This paper presents a nested codesign (combined plant and controller design) formulation that uses optimal design of experiments (DoE) techniques at the upper level to globally explore the plant design space, with continuous-time control parameter adaptation laws used at the lower level. The global design space exploration made possible through optimal DoE techniques makes the proposed methodology appealing for complex, nonconvex optimization problems for which legacy approaches are not effective. Furthermore, the use of continuous-time adaptation laws for control parameter optimization allows for the extension of the proposed optimization framework to the experimental realm, where control parameters can be optimized during experiments. At each full iteration, optimal DoE are used to generate a batch of plant designs within a prescribed design space. Each plant design is tested in either a simulation or experiment, during which an adaptation law is used for control parameter optimization. Two techniques are proposed for control parameter optimization at each iteration: extremum seeking (ES) and continuous-time DoE. The design space is reduced at the end of each full iteration, based on a response surface characterization and quality of fit estimate. The effectiveness of the approach is demonstrated for an airborne wind energy (AWE) system, where the plant parameters are the center of mass location and stabilizer area, and the control parameter is the trim pitch angle.
Optimal design of active systems presents a number of challenges when the plant and controller exhibit bidirectional coupling. If a system design exhibits bidirectional coupling, the optimal plant design depends on the controller and vice versa. Systems, such as automotive suspension systems [1–3], elevator systems in Ref. , advanced powertrain systems in Ref.  and airborne wind energy (AWE) systems in Refs.  and , have been shown to exhibit such coupling. Traditionally, combined plant and controller optimization (termed codesign) has been used to address the challenges associated with the codesign of coupled systems. In general, codesign methodologies can be broken down into four categories: sequential, iterative, nested, and simultaneous (see Ref. ). A sequential strategy completes the plant and controller optimization problems in successive order. A variant of the sequential approach, proposed in Ref. , makes use of control proxy functions to augment the objective function of the plant, thereby separating the plant and controller optimization problems into two subproblems. The iterative approach fully optimizes the plant design for a given controller, then optimizes the controller design for a fixed plant, then repeats the cycle (see Refs.  and ). A nested optimization approach contains two loops: an inner loop that completes a full optimization of the controller while keeping the plant fixed and an outer loop that completes an iteration of the plant optimization (see Refs. [1,2], and ). This approach has been proven to converge to an optimal design under limited circumstances in Ref. . In a simultaneous optimization strategy, both the plant and controller optimization problems are carried out at the same time (see Refs.  and ). An efficient decomposition-based variant of simultaneous optimization is proposed in Ref. .
Two key gaps exist in the codesign literature:
The complexity of systems that can be considered by existing codesign tools is limited—often substantially—by restrictions on the underlying optimization tools. For example, LQR-based optimal control design techniques assume a linear system. Plant optimization techniques that use sequential quadratic programming require an accurate estimate of the gradient and Hessian of the objective function or Lagrangian for constrained problems. Practical implementation of Pontryagin's minimum principle (PMP)-based techniques requires a relatively simple, closed-form expression for system dynamics.
Existing codesign techniques do not leverage the ability to adjust control parameters in real time. In a nested codesign framework, in particular, the ability to adjust controller parameters during the simulations/experiments can dramatically reduce the time and cost of the optimization process.
This paper presents a novel nested codesign methodology that addresses the above issues. The framework presented here fits into the nested methodology proposed by Fathy et al. , and the techniques used at both the plant and control design levels are specifically tailored to meet address the gaps in the state of the art discussed above. Since the nested methodology involves a complete controller optimization for every iteration of the plant design, it represents an ideal framework for adapting control parameters during simulations or experiments (where the plant design remains fixed over the course of a simulation or experiment). Table 1 provides a brief comparison of existing nested codesign strategies, including the strategy proposed herein. The current proposed framework presents the following advantages:
The use of online adaptation mechanisms to adjust control parameters enables a natural extension of the codesign strategy to the experimental realm, where plant designs can only (typically) be adjusted at discrete iterations, but control parameters can be adapted continuously.
The proposed framework relies on optimal design of experiments (DoE) techniques that explore a global design space and are not restricted to linear systems or those for which the system can be represented through a small number of closed-form equations.
