Contributors: Gert-Jan Marseille (KNMI), Kirien Whan (KNMI), Janet Wijngaard (KNMI)
Issued by: Gert-Jan Marseille and Kirien Whan (KNMI)
Issued Date: 11/05/2021
Ref: C3S_D435_LOT3_KNMI_PUG_footprints_v0.1
Official refence number service contract: 2020/C3S_435_Lot3_KNMI
Acronyms
Acronym | Description |
C3S | Copernicus Climate Change Service |
CDS | Climate Data Store |
MLR | Multiple Linear Regression |
SIS | Sectoral Information System |
WISC | Windstorm Climate Service |
Introduction
Executive Summary
A storm footprint is defined as the maximum 3-second 10-m wind gust speed (in m s-1) over a 72-hour period at each model grid point for a significant winter storm. As such, a storm footprint shows the spatial distribution of maximum wind gust speed for a storm crossing the area of interest. For European storms, the area of interest spans the area defined by 15W to 25E, 35N to 70N though in fact the storms have been tracked over their full course from formation over the Atlantic.
The operational footprints, available from the Climate Data Store (CDS), are based on statistical downscaling of the ERA5 dataset, using the main high resolution ERA5 field in each case. The operational footprints are derived for storm tracks identified by the Storm Tracking module as part of the operational service.
The storm footprints serve as input for the calculation of economic loss caused by the storm by means of so-called Tier 3 indicators, which indicate the effect of windstorms on the infrastructure and financial system of a sector.
Scope of Documentation
This document describes the C3S Storm Footprint dataset using the standard C3S format for product descriptions, i.e., in terms of product target requirements, product overview, input data and method. It is based on statistical downscaling, in contrast to dynamical downscaling, which was used in the 'Proof of Concept' contract called WISC (Windstorm Climate Service). In dynamical downscaling, a high-resolution regional model is applied using reanalysis model fields as boundary conditions. The statistical downscaling approach is outlined in the Method section, section 2.4.
Version History
The operational footprints are based on statistical downscaling of the ERA5 dataset, using the main high resolution ERA5 field in each case. The operational footprints are derived from the new ERA5 storm tracks produced for the operational service. Since these yielded a slightly different set of storms from the earlier C3S Proof of Concept (WISC) tracks, WISC tracks were also used to provide statistically downscaled storms from ERA5 to match those provided by WISC. This will allow for comparisons with the equivalent statistically downscaled WISC footprints.
Product Description
Product Target Requirements
Wind gust (m s-1) is the prime input parameter for footprint generation. It was found that wind gust from reanalysis (ERA-Interim and ERA5) underestimates measured wind gust on average. Dynamical and statistical downscaling are methods to improve on this. For the operational service, statistical downscaling was selected as statistical downscaling proved a suitable alternative as is computationally efficient in comparison.
Product Overview
Data Description
The C3S storm footprint dataset consists of footprints from all identified winter storms, by the Storm Tracking module, over the period 1979-2021. Storm footprint are available in NetCDF format. Figure 1 shows an example of the storm footprint for storm Christian which crossed Europe around 28 November 2013, obtained from the statistical downscaling method using the Multiple Linear Regression (MLR) technique, starting from ERA5 model fields.
Figure 1: Storm footprint for storm Christian (28/10/2013) as obtained from statistical downscaling using the Multiple Linear Regression (MLR) technique.
Table 1. Overview of key characteristics of the storm footprints.
Data Description | |
Dataset title | Storm footprint |
Data type | Reanalysis post-processed |
Topic category | Meteorological geographical feature |
Sector | Wind damage |
Keyword | Storm footprint. |
Dataset language | English |
Domain | 20W to 35E and from 35N to 70N |
Horizontal resolution | Gridded at 1.0km, based on 31km resolution of the ERA5 dataset |
Temporal coverage | 1979-01-01/to/2021-03-31 |
Temporal resolution | Maximum 3s wind gust for the 72-hour period is shown at each grid point based on hourly assessments |
Vertical coverage | Single level |
Update frequency | No updates expected |
Version | 1.0 |
Model | ERA5 |
Experiment | n/a |
Provider | KNMI |
Terms of Use | https://cds.climate.copernicus.eu/api/v2/terms/static/licence-to- use-copernicus-products.pdf |
Variable Description
Table 2. Overview and description of variables.
