Performance evaluation, prediction analysis and optimization of experimental ORC using artificial neural networks (ANN)

• Machine Learning in ORC : Artificial Neural Networks (ANN) are increasingly used in ORC for predicting unknown or unmeasured experimental data, aiding in optimizing system performance and minimizing resource consumption. • Experimental Prototype : A 2 kW ORC prototype was developed and validated using 101 sets of experimental data to determine optimal operating parameters for maximum output work (W exp ) and thermal efficiency (η thermal ). • Key Input Variables : Seven input variables were used in the ANN model, with working fluid mass flow rate (ṁ) and expander inlet pressure (P1) identified as having the most significant impact on W exp and η thermal according to Pearson correlation. • Prediction Accuracy : The ANN model showed high accuracy in predicting results, with 10 neurons and two layers being the optimal configuration for the system. • Error Analysis : The absolute error for W exp was less than 250 Watts, while the relative error was under 40%. Similarly, the thermal efficiency (η thermal ) showed an absolute error under 5% and a relative error of less than 40%. • Optimization : The Pareto Front analysis yielded an ideal point of 2209.8 Watts for W exp and 11.91% for η thermal , indicating a trade-off between power output and thermal efficiency. The Organic Rankine Cycle (ORC) system stands out as the most efficient solution for converting low-grade thermal energy, making it highly effective for dispersed power generation and adaptable to various heat sources, such as solar energy, geothermal, biomass, and waste-heat recovery at different temperatures. Unlike traditional Rankine cycles, ORC systems use refrigerants or mixed fluids as working fluids, which have lower boiling points than water and are environmentally friendly, allowing efficient power generation on a smaller scale and at lower temperatures (above 200°C). While many experimental studies on ORC have been conducted, significant gaps remain in accurately predicting unknown or unmeasured data and identifying optimal operating conditions. This research addresses these challenges using machine learning, specifically an artificial neural network (ANN), a self-learning and nonlinear method capable of approximating complex functions, making it ideal for ORC prediction models. The novelty of this study lies in developing a 2 kW ORC prototype and applying ANN to predict and optimize performance using 102 experimental data sets—reducing experimental resource requirements and enhancing model accuracy. Additionally, a multi-objective optimization approach is used to simultaneously maximize net output work and thermal efficiency, setting a benchmark for efficient, low-cost, and sustainable ORC system designs. The benefits of this research include advancing predictive modeling for ORC systems, improving resource efficiency, and providing insights into optimized ORC operations for real-world applications.