The measurement system, which has been used for the measurement campaigns within this paper, has two main, independent components: a powerful carrier system and a sensor unit, which consists of a commercially available ultrasonic anemometer and a self-developed compensation and data acquisition unit.

The foldable, commercial carrier system is a battery-powered octocopter with a flight time of 25 min and a maximum take-off weight of 12.5 kg. Including the sensor unit, the complete system only weighs 8.5 kg and therefore has a considerable performance reserve. Flights in turbulent air as well as during gust speeds of 25 ms-1 have successfully been tested. A real-time-kinematics (RTK) GPS is included to perform high-accuracy positional navigation and speed estimation. The open-source flight controller has been adapted for an easy setup of specific measurement strategies, which are then autonomously being followed. Although a completely unobserved operation is technically possible, European laws at this moment require an operator to be within sight.

The Gill WindMaster 3D ultrasonic anemometer is placed on top of the compensation unit centred above the rotor plane. Mounting the sensor on top of the UAV has several advantages. First of all, the rotational symmetry of the system allows wind measurement independent from the yaw angle and wind direction. Additionally, this setup results in a horizontally centred mass during hovering and therefore leads to relatively small moments to be compensated for by the UAV. This improves flight performance and flight time. Besides, the downwash above the rotors is less turbulent than below.

The distance of the sensor's measurement volume to the rotor plane is 750 mm and is considered a trade-off between manoeuvrability and reasonable interaction between the wind sensor and propeller-induced flows.

Except for the power supply, the self-developed compensation and data acquisition unit is completely independent from the UAV. If requirements concerning the carrier system change, the compensation unit as a whole can be reapplied easily onto a new aircraft. It weighs 420 g and contains all necessary sensors as well as an additional RTK GPS for accurate position and speed estimation by means of sensor fusion. Based on analytical calculations and various synthetic experiments, a compensation algorithm was developed that efficiently reduces measurement errors due to movements of the airborne system as well as its rotors. Additional telemetry transmits measurement data such as wind speeds and directions live to a ground station for in situ analysis. The anemometer data are additionally saved to the internal storage at a rate of 10 Hz.

All following calculations and measurements have been evaluated based on data that have been processed by the compensation unit. The system validation in general was conducted on several levels of detail, beginning with the Guide to the Expression of Uncertainty in Measurement (GUM; Internationale Organisation für Normung, 1995) to evaluate the standard uncertainty of a single point of measurement. The GUM allows the calculation of the standard uncertainty without the necessity of a true reference value. Error estimation is done by creating a mathematical model of the WindLocator, including relevant influences and their uncertainties, and combining them into the system's standard uncertainty, which is ± 0.37 ms-1 in our case.

After several synthetic tests with a fixated UAV to evaluate rotor influences (Fig. 2), the WindLocator's compensation unit was tested during an indoor flight under zero-wind conditions (Fig. 3).

Utilising the internal barometer, an altitude of around 4 m was maintained during our test, and pitch and roll axes for minimum horizontal movements were automatically stabilised. Nevertheless, small sensor inaccuracies made pilot interventions necessary to remain at a sufficient distance from walls. After compensation, the wind data are given out in a global north–east–down coordinate system and are independent from the specific orientation of the UAV.

Figure 4 shows the data in all three measured directions at a resolution of 10 Hz. Peaks, e.g. in the vN direction at second 17 (-1.5 ms-1) and 55 (-1.31 ms-1), are a result of the UAV's horizontal translation due to operator intervention.

As expected, mean wind speeds during the indoor flight are close to zero. SDs of up to 0.23 ms-1 meet our expectations according to the GUM but clearly show the influence of manual operator control and of the sensor being rather close to the turbulent downwash induced by the rotors.

After proving that under zero-wind conditions, mean values are in good agreement with our expectations, a measurement setup was created to compare the performance of the WindLocator with a stationary anemometer. In flat terrain 2 km west of Aachen (North Rhine-Westphalia, Germany), a stationary anemometer of the same type as the UAV's anemometer was mounted at a height of 3 ma.g.l. (above ground level). Data acquisition and storage for the stationary anemometer were realised at 10 Hz by a self-developed data acquisition system, which uses timestamps synced with an online time server. The UAV timestamps are derived from GPS time signals. The UAV was set to hold position at a height of 3 m. A distance of 4 m to the stationary anemometer orthogonal to the main wind direction was chosen to avoid interactions of the two measurement systems (Fig. 5).