Variables | |||
Long Name | Short Name | Unit | Description |
Windstorm footprint | - | m s-1 | For an identified storm track, the maximum 3- second 10-m wind gust ( m s-1) over the 72-hour period capturing the storm, crossing Europe. |
Input Data
The input data for the statistical downscaling method includes ERA5 10-meter wind gust, wind gusts estimated from the wind-shear between 100 m and 10 m and station elevation height.
Table 3. Overview of climate model data for input to storm footprint based on statistical downscaling, summarizing the model properties and available scenario simulations.
Input Data | ||||
ERA5 wind gust | ECMWF | Scenario | 1979-2021 | 31 km |
Windgust | (see footnote) | - | 1979-2021 | - |
Station elevation height | - | - | 1979-2021 | 1 km |
Input Data 1
ERA5 10-meter wind gust (m s-1)
Input Data 2
ERA5 wind field(m s-1) at 10 m and 100 m above the earth surface
Input Data 3
Station elevation height (m). Land stations, more in particular their height, have been used to estimate the parameters of the MLR scheme.
Method
Background
Accurate estimates of wind gusts are important for many sectors, including the insurance industry, e.g., for damage calculations and potential changes therein for instance due to climate change. It is difficult to obtain the complete picture from high-resolution observed wind gusts, due to the sparse observational network. Many locations remain unobserved, even in places with a dense observational network, such as many countries in western Europe. The relatively sparse observational network can be particularly problematic for variables that are spatially heterogeneous, such as wind gusts.
Dynamical, statistical, or combined dynamical-statistical (where a dynamical model is downscaled, bias-corrected and combined with other predictive information using statistical methods) approaches can be taken to develop high-resolution gridded estimates of wind gusts. There are positive and negative aspects of each approach. For example, a dynamical (numerical weather prediction meso- scale) model is computationally expensive to run and contains errors due to errors in the initial conditions and the physical parameterisation schemes. On the other hand, a statistical model (e.g. multiple linear regression) is limited by the strength of relationships in the training data set, the number of extreme observations in the training data set and cannot easily handle changes in the relationships over time. Despite some limitations, multiple linear regression (MLR) is one of the most popular ways of making predictions. MLR models the linear relationship between a response variable (e.g., observed wind gusts) and two or more predictor variables. This relationship (i.e. the statistical equation) is developed on a training data set (a set of observations and potential predictor variables) and then applied to new independent data.
Model / Algorithm
The MLR technique estimates wind gusts (Y) at unobserved locations in Europe. Various sets of potential predictors were tested. The preferred statistical model is one that uses three predictors (X1, X2 and X3);
- (1) ERA5 wind gust forecasts to the power of 2,
- (2) wind gusts estimated from the shear between 100 m and 10 m divided by the logarithm of the heights, also taken to the power of 2, and
- (3) station elevation height.
The application of the power of 2 is motivated by putting more weight to extreme wind gusts.
In MLR parameter Y is linked to the predictors through: \( Y = a + b_{1}X_{1} + b_{2}X_{2} + b_{3}X_{3} \) . Parameters a, b1, b2 and b3 are estimated using observed wind gusts and co-located values of the predictors. This was done for a selection storms resulting in the following relationship:
\[ Wind Gust = 10.3 + 0.0112 \ast ERA5^2 + 0.0148 \ast wgSLh^2 + 0.00355 \ast ELEV \]With ERA5 denoting predictor 1 (wind gust from ERA5), wgSLh denoting predictor 2 (wind gust estimated from wind-shear between two height levels) and ELEV denoting predictor 3 (observation station elevation height). Predictor 2 is defined through:
\[ wgSLh = u_{10} + \alpha \frac{u_{100} - u_{10}}{log \left(100/10 \right)} \]with u10 and u100 the hourly mean wind speed at 10m and 100m from ERA5, respectively, and parameter α is the median of the normalized gust with a value of 3.25. The latter follows from standard wind turbulence theory2.
