An unsupervised framework for sensible heat ratio dynamic prediction using air conditioning operation data

Sensible Heat Ratio (SHR), as a critical parameter characterizing the temperature and humidity regulation capacity of Air Conditioning (AC), holds significant value for energy efficiency, environmental comfort, and system optimization. However, traditional approaches relying on costly experiment measurements, complex theoretical modeling, or neglecting energy transfer hysteresis struggle to adapt to dynamically varying real-world scenarios. To address this, this study proposed an unsupervised dynamic SHR prediction framework. First, three hypotheses—continuity, correlation, and temporality—were introduced to establish theoretical foundations. Next, energy conservation equations for temperature and humidity within adjacent time windows were constructed to achieve unsupervised SHR calibration. Subsequently, an Enhanced Multi-Layer Perceptron (EMLP) incorporating physical value constraint, physical trend constraint, and an energy inertia module was proposed for dynamic SHR prediction. Finally, comprehensive validations were performed using an AC operation dataset (9.16 million entries) and an enthalpy difference measurement dataset (5,953 entries), encompassing parameter selection, characteristic analysis, model comparison, and noise robustness experiments. Results demonstrated high consistency with static enthalpy difference measurements (MAE = 0.075, ME = -0.039) and significant performance advantages compared to control groups in model ablation and performance comparison experiments. This research provides a low-cost, high-precision, and adaptive theoretical tool, offering an innovative pathway for the design of building and environmental systems.