Abstract : This paper combines neural networks with traditional PID control methods to construct an adaptive PID control system based on a fuzzy RBF neural network, which is then applied to a switched reluctance motor with severe nonlinearity in an oil pumping unit. The characteristics of the control system are improved by combining the advantages of parameter-adjustable fuzzy neural network PID control and a learning algorithm with a variable learning rate. Experimental results show that this control system has advantages such as fast dynamic response, small overshoot, good robustness, and significant energy saving.
Keywords : Neural network; Switched reluctance motor; PID control; Energy saving
Intermediate Classification Number : TP9 Document Identification Code: B
0 Introduction
At present, as most of my country's oilfields enter the middle and late stages of development and the number of heavy oil wells, low-yield wells and low-yield areas increases, the demand for pumping units in oilfields is increasing to ensure production. my country's oilfields have more than 100,000 pumping units in stock. According to incomplete statistics, more than 10,000 new pumping units are needed every year in recent years. The vast majority of oilfield pumping units still use beam pumping units, which are characterized by simple structure and reliable operation. However, they have the disadvantage of very low overall efficiency, low power factor and high energy consumption. New energy-saving pumping equipment that reduces oil production energy consumption and improves oil production efficiency will be the future development trend and goal of the pumping unit production industry [1]. Because long-stroke linear pumping units have the advantages of simple structure, light weight, short transmission route, high efficiency, convenient maintenance, energy saving and environmental protection and wide applicability, long-stroke pumping units have become a hot spot for development, testing and promotion in China in recent years [2].
The magnetic circuits of switched reluctance motors in long-stroke pumping units are mostly designed to be relatively saturated, and their doubly salient pole structure and switching control method result in highly nonlinear characteristics. To adapt to the nonlinearity of the controlled object, a variable parameter adaptive PID control strategy is adopted. By combining artificial neural networks with PID control laws, the adaptive, nonlinear mapping and self-learning capabilities of neural networks are fully utilized to form a highly adaptive parameter-adjustable neural network PID control strategy [3~5].
This paper, based on the original PID controller, utilizes a fuzzy RBF neural network. Taking the actual system output and system error as inputs, it integrates a set of optimal PID parameters. It employs a variable learning rate to accelerate network convergence and an RBF online identification network to identify the parameters of the controlled object online. Based on the changes in the controlled object, the controller parameters are adjusted in real time to improve the system control performance.
1. Working principle of long-stroke pumping unit
The new long-stroke linear pumping unit uses a switched reluctance motor (SRM) as the power element. Its main working principle is to drive the drive wheel through the reduction mechanism, and then drive the counterweight box and oil rod to rise and fall through the traction rope to pump oil. This greatly simplifies the transmission mechanism and improves the overall efficiency, providing a possibility for energy saving of the pumping unit [2]. The structural block diagram of the long-stroke pumping unit is shown in Figure 1.
Figure 1. Block diagram of long-stroke pumping unit
2. Ordinary PID controller
A typical PID controller mainly consists of proportional, integral, and derivative elements that linearly combine the deviation e = given value - actual output value to form the control quantity, u, and the controller output. The control rule is as follows:
The system in this paper uses an incremental PID control algorithm:
Where e(k) is the deviation at time k; Kp, KI, and Kd are the proportional constant, integral constant, and differential constant, respectively.
3. Neural Network PID Controller
The structure of the PID control system based on neural network tuning is shown in Figure 2. By adjusting the network weights, the parameters of the controller are optimized. A variable learning rate is used to accelerate the convergence speed of the network. The RBF online identification network identifies the online parameters of the switched reluctance motor of the pumping unit. According to the torque change, the parameters of the controller are adjusted in real time [3].
Figure 2. Structure diagram of the PID control system based on RBF neural network tuning
3.1 Structure of Fuzzy RBF Neural Network
The fuzzy neural network has 4 layers, as shown in Figure 3. The first layer is the input layer; the second layer is the fuzzification layer; the third layer is the fuzzy inference layer; and the fourth layer is the output layer [6]. The structure of the fuzzy neural network is 2–6–6–3.
Figure 3. Structure of a fuzzy RBF neural network
(l) Input Layer. This layer takes the input error e and the actual system output y(k) as inputs to the next layer. The activation function is:
Therefore, the outputs of this layer are e and y(k).
(2) Blurring layer. The activation function is the membership function. Therefore, the output is:
Where i = 1, 2; j = 1, 2, ..., 6. cij and bij are the mean and standard deviation of the membership functions of the j-th fuzzy set of the i-th input variable of the Gaussian function, respectively.
(3) Fuzzy Inference Layer. The output value of this layer is obtained by multiplying the fuzzy quantities in the upper layer pairwise. Therefore, the activation function of this layer, i.e., the output, is:
Here, k = 1, 2, 3, 4, 5, 6.
(4) Output Layer. This layer outputs the parameters of the PID controller. The output value of this layer is the weights multiplied by the output of the third layer using a matrix multiplication method. Therefore, the output of this layer is:
The control quantity of incremental PID control is
The objective function is:
Where r(k) is the desired output
3.2 Adaptive Adjustment of Learning Rate
The learning rate is adaptively adjusted online based on network convergence. Specifically, during convergence, if the current error is greater than the previous error, the increase in the learning rate is reduced and the iteration restarts; conversely, the learning rate is increased.
