Statistical calibration of probabilistic 100 m wind forecasts . 45

The statistical calibration method combined with a new type of wind observations at 100 m’s height were used to improve the probabilistic wind forecast provided by ECMWF global ensemble prediction system, IFS-ENS. The Doppler lidar wind measurements from 4 stations and Doppler radar wind measurements from 10 stations were utilized in this study. The radar measurements are operationally/ daily available, while lidar measurements are partly manually collected and the data availability may be limited to some measurement campaigns, for instance.

Therefore, both observation types were tested separately to ensure the suitability of the measurements on model calibration.

Firstly, wind distributions from both observation types were examined. The shape of the wind distributions are similar, having more observations in categories between 1-10 m s⁻¹, and less observations in categories for over 15 m s⁻¹ wind speeds (see Figs. 6 and 9 inPaper III). Secondly, the shape of the rank histogram for the raw ensemble was investigated by using both observation types, showing that the ensemble forecast in general is underdispersive (see Figs. 4 and 7 inPaper III). However, compared to lidar observations the rank histogram indicate that the forecast is positively biased, while compared to radar observations the forecast is negatively biased. This difference between the biases may be related to observa-tion volume. When compared to the model gridbox, the lidar measurement can be considered as point measurement (Tuononen et al. 2017), while the radar mea-surement volume is closer to gridbox size (Lindskog et al. 2004). Finally, after the calibration of the ensemble wind forecasts it can be clearly seen that the calibra-tion with both observacalibra-tion types is able to flatten the rank histogram, especially with short lead times, indicating reduced bias in the forecast.

Figure 10 shows the Spread-Skill scores for the ensemble wind forecast com-pared against lidar measurements. The raw ensemble is presented with black and calibrated forecast (with lidar observations) with yellow. The statistical calibra-tion is able to: (1) widen the spread, (2) reduce the RMSE, and (3) improve the predictability. These positive effects reduce the undesired underdispersive na-ture of the forecast and can be identified especially during thefirst 144 h of the forecast, which is an important forecast time scale from a wind energy produc-tion perspective. The Brier Score (BS) and Brier Skill Score (BSS) also indicate a similar improvement for shorter lead times. These two scores were studied with

46

three different thresholds (5 m s⁻¹, 10 m s⁻¹, 15 m s⁻¹). The results indicate that calibration has the largest positive impact on low and moderate wind speeds (Fig.

11). For the high wind speeds (over 15 m s⁻¹) the calibration had very low or neu-tral impact. Thesefindings imply that the calibration could improve the forecast also for higher wind speeds, if more observations were available. Similar results were obtained also when compared against radar winds (not shown here). The calibration is able to improve the ensemble wind forecasts of IFS-ENS. Moreover, the calibration did not weaken the forecast in any of the cases.

Figure 10Verification result for probabilistic 100 m wind speed forecasts by using lidar wind measurements. Spread (dashed; lower curve from the color pair) and RMSE (solid) presented for raw ensemble forecast (black) and calibrated forecast (yellow). The verification period is 1-31 Mar 2016. Numbers are explained in the text. Figure fromPaper III©2020 Monthly Weather Review.

(a) BS, Threshold 5m/s (b) BSS, Threshold 5m/s

(c) BS, Threshold 10m/s (d) BSS, Threshold 10m/s

(e) BS, Threshold 15m/s (f) BSS, Threshold 15m/s

Figure 11(left) BS and (right) BBS of 100 m wind speed for the threshold values 5 m s⁻¹, 10 m s⁻¹ and 15 m s⁻¹. The black lines represent the the raw IFS-ENS data and yellow lines the ensemble calibrated with lidar measurements. The veri-fication period is 1-31 Mar 2016. Figure fromPaper III©2020 Monthly Weather Review.

