Elsevier

Applied Soft Computing

Volume 50, January 2017, Pages 212-222
Applied Soft Computing

Automatic modeling of a gas turbine using genetic programming: An experimental study

https://doi.org/10.1016/j.asoc.2016.11.019Get rights and content

Highlights

  • The behavior of a real gas turbine is modeled using GP.

  • Several state-of-the-art GP algorithms are compared.

  • Results show that standard GP with local search outperforms recent variants.

Abstract

This work deals with the analysis and prediction of the behavior of a gas turbine (GT), the Mitsubishi single shaft Turbo-Generator Model MS6001, which has a 30 MW generation capacity. GTs such as this are of great importance in industry, as drivers of gas compressors for power generation. Because of their complexity and their execution environment, the failure rate of GTs can be high with severe consequences. These units are subjected to transient operations due to starts, load changes and sudden stops that degrade the system over time. To better understand the dynamic behavior of the turbine and to mitigate the aforementioned problems, these transient conditions need to be analyzed and predicted. In the absence of a thermodynamic mathematical model, other approaches should be considered to construct representative models that can be used for condition monitoring of the GT, to predict its behavior and detect possible system malfunctions. One way to derive such models is to use data-driven approaches based on machine learning and artificial intelligence. This work studies the use of state-of-the-art genetic programming (GP) methods to model the Mitsubishi single shaft Turbo-Generator. In particular, we evaluate and compare variants of GP and geometric semantic GP (GSGP) to build models that predict the fuel flow of the unit and the exhaust gas temperature. Results show that an algorithm, proposed by the authors, that integrates a local search mechanism with GP (GP-LS) outperforms all other state-of-the-art variants studied here on both problems, using real-world and representative data recorded during normal system operation. Moreover, results show that GP-LS outperforms seven other modeling techniques, including neural networks and isotonic regression, confirming the importance of GP-based algorithms in this domain.

Introduction

Gas turbines (GTs) are one of the key elements in the electrical power industry since they transform mechanical power into the desired electrical output. In particular, GTs are widely used in so-called Open Cycle Plants and Combined Cycle Plants. They are considered to be high-risk units since they operate at high speed, temperature and pressure, conditions that impose strict control requirements. The control system executes the startup and shutdown sequences, controls speed, power and temperature, while also performing functions related to safety and protection. Startup and shutdown sequences consist of a series of complex events where each step has to be satisfactorily completed before continuing to the next. Moreover, in central power generation it is of utmost importance to ensure high levels of availability and reliability of the generating units, particularly due to the high costs associated to system failure or an interruption in the availability of electrical power. Therefore, condition monitoring of GTs within power plants is highly desirable to guarantee the long-term safety of the units and their efficient operation [1].

For instance, a condition monitoring system can be connected in parallel to the control system for the early detection of potential abnormalities. This can allow supervisors to perform an orderly shutdown of the unit, thus avoiding unexpected stops or failures of a greater magnitude [1]. An important element of such a monitoring system is the model that is used to describe or predict the behavior of the unit. If the models are sufficiently accurate, they can be used to detect unwanted deviations from what is considered to be normal behavior. Indeed, many different physical models of GTs have been developed in the past [2]. These models capture the dynamics of the turbine, but their precision can significantly vary based on the specific unit that is used. In practice, it is not always possible to have complete mathematical models that describe all of the turbine behaviors accurately. Therefore, another alternative is to develop statistical or computational models that might provide a more accurate characterization of the behavior of such units.

Several authors have developed dynamic mathematical and computational models, as well as specialized test equipment for GTs [3]. The modeling of GTs for diagnostics and prognostics has been carried out with a variety of techniques, which can be categorized into two major groups: physics-based methods and data-driven methods. The physics-based methods are limited to the cases where failure mechanisms can be quantified [3], and require specialized personnel to aid in the modeling process as well as a detailed description of the particular system to be modeled with specific technical information from the manufacturer of the equipment, which are often not available. The data-driven methods attempt to derive the models directly from recorded system data without digging into the physics of the system, using tools derived from artificial intelligence, machine learning, statistical analysis and computational intelligence [3]. For instance, a popular approach is to use neural networks [4], [5].

This work proposes the use of genetic programming (GP) [6], [7], [8] to model a Mitsubishi single shaft Turbo-Generator, Model MS6001, with a 30 MW generation capacity. GP is a data-driven approach, a form of evolutionary algorithm (EA) [8] intended for automatic program or model induction. GP has recently gained popularity given its ability to generate syntactic models in a wide variety of problem domains [9], in many cases outperforming more common machine learning paradigms [10], [11], [12], [13] while using very little in the way of prior knowledge. While EAs have been used before in GT systems, they are mostly used for optimizing or improving system efficiency using previously derived models [14]. To the extent of our knowledge, this is the first work to address the GT modeling problem directly using a GP approach. Moreover, this work compares recent proposals in GP literature, including the fairly recent geometric semantic GP (GSGP) [15], [16] and improved variants of standard GP [17], [18]. The experimental work compares several variants of these methods, particularly emphasizing algorithms that combine the global search process provided by the basic GP paradigm with greedy local search operators. Such combinations have been shown to improve algorithm convergence and predictive accuracy [19], while also reducing model size and complexity [17], [18]. Comparisons are also made with seven non-GP modeling methods, showing that GP search outperforms all other methods in terms of predictive accuracy and model size, particularly on the most difficult modeling task.

The remainder of this paper proceeds as follows. Section 2 describes the power generation process of GTs in general and the particular unit studied in this work, describing how data was collected to apply the data-driven approach. Section 3 provides a general overview of GP, focusing on the specialized variants employed in this work and developed by the authors. The experimental work and results are presented in Section 4. Finally, conclusions and future work are outlined in Section 5.

