A wind speed prediction method based on the combination of multi-channel long short-term memory network and convolution neural network is proposed to improve the prediction performance of wind speed. The prediction model is composed of multiple long short-term memory sub-networks and convolution networks. Future wind speed values are calculated by each sub-network with the different input data length, which can avoid difficultly selecting the input data length in the single prediction network. Convolution and max-pooling operations of the calculation results of sub-networks are performed by the convolution network, and the prediction values of the wind speed are output by the fully connected layer of the convolution network. In order to overcome the accumulation and drift of prediction errors, the dynamic error compensation method is used to correct the prediction values. The network can be used in the ultra-short-term prediction of wind speed. Simulation results show that the network can better fit the variation trend of the actual wind speed series than the existing prediction networks based on deep learning, and the network shows better prediction performance.