An LSTM based method for stage performance degradation early warning with consideration of time-series information

Abstract

Renewable energy accommodation in power grids leads to frequent load changes in power plants. Therefore, an efficient monitoring method is necessary to increase the operational reliability of thermal power plants. As extant methods are only efficient under the stable state and perform poorly in the dynamic operation process, this paper proposes a novel method that comprehensively considers the dynamic properties and feature selection to achieve a sensitive and accurate early warning for stage performance degradation. First, the core demands for an ideal early warning are investigated through an analysis of the traditional method. Based on the findings, a frequent pattern model-based early warning method is proposed. Considering the stage characteristics, the features are determined by the fusion of data and mechanism analysis, and the corresponding model is established using a long short-term memory (LSTM) network. The feasibility and validity of this method are experimentally verified and its detection accuracy exceeds 99%. Furthermore, comparison experiments are conducted from the perspective of model characterization and feature selection. The results highlight the importance of time-series information, given that the features exhibit time-dependent characteristics. Moreover, additional features are not necessarily advantageous and a reasonable balance between effective information input and interference is crucial.