Machine learning methods applied to building energy production and consumption prediction

The utilization of renewable sources of energy is growing all over the world due to pressure for sustainable solutions. It brings benefits to the environment, but also adds complexity to the electricity grid, which faces energy balancing challenges caused by an intermittent production from this kind of generation. Having a good energy prediction is essential to avoid losses and improve the quality and efficiency of the energy systems. There are many machine learning (ML) methods that can be used in these predictions; however, every consumer is different and will behave in a distinct way. Therefore, the objective of this article is to compare the application of different ML methods, aiming to predict PV energy production and energy consumption for residential users. Four different ML methods were applied in a real dataset from the RESPOND project: Linear Regression, Decision Forest regression, Boosted Decision Tree Regression and Neural Network. After the simulation, the predicted values were compared against the real data, considering 150 days of measurement from two Irish houses. Overall, all the algorithms applied achieved mean errors below 14%, but the Boosted Decision Tree overperformed, with mean errors of 2.68% and 10% for energy consumption and energy production prediction, respectively.