In the proposed framework, G-optimal DoE is used to select a set of candidate plant designs that cover the design space while maximizing a statistical information metric. By selecting a set of candidate plant designs that cover the design space, this approach considers the possibility of multiple local optima. In the inner loop, a continuous-time adaptation law is used to find the optimal control design during the course of simulating or running an experiment for each candidate plant design. In this work, two continuous-time adaptation laws are proposed for this purpose: extremum seeking and entropy-based design of experiments. Extremum seeking (ES) is a commonly used, though local, adaptation technique for finding the maximizer or minimizer of an system whose extremum is unknown. Entropy-based continuous-time DoE, a novel contribution of this paper, attempts to overcome the limitations of ES (namely locality of the solution), while leveraging the same principles that are used in the outer loop plant DoE. Because information-based design techniques are typically performed in discrete batches (i.e., a finite number of design candidates are selected through each design of experiments), the tailoring of information-based techniques for use in continuous-time control parameter adaptation represents a unique contribution on its own. After each iteration of the full system optimization (G-optimal DoE for the plant and continuous-time control parameter optimization), the plant design space is reduced using a statistical quality of fit metric and hypothesis testing.
In this work, the proposed codesign strategy is validated on an AWE system. AWE systems replace the traditional tower of wind turbines with tethers and a lifting body to harness high altitude winds. In general, the wind at these high altitudes is more consistent and greater in magnitude than ground level winds (see Ref. ). The tethers significantly reduce the material costs associated with the tower and foundation of traditional wind turbines and allow for adjustment of the height of the turbine. A substantial quantity of literature has focused on control design for AWE systems, including the use of optimal control techniques for altitude optimization in Refs.  and , as well as the control of power-augmenting crosswind flight in Refs. [18–22]. However, there has been far less research focused on the combined plant and control design of an AWE system. Deodhar et al.  and NikpoorParizi and Vermillion  have both shown that a coupling exists between the plant and controller design (i.e., the optimal controller depends on the plant design and vice versa). Consequently, we focus on an AWE system—specifically, the Altaeros' Buoyant airborne turbine (BAT), shown in Fig. 1—as a case study. In particular, we take the longitudinal center of mass location and stabilizer scale factor (which dictates the surface area of the horizontal and vertical stabilizers) as plant parameters to be optimized, whereas we take the trim pitch angle as the controller parameter to be optimized. While these design parameters are the same as those considered in Refs.  and , the present paper differentiates itself from the aforementioned work through the unique continuous-time controller parameter adaptation that is used in the inner loop of a nested framework. Furthermore, the present paper represents a greatly extended version of our earlier conference paper, , which only considers extremum seeking as the inner loop control parameter adaptation strategy.
Nested Optimization Framework
A general block diagram of the codesign framework considered in this work is given in Fig. 2. This optimization strategy generates a batch of candidate plant designs using G-optimal DoE, which are split into training and validation points. Once those experiments are designed, they are tested via either simulations or experiments, during which an adaptive control law is used to find the optimal controller design. After finding the integral cost associated with each plant design operating at the optimal controller design (), a response surface characterization is carried out using only the training points. Using the response surface characterization, a quality of fit calculation based on the validation points, and hypothesis testing, the design space is reduced. At the next iteration, a batch of candidate plant designs is generated in the reduced design space. The process is repeated until the original design space is reduced to a sufficiently small design space. In the remainder of this section, we describe each of the elements of the codesign framework in mathematical detail.
Selecting Candidate Plant Design Parameters: G-Optimal Design of Experiments.
where each column of Z corresponds to a single candidate design point (i.e., zi represents the regressor associated with ith candidate design point). The structure of the regressor vector is chosen to account for anticipated dependencies in the plant parameters; e.g., if it is anticipated that the dependency of J on will be quadratic, then quadratic terms will be included in the regressor vector. G-optimal DoE populates the design space based on an information metric called the prediction variance, which is equal to the maximum entry in the diagonal of .
The generated set of points are split into training and validation points. Training points are used to characterize the response surface (i.e., the actual dependency of J on after experiments have been run), while validation points are used to assess the quality of fit of the generated response surface characterization.
Continuous-Time Control Parameter Optimization: Option 1—Extremum Seeking-Based Adaptation Law.
