Détection en temps réel de l'instabilité transitoire des réseaux électriques avec les mesures synchronisées de phaseurs

Abstract

Synchrophasors also called Phasor Measurement Units, coupled with a GPS system, measure the phase of the current more than 30 times per second, or a frequency that is well over 120 times that of current systems. They give the possibility of mapping in real time the phases of the electrical network which obviously facilitates their control, monitoring, understanding and optimal management. The applications of synchrophasors include extended control, system model validation, determination of stability margins, load optimization for a stable system, islanding detection, network-wide perturbation recording and visualization of the dynamic of the network. With this technology in perpetual development, the network operators have a very important amount of information to manipulate. The aim of this thesis is to propose to the network operators a systematic approach which allows predicting the instability of the electrical networks in real time by extracting the hidden signatures in the mass of data provided by the synchrophasors. Given that data mining is a set of algorithms leading to construct models from the data by seeking a maximum of concealed knowledge, we exploit in this work the random forest of decision trees and the boosting. We propose a new concept for the study of instability: the Center of Power (COP). This notion is an extension of the center of inertia and the extended area measurement approach and voltage control system. It considers the actual network power values deduced from the phasor measurements. A COP speed is associated with each area. There is an instability in the case where the speed COP of any two areas keeps moving away. Two indices were also proposed for detecting instability based on speed COPs and phasor measurements. The divergence of these indices confirms that the network has lost its synchronism. Unlike snapshot widely used in the literature, it is proposed in this thesis to use a sliding window during the analysis of instability. The latter has the advantages of promoting decision-making at any time without a fault detector and updating the database on the state of the network with the arrival of new data without losing the information contained in the previous data. To evaluate the importance of predictors proposed in the construction of the random forest and the boosting models, three scenarios are considered. In the first, the variables derived from the COP, the proposed indices and all the measured and estimated variables are selected as predictors. Increased relative importance is given to predictors directly related to COP and to phasor measurements in the construction of predictive models of real-time instability. So, in the second scenario, the emphasis is placed solely on the proposed predictors and the phasor measurements. It results that the differences between the accuracy, the reliability and the security of the prediction models obtained in both scenarios are very low. In addition, rates of misclassification and false alarm come closer. It can be deduced that the proposed predictors represent a good compromise between the small loss of precision of the model and the reduction in the size of the database on which the learning is carried out. The third scenario makes it possible to generalize and demonstrate that the proposed predictors do not depend on the network. Finally, a new adaptive approach is proposed to update the prediction model. The state of the network is determined by a double check: a first one based on the model of the random forest initially constructed and the second based on a boosting. At each boosting constructed, an indicator is assigned which depends on the weight of each of the trees of the model and related errors. During the update, this indicator is compared to the previous one and allows or denies the adoption of the new model for predicting the state of the network for future events.

Date	26 Apr 2017
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Louis-A. Dessaint (Supervisor)