In this paper, we propose a machine learning (ML)-based method to significantly improve the accuracy of micro-PMV (Predicted Mean Vote) sensor system for Human Thermal Comfort (HTC) measurements. According to the analysis, an accurate prediction of metabolic rate is critical to determine reliable PMV values. Therefore, compared to the previous work, we have divided indoor human activities into more categories (9 vs 3) which cover a wider range of typical indoor scenarios to address the limitation of metabolic rate estimation. A Random Forest ML model was trained based on a 6-axis MEMS motion sensor data (acceleration and angular velocity) to achieve an impressive accuracy of 94.29% in the classified human activities. This approach significantly improves the event diversity for better accuracy of metabolic rate prediction, resulting in better performance of PMV-based HTC assessment. Thus, this work provides the foundation to develop a new ML-enabled micro PMV sensor system to be integrated for smart energy-efficient buildings to reduce carbon footprint in the era of AIoT (Artificial Intelligence of Things).