Virtual Model Development of the Load Application System of a Wind Turbine Nacelle Test Bench for Hybrid Test Applications

Modern wind turbine systems are designed to generate and supply electricity for an operational period of 20 years or more [1]. However, past years have shown a high rate of failures in both offshore and onshore wind turbine systems [2,3]. Such frequent failures cause long downtimes requiring costly repairs which dominate the overall operation and maintenance (O&M) costs [4,5]. This has raised the need for finding ways of developing more reliable wind turbine systems. One major way of developing reliable wind turbines is by extensive system level testing and experimental validation [6].

In recent years, nacelle test benches have become an attractive method for testing wind turbine drivetrains as compared to the conventional way of field testing. The most advanced nacelle test benches, like the Dynamic Nacelle Testing Laboratory (DyNaLab) at Fraunhofer IWES, incorporate sophisticated hardware-in-the-loop (HIL) schemes to enable the emulation of field-like loads on a nacelle in a controlled testing environment [7] (Fig. 1). They offer faster and reproducible test campaigns at a lower cost compared with field testing.

However, wind turbine technologies are evolving rapidly and the operational capacities of wind turbines are breaking new records every year [8]. Consequently, the existing test benches must keep up with the growing demands of testing modern wind turbine nacelles. Upgrading the test bench operational capacity might be the obvious solution, but it is not necessarily the most feasible and realistic one. An alternative solution for complementing test bench capabilities is needed that is both economically feasible and applicable for all the existing test benches. This is where simulation techniques come into play. The advancements in simulation technology have allowed the development of complete virtual twins of the actual test setup and opened the doors for “hybrid testing” of wind turbine nacelles [9]. Combining the simulation results with the experimental measurements makes it possible to investigate the complex interactions between the device under test (DUT) and the test bench effectively. The virtual models of the RWTH Aachen University test benches [10,11] and the Clemson University test bench [12] have already demonstrated some of these advantages. These models have primarily been used for detailed analysis of the nacelle DUT. The ongoing VirtGondel research project [13] at Fraunhofer IWES aims to advance this technology further by using the virtual test bench model to augment the physical testing and provide a parallel virtual environment with load ranges beyond the test bench capacity. The virtual test bench model will also be utilized to further develop the existing methods for nacelle testing.

However, the development of such high-fidelity virtual models of a nacelle test bench for hybrid test application is challenging. The DyNaLab test bench features a hexapod system for the application of nontorque loads which comprises several servohydraulic actuators that are powered via a large hydraulic circuit. The selection of the required modeling fidelity that represents all the relevant system dynamics with optimal computation times for a system of such unique topology, size, and operational capacity is by no means trivial. Moreover, there is very limited information available in the literature for developing virtual models for a test bench of similar topology. The test bench model in Ref. [11] considers a simplified actuator dynamic model, whereas no details are given on the load application system model of the test bench in Ref. [10]. The test bench model in Ref. [12] considers a detailed model of the valve dynamics for studying the test bench control system and considers the cylinder pressure dynamics which are presented in Refs. [14,15]. However, the recommended modeling fidelity is still not evident from these works, as this was not the focus. This leaves the question of required modeling fidelity still open. For complete virtual testing, it is important to model all the relevant test bench dynamics that can influence the global system response. Therefore, a detailed study is needed to understand the important system characteristics and the required modeling fidelity.

This contribution presents the development of a virtual model of the complete nontorque load application system (LAS) of the DyNaLab test bench. Multibody simulation (MBS) is utilized for the modeling of the mechanical components of the LAS. The complete hydraulic system of the test bench is modeled using bond graph methods. Both system models are dynamically coupled via cosimulation in Simulink, which also features the LAS force control scheme using PI control. This allows implementation of the loads on the DUT in the same manner as on the actual bench. The model development methodology will be explained, and the influence of different levels of modeling depth of the system model investigated, and the model fidelity most relevant for capturing the system dynamics highlighted. The findings will aid the development of virtual models of similar test benches as the DyNaLab and shall open further doors for performing hybrid testing for wind turbine nacelles.

