翻譯-人工神經(jīng)網(wǎng)絡(luò)在短期負(fù)荷預(yù)測(cè)中的應(yīng)用

1、目 錄外文文獻(xiàn)翻譯11 緒論12 各種影響負(fù)荷預(yù)測(cè)的因素23 混合神經(jīng)網(wǎng)絡(luò)33.1 線性神經(jīng)網(wǎng)絡(luò)33.2 非線性神經(jīng)網(wǎng)絡(luò)44 神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)的確定54.1 自動(dòng)校正54.2 遺傳算法75 短期負(fù)荷預(yù)測(cè)系統(tǒng)76 仿真結(jié)果97 優(yōu)化處理107.1 基于規(guī)則系統(tǒng)107.2 模式識(shí)別系統(tǒng)10結(jié)論11外文文獻(xiàn)原文121.introduction122.variables afferting short-term load143. hybrid neurak networks153.1 linear neutal networks153.2 non-linear neural networks164. de

2、termination of network structure174.1 autocorrelation184.2 genetic algorithm195. short term load forecasting system206. simulation result217.enhancement227.1 rule-based system237.2 pattern recognition system23conclusion24外文文獻(xiàn)翻譯人工神經(jīng)網(wǎng)絡(luò)在短期負(fù)荷預(yù)測(cè)中的應(yīng)用摘要：在本文，我們將討論如何利用人工神經(jīng)網(wǎng)絡(luò)對(duì)短期負(fù)荷進(jìn)行預(yù)測(cè)。在這類系統(tǒng)中，有兩種類型的神經(jīng)網(wǎng)絡(luò)：非線性和線性

3、神經(jīng)網(wǎng)絡(luò)。非線性神經(jīng)網(wǎng)絡(luò)是用來捕獲負(fù)荷和各種輸入?yún)?shù)之間的高度非線性關(guān)系。基于arma模型的神經(jīng)網(wǎng)絡(luò)，主要用來捕捉很短的時(shí)間期限內(nèi)負(fù)載的變化。我們的系統(tǒng)可以實(shí)現(xiàn)準(zhǔn)確性高的短期負(fù)荷預(yù)測(cè)。關(guān)鍵詞：短期負(fù)荷預(yù)測(cè)，人工神經(jīng)網(wǎng)絡(luò)1 緒論短期（每小時(shí)）負(fù)荷預(yù)測(cè)對(duì)于電力系統(tǒng)的穩(wěn)定運(yùn)行是必要的。準(zhǔn)確的負(fù)荷預(yù)測(cè)對(duì)于高效的發(fā)電調(diào)度，開停機(jī)計(jì)劃，需求方的管理，短時(shí)維護(hù)安排或其他目的等是很必要的。改進(jìn)短期負(fù)荷預(yù)測(cè)的準(zhǔn)確性能為公共事業(yè)和聯(lián)合發(fā)電節(jié)省很多開支。很多種電力系統(tǒng)負(fù)荷預(yù)測(cè)方法在學(xué)術(shù)界已經(jīng)報(bào)導(dǎo)了。這些方法包括：多元線性回歸法，時(shí)間序列法，一般指數(shù)平滑法，卡爾曼濾波法，專家系統(tǒng)法和人工神經(jīng)網(wǎng)絡(luò)預(yù)測(cè)法。由于電力負(fù)荷和

4、各種參數(shù)（天氣的溫度，濕度，風(fēng)速等）之間的高度非線性的關(guān)系，無論在電力負(fù)荷預(yù)測(cè)建模或在預(yù)測(cè)中都有重要的作用。人工神經(jīng)網(wǎng)絡(luò)就是這種具有潛力的非線性技術(shù)的代表，但是由于電力系統(tǒng)的復(fù)雜性，神經(jīng)網(wǎng)絡(luò)的規(guī)模會(huì)較大，所以，當(dāng)終端用戶每天甚至每小時(shí)都在改變系統(tǒng)的運(yùn)行時(shí)，訓(xùn)練這個(gè)網(wǎng)絡(luò)將是一個(gè)重大的問題。在本文中，我們把這網(wǎng)絡(luò)看作是建立在負(fù)荷預(yù)測(cè)系統(tǒng)上的混合神經(jīng)網(wǎng)絡(luò)。這類網(wǎng)絡(luò)中包含兩類網(wǎng)絡(luò)：非線性神經(jīng)網(wǎng)絡(luò)和線性神經(jīng)網(wǎng)絡(luò)。非線性神經(jīng)網(wǎng)絡(luò)常用來捕獲負(fù)荷與各種輸入?yún)?shù)（如歷史負(fù)荷值、氣象溫度、相關(guān)濕度等）間的高度非線性關(guān)系。我們常用線性神經(jīng)網(wǎng)絡(luò)來建立arma模型。這種基于arma模型的神經(jīng)網(wǎng)絡(luò)主要用來捕獲負(fù)荷在很短時(shí)

