臺灣博碩士論文加值系統

　 用電負載預測長久以來即為設施規劃及負載管理人員所感興趣之主題
。一個正確之負載預測能夠提供有效之系統規劃。本研究是以類神經網路
中之多層後向傳導模式, 來進行台灣地區短期用電量之預測。此類神經網
路根據時間變數及與天氣有關之變數, 預測隔日之總用電需求量、尖峰用
電。根據過去之用電模式, 本研究之預測模式分為工作日、週末及國定假
日三種。本研究是以1991-1992年之用電數據來評估類神經網路之正確性
。此評估過程涵蓋一整年用電之預測, 以考慮季節性變化對預測正確性之
影響。對於工作日及週末, 本研究所發展出之模式能提供低於2%之絕對百
分比誤差。由整體成效來看, 此類神經網路模式確能提供有效且正確之預
測。

Forecasting of power demand has long been a subject of interest
to facilities planning and load management personnel. Accurate
load forecasting provides a basis for effective system
planning. In this thesis, a multilayer neural network based on
the widely used Backpropagation learning algorithm is proposed
for the short-term load forecasting of Taiwan power system. The
proposed network is designed to provide a one-day ahead
forecasting of both the total daily load and the peak power
load based on the time and weather related factors. The load
types were categorized into the normal working days, weekend
days, and holidays patterns. To demonstrate the effectiveness
of the proposed network, an extensive study was performed using
two years load data (1991-1992). Forecasting accuracy is
evaluated throughout a whole year to take into account the
seasonal effects on the accuracy of the proposed model. The
absolute percentage error of the proposed network was below 2%
for the weekday and weekend models. The overall performances of
the developed network indicates that it could be an efficient
and accurate method for short-term load forecasting.