Machine Learning for the prediction of the dynamic behavior of a small scale ORC system

Dynamic modeling plays a crucial role in the analysis of Organic Rankine Cycle (ORC) systems for waste heat recovery, which deal with a highly unsteady heat source. The efficiency of small scale ORCs (i.e. below 100 kW power output) is low (<10%). Therefore, it is essential to keep the performance of the system as stable as possible. To do so, it is helpful to be able to predict the dynamic behavior of the system, in order to perform a maximization of its performance over the time. In this paper, Feedforward, Recurrent and Long Short Term Memory networks have been compared in the prediction of the dynamics of a 20 kW ORC system. Feedforward Neural Network is the simplest among the architectures developed for machine learning. Recurrent and Long Short Term Memory networks have been proved accurate in the prediction of the performance of dynamic systems. This study demonstrates that the three architectures are capable of predicting the dynamic behavior of the ORC system with a good degree of accuracy. The Long Short Term Memory architecture resulted as the highest performing, in that it correctly predicts the dynamics of the system, showing an error prediction lower than 5% and 10% respectively for what concern the prediction 10 and 60 seconds ahead.