Identifying the key system parameters of the organic Rankine cycle using the principal component analysis based on an experimental database

•Key parameter subset of ORC is identified statistically based on experiment.•Principal component analysis of the ORC system parameters is performed.•A few machine learning regression models are developed and compared.•Key parameters are identified using exhaustive feature selection method. The organic Rankine cycle (ORC) is a promising technology for medium-and-low temperature heat utilization. However, the mechanism of how system parameters affect output have been investigated very little in the experimental aspect. Experimental investigation on the impact of each system parameter on system performance requires decoupling these system parameters. In this work, a series of experiments are conducted on a 10 kW scale ORC experiment setup. Statistical analysis is performed to identify a key parameter subset based on an experimental database. 6 system parameters, including temperature (Te) and pressure (pe) at the evaporator outlet, temperature (Tc) and pressure (pc) at the condenser inlet, expander shaft efficiency (ηSSE), and working fluid pump efficiency (ηP) are obtained. Combined with the ORC net power output and thermal efficiency, an experimental database of system operation conditions is constructed. Subsequently, the principal component analysis (PCA) of ORC is conducted based on the experimental database. Prediction models are developed using multi-linear regression (MLR), back propagation artificial neural network (BP-ANN), and support vector regression (SVR). Finally, accounting for the prediction performance of models and system parameter inter-correlation behavior, the key parameter subset is determined with the exhaustive feature selection method. The results imply that the key parameter subset is (pe, ηP, pc, ηSSE). Further removing or including more system parameters would reduce the accuracy of prediction models. In addition, the MLR models are slightly less accurate than the more sophisticated BP-ANN and SVR models.