Parameters Optimization of the Heating Furnace Control Systems Based on BP Neural Network Improved by Genetic Algorithm

The heating technological requirement of the conventional PID control is difficult to guarantee which based on the precise mathematical model, because the heating furnace for heating treatment with the big inertia, the pure time delay and nonlinear time-varying. Proposed one kind optimized variable method of PID controller based on the genetic algorithm with improved BP network that better realized the completely automatic intelligent control of the entire thermal process than the classics critical purporting (Z-N) method. A heating furnace for the object was simulated with MATLAB, simulation results show that the control system has the quicker response characteristic, the better dynamic characteristic and the quite stronger robustness, which has some promotional value for the control of industrial furnace.


Introduction
The conventional PID control has been widely applied because of it is good robustness and simple structure, in the industrial control used. Making P (proportion deviation, I (integral deviation) and D (differential deviation) into the linear combination is the basic idea of PID control so that it could control the controlled object. Therefore the system's performance is controlled by these three parameters. But the conventional PID control parameters are not adjusted online which is difficult to adapt to the object changes; except the PID parameters are also difficult to achieve the optimal state for the higher order or the strong coupling multivariable system which limited by the setting condition, as well as the object dynamic characteristic changes along with the environment [1][2][3] .
Proposed the BP neural network method with self-organized, self-learning and other advantages to adjustment the controller parameter online, it could meet the control requirements. BP neural network may easily lead to local minimum because of it is learning process is slow. In this paper, optimized the network threshold value and the weight by combining the genetic algorithm and the improved BP algorithm to avoid their falling into local minimum point [4][5].

PID Control of the Heating Furnace
Control system of the heating furnace is shown in Fig. 1, the PID control rule is often used in the system control rule.
The flow diagram of PID control process can be expressed in Fig. 2.
We can get the results: Kp = 2.236, Ki = 0.870, Kd = 0.267 By using parameter tuning method, that is the classical-the proportion of degrees to tuning the above closed-loop system parameters. the simulation curve is shown in Fig. 3 if the reference input is unit step signal [6].
Simulation curve shows that the Z-N tuning parameter control method is not better with big overshoot and long oscillation time which is difficult to achieve with online adjustment of PID parameters [7][8], so the method should not be used to control the parameters of heating furnace online.

Optimization Tuning of BP Neural Network PID Controller Parameters Based on the Improved Genetic Algorithm
The self-tuning PID controller of the heating furnace control system with neural network which does not rely on the knowledge of object model. When the network structure is established, the control function whether meet the requirements of control process that depends entirely on the learning algorithm [9][10][11][12].

Realization of Improved Genetic Algorithm
Generally BP network structure is shown in Fig. 4, the algorithm steps are: (2) Get the error by comparing the actual output and the desired output according to the error threshold and weight regulation; (3) Repeat the above two steps until the error up could meet the control requirements. Studies show that the BP algorithm gradually could adjust the weights and threshold of the algorithm which may result in slow learning process, excessively long training time, and easily fall into local minimum and not able to obtain the best distribution of weights and thresholds. To speed up the learning rate, BP has made some optimization algorithm, such as dynamic learning factor and inertial factor. These methods are more significant in accelerating the convergence speed of the network and avoiding falling into local minimum than others. The genetic algorithm does not request the objective function which has the continuity, moreover it may realize global optimization of the complex multi-peaks, the nonlinear and non-differentiable functions that easy to obtain the globally optimal solution or the second-best solution with better performance [13][14]. Combined the genetic algorithm and the BP algorithm will obtain the solutions of global optimization and accuracy [15][16][17]. The algorithm process is: (1) Generated initial population by coding the weight and threshold based on the optimized multiparameter mapping coding; (2) Calculate the fitness value; (3) If the contemporary individuals generate new individuals by crossing, selection and mutation, and if the algorithm process does not meet the GA stop conditions, please switch (2); otherwise, please switch (4); (4) Finding a better solution space for the genetic algorithm to and using BP algorithm to search out the optimal solution from all these small solution space [18][19][20][21].

PID Parameters Optimization
Classic incremental PID control algorithm is: Here are the steps of the algorithm: (1) Determined the network structure which using the 3-4-3 structure, the inputs are respectively, e (k), e (k)-e (k-1), e (k)-2e (k-1) e (k -2), and the output are the Kp, Ki, Kd; (2) Select the initial population N = 60, crossover probability Pc = 0.08, weight, range and threshold initialization [22][23][24][25][26]; Select the objective function of (absolute deviation points: The fitness function is: (6) (3) Get the sampled r (k) and y(k) and calculate the error at this time; (4) The network self-learning and online adjusting the weight and threshold; calculate the input and output of each layer neural network and obtain the three adjustable parameters Kp, Ki, Kd; Calculate the system output; (5) If the calculated fitness of the algorithm is not meet the system requirements, please turn into the Step 3; (6) Find the most optimal Kp, Ki, Kd and simulating the system. Simulation results show that the BP neural network by tuning with PID control system has a faster response, better dynamic characteristics and stronger robustness than the classic critical proportioning (Z-N) method [25][26].

Conclusion
Optimized BP neural network with self-organized and self-learning based on improved genetic algorithm carries on the optimization of controller parameter that could make up the insufficiency of BP neural network which is slow convergence of learning process and easily fall into local-level minimax. It could adjust the controller parameters online according to the object changes. So the method could satisfy the dynamic characteristics of the control object which changes along with the environmental variation that achieve good control effect.