通用迴歸類神經網路在中長期電力需求預測模式之研究

Implementation of General Regression Neural Network into Long- and Middle- Term Electricity Demand Forecasts

蔡蓉媛(Jung-Yuan Tsai)

指導教授 ： 陳雲岫

元智大學/工程學院/工業工程與管理學系/碩士(2001年)

摘要

本論文提出以GRNN（General Regression Neural Network）的方法來預測中長期電力需求。ARIMA模式是傳統的預測模式，常用來解決時間序列問題，其效果是傳統方法中最好的。近年來利用類神經網路進行預測，已逐漸普及且效果不錯。本研究使用二種類神經網路(GRNN、BPN)並與傳統模式（ARIMA），進行比較。考慮兩種電力需求模式：尖峰用電模式及平均用電模式，分別進行需求預測，使用平均絕對誤差（MAD）為衡量指標。本研究使用年份、月份、氣溫、平均每人所得、工業生產指數及前1至3期之電力負載為輸入變數，資料來源取自官方公佈資料，時間範圍從1985年至1999年，共計165筆月資料。預測結果顯示，GRNN預測之結果比BPN來得好，且BPN在進行學習時，花費較多時間，評估各種參數（隱藏層層數、學習率、慣性因子、Epochs），這些參數對網路學習速度會有影響。研究結果顯示GRNN能夠在過去的資料中有效的學習出資料之間的關連與趨勢，GRNN在中長期電力需求預測方面較BPN有好的表現。

關鍵字

通用迴歸類神經網路 ； 倒傳遞類神經網路 ； 電力需求預測 ； 平均絕對誤差

並列摘要

In this research, we propose a novel approach to forecast long-term and middle-term demand. Traditionally, ARIMA was one of the most effective statistical methods used to forecast the power load in the past. Recently, Back Propagation Network（BPN） is a self-learning method and has performed well in various areas. We implement General Regression Neural Network into electricity demand model and our objective is to forecast the peak load and average load per period. The mean absolute deviation（MAD） is index for the performance evaluation. The factors affected demand considered in the GRNN model are year, month, temperature, personal mean income, index of industrial production, and backward 1-3 terms of load demand. Data was obtained from national official source. There are 165 sets of monthly data in total from 1985 to1999. The comparisons of forecast accuracy among GRNN, BPN and ARIMA are studied. Comparison results show that the mean and variance of MAD for GRNN approach both are smaller than the BPN’s. Meanwhile, BPN takes longer time in estimating such as, hidden layers, learning rate, moment rate, epochs, etc. for which learning speed can be influenced. Result reveals that GRNN can effectively learn the trend of demand and link between data from the historical data. Also performs better than BPN in forecasting both in long-term and middle-term load demand.

並列關鍵字

General Regression Neural Network ； Back Propagation Network ； Power Demand Prediction ； Mean Absolute Deviation

參考文獻

2.劉亭宜，GRNN在晶圓製造良率學習模式之建構與分析，元智大學（2000）。

5.Caire, P., G. Hatabian and C. Muller, ”Progress in Forecasting by Neural Networks” IJCNN-1992, Baltimore, II, 540-545。

6.Goh, T. N., H. L. Ong, and Y. O. Lee, “A study on the accuracy of time series modeling of daily power demand,” Electric Power Systems Research, pp.73-77（1985）。

7.Goh, T. N., S. S. Choi, C. H. Tan, and K. C. Tan “A comparative study of Short-term forecasting of energy and peak power demand.” Electric Power Systems Research, pp. 63-71（1982）。

8.Hestergerg, T. C. and A. D. Pappalexopoulos, “A regression-based approach to short-term system load forecasting,” IEEE Transactions on Power Systems, Vol.5, No.4, 1535-1544（1990）。