Elsevier

Energy

Volume 239, Part B, 15 January 2022, 122089
Energy

Wind power conversion system model identification using adaptive neuro-fuzzy inference systems: A case study

https://doi.org/10.1016/j.energy.2021.122089Get rights and content

Highlights

  • Original ANFIS approach was developed for wind turbine (WT) output power prediction.

  • Scenarios-based validation with real data from wind farm enhanced model building.

  • Machine learning embedded fuzzy logic boosted the model identification strategy.

  • Benchmarking of existing models corroborated the superiority of preprocessed ANFIS.

  • Neuro-fuzzy tool accurately estimated WT output power in various climatic conditions.

Abstract

This study proposes an original adaptive neuro-fuzzy inference system modeling approach to predict the output power of a wind turbine. The model's input includes the wind speed, turbine rotational speed, and mechanical-to-electrical power converter's temperature. The structure of the adaptive neuro-fuzzy inference system-based model was first identified using the processed data gathered from wind turbine number 1 of a 30-MW wind farm in Nouakchott (Mauritania). Then, the proposed data-driven model was trained and validated according to two new scenarios based on the data set from four identical wind turbines operated in the same climatic conditions and the data set from the same wind turbines operated under different climatic conditions. Benchmarking involved the proposed model, existing approaches in the literature, and five adaptive neuro-fuzzy inference system-based models, including grid partition, subtractive clustering, fuzzy C-means clustering, genetic algorithm, and particle swarm optimization, on the same data set to validate their prediction performance. Compared with existing adaptive neuro-fuzzy inference system-based models, the proposed approach was proven to be a promising methodology with higher accuracy for estimating the output power of wind turbines operating in different climatic conditions. According to the results from two different scenarios, the lowest value of the fitting rate and the highest values of the normalized mean square error, normalized mean absolute error, and root mean square error for the validating period were 0.9977, 0.0047, 0.0473, and 46.5831 kW, respectively. Moreover, the proposed model showed superior forecasting performance and thus better accuracy in estimating wind power output compared to other adaptive neuro-fuzzy inference system-based models.

Introduction

Wind energy production has been increasing steadily at an annual rate of 31% over the past decade by following an exponential growth [1]. Cost-effective exploitation of wind energy assumes optimal geospatial localization of the wind turbines (WTs) to maximize the production of electrical power. A wind turbine (WT) converts the wind power into electrical power by converting the air mass pressure on the turbine's blades into rotating mechanical power, and then converted into high-rotating mechanical power using a gearbox and a converter of mechanical-to-electrical power. Fig. 1 depicts the conversion process from wind power to electrical power.

Research works for wind energy exploitation can be classified into three categories: (1) wind speed prediction using historical wind speed data measurement; (2) WT power curve or model identification based on various data that may include measurements wind speed, wind direction, ambient temperature, humidity, and so forth; and (3) WT electric power output forecasting model identification, using the meteorological and operation condition data of mechanical-to-electrical power generator. The specific operating parameters (e.g., rotational speed, lubrication, output voltage, output current, and energy loss with warm-up) of the mechanical-to-electrical power converter (MEPC) and its components' under some operating conditions determine the output power performance. Actual operating conditions involve various parameters of natural meteorological influences (e.g., ambient temperature, pressure, humidity) are also harmful to components. All those influences can result in degradation phenomena such as corrosion, blade abrasion, bearing picking, structure fatigue, cause failures, and reduce power conversion performance. Therefore, a wind power conversion system (WPCS) model should accurately predict the electric output power using the most relevant measurable input variables. On the other hand, the prediction of the WT output power is a complex issue which needs to consider additional independent climatic parameters (e.g., temperature, humidity, air density, pressure, solar radiation, aerosols) and WTs state variables (e.g., rotational speed of the MEPC, temperature of the MEPC, network voltage, WT components' condition, and so forth). Hence, it becomes necessary to develop a method that allows predicting WT output power using data on both the WTs condition and its operation environment. Unfortunately, to develop analytical models of such a system with all influent variable parameters is a complicated task, as some parameters are immeasurable. Besides, existing parametric models’ structures are not always well known, which limit their accuracy. So, it is more relevant and realistic to use data-driven non-parametric modeling approaches and artificial intelligence techniques.

