Increasing accuracy of distribution prediction via an environmental soft sensor to realize air conditioning that combines personal comfort and indoor environmental distribution

Rina Hirai, Keiichiro Taniguchi, Shohei Miyata, Yasunori Akashi

Proceedings of Building Simulation 2023: 18th Conference of IBPSA, pp.3239-3247, September 2023

https://doi.org/10.26868/25222708.2023.1303

Abstract

Air-conditioning systems are designed to provide a uniform environment (Zhou et al. 2017). However, in reality, there is nonuniformity in the distribution of indoor environment. Additionally, human thermal comfort differs among individuals (Ishiura 2021). Even for the same individual, thermal comfort changes owing to the previous movement, clothing, and current physical condition (Wang et al. 2018). Thus, combining personal comfort and the indoor environmental distribution has the potential to make more occupants feel comfortable and lead to more efficient use of the room. For example, a person feeling hot can sit in a colder area, and a person feeling cold can sit in a warmer area, allowing occupants who have different thermal comfort to stay in one room. We evaluated this system with computation fluid dynamics (CFD) simulation data of CO2 and confirmed the possibility of positive use of indoor environmental distribution (Hirai et al. 2021).CFD analysis is a popular method for predicting indoor environmental nonuniformity (Srebric et al. 2002). However, it requires many inputs and cannot adapt to changes in use. Additionally, the calculations are time-consuming, and predictions cannot be made simultaneously. Another method is the installation of many sensors. However, this is expensive, and the sensors can be placed only in limited places. Thus, we invented the “environmental soft sensor,” which predicts environmental nonuniformity from scaled values of a few points. Soft sensors—systems that predict unscaled values from a few scalable values—are often used at plants (Shang et al. 2014).In a previous study, we examined the possibility of predicting environmental nonuniformity from measurements at a few points and deep learning using CFD simulation data (Hirai et al. 2022). Subsequently, we developed a simplified environmental soft sensor for and tested in an actual room, investigated ways to build deep-learning models and datasets, and identified the optimal measurement locations using winter measurement data (Hirai et al. 2022).In the present study, we made the following three contributions:1. Reduce oscillations in results using a recurrent neural network (RNN)2. Increase the prediction accuracy by changing the inputs3. Increase the prediction accuracy and investigate the seasonal effect by changing the training datasetAccordingly, our objective was to develop an environmental soft sensor whose root-mean-square error (RMSE) between the measured value and prediction value at every location in a room is <1 ℃. First, we changed the network model from an artificial neural network (ANN) to an RNN to consider temporal continuity and reduce oscillations in the results. Second, we developed five models with different inputs to optimize the input. The model Pattern2 , which uses a sensor on the wall, the air-conditioning unit inlet, and the total heat exchanger blowout and intake as input, had the best prediction accuracy. Third, we built three training datasets containing data for different air-conditioning seasons and made predictions using the models. The results indicated that data for the intermediate season were accurately predicted, using data from the other seasons. However, for the cooling and heating seasons, the prediction accuracy was significantly increased when the data for these seasons were included in the training dataset. Additionally, when we used Pattern2 (with a sensor on the wall + air-conditioning unit inlet + total heat-exchanger inlet and outlet as input) and training dataset ③ (with 2 months data for each of the heating and cooling seasons) to make predictions for a representative day in every air-conditioning season, i.e., cooling, heating, and intermediate, the RMSE was <1 ℃ at 14 of 17 places for the three seasons.