Using Machine Learning to Predict Photovoltaic Energy Production

Renewable energy resources such as photovoltaics already play a significant role in our energy supply and will continue to do so in the future, especially with regards to fighting climate change. Forecasting the amount of production is important for multiple reasons such as ensuring grid stability, improving trading conditions or scheduling maintenance. The aim of this work is to find out which machine learning methods are well suited for predicting short term photovoltaic power production. Moreover, we aim to investigate the effects that features such as radiation, rain or module temperature have. A major part of this work was doing preprocessing, i.e. making the data provided by power plants more suitable to be used for learning. This includes multiple steps such as selecting features, filtering certain values and adding meta features. There are multiple machine learning models with varying complexity and unique strengths and weaknesses. We experimented with different models, namely linear regression, regression trees, neural networks and support vector machines.
To determine which of the tried models are best suited for predicting photovoltaic energy production, experiments with multiple configurations were run to build different models. These experiments were run on datasets from two seperate power plants in Austria, one from Burgenland and the other from Bisamberg. Furthermore, the influence of different features on the prediction quality was assessed by isolating and removing certain features. The models are tested and evaluated based on metrics caluclated by the predictions obtained from our tests. This allows us to compare them to each other and enables comparisons to similar works. In our experiments, multi-layer perceptrons ended up generally performing the best among most metrics. Regression trees and linear regression also produced good results. Regarding the impact of features, a good radiation forecast is the most important feature in making high quality predictions on produced energy. For one of the datasets, however, it was still possible to make decent predictions using all other features.