models and analysing data) from each end-user organisation was also involved the project, where necessary. Water managers from Mount Bold and Wupperverband had experience in using catchment hydrology, water quality, and reservoir models, and the ultimate goal at these sites would be to run seasonal forecasting workflows in-house. At Sau and Vansjø, the aim was to develop a user-friendly forecast output rather than for the end-users to run the workflows themselves.

Water managers were involved in the design of the forecasting tools from the start as active project members to ensure that the tools matched their interests and needs. This also meant that they were able to interpret the probabilistic forecasts and the reliability information included with forecasts and, thus, carry out an informed assessment of the value of the forecasts for decision-making. Formal co-development and assessment exercises included the following:

An initial workshop to introduce seasonal forecasting, define the main management challenges and priorities at each site, and discuss ways in which forecasts could contribute to decision-making.

The surface water variables of interest varied between sites. The Irish site focused on the timing of diadromous fish migration to inform the monitoring and management of fisheries. Elsewhere, all water managers were interested in streamflow forecasts. Water temperature was also of broad interest, as it is the most basic water quality parameter which affects a host of other biogeochemical and ecological processes, and it can be simulated relatively robustly using process-based models. This was the water quality end-point forecast in the German, Spanish, and Australian sites. At Lake Vansjø in Norway, water temperature alone would not be enough to inform decision-making, and the end-user was also eager for forecasts of water quality parameters, in particular the risk of toxic cyanobacteria blooms.

A range of different models was used to produce forecasts for the freshwater variables of interest (hereafter termed freshwater impact models). In most cases, these impact models integrate seasonal climate forecasts, knowledge of antecedent conditions, and the characteristics of the system to predict the future state. The different variables of interest were simulated as follows:

Streamflow, lake level, and lake water temperature. Details of the modelling workflow used in the lake/reservoir sites are given in Mercado-Bettín et al. (2021). In brief, a process-based catchment hydrology model was used to simulate streamflow and, in turn, provided input to a process-based lake model which simulated lake water level and temperature. Differences in site characteristics led to different models being used at different sites. For catchment hydrology, the spatially distributed mesoscale Hydrologic Model (mHM; Samaniego et al., 2010) was used at Sau, the semi-distributed SimplyQ model (Jackson-Blake et al., 2017) at Vansjø, and the semi-distributed Genie Rural (GR) models, implemented within the airGR R package (Coron et al., 2017), were used elsewhere (GR4J at Mount Bold and Burrishoole; GR6J at Wupper). Lake thermodynamics and water level were simulated using GOTM (General Ocean Turbulence Model; http://gotm.net, last access: 8 March 2022) in Vansjø and Sau and GLM (General Lake Model; Hipsey et al., 2019) in Mount Bold and Wupper reservoirs.

For historic meteorological data, we used ERA5 reanalysis data, the latest reanalysis produced by the European Centre for Medium-Range Weather Forecasts (ECMWF; Hersbach et al., 2020), which is a global dataset with 0.25∘ horizontal resolution and hourly temporal resolution. For seasonal climate predictions, we used the ECMWF's most recent long-range forecasting system, SEAS5 (Johnson et al., 2019), a global dataset with 1∘ horizontal resolution. SEAS5 seasonal forecasts are publicly available as real-time operational forecasts from 2018 (50 members) and as retrospective seasonal forecasts (hindcasts) for the period 1993–2016 (25 members). Hindcasts were used in this study. SEAS5 was bias corrected using ERA5 with quantile mapping (see Mercado-Bettín et al., 2021, for details). All climate data were downloaded and post-processed using the climate4R bundle of packages (Iturbide et al., 2019). ERA5 is a natural choice to use in combination with SEAS5 as it is used to initialise SEAS5 and, therefore, ensures consistency between variables. Local sources of meteorological data could be used instead and may be less biased than ERA5. However, ERA5 is available on a global grid and for the period of time, variables, and temporal resolution required in the project, which was not the case for local data in a number of the case studies. A workflow using ERA5 is also more easily generalisable and transferable, which was one of the project's objectives. However, in the Irish site, local meteorological data were used to bias correct the ERA5 data. In future workflows, it would be worth exploring the newly available bias-corrected ERA5 data (Cucchi et al., 2020). Alternative seasonal climate forecasting systems were used at the start of the project, before SEAS5 became available (including CFS and SEAS4), without substantial differences in seasonal climate forecasting skill.

