The large-scale integration of wind power to the grid poses some potential challenges to the power system. Accurate wind power forecasts reduce the impact of the nonlinearities and volatility of wind power generation. A two-channel deep learning model based on an improved coot optimization algorithm (ICOA) is proposed. First, the use of long short-term memory (LSTM) builds a channel for the extraction of chronological characteristics of historical power. Second, a temporal convolutional network (TCN) is adopted as a hybrid feature-extraction channel for multi-dimensional meteorological data, and by incorporating the self-attention mechanism (SA), the ability of TCN to extract internal information from meteorological features is enhanced. Finally, the ICOA that introduces nonlinear decision factor, adaptive dynamic boundary, and Cauchy mutation is used to optimize the model hyperparameters. The simulation analysis is carried out on the winter and summer measured data of a wind farm in Xinjiang. The results show that compared with the traditional LSTM model, the root mean square error and the mean absolute error of the proposed method are reduced by 10.35 % and 16.27 % on average, respectively, and the prediction accuracy is higher than that of other comparative models, which verifies the superiority of our proposed model.