针对燃煤电厂参与调峰负荷波动较大，出口SO₂浓度控制效果不佳的问题，建立了一种基于捕鱼优化算法（catch fish optimization algorithm，CFOA）优化融合神经网络的出口SO₂浓度预测模型。首先使用互信息算法筛选由机理分析得到的特征变量，并通过逐次变分模态分解对筛选后的辅助变量进行分解重构，保留相关性较大的重构分量作为输入变量。随后采用双向时间卷积网络、双向门控循环单元与多头自注意力机制构建融合神经网络模型，通过CFOA对模型超参数寻优以进一步提高精度。最后使用某660 MW燃煤电厂历史运行数据进行对比实验，实验结果表明，该模型在出口SO₂浓度剧烈波动的工况下仍能实现较好的预测效果。同多种模型对比，该模型具有更小的误差和更高的预测精度，体现出其在复杂变化环境中的鲁棒性和可靠性。

To address the issue of severe operational fluctuations during peak load regulation in coal-fired power plants and the poor performance of traditional SO₂ concentration control at the outlet， a prediction model for outlet SO₂ concentration was developed. This model integrates a neural network optimized by the catch fish optimization algorithm. First， auxiliary variables derived from mechanism analysis were selected using the Mutual Information algorithm. These selected variables were then decomposed and reconstructed using sequential variational mode decomposition， retaining only the components with high correlation as input features. Next， a fusion neural network model was constructed by integrating a bidirectional temporal convolutional network and a bidirectional gated recurrent unit with a multi-head attention mechanism. The CFOA algorithm was employed to optimize the model’s hyperparameters to further improve accuracy. Finally， a comparative experiment was conducted using historical operational data from a 660 MW coal-fired power plant. The experimental results demonstrate that the proposed model can achieve satisfactory prediction performance even under conditions of severe fluctuations in the outlet SO₂ concentration. Compared to the benchmark models， this model exhibits lower prediction errors and higher accuracy， reflecting its robustness and reliability in complex and dynamic environments.