A detailed description of the test bench LAS is given in Sec. 2. The modeling methodology is explained in Sec. 3. The implemented cosimulation framework for the test bench LAS is described in Sec. 4. The case studies performed and discussion of the results are provided in Secs. 5 and 6. The paper ends with an outlook for future work in Sec. 7 and conclusion in Sec. 8.

The DyNaLab offers electrical and mechanical tests for a wind turbine nacelle of up to 10 MW power. With its direct drive and hexapod LAS, the test bench can apply loads on the nacelle DUT in six degrees-of-freedom (DOF) to emulate the wind loads. The LAS consists of a particular configuration of a 6DOF Stewart-Gough platform, which is driven by means of six servohydraulic cylinders. Figure 2 shows the force control scheme of the LAS. These cylinders are arranged in a hexagonal configuration. The LAS can apply up to 20 MNm bending moment and 2 MN thrust and shear forces.

The internal control of the LAS transfers the desired loads at the load application point (LAP) to the individual cylinders. This transformation is governed according to an inverse kinematic approach that performs a mapping of desired LAP loads in the task space into the required amount of individual actuator loads in the joint space. Each actuator has a dedicated servovalve controller that receives the required actuator force set points to control the servovalve. Pressure sensors in each cylinder measure the generated cylinder forces which serve as the feedback signal to the actuator controller. In this manner, each actuator has a closed-loop in its respective joint space. This enables the application of nontorque loads in 5DOFs in a controlled fashion.

Each hydraulic actuator is connected to a comprehensive hydraulic circuit that ensures stable delivery of hydraulic fluid for the entire operating range of the test bench. Figure 3 shows a simplified representation of the test bench hydraulic circuit. A set of six motor powered displacement pumps supply pressurized fluid from the reservoir tank to the high pressure line. A series of accumulators are connected to the pressure lines for minimizing pressure ripples. The pressure line also consists of several relief valves for ensuring safe pressure limits. The return line delivers the low pressure through heat exchangers into the reservoir tank.

Elements of the hydraulic system of the test bench have been modeled using the bond graph method in 20-sim software [16]. The bond graph method is an energy-based modeling approach that provides a practical way of fully coupling multidomain systems with various power conversions. In the bond graph method, effort and flow variables, also known as power variables, are used to describe how the systems interact and exchange energy. This modeling approach categorizes each element of physical systems based on their ability for energy supply, storage, transformation, or dissipation. The kinetic and potential energy storage are represented by inertia (I) and capacitance (C) elements, respectively. Energy dissipation is represented by resistance (R) elements, while lossless energy transformation is depicted by transformer (TF) elements. These system components are interconnected with power bonds (denoted with a half arrow), representing the energy exchange. The sense of the half arrow gives the direction of the power. The causality of effort and flow variables is shown with a vertical stroke, which determines whether these variables are considered input or output in the respective bond graph elements. Elements under the same effort are associated with parallel junctions (0-junctions), whereas elements under the same flow are associated with series junctions (1-junctions).

The LAS is powered by a hydraulic system composed of several key subsystems such as the hydraulic supply unit, pipes, accumulators, and servocontrolled actuators. Modeling these subsystems using the bond graph approach gives the advantage of having complete control over the causal relationships between the component models. This helps in identifying and preventing algebraic loops in the system model. Furthermore, defining the coupling interface and the associated input and output variables between the hydraulic system model and the MBS model during cosimulation becomes straightforward as these variables are simply the power variables of the bond graph model at any interface. The constitutive equations for modeling these subsystems have been gathered from Ref. [17] and are described in Secs. 3.1.1–3.1.5.

where $V˙$ is the theoretical pump delivery as a function of pump displacement d and motor speed ω. The relief valve maintains the pressure in the system to a limit value. The relief valve opens fully when the line pressure exceeds the limit pressure and is fully closed when the line pressure is below the limit pressure. This is modeled as a modulated R element with line pressure as modulation signal.

where μ is the dynamic viscosity of the fluid.

where k is the specific heat ratio for an isentropic process and $Po$ and $Vo$ are the nominal pressure and volume of the accumulator.

where x_spool is the spool displacement which is modeled by a second order transfer function characterized by the valve response frequency and damping.

