To solve the problem of low accuracy of the fan state prediction model when the power plant load changes, a dynamic integrated state prediction method based on convolutional neural network ( CNN), long short-term memory ( LSTM) network and attention mechanism (AM) is proposed. Firstly, the CNN is used to divide the sample data into different classes with overlapping boundaries to achieve soft classification of wind turbine operating conditions. Then, the AM layer is introduced into the traditional LSTM network. LSTM-AM networks as sub-learners are established under different work conditions. The soft classification labels output by CNN are used as the initial weights, and the genetic algorithm is used to search for the optimal weight bias. Finally, the output of each sub-learner is multiplied with corresponding weights and summed to obtain the integrated prediction value, which could improve the prediction accuracy under different operating conditions of power plant fans. The experimental results show that, compared with each LSTM-AM sub-model and signal LSTM-AM model, the proposed CNN-LSTM-AM dynamic integrated model can reduce the relative mean square error by 11. 5% and 22. 3% when power plant fans are operating under variable loads. Results indicate that the model has better robustness and applicability