Edinburgh Explorer Variation of loads on a three-bladed horizontal axis tidal turbine with frequency and blade position

Sustainable and cost eﬀective design for tidal current turbines requires knowl-edge of the complex nature of unsteady loads on turbine components including blades, rotor and support structure. This study investigates exper-imentally the variation with frequency of rotor thrust and torque loads, of streamwise root bending moment on individual blades and of loads on foundation at the bed. Comparisons between these diﬀerent load spectra are also established. The impact of absolute rotor angular position on blade and rotor thrust loads is also examined. The study is based on measurements from a 1/15 scale, three-bladed, horizontal axis machine tested in a recirculating ﬂume, in onset ﬂows of 3% and 12% turbulence intensity. It is found that for frequencies below the rotational frequency, load spectra are correlated to spectral density of the onset ﬂow velocity. Above the rotational frequency, loads are mainly aﬀected by turbine operation phenomena. The tower shadowing eﬀect is clearly identiﬁed through frequency and angular analysis. Finally, thrust loads as experienced by the rotor alone are for the ﬁrst time compared with streamwise and transverse foundation loads. Higher frequency loads experienced by the tower are shown to be aﬀected by dif-ferent vortex shedding regimes associated with diﬀerent regions of the wake. All the experimental measurements presented in this article can be accessed from http://dx.doi.org/10.7488/ds/2338 .


Introduction
Systems for generation of electricity from tidal streams have attracted increasing interest in recent decades. Although a variety of design concepts have been developed, the most widely studied systems are horizontal axis turbines, similar in configuration to wind turbines. Several developers have now deployed such turbines at full-scale test sites (Parkinson and Collier, 2016;MayGen Ltd., 2016). However, there remain challenges for design, particularly with regard to accurate prediction of the unsteady loads that determine component specification for adequate operational life. For the similar case of a wind turbine, unsteady loading on blades, and hence on the turbine, is due to the rotation of the blades through a non-uniform onset flow and in proximity to a fixed support structure. Onset turbulence also play a role in the unsteady loading. The influence of these factors on wind turbine design have received attention through experimental studies such as the MEXICO campaign (Schepers et al., 2014) and CFD analysis, particularly suited to complex onset flows (Thé and Yu, 2017;Storey et al., 2016). Such studies have informed development of engineering tools for prediction of the time-varying loads and component load spectra (Moriarty and Hansen, 2005;Smilden et al., 2016) that are required for fatigue design.
Similar information is required for tidal stream turbines for which the operating environment comprises a depth constrained turbulent flow, potentially with free surface waves, and for which the supporting structure is typically much larger, relative to the turbine, than for wind turbines. Experimental datasets provide a crucial basis for development of understanding of the processes contributing to component loading and for validation of numerical models. Tidal turbine loading and performance have been analysed for threebladed horizontal axis turbine models of various diameter, (Bahaj, Batten and McCann, 2007;Kolekar and Banerjee, 2015;Gaurier et al., 2015;. Several studies have also addressed unsteady loading of rotors or turbines due to waves, in towing tanks Luznik et al., 2013;Galloway et al., 2014;Doman et al., 2015) and in flumes and tanks (Gaurier et al., 2013;de Jesus Henriques et al., 2014;Fernandez-Rodriguez et al., 2014;Martinez et al., 2017). The impact of turbulences on turbine performance and loading has also been explored (Chamorro et al., 2013;Mycek et al., 2014;Fernandez-Rodriguez et al., 2014;Milne et al., 2016;Blackmore et al., 2016). Durán Medina et al. (2017) carried out an in-depth analysis of the correlations between onset flow velocity and turbine power production based on spectral methods, empirical mode decomposition and time-dependent intrinsic correlation. However, there has, to-date, been limited in-depth analysis of the frequency variation of loading of horizontal axis tidal turbine blades and the turbine supporting structure. This work analy-ses the relationship between the frequency variation of principal components of loading of a horizontal axis tidal stream turbine focusing on streamwise blade loading, thrust and power developed by the three-bladed rotor, and the streamwise and transverse loading on the entire turbine assembly comprising rotor, nacelle and tower.
