The data set used in wind power prediction may have different degrees of data loss. Since the processing of missing values is often independent of prediction model training

it is impossible to fully use the time-series correlation characteristics of actual data to improve the prediction effect. Therefore

an ultra-short-term wind power prediction method is proposed considering dynamic interpolation of missing values of time-series features. Aiming at the problem of missing values in time series data

a gated recurrent unit embedded with a time-delay attenuation interpolation strategy is designed to dynamically capture the irregular time-delay relationship between the observations before and after the missing values in time series of the input feature. The autocorrelation analysis with mask determines the optimal time window length of the input features and the initial parameters of the time-delay attenuation rate function. Based on the temporal information extracted by the gated recurrent unit

the sequence-to-sequence prediction structure is constructed to coordinate the inconsistency in input feature dimensions between the historical and the prediction time points

and the wind power prediction sequence is output in the next 15 min~4 h. The experimental results show that the proposed method has higher prediction accuracy and more stable prediction performance than the traditional missing value processing and prediction methods in the case of time series features with missing values for wind power prediction.