Machine learning for electricity consumption forecasting in residential buildings within a local energy cooperative

• Main Author: Paredes, José António Santos
• Publication Date: 2024
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10773/42911
• topic: Local energy cooperative; Machine learning; Energy forecasting

The UNFCCC’s initiatives to combat climate change have spurred energy transition, leading to the emergence of zero-carbon energy markets. Accurate energy consumption forecasting is essential for optimizing the power grid and conserving resources amidst rising energy demand. This study aims to predict the next 24 hours of energy consumption for residential households within a local energy cooperative in Portugal. Leveraging historical energy consumption data and meteorological time series, the study focuses on implementing Machine Learning (ML) models, including xGBoost and LSTM algorithms. The analysis also encompasses traditional statistical forecasting models like ARIMA and Exponential Smoothing, alongside ML approaches. It investigates key features for consumption predictions and assesses the influence of meteorological data on forecast accuracy, while also clustering households into two groups based on consumption profiles. The analysis of the implemented models in unseen test data reveal that in the first cluster ARIMA had the best performance, with an MAE of 0.0783 kWh and a MASE of 0.9351, outperforming the two ML algorithms. In the second cluster, LSTM performed the best, achieving a MAE of 0.0372 kWh and a MASE of 0.9856. The study highlights the strong correlation between recent consumption data and future trends, with minimal impact from meteorological factors. The results confirm the effectiveness of statistical models in forecasting and highlight the potential for enhancing ML models in future research.