The experimental approach and flow conditions studied are summarised in Section 2, variation of rotor and blade loading are investigated in Section 3 and factors leading to unsteady loading of the support structure are analysed in Section 4. Conclusions are provided in Section 5.

Turbine model
The experimental turbine model used was a bed mounted three-bladed horizontal axis machine with a rotor diameter D = 1200 mm. The blade design was developed to produce a radial variation of the rotor thrust coefficient at a specific tip speed ratio similar to that of a full scale generic turbine . The idea behind this approach is that highest rotor loads and wakes are phenomena which are predominantly affected by the thrust characteristics of turbines. The turbine model was developed to integrate a high level of instrumentation in order to measure the following quantities: • streamwise root bending moment for each blade • rotor torque • rotor thrust • rotor absolute angular position A detailed description of the experimental model, of its design and manufacture can be found in  and a section view of the model is reproduced in Fig. 1  A primary consideration for the model design was to maximise quality of load measurements. This approach was implemented by locating the sensors as close as possible to where the loads they are measuring are applied. When considering the load path, the sensors are located 'upstream' of the shaft rotary seals so that their associated parasitic friction does not affect the load measurements. As a consequence, these sensors are in a wet environment and so were designed to be waterproof. The turbine generator was simulated by a brushless servo motor controlled in speed.
The turbine model was installed so that the rotor axis was located at mid-depth (1 m deep). The bottom of the turbine tower was mounted to a fixed mooring structure via a six-axis load cell of the type TYPR FX2.6 manufactured by the company SIXAXES.
Signal from all turbine sensors (including the six-axis foundation load cell) were synchronously sampled at 256 Hz. Sample measurement time series can be found in .

Testing facility
The study was conducted in the flow recirculating facility of IFREMER in Boulogne-sur-Mer, France shown on Fig. 2(a). The flow channel is 4 m wide, has a usable length of 18 m and was operated at 2 m depth (Germain, 2008). The blockage ratio, taking into account the rotor, the tower and support structure is 0.2. The inlet of the flume is fitted with removable flow conditioning units (see Fig. 2(b)) which produced two levels of streamwise turbulence intensity T I u : 3% with the flow conditioning units and 12% 7 without. T I u is defined by: where σ(u) is the standard deviation of the streamwise velocity component u and u is the mean of u. Measurements were taken at turbine hub location in the absence of the 8 turbine. The streamwise mean velocity was 0.814 m · s −1 for T I u = 3% and 0.819 m · s −1 for T I u = 12%. With LDA, the sampling frequency is not constant as measurements take place each time a seeding particle passes through the measurement volume. It is also different for the two velocity components measured. The mean sampling frequency for the high and low turbulence conditions were 283 and 941 Hz respectively in the streamwise direction and 71 and 258 Hz respectively in the transverse direction. The sampling durations were 256 and 200 s for the high and low turbulence conditions respectively.
In order to carry out spectral analyses of flow velocities, the LDA timeseries are first re-sampled at a constant frequency corresponding to the mean sampling frequency of the timeseries. The re-sampling is applied using linear interpolation and alternative sampling frequencies were tested to ensure that the chosen frequency was sufficient to resolve the frequency range shown.
Alternative methods to compute periodigrams without re-sampling, such as the Empirical Mode Decomposition -Hilbert Spectral Analysis lead to comparable results, as shown by Durán Medina et al. (2017) for the Empirical Mode Decomposition -Hilbert Spectral Analysis method. The spectral density of the velocity fluctuations with respect to the mean is shown in Fig.   3 for streamwise (u ) and transverse (v ) fluctuations for the two T I u levels.
