Machine Learning for Predicting Photovoltaic Power Generation: an Application on a University Campus

This paper describes the application of machine learning in an energy management system set up to monitor the use and generation of energy at a university campus. We explain how we built precise forecasting models for producing photovoltaic energy produced by solar panels at the University of Applied Sciences Landshut, Germany. We describe the practical challenges of dealing with data dropout and erroneous data when working with data from an actual in-production setting outside a simple lab scenario. Applying several data cleaning methods based on statistical methods, we obtained models that allow us to predict electrical energy produced by the solar panels with a precision of 0.97 as measured by the R
²
metric. More importantly, the process described to arrive at this result, more than the result per se, serves as a real-world example of applying data science and machine learning methodology in an in-production setting.