Forecasts of catchment discharge and lake temperature were produced four times a year for the boreal seasons of spring (March–May), summer (June–August), autumn (September–November), and winter (December–February). The fish model and GBN produce one forecast per year, for the months when seaward fish migration occurs in Ireland and for the 6-month (May-October) growing season used in ecological status classification in Norwegian lakes.

Seasonal climate forecasts are predictions of how the weather will evolve over the next season (typically 3 to 6 months ahead). Day-to-day forecasts are unreliable over such long horizons, so forecasts are instead used to say whether the next season will, on average, show broad differences to normal. Forecasts are, therefore, usually given as the probability of falling into one of three terciles: below normal, normal, or above normal. The statistical fish model uses terciles to summarise whether migration timing is likely to be early, normal, or late relative to normal. Instead of terciles, a binary classification was used to summarise water quality predictions in Norway, with the probability of being in two WFD-relevant classes (e.g. above or below good ecological status).

Quantification and effective communication of forecast quality is a crucial element of seasonal forecasting. Following the literature recommendations (Taylor et al., 2015; WMO, 2008), two kinds of forecast quality information were provided alongside forecasts:

Forecast uncertainty. SEAS5 has 50 ensemble members or 25 in hindcast mode. Each member is a priori equally likely and was used to produce 25 impact model forecasts. The divergence of members provides information about future predictability, with better agreement between members indicating higher predictability and lower forecast uncertainty. In practice, this information is given as the probability of the tercile, i.e. the percentage of ensemble members which fall into each tercile.

Following on from discussions at this workshop, the forecast presentation varied somewhat between sites, given a range of preferences. At all sites, however, the tercile probabilities and historic skill scores were categorised and accompanied by descriptive text summaries to aid in the interpretation. Managers were involved in deciding on the appropriate categories and wording. Tercile probabilities (i.e. agreement between ensemble embers) were split into four categories, i.e. very low (<35 %), low (35 %–49 %), medium (50 %–64 %), and high (65 %–100 %). For historic skill, the ROCSS text summary was either skilled or none, according to whether ROCSS was significantly positive or not. Historic skill given by MCC in Norway was summarised qualitatively as none (<0.2), low (0.2–0.39), medium (0.4–0.59), or high (>0.6). A combined confidence score was also provided, integrating the two types of forecast quality information. We opted to derive this by setting it to be the same as the tercile probability unless the historic skill was none, in which case it was also none. For water quality forecasts in Norway, if the class probability was high, then the overall confidence was the same as the historic skill; if class probability was medium, then the overall confidence was the historic skill reduced by one class. An example of a forecast can be seen for Lake Vansjø at https://watexr.data.niva.no/ (last access: 17 January 2022).

Windows of opportunity for seasonal forecasting were required for the second assessment exercise (Sect. 3.2). These were identified at each site using historic skill scores (Sect. 2.4). SEAS5 hindcasts were compared to ERA5, and impact model forecasts were compared to observations. Skill was calculated for every season in the 24-year period (1993–2016) or longer, where possible (1981–2019 for the GBN in Norway and 1993–2019 for fish migration timing in Ireland). At the Irish and Norwegian sites, real observations were used to assess skill. Elsewhere, forecasts were compared to pseudo-observations, i.e. model output derived by running models forced with ERA5 data. Skill calculated using pseudo-observations ignores impact model error and is therefore a best-case estimate. However, as seasonal climate skill is likely the largest source of uncertainty, this is still a useful first assessment of forecast performance. Statistical significance (95 % confidence) of ROCSS was then used to identify windows of opportunity, i.e. season, variable, and tercile combinations for which forecast performance was significantly better than expected from a forecast with no discriminative skill (ROCSS = 0). Windows of opportunity reported in this paper summarise those already reported in Mercado-Bettín et al. (2021) for Sau and Mount Bold and for streamflow and lake water temperature forecasting in Vansjø. In addition, we report updated results for Wupper, using an improved model calibration, and new results for Burrishoole and for lake water quality forecasting in Vansjø. The windows of opportunity identified present a useful first indication of where seasonal forecasts may be reliable enough to support decision-making, but they should be interpreted with some caution due to the small sample size (a short hindcast period is split into three terciles, i.e. eight data points per tercile), the use of pseudo-observations for some sites or variables, and the somewhat subjective 5 % significance threshold chosen to identify robust forecasting windows.

