Graph Neural Power Simulation for Incomplete Grid Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Electrical power grid simulations are computationally expensive and often rely on incomplete or inaccurate models, leading to inadequate results, as physics-based analytical solvers struggle to converge with inaccuracies in the grid model.
Innovation Solution
Training a machine learning model, specifically a graph neural network, to predict electrical behaviors by processing graphs representing electrical systems, allowing for the inference of unknown values and improving simulation accuracy even with incomplete data, by using training data from ground-truth simulation systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If physics-based analytical solvers are used for electrical power simulations, then simulation accuracy can be maintained with complete models, but computational cost becomes prohibitively expensive and convergence fails with incomplete or inaccurate models
Solution Approach 1:
The patent replaces physics-based analytical solvers with machine learning models (neural networks) to perform power flow simulations. The neural network is trained on simulation data and then used to predict electrical behaviors directly, substituting the traditional iterative physics-based solving approach with a learned mapping from system state to solution, thereby reducing computational complexity and improving convergence reliability
Solution Approach 2:
The patent performs preliminary training of the neural network model using extensive simulation data generated by physics-based solvers. This pre-computation phase creates a learned model that captures the complex relationships in power systems, allowing the model to make rapid predictions without requiring expensive real-time physics-based computations
2Measurement precision
If complete and accurate electrical grid models are used, then simulation results are reliable, but computational resources are excessively consumed
Solution Approach 1:
The patent substitutes computationally intensive physics-based analytical simulations with a pre-trained neural network model. The neural network learns the mapping from system configuration to electrical behaviors during training, enabling rapid inference that maintains accuracy while dramatically reducing computational resource consumption during actual simulation operations
Solution Approach 2:
The patent creates a simplified computational copy of the complex physics-based simulation system through the neural network. This copied model captures the essential input-output relationships of the original system but can be evaluated much more efficiently, trading off some model complexity for computational speed and resource efficiency
3Productivity
If traditional simulation methods are used with incomplete data, then computational resources are saved, but simulation results become inadequate and fail to converge
Solution Approach 1:
The patent incorporates a feedback mechanism where the neural network model is trained on simulation data and its predictions are continuously refined. The model learns from the relationship between system states and electrical behaviors, adjusting its internal parameters to improve prediction accuracy. This feedback-driven training enables the model to handle incomplete data robustly by learning from patterns in the training data
Solution Approach 2:
The patent changes the approach from using fixed physics-based equations that require complete accurate models to using a flexible neural network model with learnable parameters. The neural network can adapt its parameters during training to handle variations in data quality and completeness, making it more robust to incomplete input data while maintaining simulation efficiency
Data Source
AI summary
In one aspect, there is provided a method for training a machine learning model to process a graph that represents an electrical system to infer, from the graph, one or more unknown electrical values within the electrical system. In particular, the method includes: obtaining data defining multiple graphs, each graph representing a respective electrical system topology, obtaining, for each electrical system topology and from an electrical simulation system, simulation results indicating an electrical behavior of the respective electrical system topology, and training the machine learning model to predict electrical behaviors of electrical systems including by applying data defining each graph as input to the machine learning model to obtain respective output inferences and adjusting machine learning model parameters responsive to comparisons between the output inferences with simulation results of corresponding electrical system topologies.