The spectral densities are normalised by u 2 . The spectral density curves shown were obtained by averaging components over bins whose frequency widths are evenly distributed on a log scale. This means that high frequency bins contain a wider frequency range than low frequency bins which leads to curves being smoother toward higher frequencies. This processing was applied to all spectral density plots in this article. As expected, there is higher turbulent kinematic energy density across the frequency range for the flow with higher turbulence intensity. For T I u = 12%, u and v spectra are relatively close, implying isotropic turbulence. However, for T I u = 3%, the flow is anisotropic with u /v > 1 for frequencies below 2 Hz and this ratio increases with decreasing frequency. This difference is due to the fact that for the low TI conditions, the flow enters the test section of the flume through a honeycomb structure (shown in Fig. 2(b)). This constrains the flow in the cross-flow direction, hence limiting variance of transverse fluctuations. In the inertial sub-range (at frequencies roughly above 1 Hz here) it can be seen that all curves have a slope which broadly follows the Kolmogorov minus five thirds power law. However towards the higher frequencies, u at low TI departs significantly from that slope. This can be attributed to the anisotropy of the flow which implies that the Kolmogorov dissipation scale no longer fully applies (see Durán Medina et al. (2015) and Thorpe (2007)).

Reynolds and Froude numbers
The chord based Reynolds number associated with the blade profile at radius r is defined as follows: where ρ is the water density, λ is the tip speed ratio (TSR), R is the rotor radius, c(r) the blade chord length at radius r and µ the dynamic viscosity of water. For r/R 0.2, c(r) varies little with r compared with r/R (see table   2 of  ). From (2) Re c (r) can therefore The rotor diameter based Reynolds number is given by: Re D = 975 000 throughout this article.
The Froude number of the test channel is defined by: where g is the gravitational acceleration and h the channel water depth. For the experiments reported in this article, F r = 0.18.

Root bending moment sensor quality check
Each blade was mounted to the hub through a root bending moment sensor. More details on the sensor design and calibration can be found in . As noted previously by the authors, the sensors developed a slow drift fault during the initial test programme.
That drift progressed slowly, so providing that the signal remained within the saturation limits of the amplifier and that the sensors were 'zeroed' several times per day, measurements were deemed to be accurate within the 5% band established through calibration, especially for investigating the fluctuating component of the load. For time series which met this quality criteria, agreement between the streamwise force on the rotor inferred from the root bending sensor and the rotor thrust measured directly at the shaft was checked. The root bending moment sensors measure bending moment on the blades in the streamwise direction. The corresponding streamwise force on the blade is a function of the radial position of the centroid of the streamwise load on the blades. This 'lever arm' distance cannot be measured experimentally and was instead estimated with a blade element moment (BEM) theory model. The BEM tool is as described in Hansen (2008), with Spera high axial induction factor correction. The lever arm distance obtained by this method (see ) varies only by 2.3% over the tip speed ratio (TSR) range 4.5 to 6.0 and a constant value of 330 mm was therefore used to infer the mean streamwise force on the blade from the mean of the root bending moment signal. Comparison is shown in Fig. 4 for tests at T I u = 3%. The discrepancy varies little (between 13 and 15%) over the TSR range and can therefore be considered as consistent. The discrepancy itself is not surprising and can be attributed to the following factors.
• The lever arm distance is computed using a standard BEM approach which does not account for the blockage factor (14%), onset flow profile or flow turbulence, which are all present in the physical experiment.
• Due to space restrictions, the geometry of the flexure of the root bending moment sensors was that of a beam. Flexure deflection is mainly driven by bending moment but the sensor is also, to a lesser degree, affected by shear force to the extent that calibration exhibited a maximum error of 5% .

Rotor loads
The rotor loads are partly analysed using the standard thrust and power coefficients which are defined respectively by: where T is the rotor thrust, Q its torque, ω its averaged rotational speed and A r the rotor area.

