DEVELOPMENT OF THE DIGITAL-TWIN FOR BUILDING FACILITIES (PART 2): THE EVALUATION OF ANN MODELS TO SIMULATE ALL AIR CONDITIONING SYSTEM

Abstract

 Information and communication technology including internet of things have developed significantly in recent years. Building Digital-Twin which simulates a real product or building is expected to contribute to data-driven operations by connecting a cyber space and physical space seamlessly. The aim of this study is building a Digital-Twin of building facilities and realization of data-driven operation planning. In this paper, it is described that ANN models imitating all air conditioning system was built and its predictive accuracy was evaluated.

 An architecture of ANN model is forward propagation neural network which has an input layer, two hidden layers and an output layer. Each conditions of hidden layer nodes and input historical data are 4. A learning method to build models is back-propagation method, an algorithm is Adam and a loss function is RMSE. The predictive accuracy of models were evaluated for representative three days in summer and winter and MAE was applied as the evaluation function. The target building is the research institute which has an absorption chiller heater utilizing waste heat, a cogeneration system, AHUs and the others.

 As the result, the finding and issue are follows.

 (1) The ANN model has high predictive accuracy less than MAE 0.05. Also, the model can predict an efficiency of equipment.

 (2) In the case of this system, the predictive accuracy was improved under the condition of the ANN model had 800 hidden layer nodes and more 4 steps input historical data.

 (3) The items with high prediction accuracy regardless of the season were water temperature, pressure, and air temperature.

 (4) It was confirmed that the predictive accuracy decreased when the set point of equipment was changed. It is considered that the predictive accuracy is improved by continuous training of the ANN model.

 (5) Measured values and predicted values were significantly different when the failure occurred. It was suggested that a defect could be found by monitoring that values.