Last modified on May 05, 2023 10:39

Contributors: A. Troccoli (WEMC), L. Sanger (WEMC), C. Goodess (WEMC), J. Ogonji (WEMC), L. Dubus (WEMC), R. Vautard F Pons (CEA), X. Jin (CEA), G. Levavasseur (CEA), R. Legrand (MF), L. Grigis (MF), S. Martinoni-Lapierre (MF), C. Viel (MF), S. Parey (EDF), B. Oueslati (EDF), Y-M. Saint-Drenan (ASSOCIATION POUR LA RECHERCHE ET LE DEVELOPPEMENT DES METHODES ET PROCESSUS INDUSTRIELS, ARMINES, FRANCE), J. Mendes (MO), J. Osborne (MO), G.Guentchev (MO)

Introduction

This document describes the technical methodologies and implementation of the climate and energy indicators underpinning the C3S Energy operational service. The tasks undertaken were geared to the provision of data and software to ECMWF that is deemed to be compliant with the protocol adopted by the CDS. Regular contact with the ECMWF technical teams was maintained to assist in the installation of all software components and data on the CDS infrastructure.

A technical documentation of the workflows and their modules is provided in this report. The workflows are structured according to the three streams covered by C3S Energy – historical, seasonal forecasts and projections – and describe the inputs of each routine, the outputs and each step in the processing. They summarise all the steps in the operational chain in graphical manner, which are also meant to assist ECWMF staff support in understanding the indicators production steps.

Workflows

The workflows are the structures that bring together all the elements of the C3S Energy operational system. We have produced one for each stream – historical, seasonal forecasts and projections. Each workflow covers both climate and energy indicators, which are used as reference for the service operational production and monitoring are presented in detail in the next sections.

Figure 1: Schematic of the C3S Energy operational system describing the technical production of climate and energy indicators for the historical (left), the seasonal forecasts (middle) and the projection (right) streams

General workflow description

Several steps follow each other, with a first processing of climate data, generally consisting of bias adjustment through adjusting the models to reanalyses or observations, re-gridding, and possibly selecting model output (in the case of projections). The second step is the calculation of energy indicators using the processed climate data and energy data and conversion models developed as part of C3S Energy, such as statistical models to calculate electricity demand for each country.

The C3S Energy workflows are based on the adaptations from the two proof-of-concept services CLIM4ENERGY (C4E) and ECEM, into a single workflow. These provided the initial set of software and data to be compared with. The exact indicators for the three streams were selected based on the experience of each service, which collected a large amount of user/stakeholder feedback, as well as input from ECMWF. The ECEM methodology for assessment and bias adjustment for GHI of historical data was adapted to ERA5 data (GHI was the only variable targeted for bias adjustment for the historical stream). However, the adjustment did not prove as effective as when applied to ERA-Interim and was therefore dropped. A chain allowing bias adjustment of EURO-CORDEX models developed in C4E, followed by indicator calculations done both in ECEM and C4E, has been further developed. A series of data assessments was performed throughout the stages based on protocols defined at the beginning of the contract, and also outlined in the Data Management Plan.

Historical stream

Figure 2: Workflow for the historical stream.

Data retrieval

The HISTORICAL stream workflow is described in Figure 2. Data retrieval of ERA5 data from CDS to C3S uses the CDS API and requires Python and the CDS API Python package to be installed before running. Data is retrieved by setting the period required and variables are required to be downloaded. Currently retrievals are performed in monthly chunks. This step has been automated in the final operational system, to periodically check and retrieve data from the CDS, at a monthly frequency, for monthly chunks, and aligned also with the requirements for the seasonal forecast stream.

NUTS computation

This step is common to several variables and streams in C3S Energy (and beyond). The computation of NUTS (Nomenclature des Unités _Territoriales Statistiques¹ ) averages for countries (also referred to as NUTS0) and NUTS2 provinces/regions/ is split into three steps, with shapefiles retrieved from Eurostat.

The procedure we adopted for NUTS averaging accounts for both the land-sea mask (LSM) provided by the model (e.g. ERA5), including its fractional values along water bodies, and latitude cosine factor.
Note that while the NUTS computation is essentially common for all C3S Energy indicators and streams, we have implemented two options – one using python as described above and an analogous one using cdo – due to the diversity of platforms and data characteristics (particularly for projection data). We have also attempted to use the NUTS computation provided by the C3S Toolbox, but since the data are being ingested in the CDS now, offline solutions had to be adopted.

For offshore areas, we developed regions equivalent to NUTS, based on several available datasets, as they did not exist. We generated two sets of regions: MAR0 and MAR1. MAR0 is a country level, similar to NUT0, while MAR1 is a sub-country level, broadly equivalent to NUTS1 (see Appendix 2 for further details).

Climate Indicators

The variables and indicators retrieved and used in C3S Energy are:

• 2 Metre Temperature
• Total Precipitation
• 10m Wind Speed
• 100m Wind Speed
• Solar Surface Radiation Downwards
• Mean Sea Level Pressure

These variables have been made available for the European domain (26.5° N to 72.5° N by 22.0° W to 45.5° E), and at 0.25° resolution, at NUTS0 and NUTS2 levels.
In addition to the standard agreed variables above, additional variables were retrieved and processed earlier, to allow more in-depth testing of some of the energy conversion models (e.g. use of snow depth for hydropower modelling). These variables were:

• Runoff
• Snow depth
• Evaporation
• Mean Evaporation rate
• Lake cover
• Lake depth

However, as they did not provide any significant improvement to the energy models, they have been abandoned.

Energy Data

Electricity demand (Load) is retrieved from the ENTSO-E Power Statistics database², in the Monthly Hourly Load Values tab³.
Generation from wind, solar and hydro power, and the corresponding installed capacities
data are retrieved from the ENTSO-E Transparency Platform⁴:

• Generation: Actual Generation per Production Type 

  (product name = Aggregated Generation per Type [16.1.B&C])

• Installed capacity: Installed Capacity per Production Type 

  (product name = Installed Generation Capacity Aggregated [14.1.A])

Generation and installed capacities are used in the set-up and validation of the hydro power models, and for the validation of the wind and solar PV power models validation. However, as it appears that ENTSO-E’s Transparency Portal installed capacity data present some issues, the formal validation of the datasets has been made using EUROSTAT data from the <nrg_inf_epc> dataset.

ENTSO-E data needed to be pre-processed to provide a harmonized format across all countries. This was done with R codes.

Energy Conversion models

The energy conversion models have been developed and validated on the historical stream, using the above mentioned data from ENTSO-E. The same models have then been used for the two other streams, SEASONAL and PROJECTIONS, with some minor adaptations to account for specificities of these two streams. The energy conversion models are presented below for the HISTORICAL stream, and only differences will be presented for SEAS and PROJ, where relevant.

