Logo: to the web site of Uppsala University

A probabilistic machine learning method is applied to icing related production loss forecasts for wind energy in cold climates. The employed method, called quantile regression forests, is based on the random forest regression algorithm. Based on the performed tests on data from four Swedish wind parks available for two winter seasons, it has been shown to produce valuable probabilistic forecasts. Even with the limited amount of training and test data that were used in the study, the estimated forecast uncertainty adds more value to the forecast when compared to a deterministic forecast and a previously published probabilistic forecast method. It is also shown that the output from a physical icing model provides useful information to the machine learning method, as its usage results in an increased forecast skill when compared to only using Numerical Weather Prediction data. A potential additional benefit in machine learning for some stations was also found when using information in the training from other stations that are also affected by icing. This increases the amount of data, which is otherwise a challenge when developing forecasting methods for wind energy in cold climates.

In situ measurements representing the marine atmosphere and air-sea interaction are taken at ships, buoys, stationary moorings and land-based towers, where each observation platform has structural restrictions. Air-sea fluxes are often small, and due to the limitations of the sensors, several corrections are applied. Land-based towers are convenient for long-term observations, but one critical aspect is the representativeness of marine conditions. Hence, a careful analysis of the sites and the data is necessary. Based on the concept of flux footprint, we suggest defining flux data from land-based marine micrometeorological sites in categories depending on the type of land influence:

1. CAT1: Marine data representing open sea,

2. CAT2: Disturbed wave field resulting in physical properties different from open sea conditions and heterogeneity of water properties in the footprint region, and

3. CAT3: Mixed land-sea footprint, very heterogeneous conditions and possible active carbon production/consumption.

Characterization of data would be beneficial for combined analyses using several sites in coastal and marginal seas and evaluation/comparison of properties and dynamics. Aerosol fluxes are a useful contribution to characterizing a marine micrometeorological field station; for most conditions, they change sign between land and sea sectors. Measured fluxes from the land-based marine station Ostergarnsholm are used as an example of a land-based marine site to evaluate the categories and to present an example of differences between open sea and coastal conditions. At the Ostergarnsholm site the surface drag is larger for CAT2 and CAT3 than for CAT1 when wind speed is below 10m/s. The heat and humidity fluxes show a distinctive distinguished seasonal cycle; latent heat flux is larger for CAT2 and CAT3 compared to CAT1. The flux of carbon dioxide is large from the coastal and land-sea sectors, showing a large seasonal cycle and significant variability (compared to the open sea sector). Aerosol fluxes are partly dominated by sea spray emissions comparable to those observed at other open sea conditions.

A conically scanning, continuous-wave LIDAR is placed on an island in the central Baltic Sea with large open-water fetch, providing wind and turbulence profiles up to 300 m height. LIDAR and Weather Research and Forecasting (WRF) profiles from one year are used to characterize the marine boundary layer, at the same time performing an evaluation of the WRF model against LIDAR measurements with a focus on low-level jet representation. A good agreement is found between the average wind speed profile in WRF and LIDAR, with the largest bias occurring during stable conditions. The LLJ frequency is highest in May with frequency of occurrence ranging between 18% and 27% depending on the method of detection. Most of the LLJs occur during nighttime, indicating that most of them do not have local origin. For cases with simultaneous LLJs in both data sets the WRF agrees well with the LIDAR. In many cases, however, the LLJ is misplaced in time or space in the WRF simulations compared to the LIDAR. This shows that models still must be improved to capture mesoscale effects in the coastal zone.

