High-performance predictor for critical unstable generators based on scalable parallelized neural networks

J. Mod. Power Syst. Clean Energy (2016) 4(3):414–426 DOI 10.1007/s40565-016-0209-4 High-performance predictor for critical unstable generators based on scalable parallelized neural networks Youbo LIU1, Yang LIU1, Junyong LIU1, Maozhen LI2, Zhibo MA3, Gareth TAYLOR2 Abstract A high-performance predictor for critical unstable generators (CUGs) of power systems is presented in this paper. The predictor is driven by the MapReduce based parallelized neural networks. Specifically, a group of back propagation neural networks (BPNNs), fed by massive response trajectories data, are efficiently organized and concurrently trained in Hadoop to identify dynamic behavior of individual generator. Rather than simply classifying global stability of power systems, the presented approach is able to distinguish unstable generators accurately with a few cycles of synchronized trajectories after fault clearing, CrossCheck date: 3 June 2016 Received: 30 November 2015 / Accepted: 25 April 2016 / Published online: 14 July 2016  The Author(s) 2016. This article is published with open access at Springerlink.com & Yang LIU Youbo LIU Junyong LIU Maozhen LI Zhibo MA Gareth TAYLOR 1 School of Electrical Engineering and Information, Sichuan University, Chengdu 610065, China 2 Electronic and Computer Engineering, Brunel University, Uxbridge, Middx, London UB8 3PH, UK 3 Senior Power System Engineer, National Grid, Bearwood road, Wokingham RG6 3DU, UK 123 enabling more in-depth emergency awareness based on wide-area implementation. In addition, the technique is of rich scalability due to Hadoop framework, which can be deployed in the control centers as a high-performance computing infrastructure for real-time instability alert. Numerical examples are studied using NPCC 48-machines test system and a realistic power system of China. Keywords Transient stability, Critical unstable generator (CUG), High-performance computing (HPC), MapReduce based parallel BPNN, Hadoop 1 Introduction Transient stability has been widely regarded as one of the most concerned issues of modern power system. In the last two decades, a number of large blackouts occurred all over the world due to the loss of synchronization caused by cascading failures [1]. Insufficient online implementations and lack of timely emergency controls, such as load shedding, generator tripping and proactive islanding, are said to be the common causes of those accidents [2]. The increasing renewable energy integration brings dynamic security deterioration of power systems, which would lead to the operation risks [3]. However, the deployment of phasor measurement units (PMUs) provides a promising way to improve awareness ability of control centers for the disturbed operation scenarios. PMUs, the infrastructure of wide-area monitoring system (WAMS) of power system, is able to measure synchronized phasor data with much higher sampling frequency compared with supervisory control and data acquisition (SCADA) [4]. The measurement accuracy is also reported to be sufficiently satisfactory. Since PMUs can grasp the instant response of power High-performance predictor for critical unstable generators based on scalable parallelized neural… systems when faults occur, how to utilize the massive disturbed trajectories has been significantly investigated in the last decade. As WAMS are now being deployed in quite a few power systems, PMU is playing an ever increasingly vital role in transient stability awareness [5]. A number of researches have been carried out to evaluate the transient stability by using PMU data. PMU trajectories based indicators are considered as the efficient estimators to understand dynamic features of power systems, especially during severe disturbances. For example, Alvarez et al proposed seven trajectory based indices which are suitable for fuzzy inference on real-time dynamic vulnerability [6]. A phasor data–based energy function indicator was developed in [7] aiming at monitoring the dynamic status of power transfer paths. A real-time transient stability assessment (TSA) method based on centre-of-inertia estimation from PMU records was reported in [8]. From voltage stability aspect, a coupled single-port model was applied to establish WAMS based assessment indicator [9]. Furthermore, Makarov et al. [10] presented a review on PMU-based TSA, offering a clear roadmap for further development. Machine learning techniques have been widely applied for TSA. Most of the existing works are focused on the binary state prediction for global stability using clustering and classification. For example, support vector machine, decision tree and artificial neural network (ANN) are widely used to detect instability of power systems by using post-fault dynamic data during a few cycles [11–13]. Guo and Milanović presented a probabilistic framework to evaluate the accuracy of data mining tools applied for online prediction of transient stability [14], enabling the comprehensive analysis of performance of different implementations. However, few machine learning techniques have considered the impact of the critical unstable generators (CUGs) of power systems. The majority of the researches have focused on the identification of the global system status due to the fact that a power system normally has hundreds of generators which generate massive volumes of data [15]. As a result, it has become a challenge for standalone machine learning techniques running on single computers to deal with stability assessment taking into account CUGs identification. For example, Passaro et al. [16] employed adaptive neural network to evaluate stability for each generator, admitting that standalone neural networks can hardly solve the problem in a reasonable time. For this purpose, applying advanced computing techniques to enable high-performance training and prediction associated with PMU measured data has become a necessity. It is well known that neural network is highly adapted to classification tasks [17]. A number of researchers employed neural network to achieve high accuracy classifications in 415 both academia and industrial fields. References [18, 19] figured out that BPNN encounters low efficiency issue due to large number of sum and sigmoid calculations. Some researchers focused on speeding up BPNN using cloud computing techniques. For example, Yuan et al. [20] implemented parallel BPNN using cloud computing technique. Ikram et al. [21] also employed cloud computing to parallelize BPNN in training phase. And also some researchers focused on solving the issue using MPI [22]. However, their ideas are all based on data separation, which does not consider the accuracy loss caused by the simple data separation. Therefore, to improve the efficiency of BPNN whilst maintains classification accuracy in predicting CUGs, this paper presents a MapReduce based parallel back propagation neural network(BPNN) algorithm. The algorithm firstly employs ensemble techni (...truncated)