Development and comparison of urban morphological solar energy and building energy consumption models using ANN and linear regression

This study focuses on the development and comparison of predictive models for solar energy generation and building energy consumption in the context of Singapore, examining the causality between urban morphology and these objectives. Two modeling techniques were employed: Linear Regression (LR) and Artificial Neural Networks (ANNs). Both LR and ANN models were developed through extensive tuning and testing of hyperparameters and parameters. The primary aim was to explore the relationship between urban morphological characteristics and two key aspects: solar energy potential and building energy consumption, and to compare the performance of the two models.

Over 500 neighborhood-scale simulations, using Rhino-Grasshopper Ladybug and Honeybee, were conducted to generate data for model development, evaluating both solar energy generation and building energy consumption. These simulations provided comprehensive data on how different urban morphological parameters influence energy dynamics within an urban setting. Solar panels were carefully positioned on rooftops and facades and minimum solar radiation thresholds were set, to assess their impact on building energy consumption, ensuring realistic simulations as solar panels installed provide shading to the buildings and hence reduce the building energy consumptions. Due to the “no free lunch” theorem, the selection of parameters and tuning of hyperparameters were conducted meticulously for both models.

Key performance metrics, including Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and R-squared values, were employed to evaluate the models. ANN models consistently outperformed linear regression models across these metrics. For instance, in the context of rooftop electricity generation, the LR model achieved R-squared values around 0.91, whereas ANN models demonstrated similar or slightly improved R-squared values (0.95), even during testing phases. In contrast, for facade electricity generation, the linear regression model exhibited a decrease in performance from training (R² = 0.83) to validation (R² = 0.74), indicating potential overfitting. Conversely, the ANN models maintained stronger performance with minimal reduction in R-squared values from training (0.95) to validation (0.93), indicating better generalization capability.

The findings underscore the importance of selecting appropriate modeling techniques tailored to specific urban energy planning needs. While linear regression models offer ease of interpretation, ANNs exhibit superior capabilities in capturing complex urban morphological interactions and providing accurate predictions.