The world’s energy system is increasingly complex and distributed due to increasing renewable energy generation, decentralization of energy resources, and decarbonization of heavy industries. Energy producers are challenged to optimize operational efficiency and costs within hybrid power plants generating both renewable and carbon-based electricity. Grid operators have less time to dispatch energy resources optimally, while ensuring grid reliability and resiliency for homes and businesses.
Siemens Energy has developed AI surrogate models using NVIDIA Modulus, an open-source framework for building, training, and fine-tuning physics-informed machine learning (physics-ML) models. These surrogate models are for complex engineering systems that will power industrial digital twins of modern power grids. The systems include bushings used in power grids and gas-insulated switchgears (GIS). A critical transformer component, the bushing is an insulating structure that facilitates the passage of the current-carrying conductor through the grounded tank.
This post details how Siemens Energy built their AI expertise and leveraged accelerated physics-ML approaches at scale for complex grid systems.
Digital twin for transformer bushings
Siemens Energy is a global leader in energy technology and development that’s advancing the clean energy transition. AI is a key enabler to optimize power generation, improve the global energy supply chain, enhance energy demand predictions, and increase power grid operational flexibility. However, autonomous, interconnected power systems powered by AI can only be achieved by simulating physically accurate assets in power generation and transmission infrastructure for predictive equipment health and asset management. Siemens Energy is developing an AI surrogate for transformer bushings to be used in power grids for near real-time prediction of hotspot temperatures under varying operating and ambient conditions.
AI surrogates for transformer bushings training strategy
Siemens Energy trained a neural network using NVIDIA Modulus for heat generation within the conductor, heat propagation across insulating and supporting mediums, and heat loss due to cooling medium and ambient conditions. The model acts as a virtual sensor to predict the temperature profile of bushing assets under real-world operational conditions, such as load and ambient temperature, and saves costs from expensive sensor-based monitoring systems.
This approach used physics-informed neural networks (PINNs) to solve the static heat conduction problem within a complex multi-domain setup. This includes multiple solid-solid and solid-fluid interfaces, with each domain having unique thermal properties.
Using the virtual temperature sensor from the bushing digital twin, grid operators gain near real-time insights on temperature warnings and maximum loading indication. This occurs even under varying operating conditions, prevents thermal runaway mishaps, and reduces electricity waste due to overheated transformers.
Figure 1. Complex dense sampling using NVIDIA Modulus across various components of a transformer bushing, as a prerequisite to training PINNs
Results
Siemens Energy used 3D design files to capture precise geometric representations of components within each bushing asset. The team used a domain-specific neural network training approach, where heat transfer physics within asset components were captured separately and linked together using elliptical interface conditions. The AI models were parameterized during training based on input load and ambient conditions available with NVIDIA Modulus. Once training was complete, thermal profile inference calls for unseen load and ambient operating conditions were processed in less than one second, enabling near real-time asset health diagnostics.
To illustrate, Siemens Energy used an example of a bushing asset that connects high-voltage power grid transmission lines and grounded transformers. The Figure 2 inset shows the thermal profile of the transformer bushing at specific operating conditions. By monitoring thermal profiles, a grid operator can quickly assess and pinpoint hotspot temperatures. Further analysis determines whether temperatures are within prescribed limits set by local or national transmission grid standards.
Figure 2. Transformer with bushings connected to power lines at a 145 kV substation
By using digital twins of transformer bushings, customers can operate their grid assets according to hotspot temperature, rather than fixed-rate current, to dynamically increase or shed loads. Grid operators can run grid assets under hotter ambient conditions at lower current in a controlled manner, instead of being faced with expensive, unplanned downtime for equipment repairs. This improves grid reliability, reduces blackouts, and could help lower electricity costs for end customers.
The bushing twin operates under the maximum recommended rated conditions—less than 4% maximum error margin (<3.5℃) from internal measurement data—and is currently undergoing customer field-evaluation trials.
Digital twin for the gas-insulated switchgear
A gas-insulated switchgear functions as a “big switch” for routing or breaking high-power currents. It uses a type of gas (SF6, Clean Air) as an insulating medium for flashover prevention. This gas, often pressurized, is filled into a gas-tight metal casing that encapsulates the current-carrying conductors. Related to GIS is an air-insulated switchgear, where atmospheric air alongside a larger separation between electrical conductors acts as insulation. GIS is more compact and typically used in areas with limited space, such as cities.
Figure 3. A Siemens Energy Blue GIS using clean air instead of SF6, a potent greenhouse gas (left), and a section of a GIS depicting the layout of components (right)
Siemens Energy is actively developing digital twins for GIS using an AI-based surrogate model. The twin is intended to predict the transient thermal behavior of a switchgear in operation, taking into account different GIS operational conditions and configurations for short-term overloads.
AI surrogates for GIS training strategy
To operate a GIS safely, the operator must ensure that the temperature in each of the different components stays below the respective design limits. The electrical current passes through the conductors, and due to power loss, a fraction of the electrical power is converted into heat, which increases the temperature throughout the asset.
These temperature limits are factored into the switchgear design, and result in certain operation boundaries. However, those boundaries usually result from rather conservative steady-state considerations and don’t take into account the transient nature of the temperature dynamics. This means that there are scenarios where the operation boundaries can be safely breached for a certain amount of time (short-time overload).
To accurately assess potential short-time overload scenarios, the developed surrogate model must be able to predict the temperature dynamics in the solid components of the switchgear from a given state for several hours into the future with small error margins. Additionally, the model must be able to generalize across combinations of electrical current magnitudes and ambient temperatures.
Due to the complexity of the switchgear components with different material and thermal properties, and particularly the transient nature of the problem, an optimized graph neural network (GNN) available in NVIDIA Modulus was chosen for the architecture of the surrogate model. The training data for this data-based method was generated using CFD simulations. One particular challenge was capturing all modes of heat transfer across conduction, advection, and radiation.
Figure 4. A GNN-based training loop for the AI surrogate model for GIS simulation
Figure 5. The temperature field and derived hot-spot temperature prediction of the GNN model for a small section of the GIS
Results
In the initial proof-of-concept phase, a model for a subsection of the GIS was developed. This graph model, simplified from a complex geometric mesh with exact specifications, consisted of several thousand nodes and edges. Data was generated by running transient CFD simulations for different combinations of current and ambient temperatures, where each simulation spans 10 hours of operation, and amounted to hundreds of gigabytes of data. This generated data was then preprocessed (smoothing, compressing low-information regions, and so on) for the training of the graph model on a an NVIDIA multi-GPU setup using message-passing algorithms.After training, the model was able to accurately predict the rollout temperature dynamics for 10 hours of operation given an unseen ambient temperature and current. The complete inference rollout took less than 2 seconds with an error margin of under 0.3K on a single GPU. For comparison, this is 10,000x faster than the transient-CFD solver with minimal loss in accuracy. With additional optimizations, the size of the graph and required training data were further reduced without a significant increase in error.The development and field validation of a surrogate model for an entire switchgear system is ongoing.
Figure 6. The hot-spot temperature prediction of the GNN model on test data (left), and the error between the CFD-based prediction and the GNN prediction (right)
Summary
Siemens Energy started developing digital twin assets for individual grid assets. Currently, the company is developing prototypes for bushings and GIS, which will be validated by customers. Siemens Energy plans to couple individual asset twins into system digital twins and, later, integrate into grid twins that span multiple domains and geographies.
To learn more, watch the on-demand GTC session, Enabling More Resilient Grids with Physics and Data-Driven ML Virtual Sensors with Siemens Energy. And explore NVIDIA Modulus for building AI surrogate models with physics-ML.