Jadranko Matuško

Stochastic Model Predictive Control for Building HVAC Systems: Complexity and Conservatism

This paper presents a stochastic model predictive control (SMPC) approach to building heating, ventilation, and air conditioning (HVAC) systems. The building HVAC system is modeled as a network of thermal zones controlled by a central air handling unit and local variable air volume boxes. In the first part of this paper, simplified nonlinear models are presented for thermal zones and HVAC system components. The uncertain load forecast in each thermal zone is modeled by finitely supported probability density functions (pdfs). These pdfs are initialized using historical data and updated as new data becomes available. In the second part of this paper, we present a SMPC design that minimizes expected energy cost and bounds the probability of thermal comfort violations. SMPC uses predictive knowledge of uncertain loads in each zone during the design stage. The complexity of a commercial building requires special handling of system nonlinearities and chance constraints to enable real-time implementation, minimize energy cost, and guarantee thermal comfort. This paper focuses on the tradeoff between computational tractability and conservatism of the resulting SMPC scheme. The proposed SMPC scheme is compared with alternative SMPC designs, and the effectiveness of the proposed approach is demonstrated by simulation and experimental tests.

Neural network based tire/road friction force estimation

This paper deals with the problem of robust tire/road friction force estimation. Availability of actual value of the friction force generated in contact between the tire and the road has significant importance for active safety systems in modern cars, e.g. anti-lock brake systems, traction control systems, vehicle dynamic systems, etc. Since state estimators are usually based on the process model, they are sensitive to model inaccuracy. In this paper we propose a new neural network based estimation scheme, which makes friction force estimation insensitive to modelling inaccuracies. The neural network is added to the estimator in order to compensate effects of the friction model uncertainties to the estimation quality. An adaptation law for the neural network parameters is derived using Lyapunov stability analysis. The proposed state estimator provides accurate estimation of the tire/road friction force when friction characteristic is only approximately known or even completely unknown. Quality of the estimation is examined through simulation using one wheel friction model. Simulation results suggest very fast friction force estimation and compensation of the changes of the model parameters even when they vary in wide range.

Deep neural networks for ultra-short-term wind forecasting

The aim of this paper is to present input variable selection algorithm and deep neural networks application to ultra-short-term wind prediction. Shallow and deep neural networks coupled with input variable selection algorithm are compared on the ultra-short-term wind prediction task for a set of different locations. Results show that carefully selected deep neural networks outperform shallow ones. Input variable selection use reduces the neural network complexity and simplifies deep neural network training.

Scenario-based approach to stochastic linear predictive control

In this paper we consider the problem of predictive control for linear systems subject to stochastic disturbances. We repeatedly solve a stochastic finite-time constrained optimal control problem by using the scenario-based approach. We address the conservatism of the approach by presenting a new technique for fast scenario removal based on mixed-integer quadratic optimization. Probabilistic bounds are derived which quantify the benefits of the proposed technique. The approach is illustrated through a numerical example.

Stochastic model predictive control for optimal economic operation of a residential DC microgrid

In this paper we present power flow optimization of a residential DC microgrid that consists of photovoltaic array, batteries stack and fuel cells stack with electrolyser, and is connected to the grid via bidirectional power converter. The optimization problem aims to minimize microgrid operating costs and is formulated using a linear program that takes into account the storages charge and discharge efficiency. To account for power predictions uncertainty, optimization problem is defined in a stochastic framework by using chance constraints. Since we assume that the error in realization of power predictions will be compensated by utility grid, chance constraints are defined for power exchange between the microgrid and the utility grid. Finally, we investigate a stochastic model predictive control for the closed-loop power management in the microgrid. Performance verification of the proposed approach is performed on simulations for two-month period.

A Takagi-Sugeno Fuzzy Model of Synchronous Generator Unit for Power System Stability Application

This paper presents a Takagi-Sugeno (TS) synchronous generator unit model intended for application in an auto-tuning power system stabilizer. The model takes into account all three process variables that can affect synchronous machine dynamics regarding stability—beside usually used active and reactive power (P and Q) it includes also line reactance (xm). It is shown that the proposed model gives very good results in spite of simple third order local model used as the consequent part of the proposed TS structure. This makes the model appropriate for implementation on simple microprocessor platforms. Because P, Q and xm are included as TS model premises, it is enough to identify parameters of models in consequent part of TS model off-line. In this way possible numerical instability is avoided, which is common to adaptive PSSs that calculate controllers parameters directly from on-line identified plant parameters.

Real-time Predictive Control of 3D tower crane

In this paper a real-time Model Predictive Control (MPC) for a 3D tower crane, based on subsequent solving of three quadratic programs, is proposed. Three motions of the tower crane are considered as separate subsystems with couplings among them treated as a change in system parameters. Such linear parameter varying (LPV) system can be sampled and transformed into the corresponding polytopic model, by using a Tensor Product (TP) Model Transformation. Polytopic TP model of the tower crane is used to calculate the terminal set and the terminal cost via solving Linear Matrix Inequalities (LMI). In order to guarantee recursive feasibility, system states are kept in ellipsoidal approximation of worst case initial feasible set, while the asymptotic stability is ensured by using a dual-mode MPC strategy. The proposed approach is verified through simulation and experimental test on laboratory model of a 3D tower crane.

Application of the RBF neural networks for tire-road friction force estimation

This paper deals with the problem of the robust tire-road friction force estimation. Good information about friction force generated in contact between wheel and road has significant importance in many active safety systems in modern vehicles (anti-lock brake systems, traction control, vehicle dynamic systems, etc). Since state estimators are usually based on exact model of process, they are therefore limited by the model accuracy. A new estimation scheme based on RBF neural networks is proposed in this paper. The neural network is added to the estimator to compensate the effects of the friction model uncertainties to the estimation quality. An adaptation law for the neural network parameters is derived using Lyapunov stability analysis. The proposed state estimator provides accurate estimation of the tire-road friction force when fiction characteristic is only approximately known or even completely unknown. Quality of the estimation is examined through simulation using one wheel friction model. Simulation results suggest very fast compensation of the changes of the model parameters (< 150 ms) even when they vary in a wide range (changes of 100% and more). Possible drawback of proposed estimation scheme is the fact that neural network does not give the information what particular parameter has changed.

Estimation of VRLA battery states and parameters using Sigma-point Kalman filter

This paper describes a hybrid electrical model of valve-regulated lead-acid battery (VRLA) and its application in Matlab/Simulink environment. Based on the charge/discharge test characteristics of the telecommunication battery stack, all parameters of the hybrid electrical models are derived. After implementing the charge/discharge simulations of hybrid electrical model and comparisons with actual tests of battery stack, joint estimation of model states and parameters is carried out using Sigma-point Kalman filter (SPKF). Results of performed joint estimation correspond to model simulations and it is shown that the SPKF algorithm is good for estimation of model states and parameters. After validation of the hybrid electrical model and validation of SPKF algorithm, joint estimation of battery states and parameters is performed to charge/discharge test of VRLA battery stack using Unscented Kalman Filter (UKF) method.

Tensor product based control of the Single Pendulum Gantry process with stable neural network based friction compensation

Fast and accurate positioning and swing minimization of the containers and other loads in crane manipulation are demanding and in the same time conflicting tasks. For accurate positioning, the main problem is nonlinear friction effect, especially in the low speed region. In this paper authors propose position controller realized as hybrid controller. It consists of the tensor product based nonlinear feedback controller with additional friction self-learning neural compensator. The experimental results show that friction compensator is able to remove position error in steady state.