Application of data-driven methods for energy system modelling demonstrated on an adaptive cooling supply system

The efficient and sustainable operation of building energy systems is playing an increasingly important role in most industrialized countries. At the same time, building energy systems are becoming increasingly complex; fault-free and optimal operation, under dynamic boundary conditions, is becoming more and more challenging. There are many approaches in research to address the optimal control problem of building energy systems, such as Rule-based Control, Model Predictive Control, or Adaptive Control. However, most methods rely on models of the system dynamics with high prediction accuracies. This is especially the case in Model Predictive Control, where the model is part of a continuously executed optimization problem; but models are also required when it comes to the optimal design of Rule-based Controllers, the safe pre-training of Adaptive Controllers, or model-based fault detection. A limiting factor for the manual development of physical models, for building energy systems, are the low monetary incentives for engineering services, due to the low energy prices in most countries. In addition, the creation of such models is time-consuming and error-prone, even for domain experts. Another weakness is that changes in the system dynamics are not automatically adapted within the models. These challenges are contrasted by an increasing availability of monitoring-data and computational power in recent years; with machine-learning algorithms, these resources are used in numerous application areas to achieve very promising results. Machine-learning methods can help to obtain data-driven, self-calibrating models, which can be learned from monitoring-data. In this paper, we apply methods for automated data-driven model generation. We demonstrate how machine-learning algorithms together with structured hyper-parameter tuning can be used to model individual subsystems as well as a complete energy supply system. To represent the dynamics of the supply system, it is first decomposed into simple functional relationships, which are aggregated into the overall system after training of the comparatively simple subsystem models. We evaluate the accuracy of the data-driven subsystem models using established metrics for the evaluation of regression models, namely the R2-score and the RMSE. The considered system is integrated into a district cooling network and consists of two compression chillers and an ice storage unit. Our investigations show that the dynamics of the subsystems can be learned with high accuracies, depending on the operation mode and the selected features. The prediction of the power demand of the compression chillers is learned with R2-scores between 0.94 and 0.99 and RMSE values between 2.02 kW and 3.51 kW. Also, the prediction of the percentage of ice formation within the ice storage is learned accurately with a R2-score of 1 and RMSE values between 0.08 % and 0.72 %. The dynamics of the aggregated system also show plausible behavior and can thus be used in future work. This work is part of an ongoing research project with the aim to optimize the operation of the entire campus cooling energy supply system. Our results show that, if detailed monitoring-data are available, data-driven modelling represents a viable alternative to the labor-intensive physical modelling approach. Furthermore, we emphasize the importance of structured hyper-parameter tuning, discuss the specifics of different machine-learning algorithms, and elaborate on possible future developments in this research area.