48

4.2.2 VERIFICATION OF ICING PROFILE FORECASTS BY USING CEILOMETER OBSERVATIONS.

A new type of ceilometer based vertical profiles of atmospheric icing were used in Paper IV. The aim was to test how these new observations could be further utilized in icing model verification. The ceilometer-based observations provide wider geographical and denser vertical resolution compared to more traditional mast measurements. This enables the detailed investigation of the vertical distri-bution of atmospheric icing layers for the first time. This study focused on the lowest 2 km of the atmosphere due to the limits set by the ceilometer’s laser beam attenuation within the clouds.

Figure 12Profiles of (a) total number of cases detected by ceilometer, (b) the num-ber of icing events detected by ceilometer, and (c) the numnum-ber of predicted icing events by the model for 6 locations in Finland. All forecast lead times (+1h...+24h) are included in (c). Figure fromPaper IV©American Meteorological Society.

Atfirst the simulated and observed icing distributions were compared. Figure 12 shows(a)total number of cases observed by ceilometer,(b)number of icing events observed by ceilometer, and (c) number of forecasted icing events from 6 different locations. For three inland locations (Savilahti, Hyytiälä, Sodankylä) the total number of cases decreases rapidly from near surface upwards, stabiliz-ing at height∼500 m or above. For the rest of the stations the total number of cases is almost constant up to 500 m (Utö) or even as high as 1000 m (Kumpula and Vehmasmäki) after which it starts to decrease rapidly. For all the stations the decrease in number of cases is due to an attenuation of the laser beam within

the clouds leading to a lack of observations higher up in the studied layers. The ceilometer can’t see above thefirst continuous cloud layer. The same feature can be identified in the observed icing cases (Fig. 12b), in which the number of icing cases starts to decrease rapidly higher than 200-400 m.

Figure 12b shows that less icing cases are identified in the lowest 500 m at coastal stations (Kumpula and Utö) compared to inland stations. At coastal sta-tions the share of observed icing cases out of all observasta-tions is 12%-30% depend-ing on height. At inland stations the share of observed icdepend-ing cases is higher rangdepend-ing from 35% (Sodankylä) to 45% (Vehmasmäki). Furthermore, the highest number of observed icing cases at coastal areas are measured at∼400 m’s height. For the inland stations the peak is obtained lower, between 200-300 m. The reason for lower detection of icing cases in Sodankylä is probably a result of colder and drier climate due to the northern location.

Figure 12c shows the number of forecasted icing cases. For the inland stations the maximum number of forecasted icing cases is close to the observed one. For the coastal stations the height of the peak is forecasted much higher than observed, at a height of 550-750 m. The number of forecasted icing cases in coastal areas is generally lower than observed below 500 m. The model forecasts almost 3 times more icing cases than observed above 500 m. This feature can also indicate limitations of the observation method; however, overforecasting can not be fully excluded.

The performance diagram (Roebber 2009) of icing forecasts is presented in figure 13. This diagram shows skill scores (Wilks 2006) related to icing model per-formance, against new ceilometer-based icing observations. The perfect forecast would land on the top right corner of the diagram, having Probability of Detec-tion (POD)=1, Success Ratio (SR)=1, Threat Score (TS)=1 and bias=1. In this study the best scores for icing forecasts were obtained at inland stations at heights 100-400 m, having bias close to unity and the highest values of POD, TS and SC, indicating that the model has good performance in predicting icing. However, for the coastal areas underforecasting of icing events leads to weakening of bias closer to the surface. In general, the POD is better for inland stations compared to coastal stations. For inland the POD is varying between 0.6-0.75, whereas for coastal ar-eas the POD is as low as 0.18-0.45 in the lowest 400 m. However, the Performance diagram does not show the overall Proportion Correct (PC), which takes into ac-count correctly forecasted cases where icing is neither forecasted nor observed.

The PC is not varying much with respect to height, having values between 0.7-0.8 for all the stations and levels up to 2 km (see Fig. 3d inPaper IV). This indicates that the icing model has a good skill for separating icing and non-icing events.

