In order to better monitor the health state of power lithium batteries. A lithium battery health state prediction method based on Improved Bidirectional Temporal Convolutional Network, Long Short Term Memory Network and Attention Mechanism is proposed. The hyperparameters of the proposed method are optimized using Crested Porcupine Optimizer. Tests were conducted on the University of Maryland lithium battery charge/discharge dataset to extract capacity-related health features, and the health features with higher correlation were screened by Pearson correlation coefficient as inputs to the neural network algorithm. The Root Mean Squard Error of the proposed method is no more than 0.020, the Mean Absolute Error is no more than 0.017, and the R-Square is above 0.995 for all battery health state predictions. Higher accuracy can be achieved in lithium battery health state prediction.