Applications of Learning Vector Quantization to Direct Load Control Curves Classification Systems

• Main Authors: Bi-O Tsai, 蔡碧娥
• Other Authors: Hong-Tzer Yang
• Format: Others
• Language: zh-TW
• Published: 2004
• Online Access: http://ndltd.ncl.edu.tw/handle/23929407347467494437

碩士 === 中原大學 === 電機工程研究所 === 92 === Abstract Starting with time of use policies, Taiwan Power Company (TPC) has adopted various strategies of load management since 1979, including interruptible rates, seasonal rates, central air conditioning duty cycling control, ice storage central air conditioning systems, and paging system in central air conditional duty cycling control programs. Performance of various load management strategies had been satisfied, but in recent years the effectiveness seems saturated with little growth. However, with the power demand ever increasing and supply not easily expanding, systematic in-depth analysis of various load management strategies is significant to further upgrade their performance. As a basis of systematic in-depth analysis of various load management strategies, in this thesis, collected from TPC were the data in 2000 for training and testing an effective classification system of direct load control (DLC) curves. The database includes that for central air conditioning duty cycling control, ice storage central air conditioning systems, and paging system in central air conditional duty cycling control. Diverse statistic analytical numerical rules are combined with the learning vector quantization (LVQ) networks to extract the features of the DLC curves and classify curves accordingly. By dividing the customers into complying with DLC policies or not, results can be provided as the basis of subsequent performance evaluation of the DLC strategies. On the other side, the results can also be used as the references to improve the effectiveness of load management policies and related load factor. To verify the proposed approaches, software of Excel and MATLAB is employed to derive the features and construct the classification system. The testing results reveal that the computer programs of the proposed approaches can be developed easily with high performance. The classification rates of three DLC strategies can reach nearly 90% above, even as high as 100% for several cases. Though the classification performance varies for different DLC curves, the proposed approaches are simple and manpower-saving ones. Keywords: Learning Vector Quantization, Feature Extraction, Direct Load Control