Advancing industrial building energy measurement and verification (M&V) with deep learning: Evaluating data size and feature selection impact

This study explores the potential of advancing industrial building energy Measurement and Verification (M&V) using Deep Learning techniques, with a focus on the impact of data size and feature selection methods on model performance. Traditional M&V practices in the industry require extensive datasets, typically spanning 24 months, to ensure accurate energy savings verification post energy conservation measures (ECMs). This research challenges the norm by investigating whether a condensed four-month dataset could suffice for reliable M&V, thereby accelerating the verification process which is crucial for industrial applications. Utilizing a dataset with 30-minute intervals, the study first applies multi-linear regression across twelve feature selection methods to identify the most effective techniques based on a comprehensive set of performance metrics including Training Time, RMSE, MAE, R-squared, CVRMSE, NMBE, Mean MSE (via k-fold cross-validation), and the Standard Deviation of MSE (over 10-fold cross-validation). The top three methods-LASSO regression, Sequential Feature Selector, and Recursive Feature Elimination-were further analyzed as input features for a Deep Neural Network (DNN) model. The DNN's performance was evaluated across varying data sizes (20 %, 40 %, 50 %, 60 %, 80 %, and 100 %) and configurations ranging from one to ten hidden layers. The findings reveal that LASSO regression, when integrated with DNN, consistently outperforms in terms of CVRMSE and NMBE metrics across different data sizes and model complexities. Notably, increasing the dataset size from 20 % to 80 % markedly improves the model's predictive accuracy, underscoring the significance of larger datasets in enhancing DNN generalization and performance. Additionally, the study highlights a critical trade-off in DNN architecture: while models with fewer hidden layers (1-3) show greater performance variability in smaller datasets, increasing the number of layers does not linearly translate to better performance, illustrating the nuanced balance between model complexity and efficacy. This research contributes to the energy M&V field by demonstrating that shorter duration datasets, previously considered insufficient, can indeed provide accurate energy savings verification when analyzed with advanced deep learning techniques. This advancement not only paves the way for quicker, more efficient M&V processes in industrial settings but also offers valuable insights into the optimization of DNN models for energy data analysis.