Angular dependency
In order to investigate the impact of the absolute angular position of the rotor on its thrust loads, measurements from sensors are binned into 360 angular sectors (each 1 • wide) based on the rotor absolute angular position at 13 the time of the measurement. Values in each bin are then averaged to yield 'phase-averaged' measurements for each 1 • sector of the rotor full revolution. The bottom graph shows the corresponding relative discrepancy with respect to the torque and thrust transducer measurements. Measurements were carried out over a range of tip speed ratios with an onset flow velocity of 0.814 m · s −1 and a T I u = 3%.  Azimuthal variation of thrust is in close agreement with the sum of the three root bending moment signals (see top left frame) and this provides further confidence in the root bending moment quality control process of Section 2.4. The variations of the thrust (and Σ RBM) curve are overall small (within a 2% range) but two clear patterns can be observed. The first pattern is exhibited by three distinct lobes about 120 • apart, which is believed to be associated with the tower shadowing effect. The second pattern is a small distortion of the three lobe pattern with mainly the part of the lobe in the 60 • to 120 • sector being slightly smaller than for the two other lobes.
This pattern, which takes place once per revolution, is also observed in figure   8 of Section 3.2 on frequency analysis (peak A taking place at frequency f 0 ). This is correlated with the angular variation of the rotor rotational velocity (see top right frame) which indicates a lower speed in the 60 • to 120 • region.
Given that the onset flow velocity remains constant, this translates into a lower TSR which, in turns, yields a lower thrust as can be seen on the C T versus TSR curve shown in figure 6. This rotational velocity variation is believed to be due to either a mechanical artifact of the drive train or something to do with the motor speed control.   Figure   7 shows the variation of the streamwise velocity with respect to the hub velocity over the flume cross section where the turbine was deployed but in the absence of the turbine. The measurements do not quite cover the whole area of the rotor (indicated by the black circle) but a transverse asymmetry is clearly visible and consistent with the similar asymmetry in root bending moment loads observed in figure 5. The transverse asymmetry in the flow is due to the fact that it is generated by two impellers, located side by side in the bottom recirculation channel of the flume, which are both spinning in the same direction (see Germain (2008) for further detail on the flume layout).
The flow conditioning units fitted to the flume reduced the transverse flow asymmetry to some extend but did not suppress it completely.

Thrust and torque loads frequency analysis
Frequency dependency of rotor loads are first investigated considering the spectral density of the thrust and power coefficients defined by equation (5).
These spectral densities S C T and S C P are plotted in Figs 8 and 9 for the two different turbulence intensity regimes and also for when the turbine is rotating under dry test conditions, in air. The spectra are plotted against frequency normalised by the rotor rotational frequency f 0 . The rotor is spinning at 80 RPM, corresponding to TSR values of 6.14 and 6.18 for T I u = 3% and T I u = 12% respectively. The time series durations used to compute these spectral densities are 256, 128 and 3584 s for the dry test, the low and high turbulence conditions respectively. The normalised spectral density of the streamwise flow velocities of Fig. 3 are also shown.   For both turbulence intensity values studied, the general trend of the curves of thrust S C T and power S C P spectra are similar. For f /f 0 < 1 both curves (3% and 12% TI as green and red curves respectively) are approximately parallel to each other and roughly in-line with the velocity spectral density for the same turbulence intensity. The slope of the thrust coefficient spectra is similar to that of the velocity spectra. However, the slope of the power coefficient spectra is steeper than that of the onset velocity spectral density in the inertial sub-range. A similar observation was also reported by Durán Medina et al. (2015) who suggest that this may be related to maintaining rotor velocity constant through speed control. Nevertheless, Fig. 8 and 9 show a clear correlation between the velocity fluctuation in the onset flow and the thrust and torque loads with the TI = 12% load curves consis-21 tently above their TI = 3% counterparts by about one order of magnitude.