In the first assessment exercise, water managers were asked to choose a historic season in which seasonal climate resulted in problems in their study site. The events chosen are summarised in Table 2, along with associated surface water impacts and opportunities identified for mitigating the impacts, given a reliable enough forecast. Dry and hot seasons were chosen in Mount Bold and Wupper, with associated problems with low reservoir water levels, problems meeting demand, and poor water quality. A dry season was also selected in Burrishoole, which was accompanied by a later-than-normal salmon run. Prolonged wet periods and associated lake flooding and poor water quality were selected in Sau and Vansjø.

In the second exercise for exploring the potential for forecasts to support management, we carried out a more comprehensive assessment of whether the season, variable, and tercile combinations which could be forecast with reasonable confidence (the windows of opportunity) were considered useful for water management. Managers also highlighted additional windows which they felt would be particularly useful for supporting management, if forecast reliability were to increase.

There were few windows of opportunity in seasonal climate (Table A1). There were 0 windows in Ireland, 3 in Germany, 5 in Spain, 8 in Australia, and 10 in Norway. The 5 % significance level used means that some of these may be false positives (tests were carried out on up to 108 data slices per site, with four seasons × three terciles × up to nine variables, so we would expect on average five false positives per site).

A substantially larger number of impact model variables showed significant skill (Table 5). This suggests that antecedent conditions/inertia are responsible for much of the skill, further supported by the fact that bottom water temperature was better predicted than surface water temperature (15 versus 5 windows, respectively; Table 5), likely because of its lower sensitivity to seasonal climate (discussed in Sect. 4.2).

Opinions on the usefulness of the windows of opportunity are summarised in Table 5 (full responses are given in SI3; see the code and data availability section), together with the windows which water managers were most interested in. All the windows of opportunity for discharge were thought to be of medium or high relevance, and almost all combinations of season/tercile were highlighted as being desirable for management. For surface water temperature, spring to autumn above-normal forecasts were seen to be the most useful, due to often strong links between warm summer water and problematic algal growth. Many of the other windows of opportunity in surface water temperature were thought to be of medium or low relevance. As mentioned above, bottom water temperature was the variable that was most successfully forecast, with 15 windows of opportunity across the case study sites. However, it was also the variable that was thought to be least useful for management, with four of the windows being ranked as having low or no relevance. Overall, we found a mismatch between the variables that were thought to be most useful for management and those which could best be forecast. This can be seen, for example, in the difference between the number of current versus desired skilful windows for the different variables (Table 5; discussed further in Sect. 4.1 and 4.2).

In addition, a number of water managers commented that they would require information on more than just the most probable tercile, such as the likelihood of extremes, which are particularly challenging for management.

Responses from the Irish site are not shown in Table 5, as a large range of population statistics were explored. The following three windows of opportunity were found: early median day of migration for trout, later-than-normal day when 5 % of salmon have migrated, and later-than-normal day when 25 % of eel have migrated. These were all thought to be extremely relevant for management. The most desired windows of opportunity were for the day when 25 % of the population has migrated, the mean day of migration, and, for eel, the day when 75 % of the population has migrated, although the skill in the timing of all percentiles was of interest to check forecast consistency. Although all terciles were thought to be relevant, earlier-than-normal forecasts were considered the most useful, as acting on a wrong early forecast would have relatively minor consequences, whereas delaying action because of a later-than-normal forecast could result in fish mortality if the forecast were wrong.