The nested codesign approach involves a full optimization of the controller for every iteration on the plant design. In complex systems where it is necessary to run simulations and/or experiments to evaluate the performance index under a given design, it becomes advantageous to consider real-time control optimization strategies, which allow the control parameter to be optimized during these simulations/experiments. Extremum seeking is a common adaptive control algorithm used in continuous-time optimization, where there is limited or no knowledge of how the performance index depends upon the control variable(s) (i.e., no first- or second-order information is available). A brief summary of ES is provided here. The general block diagram of multivariable ES from Ref. , with added context to the codesign problem at hand, is shown in Fig. 3. In general, ES uses a sinusoidal or random perturbation applied to the approximated optimal controller parameter and feedback of the instantaneous performance index value to converge to the true optimal controller parameter. The control parameter update shown in Fig. 3 corresponds to a single control parameter contained within the control design vector (i.e., ). In the multiparameter ES case, L perturbation frequencies are used for the identification of L parameters (see Ref. ). The perturbation frequencies are selected to be sufficiently small with respect to system dynamics and not equal to the frequencies associated with noise. Each perturbation frequency corresponds to a single element within the control design vector. The filter oscillation frequencies ( and ) are selected to be slower than the perturbation oscillation frequency. It has been shown that with proper selection of ωi, ki, , and ai (all of the parameters associated with ES are listed in Table 2), ES will converge to a local minimizer or maximizer . ES in this work is used for minimization, which can be handled by ES by changing the integrator gain to .
Proper tuning of ε and ensures that must remain sufficiently small for a sufficiently long time period before convergence is detected.
Continuous-Time Control Parameter Optimization: Option 2—Entropy-Based Design of Experiments Adaptation Law.
One clear limitation of the extremum seeking-based adaptation law is that it is local in nature, thereby resulting in a codesign formulation where the outer loop explores the global plant design space but the inner loop is limited to local control parameter optimization. To remedy this limitation, we seek a continuous-time control parameter optimization technique that mirrors the global design space exploration performed by the outer loop plant DoE.
Ultimately, the optimal DoE techniques used at the outer loop are based on selecting designs that maximize a statistical information metric (for G-optimal design, this is prediction variance). At the end of each iteration, the design space is reduced based on a response surface characterization and quality of fit estimate. To mirror these techniques at the inner loop, we adjust the control parameters, , based on a continuous-time information metric. Furthermore, we reject parameters from the design space in continuous time, based on a continuous quality of fit update. A fundamental tenet in performing this continuous-time information-based control parameter adjustment and design space reduction is the continuous-time estimate of normalized information entropy, , which characterizes how uncertain we are about the performance index for a given value of . While this process is conceptually similar to existing literature on adaptive DoE ([30,31]), a fundamental distinction lies in the fact that the adaptive DoE results of Refs.  and  are iteration-based, not continuous time-based. The novel contribution of the entropy-based DoE is the ability to optimize a parameter vector of interest based on an information metric in continuous-time. The entropy-based control parameter optimization framework is shown graphically in Fig. 4.
Quantifying Normalized Information Entropy.
Figure 5 illustrates how information entropy evolves during a ten second sample simulation. The upper half of the figure shows the values of pc (a scalar in this case) as a function of time. In the lower portion, the entropy at the beginning (left) and end (right) of the simulation can be seen. Because values of pc greater than 0.2 were not visited during the 10 s period, no reduction in normalized entropy is observed over this range of design parameters.
Continuous-Time Control Parameter Update Law.
where for a small delay time, Td. Here, Ksel is a constant gain and is the distance between the controller design tested at time and the controller design presently tested. Inclusion of the term in Eq. (10) ensures that does not continue to get adjusted after it was recently changed by a large amount. Inclusion of the term helps to restrict near-term subsequent candidate designs not to veer too far from the present value of . To ensure that the lower level controller can track the commanded control design, the output of the entropy-based DoE is passed through a first-order filter, as illustrated in Fig. 4.
Continuous-Time Controller Response Surface Characterization and Design Space Reduction.
As discussed in the previous subsection, the entropy-based control parameter adaptation law adjusts the control parameter in the direction of maximum entropy, not necessarily in the direction of the optimal value. To estimate the optimal control parameter, , another step is needed. This step is the continuous-time response surface characterization and design space reduction, wherein (i) a continuous estimate of is maintained and (ii) control parameters that are nearly certain to be sub-optimal based on this response surface are excluded from C. It is important to note that the response surface characterization at this stage is performed with respect to the control parameters alone, for a given plant design. Once the inner loop control parameter optimization is complete, a separate response surface characterization, with respect to the plant parameters, is performed.
where Krej is tuned to capture the relationship between normalized entropy and uncertainty. It can be seen that a candidate control design will be rejected if its associated performance index exceeds the estimated optimal performance index by an amount that depends on uncertainty (the larger the uncertainty, the larger this amount will be).
