Product expert and business developer Niels Tiemessen’s last blog examines how different types of algorithm can improve automation solutions for industrial wastewater processing.
Smart algorithms are the heart of any type of automation solution. Even before the invention of the transistor, knowledge about the process was converted into if-then logic that was being hardwired using relays to operate pumps, engines and valves. Fast forward a hundred years, and process control has come on leaps and bounds, especially with the rise of Industry 4.0 using Internet of Things (IOT), Big Data and artificial intelligence (AI). However, the fundamentals are still the same. Although ultimately a means to an end – by mitigating operational risks for industrial clients – find out more about the use of algorithms within Twinn Aqua Suite for Industrial wastewater processing.
The fundamental for any type of process control is online measurement. In recent years, not only has more and more data become available, like flow, pH, and pressure, but there are also newer measurements finding their way to end users, like Total Organic Carbon (TOC). This opens new opportunities, especially in situations where the composition may be harder to predict or softsensor may not be possible.
At Aqua Suite, we are firm believers in fit-for-purpose algorithms. This means first going for the simplest, least computing or data intensive solutions, with the biggest impact, before moving towards black-box type of technologies. This simplifies the software and the requirements for lengthy training of AI. on big data sets. It also makes it much easier to explain to the end user what the decision process is, instilling trust from the people that need to use the software. Of course, for more complicated, dynamic, fuzzy and multi-variable applications, like anomaly detection, there is no other way, than to use neural networks. The following types of algorithms may be used within Aqua Suite applications:
operators like if-then or more basic feedback loops, like PI, PD or PID. These algorithms heavily depend on deep water domain knowledge and can traditionally also be found in SCADA/PLC programming. These algorithms can be great, but are always reactive, meaning that the systems are always playing catch-up. An example of this is to turn off a pump when a certain level in a buffer is reached.
these are algorithms that learn linear, or other types like exponential relationships between pre-set variables like load and aeration needed. These still depend on fundamental knowledge, but since the precise relationships can be different per reactor and depend on a lot of variables, the algorithm autonomously learns the exact relationship. An example of this is the relationship between the amount of polymer needed and the incoming dry-solids for sludge dewatering. Because of age, maintenance or other factors, decanter treating the same sludge might need slightly different amount of chemicals to reach the same result.
when taking a closer look at different types of data, you discover that many processes in the man-made water cycle follow patterns. An example is the diurnal consumption of drinking water and the subsequent flow to wastewater treatment plants. By using only a few weeks of historical data, an accurate prediction can be made up to 72h ahead, which in turn, can be used for feed-forward control, by anticipating what is to come instead of playing catch-up.
these types of advanced algorithms are becoming increasingly famous because of applications like ChatGPT. Neural networks can autonomously learn the relationships between many input parameters and prediction outcomes. This is especially useful with complex systems where typical domain knowledge would require a lot of modeling. Despite providing very useful insights, for example, by providing anomaly detection on pumps or drift detection for sensors, these models require large amounts of computing power and data. As this system is hard to explain and users require retraining when the system changes, this technology tends to be avoided for control type of applications, because of its lack of robustness for control type of applications.