Neural Network Model Predictive Control Research Articles

Neural Network Model Predictive Motion Control Applied to Automated Driving With Unknown Friction

Many innovative applications of vehicle control involve trajectory following while avoiding collisions, respecting actuator and dynamic limits, and using complex nonlinear dynamics. Additionally, these vehicle controllers must operate in the presence of difficult-to-model and uncertain dynamic forces which are often a function of the environment. To solve these problems, we present a design and experimental validation of neural network model predictive control (NNMPC), a method that uses vehicle operation data to construct a neural network model which is efficiently implemented in MPC. By learning a neural network model with a history of states and controls, NNMPC is capable of predicting vehicle dynamics in changing and complex operating conditions. We challenge NNMPC with the difficult task of automated racing near the friction limits without prior knowledge of the road-tire friction coefficient. The experimental results on an automated test vehicle demonstrate the capability of NNMPC to follow a trajectory near the limits on both high- and low-friction test courses. Furthermore, NNMPC outperforms a physics-based benchmark MPC on both the courses where the environmental latent state of road-tire friction is explicitly considered.

Speed control of PMSM based on neural network model predictive control

In order to optimize the control performance of permanent magnet synchronous motor (PMSM) servo system, an improved model predictive control (MPC) scheme based on neural network is investigated in this paper. First, the dynamic characteristics of the PMSM are approximated by echo state network (ESN) to predict the future speed. Particle swarm optimization (PSO) is used to train ESN output weights to solve the problem that instability of output weights caused by pseudo-inverse matrix in ESN weight solving algorithm, called PSO-ESN, which enhances the stability and the accuracy of ESN speed prediction. That provides future plant output for control optimization of the predictive control. Furthermore, in order to reduce the computational cost and improve the response performance of the controller, a fast gradient method (GM) is applied to minimize the quadratic performance index and solve the optimal control input sequences. The simulation results under three different working conditions show that the PSO-ESNMPC controller designed in this paper reduces the overshoot by 5.87% and the rise time by 0.036 s compared with the reference controllers and has better robustness under parameter changes and load disturbances.

Implicit Contact Dynamics Modeling With Explicit Inertia Matrix Representation for Real-Time, Model-Based Control in Physical Environment.

Model-based control has great potential for use in real robots due to its high sampling efficiency. Nevertheless, dealing with physical contacts and generating accurate motions are inevitable for practical robot control tasks, such as precise manipulation. For a real-time, model-based approach, the difficulty of contact-rich tasks that requires precise movement lies in the fact that a model needs to accurately predict forthcoming contact events within a limited length of time rather than detect them afterward with sensors. Therefore, in this study, we investigate whether and how neural network models can learn a task-related model useful enough for model-based control, that is, a model predicting future states, including contact events. To this end, we propose a structured neural network model predictive control (SNN-MPC) method, whose neural network architecture is designed with explicit inertia matrix representation. To train the proposed network, we develop a two-stage modeling procedure for contact-rich dynamics from a limited number of samples. As a contact-rich task, we take up a trackball manipulation task using a physical 3-DoF finger robot. The results showed that the SNN-MPC outperformed MPC with a conventional fully connected network model on the manipulation task.

Control of denitration system in cement calcination process: A Novel method of Deep Neural Network Model Predictive Control

Denitration is an important measure to reduce pollutant emissions in cement calcination prosses. The purpose of accurate control of the denitration system in cement calcination is to control the concentration of nitrogen oxides (NOx) emissions while reducing the use of ammonia. However, it is difficult to precisely control the denitration system because of the problems of multivariable, large time delay and uncertainty in the cement calcination process. In order to solve these problems, we propose a denitration process control strategy based on Deep Neural Network Model Predictive Control (DNN-MPC) for cement calcination system. This method proposes a time series deep belief network (TS-DBN) prediction model, which solves the problems of long delay, nonlinearity and high data complexity in the denitration system, and realizes a more precise forecast of NOx emission. The model predictive control framework is applied to solve the problems of volatility and uncertainty in the denitration system. The ammonia flow rate is introduced into the objective function to reduce ammonia escape and achieve the purpose of energy saving. In addition, we use differential evolution (DE) algorithm for rolling optimization solution, and add variable constraints that meet the actual running conditions in the solution process. Compared with PID controller dynamic setting time is reduced to 10% and ammonia consumption is saved by 1.7%. The experimental results show that the method proposed in this paper has higher accuracy and stronger robustness in controlling the NOx emission in the denitrification system of cement calcination, and at the same time reduces the amount of ammonia used.

