Skip to main content

Control of a wastewater treatment plant

Development of a reinforcement learning structure for the control of a wastewater treatment plant


Hansewasser deals with the wastewater disposal of the city of Bremen and operates the Seehausen wastewater treatment plant, one of the largest wastewater treatment plants in northern Germany. The wastewater treated by the Seehausen sewage treatment plant reaches 130,000 cubic meters every day. The treatment process removes 95% of the phosphorus, 84% of the nitrogen compounds and 99% of the carbon compounds from the wastewater. With a stable and automated control process, the effluent meets the strictest water pollution regulations and maintains efficient systems for sewage sludge treatment.

The aeration system for activated sludge process:

The typical activated sludge method consists of aeration tanks, settling tanks and the sludge return system, aiming to remove organic wastewater constituents, the nutrients nitrogen and phosphorus.

The aeration system in the activated sludge process consumes at least 50% of the total energy in a wastewater treatment plant. A precise aeration control is a crucial point to improve operational efficiency and reduce energy waste. At the bottom of the aeration tank, compressed air is pumped into the sewage in the form of fine bubbles by air diffusion devices. The aeration system maintains a certain dissolved oxygen concentration of the sewage.  Also it stirs the liquid so that the particles are in a suspended condition, which enhances the quality of circulation.

Project aim:

The activated sludge process will be optimised with the aim of reducing energy use, and improving operational efficiency.

With the wastewater treatment plant being complex and nonlinear, a reinforcement learning strategy is under research to control the aerobic volume.

First, a dynamic model is simulated to predict the behaviour of the wastewater treatment process and collect data of water quality parameters (e.g. dissolved oxygen, phosphate, nitrate, ammonium).

Then a fuzzy system is developed to represent the prior knowledge into a set of constrains (network topology), hence reducing the optimization search space.

In the end the reinforcement learning strategy is utilised to learn the relations between process conditions, nitrous oxide emissions and energy use to propose more optimal control settings for the treatment plant based on the predicted future states and sensor data.