Adaptive Digital Twin-Driven Performance Prediction and Collaborative Model Updating for Intelligent Control Systems

Abstract

To address the issue of decreased prediction reliability and control stability caused by model drift, non-stationary disturbances, and structural evolution in complex engineering systems, this paper proposes an adaptive digital twin-driven performance prediction and collaborative model update framework. This framework is based on a non-autonomous stochastic dynamic system, modeling the digital twin parameter field as a stochastic process in a time-varying function space. Simultaneously, as an optimization, stochastic differential equations and stochastic partial differential equations jointly characterize the coupled evolution mechanism of continuous disturbances, sudden shocks, and parameter drift. Based on this, a prediction-control coupled operator semigroup model is constructed. Nonlinear operator kernel expansion and spectral consistency constraints are introduced to maintain long-term consistency between the twin prediction operator and the dynamic structure of the physical system. Subsequently, a non-stationary Bayesian hierarchical prediction model and its variational recursive update mechanism are proposed. Furthermore, the digital twin model and control model are abstracted into a two-agent collaborative game system. This model constructs a continuous-time policy gradient flow and cross-regularized co-evolution equation. Through a demand-driven state transition operator and a task-level collaborative control mechanism, the framework's adaptability in multi-task and multi-scenario environments is further enhanced. Systematic results show that, compared with the current best comparative method, the proposed method reduces the mean squared error (MSE) and mean absolute error (MAE) by approximately 23.5% and $14.9{{\% }}$, respectively, in long-term prediction tasks.