Ridge regression and artificial neural network to predict the thermodynamic properties of alkali metal Rankine cycles for space nuclear power

Abstract

The alkali metal Rankine cycle system has great advantages in space nuclear power system. The investigation of thermodynamic properties of alkali metals is fundamental to develop thermodynamic model of thermoelectric conversion system. This paper build models for predicting the thermodynamic properties of alkali metals (potassium, sodium, cesium, rubidium) by statistics and machine learning. Pearson correlation analysis is performed to explain the correlation between measurable variables (temperature, pressure, specific volume, specific heat capacity) and unmeasurable variables (enthalpy, entropy). Ridge regression and radial basis artificial neural network models are established for the enthalpy and entropy of alkali metals respectively. Five evaluation indicators: residual, SSE, MSE, ME and R2 are selected to evaluate the accuracy of different types of prediction models. The practicability and extensibility of the prediction model are proved by case analysis. The results show that the prediction model of enthalpy and entropy based on radial basis neural network has high accuracy, and R2 can reach above 0.9. The entropy prediction model can even be stable at 0.999 when applied to different operating conditions and alkali metals. The new method proposed in this work solves the difficulty of experimental measurement in high temperature section to some extent, and opens a new idea for studying the properties of working medium. At the same time, the validation of the extrapolation feasibility of this method is of great significance to improve the thermodynamic properties database of alkali metals in the whole temperature range.