Enhancing energy savings verification in industrial settings using deep learning and anomaly detection within the IPMVP framework

This study advances industrial energy Measurement and Verification (M&V) practices by integrating Deep Learning (DL) techniques with automated anomaly detection, challenging traditional M&V reliance on manual non-routine adjustments. The research explores whether automated, data-driven anomaly detection can replace these adjustments, enhancing accuracy and efficiency in energy savings verification post-energy conservation measures (ECMs)—a critical need for industrial applications. Utilizing a dataset with 30-minute to weekly interval readings, CNN, DNN, and RNN models were applied across 12 datasets to identify the most effective model for baseline prediction using key IPMVP performance metrics (CVRMSE, NMBE, R2) alongside MAPE and RMSE. The baseline modelling findings indicate that DNN performs optimally at 30-minute intervals (R2 = 0.9600, RMSE = 22.82), hourly intervals (R2 = 0.9581, RMSE = 23.27), and daily intervals (R2 = 0.9347, RMSE = 28.00). CNN, however, demonstrated the best performance for weekly intervals (R2 = 0.8875, RMSE = 31.91). DNN provides the best overall performance across most intervals, offering a reliable balance of accuracy and practicality for regular energy baseline prediction. For anomaly detection and savings impact, the 30-minute RNN model achieved the highest estimated savings of 4.38 million kWh which translates to 27.35 % of the total energy consumption of 16,000,000 kWh with a low standard error (0.634 kWh), demonstrating strong predictive precision. Across all frequencies, savings estimates exceeded twice the standard error, meeting IPMVP acceptability criteria and confirming the robustness of this approach. These findings substantiate that deep learning-based anomaly detection can effectively replace traditional non-routine adjustments, providing a reliable, streamlined solution for energy savings calculations. Visualizations within the study illustrate the model’s enhancements, with comparative charts showing both original and anomaly-adjusted energy consumption and savings. This study contributes to the M&V field by demonstrating that, when integrated into the IPMVP framework, anomaly detection offers an efficient and accurate method for energy savings verification, paving the way for more streamlined, data-driven M&V processes in industrial settings. Additionally, it provides insights into optimizing deep learning models for energy data analysis, supporting quicker, more precise energy management decisions.