Electricity Demand (EDM)

Model family: ✅ Statistical □ Physical

Following the approach developed in ECEM, and based on EDF's expertise, EDM is modelled for each individual country using Generalized Additive Models (GAMs). This approach requires observed load data to set up and validate the models. The data used are from the ENTSO-E Power Statistics database, described in section 2.4. The load data are displayed in Appendix 1.

The GAM predictors of each country vary, but they basically include different terms linked to:

• Country average (NUTS0) daily temperature
• Country average (NUTS0) daily GHI
• Country average (NUTS0) daily wind speed at 10 m
• Position of the day in the time history (time history is a linear variable which is 0 at the beginning and then increases linearly until the end of the period)
• Calendar data, including for instance holidays, day before/after holidays, type of day (Monday, Tuesday … Sunday)

Depending on each country's characteristics and load dependence on climate variable, additional terms could be included, as for instance:

• smoothed temperature, to represent the inertia in load due to temperature changes
• GHI or ws@10m conditional to cold/hot temperatures, or season

The demand models give overall good results with Mean Absolute Percentage Error (MAPE) of the order of 1-3% for most countries (validation made against ENTSO-E data on a period different from the one used for training).

For each country, the length of the actual load data varies, as well as the quality/homogeneity, as can be seen in the Figure in the Appendix. For each country, the demand model then needs to be set up with specific parameters, in particular, the start and end dates of the training and validation periods. However, for all of them we adopt the same general approach. This is sub- divided into 4 steps, as described in Figure 3.

Figure 3: Electricity demand model set-up steps

1. A first GAM estimates the trend on the longest possible period. Data from ENTSO-E start in 2006, but for most countries, it is sufficiently accurate only from 2010. This also avoids taking on board the data in 2008/2009, which shows an impact from the global financial crisis. This trend is then removed, so that the resulting time series has a mean level of the period considered, but no more multi-annual trend. The trend for France, for instance, is estimated on 2010-2018 (red arrow on Figure 4. Note that in thisvparticular case, the trend is rather small). Root Mean Squared Error (RMSE) and MAPE are calculated by comparing the reconstructed load and the actual one (for each step, we will use a corresponding subscript, e.g. RMSE1 and MAPE1. for step 1).
2. The dataset is divided into 2 parts. The GAM model is trained over the first one (2010- 2014 for France, blue arrow on Figure 4), and parameters obtained are used for the following steps.
3. The model built in step 2 is then used to simulate and test the load on the verification period (2015-2018 for France, green arrow on Figure 4); thus, RMSE3 and MAPE3 are the most stringent error metrics of the models.
4. Then, the full ERA5 data are used with the GAM parameters obtained at step 2 to reconstruct the full 1979-2019 period.

Figure 4 and Figure 5 show the outputs of steps 1-2-3 and 4 respectively, for the case of France.

Figure 4: GAM estimation steps. Case of France. The red arrow indicates the period over which the potential trend is estimated (here, 2010-2018). After removing this trend, the climate dependent part of the model is trained on the blue arrow period (here 2010-2014) and validated on the green arrow period (2015-2018)

Figure 5: Load time series reconstructed over France for the full ERA5 period 1979-2019. The red line shows the training and validation periods, the black line is the full, final time series.

To set up the models, an EDF-proprietary R toolbox is necessary, which defines the calendar data. As the toolbox cannot be shared, the necessary calendar data has been prepared for the 1970-2100 period, so that it can be used for the 3 streams.

The GAM parameters of each country are stored as .rds R data files, which can then simply be read to produce demand time series from any given climate dataset/stream.

The statistics of the 3 steps above for France are given in Table 1, while Table 2 presents the training and validation periods for all the countries, as well as the error metrics for the independent validation periods (Mean Absolute Percentage Error, Root Mean Squared Error and correlation coefficient). Finally, Figure 6 presents the scatter plots of reconstructed versus ENTSO-E demand for the 3 years 2015-2018.

Table 1: Demand model performance for France.

The complete validation for all countries based on RMSE and MAPE values is provided in Table 2. The scatter plots of reconstructed versus actual load are presented in Figure 6.

Table 2: Validation scores (step 3). The correlation coefficient is calculated against ENTSOE data over the same period for all countries, 2015-2018

Figure 6: Reconstructed versus actual load. Pearson correlation coefficient indicated in each panel

Hydro Power (HRE and HRO)

Model family: ✅ Statistical □ Physical

To model the hydropower generation, for both Run-of-River (HRO) and Reservoir (HRE), we opted for a machine learning method called Random Forest (Breiman, 2001). There are two noticeable features in this model: input with multiple lags of climate variables, and two steps of random forests to select only the most important variables.

To capture the effect of snow depth on hydropower indirectly, we introduced a new approach with lag time of temperature and precipitation as its proxy, considering that they both affect the time and the amount of snow melted into the rivers. "Multiple lags" means a large number of lagged climate variables with increasing lag time. Instead of increasing on one-day basis, these multiple lags are calculated with 5-day increment to reduce the computing time, assuming that they are able to capture the same effect. The maximum lag time is 200 days for precipitation, and 100 days for temperature. Table 3 gives a summary of input used in the model.

Table 3: Main features of climate variables used in the model.

As with any statistical model, it is better to avoid using too many input variables as they can cause overfitting, and the model can essentially interpret noise as a signal. To avoid this problem, two steps of Random Forest are implemented subsequently. Figure 7 illustrates the main processes including in the 2-step model.

In the first step, so-called the preliminary Random Forest, all of those multiple lags, together with their daily values, are the input of the training set while the target is the mean daily power data. Then, the most important variables are selected by comparing their drop-out loss values. Variables with a lag difference of 30 days or less are considered having the same effect and only the ones with highest drop-out loss value will be selected.
In the second step, another Random Forest is run with the same target, but only with selected important variables from the preliminary step. This model is stored to validate or for further applications such as data reconstruction, seasonal forecast and projections.

With this setup, the hydropower models for different countries can be built using the same pathway but result in different parameters (their importance variables), characterising each specific country.

Figure 7: Flowchart of the two-step Random Forest model for hydropower.

We compare our results also with a previous model using optimal lag (De Felice et al., 2018). Optimal lag is defined as the lag time having the highest correlation with power generation.

The Out-of-Bag estimates produced by the Random Forest model – the average of error from all the samples which are not used in growing each individual tree (Breiman, 1996) – are used to compare four model set-ups:

• Simple Random Forest model with predictors as daily temperature, optimal lag of precipitation (denoted Opt) and the same model with an additional variable – the optimal lag of snow depth (OptSD)
• Two-step Random Forest model with multiple lag sequences of temperature and precipitation (Mult) and a similar model with additional multiple lag sequences of snow depth (MultSD).

The Out-of-Bag errors produced by these four model set-ups are shown in Figure 8. In general, the skill of the model for HRO is higher than that of HRE. This makes sense since HRO does not have large reservoirs which require human management and thus is mostly influenced by climate variables. On the other hand, reservoir generation is more affected by the decision making processes.

