Elsevier

Renewable Energy

Volume 87, Part 2, March 2016, Pages 892-902
Renewable Energy

Automatic identification of wind turbine models using evolutionary multiobjective optimization

https://doi.org/10.1016/j.renene.2015.09.068Get rights and content

Highlights

  • Accurate, succinct models of wind turbine dynamics are identified from normal operating data.

  • A novel evolutionary multi-objective optimization system is described.

  • The proposed method produces physically meaningful models without prior knowledge of the system.

  • The method is bench-marked against other modeling techniques.

Abstract

Modern industrial-scale wind turbines are nonlinear systems that operate in turbulent environments. As such, it is difficult to characterize their behavior accurately across a wide range of operating conditions using physically meaningful models. Customarily, the models derived from wind turbine data are in ‘black box’ format, lacking in both conciseness and intelligibility. To address these deficiencies, we use a recently developed symbolic regression method to identify models of a modern horizontal-axis wind turbine in symbolic form. The method uses evolutionary multiobjective optimization to produce succinct dynamic models from operational data while making minimal assumptions about the physical properties of the system. We compare the models produced by this method to models derived by other methods according to their estimation capacity and evaluate the trade-off between model intelligibility and accuracy. Several succinct models are found that predict wind turbine behavior as well as or better than more complex alternatives derived by other methods. We interpret the new models to show that they often contain intelligible estimates of real process physics.

Introduction

As wind energy grows across the globe and new offshore wind turbine installations encounter new operating environments, the models that inform the design and control of these multimillion-dollar machines become increasingly important. Typical multimegawatt wind turbines exhibit nonlinear behavior and are subject to wind (and sometimes wave) disturbances that are often hard to estimate. These properties make the simulation of their dynamics not only challenging but also site-dependent, because of the influence of wind, wave, and foundation characteristics. Accordingly, the first-principles models of wind turbines, such as the one embedded in the aero-hydro-elastic simulation tool FAST [1], are prone to cumulative discrepancies between prediction and reality. These models are also computationally expensive to run because of their fairly comprehensive representation of wind turbine dynamics. Although the use of engineering models is fundamental to the structural design and loads analysis process, model-based controllers preferably rely on a customized model of the real system in the field, rather than a first-principles model that may miss key elements present in the real system [2].

As an alternative to potentially inaccurate and computationally expensive first-principles models, empirical models of wind turbines are obtained from experimental data to provide a customized representation of the wind turbine. These models are usually in the form of auto-regressive moving-average (ARMAX) models [2], [3], [4], [5], neural networks [6], or fuzzy logic models [7], among others, to provide the structural flexibility for adapting the model according to the measured observations. Although these empirical models provide an effective means of estimation/prediction, they have the major drawback of lacking transparency about the physics of the process [8]. This lack of transparency obscures the knowledge of the process that is gained through their development. Ideally, the model should not only be accurate, but intelligible so that the user acquires the insight attained through the model's development. A well-formed model serves two purposes: (i) it improves knowledge of the underlying dynamics of the system; and (ii) it improves the ability of the wind turbine controller to extract power and minimize loads on the turbine.

In order to improve the intelligibility of adapted models, empirical models in the form of symbolic equations can be formulated by symbolic regression [9], [10]. In symbolic regression, the process variables, inputs, and parameters (constants) are treated as symbols and integrated as blocks to form candidate model structures. Free of restrictions from the form (structure), the search is typically conducted by genetic programming (GP) for candidate models having the best-fit outputs to the measured observations [9]. However, in the absence of a presumed model structure and guided only by the prediction error (i.e., the difference between the modeled and measured outputs), symbolic regression often yields illegible, albeit accurate, models that do not convey any of the physics of the process. The method proposed for modeling here safeguards against this potential shortcoming by two innovations. First, it uses a novel GP method known as epigenetic linear genetic programming (ELGP) that combines the flexibility of stack-based GP representations with an epigenetic encoding to allow for topological search of the candidate model structures, leading to less complex and more accurate results than traditional GP [11], [12]. Second, it uses an evolutionary multiobjective optimization (EMO) framework [13] that includes the complexity of the model as an objective in order to yield accurate models that are as intelligible as possible.

