Modeling and prediction of reflectance loss in CSP plants using a non linear autoregressive model with exogenous inputs (NARX)

Dust buildup on the surface of reflectors is a major challenge facing concentrated solar power (CSP) plants deployed in the MENA (Middle East and North Africa) region. Soiled CSP reflectors cause the efficiency of the solar field to drop. Thus, monitoring the loss of reflectance is essential to develop adequate cleaning strategies and evaluate the economics of the plants. The goal of this study is to model and predict the loss of reflectance in CSP plants. For this purpose, we modeled the loss of reflectance of second surface silvered glass mirrors exposed for 6 months in southwest Morocco using the non linear autoregressive with exogenous inputs (NARX) and Bayesian regularization. First, we adopted various architectures by varying the number of neurons in the hidden layer and adopting multiple tapped delay line. Then, we selected the optimal model based on the correlation coefficient (R) and the mean square error (MSE). The results revealed that the optimal model has 30 neurons in the hidden layer and 2 time delays, with a mean square error MSE = 0.029 and an overall R
_tot
= 0.62. To verify this model adequacy for the prediction of future reflectance data, we tested it on completely new data. The forecasting performance of the optimal model resulted in a mean square forecasting error of MSE
_f
=0.049. Generally, the forecasted reflectance values are quite good and follow the expected soiling pattern.