Using artificial neural networks to predict the performance of a liquid sodium reflux pool boiler solar receiver

Liquid metal reflux receivers (LMRRs) have been designed to serve as the interface between the solar concentrator dish and the Stirling engine of a dish Stirling power system. Such a receiver has undergone performance testing at Sandia National Laboratory to determine cold- and hot-start characteristics, component temperatures, throughput power, and thermal efficiency, for various times of day and year. Performance modeling will play an important role in the future commercialization of these systems since it will be necessary to predict overall energy production for potential installation sites based on available meteorological data. As a supplement to numerical thermal modeling, artificial neural networks (ANNs) have been investigated for their effectiveness in predicting long-term energy production of a LMRR. Two types of data were used to train ANNs, actual on-sun test data, and ersatz data. ANNs were trained on both the raw on-sun test data and on pre-formatted versions of the data to determine if pre-formatting of the input data would improve network training efficiency and predictive abilities. Usable on-sun test data were available for only a few days of performance testing. Therefore, a set of year-long ersatz data was generated using a transient numerical model driven by one-minute meteorological data from the Solar Energy Meteorological Research and Training Sites (SEMRTS) data base for Davis, CA. The ersatz data were used to train ANNs based on warm-month data, cool-month data, and year-long data to investigate the impact of using seasonal test data on long-term predictive capabilities. The findings indicated that a network trained on data from a limited time span could successfully predict annual energy output of a liquid metal receiver.