污水处理厂出水水质的准确预测对于优化运行控制和保障达标排放具有重要意义。研究提出双向长短期记忆网络(BiLSTM)与注意力机制相结合(BiLSTM-注意力机制模型)的深度学习模型,用于预测污水处理厂出水化学需氧量(COD)。该模型通过引入位置编码增强时序信息表达,设计特征注意力机制实现自适应水质参数权重学习,并采用多头注意力机制捕捉时间步间的复杂依赖关系。实验结果表明:BiLSTM-注意力机制模型对出水COD的预测评估指标明显优于其他模型,较BiLSTM模型,均方根误差(RMSE)降低幅度为13.5%,平均绝对误差(MAE)降低幅度为13.8%,平均绝对百分比误差(MAPE)降低幅度为15.3%;模型对前期时间步(0~6步)赋予更高权重,并识别出过程段溶解氧和污泥浓度等关键运行参数对出水COD的显著影响。因此,BiLSTM-注意力机制模型能有效捕捉污水处理系统的非线性时空特征,为出水水质预测和智能化运行管理提供了可靠方法和理论依据。

[Objective] The wastewater treatment process exhibits highly non-linear, time-varying, and multivariable coupling characteristics, making it difficult for traditional prediction methods to effectively capture complex spatiotemporal dependencies. Unidirectional LSTM utilizes only historical information, struggling to fully exploit bidirectional temporal features. This study aims to construct a deep learning model combining a bidirectional long short-term memory network with an attention mechanism to achieve high-precision prediction of effluent COD in wastewater treatment plants. [Methods] This study proposed a deep learning architecture integrating BiLSTM and a multi-layer attention mechanism. The model adopted a hierarchical design. First, sine-cosine positional encoding was used to embed time step position information. A feature attention mechanism was designed to achieve adaptive weight learning for different water quality parameters using a fully connected network and the softmax function. Then, a single-layer bidirectional LSTM structure was employed to simultaneously capture forward and backward temporal dependencies. A multi-head attention mechanism was introduced to capture complex interaction patterns between time steps. Subsequently, a time-step importance weighting mechanism was designed, using a quadratic growth curve to assign higher weights to recent time steps. An attention-gated fusion strategy was used to dynamically combine the LSTM output and the attention output. Finally, the final prediction was achieved through global average pooling and a fully connected network. The model training employed the Adam optimizer, Dropout regularization, L2 regularization, and an early stopping strategy. The prediction performance was compared with baseline models such as unidirectional LSTM, BiLSTM, and 1D-CNN. [Results] Experimental verification showed that the BiLSTM-attention mechanism model significantly outperformed other models in effluent COD prediction. Compared to the BiLSTM model, the root mean square error decreased from 1.17 mg/L to 1.01 mg/L, a reduction of 13.5%. The mean absolute error decreased from 0.92 mg/L to 0.80 mg/L, a reduction of 13.8%. The mean absolute percentage error decreased from 9.79% to 8.29%, a reduction of 15.3%. The validation set loss converged well during the training process. The visualization analysis of attention weights revealed the model’s decision-making mechanism as follows. Feature attention identified dissolved oxygen in the process section and sludge concentration as key influencing parameters. Temporal attention showed that the model assigned higher weights to recent time steps, conforming to the physical laws of time-series prediction, and the different heads of the multi-head attention captured different temporal dependency patterns, achieving complementary feature extraction. [Conclusion] This study successfully constructs an effluent COD prediction model for wastewater treatment plants based on BiLSTM and a multi-layer attention mechanism. The innovations are reflected in proposing a hierarchical deep learning architecture that integrates positional encoding, feature attention, multi-head attention, and gated fusion; utilizing a bidirectional LSTM structure to simultaneously leverage forward and backward temporal information, which reduces the error by over 10% compared to unidirectional models; and designing time-step importance weighting and gated fusion mechanisms to achieve refined modeling of temporal information.