In the first comparison using models with optimal lag, snow depth (optSD) improved the model's performance significantly. However, the second comparison with multiple lags showed no clear difference whether snow depth was included or not. This approach even gave better results than the first two setups in most countries, probably because it is able to capture other underlying effect(s) beside snow depth. Besides, snow depth variable is not highly accurate, or even available, in seasonal forecast and some climate projections. This justifies our choice of not including snow depth in the final model.

To reconstruct the energy generation in 1979-2014, a 2-step random forest model without seasonal division was trained on the full available dataset 2015-2018. Assume that the installed capacity for both the HRO and HRE remain unchanged in these countries from 1979 to 2018, the model was then using the climate data from ERA5 1979-2014 to reconstruct the hydropower generation for the whole period 1979-2018, including periods with missing data from ENTSO-E (Figure 9).

Additional details can be found in the recently published paper Ho et al. (2020).

Figure 8: Pearson correlation from out-of-bag estimation from the random forest models for HRO (upper) and HRE (lower) in 2015 – June 2018 with predictors: (1) daily temperature, precipitation and optimal lag of precipitation (opt – Brown); (2) daily temperature, precipitation, optimal lags of precipitation and snow depth (optSD – orange); (3) multiple lags of temperature and precipitation (mult – dark green); (4) multiple lags of temperature, precipitation and snow depth (multSD – turquoise).

Figure 9: Reconstruction of HRO and HRE generation from 1979 – 2018 (orange) with the 2-step random forest model. The observed data from 2015-2018 was provided (turquoise) to compare. The timeseries were smoothed with rolling mean of 30 days to eliminate very high variations Solar PV Power (SPV)

Solar PV Power (SPV)

Model family: ✅ Statistical ✅ Physical

The solar PhotoVoltaic (PV) capacity factor is calculated at grid point level. It is important to highlight that this quantity does not represent the power production of a single PV system. Instead, it is designed to represent the aggregated production of the PV plant installed in each pixel. For this purpose, the power production of a PV system is calculated from the meteorological data (GHI and 2-m temperature) for difference module orientations using a reference PV plant model using empirical models of the main parts of a PV system (optical losses, module, inverter). These different power values are then aggregated assuming a distribution of the different module orientations for the considered location. The distribution of module orientation is dependent on the location via the optimal tilt angle using a parameterization developed in the ECEM project.

A detailed description of the calculation steps followed for calculating the capacity factors from ECEM climate-adjusted dataset parameters GHI and T2M can be found in Saint-Drenan et al. (2018)⁵. The model has been adapted in C3S Energy to use ERA5.

Wind Power, onshore and offshore (WON and WOF)

Model family: □ Statistical ✅ Physical

Like for solar PV power, it is difficult, if not impossible, to know accurately in real-time what the EU wind fleet looks like, as new wind farms are built regularly while others are decommissioned. A simple solution to assess the effect of climate on wind power generation is to consider a homogeneous distribution of wind turbines all over Europe, calculate the capacity factor at the model grid level, and then aggregate at (NUTS0) country or NUTS2 level. In addition, all the wind turbines are considered with the same hub height, here set equal to 100 m. This will of course not reflect the actual capacity factor as can be reported by grid operators, but it is an efficient way to build a time series which is well correlated to the actual one.

Wind turbines power curves come from thewindpower.net⁶ database. This is a commercial database that has been used for ECEM too. However, currently only typical wind turbine characteristics are used, which are broadly based on the features provided by this database but are also available through other publicly available sources

In order to differentiate onshore and offshore areas, two different wind turbines were selected, based on expert knowledge. They represent current day most selected wind turbines. These wind turbines are:

• Onshore : Vestas V136/3450 (3.45 MW)
• Offshore : Vestas V164/8000 (8.0 MW)

The corresponding wind power curves are plotted in Figure 10. These power curves have been retrieved from thewindpower.net power curves data base.

Figure 10: Onshore (green) and offshore (blue) wind turbines power curves.

For the historical period, the wind speed at 100 m comes from ERA5. In the reanalysis, the LSM varies from 0 to 1, by 0.1 increments. Here, the following assumption has been made:

• LSM >= 0.5 is a Land point
• LSM < 0.5 is an Ocean point

With this choice, each grid point is either onshore or offshore, and the corresponding power output is computed. We then end up with two files for each wind speed file, one onshore and one offshore. Figure 11 shows the wind capacity factor from ERA5 on 2017-01-01 00:00 UTC.

Figure 11: Wind capacity factor for offshore (left) and onshore (right) regions, based on ERA5 wind speed at 100m, on 2017-01-01 00:00 UTC.

Energy Indicators averaging

Once the capacity factors for wind and solar power are calculated at grid level, they are aggregated to NUT2 and NUTS0 (and MAR0 & MAR1 for offshore wind) using the procedure discussed in section 2.2. This step is not necessary for the electricity demand and hydropower indicators since they are computed only for country averages (NUTS0 level), as presented above. As for spatial averaging, energy indicators are also produced and made available for longer timescales, namely daily (if the highest resolution of the indicator is sub-daily), monthly, seasonal, and annual averages. Indicators are also produced using three different units, whenever appropriate, namely capacity factor (unitless), power (MW) and energy (MWh). The energy indicator production is shown in Table 4.

Table 4: Summary table of indicators calculated. ✅: done and available; ⚙: work in progress; grey cells mean indicator is not calculated

Seasonal Forecast stream

Figure 12: Workflow for the seasonal forecast stream.

Data retrieval

Seasonal Forecast (SF) data are retrieved via the CDS API (python CDS API required), using python. Due to some field number limitations on the CDS, the raw hindcast/forecast are retrieved separately. The ECMW SY05 hindcasts were retrieved and bias adjusted in a single step (as the model version is frozen). Since MTFR SY07 is available only since November 2019, new hindcast months are retrieved and processed every month, as they are produced. The difference between the ECMWF system, Météo-France system and the Met Office system is the rolling release of the Met Office's hindcast whereas ECMWF and Météo-France keep the model version fixed for a longer period (typically 3-4 years).

The TA and WS at 10 m climate indicators are processed at a six-hourly time-step whereas cumulated climate variables, such as GHI and TP, are been processed at a daily time-step. Since ERA5 data original resolution is (close to) 0.25° spatially and hourly temporally, a spatial interpolation to the corresponding 1° resolution grid is automatically performed directly during the data retrieval (on the CDS's server side). Regarding TP and GHI raw hindcasts and forecasts, the SF provides values that are cumulated amounts since the beginning of the run. To be used properly, the codes that open the raw cumulated values include a conversion from cumulated values since the beginning of the run into the suitable daily amounts.

The hindcasts period covers the period 1st January 1993 to 31th December 2016; the 1993 start date corresponds to the major improvement of the observations and reanalysis datasets and is common to all C3S seasonal forecast systems.

