Abstract: In order to improve the short-term wind speed prediction acccuracy, a novel wind speed forecasting method is proposed. This method uses historical wind speed and meteorological data as inputs and first utilizes feature selection networks to quantify the importance of different features at each time step in the input sequence. Local temporal features are then captured by the multi-scale causal convolutional network. After that, self attention is introduced to integrate features from different convolutional layers, generating a high-dimensional feature sequence which can reflect multi-scale characteristics of the wind speed. Finally, BiLSTM is used to extract long-term temporal features of the high-dimensional feature sequence to obtain forecasting results of wind speed. Experimental results demonstrate that the proposed method can consider the dynamic effects of different input features on wind speed and fully extract both local and long-term temporal features of the wind speed sequence. The normalized root mean square error and mean absolute error of the proposed model for one-hour ahead wind speed forecasting are 11.92% and 8.11%, respectively. The proposed method also achieves a high prediction accuracy with correlation coefficient of 0.9735 and coefficient of determination of 0.9477, which improves the accuracy of short-term wind speed forecasting effectively.