Energy-saving control of multi-zone purification ventilation system based on a novel multi-task learning framework

It is necessary to conduct a regular commissioning of purification ventilation systems to prevent issues such as filter clogging and imbalances in airflow to maintain optimal energy efficiency. This process should be performed at least once every three months, or after each switchover in on-duty cleanrooms. However, the traditional method consisting of testing, adjusting, and balancing (TAB) is time-consuming and inefficient, leading to a significant challenge for multi-area cleanrooms where airflow from neighboring rooms interferes with each other. To improve commissioning and energy efficiency, a comprehensive experimental platform with full airflow was specifically developed in this study to support key operational data. By leveraging the experimental platform to gather data, a multi-zone cleanroom ventilation steady-state prediction model (MCV-SPM) was developed, capitalizing on the multi-dimensional data-sharing capacity with the assistance of the multi-task learning (MTL) framework. Compared with traditional single-task models, this MTL-based model integrates two types of tasks, air balance and differential pressure control, using a shared bottom layer structure, which realizes cross-task information sharing and knowledge transfer, and improves the model's prediction accuracy and operation efficiency. Furthermore, a predict-then-optimize (PTO) control strategy is employed, where an optimization algorithm iteratively refines control parameters based on the MTL model's predictions. The mean error in the prediction of airflow for this system is only 1.8%, while the mean error in differential pressure is as low as 1.8 Pa, representing a notable enhancement over the 5.3% and 8 Pa of the existing methodology. Then, the proposed post-prediction optimal control strategy requires only 3% of the manual commissioning time (17.21 s) to reduce the average fan frequency by 7.5%, which corresponds to an estimated energy savings of 20.9%. This study presents a novel approach to improving the intelligent and efficient operation of purifying ventilation systems. The experimental platform utilized in the study adheres to the standards set forth for pharmaceutical cleanrooms. Further research may be directed towards the improvement of real-world data validation, the expansion of training datasets, and the investigation of the model's applicability across a range of industrial contexts.