针对污水处理工艺强耦合、数据非线性及高波动性的挑战，通过预测曝气量实现实时调控，从而降低生化段曝气能耗。基于此，提出了一种基于完全集成经验模态分解与自适应噪声（complete ensemble empirical mode decomposition with adaptive noise, CEEMDAN）、长短期记忆网络（long short-term memory, LSTM）和Transformer的复合模型（ceemdan-LSTM-transformer, CLT），用于精准预测曝气量。首先使用CEEMDAN将曝气量分解成若干个新的分量，再将分量与原始变量合并送入LSTM，捕获分量和原始变量的深层次的时序信息，接着引入Transformer模型并行处理数据，关注序列的所有部分，更好的理解数据中长距离依赖关系同时提高模型精度和泛化能力；最后，以实际工艺数据进行消融、对比、泛化实验，实验结果显示，CLT模型的决定系数R²达到99%，在泛化试验的决定系数为98%，泛化能力优异，该模型能够实现曝气量的精准预测，为降低曝气能耗提供了方法和参考。

To address the challenges of strong coupling of wastewater treatment processes, data nonlinearity and high volatility, real-time regulation is achieved by predicting the aeration amount, thus reducing the aeration energy consumption in the biochemical section. Based on this, hybrid model termed CLT (CEEMDAN-LSTM-Transformer) was proposed to achieve precise prediction of aeration volume. This model integrates CEEMDAN (complete ensemble empirical mode decomposition with adaptive noise), LSTM(long short-term memory) networks, and the Transformer architecture. First, the aeration volume was decomposed using CEEMDAN to generate new components. These components were combined with the original variables and input into LSTM to capture deep temporal information. Then, the Transformer model was applied for parallel data processing, focusing on all parts of the sequence to improve understanding of long-distance dependencies. Finally, ablation, comparative, and generalization experiments were conducted using actual process data. The coefficient of determination (R²) of the CLT model reaches 99%, and the coefficient of determination in the generalization test is 98%. Superior generalization abilities are demonstrated. The model effectively achieves precise prediction of aeration volume, providing methods and references for reducing aeration energy consumption.