Accurate prediction of wastewater treatment plants (WWTPs) influent quality can provide valuable decision-making support to facilitate operations and management. However

since existing methods overlook the data noise generated from harsh operations and instruments

while the local feature pattern and long-term dependency in the wastewater quality time series

the prediction performance can be degraded. In this paper

a discrete wavelet transform and convolutional enhanced Transformer (DWT-CeTransformer) method is developed to predict the influent quality in WWTPs. Specifically

we perform multi-scale analysis on time series of wastewater quality using discrete wavelet transform

effectively removing noise while preserving key data characteristics. Further

a tightly coupled convolutional-enhanced Transformer model is devised where convolutional neural network is used to extract local features

and then these local features are combined with Transformer's self-attention mechanism

so that the model can not only capture long-term dependencies

but also retain the sensitivity to local context. In this study

we conduct comprehensive experiments based on the actual data from a WWTP in Shaanxi Province and the simulated data generated by BSM2. The experimental results show that

compared to baseline models

DWT-CeTransformer can significantly improve the prediction performance of influent COD and -N. Specifically

MSE

MAE

and RMSE improve by 78.7%

79.5%

and 53.8% for COD

and 79.4%

70.2%

and 54.5% for -N. On simulated data

our method shows strong improvements under various weather conditions

especially in dry weather

with MSE

MAE

and RMSE for COD improving by 68.9%

48.0%

and 44.3%

and for -N by 78.4%

54.8%

and 53.2%.