Model Order Reduction for Waste Heat Recovery: An Integrated Autoencoder-LSTM Neural Network Approach

An evaporator is a critical component in waste heat recovery systems, particularly within organic Rankine cycle technologies. In this paper, an approach for the simulation of two-phase evaporators is presented. This method aims at dimension reduction and prediction of temporal measurements, leveraging an autoencoder integrated with long short-term memory neural networks. The autoencoder-LSTM integration effectively reduces dimensions by compressing data representations and enhances temporal predictions, capturing dynamic system behaviors efficiently. The methodology effectiveness in predicting system dynamics is demonstrated, including a case study on high-fidelity data generation. In the absence of explicit system equations and with only snapshot matrices available, this method is preferred over projection-based model order reduction methods for its enhanced capability to capture and predict complex, non-linear temporal dynamics directly from data. The potential of inte-grating deep learning techniques with conventional engineering simulations to address the challenges of complex energy systems is highlighted. The results indicate that architectures with varying numbers of neurons in the examined bottleneck layer can accurately reconstruct and predict the original system's transient and steady-state behaviors with an error rate below 8%.