Matrix Converter Control Using Q-Learning to Cut MPC Search Load
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Model Predictive Control (MPC) for matrix converters is resource-intensive due to the need for a greedy search over all possible switching states, becoming infeasible as the number of states increases, necessitating a method to boost computation speed and reduce search load.
Innovation Solution
Implementing a Q-data structure trained using Q-learning to map reward values for switching states, allowing the selection of a subset of states based on sorting and evaluation, thereby reducing the search space and computation time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Model Predictive Control performs a greedy search over all possible switching states, then control accuracy is improved, but computation time and resource consumption increase significantly
Solution Approach 1:
The patent segments the control process into two distinct phases: an offline training phase where a neural network is trained using reinforcement learning to learn optimal switching patterns, and an online execution phase where the pre-trained network rapidly predicts switching states. This segmentation transfers the computationally intensive work to the offline phase, enabling fast real-time control without sacrificing accuracy
Solution Approach 2:
The patent performs preliminary action by pre-training the neural network controller offline using reinforcement learning algorithms. The network learns optimal switching strategies in advance through simulated or real-world training, storing this knowledge in its weights and parameters. During actual operation, the pre-trained network immediately applies this learned knowledge without requiring real-time exhaustive search, thus reducing computation time while maintaining control accuracy
2Adaptability or versatility
If the number of switching states increases in future power converter architectures, then system capability is improved, but the greedy search becomes infeasible due to exponential growth in computation load
Solution Approach 1:
The patent substitutes the mechanical exhaustive search process with an intelligent neural network-based prediction system. Instead of systematically evaluating every possible switching state through computational search, the trained neural network directly predicts the optimal switching state based on learned patterns from training data, replacing brute-force computation with intelligent inference that scales better with system complexity
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
A method of controlling a matrix converter system is provided. The method includes receiving an operating condition and consulting a trained Q-data structure for reward values associated with respective switching states of the switching matrix for an operating state that corresponds to the operating condition. The Q-data structure is trained using Q-learning to map a reward value predicted for respective switching states to respective discrete operating states. The method further includes sorting the reward values predicted for the respective switching states mapped to the operating state that corresponds to the operating condition, selecting a subset of the set of the mappings as a function of a result of sorting the reward values associated with the switching states of the operating state, evaluating each switching state included in the subset, and selecting an optimal switching state for the operating condition based on a result of evaluating the switching states of the subset.