Heating, Ventilation, and Air Conditioning (HVAC) systems account for a significant portion of global energy consumption, making their efficient operation critical for energy savings and demand-side management. Non-Intrusive Load Monitoring (NILM) techniques provide a promising approach for disaggregating HVAC loads from aggregate energy consumption using smart meter data. However, most NILM research focuses on high-frequency data, which is often impractical for large-scale deployment due to hardware and infrastructure constraints. This study addresses the challenge of temperature-dependent NILM using low-frequency data by evaluating three algorithms: Bayesian Disaggregation, Time-Frequency Mask Estimation, and BI-LSTM. The Bayesian approach models energy consumption as a probabilistic distribution to estimate HVAC loads, the Time-Frequency Mask method applies spectral transformations for enhanced signal separation, and BI-LSTM leverages deep learning to capture long-term energy dependencies. Using the ADRENALIN and AMPds2 datasets, we compare these models based on accuracy, computational efficiency, and applicability across residential and commercial buildings. The results indicate that the Time-Frequency Mask Estimation model provides the most consistent accuracy, while Bayesian Disaggregation performs well in environments with clear seasonal variations. The BI-LSTM model demonstrates stable performance but struggles with dataset inconsistencies. This study contributes to the field by providing a comparative analysis of NILM algorithms for low-frequency HVAC load disaggregation and offering insights into model selection for real-world applications. Future research should explore hybrid approaches that integrate spectral transformations with advanced deep-learning architectures to enhance NILM accuracy and generalizability.