對於電力事業而言,精準的負載預測可避免造成限電的危機與資源的浪費,並可以提升電力系統運轉的穩定性與安全性。由於變電所每日供電情況不相同,且負載量不固定,然而沒有一種演算法在所有情況下都能得到最佳的預測效果,因此本論文提出以灰色模糊適應性共振網路作為負載預測的模型。灰色模糊適應性共振網路學習速度快,在資料量較少情況下學習內在聚類規則並以灰關聯分析衡量資料間關聯程度的大小,透過灰關聯度和經由交叉驗證法最佳化後的警戒值作共振測試,調整網路的可塑性與穩定性,達到高度的預測精確性。接著,本研究在實驗分析中將探討所提出之預測模型與模糊適應性共振網路及模糊時間序列預測模型三種不同演算法的效能,找出最適合於短期負載的預測方法。本論文將實驗流程分成兩個部分:第一部份的實驗共分為五組實驗樣本,以五組實驗樣本在三種預測模型下所得到的預測結果,討論其差異性及優劣;第二部份的實驗是以北部某一變電所配置四台主變壓器夏季之每日負載量作為實際預測。最後在實際負載預測中均方根誤差為0.2448、絕對平均誤差為0.58%,由結果可證實本論文提出之預測模型在短期負載預測中確實有良好的預測效能。
For electric utility, accurate load forecasting can avoid the crisis of restricting the use of electricity and the wasting of resources, and can enhance the stability and safety of power system operation. The power supply of substations changes daily, and the load is unfixed, however, there is no algorithm having optimal predictive validity in all situations. Therefore, this thesis proposes the grey fuzzy adaptive resonance network as the load forecasting model. The grey fuzzy adaptive resonance network has fast learning, the intrinsic clustering rule is learned in the case of low data volume and the correlation grade of data is measured by grey relational analysis. The grey relational grade and the cross-validation method optimized vigilance value are used for resonance test. The plasticity and stability of network are adjusted to reach high prediction accuracy. Afterwards, this study will discuss the effectiveness of three algorithms, the predictive model proposed, the fuzzy adaptive resonance network and fuzzy time series forecasting model in the experimental analysis, so as to find out the most suitable method for short-term load forecasting. The experimental process is divided into two parts in this thesis. There are five groups of experimental samples in the first part of experiment. The difference and quality are discussed according to the forecast results of the five groups of experimental samples in three predictive models. The second part of experiment uses the daily capacity of a substation equipped with four main transformers in the north in summer as actual forecast. Finally, according to the actual load forecasting, the root mean squared error is 0.2448, and the mean absolute percentage error is 0.58%. The results prove that the predictive model proposed in this thesis actually has good prediction efficiency in short-term load forecasting.