5、間期限內(nèi)的變化。最終的負(fù)荷預(yù)測(cè)系統(tǒng)是兩種神經(jīng)網(wǎng)絡(luò)的組合。要用大量的歷史數(shù)據(jù)來訓(xùn)練神經(jīng)網(wǎng)絡(luò)，以減小平均絕對(duì)誤差百分比 (mape)。一種改進(jìn)的反向傳播學(xué)習(xí)算法已經(jīng)用來訓(xùn)練非線性神經(jīng)網(wǎng)絡(luò)。我們使用widrow -霍夫算法訓(xùn)練線性神經(jīng)網(wǎng)絡(luò)。當(dāng)網(wǎng)絡(luò)結(jié)構(gòu)越簡(jiǎn)單，那整個(gè)系統(tǒng)的訓(xùn)練也就越快。為了說明這個(gè)基于實(shí)際情況的負(fù)荷預(yù)測(cè)系統(tǒng)的神經(jīng)網(wǎng)絡(luò)的性能，我們采用一個(gè)公共機(jī)構(gòu)提供的實(shí)際需求數(shù)據(jù)來訓(xùn)練系統(tǒng)，利用三年（1989，1990，1991）中每小時(shí)的數(shù)據(jù)來訓(xùn)練這個(gè)神經(jīng)網(wǎng)絡(luò)，用1992年每小時(shí)的實(shí)際需求數(shù)據(jù)用來驗(yàn)證整個(gè)系統(tǒng)。這文章內(nèi)容安排如下：第一部分介紹本文內(nèi)容；第二部分描述了影響負(fù)荷預(yù)測(cè)結(jié)果的因素；第三部分介紹

6、了混合神經(jīng)網(wǎng)絡(luò)在系統(tǒng)中的應(yīng)用；第四部分描述了找到最初網(wǎng)絡(luò)結(jié)構(gòu)的方法。第五部分詳細(xì)介紹了負(fù)荷預(yù)測(cè)系統(tǒng)；第六部分給出了一些仿真結(jié)果；最后，第七部分介紹了系統(tǒng)的優(yōu)化處理。2 各種影響負(fù)荷預(yù)測(cè)的因素以下是一些影響負(fù)荷預(yù)測(cè)的因素：溫度濕度風(fēng)速云層日照時(shí)間地理區(qū)域假期經(jīng)濟(jì)因素顯然，這些因素的影響程度取決于負(fù)荷的類型。例如：溫度變化對(duì)民用和商業(yè)負(fù)荷的影響大于它對(duì)工業(yè)負(fù)荷的影響。相對(duì)較多民用負(fù)荷的區(qū)域的短期負(fù)荷受氣候條件影響程度大于工業(yè)負(fù)荷較多的區(qū)域。但是，工業(yè)區(qū)域?qū)τ诮?jīng)濟(jì)因素較為敏感，如假期。如下一個(gè)例子，圖2.1表示了午夜開始的一天中負(fù)荷的變化。圖2.1 一天中負(fù)荷變化的示例3 混合神經(jīng)網(wǎng)絡(luò)我們所研究的負(fù)

7、荷預(yù)測(cè)系統(tǒng)由兩類網(wǎng)絡(luò)組成：arma模型的線性神經(jīng)網(wǎng)絡(luò)和前饋非線性神經(jīng)網(wǎng)絡(luò)。非線性神經(jīng)網(wǎng)絡(luò)常用來捕獲負(fù)荷與各種輸入?yún)?shù)間的高度非線性關(guān)系。我們常用線性神經(jīng)網(wǎng)絡(luò)來建立arma模型，這種基于arma模型的神經(jīng)網(wǎng)絡(luò)主要用來捕獲負(fù)荷在很短時(shí)間期限（一個(gè)小時(shí)）內(nèi)的變化。3.1 線性神經(jīng)網(wǎng)絡(luò)一般的多元線性的調(diào)整參數(shù)p和獨(dú)立變量x的關(guān)系是：其中： -時(shí)刻的電力負(fù)荷 -時(shí)刻的獨(dú)立變量 -時(shí)刻的隨機(jī)干擾量 -系數(shù)線性神經(jīng)網(wǎng)絡(luò)能成功地學(xué)習(xí)歷史負(fù)荷數(shù)據(jù)和獨(dú)立變量中的系數(shù)和，widrow-hoff已經(jīng)決定了這些系數(shù)。這個(gè)模型包括了先前所以數(shù)據(jù)高達(dá)p的延遲，如上所示，這些數(shù)據(jù)不是獨(dú)立的，它與負(fù)荷有不用程度的相關(guān)性。相關(guān)性