As the frequency increases (f /f 0 > 1) the load spectra converge and tend to a plateau at high frequency, departing from the trend of the velocity spectral density. This plateau is thought to be due to background noise associated with the sensors. This is corroborated by the fact that the level of the plateau is independent of the turbulence intensity of the flow.
Multiple specific peaks are observed in the S C T spectra for both turbulence intensities. For the sake of convenience these have been labeled in Fig.   8. The highest peak for both curves is peak C, which takes place at three times the rotational frequency. This is due to the shadowing effect of the tower, taking place when each blade passes the tower i.e. three times per rotor revolution. Peaks D, E and F take place at 6f 0 , 9f 0 and 12f 0 respectively and their amplitude decreases with increased frequency indicating that they are harmonics of the 3f 0 peak. Peak A, B and G take place respectively once, twice and 18 times per revolution and could not be explained by any obvious hydrodynamic phenomenon. It was therefore decided to investigate S C T with the turbine model taken out of the flume ('dry test') which corresponds to the thick blue line in Fig. 8. The general trend of that curve is flat since there cannot be any correlation with onset flow velocity fluctuations and this is indicative of the flat, broad band noise floor inherent to the sensor. The S C T values appear to increase at the lowest frequencies. This is believed to be due to noise in the spectral density whose computation for low frequencies involves no or only little averaging (as explained in section 2.2). The curve exhibits two clear peaks (labeled A and G). Peak A occurs once per revo-lution and is attributed to a mechanical artifact of the turbine model drive train, e.g. a slight misalignment between the motor and turbine shafts or a slight imbalance in the rotor. It can be seen that this affects S C T for TI = 3% and to a much lesser extent at TI = 12%. This makes sense as ambient level of load fluctuation is significantly higher for TI at 12% than at 3%. Peak B is lower than peak A and appears to be a harmonic of peak A. Peak G occurs at 18f 0 and, as this is quite high for a mechanical artifact, this is thought to be either due to vibration induced by pole cogging in the motor or a function of the motor speed controller or possibly some electromagnetic interference. Irrespective of the phenomenon at play, it is an artifact of the turbine model itself and not of its interaction with the flow. However, it is evident that it is affecting the S C T curves corresponding to the turbine in the flume on which peak G is clearly of higher magnitude than a harmonic of the lower frequency hydrodynamic forcing.
The analysis of the peaks in the S C T curves of Fig. 8 broadly also applies to the S C P curves of Fig. 9. However, peak C is of different magnitude and further peaks are observed (D, E and G). The peak associated with the tower shadowing effect (peak C) is less pronounced than for S C T which can be explained by the fact that this phenomenon is mainly a streamwise wake effect and affects therefore more significantly rotor thrust than torque. The 'dry test' S C P curve exhibits three extra peak (D, E and G) compared with its S C T counterpart. The reason for peak D, which takes place at 6f 0 , is not fully understood but is obviously a 'dry' phenomenon associated with the turbine model. It nevertheless clearly affects S C P when the turbine is in the flow. Peak E occurs at 12f 0 and is believed to be a harmonic of peak D.
Peak F takes place at 18f 0 and corresponds to peak G in Fig. 8. Peak G (of Fig. 9) occurs at 36f 0 and is believed to be a harmonic of peak F.

Frequency variation of blade root bending moment
As mentioned in section 2.1, each turbine blade is fitted at its root with a sensor measuring fluctuation of bending moment due to the streamwise load (contributing to thrust) experienced by the blades. The root bending moment (RBM ) signals are rendered dimensionless using a similar approach as for the thrust coefficient C T (see (5)): where D is the rotor diameter and the other symbols are the same as for equation (5). Figure 10 shows the spectral densities of C RBM for each blade.
It can be seen that the load spectra are nearly identical for all blades, except for small discrepancies at both ends of the frequency spectrum. At high frequency, the difference is likely to be due to slight mechanical and/or electronic inconsistencies of the RBM 2 sensor with respect to the two others, leading to a higher background noise level.
