为提高径流预测精度，研究提出海洋捕食者算法（MPA）与长短期记忆（LSTM）神经网络相结合的径流预测方法。通过6个仿真函数对MPA、粒子群优化（PSO）算法进行测试，利用MPA优化LSTM隐藏层神经元数、训练次数等关键参数，基于主成分分析（PCA）降维和不降维处理分别建立PCA-MPA-LSTM、MPA-LSTM径流预测模型，利用云南省落却站实测数据对PCA-MPA-LSTM、MPA-LSTM模型进行训练及预测，结果与PCA-LSTM、LSTM、PCA-MPA-SVM、MPA-SVM、PCA-MPA-BP、MPA-BP模型的训练、预测结果进行比较。结果表明：①MPA仿真效果优于PSO算法，具有较好的寻优精度和全局搜索能力。②PCA-MPA-LSTM、MPA-LSTM模型对实例拟合、预测的平均相对误差分别为1.18%、2.35%和1.94%、1.96%，预测效果优于其他6种模型，具有较好的预测精度和泛化能力。③采用MPA优化LSTM关键参数能有效提高LSTM泛化能力和预测精度；数据降维模型的预测精度优于对应未降维模型的预测精度，数据降维处理能有效改善模型的预测效果。

To improve the accuracy of runoff prediction, research and propose a runoff prediction method that combines the marine predator algorithm (MPA) and the long and short-term memory (LSTM) neural network. Test MPA and particle swarm optimization (PSO) algorithms through 6 simulation functions, use MPA to optimize key parameters such as the number of neurons in the hidden layer of LSTM, training times, etc., and establish PCA based on principal component analysis (PCA) dimensionality reduction and non-dimensionality reduction processing. -MPA-LSTM and MPA-LSTM runoff prediction models, using the measured data from the Luoque Station in Yunnan Province to train and predict the PCA-MPA-LSTM and MPA-LSTM models, and the results are consistent with PCA-LSTM, LSTM, PCA-MPA-SVM, Compare the training and prediction results of MPA-SVM, PCA-MPA-BP and MPA-BP models. The results show that: 1) MPA simulation effect is better than PSO algorithm, with better optimization accuracy and global search ability. 2) The average relative errors of PCA-MPA-LSTM and MPA-LSTM models for instance fitting and prediction are 1.18%, 2.35%, 1.94%, and 1.96%, respectively. The prediction effect is better than the other 6 models and has better prediction Precision and generalization ability. 3) Using MPA to optimize the key parameters of LSTM can effectively improve the generalization ability and prediction accuracy of LSTM; the prediction accuracy of the data dimensionality reduction model is better than that of the corresponding non-dimensionality reduction model, and the data dimensionality reduction processing can effectively improve the prediction effect of the model.