A polyhedral neural network approximation method for parameterizing feasible set of control system

By constructing a polyhedral neural network approximation method, the problems of conservatism and low efficiency in feasible set calculation in time-varying and nonlinear control systems are solved, and fast and adaptive approximation of parameterized feasible sets is achieved, thereby improving the computational efficiency of online feasibility assessment.

CN122362853APending Publication Date: 2026-07-10NORTH CHINA ELECTRIC POWER UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
NORTH CHINA ELECTRIC POWER UNIV
Filing Date
2026-04-23
Publication Date
2026-07-10

AI Technical Summary

Technical Problem

Existing technologies for calculating feasible sets in the context of time-varying and nonlinear control systems suffer from problems such as high conservatism, poor adaptability, and low computational efficiency, making it difficult to achieve rapid online evaluation.

Method used

A polyhedron neural network approximation method is constructed. Through a discrete-time parameterized model, a neural network is built using time-varying parameters input to the polyhedron system matrix and right-hand term vector. A differentiable loss function is constructed based on the support point deviation between the polyhedron and the parameterized feasible set. Backpropagation training is then performed to achieve fast approximation of the parameterized feasible set.

Benefits of technology

It achieves fast and adaptive approximation of parameterized feasible sets, significantly improves the computational efficiency of online feasibility assessment, overcomes the conservatism and computational complexity of traditional methods, and is applicable to control systems with time-varying parameters and mixed constraints.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122362853A_ABST
    Figure CN122362853A_ABST
Patent Text Reader

Abstract

This invention provides a polyhedral neural network approximation method for parameterized feasible sets of control systems, relating to the field of automation control technology. The method includes: establishing a discrete-time parameterized model of the control system; defining state variables, control inputs, time-varying parameters, and constraints in the discrete-time parameterized model; constructing a neural network using the time-varying parameters as input and the coefficient matrix and right-hand side vector of the polyhedron as outputs; using the polyhedron as a geometric approximation of the parameterized feasible set; constructing a differentiable loss function based on the support point deviation between the polyhedron and the parameterized feasible set in the sampling direction; randomly sampling the time-varying parameters and directions, obtaining support points based on the constraints, calculating the differentiable loss function, and performing backpropagation to update the neural network; inputting the predicted time-varying parameters into the trained neural network, obtaining the corresponding polyhedral parameters through one forward propagation, and outputting the approximate feasible set result for feasibility assessment, thus improving computational efficiency and accuracy.
Need to check novelty before this filing date? Find Prior Art