Section snippets

Gas turbine operation and case study

This section provides a brief introduction to GTs and how they are used for power generation, as well as a detailed description of how the dataset used in this study was extracted.

Genetic programming

GP is part of the larger research area known as evolutionary computation, which deals with the development of global search and optimization algorithms that are designed based on some of the core principles of neo-Darwinian evolutionary theory [6], [7], [8]. However, the GP paradigm distinguishes itself from other EAs in several key respects. GP is intended to solve problems that can be broadly defined as automatic program induction, using a supervised learning methodology. In other words, the

Experiments

This work focuses on applying GP to model a GT unit using a data-driven approach following a supervised learning formulation. Therefore, we will use the datasets described in Section 2 to pose two symbolic regression problems and solve them with the variants of GP described in Section 3. Hereafter, we will refer to the problem of modeling the fuel flow of the GT unit as the Fuel problem, and the problem of modeling the exhaust gas temperature as the Temperature problem.

Five different algorithms

Conclusions and future work

This paper addresses the problem of modeling key aspects of a GT power system, namely the fuel flow of the system and the temperature of the exhaust gas. Both variables are critical to the proper functioning and performance of the GT power generator. In particular, this paper describes a data-driven approach to generate the models using GP, a paradigm of evolutionary computation for automatic program induction. Recent state-of-the-art variants of GP are tested. These variants integrate a

Acknowledgements

This research was partially supported by CONACYT Basic Science Research Project No. 178323, TecNM (México) Research Project 5621.15-P, as well as by FP7-Marie Curie-IRSES 2013 European Commission program with project ACoBSEC with contract No. 612689. Third and fourth authors are supported by CONACYT doctoral scholarships, Nos. 294213 and 302526, respectively. The authors also thank Ing. Sergio Ruíz González (Superintendente General) and Ing. Genero Mena Ramírez (Supervisor de Instrumentación y

References (43)

  • U. Sarwar et al.

    Time series method for machine performance prediction using condition monitoring data

  • H.R. DePold et al.

    The application of expert systems and neural networks to gas turbine prognostics and diagnostics

    J. Eng. Gas Turbines Power

    (1999)
  • M. Delgadillo et al.

    Modelling and dynamic simulation of gas turbine

  • J.R. Koza

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

    (1992)
  • W.B. Langdon et al.

    Foundations of Genetic Programming

    (2002)
  • R. Poli, W.B. Langdon, N.F. McPhee, A Field Guide to Genetic Programming, Published via http://lulu.com and freely...
  • J.R. Koza

    Human-competitive results produced by genetic programming

    Genet. Program. Evol. Mach.

    (2010)
  • M. Castelli et al.

    Energy consumption forecasting using semantic-based genetic programming with local search optimizer

    Intell. Neurosci.

    (2015)
  • M. Castelli et al.

    Prediction of relative position of CT slices using a computational intelligence system

    Appl. Soft Comput.

    (2016)
  • A. Moraglio et al.

    Geometric semantic genetic programming

  • L. Vanneschi et al.

    A survey of semantic methods in genetic programming

    Genet. Program. Evol. Mach.

    (2014)
  • Cited by (25)

    • Ammonia-hydrogen-air gas turbine cycle and control analyses

      2022, International Journal of Hydrogen Energy
      Citation Excerpt :

      Nevertheless, those numerical models are neither specialized for NH3–H2/air nor CH4/air gas turbines. Similarly, although the literature [29–32] is enriched with simulation and numerical modeling methods (such as neural network, genetic algorithm, and T-S fuzzy modeling), the challenge of developing a system-based controller, which is adaptable to the distinctive parameters of NH3–H2/air or CH4/air gas turbines, remains significant. Therefore, this paper presents a ‘lightweight’ code (settling time could be reduced to 1s) that serves as a controller for the NH3–H2/air and CH4/air gas turbines, to be utilized for experimental and industrial setups.

    • Real-time novelty detection of an industrial gas turbine using performance deviation model and extreme function theory

      2021, Measurement: Journal of the International Measurement Confederation
      Citation Excerpt :

      In addition, the nonlinear physics also demands high computational cost, which makes physical-based model unsuitable for real-time monitoring. With large datasets acquired from the health monitoring systems, data-driven models draw more attention recently and are commonly constructed through machine learning methods like autoassociative neural networks [19], nonlinear autoregressive exogenous [20], genetic programming [21], convolutional neural network [22] among others. A data-driven model could learn the relationships between the variables of a system directly without a priori knowledge of the GT specification or struggling with the complicated dynamic equations.

    • Fuzzy modeling and fast model predictive control of gas turbine system

      2020, Energy
      Citation Excerpt :

      Furthermore, the gas turbine model was identified by the state-of-the-cut genetic programming method. Unlike other machine learning paradigms, this method used the evolutionary algorithm to determine the model parameters automatically [27]. However, the introduction of genetic algorithm made the modeling process easily fall into local optimum which would lead to the deviation of parameters.

    • Novel fuzzy modeling and energy-saving predictive control of coordinated control system in 1000 MW ultra-supercritical unit

      2019, ISA Transactions
      Citation Excerpt :

      The nonlinearity, coupling and uncertainty of complex system are sometimes ignored in the conventional modeling methods [6–9]. To overcome this deficiency, a variety of intelligent methods have emerged and made great strides, such as fuzzy identification [10–12], neural network [13,14], support vector machines [15], swarm intelligence algorithm [16,17] and so on. Among them, fuzzy identification proposed by Takagi–Sugeno had shown its exceptional places of high precision and universality [18].

    View all citing articles on Scopus
    View full text