建築設備のデジタルツイン生成に関する研究（第1報）：運転データに基づく熱源設備を摸するANNモデルの予測精度の検証

• 松田 侑樹

  ダイダン(株)

• 大岡 龍三

  東京大学生産技術研究所

書誌事項

タイトル別名

• DEVELOPMENT OF THE DIGITAL-TWIN FOR BUILDING FACILITIES (PART 1): VERIFICATION OF PREDICTIVE ACCURACY OF ANN MODELS FOR HEAT SOURCE SYSTEM BASED ON OPERATION DATA
• ケンチク セツビ ノ デジタルツイン セイセイ ニ カンスル ケンキュウ(ダイ1ポウ)ウンテン データ ニ モトズク ネツゲン セツビ オ モスル ANN モデル ノ ヨソク セイド ノ ケンショウ

抄録

 Information and communication technologies such as artificial intelligence and internet of things have developed significantly in recent years, and the concept of “Digital-Twin” that imitate actual world in a cyber space has been created. It is expected that new value will be created by connecting actual world and cyber space seamlessly. Effort to save energy using ICT have been conducted for building facilities, and among them, building automation systems, central monitoring systems, building energy management systems and simulations are representative. Simulations have important parts for planning and design as WHITE-BOX models. On the other hand, ANN models are expected accurate reproduction performance as BLACK-BOX models. We aim building "Digital-Twin", and evaluated the building method of ANN models based on operation data.  In this report, the target of ANN models is absorption chiller heater utilizing waste heat, it is evaluated that the parameters such as mini-batch size, construction of hidden layers and number of output nodes influence predictive accuracy and calculation time. The heat source system has a micro cogeneration system, solar collectors, a cooling tower and a heat exchanger as peripheral equipment, chilled and hot water which is generated by them is used in air handling units.  The influence of mini-batch size are shown below, the details are in Chapter 3. Increasing the mini-batch size increased the learning times for convergence, however the calculation time for 20,000 epochs was reduced up to 60%, on the other hand, the predictive accuracy improved as decreasing the mini-batch size. Therefore it is suggested that the mini-batch size with well-balance of predictive accuracy and calculation time is 64.  There are two composition parameters of ANN models, one is composition of hidden layers, and the other is composition of output layer. The result of evaluation is shown below, and the details are in chapter 4. However, considering the preceding result, the composition of ANN models which mini-batch size is 64 is evaluated. The number of nodes in the output layer part of ANN models has less influence on predictive accuracy than the composition of the hidden layer. When the nodes of hidden layers had 50-50 or more, the influence by the number of nodes in output layer could be ignore, there was no difference in predictive accuracy, moreover the coefficient of determine were as high as about 0.99 for outlet temperature and about 0.97 for gas and power consumption.  As a conclusion, in order to imitate a heat source system with hourly data for one year using ANN, it is well-balance that the ANN model has 50-50 of hidden layers, 5 of output nodes and 64 mini-batch size. In general heat source equipment, the types and number of input and output information are mostly same, therefore this method can be applied to a lot of heat source equipment. However, more complex ANN models are needed to improve the predictive accuracy for gas and power consumption.

収録刊行物

• 日本建築学会環境系論文集

  日本建築学会環境系論文集 85 (770), 267-275, 2020

  日本建築学会

キーワード

• 深層学習
• 人工ニューラルネットワーク
• デジタルツイン
• 最適化
• 中央熱源システム

詳細情報

• CRID: 1390848250108365696
• NII論文ID: 130007838033
• NII書誌ID: AA11830377
• DOI: 10.3130/aije.85.267
• ISSN: 1881817X
• ISSN: 13480685
• NDL書誌ID: 030396714
• Web Site: https://ndlsearch.ndl.go.jp/books/R000000004-I030396714
• Web Site: https://www.jstage.jst.go.jp/article/aije/85/770/85_267/_pdf
• 本文言語コード: ja
• データソース種別:
  • JaLC
  • NDL
  • Crossref
  • CiNii Articles
  • KAKEN
• 抄録ライセンスフラグ: 使用不可