Research on Commercial Building Thermal Model Based on CNN-LSTM

In this paper, we propose an enhanced model based on a convolutional neural network combined with a long short-term memory network (CNN-LSTM) to tackle the challenges of high computational complexity and the difficulty of accounting for regional temperature differences and interactions in the thermal modeling of commercial buildings. While the existing Equivalent Thermal Parameters (ETP) model is computationally efficient, it fails to capture the temperature variations within the building and the heat exchange between different regions. Conversely, the Resistance-Capacitance (RC) thermal network model offers accurate representation but incurs significant increases in parameter identification and computational demands as the number of regions grows. To address these issues, this study leverages the spatial feature extraction capabilities of CNN s and the time-series modeling capabilities of LSTMs to standardize the building topology into a thermal feature map, which includes parameters such as temperature, temperature variation, outdoor temperature, and heat. We construct a linearized convolutional network structure devoid of activation functions to ensure compatibility with linear programming solvers. The model is trained using one million data sets generated through simulation, significantly enhancing computational efficiency and practicality while effectively considering regional temperature differences and interactions. This approach provides a robust modeling tool for optimizing energy savings in commercial buildings.