應用深度學習於短期負載預測之研究

The Study of Short-Term Load Forecasting Using Deep Neural Networks

陳昱斳(Yu-Ching Chen) ； 陳文輝(Wen-Hui Chen)

《先進工程學刊》 13卷3期 (2018/10) Pp. 131-136

摘要

在這個智慧電網盛行的時代，負載預測對電力公司來說相當重要的。使用者若能提前預知未來用電狀況或是電力監控系統具有耗能提醒的功能，則可以提高國民或電廠的節電意識，進而相對減少不必要的能耗。本研究主要是就變電所於2012年～2018年之變壓器紀錄資料，擷取出與負載相關性較高的資料特徵並將其整理分析，將資料劃分為全年非工作日、夏季工作日和冬季工作日三種，接著透過資料構建神經網路預測模型。實驗結果顯示，實際值與預測值的用電趨勢具有一致性，其中，使用卷積神經網路結合長短期記憶之模型，預測夏季工作日之平均絕對百分比誤差率最低，達到了4.67%，屬於高準確的預測。

關鍵字

電力負載預測 ； 深度學習 ； 卷積神經網絡 ； 長短期記憶

並列摘要

In this age of smart grids, power load predictions are incredibly important for the electric power industry. If consumers are able to accurately foresee their future power usage, if power monitoring systems have warning mechanisms, we could potentially raise awareness on saving electricity and thereby, reducing needless power waste. This research is based on a transformer station's power usage between 2012 and 2018. By analysing past power loads and other relevant factors in relations to each season. Information is divided into three categories, yearly, summer and winter; utilising these datasets, we create the prediction models of neural networks. Experiments results show that there is indeed correlation between the actual usage and the predicted usage; of which, using the model of convolutional long-short term memory to predict by summer dataset, MAPE yields the lowest difference rate of 4.67%, which is considered a highly accurate prediction.

並列關鍵字

load forecasting ； deep neural networks ； convolutional neural networks ； long short-term memory