Percorrer por autor "Kreuzer, Laurenz"
A mostrar 1 - 1 de 1
Resultados por página
Opções de ordenação
- Data-driven solar surplus forecasting for UPACs using a deep neural networkPublication . Kreuzer, Laurenz; Nogueira, MiguelThis thesis addresses MEO Energia’s operational challenges in solar surplus forecasting for Units of Production for Self-Consumption (UPACs). These UPACs, including private households and small companies with solar panels, generate renewable energy for self-consumption and feed surplus electricity back into the Portuguese national grid. To comply with regulatory requirements, avoid financial penalties, and optimize operations, MEO Energia must report solar surplus production to the National Energy Network 48 hours in advance. To address these challenges, a predictive model based on a feedforward Deep Neural Network (DNN) was developed. The model generates forecasts at 15-minute intervals over a 7-day horizon, integrating historical production records with weather data obtained through a free weather API. An embedding layer enhances performance significantly by extracting latent patterns from supplier-specific IDs, allowing the model to adapt to the unique characteristics of individual UPACs. The model’s performance was evaluated using a six-month holdout dataset. Results indicate reliable short-term predictions, achieving daily aggregate errors of ±6% with the latest weather forecast data. Accuracy decreases with extended forecast horizons, reaching ±11% deviations for 48-hour forecasts, highlighting dependency on weather forecast precision. This research demonstrates the potential of DNN-based models to improve solar surplus prediction accuracy and operational efficiency. Current performance matches that of existing methods while offering significantly increased efficiency. Long-term improvements, driven by enhanced data quality and coverage, are expected to further optimize accuracy and position MEO Energia as a stronger competitor in the renewable energy market.