Water managers were generally enthusiastic about the forecasting tools developed and their potential to assist them in preparing for the coming season. They identified actions that could be taken, given a reliable enough forecast, to help reduce the negative impacts of otherwise unforeseen events. They were well aware of the limited skill of many of the forecasted variables and were generally comfortable with the idea of working with probabilistic forecasts. For most sites, the act of setting up the impact models was in itself a valuable process, and managers were often enthusiastic about the new system knowledge gained in doing so and for the workflows to be more generally useful (for forecasting at shorter timescales, for example).

Despite general enthusiasm, no-one felt that forecasts could be incorporated directly into operational management straight away. In all cases, forecasts would only be used qualitatively in the first instance, to provide a general indication of how conditions might evolve, rather than to drive operational models (Sect. 3.1.2), matching the findings of Bruno Soares et al. (2018). Trust and lack of personal experience were key issues raised by most managers. The other key limitation, raised at all sites, was forecast quality (i.e. high forecast uncertainty and low skill; see Sect. 2.4). The similarity in responses across the contrasting study sites suggests that these manager viewpoints are likely to be more widely applicable. Forecast uncertainty was a barrier at most sites, as there was often close to an even distribution of probabilities across terciles for the selected historic events (Sect. 3.1.1), making it difficult for managers to know whether to act based on forecasts. While a reduction in forecast uncertainty (i.e. increased sharpness) could help in increasing the uptake of seasonal forecasts in operational water management, studies have shown that even highly uncertain forecasts can significantly improve reservoir management, as long as forecast uncertainty is explicitly accounted for. This could be done, for example, within an optimisation framework (e.g. Ficchì et al., 2016). More widespread use of optimisation routines in water management could also therefore increase the value of seasonal forecasts for management. Increasing the skill of forecasting systems remains a top priority, however (discussed further in Sect. 4.4), as, no matter how sharp the forecasts, they should not be used to inform management if the forecasting system has been proven to have no skill.

There was general enthusiasm for many of the windows of opportunity, i.e. the variables, seasons, and terciles for which the forecasting systems showed the most potential. However, there was often a mismatch between what could best be predicted and what was considered most useful (Table 5). Seasonal discharge forecasts were of particular interest, for example, but there were few discharge windows of opportunity. Bottom water temperature, meanwhile, could be forecast reasonably well at many sites, yet had limited management relevance.

At all sites, the sustained uptake of project outputs into the future requires operationalisation and maintenance, which is often a challenge as ongoing funding is required. An additional barrier to the direct uptake of the tools is that they do not include reservoir operations or interactions between management choices and catchment/lake conditions. However, at many sites there are plans for the knowledge and workflows generated to be used beyond the project duration. In Sau reservoir in Spain, for example, researchers are now providing monthly reports to the end-user for seasonal forecasts with a 1-month lead time, while in the Wupper reservoir in Germany, Wupperverband are incorporating the workflow, using a variety of climate products, into their longer-term management plans. At the Irish site, parts of the workflow (published in the fishcastr R package developed during WATExR; see code and data availability section) are already being used for other ongoing analyses of fish migration, but additional funding is required for the operationalisation of the forecasting system and expansion to a wider area. At the Mount Bold reservoir, SA Water were encouraged by the WATExR project results and have decided to invest in an internal follow-on project to establish a seasonal forecasting methodology.

As mentioned in the introduction, seasonal predictability in freshwater impact model predictions derives from knowledge of initial conditions and of climate over the target season. At the majority of study sites, there were a number of windows of opportunity where freshwater variables could be forecast with reasonable skill (Sect. 3.2). Seasonal climate forecasts themselves had very limited skill at these extratropical latitudes, so it seems likely that the impact model windows of opportunity were primarily due to models capturing how initial conditions and system inertia influence the target season. This could explain the better performance of bottom water temperature forecasts compared to surface water temperature and discharge forecasts, as the latter two are more sensitive to seasonal climate. Although further work is needed to confirm the sensitivity of the different variables to seasonal climate and initial conditions (and will be the topic of an upcoming paper), the initial indication is that those variables that are most sensitive to climate are the hardest to generate reliable seasonal forecasts for (due to low seasonal climate model skill in our study areas). However, in this case, they were also the variables which were most useful for management, given the importance of streamflow for reservoir operations and of surface water temperature in controlling algal blooms.

