A power generation plant with unconventional systems requires a delicate balance in its control systems to function properly.
The plant is managed by a team of engineers who take actions on various actuators such as valves and pumps. These actions have direct implications on the systems (pressure, temperature, and flows changes) but also indirect implications on adjacent systems.
It is crucial to maintain certain values within safety ranges (pressure and temperature), while also aiming to achieve maximum performance. Additionally, external factors cause temperature variations, which, in turn, affect pressure. It is highly complex to keep all systems running at full capacity without incidents, which is why the control engineering team is quite extensive, and managing their shifts is not trivial.
The plant is fully monitored with multiple sensors, and every action is recorded in the system. There is a two-year historical data available.
A control optimization system is proposed, which learns from the engineers’ actions on the system to recommend the best course of action for each case while adhering to safety ranges.
This control optimization model is based on simulations using past data and uses these simulations to test multiple scenarios and learn quickly. With this model, it is expected that the recommendations will be at least as good as those made by the engineers, and preferably better in terms of safety.
Proof of concept
The proof of concept starts by recommending the optimal value for just one of the valves. After data cleaning, rules regarding safety ranges are modeled, and simulations are generated using the existing historical data.
The proof of concept demonstrates that an optimal valve opening value can be recommended based on sensor values. This value closely aligns with what the engineers provide, and when it deviates, it is usually to minimize risk and strictly adhere to safety values, even if it leads to a loss in performance.
The engineers acknowledge that if they had followed the recommendations, some minor incidents could have been avoided.
The evaluation is positive, but it is noted that the model should take a bit more risk to improve performance, even if it slightly exceeds safety parameters.
A second version of the proof of concept is delivered, which achieves better performance but with higher risk. However, it remains sufficiently cautious. The feedback for this version is very positive.
Minimum viable product
The minimum viable product is developed, which provides recommendations for most of the actuators, although there is a percentage that still lacks recommendations.
All the previous steps are followed, including simulation, model training, and testing. A preliminary version is delivered with the option to receive feedback from the engineers.
After several months of use, all the feedback data is collected, along with general comments. The model is adjusted accordingly, and the final version of the minimum viable product is delivered.
Final version and outcome
A few more actuators are added, although not covering 100% of them, and the model is retrained.
This version also provides software that interacts with the system, applying the recommendations. These recommendations can be overridden by the engineers.
Minor incidents have been reduced by 70%, and major incidents by 90%. Performance averages are similar, if not better, although it is challenging to guarantee this because some recommendations are overridden by engineers.
Fewer engineers are now needed to keep the plant running, resulting in more relaxed shifts, allowing them to focus on innovation tasks to continuously improve the plant.
As always, to maintain industrial secrecy, the data shown here is altered and/or anonymized. Therefore, everything is based on real facts except for the parts that we have completely made up
If you find this use case interesting and believe it could be applied in your case, do not hesitate to schedule a meeting with us.