Abstract：Aiming at the problem of high retrieval difficulty caused by the complexity and diversity of power system data, this paper studies an adaptive retrieval method for large fluctuation data entries in the power system. Based on the rate of change in power system output, a two-component one-dimensional mixture Gaussian model is selected to construct a probability distribution model for power system fluctuations. Compare the power system fluctuation data simulated by the probability distribution model with the measurement data, identify the large fluctuation data entries in the power system based on the judgment threshold, and construct a data entry retrieval library. Use hash functions to obtain hash features of large fluctuation data entries in the retrieval database and generate binary codes. When retrieving large fluctuation data entries, generate binary codes for user search terms, calculate the Hamming distance between the binary codes of search terms and the binary codes of entries in the search database, and weight them. Use the weighted Hamming distance to sort the data entries and obtain adaptive search results for large fluctuation data entries. The experimental results show that this method can adaptively retrieve large fluctuation data entries in the power system based on user input search terms. The normalized cumulative loss gain of the retrieval results is higher than 0.9, and the retrieval time is less than 500 ms.

Abstract：To address the issue of poor recognition performance of existing named entity recognition methods in texts from fields such as electric power safety regulations, this paper introduces a method for named entity recognition in power safety based on machine reading comprehension. Firstly, a pre-trained model is used to encode the text to be recognized to obtain the vector representation of the text. Secondly, a hierarchical attention mechanism is utilized to capture the hierarchical relationships among nested entities and re-allocate the attention weights of the text sequence. On this basis, a classifier is employed to predict the entity scope in the text, and the final entity recognition results are obtained. The method is validated on the ACE 2005 and OntoNotes 4.0 public datasets, achieving optimal recognition performance compared to mainstream approaches. In the context of entity recognition for power safety scenarios, the method attains an accuracy rate of 89.3%, enabling precise identification of named entities in the power safety domain.

Abstract：The dynamic aggregation of commercial load clusters is crucial for enhancing the flexibility of power grid dispatch, optimizing demand-side management, and promoting the integration of renewable energy. This paper selects commercial load features using Canonical Correlation Analysis (CCA) and employs DBSCAN and K-means clustering algorithms to classify loads, forming load clusters suitable for different scenarios. Furthermore, three load aggregation criteria are proposed, namely, regulation speed-based, load stability-based, and economic-based standards. The characteristics, applicability, and potential applications of commercial load aggregation under different standards in power dispatch are analyzed.

Abstract：Electrical fires in high-rise buildings are difficult to predict and can cause significant damage. To address this issue, this paper proposes a multi-modal data fusion model for early detection of electrical fires in high-rise buildings. The model integrates data from three different types of sensors: temperature, CO gas concentration, and smoke, leveraging the complementary advantages of each modality. Initially, the gated Multi-Layer Perceptron(gMLP) is used to capture the intrinsic patterns of the three modalities' data, facilitating feature extraction. Subsequently, a fusion method based on multi-head attention is employed to merge the effective information from different modalities, achieving feature fusion and identifying electrical facilities with potential fire hazards. Experiments conducted on a multi-modal dataset under scenarios of no hazard and various electrical facilities with potential fire hazards demonstrate that the multi-modal data fusion model achieves high accuracy in early prediction, highlighting the superiority of the fusion approach.

Abstract：To facilitate the participation of large-scale, flexible air conditioning loads in demand response programs, enabling "source-load interaction" and ensuring the safe and economical operation of the power grid, various research institutions conducted real-time control simulation and practical studies on building air conditioning demand response. However, accurately estimating and forecasting air conditioning load remains a significant challenge. Current mainstream approaches include model-driven and data-driven methodologies. The model-driven approach relies on air conditioning load modeling, which struggles to capture the complex variations of the load. On the other hand, the data-driven approach depends on extensive data for model training, but often fails to account for the diverse characteristics of air conditioning loads. Therefore, this paper aims to integrate both model-driven and data-driven approaches to intelligently fit air conditioning loads, thereby improving the accuracy and adaptability of air conditioning load forecasting and generation. The paper proposes a load generation method and model AirGAN that combines mechanism models with Generative Adversarial Networks (GANs). This method continuously adjusts the hyperparameters of the physical model to better match the actual air conditioning load characteristics using virtual data generated by the GAN generator. Additionally, the GAN discriminator is employed to evaluate the load predicted by the mechanism model, thereby training the mechanism model to enhance its prediction accuracy.