The low skill of the seasonal climate forecasts is typical of skill over much of Europe, parts of North America, Russia, northern China, and other mid-latitude areas (Johnson et al., 2019; MacLachlan et al., 2015). Seasonal forecasting to support water management in these areas with low climate model skill will, therefore, be largely reliant on initial conditions as the main source of seasonal predictability. High-quality forecasts, which can be used to inform management, are then most likely in catchments/lakes where initial conditions exert a larger influence. This is the case in larger systems and for variables or ecological species which are less sensitive to seasonal climate. For streamflow forecasts, for example, initial conditions provide much of the forecasting skill, and predictability is highest, in slower-responding catchments with larger water storage and groundwater contributions (Pechlivanidis et al., 2020; Girons Lopez et al., 2021; Donegan et al., 2021; Harrigan et al., 2018). Many successful streamflow forecasting systems do not include seasonal climate model forecasts, showing that a great deal can be achieved using only historic information, in some cases/seasons outperforming predictions which use seasonal climate model forecasts (e.g. Peñuela et al., 2020; Arnal et al., 2018). In lakes and reservoirs, the storage buffering effect has also been shown to reduce the importance of streamflow forecast skill, particularly when the reservoir capacity is large compared to variability in inflow (Maurer and Lettenmaier, 2004; Turner et al., 2017). Studies looking at sources of predictability for seasonal water quality and ecology are currently lacking, but similar concepts will likely hold.

Where seasonal climate forecasts are skilful, they undoubtedly have the potential to provide added value to surface water impact model forecasts. All water managers in this study were very enthusiastic about the potential benefit of skilful climate forecasts for improving their operations and were particularly interested in variables which are more likely to be sensitive to seasonal climate (e.g. discharge and surface water temperature). Several studies have shown that skilful seasonal climate forecasts can lead to sometimes large improvements in streamflow forecasting ability (e.g. Shukla and Lettenmaier, 2011), which may, for example, have economic value for reservoir operations (Maurer and Lettenmaier, 2004; Turner et al., 2017).

However, in our study catchments, seasonal climate models did not produce skilful forecasts for the selected historic events (Sect. 3.1), and there were few windows of opportunity for seasonal climate (Sect. 3.2). In areas where seasonal climate model skill is low, the extra resources required to work with seasonal climate data may not be worth potentially marginal performance gains, particularly as poor seasonal climate forecasting skill may reduce trust in any skilful impact model forecasts that use seasonal climate as input (Sect. 3.1.2). Particularly in larger catchments and lakes, which are less sensitive to seasonal climate, it is likely that attention would be better spent on developing simpler benchmark systems. Methods inspired by ensemble streamflow prediction are likely candidates, potentially made more nuanced by, for example, using the North Atlantic Oscillation (NAO) index or other climate signals to condition the forecast (e.g. Donegan et al., 2021; Najafi et al., 2012; Sabzipour et al., 2020), or longer-term climate projections (Gronewold et al., 2017).

In systems that are particularly sensitive to meteorological forcing at seasonal timescales (e.g. small catchments and lakes/reservoirs with short residence times) and where the benefits of any windows of opportunity are large (e.g. a large potential cost saving or particularly sensitive drinking water source/habitat), the potential benefits of incorporating seasonal climate data are greater. In this case, it may be worth incorporating seasonal climate data into the forecasting workflow, as long as there are some windows of opportunity at lead times of interest for management.

A key barrier to the use of seasonal forecasts in operational management is forecast performance, in particular poor historic skill (Sect. 4.1). To help improve performance, we see the need for progress to be made on the following two fronts:

Improvements in seasonal climate model skill. Seasonal climate models are under active development, and recent advances in the prediction of climate teleconnections, such as the NAO in Europe (Wang et al., 2017; Scaife et al., 2014), may lead to improvements in coming years.