Traditional neuralnetwork algorithms applied to post-earthquake reconstruction (for the convenience of the following text, post-earthquake reconstruction is abbreviated as PER) engineering cost models have problems of low convergence, slow operation speed, and low accuracy of engineering cost prediction results. In order to change this situation, this paper applied the improved neuralnetwork algorithm to the PER project cost model, and applied the neural network refined particle swarm optimization method to optimize the initial neural network weight, so as to avoid the local optimization of neural network in the training process. The prediction results of neural network based on particle swarm optimization were compared with those of traditional neural network. Through experimental analysis, this article concluded that the improved neuralnetwork algorithm had a higher accuracy in predicting the cost of PER projects. Its accuracy in predicting the engineering cost of 120 samples was much higher than that of 60 samples. Moreover, when predicting the engineering cost of 120 samples, the error values of different samples were all within 2%. The improved neural network technology can greatly improve the accuracy and stability of engineering cost prediction. The improved neural network technology has greatly improved its performance compared to regression analysis, fuzzy mathematics, grey prediction, and traditional neural network algorithms. The cost model of PER engineering based on improved neuralnetwork algorithms has a very broad application space for PER in the future.