DEEP LEARNING BASED INTELLIGENT METHODOLOGY FOR PHOTOVOLTAIC POWER QUALITY CONTROL WITH ENERGY STORAGE

Power imbalances instigated from Photovoltaics (PV) avert the incorporation of large-scale PV power into the grid and makes it a challenging resource as the intermittencies caused result in frequency variations and voltage interruptions which eventually engenders electricity blackouts. Furthermore, it compromises the associated battery's operation as it results in irregular battery charging/discharging. The fluctuating power also leads to the requirement of larger and impractical storage systems. Therefore, the power quality from the PV needs to be controlled to allow a stable output power supply to the load and also to enhance the operation of the connected Battery Energy Storage System (BESS). This paper presents a novel neural network (NN) model predictive control (MPC) method for PV power control and firming with BESS. Instead of creating a mathematical model of the plant for the prediction optimization step in an MPC, the proposed controller generates a NN plant model. Due to its innate characteristic, a NN model better captures the dynamics of the plant, delivers highly accurate predictions, and resolves the mathematical model complexity issues that occur in highly complicated plants. The proposed controller effectively firms the varying solar power and stabilizes the battery state of charge.

Neural network predictive control for smoothing of solar power fluctuations with battery energy storage

Power fluctuations caused by Photovoltaics (PV) prevent the penetration of large-scale PV power into the grid as it causes multiple instabilities such as frequency deviations, voltage fluctuations, and decreased output power quality. Additionally, it severely compromises the associated battery’s performance and reduces its operational life span that can lead to the requirement of larger batteries thereby increasing the overall system cost. In this paper, a novel neural network model predictive control (MPC) approach for photovoltaic power smoothing with battery energy storage system is proposed. As opposed to the conventionally used MPC that utilizes the mathematical model of the plant for its predictive optimization, the proposed controller generates a Neural Network (NN) model of the plant. In comparison to the mathematical model, a NN better encapsulates the dynamics of the plant and can also provide higher accuracy predictions. Furthermore, the precision of the NN plant model is further increased as the collected input–output plant data increases. The NN model also solves the issues related to mathematical complexity of the MPC model that arises due to the increasing complications in the plant. Whereas the inherent characteristics of a NN allows it to model highly complex plants with a relatively simpler approach. The proposed controller is capable of firming the solar power by employing the inputs from our NN plant model and also optimizes the battery state of charge under a variety of practical constraints which consequently promotes enhanced battery life. Furthermore, this study also proposes a novel NN architecture for accurate PV power forecasting. In comparison to the popularly used fuzzy logic controller, the proposed controller manages to significantly reduce the battery charging levels and state of charge.

An Adaptive RBF-NMPC Architecture for Trajectory Tracking Control of Underwater Vehicles

An adaptive control algorithm based on the RBF neural network (RBFNN) and nonlinear model predictive control (NMPC) is discussed for underwater vehicle trajectory tracking control. Firstly, in the off-line phase, the improved adaptive Levenberg–Marquardt-error surface compensation (IALM-ESC) algorithm is used to establish the RBFNN prediction model. In the real-time control phase, using the characteristic that the system output will change with the external environment interference, the network parameters are adjusted by using the error between the system output and the network prediction output to adapt to the complex and uncertain working environment. This provides an accurate and real-time prediction model for model predictive control (MPC). For optimization, an improved adaptive gray wolf optimization (AGWO) algorithm is proposed to obtain the trajectory tracking control law. Finally, the tracking control performance of the proposed algorithm is verified by simulation. The simulation results show that the proposed RBF-NMPC can not only achieve the same level of real-time performance as the linear model predictive control (LMPC) but also has a superior anti-interference ability. Compared with LMPC, the tracking performance of RBF-NMPC is improved by at least 43% and 25% in the case of no interference and interference, respectively.