50 CHAPTER 4. SUMMARY OF THE RESULTS

0 0.2 0.4 0.6 0.8 1

0 0.2 0.4 0.6 0.8 1

10.0 5.0 3.0 2.0 1.5 1.2

Probability Of Detection

Success Ratio (1-FARatio)

bias Threat score x < 100m 100m < x < 400m 400m < x < 1000m 1000m < x < 2000m All stations Coastal Inland Lapland

Figure 13Performance diagram highlighting four different skill scores: POD (ver-tical axis), Success Ratio (horizontal axis), frequency bias (straight isolines) and threat score (curved isolines). Colours indicate different regions. Symbols indi-cate different heights: below 100 m (cross), 100 - 400 m (circle), 400 - 1000 m (square) and 1000 - 2000 m (triangle). The perfect score is at top right corner of the diagram. All forecast lead times (+1h...+24h) are included. Figure from Paper IV©American Meteorological Society.

5 CONCLUSIONS AND FUTURE PERSPECTIVES

The focus of this thesis was to provide a new meteorological solution to support the wind energyfield in Finland. Thefirst step towards this goal was taken when the Finnish government ordered new Wind and Icing Atlases for Finland, to exploit wind energy resources in detail. These resource maps are needed for commu-nity and wind park planning. Different types of solutions are needed to support daily wind energy production. Therefore, in the second part of this thesis a more forecasting orientated approach was taken to answer the questions: How could a new type of observations improve(i)probabilistic wind forecasts together with statistical calibration methods or (ii)icing model verification and lead to better icing model validation. The following sections summarise the conclusions to the research questions of this thesis.

WHICH AREAS IN FINLAND ARE THE MOST PROMISING IN ORDER TO PRODUCE WIND ENERGY?

Papers I andII presented the new ways to generate the Wind and Icing Atlas.

Traditionally these types of atlases have been based on wind and icing observa-tions from high masts and interpolating the resource estimates from nearest ob-servation points. However, the fast growing wind energy industry requires more detailed information to support the planning. Therefore, the approach to utilize finer resolution mesoscale weather models was taken. Instead of selecting one year for simulations, altogether 72 months were simulated to create a wind and icing climatology for Finland. This is the largest corresponding simulation per-formed to produce a wind atlas so far. The subset of months was chosen based on mean wind to represent the past 30 years, as simulation of the entire period was not possible due to computational limitations. The simulations were conducted by nesting ERA-Interim re-analysis to synoptic scale model HIRLAM andfinally to meso scale model Harmonie-AROME. This same subset of representative months was also utilized to produce the new Icing Atlas for Finland after studying that the same period was representative with respect to temperature and relative humidity.

The main results of the Wind Atlas including windfields and power produc-tion estimates for three heights (50 m, 100 m, 200 m) are presented as freely available dynamic maps. In addition, most potential areas for wind energy were downscaled with the WAsP model, enabling even more detailed investigation of wind conditions. The windiest and hence most productive areas for wind energy are in coastal areas and on high fjelds and hills. Unfortunately, these same areas are also at high risk to experience icing in wintertime. Active icing is more fre-quently experienced in coastal areas and high hills. Yet, the risk for passive icing is high also in inland areas. Thus, the highest power production loss is expected

52

in the same areas where the monthly hours of passive icing is high.

The diversity of end users for this type of atlases is wide. The atlases support political decision making and wind park planning, by highlighting the most signif-icant areas with respect to wind energy production. Hence, they bring additional information for urban and land use planning. Icing itself raises a different type of questions, whether investments in anti- or de-icing systems is needed. This is a difficultfinancial question when balancing between investment costs and possible production loss. However, during the planning phase a proper observational cam-paign on the selected wind park site is always needed to ensure the best possible operation in future.

HOW MUCH ENSEMBLE WIND FORECASTS COULD BE FURTHER IMPROVED BY US-ING STATISTICAL CALIBRATION METHODS COMBINED WITH NEW TYPE OF GROUND-BASED REMOTE WIND OBSERVATIONS?

