PV Performance Modeling with Noisy and Incomplete Datasets

Abstract

As photovoltaic (PV) power generation gains increasing traction globally as an energy source, there is a growing need to be able to accurately predict power output of PV systems. Accurate predictions enable the module characterization, forecasting, and performance analysis that drives decisions to invest in solar energy or not. These predictions rely on mathematical models of PV systems which relate PV system performance to irradiance and PV cell temperature via model parameters such as the power output under reference conditions, temperature derating coefficient, and the degradation rate of the PV module. In this dissertation, it is shown that the most accurate predictions can be made with the use of laboratory PV module characterization and on-site weather and irradiance measurements, while two new methods are proposed to improve the applicability of the PVWatts irradiance-to-power model to existing PV power production datasets. The first proposed method is the use of a goodness-of-fit metric with a simple function fitted to daily PV power production data to detect and remove cloudy days without the use of irradiance-based clear-sky detection. Filtering based on the clear-sky detection method and time of day is demonstrated to improve the applicability of the PVWatts model to existing PV power production datasets even in the absence of on-site weather and irradiance data. The second proposed method combines models to remove variables and uses a statistical fitting approach to enable the analysis of PV array performance from datasets that otherwise lack sufficient information to inform model parameters. The viability of this approach is demonstrated for both on-site and remote weather and irradiance datasets. Both of these techniques will be discussed along with their impact on data interpretation and prediction fidelity.