Artificial intelligence techniques, such as artificial neural networks (ANNs) [2] and adaptive neuro-fuzzy inference systems (ANFIS) [3] are known for their ability to model complex non-linear systems of which structure model is difficult to identify. In the WPCS, this modeling technique would enable build a model of inference relationships between the input variables characteristic of the wind power including the meteorological condition variables and the power measured at the WPCS output. Model settings will then be identified and adjusted using historical data available and a learning process. From all the data-driven modeling approaches and methods for WPCS output power assessment and prediction, it emerges that in most cases the ANFIS-based models show prominent performance. Besides, few of these works have examined the influence of the type of membership-shape functions and the selection of input parameters on the model's performances. The choice of the fuzzy subsets number, the shape of the membership functions and the input variables are essential to determine the suitable structure for the model and can improve the performance of the model and the accuracy of the prediction significantly [4].

Based on the literature review (see Section 2) of data-driven techniques developed for WPCS output power prediction, few have tried to deepen a specific technique to enhance its performance. Most of researchers suggested new methods combining different techniques, which in some cases may result in limitation due to operation conflict. In our case, we aimed to investigate the ANFIS-based modeling approach for the operation mechanism. Thus, a comprehensive analysis was performed to enhance the modeling performance through the tuning of membership functions and the model's parameter identification strategy based on an iterative refining process considering spatio-temporal meteorological and operating data collected from a 30-MW wind farms. To the best of our knowledge, ANFIS-based modeling approach still has a significant unexplored margin for improvement in the current issue.

Considering the afore-noted scarcity of the literature in this field, the present study aimed at fulfilling the relevant gap by focusing upon a new ANFIS-based modeling approach for identifying the WPCS output power model. To that extend, real data were collected from a wind farm of 30 MW for 10-min increments over a one-year period. This study was carried out using real data from sites located in the regions from Nouakchott to Nouadhibou in Mauritania, one of the African Sahel countries characterized by the highest wind energy potential. The sites encompass a 30-MW wind farm consisting of 15 WTs of 2 MW each and a 100-MW wind farm which development underway in the Northwestern coast of Nouadhibou. In this study, wind speed, rotational speed of the MEPC, and temperature of the MEPC, accounting for more than 70% of the WT output power, were selected as input variables for the WT output power prediction model. These three variables were determined using principal component analysis (PCA) on a set of model variables encompassing wind speed, rotational speed of the MEPC, temperature of the MEPC, ambient temperature, relative humidity, wind direction, air density, pressure, and network voltage. It is noted that the energy lost in the process of conversion from mechanical to electrical power is converted into heat dissipated in the bearings, gears, as well as generator electromagnetic interfaces of the WPCS. Thus, a significant part of these losses was considered in the present model in terms of the temperature measured on the generator bearing. The proposed model was validated using measurements data collected on other WTs of the same characteristics and operating under different conditions. We investigated the effectiveness of the model through benchmarking with existing ANFIS-based models using the same data set.

The content of the present paper is organized as follows: Section 2 summarizes some important previous attempts on modeling of wind power conversion systems. Section 3 recalls the general structures of the WT output power model and the ANFIS models, and then presents the general scheme of the ANFIS network and the learning algorithm used to train this model. Section 4 describes the ANFIS-based model identification approach implemented in this study. Section 5 introduces the statistical performance indicators (normalized mean square error (NMSE), normalized mean absolute error (NMAE), root mean square error (RMSE), and fitting rate (R)) used to corroborate the accuracy of the ANFIS-based model identification. Section 6 provides the WT output model validation using real operating data followed by the ANFIS model identification and implementation. Section 7 discusses the benchmark results for those attained using the proposed approach and other ANFIS-based techniques (e.g., ANFIS-GP, ANFIS-SC, ANFIS-FCM, ANFIS-GA, and ANFIS-PSO). Lastly, Section 8 summarizes the findings obtained within the framework of this study.