Ensemble predictions have been proven to bring additional value for the end users, in the form of estimations of uncertainty related to the forecast itself. However, many times the ensemble forecasts are either over- or underdispersive and do not provide realistic uncertainty estimates. To improve the ensemble forecast skill the statistical calibration method has been developed. InPaper III, the statistical calibration method based on Box-Cox t distribution and GAMLSS package in R, was utilized to improve the probabilistic 100 m wind speed forecasts produced by ECMWF. Due to lack of wind observations by traditional mast measurements at 100 m’s height, the new type of ground-based remote observations from Doppler lidars and Doppler weather radars were used. Similar tests including verification and calibration were conducted for both observation types separately, to test the suitability of the new observations for future operational activities. Calibration coefficients were calculated separately for each month and lead time. These coef-ficients were used for the next month, to calibrate the ensemble forecasts.

The raw ensemble predictions for 100 m wind speed was found to be gener-ally underdispersive, when comparing the forecast against both new observation types. However, when comparing against lidar measurements the ensemble was overpredicting, whereas comparison with radar indicated a negative bias. How-ever, ensemble calibration with both observation types are able to enhance the ensemble forecasts by enlarging the ensemble spread, i.e. reducing the underdis-persivity, and by decreasing the ensemble RMSE.

Based on the results of this study, the following recommendations for the op-erational usage of this method are given: (i) weekly update of the training co-efficients, would tackle the problems that may be related to seasonal change or changes in model version,(ii)longer training period from 1 to 3 months, would

add the probability of high wind speed cases within the training period, and(iii) more frequent forecast cycling would reduce the impact of the diurnal cycle on coefficients for each lead time. However, more frequent updates of the coeffi-cients and longer training periods are computationally costly and require larger data storage.

Since 2017 FMI has been running the Meterological Cooperation on Oper-ational Numeric Weather Prediction System (MEPS) together with Sweden and Norway, Estonia joining the collaboration in 2019 (MetCoOp, Müller et al. 2017).

MEPS is a meso-scale ensemble prediction system based on the Harmonie-AROME model, with a 2.5 km horizontal grid size. The ensemble in constructed in contin-uous manner by updating 5 new members hourly. This gives more flexibility for ensemble users to choose an ensemble size with a proper time lagging period. The baseline configuration is to use a 6 h time lagging window, which lead in total to 30 ensemble members. Consequently, it would be highly interesting to test the cali-bration of the Harmonie-AROME model forecasts as well, and to get the calicali-bration system operationally up and running. However, there still are some open questions related to operational usage. The main problem is the huge data amounts related to storing all 30 members, multiple times per day for multiple months. The next difficulty is related to ground-based remote wind observations. At the moment the lidar measurements have two disadvantages with respect to operational usage:

the small number of observation points and data not being collected daily. On the other hand, the radar measurements are operationally collected from all 10 sites, but this measurement technique suffers from a larger representative error between the model and observation, yet leading to improvements. The method developed in this thesis opens new opportunities to further add to the value of probabilistic wind forecasts for users in the wind energy sector.

CAN THE NEW TYPE OF PROFILING CEILOMETER OBSERVATIONS BRING ALONG NEEDED HELP TO ICING MODEL VALIDATION?

The usability of the new type of icing observations based on the profiling ceilometer is investigated in the icing forecast verification inPaper IV. The ceilometer pro-files were available from six locations representing different climatic conditions in Finland during one icing season.

The results showed that icing is predicted more accurately at inland stations than at coastal stations. The vertical structure of icing is better captured at inland stations especially below 500 m’s height. For the coastal stations too few cases are predicted at the lower atmosphere, and overall, the model tends to predict icing higher up than observed. In total, the model has a tendency to overpredict icing compared to observed cases. However, above 500 m’s height part of this apparent

54

overforecasting is due to lack of observations, which originates from the ceilome-ter’s inability to see through clouds. Nevertheless, clear sky situations are well predicted. Overall, these new ceilometer-based icing observations were proven useful in icing model validation and the results support their wider use to explore and correct errors in NWP and icing model output.

