As wind energy increases its share of total electricity generation and its integration into the power system becomes more challenging, accurately representing the spatio-temporal variability in wind data becomes crucial. Wind fluctuations impact power and energy systems, e.g. energy system planning, vulnerability to storm shutdowns, and available voltage stability support. To analyse such fluctuations and their spatio-temporal dependencies, time series of wind speeds at an hourly or higher frequency are needed. We provide a comprehensive evaluation of the global and mesoscale-model-derived wind time series against observations by using a set of metrics that we present as requirements for wind energy integration studies. We also perform a sensitivity analysis to find the best model setup of the Weather Research and Forecasting (WRF) model, focusing on evaluating the wind speed fluctuation metrics. The results show that using higher spatial resolution in the WRF model simulations improves the representation of temporal fluctuations; however, higher-spatial-resolution simulations often lower the correlations of wind time series with measurements. Thus, we recommend finer-spatial-resolution simulations for modelling power ramp or voltage stability studies but ERA5 rather than mesoscale simulations for studies where correlations with measurements are essential. We also show that the nesting strategy is an important consideration, and a smoother transition from the forcing data to the nested domains improves the correlations with measurements. All mesoscale model simulations overestimate the value of the spatial correlations in wind speed as estimated from observations. Still, the spatial correlations and the wind speed distributions are insensitive to the mesoscale model configuration tested in this study. Regarding these two metrics, mesoscale model simulations present more favourable results than ERA5.

Many wind energy applications use meteorological data derived from atmospheric models; for example, in the production of wind resource atlases

Several studies have validated meteorological datasets for modelling weather-dependent wind power generation and its highly fluctuating behaviour. These works use data provided by global atmospheric reanalysis

Validation studies of time series from existent high-spatial-resolution datasets (of the order of a few kilometres) produced by mesoscale NWP models can be found in the literature for wind power integration studies.

Fewer articles present model development focused on time series for wind integration studies.

This work focuses on modelling wind speed time series suitable for power and energy system applications. It adds to the literature by (1) investigating the impact of the interaction between the mesoscale model and its forcing data on the quality of the resultant time series and (2) providing a comprehensive evaluation of the different datasets, with a focus on how well they can represent temporal and spatial correlations over northern Europe. We perform a sensitivity study of the WRF model in multiple configurations, varying the influence of the forcing global reanalysis in the simulations to understand its role and distinguish the model configuration that outperforms various time series aspects. The results are also compared to ERA5 and NEWA mesoscale data. We hypothesize that these modelling aspects, defined by the nesting choice, size, and position of the domains, impact the accuracy of the time series more than the horizontal resolution of the model simulations.

This paper is structured as follows: Sect.

The simulations used in this work were produced by the WRF mesoscale model in two different versions, using the configuration for the model physics and model dynamics as in the New European Wind Atlas (NEWA)

All simulations use nudging to the forcing reanalysis in the outer domain with time tendencies computed from reanalysis data at 6 h, as outlined in

The purpose of the nesting experiments (hereafter named “10 km”, “6 km”, “5 km”, and “3.3 km”) is to vary the influence of the forcing data in the innermost domain by changing the grid arrangement of the simulation or to verify the impact of the grid spacing for similar arrangements. The WRF model domains used in the simulations are presented in Fig.

Location of the domains used in the WRF model simulations for two configurations: (1) single nest (domains 1 and 2, in blue) and (2) two nests (domains 1, 2, and 3, in orange). The red dots show the wind speed measurements locations. Base map created with Natural Earth.

Experiment names and the WRF model configuration. The nesting ratio refers to the grid of the relative parent domain, as shown in Fig.

Time series from NEWA and ERA5 reanalyses were included in the comparisons. All time series (from the simulations and the existent reanalyses) were extracted using a horizontal linear interpolation and a logarithmic vertical interpolation for each measurement location and its respective height. The WRF model simulations and ERA5 and NEWA reanalyses use the two closest levels for the vertical logarithmic interpolation. The NEWA dataset and the WRF model simulations have outputs in several fixed levels ranging from 25 to 250 m. From the ERA5 dataset, two fixed height levels (10 and 100 m) were used, and extrapolation is assumed for sites taller than 100 m. No assumption about the atmospheric stability condition is used, and the same process is applied at every time step throughout the entire year.

Data from 14 met masts over northern Europe (Fig.

Observational datasets. Type: meteorological masts (M), lidar (L); location: coastal (C), land (L), offshore (S), forest (F). Availability [%] refers to the valid data within time coverage during 2009 after the quality control and minimization of flow distortion.

Example of the analysis done for one location

Different qualities in wind speed time series are required for applications in power and energy system studies, as described in Sect.

Figure

This section presents the results for each metric described in Sect.

In Fig.

Correlation of simulated time series to measurements for the various experiments in Table

The ACF results (Fig.

Error in autocorrelation function (ACF) at lag

As in the ACF analysis, the standard deviation (SD) of the first difference (Fig.

