Comparison between deep learning and fully connected neural network in performance prediction of power cycles: Taking supercritical CO2 Brayton cycle as an example

AI is becoming increasingly important in promoting the energy revolution of carbon-neutral to achieve sustainable development. Induced by the large implementation of renewable energy, the more complexities and uncertainties in the future carbon-neutral energy systems make their designs hard accessible to the conventional methods, so machine learning (ML) especially the neural network becomes under focus. Here, we design a deep learning architecture based on convolutional neural networks (DL-CNN) known for its powerful predicting ability, and first utilize it in a case study of performance prediction of supercritical CO2 Brayton cycle. The design paradigm of DL-CNN architecture for performance prediction of power cycle is proposed. We also summarize the commonly used fully connected neural network (FC-NN) in related studies of power cycle design. Through systematically comparing the prediction performance of DL-CNN and FC-NN, their respective advantages and application scenarios in energy system design are discussed. In addition, a multiobjective design approach based on DL-CNN combined with random search is proposed and proved to be feasible by comparing with genetic algorithm. The results show that our proposed DL-CNN model is much more competitive than FC-NN model when the training data is sufficient and the prediction condition is complex, in which the prediction accuracy can achieve 99.6%. In the future, our deep learning model may help solve the complex design problems of hybrid carbon-neutral energy systems.

Automated summary

AI plays a crucial role in driving the carbon-neutral energy revolution, with a focus on neural networks. The study introduces the DL-CNN architecture for performance prediction in power cycles, showing that DL-CNN outperforms FC-NN in terms of prediction accuracy.