The advancement of deep learning (DL) has substantially improved the power grid’s ability to model and optimise diverse and highly complex tasks. Although intelligent transformation is an inevitable trend

traditional approaches increasingly reveal limitations in computational efficiency

nonlinear nonconvex optimisation

and multi-layer control mechanism design. To overcome these bottlenecks

effectively integrating advanced technologies—particularly DL—into Virtual Power Plants (VPPs) has become essential. As a result

recent research has shifted toward leveraging digital twins

communication network topology modelling

Large Language Model (LLM)-based cloud platforms and hierarchical distributed control strategies. These approaches aim to address the core challenges of deploying DL in practical applications

thereby driving VPPs toward a higher level of modernisation. This review provides a systematic assessment of existing research on VPP optimisation. Specifically

it examines key technological paradigms

including digital twins

communication network topology

LLM-based cloud platform construction and hierarchical distributed control

analysing the principal methods and major achievements through which these technologies shape current practice. With respect to DL–market integration mechanisms

the paper further identifies the limitations and key challenges of existing studies and outlines potential future research directions in methodological development and the expansion of optimisation objectives

thereby offering insights for advancing the field.