Research AI model predicts energy use from 2,000 records in seconds

2,000 hourly energy and emissions records fed a model that trained and ran in only a few seconds. Yongting Liu and colleagues used records drawn from smart meters, building management systems and industrial grids to build an AI system that predicts energy consumption and helps operators plan ahead of power generation and emission formation.

At its core, the system combines a transformer-based generative adversarial network, or GAN — a machine-learning setup in which two neural networks work against each other to produce realistic data — with Bayesian statistical optimization. The GAN learns complex patterns in energy use; the Bayesian method optimizes the system for its best performance.

That combination targets problems that conventional predictive models can struggle with: nonlinear consumption patterns, sudden spikes and troughs, and several external factors acting at once. Before testing, the researchers removed references to missing data and deleted outliers to normalize the input. The resulting predictions closely matched the observed energy data.

In concrete terms, the model could give energy managers a rapid rehearsal space. They could generate scenarios to test demand-shifting measures, explore how to integrate renewable generation and identify inefficiencies before actual power generation and emission formation. The approach could apply at industrial sites with internal power generation and across wider energy systems.

The result remains a research-stage model test, not a measured emissions cut in a live power system. The study used 2,000 prepared records and reports that training and running took only a few seconds; it does not describe a live deployment. Its practical value, for now, is the speed with which operators could examine possible choices before putting them into operation.