Towards good modelling practice for parallel hybrid models for wastewater treatment processes

Area of Science:

• Environmental Engineering
• Wastewater Treatment Technologies
• Computational Modeling

Background:

• Mechanistic models in Water and Resource Recovery Facilities (WRRFs) often have limitations in representing complex biological processes.
• Data-driven models struggle with extrapolation beyond their training data.
• Hybrid models offer a promising approach to leverage the strengths of both mechanistic and data-driven methods.

Purpose of the Study:

• To formulate and evaluate parallel hybrid models (HMs) for WRRFs by integrating mechanistic and data-driven components.
• To investigate the impact of calibration effort on mechanistic models versus compensation by neural networks (NNs).
• To compare the performance of different data-driven models, specifically Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNNs).

Main Methods:

• Constructing parallel HMs by training NNs on the residuals of mechanistic models for effluent nitrate prediction.
• Utilizing the Benchmark Simulation Model no. 1 and a large pilot-scale WRRF for experimentation.
• Comparing HMs developed with different calibration/training datasets and mechanistic model versions.
• Testing LSTM and CNN as the data-driven components within the HM framework.

Main Results:

• The parallel HM effectively addressed limitations in mechanistic model predictions and data-driven model extrapolation.
• Training the NN on an independent validation dataset yielded superior results compared to using the calibration dataset.
• The most effective HM utilized a mechanistic model with default, uncalibrated parameters.
• The CNN-based HM achieved a root-mean-squared error (RMSE) of 1.58 mg NO3-N/L, significantly outperforming the LSTM-based HM (RMSE = 4.17 mg NO3-N/L).

Conclusions:

• Parallel hybrid models demonstrate significant potential for improving effluent nitrate predictions in WRRFs.
• Optimizing the balance between mechanistic model calibration and NN compensation is crucial for HM performance.
• Convolutional Neural Networks (CNNs) are a highly effective data-driven component for hybrid modeling in this context, outperforming LSTMs.