In regions of little topography, the wind gust estimates exhibit the smoothness of ERA5, while high- resolution features are present in mountainous regions.
More details on MLR including verification results are found in 2(Van den Brink and Whan, 2018). Statistical downscaling has the advantage over dynamical downscaling that it is very simple once the weight for the various predictors in the MLR equation has been established. Hence, the computation time is negligible when compared to running a computationally intensive mesoscale model over the storm period (dynamical downscaling).
Validation
The validation of the MLR technique has been performed over land using wind gust data from land stations. A total of 17 storm footprints and corresponding land station data were used for validation. Figure 2 shows the locations of wind gust data over land and the mean monthly maximum wind gust between 2000-2018, the period capturing the 17 storm footprints.
Figure 2: The location of stations used in the regression analysis, and their mean winter monthly maximum wind gust between 2000-2018. Values are arranged before plotting, so that larger values are on top, in densely observed areas.
Figure 3 shows a QQ-plot3 of the observed and predicted monthly maxima winds during the 17 storm footprints. From the figure it is clear that the MLR model over-estimates low wind speeds. However, the MLR model is better able to simulate extreme winds, compared to the raw ERA5 wind gust estimates. ERA5 does not simulate extreme wind gusts well. For example, of the monthly maxima during the 17 storm footprints, wind gusts > 40 m s-1 are observed 47 times. In ERA5 there are only 6 cases that exceed this threshold, while in the MLR predictions there are 45 cases.
3The quantile-quantile (q-q) plot is a graphical technique for determining if two data sets come from populations with a common distribution. A q-q plot is a plot of the quantiles of the first data set against the quantiles of the second data set. By a quantile, we mean the fraction (or percent) of points below the given value. That is, the 0.3 (or 30%) quantile is the point at which 30% percent of the data fall below and 70% fall above that value. A 45-degree reference line is also plotted. If the two sets come from a population with the same distribution, the points should fall approximately along this reference line. The greater the departure from this reference line, the greater the evidence for the conclusion that the two data sets have come from populations with different distributions. (https://www.itl.nist.gov/div898/handbook/eda/section3/qqplot.htm)
Figure 3: QQ plot of observed (see Figure 2) and predicted monthly maxima wind gusts during the 17 storm events from the raw ERA5 forecasts (black) and the MLR model, equation in section 2.4.2 (blue).
Known issues
- During publication an issue was identified with a small subset of the Windstorm footprints NetCDF files. The issue only affects the NetCDF metadata labels FX:long_name "Wind Gust estimates: MLR monthly maximum wind gust". These are incorrectly labelled and should instead state max_wind_gust:long_name = "maximum 10-m wind gust". All files contain the correct variable data and the wind gusts were calculated from a 72-hour maximum as stated in Table 2.
Concluding Remarks
Storm footprints serve as input for economic loss calculations as part of the Windstorm Service. The validation results presented in section 2.4.3 give good confidence that statistical downscaling through the Multiple Linear Regression (MLR) model, described in section 2.4.2, provides good estimates of the strongest wind gusts which are of most relevance for economic loss and risk calculations. The MLR model has proven robust for new (recent) storms entering the Windstorm Service.
The use of storm footprints based on the statistical downscaling approach is limited to land areas. This does not limit the purpose of the Windstorm Service since damage and economic loss from windstorms is mainly over land areas. Hence, high quality storm footprints over land suffices for economic loss estimate calculations.
References
An effective parametrization of gust profiles during severe wind conditions, Henk W van den Brink, Published 28 November 2019, Environmental Research Communications, Volume 2, Number 1.