where δ is the seal clearance and l_n is the seal contact length. Two modulated effort sources are linked to the piston side. The first one applies the bumper force on the piston and the second one applies the forces from external loads on the piston.

where F_r is the friction force, F_c is the Coulomb friction force, and F_s is the static friction force. ν_s is the Stribeck velocity with n as the exponent that affects the slope of the Stribeck curve. The velocity between the sliding surface is defined by ν.

where $g(ν)$ is the Stribeck function that expresses the Coulomb friction and the Stribeck effect. σ₀, σ₁, and σ₂ represent the average stiffness coefficient of bristles, the average damping coefficient of bristles, and the viscous friction coefficient, respectively. z represents the mean deflection of the elastic bristles. The first and second term of Eq. (13) represent the force of friction arising due to the elastic bending of the bristles while the third term represents the viscous friction. Both friction models defined by Eqs. (10) and (13) can be modeled as an R element in the bond graph model of the cylinder. However, in the presented work, the friction is modeled in the MBS system in the translational joints of each actuator. This results in two main advantages. First, the normal force is directly calculated in the MBS system. Second, the more robust HHT integrator of MSC Adams can be used to solve the differential equations.

The actual hydraulic circuit of the test bench features several components that are there for safety and maintaining steady operation. By assuming that the entire system operates fault free and ignoring thermal effects, several components like filters, heat exchangers, check valves, and emergency drains can be ignored. Moreover, the goal of the virtual model is realistic simulation of the LAP loads. Therefore, only those aspects of the hydraulic circuit have been considered that are required for a realistic generation of the LAP forces under normal mode of operation. Figure 8 shows the selected abstraction of the test bench hydraulic circuit along with the equivalent bond graph model.

The individual accumulators are lumped together as an equivalent model representing the combined volume of all the accumulators. The pipes are modeled as a lumped model as explained in Sec. 3.1.2. The fluid supply system assembly to which the hydraulic actuator components are connected closely resembles the actual test bench system. The hydraulic circuit model has defined interfaces that allow it to connect with external models that are located at the spool valve input signal, the piston force input and the piston velocity output.

A high fidelity MBS model of the test bench has been developed that models all the components of the test bench drive system and LAS. Figure 9 shows the complete MBS model of the test bench and the calibration unit modeled using mscadams software [20]. The calibration unit is a steel reaction structure that serves as a DUT. The DUT structure with platform, flange adaptors, coupling, and motor rotor are modeled as flexible bodies. These bodies are created by modal reduction of their respective FE models using the component mode synthesis (CMS) method [21]. All components in the model are connected with each other and the ground using constraints according to their configuration in the actual system. Table 1 provides information on the MBS model fidelity. The choice of model fidelity for the test bench system and the DUT was mainly driven by the need to achieve the maximum accuracy possible without raising computational costs. More details on the test bench MBS model are provided in [22,23].

The high-fidelity test bench MBS model, the hydraulic system model, and the actuator controllers are coupled via cosimulation. This will allow emulation of the actual LAS of the DyNaLab test bench. Simulink serves as the cosimulation interface. The bond graph model of the hydraulic circuit developed in 20-sim is exported to Simulink as a functional mockup unit (FMU). The multibody simulation model of the test bench is imported as an Adams-Simulink block.

The hydraulic system model in 20-sim is fully coupled to the MBS model in Adams at the hydraulic actuator interface (as shown in Fig. 10). The 20-sim model receives the piston reaction force as input variables and delivers the piston velocities as output variables. The MBS model in Adams receives the piston velocities as input variables and applies them as constraints to the respective actuators. The resulting piston force reactions are delivered as output variables. Apart from these coupling variables, additional outputs are extracted from the MBS model and the 20-sim model for postprocessing of the results. These include the LAP forces, LAP displacements, piston displacements, and cylinder chamber pressures. Both the hydraulic model in 20-sim and MBS model in Adams communicate by passing the input and output variables back and forth with a communication interval that corresponds to the simulation step size. The hydraulic model in 20-sim uses the fourth order Runge–Kutta method for time integration with a step size of 60 μs. The Adams model uses the HHT integrator with a step size of 6 ms. Simulink provides the necessary rate transitions between the coupled models and performs the time integration using the 4th order Runge–Kutta method with a step size of 60 μs.

