Novel machine learning approach for solar photovoltaic energy output forecast using extra-terrestrial solar irradiance

• Extra-terrestrial Solar Irradiance has been validated for PV output forecasting. • The machine learning approach successfully captures huge intra-daily PV output variations. • Study paves the way to develop a simple and effective PV output forecasting approach. A B S T R A C T The inherently intermittent nature of solar irradiance and other meteorological variables means that accurate forecasting of the photovoltaic power output is essential for planning and balancing photovoltaic power systems. This study proposes a novel approach to predicting one-week-ahead half-hourly photovoltaic power output in the United Kingdom using sloped extra-terrestrial irradiance and weather data (e.g., cloud-cover and temperature) as input parameters. A Non-linear Autoregressive Exogenous Neural Network is trained on a three-year historical dataset from two photovoltaic plants in the United Kingdom with capacities of 53 and 103 MWp. The forecasting model captures huge intra-daily variations of photovoltaic output, which is particularly useful to balance the supply and demand of the electricity system. The result of the study validates the concept of using sloped extra-terrestrial irradiance as an input parameter and suggests that meteorological conditions will dictate the accuracy of predictions. Findings also indicate that the use of sloped extra-terrestrial irradiance in conjunction with cloud-cover presented the optimal combination of input parameters as these provided the simplest and most cost-effective model without reducing accuracy. The approach can have universal value as it only requires coordinates and weather data. There is now a strong imperative to use the model in other locations where the weather is more stable