Impact of wind farms on disturbance detection and classification in distributed generation using modified Adaline network and an adaptive neuro-fuzzy information system

Highlights

• A modified Adaline and ANFIS are used for disturbance detection in distribution generation.

• Training of the Adaline is done using a robust decoupled Gauss–Newton algorithm.

• Impact of wind velocity of the wind farm on islanding and non-islanding cases is studied.

• Power quality indices are used to classify disturbances.

• Comparison with other techniques is shown to validate the superiority.

Abstract

A new disturbance detection and classification technique based on modified Adaline and adaptive neuro-fuzzy information system (ANFIS) is proposed for a distributed generation system comprising a wind power generating system (DFIG) and a photovoltaic array. The proposed technique is based on a fast Gauss–Newton parameter updating rule rather than the conventional Widrow–Hoff delta rule for the Adaline network. The voltage and current signals near the target distributed generation (DG), particularly the DFIG, whose speed varies from minimum to the maximum cut-off speed, are processed through the modified Adaline network to yield the features like the negative sequence power, harmonic amplification factor (HAF), total harmonic distortion (THD), etc. These features are then used as training sets for the ANFIS, which employs a gradient descent algorithm to update its parameters. The proposed technique distinguishes the islanding condition of the distributed generation system with some other disturbances, such as switching faults, capacitor bank switching, voltage swell, voltage sag, distorted grid voltage, unbalanced load switching, etc. which are referred to as non-islanding cases in this paper.

Introduction

Increased demand for electricity has necessitated the larger penetration of renewable energy sources to the power grid. Until the time, DG is supplying generated power to the network loads along with the utility, it is considered to be in a safer condition. On the contrary, if the DG is supplying power to loads (local loads), in the absence of the utility (which is considered as islanding condition according to the IEEE std. 1547.1), several negative impacts are produced on the utility power system as well as the DG. In addition to this, the switching events such as capacitor bank switching, load switching, fault switching, etc., are mostly unintentional and have a great impact on the power systems. Some of the hazards may be related to safety personnel's of the utility and public, quality of power fed to the consumers, damage to the DG if the utility is restored in a wrong manner [1], [2], lack of grounding, change in the fault power, reclosing due to out of phase, voltage and frequency control [1], [3], [4], etc. Therefore it is most important task of any disturbance detection technique, primarily to quickly detect the occurrence of the disturbance and then to correctly classify the nature of the disturbance.

Disturbance (islanding and non-islanding) detection techniques [3], [4], [5], [6], [7], [8], [9], [10] are broadly classified into active and passive methods. Active methods concentrate on the injection of a small disturbance into the power system and then monitoring its response. On the contrary, passive methods concentrate on the measurement of local parameters of the power system and setting some threshold for these parameters based on the measured values. As a matter of fact souvenired from the literature survey, it reveals that on a generalized aim to condense the effect of the non-detection zone, various methods have already been proposed. For instance, some of the latest active islanding detection schemes such as signal injection at elevated frequencies [28], methods based on positive-feedbacks [29], passive Bayesians [30], etc. have proven to have localized the non-detection zones, but only for the inverter based DG's [24]. In contrast, techniques based on fuzzy rule base [20], decision tree [31], and pattern recognition [25], [32] have been incorporated for the synchronous-based DG's. Several passive methods such as voltage/current thresholds [33], rate of change of frequency (ROCOF) [21], [22], [23], rate of change of voltage (ROCOV) [21], [22], [26], change in frequency [20], harmonic assessment based methods [15], [18], etc. can be addressed from the literature, which are very much convenient for utilization in DG networks, but again at a cost of increased non-detection zones.

In recent years, with the development of artificial intelligence techniques, artificial neural networks and Adalines [11], [18], [19] have been used for tracking time varying harmonics buried in noise. The Adaline structure as shown in Fig. 1, also known as a linear combiner, is preferred due to its simple structure and ease of computation in comparison to other known and widely used techniques. Therefore, it is preferred as a real-time computational approach for the estimation of fundamental and harmonic phasors in power networks. However, one of the drawbacks in the earlier least mean square (LMS) based learning algorithms is the tracking error, introduced due to the non-stationary nature of the signal, i.e. sudden changes in frequency, phase and amplitude of the fundamental and harmonic components, and the presence of noise in the power signal. This paper, therefore, presents a modified Adaline structure, where the parameters are estimated by minimizing a weighted squared error cost function. Instead of conventional improved Widrow–Hoff rule, the unknown parameters are updated by a recursive Newton-type algorithm weighted by a variable forgetting factor. The technique proposed in this paper is classified as a new passive islanding detection method based on the modified Adaline [11], that uses a fast Gauss–Newton recursive weight updating rule (GNADA) [27] and adaptive neuro-fuzzy information system (ANFIS). It is to be noted that an analogous Gauss–Newton approach has already been investigated in [27], but suffers from inaccuracies as the fundamental frequency estimation does not take into account a frequency correction factor that is associated with the changing forgetting factor. Therefore, the earlier technique has been modified in the present paper to make the computation of frequency and phasors more accurate and robust for islanding detection in distributed generation systems.

The proposed method is very efficient in accurately classifying the various disturbances such as islanding and non-islanding (i.e. capacitor bank switching, unbalanced load switching, fault switching, etc.), even with a very large variation of wind speed and active power mismatch with a very high classification accuracy as compared to some of the traditional techniques. The paper is organized in ten sections. Apart from the introduction in Section 1, the description of the system under study has been done in Section 2. In Section 3, the proposed hybrid Gauss–Newton linear combiner (HGNLC) algorithm is described for target feature extraction. Section 4 presents the computer simulation results showing clearly the impact of wide range of variation of wind speed and active power mismatch on the various parameters like the HAF_i, HAF_v, THD_i, THD_v, and negative sequence power (NSP) at the target DG terminals, particularly for the DFIG based wind power generating system. Section 5 is dedicated to describe the ANFIS training and testing procedure whereas, Section 6 describes the non-detection zone of the islanding. Section 7 describes the performance evaluation of the proposed method, and the concluding remarks are presented in Section 8.