There is a difference in the grid used by ECMW SY5 on one hand and for MTFR SY07 and METO SY14 on the other hand: 181° longitude grid centred on 0° for ECMWF and a 180° longitude grid centred on +0.5°, +0.5° grid for MTFR and METO. After exchanges with ECMWF, it should be noted that this difference will persist in the whole ECMW SY05 dataset. That minor issue was obviously taken into account in the SF compilations and will also be clearly flagged for every CDS's users.

Climate Indicators

The four indicators (TA, TP, WS at 10 m and GHI) are all at 1° spatial resolution and their temporal resolution is their native temporal resolution of the SF systems: six-hourly for TA and WS, 24-hourly for TP and GHI.
Both ECMW SY05 and METO SY14 forecasts/hindcasts cover a 215 days period (i.e. 860 six- hourly steps). MTFR SY07 forecasts/hindcasts cover a 211 days period (i.e. 844 six-hourly steps).

Bias adjustment of the climate indicators

Both hindcasts and forecasts of SF systems are markedly affected by bias. In mid-latitudes, mean error commonly ranges from ±4° for temperature, ±3 m/s for wind and up to 5 mm/day for total precipitation (see figure below). Raw bias maps and time series plots are available for all initialization months, models and variables on the CDS's public partition:

/data/public/C3S_ENERGY/SEAS/<clim,ener>/<ec,mo,mf>//FIG/*mean_error_ H_raw_1993-2016*

Figure 13: Raw mean bias for TA (K, METO SY14), TP (mm/day, ECMW SY05) and WS (10m, m/s, MTFR SY07). Initialization on November, validity Dec-Feb.

The inherent systematic bias in seasonal forecast is reduced through a quantile mapping (QM) approach. QM was chosen as it is a robust and widely used approach. The bias-adjustment is trained over the hindcast period. For each variable, hindcasts cumulative distribution function (CDF) is constructed. Then same process is applied to ERA5 since it is taken as the reference (graphical explanation on Figure 14). Finally, with the same statistical transform enabling to pass form hindcast CDS to ERA5 CDS, the same treatment is applied to the raw forecast's CDF in order to obtain the bias adjusted forecast.

The quantile mapping method may be summarized in two steps:

1. Determine the percentile rank of a biased forecast value using its own empirical CDF;
2. Retrieve and replace the reference value associated with the same percentile within the reference CDF.

Few technical steps are also important to consider:

• Percentile computation: Given a RUN-month and a forecast length, each percentile is computed from a sample including all the hindcast members for the whole period and within a 31 days climatic window that includes the 15 days before and after the d-day. If the parameter is given in a sub-daily basis (instantaneous values such as temperature and wind) the values are taken at the same validity hour of the As an example, the resulting samples for ECMWF sys5 hindcasts bring together 18600 realizations (31 days * 24 years * 25 members). This procedure aims at enforcing the statistical robustness of the percentiles and therefore the pertinence of the use of QM in our case. The same procedure is applied to the reference ERA5 data. Around a validity time, the sample is formed over the whole period and the climatic window. An ERA5 reference sample contains 744 realizations (31 days*24 years).

• Mapping: Practically only 99 percentiles (1%, 2%, …, 99%) are computed and stored from references and hindcasts CDF (respectively [p_r,1, p_r,2, …, p_r,99] and [p_f,1, p_f,2, …, p_f,99]). Each biased forecast values f will be first ranked over the percentiles to find ‘i’ given as: p_f,i<f<p_f,i+1. The estimated unbiased forecast, is then calculated from a linear interpolation of p_r,i and p_r,i+1: 

  $$\hat{f} = p_{r,i} + \frac{f-p_{f,i}}{p_{f,i-1}-p_{f,i}}(p_{r,i+1}-p_{r,i})$$
• Dealing with the extreme forecast values: If a forecast value f exceeds the extreme percentile of an hindcast CDF, this method cannot be applied A closer look to the extreme value requires custom data handling:
  • if f>p_f,99+threshold, = 99+threshold. Indeed, for instance regarding the amount of precipitation, a new extreme value can be far away from the upper percentile value, we must reject the very high forecast value and limit them to a «reasonable» extreme value. The process is the same for low extremes.

Hindcasts bias adjustment quality assessment

After bias adjustment, the hindcast and the reference reanalysis are compared to check that they share a similar distribution. As for an example, TA at step +180 days of ECMWF forecast show a raw hindcast distribution (black) that is colder and with a different shape than the reference distribution (red). After bias adjustment (blue distribution is bias-corrected), the distributions are quite similar and the differences are mostly due to (random) forecasting errors.

Figure 14: Schematic for the Quantile Mapping bias adjustment approach


Figure 15: Example of comparison between original (dark grey) and bias-adjusted (blue) hindcast distribution, with reference to ERA5 (red).

When displayed year by year and by members, hindcast three-month averages give a view of the hindcasts dispersion around the reference values. The raw hindcast often appears far to the ERA5 reference values and after bias-adjustment, the values are often closer to the 'observations'.

In Winter, raw MTFR SY07 and METO SY14 hindcast show a strong hot bias of approximately +1.5°C. On the contrary, ECMW SY05 suffers from a –1.5°C cold bias (Figure 16).

Figure 16:Raw hindcasts by year and by members (black dots) and reference ERA5 values (blue dots). TA (°C) for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec-Feb

After bias adjustment the members are more centered around the ERA5 reference values. The very low values of the month of November for MTFR SY07 are potentially due to an error (this is currently under investigation).

Figure 17:Bias-adjusted hindcasts by year and by members (black dots) and reference ERA5 values (blue dots). TA (°C) for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec-Feb

Concerning WS, MTFR SY07 has the largest bias, around +1m/s, ECWW SY05 is slightly too strong but is relatively near to the reference. METO SY14 is too weak by –0.5m/s (Figure 18).

Figure 18:Raw hindcasts by year and by members (black dots) and reference ERA5 values (blue dots). WS- (m/s) for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec- Feb.

After quantile mapping, all the three models' values are well centered around the reference, no obvious bias is visible (Figure 19). 

Figure 19: Bias-adjusted hindcasts by year and by members (black dots) and reference ERA5 values (blue dots). WS (m/s) for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec-Feb. 

For GHI in winter, all the three models are strongly under-estimating the truth (Figure 20).

Figure 20: Raw hindcasts by year and by members (black dots) and reference ERA5 values (blue dots). GHI (daily mean amounts in W/m**2) for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec-Feb.

After bias-adjustment the bias is reduced but still visible for METO SY14 (Figure 21).

Figure 21: Raw hindcasts by year and by members (black dots) and reference ERA5 values (blue dots). GHI (J/m**2) for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec- Feb

Figure 22: Bias-adjusted hindcasts by year and by members (black dots) and reference ERA5 values (blue dots). GHI (daily mean amounts in W/m**2) for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec-Feb.

