Data inheritance–based updating method and its application in transient frequency prediction for a power system

Abstract

Summary Transient frequency prediction helps formulate and take emergency control measures in a timely way, which is of great significance to the security and stability of a power system. Conventional frequency prediction methods encounter challenges when taking both speed and accuracy into account. Machine–learning‐based prediction methods cannot readily achieve satisfactory prediction accuracy based on historical samples of transient faults; only by updating the prediction model online can these methods accommodate a transient fault, which may occur at any time. Aiming at the online update of the prediction model, the concept of data inheritance is presented and applied to the process of model updating. Based on this update method, a transient frequency prediction method considering model inheritance is proposed. In this prediction method, the historical transient fault samples of the power system are used to train the initial prediction model; in this way, the prediction performance can be quickly improved according to the newly incremented transient fault samples by inheriting the historical prediction model. Validations on the New England 39‐bus and NPCC 140‐bus systems demonstrate the effectiveness of the proposed method.

Recent Developments in Machine Learning for Energy Systems Reliability Management

“…In the literature, ensembles of ELMs are used mostly for stability classification [19], [29], [32], [35], [56], [75], [76], or combined with random vector functional links [31], [33], [77]. It was also used for regression, to predict load stability margins [78], the fault-induced voltage recovery [30], [79], and maximum frequency deviation and time [80].…”

“…Regarding neural network models, in [65], the model is updated with misclassified samples when a certain number of errors occurred. Online sequential ELM models were also proposed, as in [32], [80], because they are fast to train and can be updated regularly. Another approach is to use a recurrent neural network.…”

Generic System Frequency Response Model for Power Grids With Different Generations

“…This means that b J is the frequency droop coefficient of the generator. Furthermore, from (13) and 15, we have…”

“…Although it reduces the computation burden to a certain extent, it still encounters the same problems as the full model time domain simulation method when dealing with large scale power grid. The accuracy of the artificial intelligence method relies on a large amount of measured data, currently making the method difficult to be promoted and applied in real power grids [11]- [13]. The single-machine equivalent model method has the least computation cost among the methods, and is suitable for online analysis [14]- [18].…”

A Transfer and Deep Learning-Based Method for Online Frequency Stability Assessment and Control

“…This method dramatically reduces the complexity of power system models and realizes the fast prediction of dynamic frequency. However, the prediction error is inevitable due to neglecting the impact of load and system topology on dynamic frequency [6]. In [7], an analytical frequency nadir prediction model is proposed to predict frequency nadir and time when it reaches.…”