Forecasting Short-Term Photovoltaic Energy Production to Optimize Self-Consumption in Home Systems Based on Real-World Meteorological Data and Machine Learning

Abstract

Given the growing number of residential photovoltaic installations and the challenges of self-consumption, accurate short-term PV production forecasting can become a key tool in supporting energy management. This issue is particularly significant in systems without energy storage, where excess production is fed back into the grid, reducing the profitability of prosumer investments. This paper presents an approach to forecasting short-term energy production in residential photovoltaic installations, based on real meteorological data and the use of machine learning methods. The analysis is based on measurement data from a functioning PV installation and a local weather station. This study compares three models: classical linear regression, Random Forest and the XGBoost algorithm. The method of data preparation, the model training process and the assessment of their effectiveness based on real energy production measurements are presented. This paper also includes a practical calculation example and an analysis of selected days in order to compare the forecast results with the actual production. Of the three models compared, the highest accuracy was achieved for XGBoost, with an MAE = 1.25 kWh, RMSE = 1.93 kWh, and coefficient of determination R2 = 0.94. Compared to linear regression, this means a 66% reduction in MAE and a 41% reduction in the Random Forest model, confirming the practical usefulness of this method in a real-world environment. The proposed approach can be used in energy management systems in residential buildings, without the need to use energy storage, and can support the development of a more conscious use of energy resources on a local scale.