Probabilistic Intraday Wastewater Treatment Plant Inflow Forecast Utilizing Rain Forecast Data and Sewer Network Sensor Data

Abstract

Forecasting of wastewater treatment plant inflow dynamics constitutes an enabler technology for wastewater treatment process optimization using model predictive control. However, accurate inflow prediction is still challenging, especially for strong rainfall events, where complex system dynamics and missing information on future rainfall represent limiting factors. We propose a seasonal probabilistic time series model for modeling the short‐term wastewater inflow accurately while providing quantification of forecast uncertainty. To ensure suitability for practical implementation, the unconstrained parameters of the predictive distribution are modeled as linear functions of the input variables in the framework of Generalized Additive Models for Location Scale and Shape. Non‐linear effects are approximated by Rectified Linear Units, accounting for buffering within the sewer network and flow‐dependent catchment response time. In addition to water level measurements from within the sewer network and rain rate measurements, rain forecasts are incorporated as exogenous regressors, where historical rain forecasts are used for model calibration. The model performance is evaluated on historical data from a German wastewater treatment plant using deterministic and probabilistic scoring rules. We benchmark against an autoregressive time series model and a long short‐term memory artificial neural network. Our results show that the proposed model unites the benefits of high prediction accuracy of the neural network and enhanced intelligibility of the autoregressive model, but accurate real‐time rain forecasts are mandatory for successful real‐world implementation.