Error in the standard deviation of first difference (SD of first diff) between the simulated time series and the measurements for the various experiments in Table

Earth mover's distance (EMD; m s

The analysis of the wind speed distribution (Fig.

Probability density function (PDF) and Earth mover's distance (EMD) between simulations and measurements (OBS) for the various experiments in Table

The metric to assess the spatial correlations (Fig.

Correlation versus distance for each pair of sites (Table

Boxplots of the metrics for all stations as a function of the model experiment:

The boxplots in Fig.

The ranking for correlation to measurement (Fig.

The boxplot of error in autocorrelation function (Fig.

The EMD boxplot (Fig.

The fitted spatial correlations (Fig.

To check the consistency of the results in different periods, we recalculated all the metrics for the winter (January–March) and summer (July–September) months. The results (not shown) keep a similar ranking order to the annual time series for all metrics except for the EMD and spatial correlations. In any seasonal period considered, the EMD values range from approximately 0.3–0.4 m s

Ratio of simulated and observed characteristic length scales

All five WRF model experiments were repeated using the WRF model version 3.8, although only the 10 km_v3 one was included in the plots for comparison with the 10 km (WRF version 4.2.1). The rank is unchanged from that with WRF V4.2.1 (with minor differences) for the correlations to measurements, error in ACF, and the error in SD of the first difference (not shown). However, the ranking order is changed for the EMD and the spatial correlations. Nevertheless, the conclusions for these two metrics do not show a clear impact from grid spacing or model nesting. The direct comparison between the 10 km and 10 km_v3 shows no clear effect from the two WRF model versions, and both experiments present a similar position in the rank for most metrics (Fig.

Lastly, an experiment testing different forcing data (10 km_erai) was included to compare simulations forced by ERA5 versus ERA-Interim reanalyses. For all metrics, the 10 km and 10 km_erai present a close position in the ranking (Fig.

To investigate how to improve the mesoscale modelling of wind time series over northern Europe for power and energy system purposes, we performed a sensitivity study to various WRF model setups, including varying nesting configuration (one or two inner domains), nesting ratio (

We found that the model configuration affects the value of the wind time series correlations with measurements metrics more than the grid spacing. Thus, we recommend ERA5 reanalysis over the mesoscale simulations for studies where the correlations with measurements are essential. However, when producing mesoscale simulations for power and energy system purposes, a smoother resolution jump from outer to inner domains benefits the simulations by keeping them more correlated to the forcing reanalysis. This is especially relevant when the wind speed time series are combined with other series data (e.g. electric load or price time series). Finer-spatial-resolution simulations such as NEWA and the 3.3 km experiment may be best for applications where temporal variability has to be well modelled, such as power ramp analyses or voltage stability studies. For more accurate simulations in terms of wind speed distribution and spatial correlations, NEWA presents more favourable results than ERA5.

The value of the metrics at the considered sites shows more accurate results for offshore and coastal than for inland locations in all metrics, except for the standard deviation of the first difference. Simulated sites in forest landscapes generally have more significant errors, especially when measurements are taken at lower heights (i.e. less than 50 m tall). This could be due to model deficiencies in simulating boundary layer processes near the ground in more complex terrain, in agreement with

The evaluation of correlations to measurements indicates that strengthening the influence of the forcing from the reanalysis data on the mesoscale model simulation can be achieved by using a smooth transition between the computational domains. Thus, a nest transition from 15 to 5 km (domain 1 to domain 2) is more effective than using 30 to 6 km (considering the forcing data resolution close to 30 km) for maintaining the high correlations from the reanalysis in the inner domain. A comparison between simulations using three domains (

The ranking order in the autocorrelation function and standard deviation of the first difference is a function of decreasing spatial grid spacing rather than the nesting arrangement. This is probably a consequence of the higher frequency of occurrence of convective processes in finer grid spacing domains, as it is discussed in

Due to computational cost, many other details related to the model setup have not been tested. For example, we used the same size and position of the innermost domain for all simulations. Therefore, we did not test the sensitivity of the simulated time series to these aspects.

The WRF model is an open-source code and can be obtained from

The following projects and organizations have kindly provided the tall mast data used for the evaluations: Cabauw data were supplied by the Cabauw Experimental Site for Atmospheric Research

The WRF model simulations were initialized using ERA5 (

GL, ANH, and MJK worked on the conceptualization, methodology, and writing (review and editing). GL processed the data, ran the model simulations, and wrote the original draft. ANH and MJK supervised the work.

The contact author has declared that none of the authors has any competing interests.

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

The authors acknowledge support from the PSfuture project, Denmark (La Cour Fellowship, DTU Wind Energy). The authors would like to thank Oscar Manuel Garcia Santiago for supporting the coding of time series extraction from non-regular grids.

This research has been supported by the Danmarks Tekniske Universitet (La Cour Fellowship).

This paper was edited by Julie Lundquist and reviewed by two anonymous referees.