Today the icing model is part of FMI’s daily operations, generating icing fore-casts and warnings for wind energy producers. However, icing model forefore-casts are sensitive to NWP model output. Uncertainty related to predicted wind speed, temperature and liquid water content will have an effect on the icing forecast.

The icing forecast fails if the clouds are wrongly predicted temporally or spatially.

To tackle this uncertainty the probabilistic weather forecasts can bring additional value (Mylne 2002). One of the future plans is to start running probabilistic icing forecasts, by using MEPS. The ensemble forecast would better take into account the uncertainty associated with the location of the cloud and thus also the uncer-tainty of icing. In addition, since this study took place the FMI’s ceilometer net-work has grown to 24 stations. To conclude, the profiling ceilometer observation network will provide great benefit in future in the field of probabilistic weather forecast verification and validation due to the high resolution impact of these ob-servations.

REFERENCES

Aguado, E. and J. E. Burt (2010).Understanding weather and climate. Upper Saddle River, N.J., Pearson Prentice Hall.

Badger, J., I. Bauwens, P. Casso, N. Davis, A. Hahmann, S. B. K. Hansen, B. O.

Hansen, D. Heathfield, O. J. Knight, O. Lacave, G. Lizcano, A. B. i Mas, N. G. Mortensen, B. T. Olsen, M. Onninen, A. P. V. Loon, and P. Volker, eds.

(2019). Global Wind Atlas 3.0. (Last accessed: 18.8.2020). URL:

https : / / globalwindatlas.info/

.

Bénard, P., J. Vivoda, J. Mašek, P. Smolíková, K. Yessad, C. Smith, R. Brožková, and J.-F. Geleyn (2010). “Dynamical kernel of the Aladin–NH spectral limited-area model: Revised formulation and sensitivity experiments”.Quarterly Journal of the Royal Meteorological Society136:646, pp. 155–169. DOI:

10.1002/qj.

522

.

Bengtsson, L., U. Andrae, T. Aspelien, Y. Batrak, J. Calvo, W. de Rooy, E. Gleeson, B.

Hansen-Sass, M. Homleid, M. Hortal, K.-I. Ivarsson, G. Lenderink, S. Niemelä, K. P. Nielsen, J. Onvlee, L. Rontu, P. Samuelsson, D. S. Muñoz, A. Subias, S.

Tijm, V. Toll, X. Yang, and M. Ø. Køltzow (2017). “The HARMONIE–AROME Model Configuration in the ALADIN–HIRLAM NWP System”.Monthly Weather Review145:5, pp. 1919–1935.DOI:

10.1175/MWR-D-16-0417.1

.

Bergström, H. and S. Söderberg (2008). “Wind mapping of Sweden”.Summary of results and methods used, Elforsk rapport9:04.

Bernabò, P., F. Cuccoli, and L. Baldini (2015). “Icing hazard for civil aviation”. In:

2015 IEEE Metrology for Aerospace (MetroAeroSpace). IEEE, pp. 295–300.DOI:

10.1109/MetroAeroSpace.2015.7180671

.

Berrisford, P., D. Dee, P. Poli, R. Brugge, M. Fielding, M. Fuentes, P. Kållberg, S.

Kobayashi, S. Uppala, and A. Simmons (Nov. 2011). “The ERA-Interim archive Version 2.0”. 1. (Last accessed: 21.10.2020), p. 23. URL:

https : / / www . ecmwf.int/node/8174

.

Bouteloup, Y., Y. Seity, and E. Bazile (2011). “Description of the sedimentation scheme used operationally in all Météo-France NWP models”. Tellus A 63:2,

Bouteloup, Y., Y. Seity, and E. Bazile (2011). “Description of the sedimentation scheme used operationally in all Météo-France NWP models”. Tellus A 63:2,

In document Meteorological Solutions to Support Wind Energy Production in Finland. (sivua 46-63)