Plant Response Surface Characterization and Design Space Reduction.
Since the overall goal of the codesign framework is to converge to an optimal system design, it is important to reduce the size of the design space at each iteration so as to focus only on plant designs that could possibly be optimal. The main idea of the design space reduction is to reject, with confidence, portions of the design space that produce inferior performance. The response surface computed in the outer loop yields two important quantities that can used to shrink the design space. The first is the approximated performance index for a given plant design operating at the optimal controller design (). The second is the quality of fit metric, S, which characterizes how well the response surface approximates the true system performance index at validation points.
Design points are rejected whenever . For this work, a 95% confidence interval, which corresponds to , was chosen. The reduced plant design space, which excludes all points rejected by the Z-test, is denoted by P.
Codesign Case Study on an Airborne Wind Energy System
The general optimization strategy detailed in Sec. 2 has been applied to the Altaeros Energies BAT, using the numerical model from Ref.  for this case study. The BAT is a good candidate system for this optimization framework because it gives rise to complex dynamics with many nonlinearities, as well as coupling between plant and controller parameters. In fact, the works from Refs.  and  demonstrate bidirectional coupling for the BAT.
In this case study, we focus on two plant parameters and one controller parameter. It is important to note that this framework can be extended to many parameters; however, the restriction to two plant parameters aids in the visualization of the optimal DoE and design space reduction processes in two dimensions. The specific plant parameters of interest are the stabilizer reference area scale factor (KA) and the longitudinal location of the center of mass (xcm). Therefore, . The controller parameter of interest is the pitch angle setpoint (also called the trim pitch angle), denoted by .
Table 3 provides a summary of tuning guidelines for both adaption laws presented in Sec. 2. The tuning guidelines for ES are based on ES tuning guidelines from Ref. . For the entropy-based DoE adaption, general tuning guidelines are provided in the table. However, when tuning these parameters, it is useful to consider the units in the given application.
Moving forward, the aforementioned codesign framework will be extended to experimental codesign using the lab scale, water channel-based platform detailed in Ref.  at UNC-Charlotte. Thus, the results that are presented here provide a foundation for the extension to the experimental setup at UNC-Charlotte.
Here, τ represents a vector of generalized forces, V is the wind speed, and ψwind is the wind direction. The angle of attack (α) and side-slip angle (β) describe the orientation of the apparent wind vector with the body-fixed coordinates of the BAT. Aerodynamic forces and moments are functions of both α and β.
Here, CD, CL, CS, CMx, CMy, and CMz represent the drag, lift, side force, roll moment, pitching moment, and yaw moment coefficients of the BAT, respectively.
In general, an AWE system contains a flight controller that is used to maintain the attitude of the system to desired setpoints. The flight control architecture for the BAT is shown in Fig. 8. The flight controller of the BAT is augmented with an adaptation law that determines the trim pitch angle at which the system should operate at. Two potential adaptation laws have been presented in Sec. 2. The altitude setpoint (zsp) and the roll angle setpoint () for this work are taken to be constant.
where tf = 240 s.
where , and . These velocities correspond to flow velocities in the lab-scale platform described in Ref. . Through the dimensional analysis discussed in Ref. , these lab-scale flow velocities are equivalent to wind speeds of (i.e., a scale factor of 10) and an oscillation frequency of (i.e., a scale factor of 1/10) on the full-scale BAT, which has 100 times the characteristic length of the models used in the lab-scale setup.
Airborne Wind Energy System Codesign Results
The optimization described in Sec. 2 was applied to the system described in Sec. 3. The summary of results is broken into three subsections: The first section shows the results generated using ES for the inner loop optimization, while the second shows the results when using continuous-time, entropy-based adaptation. The last section provides a comparative assessment of these results.
Optimization Results Using Extremum Seeking.
The results presented in this section were generated using extremum seeking for the control parameter optimization. Figure 9 shows the convergence of the control parameter (trim pitch angle) for a sample plant design. The combined plant/controller optimal designs at each iteration are shown in Table 4 in the first three columns. The final column of Table 4 also lists the total reduction in the size of the plant design space (P) after each iteration, as a percentage of the area of the original plant design space. From these results, it can be seen that the design space has been reduced by 99% by the end of the fourth iteration. In Fig. 10, the candidate design points from the G-optimal DoE at the first four iterations are displayed, along with a contour plot of the response surface characterization. From this series of plots, the shrinking of the space can be visualized.
