[Drift Correction Method of Wastewater Treatment Model Based on Transfer Learning Across Time Scales]

Data is the core foundation of intelligent operation and maintenance, but currently, there is generally insufficient data for wastewater treatment plants, and the status of wastewater treatment systems dynamically evolves with the changes in the internal and external environment. The intelligent operation and maintenance of wastewater plants face difficulties in modeling and model drift caused by system evolution. In response to this issue, the summer and winter seasons with significant differences in wastewater temperature, wastewater quality, and microbial status were selected as typical comparison scenarios. The mechanism model was combined with neural networks to establish a wastewater treatment model drift correction method based on cross-time scale transfer learning. Firstly, in response to the problem of insufficient data, an activated sludge model （ASM） was established and calibrated. Summer operating data was used as input to simulate and calculate operating parameters and effluent data, generating a simulated operating dataset to achieve data augmentation and quality improvement. This was used to train a multi-layer perceptron neural network （MLP） model. The results showed that the average simulation accuracy of the MLP model for summer effluent COD, ammonia nitrogen, total phosphorus, etc., was all over 95%. This indicates the feasibility of training MLP models based on ASM-generated data. Then, the MLP model was used to guide the operation of the pilot A²O project. Experimental data analysis showed that the model drift phenomenon was significant in the field of wastewater treatment. During the operation of the pilot plant guided by the summer model, the accuracy of the predicted values gradually decreased, and the average prediction accuracy of the model for effluent COD gradually decreased from 98.14% to 75.18%. The phenomenon of model drift required effective correction to maximize the effectiveness of the model. In response to the problem of model drift caused by a significant decrease in simulation accuracy in winter operating conditions, a transfer learning approach was introduced. The winter measured data was used as the target domain dataset, and the MLP model was used as the pre-trained model for transfer learning. The experimental results showed that transfer learning methods can significantly improve model performance. After transfer learning, the average simulation accuracy of the MLP model for effluent COD, ammonia nitrogen, total nitrogen, and total phosphorus was relatively improved by 28.58%, 184.44%, 207.56%, and 100.51%, with absolute improvement values of 21.49%, 60.79%, 58.14%, and 46.74%, respectively. This indicates that the cross-time scale transfer learning method proposed in this study can significantly improve model performance, effectively solve model drift problems, and achieve a model-following response to the dynamic evolution of wastewater treatment systems. This study indicates that transfer learning based on pre-trained models only requires a small amount of engineering data and computational complexity to achieve model updating and correction. Compared to model retraining, this method reduces computational complexity and reduces the dependence on engineering data during data-driven model updates.