Physics-Informed vs. Deep Learning: Indoor Temperature Prediction with Different Data Availability

Reducing energy consumption in the building sector is crucial for global sustainability. Achieving energy efficiency requires advanced technologies and robust building models that address uncertainties and environmental variations. White-box models, based on physical principles, provide interpretability but require significant expertise and resources. In contrast, black-box models rely on historical data, lacking interpretability and being dataset-dependent. To bridge this gap, Scientific Machine Learning integrates physical insights into machine learning frameworks, ensuring interpretability while maintaining the accuracy and computational efficiency of models like neural networks. This work incorporates physical knowledge through Ordinary Differential Equations into the neural network framework, developing a physics-informed model to predict indoor air temperature based on external conditions, system thermal powers, and building internal gains. Datasets from four cities with diverse climatic conditions were used, and the model was trained on varying amounts of data, from two weeks to two years. This approach offers a novel exploration of model performance under different data availability and multiple scenarios. A comparative analysis with a Long Short-Term Memory neural network shows that, especially with limited training data, the Physics-Informed Neural Network outperforms the conventional model, with a Mean Absolute Error up to 0.69°C lower. This advantage is due to the incorporation of physics-based constraints, reducing reliance on large datasets. Additionally, the Physics-Informed Neural Network demonstrates stable accuracy across seasonal and uncontrolled dynamics conditions, highlighting its potential for temperature prediction and building control applications.