Comparison of Model Predictive Controller and Neural Network Controller on Nonconventional Distillation Demethanizer/Deethanizer Rectifier/Column System

Nonconventional distillation demethanizer/deethanizer rectifier/column offers an economic advantage compared to conventional configurations by using sidestream from second column to remove the use of reboiler. However, this configuration brings a complex system that needs attention in its process control to ensure its operability. This work explores the use of neural network model predictive control (NNMPC) for this system. Aspen HYSYS was used to simulate the process, generate input-output data to develop the plant model and to conduct performance tests. MATLAB was used to conduct model identification, design the MPC and implement the multivariable control action. Sufficient time series data, where reflux valve positions and distillate compositions from each column act as input and output respectively, can be obtained from simulation and then utilized to train a nonlinear autoregressive exogenous (NARX) neural network which will be used as a predictive model. Model predictive Control based on this neural network model is designed to control distillate compositions for each column by the changes of reflux valve positions for each column as control signal. Least square algorithm from MATLAB function wass employed to optimize the cost function. A comparison has been made between NNMPC and conventional MPC configurations with first order plus dead time (FOPDT) model to test the performance of the controller based on disturbance rejection and set-point tracking test. Quantitative calculation was done using integral absolute error (IAE) with trapezoidal rule used as numerical integral method. NNMPC shows better performance than MPC with FOPDT model in set-point tracking test. However, NNMPC has a disadvantage on the cost function calculation algorithm, which did not account for the difference between plant and model value thus cause the MPC with FOPDT model to give better performance on disturbance rejection. Bias addition on NNMPC cost function algorithm made the controller has better responses to disturbance rejection compared to MPC.

A data-driven neural network model predictive steering controller for a bio-inspired quadruped robot

Abstract This paper proposes a particular type of nonlinear data-driven Model Predictive Control (NMPC) strategy, called Neural Network Model Predictive Control (NNMPC), applied to a simulated neuro-inspired quadruped robot. The locomotion control is realised using a central pattern generator (CPG) implemented through oscillators synchronized through environmental feedback. The NMPC provides a descending command to the robot for steering control. This is realized by regulating a parameter governing the dynamics of the CPG structure. In order to test the performance obtained applying the NMPC, the results are compared with those obtained using a linear MPC. Carrying out a comparative analysis, the differences between the two methods will be highlighted.

Neural Network Model Predictive Control (NNMPC) Design for UPFC

Neural Network Model Predictive Control (NNMPC) is like almost like the model predictive control but the used inboard plant is designed based on using the concept of the artificial neural network to predict the behavior of the plant. The predicted values are fed to the optimizer in order to obtain better control variables. This type of controller will be used instead of the conventional controller in the most versatile FACTS devices, which is the Unified Power Flow Controller (UPFC). UPFC has the capability of controlling the transmission line parameters and consequently the flow of the active and reactive power in the transmission line. So, this type of adaptive controller, which is based on Artificial Neural Network (ANN) concept, will be implemented in UPFC, and will be investigated to ensure its robustness, effectiveness and the capability to accommodate any sudden load change in the system of Single Machine to Infinite Bus (SMIB). In addition, the dynamic performance of NNMPC will be compared with another type of adaptive controller scheme called Model Productive Controller (MPC).

Vision-based neural predictive tracking control for multi-manipulator systems with parametric uncertainty

To deal with the coordination problem for multi-manipulator trajectory tracking systems with parametric uncertainties, this paper proposes a two-layer control scheme incorporating a model predictive strategy and an extended state observer. In the kinematic layer, the visual information is implemented and a visual servoing error model is derived by the image-based visual servoing strategy. A recurrent neural network model predictive control approach is proposed to obtain velocities which are regarded as the reference signals for the dynamic layer. For dynamics, a linear time-varying dynamic model of the multi-manipulator system coupled with the object is established, where the parametric uncertainty is recognized as an added disturbance. An extended state observer is sequentially designed to estimate the disturbance by using pole placement method. The input-to-state practical stability of the system is further analyzed with a bounded disturbance. Finally, simulations and comparison are given to verify the effectiveness and robustness of the proposed algorithm.

