Performance and parameter prediction of SCR–ORC system based on data–model fusion and twin data–driven

This paper introduced an integrated approach of data–model fusion coupled with twin data–driven to address the bottlenecks of low accuracy and poor stability in machine learning for performance prediction and parameter reverse prediction in the Separation Compression Recirculation Organic Rankine Cycle (SCR–ORC). Initially, a cyclic database of SCR–ORC for five distinct working fluids was established through rigorous thermodynamic calculations. Subsequently, a Back Propagation Neural Network (BPNN) data–driven scheme of Grey Wolf Optimization (GWO) was developed for precise predictions of SCR–ORC thermal efficiency, exergy efficiency, and key parameters. Finally, utilizing the prediction results, Response Surface Methodology (RSM) was employed to analyze interactions among various parameters affecting the performance of the SCR–ORC, and further research on parameter optimization was conducted. The results indicate that GWO–BPNN achieves prediction errors 1–2 orders of magnitude lower than BPNN in performance prediction; regarding parameter prediction, the relative errors in GWO–BPNN predictions exhibit greater stability, with relative errors for each parameter prediction ranging from 0.01 % to 1.12 %. The proposed method notably enhances prediction accuracy, and the optimization outcomes of the agent model closely align with those of the traditional thermodynamic model.