Refractory furnace temperature control method

Abstract

The present invention provides a refractory furnace temperature control method, the method comprising: obtaining a BP neural network constructed with refractory furnace data as input and PID configuration parameters as output; using a differential evolution algorithm to solve the hidden layer weights, initial bias values ​​and hyperparameters of the BP neural network; wherein, in the solution process, the scaling ratio of the mutation process in the differential evolution algorithm changes linearly with the number of iterations in a negative correlation; assigning the hidden layer weights, initial bias values ​​and hyperparameters obtained by the solution to the BP neural network, and training the BP neural network; inputting the real-time collected refractory furnace data into the trained BP neural network to obtain real-time PID configuration parameters; inputting the desired temperature into the PID controller, and configuring the real-time PID configuration parameters to the PID controller to control the temperature of the refractory furnace. The present invention can quickly and accurately complete the initialization of the BP neural network and control the PID controller in real time.

Description

Technical Field

[0001] The invention relates to the technical field of industrial control, and in particular to a temperature control method for a refractory furnace. Background Art

[0003] In recent years, with the advancement of material technology, the fire resistance of building components in my country has been continuously improved. The building component refractory furnace is an important equipment for evaluating the fire resistance of these components. It can simulate the temperature environment of the building fire scene and test the fire resistance of different building components. Under the strict domestic standard requirements, the refractory furnace must reach the corresponding temperature within the specified time, and the control error must be within the standard range. Therefore, the control accuracy of the heating furnace in the fire resistance test becomes a key factor in the real simulation of the fire scene.

[0004] At present, the temperature control of refractory furnaces mainly relies on PID control algorithm (proportional-integral-differential controller). However, considering the time-varying characteristics and multivariable control of building component refractory furnaces, traditional PID control is difficult to effectively handle multiple input and output variables and system characteristics that change over time. In addition, the PID control algorithm has poor robustness to external interference, especially when the external environment changes or the internal parameters of the system change, the PID controller is difficult to maintain stability, resulting in poor stability of the control system and low control accuracy. Summary of the invention

[0005] The refractory furnace temperature control method provided by the present invention optimizes the initial value of the BP neural network by adopting a differential evolution algorithm, and can quickly and accurately complete the initialization of the BP neural network and control the PID controller in real time through a linear decreasing adjustment of the scaling ratio.

[0006] The present invention provides a refractory furnace temperature control method, the method comprising:

[0007] Obtain a BP neural network constructed with the expected temperature, actual temperature, temperature deviation and system control variables of the refractory furnace as input and PID configuration parameters as output;

[0008] A differential evolution algorithm is used to solve the hidden layer weights, initial bias values ​​and hyperparameters of the BP neural network; wherein, during the solving process, the scaling ratio of the mutation process in the differential evolution algorithm changes linearly with a negative correlation with the number of iterations;

[0009] Assigning the hidden layer weights, bias initial values ​​and hyperparameters obtained by solving the problem to the BP neural network, and training the BP neural network to obtain a trained BP neural network;

[0010] The expected temperature, actual temperature, temperature deviation and system control variables collected in real time are input into the trained BP neural network to obtain real-time PID configuration parameters;

[0011] The desired temperature is input into the PID controller, and the real-time PID configuration parameters are configured into the PID controller to control the temperature of the refractory furnace.

[0012] Optionally, the adopting of the differential evolution algorithm to solve the hidden layer weights, bias initial values ​​and hyperparameters of the BP neural network includes:

[0013] Obtaining a first ratio of the current number of iterations to the target number of iterations, and calculating the remaining proportion outside the first ratio;

[0014] Get the first difference between the maximum and minimum values ​​of the scaling factor;

[0015] Obtaining the product of the first difference and the remaining proportion to obtain a value to be superimposed;

[0016] The sum of the minimum value of the scaling ratio and the value to be superimposed is taken as the scaling ratio required for the current iteration.

[0017] Optionally, the adopting of the differential evolution algorithm to solve the hidden layer weights, bias initial values ​​and hyperparameters of the BP neural network includes:

[0018] Among the multiple individuals of the population retained in the previous iteration, select the individuals to be mutated and perform mutation processing to obtain a mutation vector;

[0019] According to a preset crossover probability, the individual to be mutated and the mutation vector are crossed to obtain a mutated individual;

[0020] The individuals to be mutated and the mutated individuals are selected to obtain individuals in the population retained in the current iteration.

