What is the difference between a Data + AI Digital Twin and a Static, Equation based model?
Static, equation-based model
Scientists and engineers have been constructing equation-based models to predict the output/performance of physical systems since accessible computing became commonplace about 50 years ago.
Fundamental traits of static models:
- Designed by expert with deep technical knowledge of its working principles
- ‘Average’ model – design is actually based on component specification – using chemistry, physics, and mathematics rules. So, for any potential physical system, the model can predict the performance of an idealized version of the system.
- Once designed, the model can predict the system’s performance within the range of conditions envisaged by the technical expert, but is otherwise static.
- By its nature, a static model cannot adapt. Therefore, any operating condition unaccounted for during design phase – fails to predict the performance of its physical counterpart in real time/field conditions
- Model may be developed prior to the existence of the corresponding physical system
- Model is usually, not affected adversely when exposed to inaccurate or faulty data
Digital Twin Machine Learning Model
With the proliferation of AI (artificial intelligence) over the last fifteen years, models have been designed based on the technology, using data.
With inputs based on substantial data, and outputs measured by sensors, creates a new design of these machine learning models, also known as adaptive models or Digital Twins. Provided both inputs and desired outputs are known, the designer does not require deep technical knowledge of the principles behind the physical system that is being modelled. But having expert level insight into the physical counterpart’s principles does help with input selection, model architecture selection, and optimization of performance of the digital twin model.
Digital Twin models broadly have these characteristics:
- Model structure is designed by an ML expert
- Model performance is improved by training it with ever larger amounts of data, for extended time, and careful selection of the model structure and configuration parameters (hyperparameters)
- Inaccurate or faulty data in the training set can significantly degrade the model
- Over-fitting the model to the available training data can reduce its generality (applicability to the real, field conditions)
Advantages of Machine Learning based Digital Twin Models
- Adaptability – These models can adapt to component tolerances, transportation and installation damage to components, local environmental conditions, local geographic conditions, damage history, changes in performance modes, and finally in changes in the nature of faults occurring over time
- Continuous learning – Can learn performance patterns driven by principles which are unknown to (or dismissed by) the model designer
- Models can change their nature as the physical system ages, or as environmental conditions change
For solar PV plants with a large number of very similar physical systems, individual models can be trained for each of the systems. These unique models are trained to recognize the fact that, given the same inputs, the performance of their corresponding physical systems vary due to their tolerances and operational history.
Why are Digital Twins important for Solar Asset Owners and Managers?
- With geographic distribution of assets, the need to monitor, and control plants remotely calls for a digital model that is tuned to local environmental conditions
- As the proportion of electrical energy produced by solar increases, (~4% now, projected by some to be 30% by 2030, 40% by 2040) demands for accurate forecasting and energy injection by grid operators also increase. Highly accurate, high fidelity plant models, built up from individual Digital Twins of sub-systems are the need of the hour, and the upcoming decades.
- Growing plant sizes mean more focused effort on part of the expensive, limited human engineering resources, on acute problem areas, and not on normally operating parts. The industry has realized the value of system specific models rather than single, simple idealized plant models
- A specific characteristic of PV plants is that with age, its behaviour changes substantially, and in many ways. This creates a need for an adaptive model like a digital twin
Quadrical Ai’s Digital Twin Technology
Digital Twins are at the core of our asset management solutions – CMMS, Snap-On Advanced Analytics, and Plant Audits.
We can leverage Digital Twins to create individual models for each sub-system making up your plant.
They allow us to:
- Identify any of the subsystems (e.g., inverter group, SCB group, or even an individual panel string) is behaving abnormally.
- Understand the difference between the actual performance of the underperforming subsystem (obtained from the live sensors), and the ideal performance of the subsystems if configured or operating correctly (obtained from a digital twin model performing correctly).
- Use our platform to accurately pinpoint the level in the hierarchical structure of the plant where the error is occurring, by comparing actual sensor measurement at multiple levels of the plant, with the predicted outputs that should be occurring based on the digital twin models for the corresponding components
- Compare model performance of the panel when clean, with actual performance of the somewhat dirty panel. This gives both an estimate of the current losses and a projection as to when the costs involved in cleaning the panel should be incurred
Using Time-Series models, Classification models and combining with weather forecasts, we can also:
- Identify progressive degradations in individual subsystems,
- Identify the nature of some problems experienced with associated data patterns
- Combining with weather forecasts, predict near term (the period the forecast covers) energy output of the plant
Most Digital Twin models (models based on ML technology) are based on boosted decision tree or neural network models. After sensible initial architectural choices have been made, accuracy of the models depends on the accuracy data used for training and the quality of the practitioners.
At Quadrical, we are using a combination memory models, forecasting models, and model ensembles to improve the accuracy of error reporting, and of forecasts, by adjusting weights associated with input data, and weights associated with output estimates based on confidence and accuracy predictions associated with input data and the analysis process.
- Confidence can be developed through running multiple Digital Twins at different levels of the plant hierarchy and comparing results.
- Similarly, running multiple differently architected models for the same physical component helps improve confidence that a deviation from the actual physical system measurement – if a genuine fault (all Digital Twin measurements are consistent while the physical measure diverges), versus an issue with the Digital Twin output (cases where the physical measure and some Digital Twin measures align, while the output of one Digital Twin varies).
Anomaly models can help weight training data so data that has high confidence is weighted significantly, while more problematic data can be represented and used for training exotic events, but does not degrade the models owing to using low weights.
Quadrical is also investigating using more powerful models (e.g., LMUs and GANs) that provide more insight into the nature of the faults that our Digital Twin models are reporting. For example, GANs are capable of distinguishing between normally occurring outlier performance, and the abnormal measurements resulting from a fault or failure.
Purpose-Specific – How is a Digital Twin designed to specifically look for particular loss types?
Our technology compares four measures of performance
- Sensor measurements from the actual physical system (e.g., panel string power outputs, current measures at the SCB, etc.)
- Corresponding measures from a Digital Twin for that actual physical system
- The difference in measurements from (1) and (2), allows us to identify when the panel itself is experiencing problems – problems more substantial than dirt on panel need to be addressed
- Clean panel Digital Twin reproduces the performance of a panel under the same environmental conditions (e.g., irradiance measures, temperature, etc.) if the panel were clean. The difference between (3) and (1) gives an estimate of the energy lost owing to dirt on the panel. This is how a ‘soiling DT’ gets built.
We also use Digital Twins to monitor reliability of panel cleaning data used to train Digital Twins in (2) and (3) above. Specifically, classification models can be trained to identify the occurrence of a panel cleaning event (planned, or unplanned like rain).
Finally, we use degradation models to predict the average soiling degradation occurring in sectors of the solar plant for those customers who want a regular cleaning schedule, which does not clean all parts of the plant at the same frequency.
Future Outlook for Digital Twin Technology
- Use of advanced machine learning models (deep and wide networks, generative adversarial networks, ODENets, etc)
- Ensembling of digital twins to enhance confidence in model architecture
- Confidence (of predictions) and Causality
- Generating data for simulating or testing responses to infrequent events
- Determining data sets where substantial system performance changes occur
- Precursor models (e.g., weather models) to generate complete data sets
- Ability to simulate system response to events that have not yet occurred
- Robustness of response – Quick responses driven by models
- Generation of test environments
Modeling dangerous behaviours (energy storage systems)
I would like to acknowledge help from Shreyasi Halder and Kitty Chachra in creating this post.