Implementation of a neural network MPC for heat exchanger network temperature control

Heat exchanger network (HEN) control is considered to be a difficult task due to its nonlinear behavior, complexity, presence of disturbances and noise. The optimal operation of HEN implies the implementation of a robust control system able to overcome all these issues. This work presents an advanced technique based on an artificial neural network model predictive control (NNMPC) and compares the controlling performance to three widely used controllers: two PID and a linear MPC. The goal is to control an outlet oil stream temperature of a HEN containing four heat exchangers in counter-current. A lumped parameter model was used to describe the HEN system. The data generated were used to train a neural network in the Matlab environment using the Levenberg-Marquardt algorithm. Closed-loop negative and positive step responses were simulated to investigate set point tracking and disturbance rejection. The ability to handle constraints was evidenced in MPC, as well as faster settling time and minimum overshoot.

Speed control of permanent magnet synchronous motor using neural network model predictive control

Model predictive control has been widely used in the industry. This can control the multivariable system with constraints on input and output variables but it needs online computation solver, and creates the non-convex solution in nonlinear plant due to the parameter uncertainties. The online computational problem and non-convex solution of the model predictive control are achieved via neural network model predictive control. The paper explores the speed control of permanent magnet synchronous motor (PMSM) by using neural network model predictive control (NNMPC) technique. The multi-layer artificial neural network is used to identify the dynamics of PMSM. The set point speed tracking control of PMSM is identified by using neural network model predictive control strategy. By using the set of input and output data obtained from the system, the multi input-output feed-forward neural network model is created. Levenberg-Marquardt algorithm is used to train the process models of the PMSM. That provides future plant output for control optimization of the predictive control. The overall system is developed and tested in the MATLAB/Simulink. To evaluate the efficiency of the controller proposed, it is compared with a constrained model predictive controller through the studies of simulation. The overshoot and settling time of the speed response of the PMSM are measured and analyzed for NNMPC and constrained MPC.

Application of neural networks for optimal-setpoint design and MPC control in biological wastewater treatment

This paper addresses both the design of an optimal variable setpoint and a setpoint-tracking control loop for the dissolved oxygen concentration in a biological wastewater treatment process. Although exact knowledge of influent changes during rain/storm events is unrealistic, we take advantage of the fact that during dry weather conditions the influent changes are periodic and thus predictable. Specifically, a nonlinear optimization procedure utilizes dry weather data to decide on a nominal fixed setpoint, or a weighting gain, or both; during weather events an algorithm uses the optimization solution(s) together with the ammonium predictions to adjust the setpoint dynamically (responding appropriately to significant changes in the influent). A constrained nonlinear neural-network model predictive control tracks the setpoint. Simulations with the BSM1 compare several variations of the proposed methods to a fixed-setpoint PI control, demonstrating improvement in effluent quality or reduction in energy use, or both.

Neural network model predictive control optimisation for large wind turbines

Energy poverty limits the economy and social development throughout the world. Wind turbine reduces the energy costs and facilitates the development of renewable energy industry, which provides an effective solution to energy crisis and environment pollution and develops rapidly in recently years. In this paper, a radial basis function neural network (RBFNN) optimisation model predictive control (MPC) was proposed for large wind turbines. In accordance with the complexity and uncertainty of wind turbine operation, a linear model based on the blade element momentum theory was established and the influencing factors of the proposed model were evaluated. The MPC taking into full account three degrees of freedom control multivariate was enforced by RBFNN prediction model, which meets the requirements of specified operation region. Additionally, the RBFNN prediction model with the memory of complicated rules and changed trend was trained by a great deal of historical data. The RBFNN in combination with MPC solves global optimisation problems and improves the dynamic performance of system. Simulation results for three-bladed 5 MW onshore wind turbine verified the effectiveness of the proposed method and confirmed the fact that the fatigue loads were significantly reduced in the turbine tower.

Optimized deep learning neural network predictive controller for continuous stirred tank reactor

In this paper, a deep learning neural network model predictive controller (DLNNMPC) is designed to analyse the performance of a non-linear continuous stirred tank reactor (CSTR) that performs parallel and series reactions. The data generated employing the state space model of CSTR is used to train the designed deep learning neural network controller. Deep Learning Neural Network (DLNN) progresses the training with its weights tuned by the proposed hybrid version of evolutionary algorithms – Particle Swarm Optimization (PSO) and Gravitational Search Algorithm (GSA). The developed hybrid PSO – GSA based DLNN model of continuous stirred tank reactor is employed in this paper for model predictive controller design. The effectiveness of the proposed DLNNMPC tuned by hybrid PSO – GSA for CSTR is validated for its performance on comparison with that of other designed Proportional – Integral (PI) and Proportional – Integral – Derivative (PID) controllers as available in early literatures for the same problem under consideration.