Each individual hydraulic actuator model in 20-sim is controlled by a dedicated PI controller which is modeled in Simulink. The difference in the actuator set point force and the cylinder pressure calculated force is sent to the PI controller as an error signal. The PI controller attempts to reduce the error by controlling the spool positions. This way, the force is being controlled in the joint space that translates into the desired LAP forces in the task space according to the inverse kinematics of the hexapod system. This force control scheme is similar to the one implemented in the actual test bench.

Several case studies are performed with different model variations, which are listed in Table 2. In this way, the modeling features that are most relevant for force generation at the LAP can be identified. The base model represents an idealized hydraulic circuit that ignores the fluid inertial effects in the pressure line, the spool valve dynamics, and the frictional losses in the hydraulic cylinder and pipes. This is used as a reference, and all subsequent modeling details are added individually to the base model for comparison. Figure 11 shows the load profile considered for all the cases. The test load profile involves stepped loading in the beginning and sinusoidal load toward the end. This makes it possible to understand the system response for the static and dynamic load regimes. The tests are performed for three loading directions. These directions correspond to the thrust forces (⁠$Fx$⁠), the yaw forces (⁠$Fy$⁠), and the pitch forces (⁠$Fz$⁠) that are applied at the LAP. Though less realistic than actual field loads, such unidirectional load scenarios reveal key system characteristics that might otherwise be difficult to identify in the case of mixed load scenarios.

Figure 12 shows the influence of modeling pipe in the hydraulic circuit. Three variations of pipe models have been used: the first variation considers the fluid inertia (green); the second variation considers the line friction (yellow); and the third variation combines the fluid inertial effects and line friction (red). The first two pipe model variations focus on the individual effects of fluid inertia and pipe friction, whereas the third model variation focuses on their combined effect on the system response. In all variations, the pipe capacitance has been modeled according to Eq. (2). It can be observed that, with the exception of the transition regions, the deviations are insignificant when compared to the base model.

Figure 13 shows the influence of modeling the cylinder friction. Noticeable changes are observed for the static load case regions as the models with friction result in lower LAP force in axial direction when compared to the base model. For any given force set point for an actuator, the cylinder forces that are feedback to the PI controller are calculated from the pressure difference of the cylinder chambers in the joint space. The presence of friction raises this differential force. This results in cylinder forces reaching the set point values while the actual force applied at the piston end is less than the set point force (by a factor of the frictional force in the system). As a result, the PI controller assumes that the set point has been achieved by assessing the pressure differential force. This consequently leads to the difference in the generated LAP forces in the task space. The modeled friction has a maximum value of 0.5% of nominal actuator force, which has caused up to 12% deviations in the LAP forces. The deviations are greatest in the axial load case and are very minor for the remaining two directions. This is possibly due to larger movement of the pistons for the axial load case as compared to the pitch and yaw load cases. The friction models will return higher frictional forces for larger piston movements.

Figure 14 shows the influence of modeling cylinder leakages. Deviations in the static load regions can be observed, as the model with the highest leakage rate tends to have a continuous drop in force levels. At any stationary force level, the leakage causes pressure drops that lead to a drop in the actuator forces. The PI controller continuously attempts to maintain the force levels and therefore, due to active controller actions, the resulting forces are noisier.

Figure 15 compares the models with varying levels of valve dynamics. The results show significant deviations in the dynamic load regions and some noticeable noise in the transient regions of the step loads. The PI controller actions can possibly influence the valve response frequencies that can lead to deviations in the LAP forces. The results have shown that changes in valve response frequency of as much as 5% can cause up to 20% variations in dynamic LAP forces.

From the presented results of the case studies, it is evident that certain elements of the hydraulic circuit have a more significant impact on the resulting LAP forces than the others. The cylinder friction and valve dynamics when modeled can significantly affect the model behavior. The cylinder leakage has shown minor changes in the constant force regions, whereas the modeling of the pipe has shown negligible changes in the results.

