Press this link to see the recording of the session.
Moderator: Hans Alderink (Deltares)
The speakers of this session are: Burcu Yazici (Turkish Water Institute (SUEN), Istanbul/TURKEY), Alex Van der Helm (Waternet) and Lluis Echeverria (Eurecat Technology Centre, Unit of Applied Artificial Intelligence and Universitat de Lleida)
Smart Tools for Efficient Water Use and Water Reuse in Agriculture
Presenting author: Burcu Yazici
Konya Closed Basin (KCB) is one of the hot spots in Turkey due to intense agricultural activities and high irrigation demand. Agricultural areas cover around 60% of the basin, and water thirsty crops (e.g. alfalfa, corn, etc.) are among the most widely grown crops. Over 90% of the total available water is consumed for irrigation and is mainly met from groundwater bodies, since surface water resources are scarce and largely seasonal. Current water demand of agriculture alone is 4725 hm3, whereas the total water potential of the basin is only 4430 hm3. Unsustainable irrigation methods and unlicensed wells have led to groundwater depletion (> 30 m) over the past decades which led to formation of numerous sink-holes in the basin. On top of it, climate change models predict 20-30% drop in precipitation for the semi-arid basin. Therefore, current agricultural practices need to be revised and water reuse should be taken into account in order to sustain the water resources and ecosystem of the basin. To this end, a case study is being carried out in KCB as part of the WATERMED4.0 project. Briefly, the project promotes smart farming to enhance efficient use of conventional and non-conventional water resources in the Mediterranean agriculture via IoT technologies. The pilot study in KCB aims to achieve the following results: evaluation of the suitability of using reclaimed water for root/herbaceous plants, enhancement of irrigation efficiency via modern irrigation technologies and smart agriculture tools (sensors, etc.), and last but not least, assessment of socio-economic aspects.
The study is testing two of the most widely grown and most water-consuming plants in the basin: corn and sugar beet. Half of each crop area is irrigated with reclaimed water, and groundwater is applied to the other half. Considering lack of presence of modern irrigation techniques and excessive groundwater withdrawals in the basin, drip irrigation system is employed for both type of crops to allow for a comparison of water use efficiency. Weather and soil conditions are continuously monitored by the deployment of a climate station and soil moisture sensors which are directly linked to the WATERMED platform. Continuous monitoring as well as historical climatic data will be utilized to prepare irrigation plans in order to improve water use efficiency of crops. Additionally, soil and plant analyses will allow for developing an optimum fertilization plan, cutting down excess use of nutrients. Once the first-term results are received from the study area, it will be possible to assess how efficient the actual management of water and fertilizers are, and define the gaps that need to be optimized to improve productivity, which will be tested further in the second growing season (2022). The next step will be to provide recommendations not only on efficient and optimum water/fertilizer use, but also to deliver best management practices for water reuse in irrigation, apart from providing user-friendly interfaces for end users.
Artificial Intelligence for Wastewater Treatment – The development of a digital twin used to train a control optimization agent
Presenting author: Alex van der Helm
Optimising operations in large Wastewater Treatment Plants (WWTPs) is strategically important for water utilities to reduce operational costs (especially energy consumption), to reduce the climate footprint, and to meet ever more stringent water quality targets. For Waternet, the water cycle utility of Amsterdam and surrounding areas, minimising nitrous oxide (N2O) emissions (a potent greenhouse gas (Law et al., 2012)) from wastewater treatment is expected to contribute significantly to the further reduction of Waternet’s climate footprint.
Waternet’s Amsterdam West WWTP has a capacity of ± 1 Million population equivalent and serves the city of Amsterdam. In order to explore minimisation of N2O emission and energy consumption for wastewater treatment while meeting discharge limits, Waternet has dedicated one of the seven identical treatment lanes of Amsterdam West WWTP as research lane. At the research lane, additional sensors have been installed next to existing sensors. Use of these sensor data streams and data fusion will enable new data driven control strategies and decision support based on AI. These technologies are expected to improve operational efficiency and asset management, in order to reduce climate impact and costs (Yuan et al., 2019).
