Study On Building Data-Driven Predictive Control Method And Its Implementation

As for the model predictive control(MPC)research on building system,there exist some problems,such as high cost of parameters estimation of resistance-capacitance(RC)model,imperfection of data-driven predictive control(DDPC)process,and the lack of research on predictive controller implementation.Therefore,this dissertation proposes a framework which covers the whole process from DDPC controller design to implementation.Buildings of different geometry are selected to validate the proposed framework.This dissertation then conducts studies such as,simplification of the RC model structure and parameters estimation,development of the data-driven building thermal dynamics models,improvement of the DDPC design flow and implementation of the DDPC controller in embedded hardware.Firstly,based on the building modeling technology and knowledge of MPC,the general form of mathematical optimization problems in predictive control of buildings is summarized.This dissertation then analyzes the principles and scopes of artificial intelligence(AI)technology,and finds the integration points with the tasks in the building control and energy management field.Combined with the specific research content of the resistance-capacity(RC)model identification,improvement of the DDPC strategy and implementation of the DDPC controller,etc.,a full-process framework from design to implementation of building predictive controller is developed.Secondly,according to different levels of simplified heat transfer assumptions,the corresponding RC network models are developed and identified for a single-zone building.This dissertation compares the prediction performance of different RC models and analyzes the key factors of the model performance to further select a suitable control-oriented single-zone RC model structure.A new meta-heuristic optimization algorithm(so-called BSAS)is proposed for the parameters estimation task.The results of the proposed optimization algorithm with those of commonly used optimization algorithms(e.g.genetic algorithm)on the model identification task are compared.The proposed BSAS algorithm is able to perform the parameters estimation task and reduce its computational cost.Thirdly,this dissertation proposes a hybrid optimization strategy and introduces a PID-based feedback correction strategy to improve the DDPC controller design process.An RC model and data-driven models,including an auto-regression with exogenous inputs(ARX)model,an artificial neural network(ANN)and a support vector machine(SVM)model are developed for building thermal dynamics prediction for a single-zone electric heating building.Based on the DDPC controller design process and models,the MPC and DDPC controllers are developed and compared under three temperature control scenarios to verify the feasibility of the DDPC controllers.The validation of DDPC controllers is conducted on the co-simulation platform of Energy Plus.Results show that the DDPC controllers based on the ARX and ANN models can achieve the similar control performance as the MPC controller based on RC model.The DDPC controller developed based on the SVM model has small fluctuations in temperature control performance during strong weather disturbances.Based on multi-layer perception(MLP)models developed by the deep learning platform,the DDPC design sub-framework is applied to the temperature control of two three-zone buildings under multiple scenarios.The results show that DDPC controllers based on MLP models can achieve good temperature control performance on complex buildings.And there are small fluctuations in the temperature profiles of the Case 2 building when the temperature set-point is changedFinally,in order to address the high cost of implementing building predictive control research,a predictive controller implementation scheme based on inexpensive embedded hardware is developed.By analyzing the mathematical form of the receding horizon optimization(RHO)problems in the ANN-based DDPC controller,this dissertation proposes an accelerated first-order optimization strategy for the platform with limited computational power.For a single-zone electric heating building,the DDPC controllers are successfully implemented on the embedded hardware(STM32F103C8T6 development board)based on the proposed accelerated strategy and the introduced PID feedback correction strategy.The temperature control performance and computational cost of DDPC controllers driven by the proposed accelerated strategy are studied under three different scenarios.The results of the hardware-in-the-loop(HIL)co-simulation show that the embedded DDPC controller can achieve good temperature control performance in multiple scenarios,while the accelerated algorithm can save up to 54% of the computational cost.In summary,the proposed full-process design and implementation framework for building DDPC controller as well as its conduction on development of the building thermal dynamics models,design and implementation of DDPC controllers reduces the modeling cost in building predictive control,improves the application capability of the predictive controller,and provides a viable solution for the engineering practice of predictive control in building energy management.