ARIMA-BPN 時間數列神經網路

ARIMA-BPN Times Series Neural Networks

本研究提出一個結合 ARIMA 與倒傳遞網路 （Back-Propagation Network,BPN） 優點的ARIMA-BPN 神經網路，它是以BPN 為模型，將ARIMA 模式的輸入，包括前時p 個時刻的數列值與前q 個時刻的數列殘差值做為輸入值，組成yt = f （yt-1, yt-2, …, yt-p, åt-1, åt-2, …, åt-q） 的非線性函數，以建立更準確的時間數列預測模型。因為數列殘差值在BPN 的訓練過程中會因網路連結權值的調整而改變，因此必須修改BPN 的演算法來適應此需求，即藉由不斷更新每次預測所得之殘差值做為網路的輸入值。本研究以六個人為設計的例題，及四個現實世界的例題來比較ARIMA、BPN 和ARIMA-BPN 三者的效能。研究結果顯示，ARIMA-BPN 神經網路演算法在部份例題比ARIMA與BPN 方法更準確。

In this paper we propose an ARIMA-BPN algorithm combining theadvantages of ARIMA and Back-propagation networks (BPN). Thealgorithm is based on BPN and its inputs are the same as ARIMA. It cangenerate a non-linear function yt = f (yt-1, yt-2, …, yt-p, åt-1, åt-2, …, åt-q) tocreate an accurate model to predict time series. The BPN algorithm mustbe modified because residuals will be changed when the weights arechanged during continuous BPN training. Therefore, the continuouslyupdated residuals are used as the inputs of ARIMA-BPN. This studyexamined 6 artificially designed cases and 4 real world cases to evaluatethe abilities of the ARIMA, BPN, and ARIMA-BPN. The results showedthat ARIMA-BPN is the most accurate method of the three.