Figure 23: Bias-adjusted hindcasts by year and by members (black dots) and reference ERA5 values (blue dots). For raw TP- hindcasts in Winter, MTFR SY07 is affected by a strong wet bias (it is also the case for METO SY14 in a lesser extent). ECMW SY05 seems to show a slight dry general bias.

Figure 24: Raw hindcasts by year and by members (black dots) and reference ERA5 values (blue dots). TP- (mm/day) for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity
Dec-Feb

After bias adjustment, TP- hindcast are more around the reference values but it is worth to notice that some years are still at the extremities or even out of the members range for wet (1994, 2013) or dry (2005) years and for the three models.GHI (J/m**2) for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec-Feb

Figure 25: Bias-adjusted hindcasts by year and by members (black dots) and reference ERA5 values (blue dots). TP- (mm/day) for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec-Feb

To continue the performance assessment, the Relative Operating Characteristic (ROC) curves are studied. ROC evaluates the capacity of the models to detect an event. The three chosen events are three 'observation' (reanalysis) thresholds: lower (blue on the plots), middle (black) and upper (red) terciles values. Four hindcast forecasting thresholds were used: 20%, 40%, 60% and 80%. Each point of ROC is defined by the Hit Rate (HR) and the False Alarm Rate (FAR). The values of HR and FAR are computed from three-months averaged values among the 24 years hindcast period.

The more the curve is near to the upper left corner the better it is, i.e. FAR close to zero and HR close to 1. Also the higher the Area Under the Curve (AUC), the better the forecasting skill. It should be noted that AUC less or equal to 0.5 means that forecasts do not discriminate between HR and FAR.

For TA, the AUC are around 0.6 in Winter, averaged on the entire Europe domain. This means that the correct forecast rate is just a little bit higher than the false alarm rate, the forecast skills are low, but still the seasonal forecast beats random climatological forecasts (not shown).

Figure 26:ROC curves. TA- for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec-Feb

Skills are a little bit better in Summer, with AUC up to 0.67 for the METO SY14's upper tercile threshold.

.

Figure 27: ROC curves. TA for ECMW SY05 (left) and METO SY14 (right). init June, validity Jul-Sep

For WS- (10m), the values are not much higher than 0.55, the skill is very poor.

Figure 28: ROC curves. WS for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec-Feb

Skill is even worse in Summer with hardly any skill shown.

Figure 29: ROC curves. WS- for ECMW SY05 (left) and METO SY14 (right). init June, validity Jul-Sep

Concerning GHI in Winter, the skills are quite low, with around 0.55 AUC values.

Figure 30: ROC curves. GHI for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec-Feb

Here again, Summer skills are not good, with METO showing no signal and ECMW only reaching an AUC of 0.55.

Figure 31: ROC curves. TA- for ECMW SY05 (left) and METO SY14 (right). init June, validity Jul-Sep

For TP in winter, METO SY14 and MTFR offer a little seasonal signal and ECMW SY05 show very poor skill.

Figure 32: ROC curves. TP for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec-Feb

In Summer, when averaged on the whole Europe, both ECMW SY05 and METO SY14 do not give any pertinent seasonal signal.

Figure 33: ROC curves. TP for ECMW SY05 (left) and METO SY14 (right). init June, validity Jul-Sep

The ROC figures are averages on the entire Europe, the skills are generally low, but there might exist some relatively good skills area hidden by lower ones as can be seen in the Anomaly Coefficient Correlation (ACC) maps. ACC is a skill score that shows how the forecast anomalies are correlated to the observed ones. The ACC spreads between 1 (perfect correlation) and –1 (anticorrelation). A 0 value of the ACC means very poor skill.

For TA in winter and fall (not shown), models usually show very poor skills on the continent vi ACC always < 0.4 (even if Scandinavia is slightly better forecasted). The skills are higher on the oceans and the norther the better. The Mediterranean basin appears to be the worst region for seasonal forecasting temperatures with <+/-0.2 ACC.

Figure 34 : Anomaly Correlation Coefficient maps. TA- for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right) init November validity Dec-Feb

In summertime, the skills are better, especially on the eastern Mediterranean basin (ACC>0.6). Still, western Europe (barely Brest-->Helsinki) remains with very low seasonal predictability (ACC<0.2).

Figure 35: Anomaly Correlation Coefficient maps. TA- for ECMW SY05 (left) and METO SY14 (right). init June validity Jul-Sep

Concerning WS (10m) the skill is quite low in Winter; the ACC are generally under 0.4 and often null or negative. The UK and Denmark areas seem slightly better, with ACC up to about 0.6. METO SY14 is better than the other onshore Germany to occidental Russia.

Figure 36: Anomaly Correlation Coefficient maps. 10m WS for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right) init November validity Dec-Feb

In Summer the skill is significantly lower, as shown for ECMW on the Northern Europe (negative ACC), METO SY14 is in accordance with ACC's mean value on the domain near to 0.

Figure 37: Anomaly Correlation Coefficient maps. WS for ECMW SY05 (left) and METO SY14 (right). init June validity Jul-Sep

The GHI exhibits low skill over most of the domain, only Scandinavia and the Azores are over 0.4.

Figure 38: Anomaly Correlation Coefficent maps. GHI for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right) init November validity Dec-Feb

In Summer the skill is a little better in the neighborhoods of Ukraine (ACC>0.4 and locally 0.6) or between Iceland and Norway, but much worse from France to Finland with null to strongly negative ACC.

Figure 39: Anomaly Correlation Coefficient maps. GHI for ECMW SY05 (left) and METO SY14 (right). init June validity Jul-Sep

TP is generally poorly forecasted over Europe, but the Maghreb and Turkey exhibit ACC > 0.5 for MTFR and Finland for METO.

Figure 40: Anomaly Correlation Coefficient maps. TP for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right, init November validity Dec-Feb

The skill is again worse in Summer, especially on around Denmark for ECMW SY05 and on a line from Galicia-Denmark-Sea of Kara for METO SY14 with negative ACC.

Figure 41: Anomaly Correlation Coefficent maps. TP for ECMW SY05 (left) and METO SY14 (right). init June validity Jul-Sep

Forecasts operational chain

The forecast of the four climate indicators are bias adjusted via the quantile mapping method. Then anomalies (from bias-adjusted hindcasts reference) and ensemble mean (average of the forecast members values) are computed and released on a step by step basis. The three terciles (lower, middle, upper) probabilities are also computed and released step-by-step. Those five statistics are made available separately, providing five gridded NetCDF (and csv for NUT0 aggregated values) files by indicators. The bias-adjusted forecasts are updated every month and are computed just after their raw version's releases. 2 meters dewpoint temperature and mean sea level pressure are also available in order to serve as input values for cases studies.

