Low-voltage distribution networks are experiencing increasing phase current unbalance due to the widespread installation of single-phase photovoltaic systems and electric vehicle chargers. These unbalance results in increased power losses, transformer thermal stress, and accelerated asset degradation. This article proposes a decentralized drop control method using three independent single-phase battery energy storage systems mounted on LV transformers. The proposed control scheme regulates the charging and discharging of each BESS based on the deviation between individual phase currents and the average phase current, while explicitly considering the state of charge limitation. This method relies only on local current measurements and does not require centralized communication. The effectiveness of the proposed method was validated using a real residential low-voltage distribution network with high PV and EV usage. Simulation results demonstrate a significant reduction in the current unbalanced factor, improved phase current symmetry, reduced feeder losses, and relief from transformer thermal stress. Technical and economic assessments further demonstrate that the proposed strategy is a more cost-effective alternative to traditional grid strengthening in low-voltage networks.
This research presents a blockchain-based log management system that integrates an LLM-assisted query mechanism to enhance security and provide contextual insights into storage event logs. The modules for decentralized storage event logs, including registered users, log retrieval and storage, and log auditing, are built on a blockchain deployed via an Ethereum smart contract, which enhances the following. EdDSA-based digital signatures (Ed25519) are applied to audit log events to ensure integrity. Additionally, log files are encrypted using asymmetric encryption (Curve25519) to protect confidentiality, while the decentralized IPFS (InterPlanetary File System) stores log file metadata on the blockchain, reducing storage size. In the querying module, an LLM-based natural-language model is integrated. Moreover, we organize the logs into semantic groups and hourly summary tables for LLM-assisted analysis. Users can analyze log data without complex query rules. The experiment covers four evaluations: First, functional correctness. The storage event log modules correctly support secure log management and verify log audit. Second, IPFS delivers practical performance with average upload times (0.41 s) and download times (0.19 s). Third, blockchain operations require reasonable gas usage. Fourth, the Llama 3.1 (8B) achieves the best querying performance, with Correctness (0.80), Faithfulness (0.53), and Helpfulness (0.62) supporting practical log analysis.
A multi half-bridge LLC resonant inverter with three induction coils is used to harden metal workpieces in this research. The system operates at 80–100 kHz and outputs 3 kW of power, based on to load variations. Power is controlled through changing the duty cycle of each inverter module's output voltage waveform and a phase-locked loop for frequency tracking. This method allows soft-switching and improves system efficiency. Stable operation and improved performance demonstrate that proposed system is suitable for induction hardening. The multi-inverter design improves power distribution, system dependability, flexibility, and scalability for many industrial applications.
AI-powered conversational interface designed to enhance industrial data accessibility in Warehouse Management Systems (WMS). The system addresses the challenge of enabling non-technical users to retrieve complex warehouse data without requiring SQL knowledge or database expertise, supporting data-driven decision-making in manufacturing operations. Our approach employs WangchanBERTa, a pre-trained Thai language model, for Natural Language Processing to interpret user intents and extract relevant entities from Thai questions. The system then converts natural language queries into SQL commands using Natural Language to SQL (NL2SQL) techniques to retrieve information from WMS databases. The proposed architecture consists of five key components, and the system's effectiveness is evaluated through implementation in a footwear manufacturing facility, where sample queries regarding inventory, stock levels, and operational metrics are tested. Experimental results show the Intent Classification model achieves a Macro F1-Score of 87.84% and an Accuracy of 94.01%, demonstrating robust performance across multiple intent categories. System generates correct SQL queries for common warehouse queries, including product color, quantity, and stock availability. While certain complex intents exhibit lower accuracy due to contextual ambiguity and limited training data. This research provides a replicable framework to be extended to other industrial domains, thereby enhancing operational efficiency and human-AI collaboration in Industry 5.0.
With the increasing complexity of modern power distribution networks and the rising demand for reliable electricity, accurate load forecasting at the distribution level has become essential for efficient energy management. This study presents a performance-focused analysis of machine learning based short-term load forecasting (STLF) of a substation distribution transformer (SDT). Two STLF methods are proposed: (i) Direct load forecasting method: directly forecasts the SDT load using a dedicated SDT load model; (ii) Feeder aggregation method: forecasts individual feeder load using feeder specific load model and aggregates them with optimized weights to estimate the SDT load. Both methods perform real and reactive power forecasting with a half-hour-ahead prediction horizon and five-minute resolution. The models are trained and tested on data collected from an actual SDT and its connected feeders, ensuring real-world relevance and reliability. The proposed STLF methods offers practical insights for utility operators and system planners by supporting better informed decision-making in short-term load management.
