WC142 MayJune 2025 - Magazine - Page 34
INNOVATION
What are AI and ML?
AI is a broad term that refers to systems designed to replicate
or augment human decision-making. These systems may use a
combination of tools, including rule-based logic, optimization
algorithms, and ML. Think of AI as the full decision-making
framework—while ML is just one tool within that system.
ML itself is a type of model, similar in purpose to the kinetic-based mechanistic models commonly used in wastewater process
design and simulation. However, instead of relying on biochemical reaction kinetics or stoichiometry, ML models learn patterns
What AI and ML can do in wastewater treatment
WWTPs generate massive datasets—ranging from flow rates
and pump statuses to sensor-based water quality metrics and
laboratory results. These datasets differ in frequency (hourly,
daily, weekly), format (numerical, categorical), and relevance
to operational decisions. As a result, much of this valuable
information remains underutilized, and operators are often left
overwhelmed rather than empowered.
This is where AI and ML can help. By uncovering patterns
and relationships in complex datasets, ML can support more
adaptive, efficient, and insight-driven operations across the plant.
One of the most impactful
ADVANTAGES
CHALLENGES & OPPORTUNITIES
applications is process optimization, where ML enables real-time
improvement of core treatment
REINFORCEMENT LEARNING
WORKFORCE READINESS AND TRAINING
processes by recommending
optimal setpoints based on current
operating conditions. For example,
PROCESS OPTIMIZATION
INTERPRETABILITY
ML can support dynamic chemical
dosing for phosphorus removal
by adjusting feed rates based on
SOFT SENSING
DATA QUALITY
influent loading and target effluent
concentrations—reducing chemical
PREDICTIVE MAINTENANCE
PERCEPTION OF FULL AUTOMATION
use and sludge generation. It can
also optimize polymer dosing for
sludge dewatering by predicting the
AI and ML can make signi昀椀cant contributions by helping plants move from reactive operations to data-driven,
relationship between polymer adadaptive decision-making.
dition, feed solids, and equipment
performance. Aeration—a major
energy consumer in WWTPs—can
directly from historical data—such as flow rates, environmental
also be optimized, improving energy efficiency. These targeted
conditions, and water quality parameters measured through senstrategies can even extend to energy load shifting and plant-wide
sors or routine lab measurements—and use those patterns to make
optimization.
predictions.
Importantly, ML process optimization does not need to inLike mechanistic models, ML models can be used as standalone
volve direct automated control. In many cases, ML can operate
“soft sensors,” providing insights that support better operational
in an advisory capacity, where it serves as a decision-support
decisions. But they do not act on their own—they must be embedtool for operators. By offering data-driven recommendations
ded into a broader control or decision-support system.
or visual alerts, ML enhances the operator’s ability to interpret
For example, in an AI-based aeration control system, ML
changing conditions and make timely, informed decisions.
might be used to predict future ammonia concentrations based
In other words, ML is not a replacement for operational
on real-time upstream conditions. The AI system can then
expertise—it is an extension of it.
use this prediction, along with engineering constraints, to
Beyond optimization, ML also enables soft sensing, where
recommend or adjust dissolved oxygen (DO) setpoints. In this
it estimates variables that are difficult, time-consuming, or
setup, ML serves as a forecasting engine, while AI governs the
expensive to measure directly—such as oxygen transfer effioverall control strategy.
ciency (OTE), biofilm thickness, greenhouse gas emissions, or
In short, ML enhances operator awareness and process responalkalinity and volatile fatty acids (VFA) in anaerobic digestion.
siveness. It doesn’t aim to replace domain knowledge or automaThese insights enhance process visibility and decision-making
tion logic, but rather to complement them—unlocking the value
without requiring additional instrumentation. In the context
of data that is often underutilized.
of greenhouse gas (GHG) emissions, ML models can dynami-
34
WATER C AN ADA • M AY/JUNE 2025
WAT E R C A N A D A . N E T