8、學(xué)習(xí)用來決定模型中包含的最重要的參數(shù)，決定了許多參數(shù)會(huì)被去掉。這樣就減少了給定精度模型的大小和運(yùn)算時(shí)間或是提高了給定規(guī)模大小的模型的精度。3.2 非線性神經(jīng)網(wǎng)絡(luò)為了能進(jìn)行非線性預(yù)測(cè)，要建立一個(gè)類似線性模型的非線性模型，如下表示：其中：是由人工神經(jīng)網(wǎng)絡(luò)決定的非線性函數(shù)前饋神經(jīng)網(wǎng)絡(luò)用層來表示，通常有一個(gè)隱含層（在某些情況下有2層），層和層之間是充分聯(lián)系的，每一層有一個(gè)偏置單元（輸出層除外）。輸出是每個(gè)單元的加權(quán)輸入的總和（包括偏置），中間是通過指數(shù)激活函數(shù)來傳遞。我們已經(jīng)應(yīng)用了修正的反向神經(jīng)網(wǎng)絡(luò)。錯(cuò)誤的是定義了輸出單元的計(jì)數(shù)值和實(shí)際值或理想值之間的偏差的平方，這個(gè)定義使函數(shù)在微分的時(shí)候發(fā)生錯(cuò)誤。不

9、像線性的時(shí)間序列模型那樣在每個(gè)滯后變量有一個(gè)裝有系數(shù)，非線性神經(jīng)網(wǎng)絡(luò)滯后輸入變量的選擇和裝有系數(shù)的數(shù)量是獨(dú)立的，而網(wǎng)絡(luò)的規(guī)模，是有由層數(shù)和隱含層單元的數(shù)目決定的。此外，在線性回歸模型中，如果輸入變量是無關(guān)的，那么它的回歸系數(shù)是零。但是在非線性神經(jīng)網(wǎng)絡(luò)中者不一定是真實(shí)的；一個(gè)輸入變量可能不重要但是仍可能有權(quán)重；這些權(quán)重將會(huì)影響到下層的傳遞，對(duì)于隱含單元來說也是重要的。所以，在傳統(tǒng)的反向傳播神經(jīng)網(wǎng)絡(luò)中，沒有自動(dòng)消除無關(guān)輸入節(jié)點(diǎn)和隱含節(jié)點(diǎn)的功能。但是，在實(shí)際預(yù)測(cè)中有必要建立一個(gè)簡(jiǎn)約模型，它能解決實(shí)際問題，但不會(huì)太簡(jiǎn)單也不會(huì)太復(fù)雜。如果神經(jīng)網(wǎng)絡(luò)太?。ㄝ斎攵松倩蚴请[含單元少），就不夠靈活來捕獲電力系統(tǒng)的

10、動(dòng)態(tài)需求變化。這就是我們所知的“欠擬合”現(xiàn)象。相反地，如果神經(jīng)網(wǎng)絡(luò)太大，它不僅可以容納基本信號(hào)，還可以容納訓(xùn)練時(shí)的噪聲，這就是我們所知的“過擬合”現(xiàn)象?！斑^擬合”模型可能在訓(xùn)練時(shí)顯示較低的錯(cuò)誤率，但不能以偏概全，可能在實(shí)際預(yù)測(cè)時(shí)會(huì)有較高的錯(cuò)誤率。非線性模型可以產(chǎn)生比線性規(guī)劃更高的準(zhǔn)確度，但是要更長(zhǎng)的訓(xùn)練時(shí)間。較大的神經(jīng)網(wǎng)絡(luò)容易出現(xiàn)“過擬合”，預(yù)測(cè)需要簡(jiǎn)約模型的一般化概括。非線性神經(jīng)網(wǎng)絡(luò)的大小可以通過檢查相關(guān)性系數(shù)或是通過遺傳算法來選擇最優(yōu)的輸入變量來減小。線性模型相對(duì)于非線性模型來說是一個(gè)令人滿意的模型，而非線性模型是用來決定輸入?yún)?shù)的。用反向傳播來訓(xùn)練大型的人工神經(jīng)網(wǎng)絡(luò)是很耗費(fèi)時(shí)間的，很多用

