• • The proposed T-CGAN achieves superior performance in generating 8760-hour annual PV output scenarios, overcoming the gradient vanishing and mode collapse issues that plague LSTM-GAN and SeqGAN when directly trained on long sequences; this enables reliable long-term power system planning and annual operational scheduling without the boundary feature loss and cross-month correlation degradation inherent in short-sequence stitching approaches.
• • The multi-scale periodic attention mechanism, incorporating a learnable periodic bias matrix P[i,j] = cos(2π|i-j|/T) with adjustable strength parameter β, effectively captures both diurnal (short-term) and seasonal (long-term) periodic patterns in PV output; this is critical for accurately representing the deterministic periodic components that dominate PV generation variability across annual cycles.
• • The Wasserstein distance with gradient penalty as the adversarial loss function ensures stable training convergence and mitigates mode collapse, a common failure mode in standard GAN training for time-series generation; this stability is essential for industrial deployment where model retraining and reproducibility are required.
• • Comparative experiments against three typical models (WGAN-GP, LSTM-GAN, and SeqGAN) demonstrate the proposed method's superiority across statistical metrics, temporal correlation, and scenario validity, confirming that Transformer-based architectures with periodic attention are better suited for long-duration renewable energy scenario generation than recurrent or convolutional alternatives.