Aiming at short-term wind power time series, a class of dictionary-based sparse coding prediction method is proposed. The historical wind power time series data are composed to the time-lagged input-output data pairs to build the prediction model, while the two dictionaries are respectively constructed by regarding the all input and output data vectors as atoms, without the training phase of the model. For the time-lagged test input data vectors need predicting, convex optimization algorithms with norm or elastic net regularization using sparse decomposition technology are applied to calculate the weights by sparse coding. Furthermore, by the dictionary which is constructed using the historical output data, the corresponding predicted outputs are obtained. Simultaneously, adaptive dictionary updating strategies are given, where test data is successively added to the dictionary in real time while maintaining the dictionary size by using three algorithms, so as to further improve the prediction accuracy of the model. In order to verify the effectiveness of this method, different sparse coding methods are firstly applied to the prediction of Santa Fe chaotic time series, and then applied to the indirect prediction for short-term wind power respectively. Under the same condition, compared with the SVM method, experimental results show that different sparse coding methods have achieved good prediction results, and the sparse coding methods using elastic net regularization have higher prediction precision and show their effectiveness. This work is supported by National Natural Science Foundation of China (No. 51467008).