Figure 42: Anomaly forecast maps TA (K) for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init November 2019 validity Dec-Feb 2019/2020

The climate parameters are then used with the energy conversion models to calculate the energy indicators. While wind and solar PV production are embedded in the SEAS chain, hydropower and demand models are rune separately, with the following differences with respect to the HISTORICAL stream:

Hydropower generation: as the predictors include lagged temperature and precipitation, priori to the forecast date, these are taken from ERA5. This means the SE AS forecasts for HRE and HRO can be run only when data from ERA5 are available until the end of the month M-1 for the seasonal forecast of month M. In practice, using ERA5T means the seasonal forecasts for HRE and HRO can be run approximately from the 5th of each month.

Electricity demand: similarly, one of the predictors of the demand models is a smoothed temperature (with a smoothing parameter specific to each concerned country), which mimics the inertia of buildings. Like for hydropower, the ERA5 data is taken to calculate this smoothed temperature. Similarly, it means the seasonal forecast for demand of month M can be run once the full M-1 month of ERA5 is available.

Figure 43: Anomaly (left, mm) and ensemble mean (right, mm/month) forecast maps for TP. Three- month averages, initialization November 2019 validity Dec-Feb 2019/2020

Energy Indicators quality assessment

The hindcasts values are used to adjust the parameters of the energy models. As an example, daily mean and country aggregated TA and WS, along with daily and country aggregated GHI are the inputs for the SF demand indicator. For hydro-power, HRO and HRE require daily and country aggregated TP and daily mean and country aggregated TA. The forecast GHI and TA values are the input data to calculate the solar PV power (SPV) indicator.
To assess the quality of hindcast period energy indicators, the historical stream dataset was used. The raw/bias-adjusted energy hindcasts presented here are the raw/bias-adjusted on which where applied the seasonal energy models.

Offshore Wind Production (WOF)

The hindcasts by year and by members are similar in their shape to the WS hindcast, but the maximum values are more limited, this may be due to the fact that the wind turbine is turned off for winds stronger than 25.5 m/s.

Figure 44: Raw hindcasts by year and by members (black dots) and reference ERA5 values (blue dots). WOF (CFR: kWh/kWh) for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec-Feb

After bias adjustment, the values are nearer and more centered around the reference ERA5 values.

Figure 45: Bias-adjusted hindcasts by year and by members (black dots) and reference ERA5 values (blue dots). WOF (CFR: kWh/kWh) for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec-Feb

The patterns of the ACC are also similar to the WS ones, but a careful look reveals little differences like in the North Sea where the ACC are lower for WOF than for WS. Offshore UK and Portugal and, to a lesser extent, the German coasts of the North Sea exhibit a little bit of seasonal forecast skill. On the Mediterranean basin the skill is lower and more varying on relatively short distance. This is by itself is a satisfying result which shows that the wind power model does not harm the skill too much.

Figure 46: Anomaly Correlation Coefficient maps. WOF for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init November validity Dec-Feb

Onshore Wind Production (WON)

The raw hindcasts by year and by members are similar in their shape to the WS hindcast.

Figure 47: Raw hindcasts by year and by members (black dots) and reference ERA5 values (blue dots). WON (CFR: kWh/kWh) for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec-Feb

Similarly, to the WOF hindcasts, the bias-adjusted hindcast are closer to the reference ERA5 values.

Figure 48: Bias-adjusted hindcasts by year and by members (black dots) and reference ERA5 values (blue dots). WON (CFR: kWh/kWh) for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec-Feb

UK and the Northern parts of Germany and Polonia are relatively well forecasted with ACC often over 0.5. Iceland, Italy and the south of the central Europe show very poor skill with ACC under 0.2 and even negative on the Balkans.

Figure 49: Anomaly Correlation Coefficient maps. WON for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init November validity Dec-Feb

Solar Photovoltaic Production (SPV)

After bias adjustment, both MTFR SY07 and ECMW SY05 are affected by a positive bias in Winter, METO SY14 is much closer to the reanalysis energy values.

Figure 50: Bias-adjusted hindcasts by year and by members (black dots) and reference ERA5 values (blue dots). SPV (CFR: kWh/kWh) for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init Nov, validity Dec-Feb

Still in winter, MTFR SY07 and ECMW SY05 are correct (ACC>0.5) in Scandinavia. ECMW SY05 is not very skillful on the rest of the domain. MTFR SY07 show some predictability over Tunisia and Libya whereas METO SY14 is not that bad on a Catalonia-Petersburg line.

Figure 51: Anomaly Correlation Coefficent maps. SPV for MTFR SY07 (left), ECMW SY05 (center) and METO SY14 (right). init November validity Dec-Feb

Projection stream

Figure 52: Workflow for the projection stream.

Data retrieval and bias adjustment

We have considered climate projection data for 12 combinations of GCM and RCM from the EURO-CORDEX project, for both RCP4.5 and RCP8.5 scenarios, two of which also include data for the RCP2.6 scenario. The selected climate variables were temperature at 2 m (tas), wind speed at 10 m (sfcWind), solar radiation at surface (rsds), the three of them with a 3-hourly time step, and daily precipitation (pr). The relatively high temporal resolution is required for the conversion into energy indicators since 3-hourly data are necessary (for temperature, wind and solar radiation) in order to compute wind and solar capacity factors. Note that only data at the daily frequency are normally made available on the on the Earth System Grid Federation (ESGF) portal and only a small fraction of datasets is available at 3-hourly time step on the ESGF portal. The original EURO-CORDEX grid resolution is 0.11 degrees (~ 12 km) for all datasets. Table 5 shows the list of the 12 EURO-CORDEX simulations considered here, the corresponding contact institutes and the period covered by the climate simulations.

Table 5: List of the 11 EURO-CORDEX simulations and the institutes that provided the data.

Data files were provided in NetCDF format. As a first step, a quality check was performed on the data files, considering time series length and consistency of initial dates and calendar. The performed corrections included:

• correcting variable names, times, grid metadata
• filling missing data with repeated values or linear interpolation
• converting radiation data from instantaneous to accumulated values
• correcting normalization factors
• truncating data to full years

After the initial quality check, climate projections were adjusted to remove the intrinsic bias which affects every climate model, due to simplifying assumptions and approximations in the model physics. The bias adjustment approach applied here is the Cumulative Density Function (CDF) transform (CDF-t) method, first introduced by Michelangeli et al. (2009), and modified recently, for precipitation (Vrac et al., 2016; Bartok et al., 2018). CDF-t is a non-parametric technique aimed at matching the CDF of the climate simulation of a certain variable to the CDF of the same variable provided by observations or reanalysis, over a training period over which these are available, and using this relation in combination with the relation between a reference and a future period. The training reference period used here was the 1979-2018 period from the ERA5 reanalysis dataset.

The ERA5 data was first prepared by subsampling 3-hourly data at 00, 03, 06, 09, 12, 15, 18, 21 UTC or averaging between these hours in the case of solar radiation. The grid resolution was changed from 0.11 to 0.25 degrees (~ 25 km) to match ERA5's spatial resolution. This change was made within the bias adjustment program. Since EURO-CORDEX data is available over a finer resolution grid (0.11°, rotated polar grid) than ERA5, an average was made over EURO-CORDEX grid points falling within each ERA5 grid.

