Scholarlink

<p>Residential buildings present highly variable occupant behaviors that critically influence energy use. However, most existing building energy modeling studies rely on oversimplified occupancy schedules that fail to capture the diversity and temporal structure of real household behaviors. This mismatch can lead to significant estimation errors and suboptimal control strategies. To address this gap, we propose a deep generative framework that is based on an autoregressive model implemented with long short-term memory (LSTM) networks to synthesize realistic daily occupancy sequences across four behavioral states (sleeping, absent, present, and cooking). Our model is trained on one-day time-use records from the American Time Use Survey (ATUS), encoded as 96-step categorical time series. We evaluate the generated schedules using extensive quantitative metrics, which include quarter-hourly marginal distributions, first-order transition matrices, <i>n</i>-gram coverage, Pearson correlation coefficient, and root-mean-square errors (RMSE). Results demonstrate that our model accurately reproduces both marginal and temporal properties of real ATUS schedules, while also producing diverse, novel patterns that extend beyond cluster-based prototypes. Finally, we demonstrate the real-world applicability of the generated occupancy schedules through case studies in building energy simulations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An autoregressive-based framework for simulating stochastic occupant behavior in residential buildings

  • 期刊论文

  • 发表日期：

    2026年2月1日

查看全文
  • Jiuyi Xu,
  • Ryunhee Kim,
  • Yunyang Ye,
  • Yangming Shi

摘要

Residential buildings present highly variable occupant behaviors that critically influence energy use. However, most existing building energy modeling studies rely on oversimplified occupancy schedules that fail to capture the diversity and temporal structure of real household behaviors. This mismatch can lead to significant estimation errors and suboptimal control strategies. To address this gap, we propose a deep generative framework that is based on an autoregressive model implemented with long short-term memory (LSTM) networks to synthesize realistic daily occupancy sequences across four behavioral states (sleeping, absent, present, and cooking). Our model is trained on one-day time-use records from the American Time Use Survey (ATUS), encoded as 96-step categorical time series. We evaluate the generated schedules using extensive quantitative metrics, which include quarter-hourly marginal distributions, first-order transition matrices, n-gram coverage, Pearson correlation coefficient, and root-mean-square errors (RMSE). Results demonstrate that our model accurately reproduces both marginal and temporal properties of real ATUS schedules, while also producing diverse, novel patterns that extend beyond cluster-based prototypes. Finally, we demonstrate the real-world applicability of the generated occupancy schedules through case studies in building energy simulations.