Section snippets

Literature review on modeling of wind power conversion systems

To predict the wind speed data series used to evaluate wind potential at specific locations, data-driven models have also been worked out. Khosravi et al. [5] conducted a benchmarking of the effectiveness of three models to predict wind speed, wind direction and then the WT output power. They considered a multilayer feed-forward neural network (MLFFNN), a support vector regression with a radial basis function (SVR- RBF), and an ANFIS optimized by a particle swarm algorithm (ANFIS-PSO), with an

Fuzzification of the WT output power

Developing an ANFIS model for a dynamic process requires fuzzy variables at the model input and output. The fuzzy variables exploit prior knowledge on the system's physical behaviors and by the fuzzification process. The power produced by the WT depends fundamentally on the wind speed and is modeled in several research works [27] as described in Eq. (1).P={Prf(v),vcivivrPr,vrvivco0,0vivciorvivcowhere Pr is the rated output power (W); vci is the cut-in wind speed (m/s); vco is the

ANFIS-based model identification

The prediction model for the WT output power is summarized by the flowchart given in Fig. 4. Preprocessed data from a benchmarked WT called WT01 is used to identify the membership functions and the rules. Next, the model is trained and validated. Thus, the premise and consequent parameters are tuned during every cycle of the training process using a hybrid learning algorithm, which combines the least-squares method and the backpropagation gradient descent method [30] as explained next in this

Model performance indicators

To assess the performance of the proposed model, the performance criteria were NMSE, NMAE, RMSE, and the fitting rate (R) given in Eqs. (28), (29), (30), (31) [2]. These statistical indicators were used for benchmarking the performance of the proposed model with other ANFIS-based WPCS output power model identification from the literature. The smaller values of NMSE, NMAE, and RMSE stand for better accuracy of the model. The fitting rate (R) is the correlation coefficient that expresses the

Results and discussion

The identification of the WT output power model was conducted based on the ANFIS approach by using the IF-THEN logical statement. In this study, ANFIS-based modeling and all procedures were performed using the Fuzzy Rule Editor of MATLAB® software. For this aim, the proposed model was established as a soft-computing tool, rather than a deterministic approach. The process of tuning the model parameters for the ANFIS was developed considering the hybrid learning procedure as described in Sections

Conclusions

An original methodology for predicting the output power of a WT was introduced, based on real operational data (collected from a 30-MW wind farm in Nouakchott (Mauritania)) and a set of important input variables chosen by the PCA method, which accounted for 70% of the significant elements.The most important input factors in this study were the wind speed, MEPC's rotational speed, and MEPC's temperature, while the model was designed to forecast WT output power.With a comprehensive comparison of

Credit author statement

Boudy Bilal (First author): Supervision, Project administration, Funding acquisition, Conceptualization, Methodology, Software, Formal analysis, Investigation, Resources, Data Curation, Writing-Original Draft, Writing-Reviewing and Editing, Visualization, Kondo Hloindo Adjallah: Supervision, Funding acquisition, Conceptualization, Formal analysis, Investigation, Reviewing, Visualization, Alexandre Sava: Methodology, Software, Formal analysis, Investigation, Writing-Reviewing and Editing,

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements & Funding

The authors would like to acknowledge the Ministry of Petroleum, Energy, and Mines, National Industrial and Mining Company of Mauritania and Mauritanian Electricity Company for providing the data used in this study. Also, the authors would like to acknowledge the Islamic Development Bank (IDB) for funding allocated for this research.

References (34)

Cited by (9)

  • Wind turbine output power prediction and optimization based on a novel adaptive neuro-fuzzy inference system with the moving window

    2023, Energy
    Citation Excerpt :

    Further, the proposed model yields the lowest NMAE, between 0.0314 and 0.0375 and between 0.0347 and 0.0557, for the training and validation periods, respectively. In contrast, for the validation period, the NMAE equals 1.31 in Ref. [48] while it lies between 0.0571 and 0.1171 in Ref. [49], and for the training period it equals 0.0401 in Ref. [65] while it lies between 0.0955 and 0.1176 in Ref. [73]. Further, the fitting rate (R) is the highest as, for the training period, this indicator lies between 0.9985 and 0.9988 for the different time series horizons in this study.

  • A novel self-learning framework for fault identification of wind turbine drive bearings

    2023, Proceedings of the Institution of Mechanical Engineers. Part I: Journal of Systems and Control Engineering
View all citing articles on Scopus
View full text