Energy Informatics最新文献

The rapid advancement of smart grids necessitates robust dynamic assessment of wireless communication link quality, which faces dual challenges: complex electromagnetic interference (EMI) and the need for effective multi-source temporal data correlation modeling. Traditional methods relying on manual expertise and existing data-driven models often inadequately capture intricate multi-source temporal features. To address these limitations, this paper proposes a novel contrastive learning-based model for wireless link quality assessment in smart grids. Our framework employs Link Quality Indicator (LQI), Received Signal Strength Indicator (RSSI), and Signal-to-Noise Ratio (SNR) as multi-view inputs. A cross-view semantic alignment strategy is introduced to extract noise-robust shared features across these heterogeneous indicators. Furthermore, we design a hybrid attention temporal encoder integrating Long Short-Term Memory (LSTM) networks, adaptive channel attention, and global temporal attention modules. This cascaded architecture achieves deep fusion of local dynamic feature enhancement and global long-range dependency modeling. Experimental validation on 48 hours of continuously collected real-world communication link data demonstrates that the proposed model outperforms baseline methods, achieving accuracy improvements of 2.5% to 7.7% with validated statistical significance. Specifically, for abnormal link states, the model maintains a high recall rate of over 92.1%, ensuring reliable fault detection. While maintaining high overall stability, we observe minor performance degradation under conditions of extreme burst noise or high rates of missing data. Crucially, it exhibits substantially enhanced robustness and generalization capability, particularly in identifying abnormal link states under challenging EMI conditions.

Accurate short-term wind power prediction (WPP) is critical for power system stability but remains challenging due to the inherent non-linearity and volatility of wind series. This study proposes a novel framework, SVMD-RIME-TCN-BiGRU, to address these challenges. First, the Maximal Information Coefficient (MIC) is used to select high-correlation features and eliminate redundancy. Second, Successive Variational Mode Decomposition (SVMD) decomposes raw data into successive intrinsic modes, effectively mitigating non-stationarity and avoiding the mode-mixing issues of traditional methods. Third, a hybrid Temporal Convolutional Network-Bidirectional Gated Recurrent Unit (TCN-BiGRU) model is constructed to extract spatiotemporal features. Crucially, the RIME optimization algorithm is introduced to automatically tune the key hyperparameters of the TCN-BiGRU, avoiding local optima. Experimental results on a Xinjiang wind farm dataset demonstrate that the proposed model achieves a Root Mean Square Error (RMSE) of 7.6882 and an R² of 0.9813. It significantly outperforms baseline models (including LSTM, TCN, and Transformer) and other hybrid variants, reducing errors by over 38% compared to the TCN-BiGRU baseline. This validates the framework’s reliability and accuracy for practical power dispatching.

Accurate solar power generation forecasting is essential for grid stability and renewable energy integration. This paper presents an enhanced solar power forecasting system achieving 94.95% accuracy ((hbox {R}^{2})) using a voting ensemble approach combined with physics-informed feature engineering. The methodology transforms 21 meteorological variables from the Kaggle Solar Energy Power Generation Dataset into 41 engineered features incorporating solar geometry, atmospheric physics, and temporal dynamics. The proposed voting ensemble combines Gradient Boosting Regressor, LightGBM, and XGBoost through simple averaging, achieving (hbox {R}^{2}) = 0.949, RMSE = 214.8 kW, and MAE = 127.7 kW with only 142.4 seconds training time. Experimental validation on 4,213 observations demonstrates superior performance compared to individual models, positioning the system within 3.05% of the target 98% accuracy threshold while maintaining exceptional computational efficiency for real-time deployment.

Substation engineering drawing parsing is essential for the automation, intelligence, and digital transformation of power systems. However, existing methods face significant challenges due to the complexity of these drawings and the limited availability of bitmap datasets. The drawings contain dense lines, specialized symbols, and intricate layouts, making it difficult for traditional object detection models to accurately identify components and text, often resulting in high rates of false positives and false negatives. Additionally, the lack of unified data standards leads to overfitting during model training, limiting generalization across diverse scenarios. To address these issues, we propose an integrated framework combining object detection and OCR for intelligent substation drawing analysis. Our method employs a dense random data augmentation matching strategy and an improved semantic alignment strategy to enhance feature robustness and model adaptability while maintaining computational efficiency. We also introduce a new dataset of 600 annotated substation engineering drawing images, covering various layout types and textual elements. Our experimental results show that our proposed method significantly outperforms existing techniques, achieving AP50, AP75, and mAP scores of 89.40%, 89.30%, and 63.00%, respectively. This demonstrates the effectiveness of our approach in accurately parsing complex substation drawings and contributes to the advancement of power systems’ automation and digital transformation.

