Solar Resource Prediction Using Data-Driven Meta Models

In this work, four different machine learning models are developed for the predictions of ambient and solar data using long range (10 years) but discrete datasets in the fitting of the models. These datasets are taken from the location of this study, i.e., Silchar, Assam, India. Machine learning-based interpolation methods such as kriging interpolation, polynomial regression interpolation, nearest neighbor interpolation, and piecewise cubic Hermite interpolation techniques are used for developing the models to compare between the actual and predicted values. MATLAB code is written to generate the comparable graphs between the predicted results and interpolated results, i.e., fitted curves with actual values. The results indicate that the median errors in predictions by the models are in the range of 5.19–11.38%, with the least error exhibited by the Kriging interpolation method. Thus, this study demonstrates that kriging interpolation method can be suitably applied for the ambient and solar resource predictions for tropical conditions with discrete datasets.