In this paper we evaluate the applicability of the proposed ELGP method in identifying wind turbine models based on experimental data collected in normal closed-loop operation from the three-bladed Controls and Advanced Research Turbine (CART3), a turbine maintained by the National Renewable Energy Laboratory (NREL). The paper is organized as follows. First, we present a brief overview of wind turbine mechanics. We then review previous system identification work. Next, the problem formulation as sought by multiobjective optimization is presented, followed by a description of the proposed ELGP method. We then detail the wind turbine identification procedure and analyze results pertaining to local and global models of the wind turbine. The paper concludes with a discussion of the intelligibility of the identified models as they inform the physics of the process.

Section snippets

Wind turbine mechanics

Identification of wind turbine models is a difficult undertaking because of the many layers of nonlinearity governing their behavior. Moreover, modern horizontal-axis wind turbines (HAWTs) are controlled using variable-speed and variable-blade pitch operation, further complicating the dynamics. Consider for instance the steady-state aerodynamic rotor torque (QR) and thrust (TR) generated by the rotor operating in freestream wind speed V, defined by:QR=12ρπR3Cq(λ,β)V2TR=12ρπR2CT(λ,β)V2where the

Related work

Most system identification attempts at modeling wind turbines have focused on producing linear time-invariant (LTI) models via ARMAX models [2], [4] or modified forms of closed-loop subspace identification (SSID) [3], [5]. Although LTI models seem to be effective in characterizing simulated wind turbine behavior at specific operating wind speeds [2], [4], they provide only localized representation. As a remedy, SSID methods have been extended to account for the time-varying, nonlinear dynamics

Problem statement

The underlying assumption of symbolic regression is that there exists an analytical model of the system that would generate the measured observations y(tk) at the sample times tk = t1,…,tN under the input, u(t), as:y(tk,u)=yˆ(tk,M*(x,u,Θ*))+ν;k=1,,Nwhere yˆ is the modeled output, ν represents measurement noise in y, x = [x1,…,xn]T is the vector of state variables, and M(x,u,Θ) denotes the correct model form embodied by the correct parameter values Θ, written M hereafter for brevity. In the

Proposed method

In symbolic regression, the search for candidate models is conducted by GP, whereby a population of computer programs that produce models of the process are evolved. Mathematical building blocks compose the genotype of each program, which is optimized by an evolutionary algorithm. The optimization process, shown in Fig. 2, starts with randomly constructed programs that are repeatedly assessed for their fitness and selectively recombined and mutated until an adequate solution is produced.

System identification of CART3

The proposed method is evaluated in application to experimental data from the CART3 system. NREL's CART3 is instrumented with numerous sensors to make the identification of various system models possible. The nature of experimental data available from this system, including its instrumentation, data collection procedure, and control system, is described first. We then describe the settings used for ELGP, including general and problem-specific settings, followed by the types of models considered

Results

The performance of the identified local models by ELGP is summarized in Table 2. The results correspond to the final model with the maximum VAF from training. The results indicate that the accuracy of the models of Ω, ω, and P is excellent in all cases except for V¯=18.0m/s, where the response of the system was flat (i.e., var(Ω) < 5e−4), obscuring the dynamics. The DTM model has a better performance for this case. The other two models, MFA and MSS, produce generally accurate outputs, but not

Discussion

The local and global models obtained by ELGP using EMO demonstrate the potential for the succinct resultant models to enhance the understanding of the characteristics observed from a process. We have shown that the models are transparent enough to link process estimation to understandable model components (Eqns. (14), (15), (16)) and accurate enough to shed light on the characteristics of the closed-loop behavior of the system. Furthermore, by studying the archive we are able to see how

Conclusion

In this work we use a novel symbolic regression system in an evolutionary multiobjective optimization framework to identify compact models of a wind turbine from operating data with minor assumptions. The models are not only accurate, but succinct and intuitive, and have been shown to embody process knowledge in several instances. This method of system identification may be a promising middle ground between conducting computationally expensive physics simulations and using black-box models