11、來減少訓(xùn)練時(shí)間的方法已經(jīng)通過評(píng)估，已經(jīng)找到一個(gè)減少訓(xùn)練時(shí)間的方法來取代使用最小二乘法來修改網(wǎng)絡(luò)權(quán)重而達(dá)到速下降搜索的技術(shù)。每一步的計(jì)算量大了，但是迭代次數(shù)卻大大減少。減少訓(xùn)練時(shí)間是我們希望達(dá)到的，不僅可以通過減少計(jì)算消耗，也可以通過研究考慮更多的可取的輸入變量來達(dá)到，從而達(dá)到優(yōu)化預(yù)測(cè)的精度。4 神經(jīng)網(wǎng)絡(luò)結(jié)構(gòu)的確定4.1 自動(dòng)校正一階線性自動(dòng)校正就是校正負(fù)荷在兩個(gè)不同時(shí)間之間的校正系數(shù)，可以用下式表示：其中：是在時(shí)的自動(dòng)校正系數(shù) 是期望值 是在時(shí)刻的電力負(fù)荷值圖4.1顯示了滯后于某個(gè)特殊電力用戶的電力需求自動(dòng)校正系數(shù)的每小時(shí)負(fù)荷變化。這個(gè)圖證實(shí)了常識(shí)經(jīng)驗(yàn)，就是在任何時(shí)候的負(fù)荷與前幾天同一時(shí)刻的負(fù)

12、荷有高度相關(guān)性。這很有趣，并且對(duì)負(fù)荷預(yù)測(cè)很多幫助，另外，滯后的自動(dòng)校正在24小時(shí)中比前整個(gè)一周都高出許多。除了前4天，負(fù)荷的相關(guān)峰值下降到0.88外，第7天又上升了。圖4.1電力負(fù)荷自動(dòng)校正系數(shù)與滯后時(shí)間的比較我們也分析了樣本負(fù)荷在時(shí)間序列上的偏自相關(guān)函數(shù)（pacf）。這衡量去除了干擾變量后和之間的依賴關(guān)系。圖4.2顯示了負(fù)荷序列的pacf。可以觀測(cè)到，負(fù)荷變化與之前的負(fù)荷有很大影響，這就表明一個(gè)小時(shí)后的負(fù)荷預(yù)測(cè)將會(huì)變得簡(jiǎn)單。圖4.2 上午1點(diǎn)負(fù)荷的pacf4.2 遺傳算法在時(shí)間序列模型中重要系數(shù)可以通過遺傳算法自動(dòng)鑒定出，不像反向傳播模型的最小平方誤差那樣，遺傳算法可以直接將mape減到最小

13、。mape就是平均絕對(duì)誤差百分比，它廣泛用于衡量負(fù)荷預(yù)測(cè)的準(zhǔn)確度。為了描述遺傳算法里的負(fù)荷預(yù)測(cè)模型，要定義一根曲線，它包括滯后值和每個(gè)滯后的系數(shù)或是，那么這根曲線可以表示為：常數(shù)項(xiàng) 第一個(gè)滯后， 系數(shù)第二個(gè)滯后， 系數(shù)滯后， 系數(shù)第一個(gè)獨(dú)立變量的滯后，系數(shù)第二個(gè)獨(dú)立變量的滯后，系數(shù) 獨(dú)立變量的滯后，系數(shù)這樣一種曲線是隨機(jī)產(chǎn)生的。然后兩根曲線被隨機(jī)選擇（與它們的mapes的概率成反比）。兩根曲線的交叉點(diǎn)被隨機(jī)選擇，而兩條母曲線通過交叉點(diǎn)復(fù)制兩條新的曲線。這個(gè)過程中產(chǎn)生了新一代的曲線。將會(huì)計(jì)算出每一條曲線的適應(yīng)值（通過一組負(fù)荷數(shù)據(jù)訓(xùn)練而產(chǎn)生的預(yù)測(cè)mape的逆值）。這些低適應(yīng)能力的將會(huì)被丟棄，高適應(yīng)