With the consistency in spectral density of the root bending moment established across all three blade sensors, the relationship between spectra of velocity, of blade loading and of rotor loading is now considered. Figure 11 compares the spectra of C RBM 1 (corresponding to blade 1 only) with that of C T , with the normalised velocity spectral density of the onset flow (u ) and with the spectral density of C ΣRBM defined as follows: where suffix i is the blade index. For f /f 0 > 0.5 the C RBM 1 curve exhibits a series of sharp peaks (labeled A to F) of decreasing amplitude and whose frequencies correspond to f /f 0 = 1, 2, 3, 4, 5 and 6 respectively. Peak A is expected to be due to the 26 tower shadowing effect which affects each blade once per rotor revolution.
This has a similar amplitude to that of the low frequency turbulence induced loads. The frequencies and amplitudes of peaks B to F suggest that they are harmonics of peak A.
The main peak exhibited by the C ΣRBM curve is at 3f 0 (peak C) corresponding to the tower shadowing effect on each blade once per revolution.
Peaks F, G and H on that curve, taking place respectively at 6f 0 , 9f 0 and 12f 0 are harmonics of peak C. Peaks A, B and I of the C ΣRBM curve are not believed to be due to hydrodynamic phenomena but to mechanical and/or motor artifacts (see figure 8 and Section 3.2).
There is agreement between the spectra of thrust and of sum of root bending moment sensors over much of the frequency range (to about 3f 0 ) but the RBM sensors have a lower noise floor and so capture greater detail on the frequency dependency to about 20f 0 . Over this range, the distinct three harmonics of peak C are observed (peaks F, G and H). Figure 12 compares the spectral density of C RBM 1 for the two levels of onset flow turbulence intensity (3% and 12%) and also shows the normalised velocity spectral density of the corresponding onset flows. The vertical gap between the two C RBM 1 curves is qualitatively consistent with that between the velocity spectral density curves. The frequency at which blade loading becomes uncorrelated with the onset flow spectrum is around 0.6f 0 . This is slightly higher than observed in Figure 11 and this is likely to be due to the  (2017), Chamorro et al. (2013)). It can be seen that for the flow with higher turbulence intensity, the peaks of the blade loading spectrum associated with the effects of the turbine structure are less pronounced. This is believed to be due to the higher level of turbulence induced loads which reduces the apparent amplitude of the peaks. Qualitatively, similar behaviour has been observed through CFD simulation with a time-varying turbulent onset velocity field imposed to represent similar magnitude of turbulence (Ahmed et al., 2017). For frequencies greater than the third harmonic of the tower shadowing effect, the slope of blade loading spectra is steeper than that of the corresponding velocity spectra. For TI = 3% this rapid decay is obscured by the narrow spike at 20f 0 and by the sensor noise floor. However, for TI = 12% a similarly steep slope is observed over a wider frequency range before the sensor noise floor is reached, resulting in lower variance over frequencies greater than 20f 0 than would be expected if loads were correlated with the onset velocity spectral density over this high frequency range.

Loads on support structure
To the authors' knowledge, in previously reported experimental investi- loads using a load cell located above waterline on the supporting tower. The latter approach cannot isolate thrust on the rotor alone from loads on the nacelle and support structure as is pointed out in Gaurier et al. (2015). Mc-Tavish et al. (2014) adopted an interesting technique to mitigate this issue by enclosing the turbine tower (which is connected to the load cell) into a hollow tube held rigidly but not mechanically connected to the tower or the load cell. This way, the tower does not experience any drag. This method does not however suppress the contribution of the drag on the nacelle to the thrust measurements.
