本研究擬探討複數區域的短期電力總負載的預測，欲比較二種預測總負載之方式，其預測結果之差異性。即以全區總負載資料直接進行預測，或以分區負載資料，分別進行預測後再加總等二種方法之比較。資料的使用為美國德克薩斯州的電力負載資料，以及當地之氣候資料來進行負載總量的預測。德州區域分割方式是藉由氣候的不同將之分成八大區域。又由於國定假日和天氣異常日的趨勢與一般日子不同，因此將國定假日以及氣候異常的日子稱為特殊日另外處理，在一般日的預測利用 Cho et al. (2015) 提出的混合型方法，然與之不同的是，本研究之長期模型係使用月平均而非週平均來進行分析、建模與預測。而模型評估的部分，利用未來 24 小時負載預測與實績負載，計算平均絕對比例誤差 (MAPE)。長期模型的訓練集為 2006 年至 2014 年的每月平均負載量，短期模型的訓練集為預測日前 14 天再加入一些溫度相似的歷史資料，經過分析、建模後，再比較 2015 年利用全區總負載，或分區負載預測加總等二種方式，所得結果之平均絕對比例誤差。

In this work, we study the short-term load forecasting (STLF in short) of electricity demand with multiple regions, and compare the difference of results obtained by combining the STLF from subregions and those obtained directly from the total load of all regions. The data used in this study are the electricity loads and corresponding temperatures in Texas. There are eight regions divided by the weather zone. Due to the different patterns for the national holidays, days with aberrant weather in contrast to the usual days, we define those as special days and the others as normal days. In normal days, we use a hybrid procedure similarly as in Cho et al. (2015) except we use the monthly average for estimating the long-term effect instead of the weekly average to make predictions. To test the effectiveness of the above approach, the future 24 hours' electricity load forecasts are made and compared to the real demands, and calculate the mean absolute percentage errors (MAPE). The training set under study for the long-term load model includes the monthly averages from 2006 to 2014, and the training set for the short-term load model includes hourly data two weeks before the forecasting date as well as some historical data ensemble similar temperature pattern with the forecasting day. We present the performances of the hybrid approach for the two types of data arrangement as described above.

論文審定書i
誌謝ii
摘要iii
Abstract iv
1 前言1
2 資料介紹與處理2
3 一般日模型3
3.1 模型介紹. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
3.2 日期週期效應. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.2.1 週期基底介紹. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
3.2.2 週期基底模型. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2.3 週期基底生成. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.3 溫度效應. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.4 殘差分析. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.5 模型選取. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4 特殊日模型17
5 預測結果18
5.1 一般日預測結果. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
5.2 一般日即時校正. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.3 特殊日預測結果. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
6 結論與建議23
7 參考文獻24
8 附錄25
8.1 附錄一. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
8.2 附錄二. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

Cho, H., Goude, Y., Brossat, X., & Yao, Q. (2013). Modeling and forecasting daily electricity load curves: a hybrid approach. Journal of the American Statistical Association, 108, 7-21.
Fan, S., & Hyndman, R. J. (2012). Short-term load forecasting based on a semi-parametric additive model. IEEE Transactions on Power Systems, 27, 134-141.
Harvey, A., & Koopman, S. J. (2008). Forecasting Hourly Electricity Demand Using Time-Varying Splines. Journal of the American Statistical Association, 88, 1228-1236.
Hsu, C. (2016). Short-term Electricity Load Forecasting. Master Thesis, Department of Applied Mathematics, National Sun Yat-sen University.
Tsay, R. S. (2005). Analysis of financial time series. John Wiley & Sons.
Electric Reliability Council of Texas. http://www.ercot.com/.