14、能力的將會(huì)繁殖下一代。突變也用來隨機(jī)修改下一代中獨(dú)特的。結(jié)果就是經(jīng)過多代的繁殖過程，曲線具有高度的適應(yīng)性（低mape值），這就是用電力負(fù)荷通過訓(xùn)練后最好的預(yù)測(cè)值。5 短期負(fù)荷預(yù)測(cè)系統(tǒng)本文的短期負(fù)荷預(yù)測(cè)系統(tǒng)是一個(gè)線性神經(jīng)網(wǎng)絡(luò)（arma模型）和非線性神經(jīng)網(wǎng)絡(luò)的組合。整個(gè)系統(tǒng)的結(jié)構(gòu)如圖5.1示。圖5.1 短期負(fù)荷預(yù)測(cè)系統(tǒng)的結(jié)構(gòu)圖在這個(gè)系統(tǒng)中，線性系統(tǒng)和非線性系統(tǒng)兩者都有第二部分中提到的影響負(fù)荷預(yù)測(cè)的幾種或全部因素作為歷史數(shù)據(jù)的輸入。數(shù)據(jù)處理器的數(shù)據(jù)是從線性和非線性神經(jīng)網(wǎng)絡(luò)的歷史數(shù)據(jù)中提取出來的，分別地，線性神經(jīng)網(wǎng)絡(luò)的輸出作為反饋，輸入到非線性神經(jīng)網(wǎng)絡(luò)中。有歷史數(shù)據(jù)和線性神經(jīng)網(wǎng)絡(luò)的輸出作為輸入，非線性

15、神經(jīng)網(wǎng)絡(luò)就會(huì)預(yù)測(cè)出一天或者一周的負(fù)荷值。這兩個(gè)網(wǎng)絡(luò)組成的最初的網(wǎng)絡(luò)結(jié)構(gòu)是基于統(tǒng)計(jì)分析和遺傳算法。如圖4.2所示，時(shí)刻的負(fù)荷值很大程度上取決于時(shí)刻的歷史負(fù)荷值。所以，準(zhǔn)確地預(yù)測(cè)1小時(shí)后負(fù)荷的會(huì)提高短期負(fù)荷預(yù)測(cè)準(zhǔn)確度。但是，一天（24小時(shí)）后或在一個(gè)星期（168小時(shí)）后的預(yù)測(cè)，在之前的幾個(gè)小時(shí)的負(fù)荷值仍然是預(yù)測(cè)值。例如，我們要預(yù)測(cè)明天上午10點(diǎn)的負(fù)荷值，顯然，我們擁有的明天上午9點(diǎn)的負(fù)荷值不是實(shí)際值，我們只有明天上午9點(diǎn)的預(yù)測(cè)值。因?yàn)樵?點(diǎn)的負(fù)荷對(duì)10點(diǎn)的負(fù)荷的影響較密切，準(zhǔn)確的預(yù)測(cè)9點(diǎn)的負(fù)荷會(huì)提高預(yù)測(cè)10點(diǎn)負(fù)荷的準(zhǔn)確度。在我們這個(gè)系統(tǒng)中，線性神經(jīng)網(wǎng)絡(luò)（arma模型）是用來預(yù)測(cè)一個(gè)小時(shí)后的負(fù)荷值的

16、。對(duì)于非線性神經(jīng)網(wǎng)絡(luò)來說，輸入層包括不同時(shí)間滯后的變量。雖然時(shí)刻的負(fù)荷受到時(shí)刻的顯著影響，但是時(shí)刻的負(fù)荷本身的準(zhǔn)確度不足夠以至影響預(yù)測(cè)時(shí)刻負(fù)荷的準(zhǔn)確度。這主要受長(zhǎng)期負(fù)荷變化的影響（見圖4.1）6 仿真結(jié)果我們可以通過公共事業(yè)公司獲得歷史數(shù)據(jù)和各種天氣數(shù)據(jù)。我們用來仿真的數(shù)據(jù)是1898，1990和1991年的每小時(shí)歷史負(fù)荷數(shù)據(jù)和當(dāng)年的每小時(shí)的溫度數(shù)據(jù)。非線性神經(jīng)網(wǎng)絡(luò)由24個(gè)子網(wǎng)組成，沒一個(gè)代表一天中一個(gè)特定的時(shí)間。相似的，線性神經(jīng)網(wǎng)絡(luò)也有24個(gè)子網(wǎng)。全部48個(gè)子網(wǎng)有很多個(gè)輸入節(jié)點(diǎn)，但是只有一個(gè)輸出節(jié)點(diǎn)。在任何時(shí)候，只有一個(gè)非線性子網(wǎng)和一個(gè)線性子網(wǎng)在工作（總共只有2個(gè)網(wǎng)）。這種獨(dú)一無二的結(jié)構(gòu)具有以

