外文翻譯---基于plc平臺的智能模型與控制技術(shù)的實(shí)施執(zhí)行（英文）

1、 Real-Time Implementation of Intelligent Modeling and Control Techniques on a PLC Platform Curtis Parrott, Student Member, IEEE and Ganesh K. Venayagamoorthy, Senior Member, IEEE Real-Time Power and Intelligent Systems

2、 Laboratory Missouri University of Science and Technology, Rolla, MO 65401 USA cap9qd@mst.edu and gkumar@ieee.org Abstract – Programmable logic controllers (PLCs) have been used for many decades for standard control

3、in industrial and factory environments. Over the years, PLCs have become computational efficient and powerful, and a robust platform with applications beyond the standard control and factory automation. Due to the

4、new advanced PLC’s features and computational power, they are ideal platforms for exploring advanced modeling and control methods, including computational intelligence based techniques such as neural networks, particl

5、e swarm optimization (PSO) and many others. Some of these techniques require fast floating-point calculations that are now possible in real-time on the PLC. This paper focuses on the Allen-Bradley ControlLogix brand

6、 of PLCs, due to their high performance and extensive use in industry. The design and implementation of a neurocontroller consisting of two neural networks, one for modeling and the other for control, and the trainin

7、g of these neural networks with particle swarm optimization is presented in this paper on a single PLC. The neurocontroller in this study is a power system stabilizer (PSS) that is used for power system oscillation da

8、mping. The PLC is interfaced to a power system simulated on the real time digital simulator. Real time results are presented showing that the PLC is a suitable hardware platform for implementing advanced modeling and

9、 control techniques for industrial applications. Index Terms – Computational intelligence, modeling, neurocontrol, particle swarm optimization, programmable logic controllers, power system stabilizer. I. INTRODU

10、CTION Programmable logic controllers (PLCs) have been used extensively in industrial applications for control for decades due to their high reliability and robust architecture [1]. The newest PLCs have moved past jus

11、t a robust platform into a new realm of high computational power and processor speed. These, along with the PLC’s highly expandable layout, makes it an ideal platform for far beyond the classical applications. Thes

12、e new applications include implementing computational intelligence based modeling, optimization and control techniques that require fast processing power to be executed in real-time. With the ability to contain analo

13、g I/O, the PLC is also ideal for interface to real-time simulation hardware, such as the real-time digital simulator (RTDS) for power systems [2]. The RTDS is a custom parallel processing hardware platform that allo

14、ws power systems to be simulated and its accessories (controllers, transformers, relays) to be tested in real-time [3]. Through the use of analog I/O, power control devices can be seamlessly tested as if they were pa

15、rt of the physical power system running on the simulator. This allows for the testing of any such control device containing low voltage I/O and allows the gauging of this control scheme as a legitimate real-world app

16、lication. The ability of the RTDS for control and protection system testing has been further explored in [2]. This makes the PLC-RTDS platform, an ideal platform for testing the viability of the PLC as a real-world

17、control platform for computational intelligence techniques. In this paper, a case study of implementing a controller based on neural networks for damping speed oscillations in generators is explored [4]. The PLC plat

18、form implements the neural networks required to realize the adaptive control and these neural networks are trained using particle swarm optimization (PSO) algorithm [5]. To the knowledge of the authors, computational

19、 intelligence techniques have not been implemented on PLCs which are known to be robust platforms for industrial controls. Power system stabilizers (PSSs) are used as an auxiliary control system to a generator’s

20、 excitation system. The purpose of the PSS is for power system oscillation damping during small and large system disturbances by providing supplementary control signals to the generator’s automatic voltage regu

21、lator (AVR) [6]. The speed oscillations can take the form of intra-area and inter-area modes in a multi-machine power system: intra area modes form where two or more synchronous machines swing together against a comp

22、aratively large power system or load center and inter-area modes involve combinations of many machines on one part of a power system swinging against machines on another part of the system [7]. Advanced power system

23、 modeling and control techniques have been explored on a wide variety of platforms, including digital signal processors (DSPs) and field programmable logic arrays (FPGAs), in great detail. However, research has negle

