An Improved Adversarial Learning Method for Cross-Scene Reconstruction of Industrial Load Symmetry Power Data Based on Denoising Diffusion

Symmetry power integrity is a core issue for power system data acquisition. However, industrial load data integrity is affected by missing values, abnormal disturbances, and low-reliability observations. A reliability-aware cross-scene industrial load symmetry power data reconstruction method is proposed based on adversarial learning. Firstly, an industrial electricity scene classification is proposed. Temporal and frequency-domain features are jointly encoded by a multilayer perceptron. The scene affiliation of the data is identified by cosine similarity to improve the cross-scene generalization capability. Secondly, a diffusion denoising generative adversarial reconstruction framework is proposed. For missing data, a conditional diffusion model is constructed with historical temporal distributions. Data structures are recovered by forward diffusion and reverse denoising processes. For low-reliability observations, original observations, first-order differences, and second-order differences are adopted to construct local shape constraints. In addition, the residual-correction guidance mechanism is introduced to estimate and correct observation deviations to improve the data reconstruction accuracy. Finally, simulations are conducted with industrial load datasets in Liaoning Province. The results validated the effectiveness of the proposed method. The average accuracies of data correction and data reconstruction reach 97.21% and 97.17%, respectively. Moreover, reconstruction accuracies exceeding 90% are maintained in cross-scene conditions involving different seasons and regions.