Construction and application of effluent quality prediction model with insufficient data based on transfer learning algorithm in wastewater treatment plants

Prediction of effluent water quality is important for the management of wastewater treatment plants (WWTPs). However, the accuracy of effluent quality prediction is poor owing to a lack of training data, and prediction accuracy is influenced. This study builds an effluent Chemical Oxygen Demand (COD) prediction model trained by insufficient data with influent and biochemical parameters as inputs, which is integrated with transfer learning algorithm Two-stage TrAdaBoost.R2. The model trained by the source WWTP is transferred to the target WWTP through transfer learning, it obtains better performance than boosting algorithm AdaBoost.R2 and de-cision tree algorithm. The coefficient of determination (R2) improves from 0.51 to 0.8, and the mean square error (MSE) decreases from 19.69 to 8.29. The application of Two-stage TrAdaBoost.R2 effectively excavates the data characteristics, which demonstrates the effectiveness of transfer algorithm. Moreover, the transfer algorithm is explored in other effluent indexes (TP, NH3-N, pH), and NH3-N obtains better prediction accuracy than that of TP and pH, possibly attributed to more useful knowledge of NH3-N obtained from source data to get excellent predicting effectiveness. The results indicate that the integrated method proposed in this study can relieve inferior predicting ability caused by insufficient data effectively. Thus, this provides a theoretical basis and technical support for the stable operation of WWTPs.