Accelerating Battery Charging Simulations with Altair romAI - Part 1
Altair EmployeeIntroduction
As the world shifts toward electric solutions, battery charging technology has become a cornerstone of modern innovation. At the heart of this technology lie power converters - sophisticated devices that manage the crucial task of delivering power to batteries efficiently and safely.
Designing these power converters presents significant challenges. Engineers must rely on high-fidelity simulations to verify their designs and ensure optimal performance. However, a major bottleneck emerges when these validated models need to be incorporated into battery charging simulations, which can be extremely time-consuming to execute.
This is where Altair romAI steps in. By leveraging neural networks to create Reduced Order Models (ROMs) from existing simulation data, romAI offers a powerful solution to this computational challenge. In this tutorial, I'll walk through a practical example: creating and validating a ROM from an LLC resonant power converter and demonstrating how it can deliver instant results in extended battery charging simulations that would traditionally take hours or days to complete.
Model Overview
The foundation of my ROM begins with a detailed LLC resonant power converter model created in Altair PSIM, a specialized tool for power electronics simulation. The accompanying video demonstrates the model's architecture and, crucially, identifies the key signals that will form the basis of the ROM:
- Inputs:
- Reference current Iref: The current setpoint imposed by the controller on the converter.
- Battery voltage Vbat: The voltage level at the battery terminals. This parameter heavily affects the system dynamics.
- Outputs:
- Battery current Ibat: Battery flowing from the converter into the battery terminals.
- Switching frequency fsw: Resonant converters adjust their switching frequency relative to their resonant frequency to control power flow.
- Internal States: In this case, internal states are equal to both outputs. Declaring a signal as a state ensures it keeps continuous behavior and gives additional information about the system to romAI.
This is the resonant LLC converter high-fidelity model, created in Altair PSIM:
This is an example of the system’s response to a step current reference from 8 to 4 A at t = 5 ms:
Note: It’s important to consider the voltage at the battery terminals as an input to the ROM because, even though it is a parameter which normally fluctuates slowly, it heavily affects the dynamics of the system, as can be seen here:
These signals are essential for the ROM training process in romAI Director, as they will be used to replicate the converter's behavior accurately while maintaining computational efficiency.
Generating Training Data through Design of Experiments
While a high-fidelity PSIM model serves as the foundation, creating an effective ROM requires comprehensive training data that captures the converter's behavior across various operating conditions. To systematically generate this dataset, I leveraged Altair HyperStudy to design and execute a Design of Experiments (DOE) study.
The DOE approach ensures the training data encompasses diverse scenarios within the converter's operational design space, capturing how the system responds to different input combinations. Two key parameters were varied in the study:
- Reference current: A dynamic parameter that transitions between two values during simulation, both within the range of 4 - 11 amperes
- Battery voltage: A constant parameter varying between 2.5 and 4.2 volts
To ensure coherence from all simulations run, several constraints were defined to ensure an appropriate value of I_in, the initial current reference, and Istep, the reference change occurring at t = 5 ms. These are the constraints:
- Current reference should stay beneath upper limit: I_in + Istep <= 11
- Current reference should stay above lower limit: I_in + Istep >= 4
- Reference step change should be big enough to produce a meaningful response: |Istep| >= 1
For this case study, we executed 50 simulations to generate our training dataset. One of the significant advantages of the DOE framework is its scalability - we can easily adjust the number of simulation runs (whether 10, 50, 200, etc.) to fine-tune the balance between model accuracy and training data generation time. These are a part of the results of some of the simulations produced by the study:
The accompanying video demonstrates the step-by-step process of configuring the DOE in HyperStudy, showing how we structured the experiment to capture the converter's full behavioral spectrum.
Note: There’s a PSIM connector available for HyperStudy which must be installed separately from PSIM > Utilities > HyperStudy Setup > 2024.1. This connector simplifies running PSIM simulations, identifying design variables and reading results files.