[0021] Optionally, selecting an individual to be mutated from a plurality of individuals of the population retained in the previous iteration and performing mutation processing to obtain a mutation vector includes:

[0022] Among the multiple individuals of the population retained in the previous iteration, the individual to be mutated is selected and the first individual, the second individual and the third individual corresponding to the individual to be mutated are randomly selected; wherein each individual at least includes the potential solution vector formed by the hidden layer weights, the initial value of the bias and the hyperparameters of the BP neural network;

[0023] Multiplying the difference vector between the first individual and the second individual by the scaling factor required for the current iteration to obtain a first vector;

[0024] The third individual is added to the first vector to obtain the mutation vector.

[0025] Optionally, the step of crossing the individual to be mutated and the mutation vector according to a preset crossover probability to obtain the mutated individual includes:

[0026] Construct a second vector corresponding to the mutated individual;

[0027] Obtain a first random number corresponding to each element of the second vector; wherein the first random number is not greater than 100%;

[0028] Obtain a second random number corresponding to each element of the second vector; wherein the second random number is not greater than the number of elements in the first vector;

[0029] When the first random number is less than a preset crossover probability or the second random number is equal to the position index of the element in the second vector, the element at the corresponding position in the mutation vector is filled into the second vector as the element of the mutated individual;

[0030] When the first random number is not less than the preset crossover probability or the second random number is not equal to the position index of the element in the second vector, the element at the corresponding position in the individual to be mutated is filled into the second vector as the element of the mutated individual.

[0031] Optionally, the individuals to be mutated and the mutated individuals are selected to obtain individuals in the population retained in the current iteration, including:

[0032] Using the time multiplied by the absolute error integral criterion as the fitness function, the individual to be mutated and the individual after mutation are measured to obtain a first deviation of the individual to be mutated and a second deviation of the individual after mutation;

[0033] When the first deviation is greater than the second deviation, selecting the mutated individual as an individual in the population retained in the current iteration;

[0034] When the first deviation is not greater than the second deviation, the individual to be mutated is selected as the individual in the population retained in the current iteration.

[0035] Optionally, the individuals to be mutated and the mutated individuals are selected to obtain individuals in the population retained in the current iteration, including:

[0036] When the current iteration reaches the target number of iterations, or the population retained in the current iteration converges, the solution process is terminated;

[0037] When the current iteration does not reach the target number of iterations, or the population retained in the current iteration does not converge, the solution process continues.

[0038] Optionally, the step of inputting a desired temperature into a PID controller and configuring the real-time PID configuration parameters into the PID controller to control the temperature of the refractory furnace includes:

[0039] The output data of the PID controller is input into the temperature control system and the Smith predictor controller to realize the temperature control of the refractory furnace.

[0040] Optionally, assigning the solved hidden layer weights, bias initial values, and hyperparameters to the BP neural network, and training the BP neural network to obtain a trained BP neural network includes:

[0041] Assigning the hidden layer weights, initial bias values ​​and hyperparameters obtained by solving to the BP neural network to obtain an initialized BP neural network;

[0042] Taking the expected temperature, actual temperature, temperature deviation and system control variables of the refractory furnace as input and the corresponding PID configuration parameters as true values, a training set of the BP neural network is constructed;

[0043] The initialized BP neural network is trained using the training set of the BP neural network to obtain the trained neural network.

[0044] Optionally, the step of inputting the real-time collected expected temperature, actual temperature, temperature deviation and system control variables into the trained BP neural network to obtain real-time PID configuration parameters includes:

[0045] The expected temperature, actual temperature, temperature deviation and system control variable collected in real time are input into the trained BP neural network, and the proportional coefficient, integral coefficient and differential coefficient output by the neural network are obtained as the real-time PID configuration parameters.