Optimization Results Using Continuous-Time Entropy-Based Adaptation.
The results presented in this section were generated using continuous-time, entropy-based adaptation for the control parameter optimization. Figure 11 shows the convergence of the control parameter (trim pitch angle) for a sample plant design (the same sample plant design that was used in Sec. 4.1). The combined plant/controller optimal designs at each iteration are shown in Table 5, in the first three columns. The final column of Table 5 also lists the total reduction in the size of the plant design space (P) after each iteration, as a percentage of the area of the original plant design space. From these results, it can be seen that the design space has been reduced by 99% by the end of the fourth iteration. In Fig. 12, the candidate design points from the G-optimal DoE at the first four iterations are displayed, along with a contour plot of the response surface characterization. From this series of plots, the shrinking of the space can be visualized.
Comparison of Adaptation Strategies.
Table 6 displays a comparison of the system response for multiple configurations, using both continuous-time control parameter optimization strategies. For each optimization strategy, four configurations are compared. The first design is a suboptimal design configuration in terms of both the plant and controller. The plant design was selected as the worst tested plant design throughout each iteration of the nested strategy. The control design was chosen by gridding the controller design space and selecting the worst performing trim pitch angle. The second is the same plant design as the first configuration, while operating at the optimal control design, which is discovered using the corresponding controller optimization strategy. The third is a random plant and controller configuration. The fourth configuration is the optimal design configuration resulting from the nested optimization strategy.
From Table 6, it can be seen this nested codesign framework results in substantial performance improvement and design space reduction, regardless of the chosen optimization technique. Specifically, the nested approach with ES for the inner loop optimization resulted in performance improvement when the optimal and worst tested configurations are compared. Similarly, the entropy-based DoE approach resulted in a performance improvement. Visualization of the performance improvement when using ES for the inner loop optimization can be seen in Fig. 13. Similarly, the components of the performance index for the entropy-based approach can be seen in Fig. 14.
When comparing the results from both controller optimization approaches, one can see that the ES-based result leads to a slightly higher (worse) performance index value than the entropy-based approach, suggesting that ES might be leading to a local, rather than global, optimum. The last column of Table 6 displays the convergence time associated with the controller optimization on the inner loop. The convergence time associated with entropy-based DoE is faster than that of extremum seeking. One key advantage of extremum seeking over entropy-based DoE is the existence of established and well-defined convergence criteria. However, these convergence criteria only guarantee local convergence, and convergence times can be lengthy. Entropy-based DoE attempts to overcome the local nature of extremum seeking by employing an information-based design metric for design point selection, but rigorous convergence guarantees remain a topic of future research.
Assessment of Net Energy Production.
From the results presented in Table 7, there is a tradeoff between robust flight control and energy harvesting. This is due to the nature of the control parameter of interest in this application. In particular, pitching up reduces the component of wind presented to the turbine, causing a reduction in energy production. On the other hand, adjustment of the plant parameters alone has minimal impact on net energy production.
Conclusions and Future Work
In this work, a novel nested codesign framework that leverages optimal DoE techniques to adjust plant parameters in an outer loop and continuous-time adaptation to perform control parameter adaptation in an inner loop. The use of DoE techniques enables global exploration of complex design spaces, whereas the continuous-time adaptation enables direct extension of the proposed strategy to the experimental realm, where controller parameters can be adjusted during experiments. The results presented here display the capabilities of the nested codesign framework on an AWE system. From the results presented based on the numerical model of the Altaeros BAT, the plant design space was reduced to less than 1% of the original design after four iterations.
In future work, development of convergence criteria of the entropy-based DoE adaptation law and optimizing the system over a variety of environmental conditions will be explored.
This work was supported by eentitled “CAREER: Efficient Experimental Optimization for High-Performance Airborne Wind Energy Systems.”
National Science Foundation award number (1453912).
control design space
normalized information entropy
reference area scale factor
performance index value while operating at the optimal controller design () and a candidate plant design ()
instantaneous performance index value
response surface characterization
number of validation points
number of training points
plant design space
candidate plant design
candidate control design
quality of fit metric for response surface characterization
x-location of center of mass
test statistic used for hypothesis testing
trim pitch angle