The rapid growth of electric vehicle adoption presents new challenges for distribution system operators and charging station owners. Distribution system operators must correctly manage charging demand, optimize infrastructure usage, and support grid stability, while charging station owners seek to analyze user behavior and predict future demand to improve operational efficiency. Meeting these goals requires accurate analysis and forecasting of charging behaviors. A major bottleneck, however, lies in the heterogeneity of available electric vehicle charging datasets: each dataset comes with its own structure, quality issues, and missing information, requiring time-consuming and error-prone preprocessing before any analysis or forecasting can be performed. To overcome this limitation, we introduce EV-Insights, an open-source framework designed to provide standardized services for data ingestion, preprocessing, analysis, and forecasting by supporting the integration of real-time data, synthetic data, and public datasets. Once data is integrated, users can easily generate insights on charging behavior and extend the framework with new analyses or forecasting models through modular interfaces. We evaluated EV-Insights using seven real-world public datasets comprising over 3 million charging sessions, demonstrating its potential to uncover valuable insights and support informed decision-making. Ev-Insights is available as open source at https://github.com/EV-Insights

As power systems move toward digitalization and low-carbon transformation, improving the intelligence of substation design processes has become increasingly critical. Traditional equipment parameter selection relies heavily on manual experience and fragmented document retrieval, leading to inefficiencies, inconsistencies, and limited scalability. This paper proposes an intelligent parameter recommendation method tailored for substation engineering, integrating domain-specific knowledge graphs with adaptive graph neural networks (GNNs). The framework first extracts structured equipment information from multi-voltage substation design drawings using entity disambiguation, then constructs a hierarchical knowledge graph to represent inter-device relationships. A natural language interface captures user queries and encodes them into context-aware instruction vectors. These are used to guide a hybrid reasoning process that combines fuzzy rule matching and GNN-based relation inference. Case studies using real-world 10 kV/110 kV substation projects demonstrate that the proposed method significantly outperforms existing knowledge graph-based baselines in both accuracy and interpretability. The results show that this work is superior to the comparison baseline model in both ACC and AUC indicators, and support intelligent decision-making throughout the equipment lifecycle. This work provides a scalable solution for knowledge-driven substation design automation in the era of smart grids.

This paper presents a novel hybrid machine learning method for enhanced islanding detection in distributed generation systems. The proposed approach integrates distinct feature extraction from voltage and current signals at the point of common coupling with an optimized Extreme Gradient Boosting (XGBoost) classifier to accurately differentiate islanding events from normal grid disturbances. Validated on a public residential microgrid dataset, the method demonstrates superior performance by achieving a high detection accuracy of 97.83%, effectively eliminating the non-detection zone, and maintaining a detection time of under two cycles. This approach provides a robust, non-intrusive, and computationally efficient solution for anti-islanding protection, significantly outperforming conventional passive techniques, while the use of a publicly available dataset ensures full reproducibility and offers a benchmark for future research.

Access to reliable and affordable energy is essential for manufacturing facilities, where even brief disruptions can result in significant financial losses. Solar-plus-storage microgrids offer a promising solution by enhancing energy resilience and reducing costs. However, the design and implementation of such systems remain challenging, especially for non-expert users such as building owners and facility managers. This study evaluates the usability and design effectiveness of two widely used microgrid planning tools, System Advisor Model (SAM) and HOMER Grid, through the lens of a novice user. Drawing on energy consumption data from a real manufacturing facility, two use cases are developed and modeled using both tools. The assessment framework is based on the IEEE Standard 1061 for Software Quality Metrics Methodology, emphasizing usability, system functionality, and component selection support. Both tools generate feasible system configurations but lack intuitive, step-by-step workflows and provide limited guidance on selecting photovoltaic modules, battery chemistries, and inverters. SAM provides more detailed modeling support, but its complexity can hinder adoption by non-technical users. We propose improvements to enhance accessibility, including guided design templates, time-aligned system sizing, prioritized input visualization, and integrated component recommendations. These enhancements could be implemented via SAM’s open-source development kit. The findings highlight the importance of usability and decision-support features in accelerating microgrid adoption, particularly for practitioners and facility managers without specialized expertise. Improving accessibility of these tools can directly support faster deployment of renewable microgrids, enhance resilience for commercial and industrial facilities, and contribute to broader decarbonization goals.

This paper proposes a Hybrid Stackelberg-Markov framework for adaptive load scheduling and dynamic pricing in smart grids. The framework integrates a Stackelberg game to model the interaction between the utility and consumers with a Markov process that captures consumer behavioral dynamics. By combining economic incentives with behavioral adaptation, the model achieves a balance between reducing the peak-to-average ratio (PAR), lowering consumer costs, and increasing utility profit. Simulation results demonstrate that the proposed approach reduces PAR by 43% compared with the baseline, decreases average consumer costs by 28%, and improves utility profit by 10%. The behavioral state analysis further shows that most consumers transition into the Content state, indicating long-term acceptance of dynamic pricing strategies. Moreover, the computational analysis confirms faster convergence and reduced run time compared with conventional demand response schemes. These results establish the proposed framework as a scalable and practical demand response solution for modern smart grids.