What Are the Uses of Neural Networks for Prediction?
The human brain is a highly complex, non-linear, parallel information processing system. The artificial neural network, which is modeled on the working mechanism of the human brain, has strong non-linear input-output mapping capabilities. If artificial neural network is applied to predictive control, it will form Neural Network Predictive Control (NNPC).
- Artificial Neural Network (Artificial Neural Network, ANN for short) can be generally defined as: a complex network computing system composed of a large number of simple highly interconnected neurons. It is one of the main branches of intelligent control technology. It is based on the results of modern neuroscience research. Neural networks reflect some basic characteristics of human brain function and are a very important method for simulating artificial intelligence. A general form of a neural network is a machine that models the way the human brain performs a specific task or function of interest. Artificial neural networks can be implemented in hardware or software; both can be viewed as a computing model or a cognitive model. Therefore, in essence, artificial neural networks, parallel distributed processing, and neural computers are the same concepts. In the field of control, neural networks play an extremely important role. With the continuous maturity and improvement of neural network theoretical research, neural networks have been used in many aspects of the control field, such as process control, production control, pattern recognition, and decision support. and many more. [1]
- Predictive control, that is, model predictive control, is based on various predictive models.
- Neural networks have the advantages of function approximation ability, self-learning ability, complex classification function, associative memory function, fast optimization computing ability, and strong robustness and fault tolerance brought by highly parallel distributed information storage. God combined the network with model predictive control to provide a powerful tool for solving the control of complex industrial processes. [2]
- There are two forms of neural network predictive control:
- 1) Both the rolling optimization controller and the predictive model controller use neural networks;
- 2) The rolling optimization controller uses other algorithms (dynamic matrix, etc.), and the predictive model controller uses a neural network, as shown in Figure 2. [1]
- According to the method of obtaining the control law, neural network predictive control can be divided into the following two types.
- (1) Neural network predictive control based on linearization method or iterative learning. The linearization method has always been a common method for dealing with nonlinear problems. Through various linearization approximations, the solution of the nonlinear control law can be simplified and its real-time calculation speed can be improved.
- (2) Neural network predictive control based on neural network controller. This method is based on two neural networks, one is a modeling network that is used for the dynamic modeling of the process to obtain a predictive signal for the process; the other is a control network that adjusts the whole according to the driving signal corresponding to the predictive control objective function The weight of the network to obtain an approximation to the predictive control law function. [1]
Neural network predictive control structure
- The structure of the neural network predictive control system is shown in Figure 3.
- Figure 3 Structure of neural network predictive control system
- Take the secondary performance indicator function:
Neural network predictive control algorithm steps
- As can be seen from the above, the steps of the neural network predictive control algorithm can be summarized as (j = 1,2, ..., P):
- (1) Calculate the expected input reference trajectory
- (2) The neural network prediction model outputs y * (k), and the prediction output is generated by the filter
- (3) Calculation of prediction error:
- (4) Find the quadratic performance function min J (P, L, r); obtain the optimal control law u (k + j1), and use u (k) as the first control signal as the controlled object Input and go to step (2). [2]