Furthermore, in almost all cases, the $Fy$ and $Fz$ hub forces showed more noise as compared to the $Fx$ hub forces. One possible reason for this could be linked to the higher stiffness of the DUT structure in the Y and Z directions as compared to the X direction. This higher stiffness leads to higher-frequency pressure fluctuations in the hydraulic actuator model, which causes the actuator PI controller to react more aggressively, eventually leading to more noise in hub forces.

Based on the findings of the case studies, features such as cylinder friction and valve dynamics have been incorporated into the base model. A preliminary validation is performed for the virtual model by comparing the hub forces from simulations with experimental results for a sinusoidal load case. In the experimental setup (shown in Fig. 9), multiple load cells are installed between the DUT and the interface flange of the hexapod. This allows a direct measurement of forces at the load application point between the DUT and the hexapod unit during experiments. Figure 16 shows the comparison of hub forces simulated by the virtual model and those determined during experiments via the load cells. The virtual model can reproduce the dynamic axial forces with decent accuracy, having less than 4% deviation from the experimentally measured axial forces. However, the model appears to overestimate the generated $Fy$ and $Fz$ forces as compared to the experimental measurements of the corresponding loads. These deviations could be linked to several reasons, such as uncertainties in the model parameters of the hydraulic system, uncertainties in the MBS model, and uncertainties in the controller model. The actuator controller used in the virtual model is an abstraction of the actual controller on the test bench. Therefore, the modeled controller can have differences in dynamic behavior as compared to the actual controller on the test bench. Further investigations are needed to optimize the PI parameters that can make the controller more robust for all load directions.

Although several aspects were covered in the presented case studies, some modeling features that might require further investigation include, cylinder expansion, seal clearance variations, fluid viscosity variations, spool valve friction and hysteresis. In physical systems, the cylinders do expand at higher pressures leading to changes in the volume and seal clearance. This could influence the pressure dynamics and the leakage rates leading to fluctuations in actuator forces. The spool valves also have some level of friction and hysteresis effects that might have an influence on the actuator load response. The viscosity of fluids changes with variations in temperature. Changes in fluid viscosity can lead to changes in friction and leakages, which can influence the loads applied by the actuators.

In the presented work, only unidirectional load cases were considered to investigate the dynamic response of the LAS. Future work will involve further investigations using different types of mixed load cases that closely resemble the loads that are typically encountered by a wind turbine during field operation. Such investigations can help in estimating the capability of the test bench to reproduce field-relevant dynamic wind loads.

Furthermore, aspects concerning the behavior of the actuator PI controller require detailed investigations, which weren't the focus of this paper. However, it can have a strong influence on the generated hub forces, both in terms of response time and deviation from target setpoints. Further investigations are required to highlight the key aspects of developing actuator controller models that are both robust and resemble the behavior of the controllers on the actual test bench.

This contribution has provided insights into the modeling of the complete load application system of a multimegawatt wind turbine nacelle test bench. The important modeling aspects of a hydraulic circuit were revealed by a series of case studies with different modeling variations. The results showed that modeling of the actuator friction and valve dynamics has significant influence on the generated loads at the load application point. Introducing friction with stick-slip effects with a maximum value of 0.5% of the nominal actuator force can lead to up to 12% deviations in the LAP forces. Variations in valve dynamic response of as much as 5% can lead to up to 20% variations in dynamic LAP forces. Therefore, it is recommended to include these features in the system model. These findings can be relevant for modeling the load application systems of similar types of nacelle test bench systems.

The authors would like to thank the involved colleagues at Fraunhofer IWES for their contributions to the VirtGondel project. The funding of the VirtGondel project by the German Federal Ministry for Economic Affairs and Climate Action (BMWK) (No. 03EE2018) is kindly acknowledged.

The datasets generated and supporting the findings of this article are obtainable from the corresponding author upon reasonable request.

CMS =

component mode synthesis

DOF =

degree of freedom

DUT =

device under test

FMU =

functional mockup unit

HiL =

hardware-in-the-loop

LAS =

load application system

LAP =

load application point

MBS =

multibody simulation

O&M =

operation and maintenance