Wind energy integration research generally relies on complex sensors located at remote sites. The procedure for generating high-level synthetic information from databases containing large amounts of low-level data must therefore account for possible sensor failures and imperfect input data. The data input is highly sensitive to data quality. To address this problem, this paper presents an empirical methodology that can efficiently preprocess and filter the raw wind data using only aggregated active power output and the corresponding wind speed values at the wind farm. First, raw wind data properties are analyzed, and all the data are divided into six categories according to their attribute magnitudes from a statistical perspective. Next, the weighted distance, a novel concept of the degree of similarity between the individual objects in the wind database and the local outlier factor (LOF) algorithm, is incorporated to compute the outlier factor of every individual object, and this outlier factor is then used to assess which category an object belongs to. Finally, the methodology was tested successfully on the data collected from a large wind farm in northwest China.

In recent years, the energy storage system (ESS) has been demonstrated to be involved in many aspects of the integration of wind power. For ESS application, ESS allocation of the installation location, power rating, and energy rating is the first concern. Different from previous studies, this study emphasises the significance of the ESS operation in the study of ESS allocation. A bi‐level‐programming‐based model is proposed to take the interaction of allocation and operation into consideration at the same time, with the external level optimising allocation and the internal level optimising operation. The complexity assessment and solution algorithm of the model is also discussed. Next, a genetic numerical algorithm is proposed to solve the bi‐level model. The authors’ results were tested on a modified IEEE 39 bus system and a provincial regional power system to verify both the flexibility and applicability of the proposed model and algorithm. This model is useful for various types of ESS and provides a foundation for ESS application.

Existing machine-learning research on power grids relies on online measurements without missing data. We propose a missing data reconstruction model based on generative adversarial networks to supplement existing methods. This model fits better spatio-temporal data with several improvements over previous approaches. First, the loss function considers both distribution and value differences, leveraging all available information to minimize differences between original and generated data. Then, a deep-learning architecture incorporating convolutional neural layers and nonlocal blocks is developed to extract the spatial-temporal information in electrical feature maps. The proposed method exhibits enhanced credibility by neglecting invalid consecutive data under phasor measurement unit (PMU) failures (proven by attention maps generated in nonlocal blocks), and higher accuracy than existing models for recovering data under random data missing/PMU failure conditions (proven by numerical results). Finally, the proposed data reconstruction model is effectively applied to an online framework for transient stability assessment.

Online transient stability assessment (TSA) is crucial for ensuring the security of modern grids. However, problems with limited sample sizes and data imbalance hinder the performance of data-driven TSA classifiers. Addressing this challenge, this article presents a generative adversarial network (GAN)-based model to generate instability samples. Unlike existing methods, our approach moderately alters the long-tailed distribution of instability moments within the raw dataset, producing a more diverse database for TSA tasks. A convolutional neural network-based supervised model, mapping the relationship between fault-clearance electrical characteristics and instability moments of the power system, is incorporated to drive the GAN model to generate realistic yet rare instability cases. Numerical results have proven the superiority of the proposed model over existing case generation methods in terms of realistic and diverse sample generation. In addition, the effectiveness of the proposed data augmentation scheme is demonstrated for online TSA applications.

Transient stability assessment is examined in a data driven framework. The original transient stability data are embedded into a low-dimensional representation space using a deep belief network (DBN) based nonlinear representation learning method. Specifically, unsupervised pre-training is used to learn the data distribution first, and then the expected classification accuracy (ECA) index is used to fine-tune the parameters of the DBN. The structure of power grid is also considered in the learning process. In the representation space, a simple linear model is utilized to classify the unstable cases from stable ones. The proposed method is demonstrated in a regional power system in central China and gets remarkably better testing results compared with a SVM benchmark. The most unique advantage of the proposed approach is that it can learn high level abstract representations automatically to avoid the potential negligence and mistakes introduced by human feature engineering, hence gain more accurate and robust results.