The wind and turbulence fields over a small, high‐latitude sea are investigated. These fields are highly influenced by the proximity to the coast, which is never more than 200 km away. Simulations with the WRF model over the Baltic Sea are compared with a simplified, stationary wind model driven by the synoptic forcing. The difference between the models is therefore representative of the mesoscale influence. The results show that the largest wind‐field modifications compared with a neutral atmosphere occur during spring and summer, with a mean monthly increase of up to approximately 1 ms−1 at typical hub heights and upper rotor area (120‐170 m height) in the WRF model. The main reason for this is large‐scale low‐level jets caused by the land‐sea temperature differences, likely increasing in strength due to inertial oscillations. These kind of events can be persistent for approximately 12 hours and cover almost the entire basin, causing wind speed and wind shear to increase considerably. The strongest effect is around 2000 to 2300 local time. Sea breezes and coastal low‐level jets are of less importance, but while sea breezes are mostly detected near the coastline, other types of coastal jets can extend large distances off the coast. During autumn and winter, there are fewer low‐level jet occurrences, but the wind profile cannot be explained by the classical theory of the one‐dimensional model. This indicates that the coastal environment is complex and may be affected by advection from land surfaces to a large degree even when unstable conditions dominate.

Global oceans are an important sink of atmospheric carbon dioxide (CO2). Therefore, understanding the air-sea flux of CO2 is a vital part in describing the global carbon balance. Eddy covariance (EC) measurements are often used to study CO2 fluxes from both land and ocean. CO2 are usually measured with infrared absorption sensors, which at the same time measure water vapor. Studies have shown that presence of water vapor fluctuations in the sampling air potentially result in erroneous CO2 flux measurements due to cross-sensitivity of the sensor. Here we compare measured CO2 fluxes from both enclosed path Li-Cor 7200 sensors and open-path Li-Cor 7500 instruments from an inland measurement site and a marine site. We also introduce new quality control criteria based upon a Relative Signal Strength Indicator (RSSI). The sampling gas in one of the Li-Cor 7200 instruments was dried by means of a multi-tube diffusion dryer so that the water vapor fluxes were close to zero. With this setup we investigated the effect that cross-sensitivity of the CO2 signal to water vapor can have on the CO2 fluxes. The dryer had no significant effect on the CO2 fluxes. We tested the hypothesis that the cross-sensitivity effect is caused by hygroscopic particles such as sea salt by spraying a saline solution on the windows of the Li-Cor 7200 instruments during the inland field test. Our results confirm earlier findings that sea salt contamination can affect CO2 fluxes significantly and confirm earlier findings, that drying the sampling air for the gas analyzer is an effective method to reduce this signal contamination.

The problem of icing on wind turbines in cold climates is addressed using probabilistic forecasting to improve next-day forecasts of icing and related production losses. A case study of probabilistic forecasts was generated for a 2-week period. Uncertainties in initial and boundary conditions are represented with an ensemble forecasting system, while uncertainties in the spatial representation are included with a neighbourhood method. Using probabilistic forecasting instead of one single forecast was shown to improve the forecast skill of the ice-related production loss forecasts and hence the icing forecasts. The spread of the multiple forecasts can be used as an estimate of the forecast uncertainty and of the likelihood for icing and severe production losses. Best results, both in terms of forecast skill and forecasted uncertainty, were achieved using both the ensemble forecast and the neighbourhood method combined. This demonstrates that the application of probabilistic forecasting for wind power in cold climates can be valuable when planning next-day energy production, in the usage of de-icing systems and for site safety.

Sweden has good conditions for wind power. However, most of Sweden (ca. 70%) is covered by forest. Forests decrease wind speeds and create turbulence, something which is not favourable for wind power. Several Swedish wind maps have shown that forests in Nordic countries can be well suited for wind power (e.g. Bergström and Söderberg 2011, Byrkjedal and Åkervik 2009).

At the same time, there is uncertainty over wind conditions over forests at very high altitudes (ca. 150 m above ground). How good do wind resource assessment models agree with measurements? How much energy is a wind turbine in forest going to produce and which loads will a wind turbine in forest experience?