In this experiment, the model was fitted with a torque and thrust transducer located between the shaft and the hub, and the entire turbine and support structure assembly was mounted on a six-axis load cell at the base of the tower. Data from this arrangement provided the opportunity to compare thrust loads experienced by the rotor alone with those acting on the turbine as whole. Such a comparison for mean load values over a range of TSR is shown in Fig. 13. The vertical position of the centroid of the streamwise force is also computed, using both F x and the moment along the transverse axis M y " from the load cell and this is shown in the bottom frame.
As would be expected, Fig. 13 clearly shows that the mean streamwise force F x on the overall turbine exceeds the mean thrust load on the rotor alone. The standard deviation of turbine force F x is also larger than that of shaft thrust. This can be explained by the fact that the main additional contribution to F x compared to rotor thrust loads is drag on the tower. The tower is located in the near wake of the rotor (0.4D downstream) and is therefore subjected to a more turbulent flow than the rotor itself. This leads to more variability in the tower drag force than in the rotor thrust, hence the larger standard deviation in F x than in thrust.
Mean force on the rotor is around 10% less than the mean force on the turbine. However this disparity (middle graph) decreases with increasing TSR. To understand this trend, it is useful to consider BEM theory, as derived in detail in Hansen (2008) (chap. 4), which implies that: where T is the rotor thrust, a the axial induction factor, u the onset velocity and ∝ means "proportional to". The streamwise flow velocity u 1 directly in the wake of the turbine is given by: Given that the drag force F D on the tower is proportional to u 2 1 : Combining (8) and (10) yields after simplification: From BEM, the right-hand side of (11) decreases with TSR which is consistent with the middle graph of Fig. 13.
The bottom graph of Fig. 13 shows that the vertical position of the centroid of F x is slightly above the rotor axis, even though the streamwise loads on the tower are applied below the rotor axis. This is the case because F x is dominated by rotor thrust and the thrust on individual blades (see Fig. 5 (bottom right)) is larger whilst blades are rotating through the upper half of the depth.
It is also interesting to compare rotor thrust and turbine loading in terms of spectral density. Fig. 14 shows the spectral density of these forces together with that of the transverse force F y due to the whole turbine as measured by the load cell. Although the area on which force acts differs, all three forces are normalised by 1 2 · ρ · A r · u 2 (where A r the rotor area) to facilitate comparison. For frequencies less than 5 Hz, turbine streamwise loading (F x ) is nearly identical to rotor thrust. For higher frequencies the F x spectrum is of higher magnitude than the rotor loading spectrum. This is partly due to the higher noise floor of the load cell (∼ 5×10 −6 for both F x and F y components reached at about 30 Hz) compared to the thrust sensor (∼ 4 × 10 −7 reached at 60 Hz). The F y curve is consistently lower than the F x one despite the fact that the flow is nearly isotropic, with the level of turbulence similar in the x and y directions, as can be seen in figure 3 (T I u = 12%). This is partly due to the disparity between projected area of the tower and nacelle associated with transverse force and the much larger swept area of the rotor which determines the rotor thrust.
All the major peaks of the thrust curve are explained in section 3. The F y curve exhibits a peak at the blade frequency (3f 0 ) which is most likely due to the tower shadowing effect. The two narrow peaks, at 2.18 and 1.21 Hz do not appear on the thrust or F x curves. These can be associated with vortex shedding behind the tower which is subjected to two flow regimes; the rotor wake across the upper portion and the bypass in the portion below the swept area of the rotor. To assess whether this is due to shedding the normalised shedding frequency is considered. Rearrangement of the Strouhal number yields the streamwise velocity u tow to which the tower is subjected: where f tow is the frequency of oscillation of the vortices, D tow the tower diameter and St the Strouhal number. Streamwise velocity in the turbine wake was measured in the vertical plane above the turbine axis at 0.5D downstream of the rotor and is reported in Payne, Stallard and Martinez (2017) (Fig. 18). This velocity varies from 0.5 m · s −1 to 0.9 m · s −1 from the wake to the bypass region. The tower is actually located 0.4D downstream of the rotor and below the turbine axis where the wake and bypass velocities might be slightly different than the ones measured above. However, assuming 0.5 u tow 0.9 m · s −1 is a reasonable starting point to estimate the Strouhal numbers for each section of the tower. For this velocity range the Reynolds number of the flow incident to the tower, Re tow , is in the range 5.1 · 10 4 Re tow 9.1 · 10 4 . The Strouhal number varies little for Reynolds numbers ranging from 3 · 10 4 to 2 · 10 5 and is equal to 0.19 (Bearman, 1969).
