Advances in artificial intelligence (AI) are making a large impact in many fields, and the water sector is no different. With operators facing constant pressure to boost plant efficiency and to deliver maximum value to the consumer with minimal environmental impact, those that view AI as a critical differentiator stand to reap a wide range of business advantages.
Through these technologies, operators can transform performance without huge investment in hardware or training. Here, the Ada Mode team discuss the application of human-in-the-loop AI to continuously optimise process performance within the water sector.
Optimising performance of industrial plants requires continuous expert tuning of numerous controls and consideration of multiple interacting performance metrics. Machine learning (ML) algorithms can be trained to learn complex behaviours relating to system performance, and then be exploited with optimisation algorithms to automatically suggest control strategies to continuously optimise plant performance. This approach allows plants to improve their efficiency without the need for CAPEX, thereby maximising the value of existing assets.
This process is known as machine learning control and can be successfully applied to many cases across the water industry including:
Reducing energy and chemical consumption of various processes
- Minimise the energy consumption of pumps, blowers, mixers and compressors
- Optimise chemical use to achieve compliance and simultaneously reduce OPEX
Process management of water treatment plant
- Dynamic management of process to control treatment stage outputs based on varying demand and external environmental factors
Biological process optimisation
- Improve odour management accounting for wind speed, direction and ambient temperatures
- Optimise resource recovery including biogas and water reuse
- Optimise aeration to reduce OPEX while improving output quality
Minimising environmental emissions
- Reduce CH4, CO2, N2O release through biological processes and denitrification optimisation
- Reduced effluents of pollutants into the receiving waters
- Reduced costs from nonconformities and limit breaches
Flood and blockage management
- Dynamic control of assets to minimise flooding and overflow events
To implement such technology, a machine learning model must be trained to predict or forecast industrial process performance based on plant control inputs and relevant contextual information. This model can then be expanded with an optimisation process, creating a digital twin of the system which recommends control adjustments to operators, with the aim of optimising system performance.
When implemented on continuous process systems, this approach can simplify plant control, alleviate pressure on experienced system controllers and identify beneficial system optimisations.
AI Support for Plant Operators
The responsibility that plant controllers carry is significant. They must govern an often vast, non-linear and time-dependent system to ensure that multiple performance targets are met. This may require hitting a manufacturing quality threshold, minimising emissions below a target allowance, limiting waste to within enforced/legal bounds, all while ensuring power usage is tolerable. This cycle of continuous multi-dimensional, multi-faceted optimisations must be conducted to a degree that allows the plant to maintain healthy performance across a diverse range of operating modes and external conditions.
Applying machine learning control best practices that guide control policy decision making, can lessen this pressure on plant controllers and enable long-term plant performance improvements. The suggested policies provide an informed second opinion on control strategy, resulting in reduced operator stress, knowledge requirements and risk. The optimisation step is also likely to find control sweet spots, simultaneously unlocking a transformation in performance, reduced emissions, power draw and chemical consumption within key water treatment processes.
Traditionally, system control becomes more challenging during periods of volatile/uncontrollable environmental conditions. However, machine learning control frameworks can consider such relevant contextual information. This compelling advantage enables the digital twin to be robust and consistent during periods where plant controllers are most unsure. Examples of these uncontrollable factors may include; local agricultural activity, ambient temperature and other weather conditions, fluctuations to government emissions thresholds, or variability in water processing plant influent chemistry. By reviewing relevant contextual information the digital twin can provide robust advice all year round in a variety of scenarios, boosting performance during challenging periods.
Multiple Metrics and many Controls
As industrial processes are almost always multi-dimensional and non-linear, machine learning control can handle many parameters within its performance prediction engine, with the only concern being the speed of the optimisation stage.
Balancing trade-offs in performance is a significant challenge for operators. For example, within biogas production, a gas upgrade plant’s performance is a function of gas quality, emissions and power consumption. This process can be optimised by reducing the three parameters to a single cost of operation. This streamlined approach simplifies multi-objective optimisations to a single-target problem allowing for a broad range of optimisation techniques. Optimisation targets can include costs, emissions, waste, output quality, or a combination of factors. Dynamic refocusing of plant performance objectives is also possible, e.g. in response to changes in energy price. By reducing the performance metric to a univariate form, processes can become applicable across many industrial challenges.
Requirements and Limitations
While AI is capable of many great things, it will never replace real human experts. The performance prediction engine can be trained on huge volumes of operational history, but the possibility of the AI making mistakes still remains – particularly when facing novel conditions. Within dynamic industrial processes the development of novel system behaviour is of high probability. Instances such as faulty input materials in manufacturing, record breaking temperatures or workforce restrictions can have unforeseeable effects on the performance of related systems. Machine learning algorithms cannot be expected to perform when extrapolating to new conditions, whereas human expertise is more transferable. Human-in-the-loop AI systems support experienced operators through continuous, sophisticated and automated plant monitoring, alerting operators to issues which require their attention at an early stage.
A Case Study in Anaerobic Digestion
Ada Mode has applied this control strategy to optimise the performance of a gas-upgrade plant for an anaerobic digester. The plant is responsible for refining raw biogas to meet grid standards. Potential savings of the approach in testing reached ~ £100,000 per year and included reductions in ~5,000kg of emitted methane to the atmosphere.
The performance of the process depends on three key variables: CO2 concentration in the processed output gas, CH4 Concentration in emitted gas (methane slip), and the plant’s power draw.
System operators adjust the controls in an attempt to minimise the three metrics simultaneously. A number of feedback loops within the system present a major challenge to on-the-fly optimisation by human operators. In our approach, we designed a utility function to attribute financial costs to each of the metrics, returning a single parameter to be the target of optimisation. This utility function involved calculating the market value of the lost methane, the cost of the propane required to upgrade the gas to grid standards due to excess CO2 in the raw output, and the cost of power drawn in operation. The sum of these three calculated values gives the overall process cost to the gas-upgrade plant.
With a combination of expert domain knowledge, performance forecasting AI and optimisation, a digital twin can be developed to support system controllers by suggesting optimal control strategies. The automated controller can enable consistent performance in varying operating modes and environmental conditions. The methodology can reduce costs, improve quality, limit emissions, prevent waste and more. Applications to a single system of a biogas plant suggest significant performance gains and further applications to the wider water sector are numerous.
To find out more please visit ada-mode.com or email us at firstname.lastname@example.org