[Paper Review] Convolutional versus Dense Neural Networks: Comparing the Two Neural Networks Performance in Predicting Building Operational Energy Use Based on the Building Shape

Farnaz Nazari, Wei Yan|arXiv (Cornell University)|Aug 29, 2021

Building Energy and Comfort Optimization40 references4 citations

TL;DR

This study compares Dense Neural Networks (DNN) and Convolutional Neural Networks (CNN) for predicting building operational energy use based on building shape. Using architectural geometry as input, the DNN outperformed CNN in accuracy, training speed, and simplicity, though CNN offered visual interpretability through image-based representations.

ABSTRACT

A building self-shading shape impacts substantially on the amount of direct sunlight received by the building and contributes significantly to building operational energy use, in addition to other major contributing variables, such as materials and window-to-wall ratios. Deep Learning has the potential to assist designers and engineers by efficiently predicting building energy performance. This paper assesses the applicability of two different neural networks structures, Dense Neural Network (DNN) and Convolutional Neural Network (CNN), for predicting building operational energy use with respect to building shape. The comparison between the two neural networks shows that the DNN model surpasses the CNN model in performance, simplicity, and computation time. However, image-based CNN has the benefit of utilizing architectural graphics that facilitates design communication.

Motivation & Objective

• To evaluate the performance of DNN and CNN in predicting building operational energy use based on building shape.
• To assess the trade-offs between model accuracy, computational efficiency, and interpretability in energy prediction.
• To explore the potential of deep learning models in supporting early-stage building design decisions.
• To determine whether image-based CNNs offer advantages in design communication despite lower performance.

Proposed method

• Input data consisted of building shape geometry, converted into 2D representations for CNN and vectorized features for DNN.
• A DNN was trained using fully connected layers with ReLU activation and dropout regularization to prevent overfitting.
• A CNN was implemented using 2D convolutional layers, pooling, and fully connected layers to process shape images.
• Both models were trained on a dataset of building shapes and corresponding operational energy use values from simulation.
• Model performance was evaluated using mean squared error (MSE) and R-squared metrics, with unit-free MSE values reported.
• Training and inference times were recorded to compare computational efficiency.

Experimental results

Research questions

• RQ1 How do DNN and CNN models compare in predicting building operational energy use based on shape?
• RQ2 Which model architecture achieves higher prediction accuracy for the same input data?
• RQ3 What are the differences in training and inference speed between DNN and CNN models?
• RQ4 Can image-based CNNs enhance design communication despite lower performance?
• RQ5 How do architectural geometry and shape features influence energy prediction outcomes?

Key findings

• The DNN model achieved superior prediction accuracy compared to the CNN model, as measured by lower mean squared error (MSE).
• The DNN model demonstrated faster training and inference times, indicating better computational efficiency.
• The DNN was simpler to implement and required fewer hyperparameters than the CNN.
• Despite lower performance, the CNN model provided visual interpretability by processing building shape as images, aiding design communication.
• The corrected, unit-free MSE values confirmed the robustness of the performance comparison.
• Both models were effective, but the DNN was more suitable for practical energy prediction tasks due to its balance of accuracy and speed.