Transfer learning for forecasting hourly indoor air temperatures of buildings with electrochromic glass

Abstract

This study aimed to employ transfer learning with a fully connected feed-forward neural network for forecasting the indoor air temperatures of adaptive buildings with electrochromic (EC) glass. This study predicted indoor air temperatures for an intermediate season requiring heating and cooling. Forecasting indoor air temperature can help control the EC glass to avoid overheating the interiors. The forecasting times for the predictions varied from 1 to 5 h between early morning and noon, which is when the interior is often overheated. The pretrained model was created using multilayer perceptron learning with the simulation data of a source building in Tokyo and transfer learning with feature-based extraction models that used datasets from the simulation of target buildings in Tokyo and Fukuoka. Further, the effects of facade orientation were investigated. The root mean squared error (RMSE) of the pretrained model varied from 0.027 to 0.935 when predicting the indoor air temperatures from 1 to 5 h. The RMSE of the transfer learning models using the pretrained model with the same and different orientations varied from 0.022 to 1.205 and from 0.9301 to 2.566. This study demonstrated that utilizing predicted indoor air temperatures to control EC glass can help protect against overheating.