Four 10 min measurements with ∼ 6000 data points each have been conducted, with a short break to switch batteries after the second measurement point. Unlike the indoor tests, all three wind components are combined into a resulting wind speed v for every point of measurement to improve comparability to the stationary anemometer. However, the vertical component vD in general has a minor impact on the resulting wind speeds. v=vN2+vE2+vD2

The diagram in Fig. 6 shows exemplarily the compensated for wind speeds of the WindLocator in comparison to the stationary reference as well as the corresponding regression plot.

For all measurement points (see Table 4), a very good agreement of the 10 min mean wind speed between the WindLocator and reference has been achieved, especially when taking into account the turbulent wind situation during such a low-altitude flight. Turbulence intensities (TIs) of up to 44 % have been calculated for the stationary reference. Although there are absolute differences of +1 % (measurement 1) to +6 % (measurements 2 and 4), the WindLocator already provides a good estimation of the prevailing turbulence intensity.

An analysis of wind directions during this experiment was not yet possible, because an accurate orientation of the stationary measurement system could not be guaranteed. This was taken into account for the next experiment at a 134 m met mast under more realistic conditions. The measurement system was tested close to a met mast on a small plateau. Four measurements of 8–10 min were conducted and are compared to the velocity data of a cup anemometer at 134 m and the directional data of a wind vane at 130 ma.g.l. The WindLocator was held at a height of 134 m based on barometer and GPS data and was then moved closer towards the met mast using the onboard camera system. Because the flight was performed without autopilot, distances to the met mast and exact height vary throughout the measurements (see Table 5). Additionally, Table 5 contains wind speed data analogue to Table 4 as well as information concerning the accuracy of wind direction estimations. For all following calculations, the WindLocator data were averaged to 1 Hz for better comparability to the met mast.

The results during the met mast experiment show a slightly different picture compared to the ground-level measurements. The turbulence intensity is still reasonably well estimated. Additionally, measurements 2 and 4 show a good correlation of the WindLocator with the corresponding reference speed. However, mean wind speed deviations for the first and third measurement not only are higher than before but also vary a lot more compared to the other measurements of that day. Significant deviations mainly occur during the first half of the measurements (Fig. 7), e.g. seconds 180 to 270 for measurement 3.

Those deviations are a result of the pilot still doing positional adjustments during the measurement point. Nevertheless, those adjustments seem not to have a critical impact on the UAV's wind direction estimation, which shows very good correlation through all measurement points with a maximum mean deviation between met mast and WindLocator of 2.4∘. As an example, the absolute wind direction and its regression plot for measurement 3 is shown in Fig. 8.

The WindLocator performed very well throughout the tests, especially concerning the calculation of averaged measurement quantities like speed and direction. When the system used its GPS-based hover mode without interference by a pilot, mean wind speed differences compared to a reference were below 0.12 ms-1 and wind direction differences were smaller than 2.4∘. The maximum absolute difference in turbulence intensity was 5.9 % for a high-turbulence-intensity measurement. Although more measurement points are necessary to finally evaluate the system's performance, initial results in comparison to scanning-lidar errors seem promising. It also has to be taken into account that the airborne measurement system and a reference cannot measure at the exact same place at the exact same time. Remaining uncertainties might always be a result of spatial deviations in the wind situation, which will be discussed in more detail in the following section.

The test site for this measurement strategy is a small hill in the south of North Rhine-Westphalia in the German Eifel and was chosen for the following reasons:

With a yearly mean wind speed of 6.5–7 ms-1 at 100 ma.g.l., the area has rather high wind speeds compared to the rest of the county. The main wind direction is southwest.

All diagram coordinates within this chapter are referenced to the UTM coordinate 32U 308450 5604720.

