基于MPSO–CNN–LSTM–Attention模型的海温预测OA

To enhance the prediction performance of ocean temperature,a hybrid model integrating a modified par-ticle swarm optimization(MPSO)algorithm,a convolutional neural network(CNN),a long short-term memory(LSTM)network,and an attention mechanism,namely,the MPSO–CNN–LSTM–Attention model,was proposed and established to predict the ocean temperature in time series.First,the variation characteristics of temperature data were extracted using the CNN;second,the extracted features were used to form a time series as the input for the LSTM;third,the MPSO algorithm was introduced to establish a dynamic nonlinear inertia weight and a bimodal feedback learning factor to determine the optimal value of the key model hyperparameters and thereby improve prediction accuracy and convergence speed;finally,information characteristics output by the LSTM layer were captured by the attention mechanism,the weight ratio was adjusted,key information was extracted,and the ocean temperature prediction was thus realized.A comparison of the time-series prediction results,based on historical temperature data,with other neural network models,in terms of three indicators—Root Mean Square Error(RMSE),Mean Absolute Error(MAE),and Coefficient of Determination(R²)—shows that the MPSO–CNN–LSTM–Attention model converges faster,has smaller prediction errors,and performs relatively better overall.