17、下優(yōu)點(diǎn)：（1） 預(yù)測(cè)速度快（2） 重新訓(xùn)練系統(tǒng)快（3） 模塊化?？梢栽谔囟〞r(shí)間根據(jù)預(yù)測(cè)精度更新系統(tǒng)（4） 預(yù)測(cè)精度高可以得出系統(tǒng)的這些優(yōu)點(diǎn)對(duì)于商業(yè)應(yīng)用來說是很重要的。根據(jù)每小時(shí)或每天預(yù)測(cè)的原則來說，預(yù)測(cè)速度很精度對(duì)于公共事業(yè)來說是非常需要的我們用1898和1990年的歷史負(fù)荷數(shù)據(jù)和溫度數(shù)據(jù)來訓(xùn)練；1991年的負(fù)荷和溫度來作驗(yàn)證。在訓(xùn)練和驗(yàn)證期間，用到了未來的實(shí)際溫度。圖6.1顯示了利用1991年第一季度的數(shù)據(jù)驗(yàn)證我們的系統(tǒng)預(yù)測(cè)24小時(shí)后的mape值曲線。圖6.1 1991第一季度mape的驗(yàn)證結(jié)果7 優(yōu)化處理 由經(jīng)驗(yàn)可知，我們發(fā)現(xiàn)只有一個(gè)傳統(tǒng)神經(jīng)網(wǎng)絡(luò)的系統(tǒng)不足夠處理我們往往遇到的那些具有多種變

18、化情況的公共事業(yè)公司。例如，當(dāng)天氣突然變化時(shí)，利用常規(guī)的數(shù)據(jù)來訓(xùn)練系統(tǒng)不能得到較好的預(yù)測(cè)效果。當(dāng)系統(tǒng)的歷史數(shù)據(jù)點(diǎn)不足夠系統(tǒng)來學(xué)習(xí)時(shí)，可以通過簡(jiǎn)單地增加相似的歷史負(fù)荷點(diǎn)到訓(xùn)練數(shù)據(jù)中來解決上述問題。我們將增加兩個(gè)附加的子系統(tǒng)到我們的短期負(fù)荷預(yù)測(cè)系統(tǒng)中，給它取名為：基于規(guī)則的系統(tǒng)和模式識(shí)別系統(tǒng)。這兩個(gè)字子系統(tǒng)在遇到上述的一些情況下會(huì)起不同的作用和完成不同的任務(wù)。7.1 基于規(guī)則系統(tǒng)模式識(shí)別，遺傳算法和人工神經(jīng)網(wǎng)絡(luò)的時(shí)間序列模型所構(gòu)成的神經(jīng)網(wǎng)絡(luò)都可用作短期負(fù)荷預(yù)測(cè)。但是，為了獲得最小的預(yù)測(cè)誤差，且在可接受的復(fù)雜程度和訓(xùn)練時(shí)間，需要知道使用這個(gè)網(wǎng)絡(luò)的特殊公共事業(yè)的使用范圍。特別是對(duì)于區(qū)域的負(fù)荷預(yù)測(cè)，這些

19、特殊地理區(qū)域和服務(wù)場(chǎng)所或多或少受到諸如溫度和假期的影響，取決于這個(gè)區(qū)域的負(fù)荷是工業(yè)負(fù)荷占重要部分，還是商業(yè)負(fù)荷，或是民用負(fù)荷，或取決于負(fù)荷是在夏季達(dá)到峰值還是冬季達(dá)到峰值等。為了使公共事業(yè)單位或其他沒背景的公司能夠成功使用人工智能的短期負(fù)荷預(yù)測(cè)系統(tǒng)，當(dāng)它達(dá)到最佳性能的時(shí)候，有必要提供根據(jù)當(dāng)?shù)貤l件來設(shè)置變化參數(shù)的規(guī)則。7.2 模式識(shí)別系統(tǒng)這個(gè)系統(tǒng)被很多公共事業(yè)單位所用來作日常負(fù)荷預(yù)測(cè)的一種方法，它給出了一個(gè)小時(shí)為單位的負(fù)荷的大型數(shù)據(jù)庫(kù)，只要找出與預(yù)測(cè)日相似的負(fù)荷記錄，將它所在那天的數(shù)據(jù)作為預(yù)測(cè)的依據(jù)。這個(gè)系統(tǒng)的問題就是如何在歷史負(fù)荷數(shù)據(jù)記錄中找出相似的記錄。有很多種可行的方式來定義相似，我們所用

20、的其中一種就是比較平均絕對(duì)誤差百分比，我們概括為：（1） 神經(jīng)網(wǎng)絡(luò)可以用來識(shí)別模式或評(píng)估相似匹配程度。（2） 這些神經(jīng)網(wǎng)絡(luò)應(yīng)該組合起來，如用時(shí)間序列法(利用延遲線) 那樣單獨(dú)來預(yù)測(cè)，就存在每一種方法矛盾錯(cuò)誤的權(quán)重。結(jié) 論在本文中，我們介紹了以用線性和非線性網(wǎng)絡(luò)組成的負(fù)荷預(yù)測(cè)系統(tǒng)為基礎(chǔ)的混合神經(jīng)網(wǎng)絡(luò)。我們已經(jīng)論證了這個(gè)系統(tǒng)是理想的，可為公共事業(yè)或是商業(yè)應(yīng)用服務(wù)的。另外本文也描述兩個(gè)子系統(tǒng)，它們作為優(yōu)化處理我們現(xiàn)有的系統(tǒng)來處理各種不平常的情況。外文文獻(xiàn)原文artificial neural networks in short term load forecastingk.f. reinschmid

