EnerVision AI driven digital twin for predictive energy intelligence and sustainable manufacturing

Small and medium-sized manufacturing enterprises form the backbone of industrial economies worldwide, yet they consistently struggle to monitor energy use, control energy cost, and meet sustainability targets because affordable, integrated decision-support tools remain scarce. This paper presents an Artificial Intelligence (AI)-driven digital twin for energy optimization that unifies, within a single low-cost platform, five capabilities that prior studies address only in isolation: interpretable regression of energy against production, ensemble and recurrent Machine Learning (ML) models benchmarked for demand prediction and forecasting, generator-activation alerting, statistical anomaly detection, and a what-if scenario simulator with a decision-support dashboard. The framework ingests daily production volume together with grid-supplied and diesel-generator energy from an operating facility and derives Key Performance Indicators (KPIs) for energy efficiency, Energy Per Unit of production (EPU), operating cost, and carbon dioxide (CO 2 ) emissions. Linear and quadratic regression models link production to energy demand and are scored with the coefficient of determination (R2) and Root-Mean-Square error (RMSE); a Random Forest (RF) regressor, an exploratory RF generator-activation alert, and a Long Short-Term Memory (LSTM) network for next-day forecasting are benchmarked against these interpretable baselines; because only six of the 59 valid daily observations were generator-active, the RF alert is not treated as a validated classifier. The principal novelty lies in integrating these complementary methods within one vendor-neutral digital twin aimed at resource-constrained enterprises. On a real case-study dataset, quadratic regression captures the production–energy relationship more closely than the linear baseline, and a load-shifting recommendation avoids generator use on the peak-demand day, reducing both cost and CO 2 emissions, while a 50% reduction in generator reliance yields substantial additional savings. A cross-validation study further shows that, in this small-data regime, higher-capacity ensemble and recurrent models do not improve out-of-sample accuracy over interpretable regression—an instructive finding that motivates the framework’s emphasis on transparent, deployable models for resource-constrained enterprises.