A Spatiotemporal Deep Learning Approach to Improve Real-Time Effluent Quality Prediction in Wastewater Treatment Plants

Integrating hydraulic topology and temporal dynamics for more accurate and interpretable wastewater quality forecasting

Wastewater treatment plants (WWTPs) are inherently dynamic systems.

The characteristics of incoming wastewater can change substantially over time, while treatment itself involves a sequence of interconnected physical, chemical, and biological processes. These variations make it difficult for plant operators to anticipate changes in final effluent quality—yet accurate real-time prediction is critical for process optimization, operational decision-making, and regulatory compliance.

Deep learning offers a promising approach to this problem. By learning patterns from large volumes of operational data, data-driven models can predict future water-quality conditions without explicitly simulating every biochemical reaction occurring inside a treatment plant.

However, this approach also introduces an important limitation.

Many existing models understand how wastewater quality changes over time, but not how wastewater actually moves through the treatment plant.

Our research addresses this gap through a new spatiotemporal wastewater quality prediction (STWQP) framework, designed to simulate hourly effluent Total Nitrogen (TN) concentrations by integrating both the temporal dynamics and physical structure of wastewater treatment.

Experimenting in Lab

Research Motivation

Traditional data-driven wastewater prediction models predominantly rely on time-series relationships.

Given a sequence of historical measurements—such as influent characteristics and operational parameters—a model learns patterns in those measurements and uses them to predict future effluent quality.

This can be effective, but it simplifies the internal structure of a WWTP.

Wastewater does not move directly from influent to effluent. It passes through a sequence of treatment units, with the condition of water leaving one process influencing what enters the next. These units therefore have explicit hydraulic flow directions and topological relationships.

When a model considers only temporal correlations, much of this spatial information is lost. The plant effectively becomes a black box: the model knows what enters, observes historical operational patterns, and predicts what eventually comes out, but has a limited representation of the physical relationships occurring in between.

Our study began with a different way of representing the problem:

What if the structure of the treatment plant itself became part of the model?

Process Topology Construction

To incorporate this information, we represented the wastewater treatment process as a directed spatial graph.

Schematic representation of treatment processes and on-line instruments in the selected WWTP

In this representation, individual treatment units become interconnected components, while the direction of wastewater flow determines how those components are linked.

This distinction matters.

Two water-quality measurements collected from different treatment units are not simply two independent variables in a dataset. Their relationship is partly determined by the treatment process itself. If wastewater flows from one biochemical unit into another, changes upstream can propagate downstream.

By encoding these connections explicitly, the model can learn not only statistical correlations among variables but also the structural dependencies between sequential treatment units.

The resulting framework therefore provides a more process-aware representation of the WWTP rather than compressing the entire treatment system into a conventional time series.

Analyzing Micropollutants and Microbial Levels in Water Samples

Spatiotemporal Framework Development

Representing spatial relationships solves only half of the problem.

Wastewater treatment is also strongly dependent on time. Influent characteristics fluctuate, treatment conditions evolve, and changes within one unit may affect downstream water quality over subsequent hours.

The STWQP framework therefore combines spatial modeling with sequential memory mechanisms capable of tracking dynamic operational patterns.

In simplified terms, the model asks two questions simultaneously:

Where is a change occurring within the treatment system?

and

How is that change evolving over time?

The spatial component captures relationships among interconnected treatment units, while the temporal component learns how water-quality and operational conditions develop across sequential measurements.

Rather than simply stacking independent algorithms, the framework integrates these two dimensions into a unified representation of wastewater treatment.

The objective was to determine whether explicitly modeling both process topology and temporal dynamics could provide more reliable predictions of hourly effluent TN concentrations than approaches relying on either dimension alone.

Organizing Data and Building the Spatiotemporal Prediction Model

Sketches of the deep learning model for WWTP effluent prediction.

Model Validation

We evaluated the framework using real-time operational data from a municipal wastewater treatment plant, with effluent Total Nitrogen serving as the primary prediction target.

TN is particularly useful for evaluating the framework because nitrogen removal depends on interconnected biochemical processes occurring throughout wastewater treatment. Accurately predicting its final concentration therefore requires the model to capture complex relationships across both treatment stages and time.

The STWQP framework demonstrated substantially stronger predictive performance than the stand-alone algorithms used for comparison.

Across model evaluations, it achieved an accuracy improvement of 38.4%–62.8%.

More importantly, this improvement was not restricted to predicting only the immediately following time step.

Multi-Step Forecasting

For a prediction system to become useful in plant operation, forecasting what happens in the next hour is only one part of the problem. Operators may also need to anticipate how effluent conditions will develop several hours into the future.

We therefore evaluated the model through multi-step forecasting from 1 to 12 hours ahead.

The framework maintained high predictive accuracy across these forecasting horizons, demonstrating robust generalization beyond immediate one-step prediction.

This matters operationally because longer forecasting windows can provide more time to identify potential changes in effluent quality and adjust treatment conditions before those changes reach the final discharge.

Rather than functioning only as a tool for describing current plant conditions, real-time forecasting can potentially support more proactive process management.

WWTP effluent multi horizon forecast performance by the STWQP model for (a) MAE of training data sets, (b) MAPE of training data sets, (c) MAE of test data sets and (d) MAPE of test data sets

Interpretability Analysis

Higher accuracy alone, however, does not completely solve one of the central limitations of deep learning.

Complex models are frequently criticized for being difficult to interpret. A highly accurate prediction is less useful for process understanding if operators cannot determine how conditions within the treatment system contributed to that prediction.

The explicit topology incorporated into STWQP provides an important advantage here.

Because the model preserves relationships among different treatment units rather than collapsing the plant into a single black box, it provides a stronger basis for evaluating intermediate water-quality variations between different biochemical structures.

This makes the framework useful not only for asking:

What will the final TN concentration be?

but also for investigating:

How are changes within the treatment process contributing to that outcome?

The distinction is important for real-world wastewater management. Plant operators do not simply need predictions; they need information that can help them understand where problems may be developing and how treatment processes might be adjusted.

Findings

Our results suggest that the physical structure of a wastewater treatment plant should not be treated as background information.

It is part of the data.

By combining hydraulic topology, temporal dynamics, and deep learning, the STWQP framework achieved higher prediction accuracy, maintained robust performance across 1–12-hour forecasting horizons, and improved the interpretability of intermediate treatment processes.

The framework also helped mitigate problems such as feature degradation and overfitting, supporting more reliable generalization when modeling complex wastewater-treatment dynamics.

Ultimately, real-time effluent prediction should do more than produce an accurate number. For a model to support wastewater treatment in practice, it must reflect enough of the underlying process to make that prediction meaningful.

By modeling not only how water quality changes over time, but also how water actually moves through a treatment plant, spatiotemporal learning offers a path toward wastewater prediction that is both more accurate and more transparent.