With these numbers, (12) yields u tow = 1.17 m · s −1 and 0.7 m · s −1 which are slightly higher but reasonably close to the velocities which can be expected in the bypass flow and rotor wake respectively. The 1.17 m · s −1 velocity corresponds to Re tow = 1.2 · 10 5 which is also within the range for which the Strouhal number is stable at 0.19. This strongly suggests that the peaks in The support structure considered here is a simple cylinder. Developers may consider different forms, such as for reduced drag and hence reduced momentum extraction or for structural requirements. For design of both blades and supporting structures, consideration should be given to the impact of the unsteady loads due to onset flow and the support structure shedding regimes on fatigue design and avoidance of coincidence with driving forces at harmonics of the rotor frequency.

Conclusions
Unsteady loading of a three-bladed horizontal axis turbine has been analysed. The system studied comprises a three bladed rotor on a bed mounted support structure subject to a nearly uniform onset flow. The frequency variation of unsteady loads on the rotor, the supporting structure and the blades are analysed and related to characteristics of the turbulent onset flow and of the turbine.
For low frequencies, the spectra of blade loading, of rotor thrust and of the streamwise force on the turbine are correlated with the spectrum of the streamwise onset velocity. However, a steeper slope is observed for the spectra of power coefficient. Correlation of streamwise load spectra with the onset velocity spectra occurs up to a frequency equal to half the rotational frequency for blade loading, but to around three times this frequency for the rotor loading. For low onset turbulence (3% turbulence intensity) blade loading peaks are clearly observed at the rotational frequency and at five higher harmonics. This is associated with the tower-shadow affect with a reduction of blade load observed once per cycle when the blade is approximately 15-20 degrees misaligned with the tower. The amplitude of the peak at rotational frequency is similar to that of the highest onset flow induced loads at low frequencies. Similar fluctuations of load are observed when aggregated across the three blades and in the rotor thrust spectrum, with each harmonic at three times that of the blade load. For higher value of onset turbulence (12% turbulence intensity) blade loading is observed at the rotational frequency and the first two harmonics but with greater variation at intervening frequen-cies. For the high frequency range, greater than fifth and third harmonic for low and high turbulence cases respectively, the slope of blade loading spectra is observed to be greater than that of the corresponding onset velocity spectra.
Although the turbine is bed-mounted and its tower is therefore generating drag force below the rotor axis, the centroid of the streamwise force over the entire turbine (rotor, nacelle and tower) is above the axis of rotation (by about 4% of the rotor diameter). This is mainly associated with the tower shadowing effect which reduces blade loads in the lower part of the rotor disc. Mean streamwise load on the nacelle and supporting structure is approximately ten percent of rotor thrust although this reduces slightly with increasing tip-speed ratio, due to variation of thrust and hence wake velocity over this range. Analysis of both streamwise and transverse force on the tower indicates that two vortex shedding regimes occur along the tower associated with the wake and bypass flow.
The analysis presented provides a benchmark for evaluating the capability of numerical models for unsteady load prediction.

Authors' Contributions
G.P. and T.S developed the turbine model specifications and design, the tank test programme and the analysis plan, G.P led the tank testing, the data analysis and drafted this manuscript, T.S was co-investigator on the project and contributed to the analysis presented and to manuscript preparation; R.M. contributed to the assembly of the physical model and to tank testing; T.B. was co-investigator for the project and contributed to specification of tank test programme and to manuscript review.