Linear and Artificial Neural Network Radiation models over North India

Abstract

Abstract The present study examines three key aspects of global solar radiation modelling based on sunshine duration. Firstly, Angstrom coefficients were calibrated for both daily and monthly datasets using the recommended values of a = 0.25 and b = 0.50 as suggested by the Food and Agriculture Organization (FAO). Secondly, it conducts a stability analysis of the estimated coefficients for both daily and monthly time scales, taking into account variations in data length at selected locations due to the availability of data during different time durations. Lastly, data-driven artificial neural network (ANN) regression model was employed to not only enhance correlation and accuracy, but also to identify non-linear relationship between the duration of sunshine and global radiation. Four meteorological stations, namely New Delhi, Lucknow, Varanasi, and Patna, were chosen for this study. These stations are located in North India, specifically between 23°52º N to 30°15³ N and 76°24' E to 88°17' E. The Angstrom-Prescott models showed a strong linear relationship between sunshine duration and global radiation for both daily and monthly data with high statistical significance (p<0.01). The monthly and daily value calibrated coefficients for the Angstrom-Prescott models showed no significant variation in predicted global radiation root mean square error (RMSE), suggesting that both types of calibration can be used interchangeably. The average values of Angstrom coefficient for the region were found to be, a= 0.264, b=0.454 for monthly and a=0.292, b=0.427 for daily sunshine-radiation data. Angstrom constant ‘b’ shows relatively high variability with locations in two climate zones of the area. For monthly data, the estimated average coefficient of determination (R2) was 0.7962 and an average RMSE 1.2377 MJ/m2-day was found in North India. For the daily data, R2 = 0.7910; the average RMSE) was estimated to be 1.6794 MJ/m2-day in the region. The double and triple layer ANN regression models showed better performance improving R2 from 0.08% to 5.67% and lowering RMSE by 0.0345-0.2575 MJ/m2-day. Thus, the data-driven ANN regression models demonstrated not only a higher accuracy but also revealed the non-linear relationships between global radiation and sunshine duration in the region. Similar results were also exhibited by higher-order (non-linear) physical models for global radiation and sunshine duration conducted in India and other parts of the world. For most of the practical applications the linear A-P model was found to be simpler and sufficient with reasonable accuracy.