As a core tool, anomaly detection based on a generative adversarial network (GAN) is showing its powerful potential in protecting the safe and stable operation of industrial control systems (ICS) under the Internet of Things (IoT). However, due to the long-tailed distribution of operating data in ICS, existing GAN-based anomaly detection models are prone to misjudging an unseen marginal sample as an outlier. Moreover, it is difficult to collect abnormal samples from ICS. To solve these challenges, a dual auto-encoder GAN-based anomaly detection model is proposed for the industrial control system, simply called the DAGAN model, to achieve an accurate and efficient anomaly detection without any abnormal sample. First, an “encoder–decoder–encoder” architecture is used to build a dual GAN model for learning the latent data distribution without any anomalous sample. Then, a parameter-free dynamic strategy is proposed to robustly and accurately learn the marginal distribution of the training data through dynamic interaction between two GANs. Finally, based on the learned normal distribution and marginal distribution, an optimized anomaly score is used to measure whether a sample is an outlier, thereby reducing the probability of a marginal sample being misjudged. Extensive experiments on multiple datasets demonstrate the advantages of our DAGAN model.

Clustering-based reactive voltage partitioning is successful in reducing grid cascading faults, by using clustering methods to categorize different power-consuming entities in the power grid into distinct regions. In reality, each power-consuming entity has different electrical characteristics. Additionally, due to the irregular and uneven distribution of the population, the distribution of electricity consumption is also irregular and uneven. However, the existing method neglects the electrical difference among each entity and the irregular and uneven density distribution of electricity consumption, resulting in poor accuracy and adaptability of these methods. To address these problems, an enhanced density peak model-based power grid reactive voltage partitioning method is proposed in this paper, called EDPVP. First, the power grid is modeled as a weighted reactive network to consider entity electrical differences. Second, the novel local density and density following distance are designed to enhance the density peak model to address the problem that the traditional density peak model cannot adapt to weighted networks. Finally, the enhanced density peak model is further equipped with an optimized cluster centers selection strategy and an updated remaining node assignment strategy, to better identify irregular and uneven density distribution of electricity consumption, and to achieve fast and accurate reactive voltage partition. Experiments on two real power grids demonstrate the effectiveness of the EDPVP.

Edge embedding is a technique for constructing low-dimensional feature vectors of edges in heterogeneous graphs, which are also called heterogeneous information networks (HINs). However, edge embedding research is still in its early stages, and few well-developed models exist. Moreover, existing models often learn features on the edge graph, which is much larger than the original network, resulting in slower speed and inaccurate performance. To address these issues, a multi-view learning-based fast edge embedding model is developed for HINs in this paper, called MVFEE. Based on the “divide and conquer” strategy, our model divides the global feature learning into multiple separate local intra-view features learning and inter-view features learning processes. More specifically, each vertex type in the edge graph (each edge type in HIN) is first treated as a view, and a private skip-gram model is used to rapidly learn the intra-view features. Then, a cross-view learning strategy is designed to further learn the inter-view features between two views. Finally, a multi-head attention mechanism is used to aggregate these local features to generate accurate global features of each edge. Extensive experiments on four datasets and three network analysis tasks show the advantages of our model.