In order to consider climate change, bias adjustment was performed on moving windows of 20 years, with a 10-year advancement. For each 20-year period, the adjustment was calculated using the training set and the 20-year period, but the adjustment was saved only for the central 10 years. Then, a new 20-year window was considered, but only with a 10-year shift respect to the previous 20-year window.

The bias adjustment method was applied separately on each climate variable. Furthermore, model and reanalysis data were separated for each month and each available hour for 3- hourly variables, to allow the computation of bias adjustment by month and hour. This step was necessary to preserve dominant frequencies in the autocorrelation, such as daily and seasonal cycles. Months and hours were then merged after bias adjustment, to obtain an adjusted dataset.

A "health check" procedure was applied on bias adjusted data to verify:

• that the bias adjusted projections had similar seasonal means over the reference periods; by mapping the absolute differences, it was found that they are an order of magnitude smaller than the variables variations;
• that the changes between end of current and past centuries were similar with or without bias adjustment. We found that while the general patterns were quite similar, changes amplitudes can be modified by the CDF-t method.

Bias adjusted datasets, daily for precipitation and 3-hourly for the other variables, were saved to NetCDF files, quality checked and standardized to be compliant with the DMP. The CMOR standardization was used for consistency with ESGF standards. Each file has a size of ~ 50-60 Gb. They are made available and uploaded to ESGF (https://esgf-node.ipsl.upmc.fr/search/cordex-ipsl/).

NUTS computation

Datasets of daily values at the country (NUTS0) and regional (NUTS2) level are made available for climate indicators following the methodology described in Section 1.1.1. The wind speed at 10 and 100 metre heights is also aggregated over maritime regions, defined at two different levels, denoted MAR0 and MAR1, in analogy with NUTS0 and NUTS1 (clusters of regions) levels.

In particular, for climate projections, the NUTS averaging is performed using CDO, using a LSM and cosine weights saved as NetCDF files.

Energy Conversion models

The computation of gridded energy variables for the projection stream relies on the same models and procedures as described in section 2.5. However, the wind power computation differs in one step, as it requires in input the wind at 100 m height, which is not a direct output of the EURO-CORDEX models and must then be obtained as a derived variable.

Wind at 100 m

The wind speed at 100 m height is estimated from the wind speed near the surface, at 10 m. To accomplish this, we use a power law defined as follows:

where the exponent α has been estimated by taking the ratio between the ERA5 reanalysis 100 m and 10 m wind speeds. This ratio is computed for each grid point at hourly frequency and then taking the average over the 2000-2017 period (the dataset available at the time of computation). The resulting map is shown in Figure 53. Even though this does not account for daily or seasonal variability in the relationship between wind at 10m and wind at 100m, it accounts for spatial features, especially the difference between onshore and offshore values, and it is therefore considered a marked improvement over the use of a constant α factor, independent of location. For completeness, the standard deviation of α has also been computed to assess the variability of the field (not shown). Variations display an expected behaviour and it is not planned for standard deviation to be used in the computation of wind power.

Currently, the wind at 100 m is an intermediate by-product of the wind capacity factor, and power, computation, and results are not saved to NetCDF file. This aspect could be revisited, although it is more efficient to provide the wind at 10 m and, separately, the field of α values used to produce the wind at 100 m.

Figure 53:Values of the parameter α averaged over the period 2000-2017

Appendix

Appendix 1: Supplementary material

Figure 54: ENTSO-E load data from the Power Statistics database

Appendix 2: Maritime Regions EU Description Data

Motivation & objective

With the increasing share of the installed Renewable Sources (RES) capacity, evaluating the effect of renewable energy on the energy supply is a very important issue which is addressed in several climate services projects such as C3S-Energy, Clim2power or C3S-ECEM. Prospective analysis are generally made at regional level (e.g. TIMES model) and it is becoming a standard practice to aggregate gridded RES power generation data into aggregated values at NUTS1 or NUTS2 level. This approach raises an issue for the consideration of the offshore wind energy as well as other Marine Renewable Energies (MRE) since there is to the best of our knowledge no commonly accepted region definition corresponding to the NUTS boundaries for the maritime area.

Expected characteristics of the maritime regions

In order to identify the expectation of the energy community with respect to maritime regions, stakeholders of the marine energy community have been contacted and their feedback collected. The main outcomes of this process are summarized below and have been used as a guidance to the choice of the maritime regions.

The policy being different, and the interconnection limited among countries, it is important that the regions be differentiated between countries.

The question arose of limiting some region according to the water depth or the distance to the coast. It turned out that with the development of floating technologies the constraint on the water depth will be less and less relevant.

The distance to the coast is a relevant issue considering its impact on the interconnection. Though, defining a predefined limit such as 12 or 24 nm has been judged difficult and it was decided as well not to consider this aspect.

The regions should allow differentiating the different meteorological regions within Europe by having a level of accuracy where climatic conditions are homogeneous within an area.

Finally, it is not the goal of this initiative to define new regions and existing/recognized regions should be used.

Description of the chosen regions:

We follow the same approach as that used in the NUTS and propose two levels: a national level and a regional level. Several possible approaches for the definition of a third sub- regional level are also discussed.
Level 0: National maritime area

We used the economic exclusion zones as a basis for the definition of the national maritime areas. The input shapefile World_EEZ_v10_20180221.shp used has been downloaded http://www.marineregions.org/downloads.php (Retrieval in December 2018). As illustrated in Figure 1, only European countries included in the bounding box (26°N, 72°N, 22°W, 45°E) have been selected. Island regions that are not expected to contribute the energy supply of continental Europe have been excluded: Azores, Canary Island, Faeroe, Madeira, Svalbard and Jan Mayen. Finally the maritime regions of UK, Guernsey and Jersey have been merged.

The resulting regions are illustrated by red areas in Figure 1, where regions colored in green represent EEZ regions close to Europe that have not been selected in our data.

Figure 55 : Map with the selected level-0 maritime region displayed in red. EEZ areas close to Europa that have not been retained are represented by green areas.

The list of regions contained in the level-0 European maritime area is summarized in the table below:

Table 6 : list of countries included in the level-0 European maritime area

The shape file of the level-0 European maritime area "EuropeanContinentalMaritimeAreas_Level0_v1.1.shp" can be downloaded with the following link: https://cloud.mines-paristech.fr/index.php/s/EAQo2mv8sVQ2VSz

Level 1: Regional maritime areas

The level-0 regions address the need of having country-wide regions. Nevertheless, this regionalization has shortcomings for several countries. Indeed, the French maritime area includes the Channel, North Sea, Atlantic and Mediterranean Seas. This is also the case of Germany where the maritime area encompasses the North and Baltic Seas. These examples show that a lower level with smaller regions is needed to have more homogeneous conditions within a region.