[0046] In the technical solution provided by the present invention, the BP neural network parameters are automatically adjusted and optimized, thereby improving the performance and adaptability of the network. In the calculation process of the differential evolution algorithm, the scaling ratio is processed in a linear decreasing manner, so that the inertia of the population can be gradually reduced, thereby having a higher exploration ability in the early stage of the search space and a higher utilization ability in the later stage. In this way, the balance between the global and local searches of the algorithm can be enhanced, and the convergence speed and search effect of the algorithm can be improved. After the optimization of the BP neural network is completed, the configuration parameters of the PID controller are updated in real time by collecting data in real time, and the control strategy is optimized by using the learning ability of the neural network to achieve adaptive adjustment of the PID controller. Therefore, the technical solution provided by the present invention can achieve higher control accuracy and system robustness in a nonlinear, large-inertia temperature control environment, thereby improving the control effect and convergence speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of a method for controlling the temperature of a refractory furnace according to an embodiment of the present invention;

[0048] Figure 2 A flow chart of a refractory furnace temperature control method according to another embodiment of the present invention;

[0049] Figure 3 The present invention is a flow chart of a refractory furnace temperature control method for conducting a refractory test according to another embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] The embodiment of the present invention provides a method for controlling the temperature of a refractory furnace, such as Figure 1 As shown, the method includes:

[0052] Step 100, obtaining a BP neural network constructed with the expected temperature, actual temperature, temperature deviation and system control variables of the refractory furnace as inputs and the PID configuration parameters as outputs;

[0053] In some embodiments, a BP neural network is constructed with the expected temperature, actual temperature, temperature deviation and system control variables of the refractory furnace as input and the PID configuration parameters as output. The neural network can adjust the PID configuration parameters in real time according to the data of the refractory furnace, which is conducive to timely control of the refractory furnace, thereby reducing the overshoot of the control system and shortening the adjustment time of the control system.

[0054] Step 200, using a differential evolution algorithm to solve the hidden layer weights, bias initial values ​​and hyperparameters of the BP neural network; wherein, during the solving process, the scaling ratio of the mutation process in the differential evolution algorithm changes linearly with a negative correlation with the number of iterations;

[0055] In some embodiments, the learning rate and inertia coefficient of the BP neural network belong to the hyperparameters of the BP neural network, which are closely related to falling into local extreme values. Improper hyperparameter settings often cause the network to stay in local extreme values, which directly affects the control effect of BP-PID control. The differential evolution algorithm can be used to optimize hyperparameters such as the learning rate and inertia coefficient in the BP neural network, thereby improving the training effect of the neural network. However, the parameter settings of the differential evolution algorithm, especially the value of the scaling ratio, will affect the movement and search behavior of the population. The choice of scaling ratio has a great influence on the performance and convergence speed of the algorithm, and finding a suitable scaling ratio value usually requires experience and repeated trials. For the problem that a fixed scaling ratio value may cause the algorithm to converge early or converge slowly, in the early stage of the algorithm, it is necessary to ensure that the exploration ability of the population is strong, and more attention should be paid to global search. However, due to the fixed scaling ratio value, the population may gather near a local optimal solution too early, limiting the global search ability, resulting in early convergence. At the same time, the fixed scaling ratio value may cause the population to converge slowly when approaching the optimal solution. In this implementation, based on the characteristics of the differential planning algorithm, a smaller inertia weight can enhance the global search capability of the population, while a larger inertia weight can enhance the local search capability of the population. By using a linear decreasing function to process the inertia weight, the inertia of the population can be gradually reduced, so that the search space has a higher exploration capability in the early stage and a higher utilization capability in the later stage. This can enhance the balance between the global and local searches of the algorithm and improve the convergence speed and search effect of the algorithm.

[0056] Step 300, assigning the hidden layer weights, bias initial values ​​and hyperparameters obtained by solving to the BP neural network, and training the BP neural network to obtain a trained BP neural network;

[0057] In some embodiments, the differential evolution algorithm outputs the position corresponding to the population with the best fitness value as the optimization result, and this position represents the optimal initial parameters of the BP neural network obtained by population optimization. The whole process realizes the automatic adjustment and optimization of the BP neural network parameters, and improves the performance and adaptability of the network. The optimized network is trained, and the operating data of the industrial heating furnace under different working conditions are classified into four categories according to the expected value, actual value, deviation, and system control variables corresponding to the four neurons in the BP neural network input layer, and the optimized neural network is trained.

[0058] Step 400, inputting the expected temperature, actual temperature, temperature deviation and system control variables collected in real time into the trained BP neural network to obtain real-time PID configuration parameters;

[0059] In some embodiments, for the trained BP neural network, by inputting the expected temperature, actual temperature, temperature deviation and system control variables collected in real time, the corresponding real-time PID configuration parameters can be output, thereby adjusting the PID configuration parameters in real time.

