Abstract: The variability and randomness of wind energy make power prediction challenging, and traditional methods struggle to fully exploit the underlying features of data when sample sizes are limited. To address this, a TCN-LSTM(Temporal Convolutional Network-Long Short-Term Memory) short-term wind power forecasting method based on improved Generative Adversarial Network sample augmentation is proposed. First, to tackle the issue of insufficient data, a TWGAN-GP ( Temporal Dynamic Maximum Mean Discrepancy-Constrained Wasserstein Generative Adversarial Networks with Gradient Penalty) model with dual constraints of TD-MMD (Temporal Dynamic Maximum Mean Discrepancy) and Wasserstein distance is introduced to generate high- quality wind power time series data for augmenting the training set. Secondly, a TCN-LSTM forecasting model is constructed, and TCN(Temporal Convolutional Networks) is used to capture long-term temporal dependencies, combined with LSTM(Long Short-Term Memory) to extract dynamic features, and to optimize hyperparameters through the RIME(Rime Optimization Algorithm) optimization algorithm. Finally, through case analysis, theproposed method demonstrates a significantly superior prediction accuracy compared to benchmark models. This method effectively resolves the forecasting challenges caused by the randomness and volatility of wind power, overcomes the shortcomings of traditional methods in feature extraction under small sample conditions, and provides a reliable approach for precise short-term wind power prediction.

Key words: wind power generation forecasting, generative adversarial networks, temporal dynamic maximum mean discrepancy, temporal convolutional networks, frost and icing optimization algorithms