Speaker
Description
This study investigates whether decentralised, AI-driven flexibility agents can collectively reduce transformer peak load in an industrial park, even when each agent has no real-time knowledge of aggregate transformer loading or co-located consumers' actions. A two-stage simulation framework is used. First, a centralised day-ahead linear programme (LP) solves the joint scheduling problem for four industrial consumers over a synthetic training year, minimising peak transformer load subject to a ±10% per-hour flexibility envelope and daily energy neutrality. Its solutions serve as labelled training data, not an operational schedule. Second, one Random Forest model is trained per industry to map local inputs (own load profile, hour-of-day, calendar features) to the LP-optimal flexibility response; no transformer measurements or other industries' data are used.
On a held-out synthetic year never seen during training, the four independently operating models reduce annual peak transformer load by 4.1%, compared to 5.7% for the LP benchmark. They eliminate all transformer overloading hours present in the unoptimised baseline, improve the Peak-to-Average Ratio, and retain 95.8% of the LP cost saving. These results offer preliminary simulation evidence that centralised optimisation data can transfer implicit system-level knowledge to independently acting agents, enabling coordinated grid impact without real-time communication infrastructure. Scalability to larger consumer populations and robustness to real industrial load variability still require further investigation before deployment conclusions can be drawn.
| Topic Area | Computational Intelligence for Power Systems |
|---|