This is the initial network learning rate.
3.3 Learning Algorithm of Fuzzy Neural Network
Since the systems are time-varying and nonlinear, the neural network needs to adjust the weights in real time, i.e., optimize wi(k) continuously. Therefore, online adjustment of the neural network is necessary. Here, the method of searching for the minimum value in the negative gradient direction is adopted.
Here, y is the dynamic factor.
3.4 Parameter Identification Based on RBF Neural Network
The long-stroke pumping unit control system employs a three-layer RBF neural network for parameter identification. X(k) = [u(k), y(k), y(k-1)]T is selected as the input vector of the identification network, where y(k) is the sampled value of the system output, i.e., the current motor output torque value. The output of the identification network is shown in Figure 2. The output u(k) of the PID controller is simultaneously transmitted to both the controlled system and the RBF identification network as the control quantity. The parameters of the fuzzy RBF control network are corrected based on the deviation between the actual output y(k+1) and the identification output. After obtaining the system sensitivity information, the RBF identification network corrects its own parameters.
4. Simulation
The dynamic response characteristics of a switched reluctance motor in a long-stroke pumping unit controlled by an RBF neural network were studied experimentally. The experimental system used a four-phase 8/6-pole switched reluctance motor with a rated power of 30 kW and a rated speed of 1500 r/min. Figure 4 shows the system response process under the control strategy proposed in this paper. Figure 5 shows the online tuning of the PID parameters by the fuzzy RBF neural network. Table 1 shows the main parameters of the switched reluctance motor in the long-stroke pumping unit based on the control method presented in this paper, along with the field test report. The data in Table 1 show that the energy-saving long-stroke pumping unit driven by the switched reluctance motor significantly outperforms the beam pumping unit in both active and reactive power; the power factor is also further improved; and the overall energy saving rate is 40.64%. Compared to ordinary pumping units, the stroke and stroke rate are greatly improved, and maintenance is very easy. Especially in terms of energy saving, the new pumping unit demonstrates many advantages.
Figure 4. System dynamic response curve (using the method described in this paper)
Figure 5. Online tuning of PID parameters using a fuzzy RBF neural network.
Table 1. Main parameters and field test results of switched reluctance motors for long-stroke pumping units.
Parameter name | Improved walking beam machine | Long-stroke energy-saving linear pumping unit |
model | CYJ14-4.8-73HB | LSCJ-14-7P |
Drive motor type | CDJT series pole changing multi-speed motor | Switched reluctance motor |
Motor rated power/KW | 50 | 30 |
Maximum stroke/m | 4.8 | 7.3 |
Input active power/kW (test) | 12.54 | 5.9 |
Input reactive power / kvar (test) | 6.63 | 0.70 |
Power factor cosφ (test) | 0.884 | 0.993 |
5. Conclusion
Neural network controllers are particularly suitable for adaptive control of nonlinear objects. This paper combines fuzzy theory, system identification, and neural networks, and utilizes their control method of adjusting PID parameters to achieve excellent control performance in the PID control system, solving the problems encountered by ordinary PID controllers in controlling time-varying and nonlinear systems. Through online identification and online adjustment of network parameters, the system can quickly and accurately track changes in its operating state. The use of a variable learning rate neural network learning and training algorithm significantly accelerates the convergence speed of the neural network. Experiments using a MATLAB simulator demonstrate that the system exhibits good dynamic characteristics, significant energy-saving effects, and good adaptability and robustness. This novel control system can greatly improve the working efficiency of oil pumping units while also achieving good energy-saving effects, attracting considerable attention from the industry.
References
[1] Zhang Xiaodong, Jia Guochao, Some thoughts on the development of oil pumping units in my country [J]. Petroleum Machinery, vol.37, NO.1, 24~27, 2008.
[2] Zhang Zhixue. Application of switched reluctance motor in energy-saving long-stroke pumping unit [J]. Petroleum Machinery vol.38, NO.5, 75-78, 2009.
[3] Xia Changliang, Xiu Jie. Adaptive PID control of switched reluctance motor based on nonlinear prediction model of RBF neural network [J]. Proceedings of the CSEE, vol.27, NO.3, Jan, 2007.
[4]Hecht-NielsenR.TheoryoftheBack-Propagation,NeuralNetwork[J].ProcIEEEonNeuralNetworks1989,Int'conf560-605.
[5] Ouyang Lu, Huang Yourui, Huang Yiqing. PID based on fuzzy RBF neural network and its application [J]. Computer Engineering, vol.34, NO.22, Nov, 2008.
[6][Zhang Hong. PID control based on fuzzy-RBF neural network optimization[J]. Journal of Xi'an University of Posts and Telecommunications, 2007, 12(5): 79-81 (in Chinese).
[7]T.Lachman,TRMohamad,andC.H.Fong.NonlinearModelingofSwitchedReluctanceMotorsUsingArtificialIntelligenceTechniques[J]IEEProl,-Electr,PowerAppl,vol.151,Jan,2004:53-60.
[8]ShufenQiandGuodongWang.PIDControlofSingleNeuronforSwitchReluctanceMotorBasedonRBFNeuralNetwork[J].IEEE(ICEEAC2010),vol.4,pp111-114,November,2010.
