Session B3 - Tools, models and real-time control

Stream flow regimes are typically negatively affected by urbanisation and climate change, causing floods, loss of baseflow, habitat degradation, and thus threatening aquatic ecosystems. Stormwater control measures (SCMs) have been applied to deliver, among other objectives, more natural flow regimes. Model Predictive Control (MPC), a real-time control (RTC) strategy, can be used to find the optimal control actions for SCMs, by predicting flow dynamics under different weather and operating scenarios. However, optimising scenarios from a hydrological and hydraulic model in MPC is computationally intensive, especially where short time-step decisions are required. Constructing a simplified predictive model that can be both accurate and rapid is needed for real-time flow regime management. Recent research has shown that a hybrid model that combines physically-based models and deep learning techniques has potential to improve the accuracy and interpretability of flow predictions, with comparable computational speed to data-driven models. In a case study conducted in the Monbulk Creek catchment in eastern Melbourne, Australia, both the proposed hybrid model and deep learning model generate more accurate and faster flow predictions than a physically-based model (SWMM; stormwater management model). The hybrid model will be further improved by integrating more physical mechanisms, and will then be investigated to see if it can inform more adaptive prediction and control of distributed SCMs.

Real time control strategies can be classified into heuristic and optimization-based approaches, with the latter being very demanding in terms of computational requirements and being operationally complex, posing challenges for their field implementation. As a result, heuristic approaches are being applied more widely at the operational level, even if there is a lack of methodologies for formulating optimal rule sets for these approaches. This study presents a framework for formulating a set of rules to control water levels in a stormwater retention basin with a rule-based control approach based on rainfall forecasts. The method uses machine learning regression decision trees to learn the control logic from the optimal behavior of the system, as defined using optimization-based model predictive control. The model was trained by applying model predictive control over 10 historical rainfall events that occurred between 2022 and 2024. Applied to a case study where the control objective was to maintain water elevation into the basin between 15.1 m and 15.4 m, to avoid overflows and damage to infrastructure, the method demonstrated that the derived rules outperformed static control and achieved performance comparable to that obtained through the direct application of model predictive control. Furthermore, the method was effective in extracting control decisions in understandable formats, such as IF-THEN rules, from the behavior of the predictive control model.

A system of four stormwater ponds in Knoxville, TN, USA, were fitted with controllable valves. The operation of the ponds was shown to impact conditions within the receiving water, leading to continued work to understand if and how in-stream objective functions can be realized. These objective functions were quantified and a computational model paired with reinforcement learning was used to determine how the ponds should be controlled to meet these objectives.

Stormwater management aims to protect waterways from habitat degradation but often falls short due to poorly defined objectives. We present a method to define ecologically relevant flow objectives and embed them in real-time stormwater control algorithms. The method involves: (i) identifying the physical habitat needs of key species; (ii) surveying the stream to map habitat features; (iii) using high-resolution hydrodynamic modelling to relate flows to habitat availability; (iv) combining these relationships across sites; and (v) integrating them into stormwater control algorithms. We applied this method to the Monbulk Creek Smart Water Network in Melbourne, Australia, which aims to use real-time-controlled rainwater tanks and urban lakes to support a vulnerable platypus population affected by urbanization and reduced baseflows. By supplementing low flows, the system increases wetted habitat and macroinvertebrate productivity, improving platypus foraging areas and food sources. This approach enables stormwater management to directly target habitat objectives such as inundating larger wood or increasing the area with sufficient water depth for platypus to swim and hide.