Beijing Automotive Group (China)

Integrating hybrid energy storage systems (HESSs) into wave energy converters (WECs) can mitigate power fluctuations of WECs across multiple timescales, provided that an effective energy management strategy (EMS) is implemented. Model predictive control (MPC) is the mainstream EMS for HESSs, as it typically yields a control solution close to the global optimum while satisfying constraints. However, MPC faces a high computational burden when the optimal control problem is nonlinear. More importantly, considering multiple competing objectives, tuning weighting factors (WFs) in the MPC cost function is also challenging. To tackle these challenges, this article proposes a learning-augmented MPC strategy to optimize energy management for the HESS in WECs. The strategy first utilizes a fuzzy logic-based asymmetric action trimming technique to reduce MPC computational time. Further, a warm-start Q-learning (QL) framework with high learning efficiency is applied to obtain the WF online tuning method. To bridge the simulation-to-reality gap in the QL framework, the article designs a neural network-based current predictor, aiming to sense the nonlinear power loss during power conversion. Finally, simulations and experiments demonstrate the superior performance of the proposed strategy in reducing energy loss, battery degradation, and MPC computational burden.

This study proposes an AI-powered intelligent property management system, forming a closed-loop management framework comprising three core modules: “AI Work Order Agent – Robotic Collaborative Inspection – Dynamic Dispatch Optimization.” The system integrates multi-modal work order generation to address the adaptation issues in domain-specific voice interactions, overcomes the limitations of complex environment monitoring via an air-ground collaborative inspection architecture, and utilizes reinforcement learning models to jointly optimize dispatch efficiency and fairness. Empirical results demonstrate that the voice-based work order system significantly enhances the efficiency and accuracy of order generation; multi-modal inspection greatly improves the timeliness of anomaly detection; and the dynamic dispatch model shortens response time while balancing employee workload. Under extreme stress testing, the system exhibits strong robustness and adaptability. Its deployment in a mid-sized community has achieved practical benefits including labor cost savings and reduced training costs, offering an innovative technological path for smart property service construction.

For photovoltaic-rich areas such as sandy wastelands and deep seas, photovoltaic DC collection schemes can effectively solve the problems of large line losses and low system efficiency in traditional AC schemes. It is an effective solution for efficient photovoltaic collection and long-distance transmission. This article focuses on the core equipment of the DC collection scheme, the DC transformer, and conducts relevant technical research. Firstly, the working principle of the LC resonant DC transformer is introduced. Then, the parameter design methods of its core components, including isolation transformers, support capacitors, resonant circuits, are theoretically analyzed. Finally, based on circuit simulation software, the theoretical calculation results are simulated and verified. The simulation results are basically consistent with the theoretical calculation values, verifying the rationality of the parameter design, and providing a basis for the selection and design of subsequent large-capacity DC transformer devices.

A machine learning-based enterprise tax risk warning model is constructed in the paper, which comprehensively utilizes supervised learning and unsupervised learning methods to realize the identification and classification of risk samples. Based on data collection and preprocessing, the model combines feature engineering and PCA dimensionality reduction technology, takes logistic regression, random forest, XGBoost, support vector machine and deep neural network as the framework of supervised model, and introduces K-means, DBSCAN, GMM and self-encoder to perform unlabeled data clustering analysis. The experimental results show that the deep neural network has the best performance in the classification task with 93.6% accuracy and 95.4% AUC value, followed by XGBoost with 92.4% accuracy, while the auto-encoder has 82.6% cluster purity in the unsupervised task, which is better than K-means and DBSCAN.The analysis concludes that the constructed model combines both high efficiency and stability, and has application value for corporate tax The analysis concludes that the constructed model is both efficient and stable, and has application value for enterprise tax risk identification.