以極速學習機建立電力系統之負載動態模型

Dynamical Load Modeling Developments in Power Systems by Extreme Learning Machine

徐書敏(Hsu, Shu Min)

指導教授 ： 朱家齊

國立清華大學/電機資訊學院/電機工程學系/碩士(2015年)

摘要

電力系統各設備元件模型的精確度，對模擬結果能否如實反映系統實際運轉之行為有著重大的影響。其中又以負載的不確定性及複雜性為最高，故建立正確且適當的負載模型，可提升電力系統穩定度模擬的真實性，加強電網分析的可靠度。一般建構負載模型的方法主要分成兩類:監錄量測法及元件聚集法。本研究以監錄量測法建構負載模型，採用極速學習機進行負載模型的識別與驗證。首先，利用極速學習機對輸入輸出資料間優越的映射能力及通用化表現，以變電所實際量測之監錄記錄檔進行負載功率的估測，此方式期望在不需先得知負載模型之架構的情況下，可僅以監錄設備量測之負載電壓、功率等數據，準確地估測出負載功率的動態響應，模擬結果可顯示極速學習機良好的預測能力。本論文亦使用極速學習機對靜態指數負載模型進行參數識別，為充分驗證此方式之參數識別的準確性，利用PSS/E自定義模型功能預設參數值，產生欲識別之事故案例進行負載模型參數估測。比較極速學習機識別之參數值與原先自定義模型之參數預設值，並將參數估測結果代入指數負載模型進行負載功率的驗證，在IEEE 9匯流排系統及IEEE 39匯流排系統之試驗中，皆可獲得理想的識別結果。最後，在應用至實際監錄量測設備得到之事件紀錄檔，進行靜態指數負載模型的參數識別。

關鍵字

負載模型 ； 監錄量測法 ； 極速學習機 ； PSS/E自定義模型 ； 模型參數識別

並列摘要

The accuracy of the model of each equipment component impacts whether the power system simulation results can exactly reflect the operating behavior of the actual system, and the complexity and uncertainty of the load models are the highest among the rest. Adequate modeling of loads which can capture essential dynamic behaviors of the power system can indeed improve the factuality of stability simulations and the reliability of electric network analysis. In the literature, two distinct categories of dynamical load modeling approaches which are commonly used: measurement-based and component based method. This thesis will focus on the measurement-based method. The extreme learning machine (ELM) is adopted for identifying and validating of various dynamical load models. The proposed approach will consist of the following two steps. First, by considering the superior capability of mapping the input-output relationship and generalization performance, the load power was predicted by ELM with measured data. Then, it is expected that the dynamic response of the load power can be accurately estimated without using pre-defined physical load models. In this thesis, ELM was also used for parameter identification of static exponential load model. The contingency events for load model parameter identification were generated with PSS/E user-defined model. The predicted dynamical load models were compared with the default values set in the user-defined model. The predicted values were also used for validating the model of load power. Simulation results of both IEEE 9 bus and IEEE 39 bus system have shown satisfactory results.

並列關鍵字

Load Models ； Measurement-based Method ； Extreme Learning Machine ； PSS/E User-defined Model ； Model Parameter Identification

參考文獻

[1] A. M. Azmy, I. Erlich, and P. Sowa, “Artificial neural network-based dynamic equivalents for distribution systems containing active sources,” IEE Proc.-Gener. Transm. Distrib., Vol. 151, no. 6, pp. 681-688, November 2004.

[2] A. Y. Abdelaziz, M. A. L. Badr, and A. H. Younes, “Dynamic load modeling of an Egyptian primary distribution system using neural networks,” International Journal of Electrical Power and Energy Systems, Vol. 29, pp. 637-649, 2007.

[3] A. Miranian, and K. Rouzbehi, “Nonlinear Power System Load Identification Using Local Model Networks,” IEEE Trans. On Power Systems, Vol. 28, NO. 3, pp. 2872-2881, August 2013.

[4] CIGRE, Modelling and Aggregation of Loads in Flexible Power Networks, WG C4.605, 2013.

[6] C.-Y. Chiou, C.-H. Huang, A.-S. Liu, Y.-T. Chen, T.-H. Li, C.-J. Lin, H.-D. Chiang, and J.-L. Yuan, “Development of a Micro-processor-based Transient Data Recording System for Load Behavior Analysis,” IEEE Transactions on Power Systems, vol. 8, no. 1, pp. 16-22, August 1993.