Student : AM Diepeveen
About the student
I have a passion for advancing Sustainable Development Goals by developing advanced AI technologies, such as AMPERE. I am driven by solving real-world problems currently faced by humanity, in order to drive a more sustainable future, ultimately leaving an impact on society through my research work.
About the Project
Description of AMPERE Research Project: AMPERE integrates hierarchical multi-agent deep reinforcement learning with evolutionary algorithms to optimise smart grid distribution at regional and national scales under a Centralised Training Decentralised Execution (CTDE) paradigm, reducing deficits, surpluses, and distribution losses using real-world IoT datasets. Overview: AMPERE addresses legacy grid volatility by decentralising decisions and scaling coordination from regional to national levels, bridging gaps where single methods or flat agents lacked hierarchical knowledge sharing. Objectives: Minimise home-level deficit and surplus and electricity lost in distribution, operationalised via rewards that minimise Mean Squared Error (MSE) during allocation decisions. Architecture: Three-tier hierarchical agent architecture, namely national, regional, and battery agents, which coordinate via a publish-subscribe communication architecture. Learning and Coordination: CTDE trains policies centrally with decentralised execution. Furthermore, a publish-subscribe pattern propagates rewards and state signals upward to enable multi-scale coordination. Research Methodology: A Design Science Research (DSR) methodology guides artefact design, demonstration, and evaluation for a complex, real world optimisation problem. Data and Context: Real-world Pecan Street IoT time-series datasets simulate realistic conditions and support transferability across first and third world contexts, including South Africa. Novel Contributions: Combines Multi-Agent Deep Reinforcement Learning (MADRL) with Separable Natural Evolution Strategies (SNES) in a hierarchical, decentralised architecture that scales from regional to national smart grids, which is a novel implementation compared to flat single method systems currently in the literature. Backend and Training Plan: A FastAPI Python backend exposes real-time prediction and agent messaging over a three-tier publish-subscribe design. The main pipeline trains MADRL with Stochastic Gradient Descent, while a comparative MADRL with SNES pipeline evolves weights and biases alongside specialised predictors for supply, demand, and distribution variables. Justification and Impact: Given South Africa’s outages, ageing smart grid infrastructure, and economic losses, decentralised hierarchical MADRL can reduce losses and improve equitable, robust service, with global applicability.