24、cted the staple of industrial control, the programmable logic controller. The PLC platform is used extensively in industry due to its very high reliability and expandability. This expandability includes a wide varie

25、ty of digital and analog I/O modules along with many different communication modules. The PLC is also designed with a powerful processor with the ability to do real-time control of a wide variety of control applicat

26、ion [1]. This paper demonstrates the potential of PLCs for implementing computational intelligence paradigms including neural networks and particle swarm optimization in real time for modeling and control of synchron

27、ous generators in a multimachine power system environment. This work is supported by the NSF CAREER Grant ECCS # 0348221 and US Dept. of Education GAANN funding awarded to Dr. Venayagamoorthy. 978-1-4244-2279-1/08/$25.0

28、0 © 2008 IEEE 1values of its output as well as past m values of its input. The inputs and outputs of the model are speed deviation of the plant (generator G1 or G3) and the output of the neurocontroller, and t

29、he estimated speed deviations respectively. Here, both n and m are chosen to be 2. The main reason for choosing three time step values is because a third order system is sufficient for the modeling the generator

30、 dynamics for this study. The model is a multi-layered feedforward neural network (Fig. 5) trained using the PSO algorithm. The input vector to the model network is X?and the estimated speed deviation at instant (k+1)

31、 is '( 1) k ω Δ + . Fig. 5. Neural network system model structure. ( ), ( 1), ( 2),( ), ( 1), ( 2) PSS PSS PSSk k k X V k V k V kω ω ω Δ Δ ? Δ ? ? ? = ? ? ? ? ? ?(1) 10, 1 i i j j j a W X= = ? ∑(2) 1 11 i i a d e? ?

32、 = +(3) 101 '( 1) i i i k V d ω= Δ + = ? ∑(4) B. Neurocontroller The neurocontroller is also a multi-layer feedforward network trained with PSO algorithm. The inputs to this system are the actual speed deviation

33、and the two previous values of a generator and the output of the neurocontroller is the supplementary control signal to the AVR, VPSS as shown in Fig. 3. The training of the neurocontroller is similar to that describ

34、ed in [4]; however PSO is used in lieu of backpropagation algorithm. C. PSO Algorithm PSO is a type of evolutionary computing technique. The algorithm is based on the simulation of the social interaction of birds wi

35、thin a flock and school of fish. Being a population based search algorithm, a swarm consists of particles which are potential solutions to the problem solved or optimized. The changes in the particles position in th

36、e search space is influenced by the past knowledge of the swarm as well as the particles own past knowledge of the search space. At initialization, each particle is randomly assigned to a point in the search space,

37、 as well as given a random starting velocity. The particle is then flown through the search space with the initial velocity. The particle is then evaluated as to how well it solves the problem at hand; this evaluati

38、on is called the particle’s fitness. This is then compared to the particle’s memory of its best solution of the problem, the pbest position. If the newest solution is better than the current pbest (the current fitn

39、ess lower than the pbest fitness), the pbest position is updated to the current position. Once all the particles have been evaluated the pbest with the lowest fitness is compared to the gbest position fitness. If th

40、is pbest’s value is lower than the current gbest fitness then the gbest position is update to this pbest’s location. This gbest represents the social aspect of the algorithm. After these updates have been done the

41、PSO equations are again evaluated, and take the form seen in (5) and (6). Also, an example of a single particle update can be seen in Fig. 6 where: xid(k) is the ith particle’s dth dimension current position; x(k+1)

42、is the ith particle’s dth dimension position after the PSO update, at the next time step; vid(k) is the ith particle’s dth dimension current velocity; vid (k+1) is the ith particle’s dth dimension velocity at the next

43、 time step; pid is the pbest position of ith particle’s; pgd is the groups best position or gbest for the dth dimension; w is the inertia weight constant; c1 and c2 are the cognitive and social acceleration constants

44、 respectively [5]. 1 12 2( 1) ( ) ( ( ) ( ))( ( ) ( ))id id id idgd idv k w v k c rand p k x kc rand p k x k+ = ? + ? ? ?+ ? ? ?(5) ( 1) ( ) ( 1) id id id x k x k v k + = + +(6) ) (k v w id ?)) ( ) ( ( 1 1 k x k p rand c