基于PSO-SABO-CGA的养殖水氨氮预测模型研究

Ammonia Nitrogen Prediction Model for Aquaculture Water Based on PSO-SABO-CGA

摘要:

氨氮含量是影响水产养殖水质的重要指标之一,其精准预测对水产养殖管理具有重要意义。 针对氨氮浓度变化受多环境因素影响、预测精度不足的问题,提出一种基于粒子群(Particle swarm optimization,PSO)-减法平均优化器(Subtract average-based optimizer,SABO)改进的卷积神经网络-门控循环单元-注意力机制(Convolutional neural network gated recurrent unit attention,CGA)预测模型(PSO-SABO-CGA)。 该方法通过融合 PSO 优化算法的全局最优引导机制,增强 SABO 算法在超参数空间中的全局探索与收敛能力,并利用该改进算法对 CGA 模型的关键超参数进行自动寻优。 试验结果表明,该研究提出的 PSO-SABO-CGA 模型在养殖水氨氮浓度预测中表现优异, 其平均绝对误差、均方根误差和决定系数分别达到 0. 062 mg / L、0. 082 mg / L 和 0. 934;相较于基线模型,PSO-SABO-CGA 模型在短期与长期预测中均显示出更高的精度与稳定性,当预测时间为 24 h 时决定系数仍高达 0. 842,为养殖水体氨氮浓度的精准预测提供了可靠方法。

Abstract:

Accurate prediction of ammonia nitrogen (ammonia-N )concentration is critical for water quality management in intensive aquaculture. However, ammonia-N dynamics exhibit pronounced nonlinear characteristics due to coupled environmental factors, making precise forecasting challenging. To address this issue, a hybrid prediction model integrating a convolutional neural network-gated recurrent unit-attention (CGA), optimized by a particle swarm optimization (PSO)-enhanced subtraction average- based optimizer (SABO), was proposed. Key environmental variables influencing ammonia-N dynamics were identified by using Pearson correlation analysis, with dissolved oxygen, nitrate-nitrogen, pH value, and water temperature selected as input features. To overcome hyperparameter optimization challenges, the PSO-SABO algorithm was developed by integrating the global memory mechanism of PSO into the update framework of SABO to enhance global exploration and accelerate convergence. Experimental results demonstrated that the proposed PSO SABO CGA model achieved superior prediction performance. For the 5-min prediction horizon, it attained a mean absolute error (MAE)of 0. 062 mg / L, a root mean square error (RMSE)of 0. 082 mg / L, and a coefficient of determination (R2 )of 0. 934. Comparative analyses revealed that PSO-SABO achieved faster and more stable convergence than the hippopotamus optimization (HO)and genetic algorithm (GA), reducing MAE and RMSE by up to 24. 26% and 25. 19% , respectively. Ablation studies confirmed the effectiveness of the GRU component and the necessity of PSO global guidance. Compared with the temporal convolutional network (TCN), bidirectional long short-term memory (BiLSTM ), backpropagation neural network (BPNN ), and standard GRU, the proposed model reduced MAE by up to 15. 07% for the 5-min horizon and maintained an R2 of 0. 842 for the 24 h horizons. These findings indicated that the PSO-enhanced global search, combined with SABO's adaptive mechanism, significantly improved hyperparameter optimization. This enabled the framework to effectively capture nonlinear ammonia-N dynamics under multi-factor interactions, offering superior accuracy, robustness, and generalization across both short-term and extended prediction horizons. This approach can provide a reliable technical foundation for intelligent water quality monitoring and precision management in aquaculture systems.