This project has investigated all these issues. Work was concentrated in the following work packages:

1. Wind resource at very high heights
2. Turbulence- and wind measurements at very high heights above forest
3. Analysis of turbulence data from forests
4. Model simulations with wind flow models
5. Model simulations with very-high-resolution weather forecast models
6. Model simulations with Large Eddy Simulation (LES) models
7. Improved specification of so-called “synthetic turbulence” over forest
8. Analysis of airborne laser altimeter measurements over forest
9. Forest’s effects on wind turbine energy production
10. Load simulations for wind turbines over forest

WP1 studies how wind speed and direction varies with height over forest (up to ca 150 m above ground and higher up). Several profile relations are studied here.  Frequency distributions of wind shear and veer are presented. WP2 describes turbulence and wind measurements that have been carried out within the project at Hornamossen. Moreover, the measurement campaign that was carried out in a line over the Hornamossen-hill within the New European Wind Atlas project is described. WP3 analyses turbulence data from Hornamossen together with turbulence data from Ryningsnäs. Of special interest is how turbulence intensity decreases with height as well as if the IEC-standard class A, B or C for wind turbines is complied with at different heights. WP4 describes the newly developed linearised wind flow model ORFEUS with a dedicated forest module. WP5 describes model simulations with WRF and the MIUU model, their sensitivity for surface roughness and turbulence parameterisations. Mean wind profiles from the models are compared to Hornamossen. WP6 describes LES simulations with Chalmers LES model and WRF-LES. LES-resultats depend to a large degree on how the turbulent vortices are initialised at the inflow boundaries of the LES model. Several different methods for that are described. WP7 describes a new turbulence model (the Segalini & Arnqvist model) that includes atmospheric stability. This is a further development of the IEC turbulence model (=Mann model) for neutral stability. Coherence of turbulent winds as well as phase profiles are other improvements of the IEC model. WP8 describes a new method to compute leaf/needle/plant area density from laser scans of the Swedish forest and how one estimates surface roughness and zero plane displacement from that. The new method is compared with two other methods. Results are also compared with official forest data (“skoglig grunddata”). The effect on the wind profile is also shown. WP9 describes the new methods for estimating AEP from the Power Curve Working Group and the IEC standard for Power Performance Testing. Effects on estimated AEP are shown. A new simple model for calculating turbulence effects on energy production is developed and compared with data from a wind farm. Within WP10 a new generic open-source wind turbine is developed and used for load simulations with aero-elastic simulations. Results show that the new coherence model for turbulence gives much smaller loads than the turbulence model of the IEC standard.

For more information on the different parts of the project the reader is referred to the report’s introduction, the ”Summary and Conclusions” of each chapter as well as the overall summary (”Executive Summary”) at the end of the report.

Streaky structures of narrow (8-9 km) high wind belts have been observed from SAR images above the Baltic Sea during stably stratified conditions with offshore winds from the southern parts of Sweden. Case studies using the WRF model and in situ aircraft observations indicate that the streaks originate from boundary layer rolls generated over the convective air above Swedish mainland, also supported by visual satellite images showing the typical signature cloud streets. The simulations indicate that the rolls are advected and maintained at least 30-80 km off the coast, in agreement with the streaks observed by the SAR images. During evening when the convective conditions over land diminish, the streaky structures over the sea are still seen in the horizontal wind field; however, the vertical component is close to zero. Thus advected feature from a land surface can affect the wind field considerably for long times and over large areas in coastal regions. Although boundary layer rolls are a well-studied feature, no previous study has presented results concerning their persistence during situations with advection to a strongly stratified boundary layer. Such conditions are commonly encountered during spring in coastal regions at high latitudes.

In this paper, we document the existence of wave-like motions above a forest canopy using data taken from a 138 m high tower placed within a forest Characteristics of the waves are examined in relation to their possible effects on wind energy. It is shown that when the wave signal is relatively clean, the phase lag between horizontal and vertical velocity is close to 90, which limits the contribution of the waves to the downward momentum flux. Numerical solutions of the linear wave equations agree with measurements in terms of wave period and the vertical shape of the wave amplitude. Linear analysis show that shear instability is the cause of unstable wave growth, and that the fastest growing unstable wave number typically has a period of 10-100 s. In addition to the shear instability, the linear analysis predicts that under certain conditions instabilities of the Holmboe kind can develop over forests.