Accurate measurement of solar radiation is essential for understanding climate patterns, assessing solar energy potential, and predicting weather conditions. Over the years, solar radiation measuring instruments such as pyranometers and solarimeters have been used to achieve this objective. However, the high cost of these standard instruments makes the technology less accessible, especially to students and academic researchers in low-to-middle-income countries. A low-cost prototype solarimeter has been developed that operates using a mini solar PV panel and a microcontroller. During testing, it was found that as temperature increased the instrument had significant accuracy deviations. As such, this study seeks to optimise the performance of the prototype solarimeter using Artificial Neural Networks (ANNs), a powerful data-driven machine learning tool for optimisation. Solar radiation data was simultaneously collected using the prototype solarimeter and a standard solarimeter. Corresponding ambient temperature was also recorded for each solar radiation measurement. The data was used to train the ANN model to learn data patterns and to predict accurate solar radiation in spite of the ambient temperature. Results of the study revealed that temperature has a negative correlation (-0.7381) with accuracy, such that an increase in temperature reduces the accuracy of the prototype solarimeter. Increased temperature caused an accuracy deviation of about 27.16%. The ANN model successfully predicts accurate solar radiation measurement with an R-squared of 0.974, RMSE of 49.24 W/m2, and an accuracy of 86.32% which represents a 13.39% improvement in the performance of the prototype solarimeter. This study's novelty stands in its attempt to use machine learning to address the temperature sensitivity of a PV-based solar irradiance instrument. The results revealed here exposes the potency of deep learning models for optimising engineering systems.

Solar PV mini-grids are increasingly being deployed in off-grid and island communities especially in sub-Saharan Africa (SSA) countries to meet household energy demand. However, one challenge of solar PV mini-grids for community energy supply is the mismatch between the PV energy generation and household energy demand. PV mini-grid energy generation is highest in the afternoon whilst household energy demand is highest in the mornings and evenings, but lowest in the afternoons. This mismatch creates redundant energy generation during peak sunshine hours when battery energy storage is full, leading to low profitability for mini-grid systems. In this study, four machine learning models have been applied on an installed 30.6 kW mini-grid system in Ghana to ascertain the level of the redundant energy. The study has revealed that redundant energy exists on the mini-grid, in the range of 56.98 - 119.86 kWh/day. Further analysis has shown that the redundant energy can support household cooking energy demand through sustainable thermal batteries. With the four machine learning (ML) models applied in predicting the redundant energy, the most accurate ML model, K-nearest Neighbour Regressor, had a root mean square error (RMSE) of 0.148 and a coefficient of determination (R2) value of 0.998.