To address the lack of systematic exploration of error generation mechanisms in traditional wind power forecasting methods

which limits effective model improvement

an ultra-short-term wind power forecasting method based on error path analysis and differentiated parameter update is proposed. A baseline forecasting model is first established

using convolutional neural networks (CNNs) as feature encoders for each input feature

combined with feature attention and temporal attention mechanisms to enhance the spatio-temporal feature representation capability. Subsequently

model forecasting errors are obtained

and a decision tree model is trained using the input feature set and these errors to identify feature combinations that significantly contribute to forecasting accuracy. Based on this

a differentiated parameter update strategy driven by error information is developed to enable adaptive optimization of model parameters

thereby maintaining the learned feature representation while enhancing the ability to mine potential information. Case studies demonstrate that the proposed method achieves superior forecasting accuracy compared to baseline methods. Further analysis of feature weight distributions validates the effectiveness of the method in promoting more balanced feature importance and enhancing physical consistency.