With the rapid integration of distributed renewable generations and the evolving complexity of load operation patterns, anomalies in grid currents are increasingly manifesting as steady-state harmonics or transient fluctuations. Traditional power quality analysis techniques, such as FFT and wavelet transform, identify anomalous voltage and current by decomposing signals and comparing them to predefined standards. Some recent approaches also apply deep learning, framing power quality issues as supervised classification tasks. Being different, this paper proposes a novel data-driven method that not only detects anomalous current behavior but also uncovers the underlying anomalous components of current, without relying on fixed standards or classification labels. By leveraging the power of Generative Adversarial Networks (GANs), the proposed method learns the distribution patterns of current data in an unsupervised manner. The approach is demonstrated on real-world current data from a transformer, showcasing its practical potential for addressing emerging challenges in modern power grid.

The fundamental cause of severe output fluctuations in photovoltaic (PV) power plants is the abrupt change in ground irradiance due to cloud cover. To enhance the accuracy of PV power forecasting under cloud cover conditions, this article proposes a second level power forecasting method for PV power plants based on an improved Pix2PixHD image restoration algorithm. First, the PV plant model is constructed based on the actual layout by applying Newton Raphson method. Second, the data characteristics of PV power output from the inverters are deeply explored to analyze the mapping relationship between PV power and irradiance. A virtual cloud image is constructed to represent the cloud cover (power loss) situation by describing the shape, thickness, and movement direction of the clouds. Subsequently, the virtual cloud images were preprocessed using the Canny edge detector, followed by restoration of the processed defective virtual cloud images using the Improved Pix2PixHD image restoration algorithm. Finally, a high-precision PV power forecasting at second level is achieved based on the linear relationship between PV power, irradiance, and pixel values of the virtual cloud image.

This article proposes a data-driven online stability monitoring method using real-time output currents. It contributes to stability judgement for grid-following inverters in weak grid. The mainstream approach relies on impedance measurement, which has a tradeoff between disturbance injection magnitude and duration and measurement accuracy. In contrast, the proposed approach, which is the first to use artificial intelligence technology for stability monitoring of inverters based on real-time current data, enables rapid, and accurate stability assessment without requiring disturbance injection, thus preserving normal inverter operation. The single variate of output currents is skillfully expanded to incorporate multiple features and a novel multiaspect feature fusion (MAFF) model is designed to extract these expanded features. Experimental results demonstrate that the proposed method achieves a sample loading interval of 10 ms, an average classification accuracy of 98%, and an average alarm delay of less than 27 ms. Furthermore, the trained MAFF model shows strong adaptability in wideband oscillation monitoring and can effectively accommodate grid-following inverters with different parameters.

Energy disaggregation is a pivotal task in non-intrusive load monitoring, involving the separation of individual appliance contributions from aggregated energy consumption. At present, deep neural networks are extensively employed for the resolution of this problem, eliciting salutary effects. Unfortunately, this resolution demands a wealth of computational and storage resources. Therefore, energy disaggregation models and schemes characterized by low demands and high performance are anticipated. In this article, we propose a novel lightweight energy disaggregation model by incorporating residual learning and deep separable convolutions while achieving comparable performance to state-of-the-art models. Furthermore, an efficient segmented prediction scheme is proposed to reduce the execution frequency in model applications and meet the real-time requirements of nonintrusive load monitoring. The experimental results on publicly available datasets demonstrate that the proposed method reduces computation by 99.17% compared to state-of-the-art models, while the average of mean absolute error for all appliances increases by only 0.727 W.

Photovoltaic power generation systems, as a mature renewable energy technology, may suffer from faults during operation, which could significantly impact system performance. Therefore, it is crucial to develop efficient and reliable fault detection methods to improve the security and stability of the system when it is working. The timely identification and diagnosis of faults can lead to decreased maintenance expenses while enhancing both the stability and efficiency of the system. However, the multitude of data types collectible in current photovoltaic power generation systems makes it challenging for traditional algorithms to acquire features closely associated with faults. Therefore, accurately detecting the types of faults present in the system poses a challenge. To this end, we introduce an attention mechanism to construct a fault detection model aimed at extracting effective dependency relationships among features. Subsequently, the occurring faults are predicted by a fully connected neural network using the extracted feature relationships. Various experimental results on available datasets confirm that the proposed fault detection model based on attention mechanism not only accurately predicts fault types but also exhibits stable performance on small sample and class-imbalanced data.

Considering the nonlinearity, large fluctuations, and rich frequency components of industrial loads, this paper proposes a short-term industrial load forecasting model. The model is based on Variational Mode Decomposition (VMD), a Fine-Grained Feature Encoding Network (FFEN), and an improved Transformer algorithm (ITR). Firstly, considering the characteristics of industrial load, VMD is selected to decompose the original data into multiple modal components. Subsequently, FFEN is established to extract features from each modal component effectively. Then, ITR is employed to predict each modal component. Integration of these components yields more accurate short-term industrial load predictions. Finally, comparative experiments validate the effectiveness of the proposed model, and the results show that the proposed VMD-FFEN-ITR prediction model has high fitting ability and high prediction accuracy and is an effective method for short-term power load prediction.

