以機器學習預測建築自動化控制系統之短期電力負載

Abstract

根據聯合國環境規劃署UNEP估計，建築物所消耗的能源和釋放的溫室氣體占全世界能源總消耗量的40%左右，因此世界各國均視建築節能為減少溫室氣體排放的一個重要手段 [1]。其中我們關注的是提高能源效率—特別是電力—以減少電力浪費。由於BAS整合了各種電器控制，溫濕度、照明、頻率、水池流量、二氧化碳濃度等等，這些大量的歷史紀錄，使統計為導向的數據挖掘方法也因此能夠應用在建築能源預測上。能源管理者可藉由預測結果進行能源控制。 本研究在一部份提出了用在電量預測的特徵轉換方式，以及預測的架構。首先透過資料視覺化的方式分析使用者行為，並從資料視覺化的結果及文獻篩選進一步挑選適合的作為訓練資料的感測器，最後以Support vector machine（SVR）及線性加權回歸（WLR）進行用電量擬合。 第二部分為應用深度學習模型再用電量的預測。由於加權回歸模型與SVR須進行一層又一層的篩選，才能找出有關連的參數、特徵才能得到較佳的結果，但篩選過程中，不同篩選方式難免有疏漏或是難以描述的部分，因此我們提出了幾種不同的架構，能夠更全面的採用更多感測器的資訊，也能在少數幾個感測器受擾動時能夠避免模型完全失去預測能力。 本研究提供了更佳準確及彈性的架構，並有一個清楚的方法及流程作為參考，讓不同樣本數大小的資料集可以依據資料的特性選擇適合的預測架構。以達到預測的目的。

According to UNEP, the energy consumption and greenhouse gas discharged by buildings are responsible for about 40% of the global energy used. Thus, the energy efficiency is an important mean of reducing greenhouse gas emission. Among the improving methods, we put our attentions on energy efficiency to cut energy waste, especially on electricity consumption. In the past decades, the rate of buildings with Building Automation System (BAS) is increasing. BAS integrates electrical consumption, temperature, humidity and so on, which depends on the building. With various kinds of record, BAS allows data mining techniques to support decision making. The first part of our research developed an approach of feature extraction and a prediction structure which will be utilized in energy forecasting. To begin with, we analyzed user behavior by data visualization. Next, we selected the appropriate sensors to obtain training data through observing the results on the last step and literature reviews. At the last, we apply support vector regression (SVR) and weighted linear regression to train a regression model. In the second part of this study, we presented some deep learning structures to forecast electricity consumption. In the last part of our research, we combined some ways to select proper sensors. In addition, we made multiple steps to train a better model. To solve difficult problems such as that features are hard to describe, we integrated Deep Learning in this chapter. To sum up, we build a flexible and accuracy architecture which different BAS data and field can be applied in. In additional, we also provide a clear method and process as an example, so that people can select the appropriate forecasting architecture based on the characteristics of their data.