Within this paper, two different measurement campaigns are presented. While the mean wind speeds are within a similar range, wind directions differ, resulting in different inflow conditions into the measuring area.

The presented campaign aims to investigate the feasibility of using a simple measurement strategy for the identification of the spatial distributions of meteorological variables (wind speed, turbulence intensity, inclination) above complex terrain. This information will be used in the further course of the project for the development of the final measurement strategy. The measurement strategy can be described as follows:

The WindLocator automatically flies to one measuring point after another and measures at each position for a specified duration.

The feasibility study within this paper addresses two basic questions on the postprocessing of the gathered measurement data. In the first step it will be discussed whether the temporal change in the wind speed during the measurement campaign has to be taken into account for the further investigation of the spatial distribution of the meteorological variables. A necessary condition for a constant spatial distribution is a constant wind direction, which will be verified at the beginning. Variations in averaged wind speeds at the stationary reference are used to estimate the impact of temporal variations within the airborne measurements in comparison to expected spatial variations. The result of this analysis is also valid for turbulence intensity, as it depends on the wind speed. Additionally, the spatial distribution of turbulence intensities is checked for plausibility. The influence of temporal changes on inclination angles is checked in a qualitative manner by comparing them to the terrain.

Assuming that the temporal change in the wind speed has a significant effect on the measurement, in a second step it will be investigated whether the ground-level (3 m) anemometer can be used as a reference to compensate for temporal changes in the respective variables. Therefore, the ground-level anemometer needs to represent the overall wind situation. This is evaluated using the correlation between ground and airborne measurement data, assuming a linear dependency between those measurements. If correlation is confirmed, wind speed measurements of the WindLocator shall be used to calculate a local speed-up factor in comparison to ground-level wind speed. This distribution is then checked for plausibility.

Figure 11 gives an overview of the measured resulting wind speeds v from the moving WindLocator and the stationary reference on the ground, exemplarily shown for measurement 1 (M1). After the data acquisition was started, the UAV heads to the first measurement point, where it is holding position for 5 min 100 m above the start level, before moving on to the next waypoint at the same height. A measured wind speed deviation between WindLocator and reference is expected because of the differences in height and horizontal position of both systems. During the battery swap after four measurement positions, no WindLocator data are available. The stationary reference in contrast measures non-stop. Measuring 16 points of 5 min each, yielding 80 min of usable measurement data, takes around 2 h in total.

Figure 12 represents all measurement points for both campaigns, showing the results of the single points measured one after another. During both measurements, wind directions are in good agreement with the mean wind direction. With mean absolute deviations of 9.9∘ (M1) and 11∘ (M2), no significant changes in wind direction during the measurement time of 2 h for each measurement campaign are found. This validates our assumption that the distributed wind field will not be influenced by a change in wind direction.

Figure 13 shows the averaged wind speeds for measurement points one after another for WindLocator and ground station. Over all UAV measurement points, an absolute variation in mean wind speeds of between 2 and 6 ms-1 has been detected (Fig. 13). As implied earlier, these variations are considered to be too high for spatial deviations due to complex terrain only, especially when taking into account the measurement height of around 100 m. These fluctuations are also a consequence of wind variation over time.

Otherwise, the stationary reference (assuming it to be an indicator of the overall wind situation) would not have shown any significant differences in wind speed over time. This is clearly not the case, especially when looking at the normalised reference wind speed, calculated by dividing the mean wind speed value of each point by the maximum mean value of all points of that measurement (Fig. 14). Normalised variations in the stationary reference and the WindLocator data are on a comparable order of magnitude (-70 % compared to the maximum wind speed). Because those are comparable and even higher than expected spatial variations (El Bahlouli et al., 2019; Wildmann et al., 2017) by up to 30 %, temporal variations clearly have to be compensated for to achieve successful measurement.