We decided to use the IHO⁷ definition of oceans and seas to subdivide the EEZ regions. A shapefile containing the intersection of the IHO seas and EEZ could be found on the website www.marineregions.org ⁸ . A description of this calculation is available on http://www.marineregions.org/sources.php#ihoeez. A subset of the global dataset was selected in order to ensure coherence with level 0 dataset. In addition, minor modifications were realized on the dataset which are described below.

Figure 56 : Maritime regions of Spain and Italy that have been reworked

In the left map of Figure 56, the Spanish part of the North Atlantic Ocean is represented. It can be observed that this region is constituted of two non-adjacent sub-regions that can be expected to have different weather conditions.

We therefore split this region in a northern and southern zone. The attribute table of the shapefile has been modified by exchanging the area "fid 494" (Spanish part of the North Atlantic Ocean) by the two following areas:

• fid 910494 (Spanish Southern part of the North Atlantic Ocean)
• fid 920494 (Spanish Northern part of the North Atlantic Ocean)

The same issue can be observed for the Italian part of the Mediterranean Sea - Western Basin in the right map of Figure 56. To address this issue the northern part of this region has been merged with the Italian part of the Ligurian Sea and the southern part to the Italian part of the Tyrrhenian Sea.

Figure 57 : Maritime regions of the UK and Denmark that have been refined

In Figure 57, it can be observed that the United Kingdom part of the North Atlantic Ocean is composed of two parts: one at the north of the UK and one small subarea at the South West of the UK. As illustrated in the map, the Southwestern part has been merged to the United Kingdom part of the Celtic Sea. The same situation can be observed on the right map of Figure 57, for the Danish part of the Baltic Sea (fid=95). This Danish region has been split in the two following sub-regions:

• 910095 – Danish Western part of the Baltic Sea
• 920095 – Danish Eastern part of the Baltic Sea

Finally, after the above-described modifications, the intersection of the Continental European EEZ areas and the IHO ocean and sea boundaries gives a total of 80 regions which are displayed in Figure 58. The name and characteristics of these regions as well as the map of these regions are listed the Annex of this document.

Figure 58: Level-1 regions calculated on the basis of the EEZ areas and the IHO sea and ocean definitions.

The shape file of the level-1 European maritime area "EuropeanContinentalMaritimeAreas_Level1_v1.2.shp" can be downloaded with the following link: https://cloud.mines-paristech.fr/index.php/s/EAQo2mv8sVQ2VSz

Level 2: Sub-regional maritime areas

In the present work, a level 2 were not calculated but some possibilities for a future work have been considered. Firstly, it can be noted that there are many countries where the level-1 regions are already very small so that a smaller regionalization is not needed. Though, in Norway, Ireland, UK, France, Spain and Italy smaller regions may be needed for some applications.

A first approach may for example consist is splitting the level-1 region in an inshore and offshore area according to the distance from the coast (see Figure 59). Indeed, this would reflect an important factor related to connection cost or maintenance constraints for offshore projects.

Figure 59 : OSPAR maritime regions (http://gis.ices.dk/sf/index.html?widget=StatRec)

The approach above will only split regions close to the shore in two parts. A possibility to further increase the level of detail may consists in integrating the regional fishery regions of the Northern Atlantic and Mediterranean, the ICES and GFCM fishery regions respectively. As illustrated in Figure 60, this would allow reducing the size of e.g. the Spanish part of Mediterranean Sea as well as the Portuguese and Spanish part of the Atlantic Ocean. Though, the impact on the very large Norwegian region is limited. In addition, it has to be first validated that these regions are relevant for the targeted applications.

Figure 60 : ICES and GFCM region definition for the north Atlantic (grey polygons) and Mediterranean Basin (green polygons) respectively

In case this subdivision is not sufficient, the ICES⁹ statistical rectangles can be used for any location above 36°N¹⁰. These rectangles which can be found on http://gis.ices.dk/shapefiles/ICES_rectangles.zip are represented in the left map of Figure 61.

Figure 61: Left map: ICES statistical rectangles. Right map: GFCM statistical rectangles

Unfortunately, the ICES statistical rectangles don't cover the entire Mediterranean Basin. To cope with this limited domain, the GFCM statistical rectangles could be used instead for the Mediterranean region. These rectangles that can be downloaded from http://www.fao.org/fileadmin/user_upload/faoweb/GFCM/Maps/GFCM_Statistical_grid.zip are represented in the right map of Figure 61. It can be observed that these rectangles are small than that of the ICES dataset. In addition, a region of the north Atlantic is not covered by the high resolution rectangles. An additional dataset or a manual extension of the available data is therefore needed here.

There is thus further work needed to choose the optimal approach for the level 2 regions. In the design, it may be interesting to into consideration or at least indicate the following information for each region:

• Type of utilization (zone natura 2000, fishery zone, sea road…)
• Type of geology
• Distance to coast and sea depth
• …

Annex 1: List of level-1 maritime regions

Annex 2: Map of the maritime regions for each country

References

Bartok, B., I. Tobin, R. Vautard, M. Vrac, X. Jin, G. Levavasseur, S. Denvil, L. Dubus, S. Parey, P.-A. Michelangeli, A. Troccoli, Y.-M. Saint-Drenan, 2018 : A climate projection dataset tailored for the European energy sector. Climate Services, 16, 2019, 100138, https://doi.org/10.1016/j.cliser.2019.100138.

Breiman, L. (1996) 'Out-of-Bag Estimation', Technical Report, pp. 1–13. doi: 10.1016/j.patcog.2009.05.010.

Breiman, L. E. O. (2001) 'Random Forests', Machine Learning, 45, pp. 5–32. doi: 10.1023/A:1010933404324.

De Felice, M. et al. (2018) 'The impact of the North Atlantic Oscillation on European hydro- power generation'.

Ho L.T., Dubus L., De Felice M. and Troccoli A. (2020) Reconstruction of Multidecadal Country-Aggregated Hydro Power Generation in Europe based on a Random Forest Model. Energies, 13, 1786. https://doi.org/10.3390/en13071786

Michelangeli, P. A., Vrac, M., & Loukos, H. (2009). Probabilistic downscaling approaches: Application to wind cumulative distribution functions. Geophysical Research Letters, 36(11).

Vrac, M., Noël, T., & Vautard, R. (2016). Bias correction of precipitation through Singularity Stochastic Removal: Because occurrences matter. Journal of Geophysical Research: Atmospheres, 121(10), 5237-5258.

Related articles

• Page:
  NEMO-ERSEM and POLCOMS-ERSEM: User Guides for Products (v1.1)
• Page:
  Sea level and ocean wave time series and indicators
• Page:
  Product user guide for sea level and ocean wave products - time series and indicators
• Page:
  ECDE indicators catalogue entry: Product User Guide
• Page:
  Climate variables for cities in Europe from 2008 to 2017 documentation