572 / 2019-01-18 13:42:14

Typical Condition Data Selection for Data-driven Model Development and Application in Power Industry

operating data; power plant; typical condition library; prediction; failure detection

全体主题 > 操作和维护

The development of modern industrial Information technology makes data resources become more and more important production factors. Based on power generation operating data and modern intelligent technology, models can be constructed to predict and estimate power generation process parameters, thus providing a basis for condition monitoring of equipment and operation optimization of power plants. However, the prediction accuracy of the model has a strong dependence on the operating condition characteristics of the selected data samples when utilizing the operating data to construct the parameter-prediction model. The distribution of the plant operating data has characteristics of being messy and random, which brings difficulties to develop data-driven models. To solve the problem, this paper proposed a method of establishing a typical operating condition library, which only contain most informative data samples. The collect such samples, the information index of operating data is firstly proposed with consideration of the factors including the variation span, distribution status, and redundancy; then the discrete particle swam optimization (DPSO) is used to search the samples that have the largest information index value. The proposed method is validated by two cases, i.e., the prediction of the NOx emission of the SCR system, and the early failure detection of an induced draft fan system. Results show that the samples in the typical condition library can produce accurate prediction models, which are beneficial to the parameter estimation and failure detection.