21、t, president b. lingstone h webster advanced systems development services, inc. 245 summer street boston, u 0221 0phone: 617-589-1 84 1abstract：we discuss the use of artificial neural networks to the short term forecasting of loads. in this system, there are two types of neural networks: non-linear

22、and linear neural networks. the nonlinear neural network is used to capture the highly non-linear relation between the load and various input parameters. a neural networkbased arma model is mainly used to capture the load variation over a very short time period. our system can achieve a good accurac

23、y in short term load forecasting.key words: short-term load forecasting, artificial neural network1 introductionshort term (hourly) load forecasting is an essential hction in electric power operations. accurate shoirt term load forecasts are essential for efficient generation dispatch, unit commitme

24、nt, demand side management, short term maintenance scheduling and other purposes. improvements in the accuracy of short term load forecasts can result in significant financial savings for utilities and cogenerators. various teclmiques for power system load forecasting have been reported in literatur

25、e. those include: multiple linear regression, time series, general exponential smoothing, kalman filtering, expert system, and artificial neural networks. due to the highly nonlinear relations between power load and various parameters (whether temperature, humidity, wind speed, etc.), non-linear tec

26、hniques, both for modeling and forecasting, tend to play major roles in the power load forecasting. the artificial neural network (a") represents one of those potential non-linear techniques. however, the neural networks used in load forecasting tend to be large in size due to the complexity of

27、 the system. therefore, training of such a large net becomes a major issue since the end user is expected to run this system at daily or even hourly basis. in this paper, we consider a hybrid neural network based load forecasting system. in this network, there are two types of neural networks: non-l

28、inear and linear neural networks. the nonlinear neural network is used to capture the highly non-linear relation between the load and various input parameters such as historical load values, weather temperature, relative humidity, etc. we use the linear neural network to generate an arma model. this

29、 neural network based arma model will be mainly used to capture the load variation over a very short time period. the final load forecasting system is a combination of both neural networks. to train them, sigxuiicant amount of historical data are used to minimize mape (mean absolute percentage error

30、). a modified back propagation learning algorithm is carried out to train thenon-linear neural network. we use widrow-hoff algorithm to train the linear neural network.since our network structure is simple, the overall system training is very fast. to illustrate the performance of this neural networ

31、k-based load forecasting system in real situations, we apply the system to actual demand data provided by one utility. three years of hourly data (1989, 1990 and 1991) are used to train the neural networks. the hourly demand data for 1992 are used to test the overall system. this paper is organized

32、as follows: section i is the introduction of this paper; section i1 describes the variables sigdicantly affecting short term load forecasting; in section iii, wepresent the hybrid neural network used in our system; in section iv, we describe the way to find the initial network structure; we introduc

33、e our load forecasting system in details in section v; and in section vi, some simulation result is given; finally, we describe the enhancement to our system in section vii.2 variables afferting short-term loadsome of the variables affecting short-term electxical load are:temperaturehumiditywind spe

34、edcloud coverlength of daylightgeographical regionholidayseconomic factorsclearly, the impacts of these variables depend on the type of load: variations in temperature, for example, have a larger effect on residential and commercial loads than on industrial load. regions with relatively high residen

35、tial loads will have higher variations in short-term load due to weather conditions than regions with relatively high industrial loads. industrial regions, however, will have a greater variation due to economic factors, such as holidays.as an example, figure 2.1 shows the loadvariation over one day,

36、 starting at midnight.figure 2.1 example of load variation during one day3 hybrid neurak networksour short-term load forecasting system consists of two types of networks:linear neural network arma model and feedforward .non-linear neural network.the non-linear neural network is used to capture the h

37、ighly non-linear relation between the load and various input parameters.we use the linear neural network to generate an arma model which will be mainly used to capture the load variation over a very short time period(one hour).3.1 linear neutal networksthe general multivariate linear model of order

38、p with independent x,iswhere:-electrical load at time t -independent variable at time t-random disturbance at time t-coefficientslinear neural networks can successfully learn the coefficient and from the historrcal load data,and the independent variables,widrow-hoff has been used to determine the co

