Experimental application of a semi-parametric model for interpretable and accurate egression analysis of building energy consumption

• Regression algorithms are fundamental tools for the analysis of energy systems • Semi-parametric regression combines regression accuracy and model interpretability • The proposed semi-parametric model comprehends a physics-based sub-model and a neural network • Our experiments employ real-world data from 18 industrial buildings • This hybrid approach offers improvements with respect to both parametric and black-box ones Regression analysis is a versatile tool with numerous applications across diverse domains. Its utility extends to several tasks, including forecasting, inverse modeling, anomaly detection, and pattern identification. Over the years, researchers have mainly focused on two regression categories: parametric and non-parametric analysis. In light of the benefits and drawbacks of both methods, this work introduces a semi-parametric approach, combining regression accuracy and interpretability. This is achieved by designing a hybrid model, that includes a physics-based sub-model and a neural network. The proposed data-driven pipeline is applied to a relevant case study from the energy sector, namely the analysis of building energy consumption, achieving high accuracy compared to the parametric approach. Results demonstrate an increase in the mean coefficient of determination, from 0.77 to 0.94, with a MAPE drop from 5.5% to 2.2%. Meanwhile, the semi-parametric model allows the assessment of the thermal behavior of the buildings, thereby offering an improvement over black-box approaches.