Modern power conversion systems (PCSs) have been evolving for over more than half a century by gradually enhancing their digital computing, communication, memory, and sensing functions. However, it is evident that state-of-the-art PCSs have largely overlooked the enticing potential of fully exploring intelligent and extensive interconnections among PCSs and even with other heterogeneous devices. This oversight can be attributed to the fact that digital control systems in PCSs have been designed exclusively for real-time control due to the limited resources of industry microcontrollers. With the rapid development of industry microcontrollers in terms of computing and communication capabilities, this article therefore introduces a novel Internet of Things (IoT) architecture, specifically tailored for PCS, named as power conversion IoT (PC-IoT). This article systematically introduces the architecture, key technologies, and prospects of PC-IoT. The designed PC-IoT offers flexibility with timely information transfer, forming the foundation for empowering the intelligent next-generation PCS. In pursuit of this objective, a multistep transmission-based method for converter time synchronization is proposed. Subsequently, a sequential transmission protocol is developed, enabling multiple converters to send data without conflicts and with low latency in a predefined order. Furthermore, to alleviate communication burden and enhance the system's scalability, we propose a fast data processing method that can be implemented in resource-constrained microcontrollers by employing multiple small-scale models. Finally, the article provides a promising outlook for future applications within the PC-IoT framework. Through experimental validation, the proposed transmission protocol has demonstrated exceptional results, including node-to-node latencies below 3 ms and effective data processing at resource-constrained microcontrollers.

As the installed capacity of renewable energy continues to increase, the grid is facing an increasingly prominent problem in terms of power quality. In particular, the power quality caused by high-impedance weak grid is more severe. Online adaptive parameters adjustment strategy can effectively improve the output quality of the grid-connected converters under time-varying grid. Adopting a fast and effective adaptive parameters adjustment method is essential to improve the output quality of the converters. This paper proposes a new data-driven online adaptive parameters adjustment method, which considers the parameters adjustment task under time-varying grid impedance as a multi-classification problem. This method uses the sequential impedance model and particle swarm optimization (PSO) to solve the optimal control parameters under time-varying grid as supervised learning labels. Subsequently, multiple-channel convolutional neural networks (MCCNN) are introduced to fully mine the temporal features of the instantaneous current data output by the converters and improve the accuracy of adaptive parameter adjustment. Finally, this method is validated through a simulation model. The empirical results demonstrate that the delay time of the proposed method from the change of grid inductance to parameter adjustment is less than 10ms, and the classification accuracy is about 96.33%.

This article illustrates a reactive power coupling interaction between high voltage direct current (HVDC) transmission system and photovoltaic (PV) station with an improved PV station reactive power profile model. Different from traditional generic model where PV station is simplified into a one-step shaped switching power source, the developed PV station reactive power profile model illustrates detailed outline of plant's output reactive power, which considers distinctions among PV power units in the station. An analysis of the proposed power coupling interaction is presented. The simulation is performed on IEEE 39 power system model integrated with an HVDC system and a PV station. The result demonstrates that the proposed interaction mode impacts and may even reroute failure path of cascading failure, confirming the necessity to bring in-station dynamics into power system failure study.

Event-related potentials (ERPs) especially P300 are popular effective features for brain-computer interface (BCI) systems based on electroencephalography (EEG). Traditional ERP-based BCI systems may perform poorly for small training samples, i.e. the undersampling problem. In this study, the ERP classification problem was investigated, in particular, the ERP classification in the high-dimensional setting with the number of features larger than the number of samples was studied. A flexible group sparse discriminative analysis algorithm based on Moreau-Yosida regularization was proposed for alleviating the undersampling problem. An optimization problem with the group sparse criterion was presented, and the optimal solution was proposed by using the regularized optimal scoring method. During the alternating iteration procedure, the feature selection and classification were performed simultaneously. Two P300-based BCI datasets were used to evaluate our proposed new method and compare it with existing standard methods. The experimental results indicated that the features extracted via our proposed method are efficient and provide an overall better P300 classification accuracy compared with several state-of-the-art methods.

This paper proposed a unified power compensation (UPC) system by using a cascaded inverter plus DC/DC boost converters as the interfacing circuit between the high voltage level power grid and the battery storage systems in order to both compensate the real power fluctuation produced from the renewable energy generation systems and provide the reactive power compensation capability without shortening the battery life cycle when needed. The proposed system can manage the state-of-charges (SOCs) of batteries to be equal to avoid the overcharge or overdischarge of single battery stack. Meanwhile, the dc-link voltages of all H-bridge modules assumed in the cascaded inverter can be easily controlled to be equal also under the condition of producing/absorbing real or reactive power. The multifold operation will assist the renewable energy generation systems to be the grid friendly power sources.