Closing the Industrial Energy Audit Gap with Machine Learning: Estimating Product-Level Utility Costs Across Product Mix and Capacity Utilization in US Manufacturers

Abstract

Incomplete submetering and aggregate utility bills limit manufacturers’ ability to estimate how electricity, natural gas, steam, and water costs vary across products and operating conditions. This paper proposes the Physics-Guided Mix-Utilization Allocation Estimator (PG-MUAE), which combines engineering-based utility-intensity estimates with a gradient-boosted residual model to estimate product-level utility costs as product mix and capacity utilization change. The method is evaluated using a clearly labeled synthetic benchmark comprising 72 simulated facilities, 36 monthly periods, and five product families, yielding 12,960 facility–month–product records. Performance on 15 held-out facilities is compared with mass-proportional allocation, Ridge regression, Random Forest, Extra Trees, and gradient boosting. PG-MUAE records an RMSE of $0.02516 per product unit and an  of 0.96415. Its RMSE is marginally lower than gradient boosting’s $0.02519, while gradient boosting achieves a slightly lower MAE ($0.01851 versus $0.01878). PG-MUAE’s RMSE is substantially lower than the mass-allocation baseline’s $0.11910. These results demonstrate performance under the specified synthetic assumptions; they are not empirical estimates for U.S. manufacturers. The study’s evaluation framework examines prediction error across product-mix and utilization conditions, reconciles product estimates with facility utility totals, and assesses sensitivity to tariffs and operating assumptions. Validation with facility-level audit, metering, and production records is needed before operational use.