Machine Learning Applied to Positive Displacement Compressors and Expanders Performance Mapping

Positive displacement compressors are critical components in today’s vapor compression refrigeration, air conditioning, and heat pumping applications and can also be applied as expanders in power generation systems, such as organic Rankine cycles (ORC). The simulation of such systems is essential to predict and optimize the performance behavior at full- and part-load conditions. To this end, comprehensive system models are built by including different sub-models corresponding to each cycle component (e.g., heat exchangers, compressor, linesets). In general, the higher the complexity of each sub-models utilized to capture the physics, the higher the computational time required to solve a simulation run. In this work, deep learning is utilized to obtain high-accuracy performance predictions of positive displacement machines. A fixed-speed two-phase injected and vapor injected scroll compressor for air-conditioning applications and an oil-free scroll expander for low-grade waste heat recovery by means of an ORC are considered as test cases. In particular, Artificial Neural Network (ANN)-based models have been developed for each of the machines and trained using experimental data collected at the Ray W. Herrick Laboratories. The results of the training and testing of the models are presented as well as a discussion of the reliability of such models for extrapolating performance. In addition, the ANN models are compared with conventional empirical and semi-empirical modeling approaches. The models have been implemented in the Python programming language by using the open-source Keras package