Machine learning-based surrogate model assisting stochastic model predictive control of urban drainage systems

Area of Science:

• Environmental Engineering
• Hydrology
• Artificial Intelligence

Background:

• Urban drainage systems (UDSs) require robust real-time control to manage stormwater inflow uncertainty and enhance resilience.
• Current uncertainty-addressing methods for UDSs are computationally intensive, limiting their practical application.
• Accurate quantification of stormwater inflow uncertainty is crucial for effective UDS operation.

Purpose of the Study:

• To develop a computationally efficient machine learning-based surrogate model (MLSM) for UDSs.
• To integrate MLSM with stochastic model predictive control for a novel real-time control strategy.
• To minimize expected system overflow by considering stochastic stormwater inflow scenarios.

Main Methods:

• Developed a machine learning-based surrogate model (MLSM) to emulate high-fidelity urban drainage dynamics.
• Combined MLSM with stochastic model predictive control for real-time operational strategy.
• Generated and reduced an ensemble of stormwater inflow scenarios using statistical methods.

Main Results:

• MLSM achieved a 99.1% reduction in computational time compared to the original model while maintaining high fidelity.
• The proposed MLSM-based stochastic control strategy significantly outperformed deterministic control in enhancing system resilience.
• The strategy demonstrated effectiveness for real-time operation in a complex UDS in China.

Conclusions:

• The MLSM significantly reduces computational burden, making advanced control strategies feasible for real-time UDS management.
• The developed real-time control strategy effectively enhances UDS resilience against stormwater inflow uncertainty.
• This approach offers a promising solution for improving the operational efficiency and reliability of complex urban drainage systems.