To address the limitations of insufficient feature information of characteristic factors, insufficient mining of complex temporal relationships between multiple factors, and low efficiency of model hyperparameter optimization in shrimp breeding and feeding prediction, a long short-term memory network model based on attention mechanism and genetic algorithm optimization (GA-LSTM-ATTN) was constructed. Firstly, based on the core factors such as dissolved oxygen, water temperature, body length and number of shrimp, the growth rate was introduced as a supplementary feature. Secondly, combined with the attention mechanism, the learning ability of the model to the relationship between multiple factors and the feeding rules at different growth stages was enhanced. Then, the genetic algorithm was used to optimize the hyperparameters such as time step, hidden layer dimension, network depth, number of training iterations and batch size before model testing. The results show that the R² (coefficient of determination) = 0.8683, RMSE (root mean square error) = 0.3703 and MAE (mean absolute error) = 0.3311 on the breeding dataset. Compared with the benchmark LSTM model, the R² is increased by 7.3%, the RMSE is reduced by 15.2%, and the MAE is reduced by 13.5%. Compared with the mainstream prediction models, the prediction accuracy of GA-LSTM-ATTN is also improved. In conclusion, the model can effectively improve the accuracy of shrimp feeding prediction, and can provide technical support for accurate feeding in actual aquaculture.
