This project leverages advanced machine learning techniques to predict urban energy demand in real-time, using a variety of urban and building features. By developing a parametric model, we simulate different urban forms and mixed-use buildings, capturing essential design parameters that influence energy consumption. Through a series of iterative simulations, we create a robust dataset that incorporates urban morphology, building characteristics, and energy-related factors, providing a comprehensive basis for energy demand forecasting.
To enhance the accuracy of our predictions, we employ sophisticated feature selection methods, such as Latin Hypercube Sampling and Random Forest analysis, to identify the most influential parameters. These techniques help reduce data complexity while maintaining essential information that improves the model’s performance. We also utilize both linear and non-linear correlation techniques, ensuring that our model can capture complex relationships between urban features, building designs, and energy consumption.
At the core of the project is the application of Graph Neural Networks (GNNs), which are well-suited for modeling urban energy systems. GNNs allow us to capture the dependencies between buildings and their interactions, enabling more accurate energy predictions. The model aggregates data at both the node (building) and graph (urban) levels, ensuring that both local and global factors are considered. This comprehensive approach results in a powerful tool for real-time urban energy management.
