應用灰理論及遺傳規劃法建構改良及混合式能源需求及消耗之預測模型

Abstract

由於高科技與經濟的發展均需要大量的能源，能源危機已成為全球的重要議題。從 世界各國的統計年報可以發現各國對能源的需求都呈現穩定成長的現象;而地球溫室效 應的問題使得氣溫漸趨酷熱或酷寒，造成有些國家水源短缺等問題，因此，對政府能源 管理單位而言，極需建立一套國家能源管理系統，以準確地預測能源需求量及消耗量， 進而制定出符合經濟效益之能源管理政策，減少不必要的成本。目前中外文獻大多是利 用計量經濟模型(econometrics model)，灰預測模型(Grey forecasting model)，混合預測 模型(hybrid forecasting model)來探討上述問題。在計量經濟模型中常用到自我廻歸整合 移動平均(Autoregressive Integrated Moving Average, ARIMA)方法來預測時間序列資 料。由於實際之能源資料是屬於非平穩型數列(non-stationary series)，需使用ARIMA 的 差分方法將其轉換成平穩型數列(stationary series)才能建構預測模型。然而 ARIMA 方法需符合一些統計假設，如預測模型須為線性模型，資料須符合常態分配;此外,由於 能源資料通常很少筆，能源管理決策者在此情況下利用ARIMA 所建構之能源預測模型無 法準確地預測未來能源趨勢的走向，而無法制定出符合經濟效益的能源管理政策。在建 構灰預測模型以及混合灰預測模型時不需任何統計假設;且只要四筆以上資料即可，故 灰預測模型比 ARIMA 模型更能符合能源資料的實際情況。由於實際能源資料取得不易 以及當實際資料量少於四筆時，利用灰預測模型建模時會有不穩定的殘差訊號 (residual sign)，雖已有文獻提出用類神經網路(Artificial Neural Network, ANN) 方法來估計殘差訊號，然而類神經網路必須要大量的資料才能建構良好的網路架構;而 在混合灰預測模型中，過去文獻上有利用灰預測模型結合ARIMA 方法在殘差資料的改良 來提升整體的預測準確度，然而ARIMA 方法有其統計上的假設問題;除此之外，影響能 源需求以及消耗之因素眾多，傳統的灰預測模型(GM(1,1))僅能使用作單一預測變數來 建模，當有多個預測變數，GM(1,1)不再適用。為了解決這些問題本計畫建立一套合理 且精確的能源需求量或消耗量預測模型以供政府能源管理單位參考之用。本計畫共分三 年完成，第一年的工作是先依據灰預測模型建構能源需求量或消耗量之預測模型，再利 用遺傳規劃法(Genetic Programming, GP)來估計残差訊號，可避免使用類神經網路在 少量資料預測殘差訊號能力的不足，來建立一套不限資料筆數且有效之能源管理預測模 型。本計劃第二年的工作是利用混合灰預測模型來建構另一個能源預測模型，本計畫利 用灰預測模型結合遺傳規劃法在少量資料以及非線性的特性提出一個新的混合灰預測 模型，可以免除如:ARIMA 在其殘差資料的假設。本計畫第三年的工作則是利用灰色關聯 在少量資料中找出影響國家能源的 N 個重要因子，再結合灰預測模型(GM(1,N)),以及 遺傳規劃法來建立一個精簡有效的預測模型。本計畫最後將利用台灣或其它國家的能源 消耗量作為實際案例，來說明本計畫所構建之能源預測模型之有效性。

With the high-tech and economical development of a country, energy demand and consumption become important issues. From the Statistical Year Book of each country, the energy demand demonstrates stable growing trend annually. The problem of global warming changes the temperature dramatically. Hence, it is very important to build an accurate energy demand or consumption model for the energy policy makers of the government. If the energy demand or consumption can be predicted accurately, some unnecessary cost will be saved. Many forecasting models are proposed such as the autoregressive integrated moving average (ARIMA) model, the grey model (GM), and the hybrid forecasting model. The ARIMA model is often utilized to predict the time-series data. Since the energy data are nonstationary, differential rule need to be utilized to transform the data into the stationary one. Moreover, the ARIMA model requires some statistical assumptions such as the model is linear, large sample size and data follow a normal distribution. The data collected on energy demand or consumption are often very few or non-normal Past studies developed two models (i.e., the grey forecasting model and the hybrid forecasting model) for forecasting the energy data Although a grey forecasting model, based on grey theory, can be constructed for at least four observations or ambiguity data, it yield large forecasting errors. To minimize such errors, the main objective of this three-year project is to build an improved grey forecasting and a hybrid grey model, respectively. The three-year project is divided into three phases. In the first year, the grey forecasting model and the genetic programming will be utilized to build a novel improved grey forecasting model through the forecasting ability of grey forecasting in small data and the detection ability of Genetic Programming (GP) in residual signs. In the second year, a hybrid grey forecasting model will be established to further improve the accuracy of the forecasted errors using the grey forecasting model and Genetic Programming (GP) in estimating the residual values. In the third year, the grey relation analysis is utilized to determine the important factors which influence the energy demand or consumption data. The grey forecasting model and GP are then utilized to build an energy forecasting model. Finally, the energy demand or consumption data from Taiwan, China or other countries were adopted to demonstrate the effectiveness of the proposed forecasting models.