Artificial intelligence-driven performance mapping: A deep learning-based investigation of a multi-vane expander in retrofitted organic Rankine cycle

Sindu Daniarta,Piotr Kolasiński,Attila R. Imre,Dawid Sowa

DOI: https://doi.org/10.1016/j.enconman.2024.118763

IF: 10.4

2024-07-08

Energy Conversion and Management

Abstract:This study delves into experimental research on a retrofitted organic Rankine cycle (ORC), focusing on the performance of the expansion process near saturated vapor states (from two-phase to superheated conditions). A multi-vane expander was selected for its relevance in small-scale partially evaporated ORC. Several sensors were installed to monitor key parameters, including pressure, temperature, vapor quality, and rotational speed of the expander. The results show that the isentropic efficiency spans a range of 0.37 to 0.49. This variability was observed within an inlet vapor quality of 0.93–0.98, pressure ratios ranging from 5.67 to 10.67, and density ratios ranging from 7.4 to 13.3. In response to the challenge of time-consuming experiments and identification, this research employed deep learning techniques to predict the isentropic efficiency. Two different prediction scenarios were explored: one based on pressure ratio and the other based on density ratio. The results show that predictions based on density ratio offer better accuracy for isentropic efficiency compared to those based on pressure ratio. Three different scenarios for prediction were explored: initially utilizing two input parameters (density ratio and vapor quality), which show root mean squared error (RMSE) and coefficient of determination ( R 2 ) values in the ranges of 0.0136–0.0148 and 0.5832–0.6474, respectively. Subsequent enhancements were achieved by introducing inlet temperature as a third input parameter, improving performance with RMSE of 0.0033–0.0034 and R 2 of 0.9780–0.9793. Moreover, incorporating rotational speed as a fourth input parameter, the prediction was performed with nearly similar RMSE and R 2 values compared to using only three input parameters; it significantly reduces computing times in some cases, achieving speeds relatively 2.25–49.70% faster. This approach demonstrated a well-fitted line and an accurate performance map, establishing its effectiveness in predicting isentropic efficiency.

energy & fuels,thermodynamics,mechanics