An Interval Autocorrelation Mix-Up of Data Augmentation Based on the Time Series Prediction for Wastewater Treatment Model

The abuse of chemical agents in wastewater treatment is very universal. However, accurate predictive models capable of addressing this issue rely on precise and abundant data. Limitations in sampling frequency and equipment often lead to insufficient data, resulting in model overfitting. To address this, the IA Mix-up data augmentation algorithm, based on correlation coefficient-weighted mixing, has been proposed. By ranking the temporal correlation of water quality data and incorporating error weight mixing into original labels, the algorithm adjusts mixed label weights according to the temporal characteristics of the original signal, preserving time-series correlation. Experimental results demonstrate an average 8.75% improvement in prediction accuracy across four neural network models, with R²-e reduced to 1–5%. Among the four prediction models, the LSTM model has the highest prediction accuracy of 89%. Compared with existing time-series data augmentation methods, IA Mix-up enhances the r value by 9.5%, improves prediction accuracy by 7%, and reduces the training-validation prediction error by 3.67%. These results indicate that the proposed algorithm effectively mitigates overfitting and enhances model performance. In actual use, the total phosphorus in the effluent meets the Class I effluent standards while saving 33% of polyaluminum chloride.