本論文提出結合模糊理論與類神經網路的方法應用於電力系統短期負載預
測。本文提出兩種不同架構的預測模式:一為利用不同於一般反傳遞網路(
Counter Propagation Network) 的架構，在反傳遞網路的隱藏節點上，
以模糊分類的原理賦予其各別的學習率，再用加權 (Weighting) 的方式
，將網路激發的節點模糊化，使其節點具有更準確的運算。本文稱之為模
糊反傳遞網路(Fuzzy Counter Propagation Network)。二則為一模糊類
神經網路(Fuzzy Neural Network)，係應用非監督式學習(Unsupervised
Learning) 的方法將輸入變數先予以分類，求得每一輸入節點高斯歸屬函
數 (Gaussian Membership Function) 的參數值，再應用倒傳遞的監督學
習(Supervised Learning) 法則，求得最佳非線性輸出入對應模式。這兩
種方法基本上是利用模糊理論來改進類神經網路學習上的缺點，避免嘗試
錯誤的方法，以減少人為的負擔，增加模式的實用性。本論文所提出的兩
種模式，皆實際應用在台電公司全系統未來一小時的負載預測，以驗證其
預測準確度所得結果，並與現有巴氏簡氏(Box-Jenkins)時間序列預測法
及傳統類神經網路模式作一比較，模糊反傳遞網路的模式中，其預測相對
誤差平均值1.96%，標準差為0.0145%，而模糊類神經網路的模式，其預測
平均值為1.57%，標準差為0.1020%，證明所提出之模式應用於電力系統短
期負載預測的可行性。

The thesis presents two fuzzy logic based neural network to
power system short term load forecasting. Employing the
structure of counter propagation network, the Fuzzy Counter
Propagation Network (FCPN) first clusters the input historical
load data, and accordingly gives different learning rates to
the hidden neurons. By weighting and thus fuzzifying the
activation function of the neurons in the network, the neurons
will possess more capability of classification. The second
approach presented is the Fuzzy Neural Network (FNN). It first
clusters the input load data using unsuper vised learning
scheme. Then based on the results of clustering, parameters of
the Gaussian membership function of each input neuron can be
determined. Finally, the optimal model for the nonlinear input-
output mapping relationship is obtained by the algorithm of
supervised learning. Basically, these two approaches are
intended to ameliorate the problems with training the convent
ional neural networks. Consequently these approaches avoid the
process of trial and errors, as well as the humam involvement
in building the forecasting models, and further enforce the
practicability of the models. To verify the forecasting
accuracy of the proposed approaches, they are both applied to
forecast the one hour ahead load demand of Taiwan Power
(Taipower) system. Results obtained are also compared with
those of Box-Jenkins time series forecasting method as well as
conventional feedforward neural network model. Average relative
forecasting errors of 1.96% ( with standard deviation of
0.0145% ) for the Fuzzy Counter Propagation Network and 1.57% (
with standard deviation 0.1020% ) for the Fuzzy Neural Network
validate our approaches in power system short term load
forecasting.