Umeå University's logo

Recognizing human behaviors offers an effective approach to reduce building energy consumption while maintaining indoor thermal comfort. This study proposed Wi-PTAB, a novel framework designed to recognize personal thermal adaptive behavior (PTAB) using channel state information (CSI) derived from Wi-Fi signals. The key PTABs essential for the framework were identified through a literature review. Building on these key behaviors, the proposed framework comprises three key steps: CSI acquisition, preprocessing and signals of interest (SOI) extraction; CSI image generation; and feature extraction and classification. During the feature extraction and classification phase, two specialized models combining a 2D Convolutional Neural Network (2D CNN) with either Long Short-Term Memory (LSTM) or Gate Recurrent Unit (GRU) were developed. To evaluate the effectiveness of the proposed Wi-PTAB framework and models, CSI data for the key PTABs were collected from two rooms and processed into four structured datasets, using varying temporal feature partitioning parameters, including window size and step size. The evaluation process involved measuring performance indicators for the proposed models on the self-developed datasets and comparing their performance to that of other neural network models. Results revealed that both models outperformed alternatives across most datasets, and model 2D CNN–LSTM achieved the highest average accuracy of 86 % on dataset III with a temporal granularity of 18. Additionally, two guidelines were established for effectively identifying the defined PTAB: (1) data must be collected with actions executed at the center of transceivers and at a 45° angle to the transceiver. (2) signal segments should be processed with a window size of 0.5 s and a step size of 0.2 s. This study provides technical and theoretical support for non-intrusive and private personal thermal comfort prediction and smart building design.