Figure 15 shows turbulence intensities measured at each single point. The mean turbulence intensity over all measurement points of M2 is 18 % and therefore slightly higher than during M1 with 15 %. This seems plausible due to the forested and steep escarpment in the upwind direction for M2. However, single turbulence intensities within the measured field seem to vary rather strongly (between 10 % and 30 %) and without obvious influences by terrain and surface. As the normal turbulence model of IEC 61400 predicts, turbulence intensity depends significantly on mean wind speed. Very low average speeds of only 2–3 ms-1 (see Fig. 13a, measurement points 1 and 2) might be an explanation for unexpected high turbulence intensity in the northeast of M1, for example. Consequently, temporal changes in wind speed have to be taken into account when measuring turbulence intensity distributions.

Although wind speeds vary significantly over time, inclination angles do show plausible results (Fig. 12). The flow and therefore the inclination angles follow the terrain quite well for M1, varying mostly between +5∘ at the windward side of the hill and -5∘ at the lee side and switching their sign at the ridge with a peak of -9∘ in the south close to the escarpment. For M2 on the other side, nearly all angles are above 0∘, especially in the north with several measured inclination angles higher than 8∘ and even up to 12.7∘. Positive inclination angles are considered to be a plausible result of winds passing the steep escarpment in the southwest. Temporal variations in wind speed seem to have a minor impact on inclination angle measurements.

All in all, temporal wind speed variations do have a significant impact while measuring wind speed distribution and therefore have to be compensated for. A simple approach would be calculating a wind speed-up value compared to a representative stationary reference. Although the stationary reference in this experiment is only 3 m high, a strong correlation (R = 0.86) between relative mean speeds of WindLocator and reference data for M1 is observable, as opposed to M2 (R = 0.32). We assume this to be an indicator that the ground-level stationary anemometer for the particular campaign M1 is a suitable reference to also track temporal changes in the overall wind situation. The remaining differences between WindLocator and the reference tend to be local wind speed deviations, e.g. due to terrain. Figure 16 shows the speed-up value over time.

Figure 17 combines the measured results on the one hand and the UAV's GPS data on the other hand into a spatial distribution. The purple arrow indicates the mean wind direction. Each red arrow represents a measurement point, showing the measured horizontal wind direction and indicating with its length the wind speed increase compared to the stationary reference. The wind speed increase is then interpolated linearly between measurement points to create a contour plot. The background shows the digital-terrain-model data.

The calculated speed-up factor of around 2 seems plausible, when assuming a logarithmic wind profile with a roughness length of 0.2 m. A variation in the speed-up factor of +30 % and -28 % compared to the mean value is calculated. The highest increase in wind speed compared to the stationary anemometer is located towards the ridge at the upwind side, which meets our expectations concerning a speed-up effect at a steep hill. Nevertheless, directly over the highest point, where inclination angles are close to zero, an unexpected decrease in wind speed-up is detected, followed by an area of higher wind speeds at negative inclination angles. Towards the plateau in the northeast direction, we do see an expected decrease in wind speeds. The results clearly show that temporal effects must be considered when dealing with turbulence intensities or averaged wind speeds in general. The simple measurement strategy with a representative ground-level anemometer can be regarded as a proof of concept, leading to an improved estimation of spatial wind deviations for M1 when comparing it to unreferenced data. The wind speed variations are rather high but comparable to other campaigns in complex terrain (El Bahlouli et al., 2019; Wildmann et al., 2017). A plausible explanation for a decrease in wind speed directly on top of the ridge has not yet been found. For the future, a CFD validation shall give insight into whether these effects are a result of measurement errors.

As seen for M2, the presented measurement strategy obviously depends strongly on the stationary reference, its positioning and expected spatial variations and therefore cannot be considered to be valid in general. A change in wind direction from 310∘ (M1) to 240∘ (M2) leads to even lower, less correlated changes in ground-level wind speeds with increased turbulence intensity, presumably as a consequence of surrounding obstacles like hedges. A more robust measurement strategy would probably make use of a more representative stationary reference of greater height.

These findings are currently being evaluated with a simulative approach to find more robust measurement strategies, independent from the terrain, location, surface and prevailing wind situation. Once this has been obtained, several of such measurements (for example for different overall wind speeds and directions) might be combined to achieve an extensive insight into the location's wind situation. In future, this could allow bankable site assessment, similarly to how CFD simulations are used today.