| Advertorial feature The better path to the Digital Twin
Physics-based modelling is the key to the correct representation of a power plant.
Machine learning (ML) has been a buzz word in recent years also in the power industry, but there are several reasons why the application of this technology has not been as successful as with the manufacturing and other industrial process industries.
One important fact is that power plants interact much more with the ambient than other industrial processes, simply due to the fact that the water-steam cycle must reject surplus heat from the condenser. In addition, the ambient conditions have a significant effect on output and heat rate of gas turbine-based power plants, and thus the entire range of ambient must be reflected suitably by their digital twins. Since combined cycle power plants are also among the most flexible producers, they may be dispatched over a very wide load range fulfilling the target output with varying plant configurations.
The resulting
abundance of possible operating modes is thus practically never covered with sufficient data in the plant historian, so that ML-based models cannot be trained to fully represent the capabilities of the
PHYSICAL MODEL
plant, even more so when it comes to actual technical limits, since ‘normal’ operation should have some margin to such boundaries. Adding the typical scatter in plant sensors, it becomes obvious that a purely data-driven model is likely to be inaccurate, and unpredictable when exposed to exceptional – but possible situations.
Simulated data derived from a good physical model ensure quality and completeness of information.
Since many years, the Austrian expert company ENEXSA focusses on thermody- namic simulation of power plants, and it has developed special distributed calculation technology that allows to manage the execution of very large batches of simulations on server farms. Utilizing EBSILON heat balance models that are initially configured based on vendor data and then tuned to accurately match selected representative historical data, the power plant process can be reliably predicted over the entire range of ambient conditions and operating modes. Due to the scalability of the system, what would have taken months on a single PC can now be accomplished within minutes. With this approach, a complete and consistent set of data
SIMULATED DATA
representing the subject power plant can be provided for the ML process.
Why then use machine learning at all? Simply because for decision making, time is of essence.
A well-trained neural net-based digital twin can produce the same key results as the thermodynamic model about four to five orders of magnitude faster than the heat balance software. For applications to plot accurate operating maps in terms of fuel consumption and capacity limits such as ENEXSA’s Visual Operations Support tool, a single display contains tens of thousands operating points. Optimizing power plant set points for the current production target or finding the optimal schedule for day–ahead or weekly scenarios may have to evaluate even millions of options.
The key quality criterion for ML remains: “I have seen it all”
Whatever you want the ML-based digital twin to predict, you have to make sure that this possibility is properly represented in your learning set. You wouldn’t want to actually expose your plant to massive degradation or even failures just to have them on record. Luckily, simulation can do the trick.
DIGITAL TWIN
Contact Information: Josef Petek, Manager Commercial Operations, ENEXSA GmbH, Parkring 18, 8074 Raaba-Grambach, Austria
www.enexsa.com
josef.petek@
enexsa.com
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34 |
Page 35 |
Page 36 |
Page 37 |
Page 38 |
Page 39 |
Page 40 |
Page 41 |
Page 42 |
Page 43 |
Page 44 |
Page 45 |
Page 46 |
Page 47 |
Page 48 |
Page 49 |
Page 50 |
Page 51 |
Page 52 |
Page 53 |
Page 54 |
Page 55
