Engineering-based generic modeling of occupant behavior for energy efficient buildings

该项目的目的是开发一种综合的方法来模拟建筑中与能源消耗相关的居住者行为（OB）。由于OB对建筑物能源消耗的影响很大，并且基于目前的研究水平，本项目申请中提出的方法应在手动操作窗户的子案例中实施。由此产生的知识随后将用于模拟建筑物中与能源有关的进一步用户行为。为此，将OB操作定义为离散时间序列。将特别注意模型的推广，使其适用于广泛的建筑物居住者。个性化和泛化能力被认为是至关重要的组成部分。这种通用的OB建模方法目前是建筑性能仿真和建筑自动化系统（BAS）中缺失的一个重要组成部分，因为OB模型的可靠性和适应性在能源转型和智能电网控制的分散能源系统背景下变得越来越重要。建模方法属于深度学习的范畴。更具体地说，将研究前馈和递归神经网络（RNN）的不同架构及其适用性。在第一个研究任务中，前馈神经网络用于预测基于物理传感器测量的用户行为，并结合表示时间序列的概率图形模型。第二种情况包括循环神经网络（RNN）的发展，以预测人类的行为。这包括对最优网络架构的研究以及门状结构的适用性，包括用于基于物理系统先前变化的时间序列和执行动作建模的长短期记忆（LSTM）。

The aim of this project is to develop a comprehensive method for modeling energy consumption-related occupant behavior (OB) in buildings. Due to the high impact of OB on a building’s energy consumption and based on the current state of the art research the methods proposed in this project application shall be implemented on a subcase of manually operable windows. The resulting knowledge will be subsequently used to model the further energy-related user actions in buildings. For that purpose, OB actions are defined as discrete time sequences. Special attention will be put on the generalization of the models to enable its applicability for a wide range of building occupants. Individualization and generalization capabilities are regarded as crucial components. Such a generic OB modeling approach is currently an important missing component in both, building performance simulation and building automation systems (BAS), since reliability and adaptability of OB-models become increasingly relevant in the context of the energy transition and smart grid controlled, decentralized energy systems.The modeling methods belong to the category of deep learning. More specifically, different architectures of feed forward and recurrent neural networks (RNN) will be researched with respect to their applicability. In the first research task, a feed forward neural network is used to predict user actions based on physical sensor measurements in conjunction with a probabilistic graphical model that represents temporal sequences. The second case includes the development of recurrent neural networks (RNN) to predict human actions. This includes research on the optimal network architecture as well as suitability of gated structures, including long short-term memory (LSTM) used to model actions based on the time-series of previous changes in physical systems and performed actions.