Aiming to address the problem of low prediction accuracy and slow convergence speed of Elman neural networks in predicting soil heavy metal content, an adaptive evolutionary model was proposed in this study. This model was based on the Elman neural network and used Bayesian regularization to optimize the objective function of the Elman neural network, improving the prediction accuracy of the network model. To overcome the shortcomings of slow convergence speed and susceptibility to local extremum in network models, the Adaptive Gray Wolf algorithm(AGWA) was used to optimize the initial parameters of the network model. And an outlier detection method based on entropy weight distance was used to remove outliers from the data, in order to reduce the interference of outliers on the prediction results. The prediction experiment was conducted using data on heavy metal content in farmland soil collected by Wuhan Academy of Agricultural Sciences. The average absolute error and average absolute percentage error of AEM model for predicting heavy metal content were 1.623 and 17.48%, respectively. Compared with Elman’s comparative model, the determination coefficient index improved by 0.394. After conducting comparative experiments with five different prediction models(AEM, Elman, SBGRNN, SIDIM, CBSA-WNN), it was found that the AEM model had the highest accuracy in predicting soil heavy metal content. The results of the ablation experiment indicated that the three improvement points(Bayesian regularization optimizing, AGWA optimizing, and removing outliers from the data by using an outlier detection method based on entropy weight distance) all contributed to improving the accuracy of predicting soil heavy metal content to varying degrees.