Simulation and Control of a Methanol-To-Olefins (MTO) Laboratory Fixed-Bed Reactor

In this research, modeling, simulation, and control of a methanol-to-olefins laboratory fixed-bed reactor with electrical resistance furnace have been investigated in both steady-state and dynamic conditions. The reactor was modeled as a one-dimensional pseudo-homogeneous system. Then, the reactor was simulated at steady-state conditions and the effect of different parameters including inlet flow rate, inlet temperature and electrical resistance temperature on reactor performance was studied. Results showed that the most effective parameter is electrical resistance temperature. Thus, it was selected as manipulating variable for controlling product quality. In the next step, dynamic simulation of the process was performed and the effect of different disturbances on the dynamic behavior of the reactor was assessed. Finally, PID and Neural Network Model Predictive (NNMP) controllers were utilized for process control, and their performances were compared to each other. The response of the control system to different disturbances and set point changes showed that both PID and NNMP control systems can maintain the process at the desired conditions. PID controller had smaller rise time and no offset compared to NNMP controller while NNMP controller had smaller overshoot.

Neural Network Based Modeling and Control for a Batch Heating/Cooling Evaporative Crystallization Process

Crystallization processes have been widely used for separation in many fields to provide a high purity product. In this work, dynamic optimization and neural network (NN) have been applied to improve the quality of the product: citric acid. In the dynamic optimization, optimization problems maximizing both crystal yield and crystal size have been formulated. The neural networks have been developed to provide NN models to be used in the formulation of not only neural network inverse control (NNDIC) but also neural network model predictive control (NNMPC) strategies. The Levenberg Marquadt algorithm has been used to train the network and optimal neural network architectures have been determined by a mean squared error (MSE) minimization technique. In addition, a neural network model has been designed to provide estimates of the temperature and the concentration of the crystallizer. These estimates have been incorporated into the NNMPC controller. In the NNDIC controller, another neural network model has been applied to predict the set point of jacket temperature. The simulation results have shown that the obtained crystal size is increased by 19% and 30% compared to that by cooling and evaporation methods respectively and the obtained yield is increased more than 50%. The robustness of the proposed controller is investigated with respect to parameters mismatches. The results have shown that the NNMPC controller provides superior control performances in all case studies.

Neural Network Model predictive control for diary Falling Film Evaporators

A nonlinear neural network model predictive control (NNMPC) is proposed for an industrial evaporator system. This research is conducted in the Abo-Greeb state enterprise of dairy products. Evaporation process and its prediction model is used in the controller design. In this paper, a nonlinear model of the evaporator system components are described, thepresentedmodel of the evaporator system is done with MatLab Simulink 2014b.NNMPC based technique for developing nonlinear dynamic models is carried out from empirical data. Results shows that MPC can be used efficiently in control such systems.

An adaptive neuro-fuzzy sliding mode controller for MIMO systems with disturbance

This paper introduces the mathematical model of ammonia and urea reactors and suggested three methods for designing a special purpose controller. The first proposed method is Adaptive model predictive controller, the second is Adaptive Neural Network Model Predictive Control, and the third is Adaptive neuro-fuzzy sliding mode controller. These methods are applied to a multivariable nonlinear system as an ammonia–urea reactor system. The main target of these controllers is to achieve stabilization of the outlet concentration of ammonia and urea, a stable reaction rate, an increase in the conversion of carbon monoxide (CO) into carbon dioxide (CO2) to reduce the pollution effect, and an increase in the ammonia and urea productions, keeping the NH3/CO2 ratio equal to 3 to reduce the unreacted CO2 and NH3, and the two reactors' temperature in the suitable operating ranges due to the change in reactor parameters or external disturbance. Simulation results of the three controllers are compared. Comparative analysis proves the effectiveness of the suggested Adaptive neuro-fuzzy sliding mode controller than the two other controllers according to external disturbance and the change of parameters. Moreover, the suggested methods when compared with other controllers in the literature show great success in overcoming the external disturbance and the change of parameters.