電子情報通信学会総合大会講演要旨
BS-1-6
Reducing overfitting in parameter estimates for high-accuracy thermal simulation of smart homes
○Hoai Son Nguyen・Yoshiki Makino・Hieu Chi Dam・Azman Osman Lim・Yasuo Tan(JAIST)
We have built a thermal model for a smart house testbed, which utilizes external environment data measured by sensors as input. Our simulation calculates the change of room temperature by calculating heat fluxes coming in and escaping a room based on a number of physical models. We estimated uncertain thermal parameters in our thermal model based on measurement data of room temperature. However, we found that least square estimation of these parameters in our model may cause overfitting since the measurement data and calculation model of simulation input have some margins of errors.
In this paper, we utilize ridge regression and cross validation in order to reduce overfitting in parameter estimates for our thermal model. Our simulation result shows that the use of ridge regression can improve the accuracy of our thermal model by 46% in mean square error.