Lessons Learned from a Field Study on Deep-Learning-Based Model Predictive Control for Optimizing the Operation of Facade Ventilation Units

The current study aims to present the lessons learned from development and implementation of a deep-learning (DL)-based model predictive control (MPC) algorithm for façade ventilation units (FVUs) in office buildings. Conducted as a field study in Germany, this research summarises the challenges and potential of DL-based MPC in reducing energy consumption during building operation. Controlled and optimized parameters included fan power, affecting airflow rate and the heating and cooling power of the FVU, as well as room temperature set points. The optimization function aimed to minimize energy consumption, with hard constraints on CO2 levels and thermal comfort. Combined deep learning models (Convolutional Neural Networks and Artificial Neural Networks), previously trained on building monitoring data, were employed to forecast energy consumption and room temperature within the MPC optimization loop.

Despite a plethora of research on MPC for optimised building control, practical implementations remain in the early stages and field research is scarce. To our knowledge, this study is the first to deploy MPC for the optimised control of FVUs in a field setting. The MPC evaluation involved measuring the actual energy consumption of the FVU in the investigated office room and comparing it to the predicted consumption under default control settings. This evaluation and the field implementation revealed several challenges.

Given the implementation under real operating conditions, challenges arose from measurement and sensor errors, building operation documentation inaccuracies, and communication issues with the building automation control network (BACnet). When integrating deep learning (DL) in MPC, it was observed that although DL models showed high accuracy for energy consumption and room temperature forecasts based on usual performance metrics, they did not produce plausible predictions for non-standard operational schedules. This emphasizes the importance of plausibility checks when using DL for MPC, highlighting that DL models based on standard operational data may be ineffective for optimizing non-standard schedules in heating and cooling control applications.

The implications for real-world implementations underscore that DL models trained on typical building monitoring data may not be adequate for optimizing atypical operational schedules. One key finding was that MPC can be used to assess DL models, as the optimizations revealed logical errors in the model. Additionally, the study highlights the necessity of sampling solution spaces before deploying even validated DL models within an MPC. This research advocates for rigorous validation and fine-tuning of DL models specifically within MPC frameworks to ensure their robustness and reliability in practical applications.