Scholarlink

Data-drivenPrediction network nonlinearNonlinearNonlinear model predictive control (DD-NMPC) algorithmsAlgorithm typically employ multi-step predictors parameterized using neuralNeural networksNeural networks to predict future outputs over a finite prediction horizon. This yields a so-called deepPrediction network prediction networkNetwork (DPN), with an architectureArchitecture defined by the structuring of predictors for different time instants. In this book chapter, we analyze and classify existing DD-NMPC formulations based on their underlying prediction networkNetwork architecture, rather than based on their predictorPredictor parameterization. This enables us to develop a unifying architectureArchitecture, i.e., the deep subspace prediction network (DSPN), which merges and extends existing architectures. Moreover, we show that DD-NMPC based on neuralNeural DSPN architecturesArchitecture can recover the well-known linear subspaceSpace predictive control algorithmAlgorithm for a sufficient number of hidden layerLayer neurons. To address the inherent computational complexityComputational complexity of DD-NMPC, we present a tailored sequential quadratic programming solverSolver for multi-step prediction networksPrediction network. Simulation results on a benchmark pendulum model show that DD-NMPC based on DSPN achieves high control performancePerformance for both noiseless and noisy data. A real-life implementation of DD-NMPC using neuralNeural DSPN with a prediction horizonHorizon of 10 for a mass–spring–damper–mass system with a sampling time of 5ms is also presented.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Prediction Networks for Data-Driven Nonlinear Model Predictive Control

  • 章节

  • 发表日期：

    2025年

查看全文
  • Mircea Lazar,
  • Mihai-Serban Popescu,
  • Maarten Schoukens

摘要

Data-drivenPrediction network nonlinearNonlinearNonlinear model predictive control (DD-NMPC) algorithmsAlgorithm typically employ multi-step predictors parameterized using neuralNeural networksNeural networks to predict future outputs over a finite prediction horizon. This yields a so-called deepPrediction network prediction networkNetwork (DPN), with an architectureArchitecture defined by the structuring of predictors for different time instants. In this book chapter, we analyze and classify existing DD-NMPC formulations based on their underlying prediction networkNetwork architecture, rather than based on their predictorPredictor parameterization. This enables us to develop a unifying architectureArchitecture, i.e., the deep subspace prediction network (DSPN), which merges and extends existing architectures. Moreover, we show that DD-NMPC based on neuralNeural DSPN architecturesArchitecture can recover the well-known linear subspaceSpace predictive control algorithmAlgorithm for a sufficient number of hidden layerLayer neurons. To address the inherent computational complexityComputational complexity of DD-NMPC, we present a tailored sequential quadratic programming solverSolver for multi-step prediction networksPrediction network. Simulation results on a benchmark pendulum model show that DD-NMPC based on DSPN achieves high control performancePerformance for both noiseless and noisy data. A real-life implementation of DD-NMPC using neuralNeural DSPN with a prediction horizonHorizon of 10 for a mass–spring–damper–mass system with a sampling time of 5ms is also presented.