Top

Constraint Screened Microgrid Black Start Strategy Generation Using ConvNeXtV2 CBAM

Authors

Files

PDF

Abstract

Microgrid black start requires coordinated inverter energization and staged load pickup under limited current capability, uncertain demand recovery, and incomplete measurements. This paper develops a ConvNeXtV2 based strategy generator with a convolutional block attention module to encode inverter histories and load restoration states, propose switching actions, and submit each proposal to an independent feasibility screen. Representation learning, sequence generation, and operational acceptance are separated so that physical constraints are evaluated independently of proposal probability. The evaluation covers 5100 operating records, including 3600 training, 900 validation, and 600 testing records, and compares six supervisory methods on common testing identifiers. The full configuration achieves 576 accepted episodes out of 600, compared with 558 out of 600 for the ConvNeXtV2 encoder baseline. Mean restoration durations within the respective accepted subsets are 98 and 106 seconds. The corresponding improvement is three percentage points in acceptance and eight seconds in restoration duration. Fifteen diagnostic panels examine operating coverage, transient response, duration distributions, computational tradeoffs, observation loss, and pickup sequencing. Joint analysis of voltage and current margins identifies cases in which satisfactory voltage recovery coexists with excessive transient current. The proposed framework connects heterogeneous restoration records to executable supervisory decisions through an explicit candidate acceptance interface. Its operating design supports the assessment of service recovery, constraint satisfaction, and decision latency within a common restoration procedure.

PDF

Details