Practical Space Conditioning System Runtime Identification and Indoor Air Temperature Forecasting Using Low-cost Sensor Measurements

Developing practical and robust thermal dynamic models for indoor temperature forecasting is hindered by fundamental challenges: critical system runtime data is often unavailable or difficult to integrate, and complex temperature forecasting algorithms typically carry high computational demands. To address these foundational modeling bottlenecks, this study focuses on developing a highly practical framework for indoor thermal dynamic modeling. The methodology is twofold: first, we propose a data-driven method to implicitly identify space conditioning system runtime status using only indoor air temperature time-series data and low-cost sensors. Second, we evaluate the efficacy of a computationally efficient linear model for forecasting passive indoor air temperatures, directly contrasting its performance with a deep learning long short-term memory (LSTM) network. Results from an environmental chamber validate the robustness of this approach. The proposed runtime identification method accurately distinguished active and passive periods with over 98% classification accuracy. Furthermore, the linear thermal dynamic model demonstrated forecasting performance comparable to the LSTM network, maintaining maximum absolute temperature prediction errors within 1.2 ∘ C on the test datasets. By resolving these critical barriers to thermal dynamic modeling, this study establishes a prerequisite framework capable of being deployed directly on edge devices. Validating the accuracy of implicit runtime identification and the efficiency of linear models lays the vital groundwork for several advanced future applications. Ultimately, this foundational modeling framework enables the future implementation of computationally lightweight, energy-optimizing model predictive control (MPC) strategies for window shading and other complex building systems.