Computer vision-driven dynamic optimal allocation strategy for load aggregator energy storage

Abstract

To address the challenges of multi-source fluctuations and redundant energy storage configuration faced by load aggregators in new power systems, this paper proposes a machine vision-driven dynamic optimization configuration method for energy storage. The study constructs a physical semantic feature extraction network incorporating a dualstream convolutional architecture and spatiotemporal self-attention, achieving dimensionality reduction mapping from video streams to power sequences via edge terminals. The joint optimization model injects denoised visual feedforward variables into the prediction hidden layer, combines lifetime degradation cost terms, and uses a linear programming solver to update the battery safety capacity configuration. Experimental results show that the edge inference latency of this crossmodal framework is 68 milliseconds, and the environmental fluctuation perception recall rate reaches 95.6%. Compared with conventional robust scheduling strategies, the capacity redundancy configuration rate decreases by 37.4%, the overall lifecycle operating cost is reduced by 27.5%, and the battery SOC steady-state tracking accuracy remains above 96%. This establishes a multi-dimensional feature fusion paradigm for distributed energy device scheduling from the underlying physical observation.