就高效數據中心與管理資訊設備而言，電源管理策略扮演相當重
要的角色。為了確切管控伺服器電源供應器的實際功耗，精確的電源
供應器特性必頇被偵測，例如輸入與輸出的電壓與電流資訊、工作環
境溫度以及風扇轉速。傳統上，輸出電流偵測是利用分流電阻來實現，
雖然此方法容易設計，但分流電阻會帶來功率損耗。設計上，阻值選
用存在權衡精確度與耗散功率的問題。此外，元件熱偏移的特性亦會
影響其精確度。
本文提出以類神經網路為基礎對伺服器電源供應器作輸出電流
之估測，將電源供應器的輸入電壓、輸入電流、功率因數、輸出電壓
以及工作環境溫度做為類神經網路的輸入數據。為了取得充足的訓練
數據量，利用自行開發之自動化量測帄台，可以有效取得上述輸入資
料。自動化量測帄台由交流電源供應器、功率量測儀、電子負載、資
料擷取裝置及高低溫詴驗機所組成。
根據三組測詴案例的結果，所估測到之輸出電流資訊誤差小於
1.3%。因此，本文提出之技術可以被用來精確的估測伺服器電源供應
器之輸出電流資訊，其優點包含提高電源轉換的效率、減少印刷電路
板的尺寸與降低偵測線路帶來的成本。

Power management strategy is very important in terms of the efficiency of data centers and the availability of information technology (IT) equipment. In order to determine the actual power consumption, accurate measurements of power supply unit’s attributes such as input/output voltage and current levels, temperatures and fan speeds are essential. Conventionally, output current measurement is realized using current shunt. Although it is simple to implement, current shunt dissipates quite an amount of power, and a tradeoff exists between precision and dissipated power. In addition, it suffers from temperature drift problem.

In this thesis, an artificial neural network (ANN)-based algorithm for estimating the output current information of a server power supply is proposed. The input data of the proposed ANN includes input voltage, input current, power factor, output voltage and ambient temperature. To generate a full coverage of the training data, an automatic testing platform (ATP) consists of a programmable power supply, a high accuracy power meter, two electronic loads, one data acquisition system and one thermal chamber is developed. Using the constructed ATP, training data of the proposed ANN can be generated efficiently.

According to the experimental results, the estimation error of the output current is less than 1.3 % in three tested cases. Therefore, the proposed technique can be utilized to estimate the output current of power supply. The advantages of the proposed method include high power efficiency, reduced board size and decreased bill of material cost from sensing circuitry.

參考文獻
[1] 梁適安，第二版，「切換式電源供給器之理論與實務設計」，全華圖書股份有限公司，2011年。
[2] Revision 1.3 of the PMBus specification, 2013.
[3] Jason Starck, Kelvin (4-wire) resistance measurement
[4] Server power management specifications – IBM xSeries system, 2014.
[5] Maxim Application Note 1108: Understanding Single-Ended, Pseudo-Differential and Fully-Differential ADC Inputs – Retrieved November 10, 2007
[6] 葉怡成，「類神經網路模式應用與實作」，儒林圖書出版，2003。
[7] M. M. Gupta, L. Jin, and N.Homma, “Static and Dynamic Neural Networks”, Wiley, 2003.
[8] Generalized Test Protocol for Calculating the Energy Efficiency of Internal Ac-Dc and Dc-Dc Power Supplies, 2014.
[9] 致茂電子，http://www.chroma.com.tw.
[10] Yokogawa Electric Corporation，http://www.yokogawa.com/index.
[11] Keysight Technologies，https://www.keysight.com.
[12] Thermotron Inc.，http://thermotron.com.
[13] 陳樹學、劉萱，「LabVIEW寶典」，電子工業出版社，2011年。
[14] 羅華強, 「類神經網路-MATLAB的應用」,高立圖書出版, 2001。
[15] 朱凱、王正林，「精通 MATLAB 神經網絡」，電子工業出版,2010
[16] Claesen, Marc; Bart De Moor (2015). "Hyperparameter Search in Machine Learning". arXiv:1502.02127.
[17] Bergstra, James; Bengio, Yoshua (2012). "Random Search for Hyper-Parameter Optimization". J. Machine Learning Research. 13: 281–305.
[18] A Server System Infrastructure (SSI) Specification_PSDG
[19] 黃崇哲，「用於相移全橋轉換器輕載效率提升之可變怠滯時間控制策略」，國立中山大學，中華民國 103 年七月。
[20] 黃建霖，「高效率數位控制全橋轉換器之研製」，國立台灣科技大學，中華民國 104 年七月。
[21] 陳重諺，「以交流阻抗為基礎之鋰離子電池殘餘容量估測技術研究」，國立台灣科技大學，中華民國 103 年二月。
[22] 黃貞翔，「利用LabVIEW實現切換式電源供應器之自動化效率量測平台」，國立台灣科技大學，中華民國 105 年一月。