Abstract: To address the common problems of insufficient temporal feature extraction and poor adaptability to special day scenarios in day-ahead electricity spot market price forecasting, this paper proposes an improved forecasting model based on the Transformer architecture. The Patch mechanism is introduced to enhance local temporal feature extraction, and the channel-independent structure is combined to improve the learning efficiency of multivariate features. In addition, the multi-head attention mechanism is adopted to capture the global price fluctuation patterns. The proposed method is verified based on the historical data of the Guangdong electricity spot market. Compared with the baseline Transformer model, the mean absolute error (MAE) of the proposed model is decreased from 32.95 to 23.88 in weekend scenarios, and from 78.33 to 70.33 in holiday scenarios. The model exhibits significantly better adaptability to the phenomenon of quantity-price deviation than the baseline model, and can accurately capture the upward trend of price floors when the bidding space exceeds 60000 MW. The proposed model achieves a significant improvement in prediction accuracy under different scenarios (especially special scenarios) and has good adaptability to quantity-price deviations.