Prediction of photosynthetically active radiation in Kampala, Uganda using artificial neural networks.

Abstract

In this research, we formulated seven models to forecast photosynthetically active radiation(PAR) in Uganda based on artificial neural networks. Our models incorporated sunshine hours, global solar radiation, and clearness index as input parameters. The dataset utilized for model training and validation spanned five years, from 2010 to 2017, excluding 2013 and 2014 due to insufficient data. The study was conducted at Makerere University’s Department of Physics in Kampala, situated at coordinates 0.35 ◦ N and 32.58 ◦ E. The global solar radiation data was obtained using the CMP6 Pyranometer, the sunshine hours data was obtained using the CSD3 sensor while the PAR data was obtained using the NILU UV radiometer all of which are already installed at the physics Department. We employed Feed Forward-Back Propagation networks and trained them using the Steepest-Descent, Levenberg-Marquardt, and Adam training algorithms. A configuration with twenty(20) neurons proved suitable for training the model when all three parameters were used as inputs. We tested activation functions like Tan-sigmoid,rectified linear unit, and log-sigmoid in the hidden layer, while the output layer employed the linear transfer function. Tan-sigmoid exhibited superior performance across all models. Subsequently, we compared predicted and measured values of photosynthetically active radiation during the training and testing phases. The testing results revealed high positive correlation coefficients of 0.985, 0.970, 0.934, and 0.982 for clearness index, sunshine hours, and global solar radiation, respectively. Mean bias errors were determined as 0.003 MJm −2 day −1 , 0.004 MJm −2 day −1 , 0.031 MJm −2 day −1 , and 0.066 MJm −2 day −1 .Corresponding root mean square errors were 0.044 MJm −2 day −1 , 0.060 MJm −2 day −1 ,0.088 MJm −2 day −1 , and 0.046 MJm −2 day −1 . To underscore the effectiveness of our developed ANN prediction model, we compared it with an empirical model proposed by Jacovides et al.2003 for predicting photosynthetically active radiation at the same study site. The empirical model demonstrated correlation coefficients of 0.982 between photosynthetically active radiation and global solar radiation, with mean bias errors and root mean square errors of 1.062 MJm −2 day −1 and 0.955 MJm −2 day −1 , respectively. This comparison highlights the superior performance of our ANN prediction model over the empirical one.