Acknowledgments

The authors would like to thank Dr. van der Veen for sharing his insights into identification of wind systems. This work is partially supported by the NSF-sponsored IGERT: Offshore Wind Energy Engineering, Environmental Science, and Policy (Grant Number 1068864), as well as Grant Nos. 1017817, 1129139, and 1331283. This work was also supported by the U.S. Department of Energy (DOE) under Contract No. DE-AC36-08GO28308 with the National Renewable Energy Laboratory. Funding for the work was

References (35)

  • Gijs van der Veen et al.

    Global identification of wind turbines using a Hammerstein identification method

    IEEE Trans. Control Syst. Technol.

    (July 2013)
  • Stephen A. Billings

    Nonlinear System Identification: NARMAX Methods in the Time, Frequency, and Spatio-temporal Domains

    (2013)
  • John R. Koza

    Genetic Programming: on the Programming of Computers by Means of Natural Selection

    (1992)
  • William La Cava et al.

    Evolving differential equations with developmental linear genetic programming and epigenetic hill climbing

  • William La Cava et al.

    Genetic programming with epigenetic local search

  • Kalyanmoy Deb

    Multi-objective Optimization Using Evolutionary Algorithms

    (July 2001)
  • J.F. Manwell et al.

    Wind Energy Explained: Theory, Design and Application

    (2009)
  • Cited by (43)

    • A new method for prediction of power coefficient and wake length of a horizontal axis wind turbine based on energy analysis

      2022, Energy Conversion and Management
      Citation Excerpt :

      Ren at al [18] and Corradini et al. [19] obtained their models with power functions of TSRs and four constant variables, while Xia et al. [20] establish the calculation model of Cp with function of pitch angle, TSRs and six constant variables. Shamshirband et al. [21] built the calculation model of Cp with third power of pitch angle and its power function; Cava et al. [22] proposed a model with integral of TSR, and Qiao et al. [23] built a nonlinear model of power coefficient based on wind speed, shaft speed and pitch angle. During the real operation of variable speed wind turbine, especially at the condition of the wind speed is lower than the rated wind speed, in order to obtain the maximum power point tracking [24], the shaft rotational speed of the wind turbine will be controlled by the rotational speed control of the generator to adapt the variable wind speed for keeping the optimal power coefficient Cp [25].

    • Controller design by symbolic regression

      2021, Mechanical Systems and Signal Processing
      Citation Excerpt :

      The particular features of ELGP that benefit the current application are its efficiency and the conciseness of equations it develops compared to GP, as demonstrated in application to numerous system identification problems [23,24]. ELGP has consistently produced smaller models with better estimation performance for systems represented by nonlinear ordinary differential equations, and in system identification of real-world benchmark problems, industrial processes, population diversity models [23], and industrial wind turbines [24]. The mechanism responsible for its superior performance is shown to result from preservation of diversity in the equation forms during optimization, precipitated from the introduction of the epigenetic layer [21,22].

    • Learning human-understandable models for the health assessment of Li-ion batteries via Multi-Objective Genetic Programming

      2019, Engineering Applications of Artificial Intelligence
      Citation Excerpt :

      Section 5 contains the experimental design and discusses the results thus obtained and the conclusions of the paper are presented in Section 6. Although tree-based data structures associated with Genetic Programming-based Learning (GPBL) have been well-studied both in classification problems and in symbolic regression problems (La Cava et al., 2016), there are only a few papers on the use of these structures in dynamic models (Villar et al., 2015). Most of these studies focus on optimizing scalar fitness functions, while the majority of the multi-objective implementations of Genetic Programming (GP) are intended to combine anti-bloat techniques with a numerical fitness function (Poli et al., 2008, 2010).

    • HYGP-MSAM based model for slewing bearing residual useful life prediction

      2019, Measurement: Journal of the International Measurement Confederation
    View all citing articles on Scopus
    View full text