39、efficient.this model includes all the previous data up to lag p.as shown above ,these data are not independent ,and have varying degrees of correlation with the load.correlation studies can be used to determine the most significant parameters to be includes in the model,allowing many to be eliminate

40、d.this reduces the size and computer time for a model of given accuracy,or increases the accuracy for a model of given size.3.2 non-linear neural networksfor non-linear forecasting,a nonlinear model analogous to the linear model is:where:f(.) is a nonlinear function determined by the artificial neur

41、al network.layered, feed-forward neural networks are used, typically with one hidden layer (although in some cases with two). the layers are fully connected, with one bias unit in each layer (except the output layer). the output of each unit is the slum of the weighted inputs (including the bias), p

42、assed through an exponential activation fiinction.our modiked backpropagation method is applied. the errors are defined to be the sum of the squares of the deviations between the computed values at the output units and the actual or desired values; this definition makes the error function differenti

43、able everywhere.unlike the linear time series model, in which there is one fitted coefficient for each lagged variable, in the nonlinear neural network forecaster tlhe selection of lagged input variables is independent of the number of fitted coefficients, the network weights, the number of which is

44、 determined by the number of layers and the number of hidden units. also, in linear regression models, if an input variable is extraneous, then its regression coefficient is zero (or, more properly, is not significantly different from zero by a t-test). however, in nonlinear neural networks this is

45、not necessarily true; an input variable may be unimportant but still have large weights; the effects of these weights cancel somewhere downstream. the same is true for the hidden units.therefore, in conventional backpropagation for nonlinear neural networks, there is no automatic elimination of extr

46、aneous input nodes or hidden nodes. however, in practical forecasting it is necessary to achieve a parsimonious model, one which is neither too simple nor too complex for the problem at hand. if the neural network is chosen to be too small (to have too few input or hidden units), then it will not be

47、 flexible enough to capture ithe dynamics of the electrical demand system; this is known as underfitting. conversely, if the neural network is too large, then it can fit not only the underlying signal but also the noise in the training set; this is known as overfitting. overfitted models may show lo

48、w error rates on the training set but do not generalize; they may then have high error rates in actual prediction. the nonlinear model can yield greater accuracy than the linear formulation, but takes much longer to train. large nonlinear neural networks are also prone to overfitting. forecasting re

49、quires parsimonious models capable of generalization. the size of the nonlinear neural network can be reduced by examining the correlation coefficients, or by using the genetic algorithm to select the optimum set of input variables. the linear model is a satisfactory approximation to the nonlinear m

50、odel for the purpose of selecting the input terms. large artificial neural networks trained using backpropagation are notoriously time-consuming, and a number of methods to reduce training time have been evaluated. one method that has been found to yield orders of magnitude reductions in training ti

51、me replaces the steepest descent search by techniques that model the network weights using a least-squares approach; the computations in each step are greater but the number of iterations is greatly reduced. reductions in training time are desirable not only to reduce computation costs, but to allow

52、 more alternative input variables to be investigated, and hence to optimize forecast accuracy.4 determination of network structureas we stated above, the neural network used in load forecasting tends to be large in size, which results in longer training time. by carefully choosing network structure

53、(i.e., input nodes, output nodes), one will be able to build a relatively small network. in our system, we apply statistical analysis and genetic algorithm to find the network "optimal" structure which is used as a base for further network turning.4.1 autocorrelation first-order linear aut

54、ocorrelation is the correlation coefficient between the loads at two different times, and is given by: where: is the autocorrelation at lag ze is the expected valuez(f) is the electrical load at time t.figure 4.1 shows the hourly variation in the lagged autocorrelation of electrical demand for a par

55、ticular electric utility. this plot confirms common sense experience, that the load at any hour is very highly correlated with the load at the same hour of previous days. it is interesting, and useful for forecasting, that the autocorrelation for lags at multiples of 24 hours remains high for the en

56、tire preceding week the peak correlation falls to about 0.88 for loads four days apart, but rises again for loads seven days apnpart. figure 4.1 autocorrelation of utility electrical load vs.lag hourswe also analyze the sample partial autocorrelation function (pacf) of the time series of load. this

57、is a measure of the dependence between zt+h and z, after removing the effect of the intervening variables zt+ , z 2, . zt+h-l . figure 4.2 shows the pacf of load series. it can be observed that load variation is largely affected by one at previous hour. this indicates that one-hour ahead forecast wo

58、uld be relatively easy. 4.2 genetic algorithm the most significant coefficients in the time series model can be identified automatically byusing the genetic algorithm. unlike the back propagation method, which minimizes the sum of squares of the errors,the genetic algorithm can minimize the mape directly.