中華民國內政部建築研究所中華民國內政部建築研究所

一、研究緣起
空調設備在建築中佔有相當比例之能耗，因此智能空調控制變成了重要研究議題，然而，如何在感受度與能耗間達到平衡，需要考慮相當多樣的因素。當今許多新型態之辦公大樓皆裝有各類感應器，而建築物與內外環境間之物理關係，則隱含於這龐大且關係複雜的感測資料中，若能引入人工智慧科技，協助資料分析，其分析結果將可提供管理人員以更有效益、更環保之方式，管理建築物內部之溫熱環境，在節流、舒適、環保等面向上達到三贏。
二、研究方法及過程
本研究蒐集過去的案例與技術文獻探討空調控制智慧化的前瞻性，並詳細的介紹了本團隊因應真實案例提出的一種可線上學習、隨環境狀況彈性調度AHU溫度設定排程的系統，來控制實際建築物中之空調，證明其於實際建築中的可行性。
本溫控AI系統的架設包括兩個階段:
第一階段:旨在從具有偏差(Data bias)和訓練樣本的小型數據集中訓練建築模型。為了克服這一挑戰，我們增加了合成數據，並嵌入一些物理約束作為訓練初始模型的先驗。
第二階段:安裝至案廠後實際操作，模型可藉由日復一日地更新自適應調整以增加效能，使實際建築物中的在線控制能確保終端溫度更穩定地接近預期溫度。
三、重要發現
本研究之溫控AI系統能根據感應器所提供的大數據調控AHU溫度設定排程，與場域專家調控溫度之方式相比，本系統能更穩定同時又以更有效率的方式達目標之舒適度，並且具有隨時間進化的特性。
此外，本團隊也透過實驗分析於第四章探討應佈署的感測器、資料更新頻率、匯聚物溫熱環境感測大數據方式，探討溫熱環境、空調設備運轉、與感測資料間之相互關係，並透過實際案場控制確保其有足夠資料量可形成溫熱環境及空調控制者之經驗資料庫，整理感測資料庫規格書於附錄十三;應用機器學習分析建築溫熱環境感測大數據回饋於建築物空調系統運轉之模型如欲推廣至其他商辦建築，可參考附錄十二之”推廣此技術至可能案場之說明書”。團隊也於第五章就實驗說明應用機器學習分析溫熱環境大數據可發展之額外效益。
四、主要建議事項
建議一
持續推動智慧建築資料開放應用：立即可行建議
主辦機關: 內政部建築研究所
本研究所發展之機器學習演算法，係應用特定建築物日常營運所產生之大量建築溫熱環境感測開放資料進行可行性驗證。研究進行過程中，深感可取得之公開資料集為發展人工智慧應用技術之必要研究資源；此外，不同建築物之日常營運資料能形成更多樣化之資料庫，有助增進機器學習演算法之準確度及適應性，為個別建築物提供更適切之演算法。因此，建議後續將推動建築物日常營運資料開放應用研究及推廣列為優先工作，以利本研究所開發演算法之發展及推廣，並提供開發其他人工智慧科技之建築應用所需開放資料集及大數據來源。

1.Origin
HVAC equipment accounts for a large proportion of energy consumption in buildings. Thus, intelligent HVAC control becomes an emerging research topic recently. However, to deal with the balance between end-users’ concerns and energy consumption, a variety of environmental issues should be considered. Nowadays, modern office buildings are more or less equipped with multi-type sensors for understanding the environment. Truly, the abstract interaction between a building and its Air Control (AC) system are recorded and embedded within the huge and multimodal sensor data. However, how to utilize these sensor data to well control the building environment is still an open challenge. In this research, we attempt to analyze the complicated sensor data with the help of contemporary AI-related technologies. By well modeling the abstract interaction through data training, we aim to provide a systematic way to visualize the effects of control factors on the environment. The analysis result may also help the building managers to figure out a more efficient way to control HVAC subject to comfortability, energy saving, and environmental protection.
2.Proposed Method
We firstly collect various previous methods and examples as reference to discuss the novelty of smart HVAC control, and then propose a data-driven HVAC control algorithm, which is able to self-improving during the online control process, to adaptively provide the AHU with a suitable setting schedule under different conditions for its temperature set point. This system then will be installed in a real environment to prove its feasibility.
The system setup process is composed by two stages below:
(1)First stage:
This stage is mainly dealing with the data bias problem and sparse data problems. To overcome the challenges, we synthesize some data with physical constrains as prior to train an initial environment simulation model.
(2)Second stage:
The well-trained simulation model from first stage will be installed as the controller for AHU in real world, and then the system will adjust this model with new data explored by itself during the control process, to make the performance improve day by day.
3.Experimental result
Our AI system is able to keep the indoor temperature in a desirable range, in a more stable and more efficient way comparing with the original schedule designed by experts, and it is sure that the system can be improved over time.
Based on the experimental result, we discuss about which kinds of sensors should be deploy, what is the suitable data updating rate, and how strong of various environment factors impact on the indoor temperature in chapter 4. We further more show some extra benefit for energy saving in chapter 5. The specifications of sensors and data based are summarized in appendix 13. For generalization to different buildings, please take the information in appendix 12 as reference.
4.Suggestions
Suggestion:
Please keep promoting the open application of smart building data. Organizer: Architecture and Building Research Institue, Ministry of the Interior
Co-organizer: None
Our research aims to develop an algorithm based on the machine learning method to assist in providing a thermal comfort indoor environment. Besides, we’ve verified the feasibility of our system in a specific office building. However, we realize the importance of an open database for the development of an AI system in the thermal control field. An open database could not only provide the important information for the machine learning, but also improve the generalization and accuracy of the system. Therefore, we would like to suggest to promote the database buildup and its application. By doing so, we could improve the development of the automatic and intelligent building thermal control system based on the machine learning method and provide a rich source to different applications.