The presented research discusses the data preparation with anomaly detection and correction (Seshan et al., 2021), as well as a recurrent neural network (RNN) architecture that serves as a digital twin of the WWTP research lane, against which a control agent is trained. The control agent learns to set a controller setpoint of the WWTP research lane (Echeverria et al., 2021). The RNN architecture used for the digital twin is a Gated Recurrent Unit (GRU) architecture, and is trained to predict the near-future evolution of sensor time series (e.g. N2O off-gas sensors, energy meters, flow meters, water quality sensors, in the influent, in the aerobic tanks and in the effluent), see Figure 1. The sweet spot in the trade-off between the size of the training set and validation error was found at a time series duration of three months using a 15 minute sampling interval, see Figure 2. As of yet, the seasonal pattern cannot be accounted for due to data availability constraints (only one year of data is available). Additionally, the sensitivity of predictions to large, but realistic deviations in the input parameters was investigated. It was found that the prediction of the digital twin is insensitive to deviations of a single parameter, however, a significant change in the predicted output occurs when all input parameters are varied at once. This is favourable for the prediction reliability, but challenging for a control agent that determines an operational strategy (control setpoints) to minimize short and medium-term energy use and N2O emission
Artificial Intelligence for Wastewater Treatment – AI-agent- based service for optimal control
Presenting author: Lluis Echeverria
The way in which control processes are carried out at a Wastewater Treatment Plant (WWTP) is not only crucial because of their direct impact on final effluent quality, but also because of their effect on the WWTP operational costs and direct and indirect greenhouse gas (GHG) emissions. WWTP processes use a considerable amount of resources, mainly energy and chemicals, produce residual sludge, and GHGs are emitted at several process stages. Thus, given the unceasing societal and economic growth in urban and industrial ecosystems, the need for more and more efficient WWTPs concerning GHG emissions, use of resources, effluent quality and costs is of utmost importance.
In this vein, our research aims to develop an AI-agent-based service, able to learn control strategies that can be implemented for more optimal WWTP control. The research is conducted in Waternet’s Amsterdam West WWTP, a full-scale WWTP with a capacity of ± 1 million population equivalent. In this scenario, the targeted process for the control service is the Activated Sludge Process , specifically its aeration stage. Usually, this is the most critical process in a WWTP, which has a direct impact on the effluent water quality and GHG emissions, demanding considerable quantities of energy. The aeration stage is controlled by setting an O2 setpoint, therefore, through its adequate calibration, optimal oxygen setpoints can be obtained, consequently minimizing GHG emissions and energy use.
We propose the use of Reinforcement Learning (RL) algorithms , a branch of the Machine Learning domain aimed at developing agents for optimal and autonomous control. These algorithms are based on a computational approach to understand and automate goal-driven learning and decision-making. Furthermore, by combining RL with the capabilities of Deep Learning as a knowledge generalization engine, Deep Reinforcement Learning (DRL)  algorithms become a powerful tool to operate and manage large, high-dimensional and complex continuous environments.
The AI-agent-based service is trained on the Waternet WWTP Digital Twin . In this setup, the DRL reward signal includes N2O emissions (a powerful GHG and the largest contributor to the GHG emissions of WWTP Amsterdam West), energy consumption, and effluent NO3 and NH4 levels, which are translated to costs (€) for its
combination. DRL agents are provided with a short- and mid-term wastewater influent flow forecast, which allows them to perform predictive control strategies to adapt the WWTP complex dynamics to future situations. Additionally, agents’ policies are implemented through a Recurrent Neural Network architecture, based on Long Short Term Memory units, which provides the agent with the ability to memorize past conditions, leading to a better process dynamics understanding and improved decision making. Finally, an initial Transfer Learning stage is aimed to supply the control service with an expert baseline knowledge (provided by the WWTP Digital Twin), thus improving the learning process.
As a conclusion, a benchmark of different DRL algorithms (i.e. DQN, PPO, A2C/A3C, etc.) and operational modes (continuous vs discrete action spaces) will be carried out to discover the most effective AI-agent-based service configuration able to provide a WWTP optimal control strategy.