應用自我組織網路於直接負載控制績效評估之研究

A Study On Performance Evaluation of Direct Load Using Self-Organizing Maps

蔡宛妮(Wan-Ni Tsai)

指導教授 ： 楊宏澤

中原大學/工學院/電機工程研究所/碩士(2002年)

https://doi.org/10.6840/cycu200200082

摘要

中文摘要
由於經濟快速成長與生活水準提高，以致使夏季尖峰時期冷氣負載急遽增加，因而造成負載因素持續惡化，需求面管理乃成為電力公司營運的重要策略，以有效利用現有發電設備並緩和電源開發之壓力。然而，近年來各項負載管理措施實施成效趨於平緩，進行各項負載管理措施特別是空調直接負載控制之績效評估，即成為電力公司據以參考改善相關負載管理策略與訂定未來負載管理目標值之重要手段。
為了評估各空調直接負載控制措施之實施績效，本文建構一自動分類直接負載控制曲線之系統，以鑑別實施效果不彰與實施狀況良好之負載。對於實施狀況良好之用戶而言，經由本文所擬定之績效評估模式，估算其抑低負載量與轉移用電量；對於實施效果不彰之用戶則予以剔除。為有效達成直接負載控制曲線之分類目標，首要工作在於負載特徵之擷取。是故，本研究利用非線性主要成份分析與週期性分析兩種特徵擷取方法，擷取各負載之重要特徵，其中非線性主要成份分析法乃應用對偶式類神經網路技巧擷取其負載特徵，而週期性分析則擷取其週期性特徵，以供各負載分類之重要依據。
本文利用自我組織映射網路建構直接負載控制曲線之分類系統，其特性在於將輸入資料以投影方法轉換至低維度群聚空間上，所產生的拓撲結構可充分反應輸入資料之分佈型態。另外，為瞭解該資料之最佳分類數目，本文利用分類適確性Davies Bouldin(DB)指標以指出該資料最佳之分類群聚。透過上述之分類系統，可將實施狀況良好之用戶加以篩選以進行績效評估，供電力公司改善直接負載控制策略之成效參考。
為驗證本文所提直接負載控制曲線分類系統與績效評估之可行性，本研究引用89年台電夏季實施中央空調負載管理措施資料做為測試資料，其中包含中央空調週期性暫停用電、中央空調遙控降載週期性暫停用電與儲冷式空調系統。同時，為測試本分類系統鑑別實施不彰之用戶準確性與有效性，本文亦採用主要成份特徵擷取方法與倒傳遞類神經網路進行比較，以觀察本文提出分類系統在實測資料上的成效。本文利用Matlab與Borland C++ Builder軟體撰寫分類系統與績效評估模式之計算機程式，以驗證所提方法之準確性與效果。經由測試結果顯示，本文所提出之方法可正確分類各措施之負載控制實施成效，並可進一步藉由績效評估模式，獲得較合理與準確之績效值，提供未來擬訂各項直接負載控制措施目標值之參考。

關鍵字

週期性分析 ； Davies Bouldin分類適確性指標 ； 空調負載直接負載控制 ； 自我組織映射網路 ； 非線性主要成份分析

並列摘要

Owing to rapid growth of economy and living standards on Taiwan, the summer peak loads, mainly attributed to the air-conditioner loads, have increased drastically faster than the average loads. As a result, the load factors become worse and worse, a fact that makes the demand side management (DSM) evitable for the utilities. The DSM provides solutions to efficient utilization of existing power apparatus and reduction of the pressure of installing new one for the utilities. However, in recent years the effectiveness of DSM strategies is slowing down. Performance valuation of the existing DSM strategies, particularly for the air-conditioner direct load control, would be an important approach to further improving the DSM strategies and determining the targets for the future DSM.
To evaluate the effectiveness of the direct load control (DLC) programs, this thesis establishes an automatic classification system of the DLC curves. Through the system, identified are either the DLC curve that complies with the control model of a specific DLC program or the other one that does not. For the one complying with the DLC model, a simple program can then be used to evaluate the performance of the DLC program, i.e., the amounts of load reduction or shifting of the customers that commit the DLC programs. The load curve that does not comply with the DLC model would be excluded from the database and not taken into account in the performance evaluation. Nevertheless, to achieve the above purpose of DCL curve classification, the significant features of load curves must be first extracted. In the thesis, the significant features are extracted by using the nonlinear principal component analysis (NLPCA) and periodic analysis methods proposed. The nonlinear feature extraction method of NLPCA is implemented by dual multi-layer neural networks model. The periodic proposed analysis method is utilized to extract the features by analyzing the cyclic nature of the DLC load curves. The extracted features are then taken as the basis of the classification system of the DLC curves.
In accordance with the features extracted, the thesis presents an efficient approach to clustering the DLC curves by using the self-organizing maps (SOM) networks in the classification system. The unsupervised learning based SOM creates a set of prototype vectors representing the template data sets and carries out a topology preserving projection of the prototypes from multidimensional input space onto a low-dimensional grid. In the SOM, to achieve better classification, the Davies Bouldin (DB) index is employed to determine the best number of the clusters. Therefore, the complying DLC models identified by the classification system are then used to evaluate the performance of the DLC programs for the utilities.
To verify the proposed approaches, the practical DLC database in Year 1999 obtained from Taiwan Power Company (TPC) are employed. The database includes the data for central air conditioning duty cycling control, paging system in central air conditioner duty cycling control, and ice storage central air conditioning systems. To demonstrate the efficiency and effectiveness of the proposed approach for the practical data evaluations, comparisons with the methods using existing principal component and artificial neural networks methods in the classification system are presented. MATLAB and Borland C++ Builder implement the programs of the DLC classification system in the thesis. The testing results reveal that a very low degree of misclassification for the TPC DLC database has been obtained. Based on the classification results, the performance evaluation model can, therefore, estimate the effectiveness of the DLC programs more reliably and accurately for the future target settlement of the DLC strategies of the utilities.

並列關鍵字

Periodical analysis ； Davies Bouldin Index of Cluster Validity ； Direct Load Control ； Self-Organizing Maps ； Nonlinear Principal Component Analysis

參考文獻

[1] M. Delgando and F. Rey, “Demand-side Management Alternatives,” Proceedings of the IEEE, Vol. 73, no. 10, pp. 1471-1488, Oct.1985.

[2] F. Kreith, “Integrated Resource Planning,” Journal of Energy Resources Technology, Vol. 115, pp 80-85, June 1993.

[3] S. Rehman and K. Rinaldy, “An Efficient Load Model for Analyzing Demand Side Management Impact,” IEEE Trans. on Power Systems, Vol.8, no.3, pp.1219-1227, August 1993.

[11] S. Majumdar, D. Chattopadhyay, and I. Parikh, “Interruptible Load Management Using Optimal Power Flow Analysis,” IEEE Trans. on Power Systems, Vol.11, no.2, pp.715-720, May 1996.

[12] M. W. Gusftafson, J. S. Baylor and G. Epstein, “Estimating Air Conditioning Load Control Effectiveness Using An Engineering Model,” IEEE Trans. on Power Systems, Vol.8, no.3, pp.972-978, August 1993.