Assessment of deep learning techniques for prognosis of solar thermal systems
Renewable Energy, 2020•Elsevier
Abstract Solar Hot Water (SHW) systems are a sustainable and renewable alternative for
domestic and low-temperature industrial applications. As solar energy is a variable
resource, performance prediction methods are useful tools to increase the overall availability
and effective use of these systems. Recently, data-driven techniques have been successfully
used for Prognosis and Health Management applications. In the present work, Deep
Learning models are trained to predict the performance of an SHW system under different …
domestic and low-temperature industrial applications. As solar energy is a variable
resource, performance prediction methods are useful tools to increase the overall availability
and effective use of these systems. Recently, data-driven techniques have been successfully
used for Prognosis and Health Management applications. In the present work, Deep
Learning models are trained to predict the performance of an SHW system under different …
Abstract
Solar Hot Water (SHW) systems are a sustainable and renewable alternative for domestic and low-temperature industrial applications. As solar energy is a variable resource, performance prediction methods are useful tools to increase the overall availability and effective use of these systems. Recently, data-driven techniques have been successfully used for Prognosis and Health Management applications. In the present work, Deep Learning models are trained to predict the performance of an SHW system under different meteorological conditions. Techniques such as artificial neural networks (ANN) recurrent neural networks (RNN) and long short-term memory (LSTM) are explored. A physical simulation model is developed in TRNSYS software to generate large quantities of synthetic operational data in nominal conditions. Although similar results are achieved with the tested architectures, both RNN and LSTM outperform ANN when replicating the data's temporal behavior; all of which outperform naïve predictors and other regression models such as Bayesian Ridge, Gaussian Process and Linear Regression. LSTM models achieved a low Mean Absolute Error of 0.55 °C and the lowest Root Mean Square Error scores (1.27 °C) for temperature sequence predictions, as well as the lowest variance (0.520 °C2) and relative prediction errors (3.45%) for single value predictions, indicating a more reliable prediction performance.
Elsevier