[0060] Step 500, inputting a desired temperature into a PID controller, and configuring real-time PID configuration parameters into the PID controller, so as to control the temperature of the refractory furnace.

[0061] In some embodiments, after adjusting the PID configuration parameters in real time, the desired temperature is input into the PID controller to achieve control of the temperature of the refractory furnace.

[0062] In the technical solution provided by the embodiment of the present invention, the BP neural network parameters are automatically adjusted and optimized, thereby improving the performance and adaptability of the network. In the calculation process of the differential evolution algorithm, the scaling ratio is processed in a linear decreasing manner, so that the inertia of the population can be gradually reduced, so that the search space has a higher exploration ability in the early stage and a higher utilization ability in the later stage. In this way, the balance between the global and local searches of the algorithm can be enhanced, and the convergence speed and search effect of the algorithm can be improved. After the optimization of the BP neural network is completed, the configuration parameters of the PID controller are updated in real time by collecting data in real time, and the control strategy is optimized by using the learning ability of the neural network to achieve adaptive adjustment of the PID controller. Therefore, the technical solution provided by the embodiment of the present invention can achieve higher control accuracy and system robustness in a nonlinear, large inertia temperature control environment, thereby improving the control effect and convergence speed.

[0063] As an optional implementation, the method of using a differential evolution algorithm to solve the hidden layer weights, bias initial values, and hyperparameters of a BP neural network includes:

[0064] Obtain a first ratio of the current number of iterations to the target number of iterations, and calculate the remaining proportion outside the first ratio;

[0065] Get the first difference between the maximum and minimum values ​​of the scaling factor;

[0066] Obtaining the product of the first difference and the remaining proportion to obtain a value to be superimposed;

[0067] The sum of the minimum value of the scaling ratio and the value to be superimposed is taken as the scaling ratio required for the current iteration.

[0068] In some embodiments, the scaling ratio required for the current iteration may be calculated, for example, as follows:

[0069]

[0070] Among them, S max is the maximum scaling ratio, S min is the minimum scaling ratio, S t is the scaling ratio required for the current iteration, T is the target number of iterations, and t is the current number of iterations.

[0071] As an optional implementation, the method of using a differential evolution algorithm to solve the hidden layer weights, bias initial values, and hyperparameters of a BP neural network includes:

[0072] Among the multiple individuals of the population retained in the previous iteration, select the individuals to be mutated and perform mutation processing to obtain a mutation vector;

[0073] According to a preset crossover probability, the individual to be mutated and the mutation vector are crossed to obtain a mutated individual;

[0074] The individuals to be mutated and the mutated individuals are selected to obtain the individuals in the population retained in the current iteration.

[0075] As an optional implementation, the selecting an individual to be mutated from a plurality of individuals of the population retained in the previous iteration and performing mutation processing to obtain a mutation vector comprises:

[0076] Among the multiple individuals of the population retained in the previous iteration, select the individual to be mutated and randomly select the first individual, the second individual and the third individual corresponding to the individual to be mutated; wherein each individual at least includes the potential solution vector formed by the hidden layer weights, the initial value of the bias and the hyperparameters of the BP neural network;

[0077] Multiplying the difference vector between the first individual and the second individual by the scaling factor required for the current iteration to obtain a first vector;

[0078] The third individual is added to the first vector to obtain a mutation vector.

[0079] In some embodiments, when calculating the mutation vector, the calculation may be performed as follows:

[0080] U i =X r1 +S t (X r2 -X r3 )

[0081] Among them, U i is the number corresponding to the i-th individual (i.e., X i )’s mutation vector, X r1 , X r3 , X r2 They are different individuals, i≠r1≠r2≠r3.

[0082] As an optional implementation, the step of crossing the individual to be mutated and the mutation vector according to a preset crossover probability to obtain the mutated individual includes:

[0083] Construct a second vector corresponding to the mutated individual;

[0084] Obtain a first random number corresponding to each element of the second vector; wherein the first random number is not greater than 100%;

[0085] Obtain a second random number corresponding to each element of the second vector; wherein the second random number is not greater than the number of elements in the first vector;

[0086] When the first random number is less than the preset crossover probability or the second random number is equal to the position index of the element in the second vector, the element at the corresponding position in the mutation vector is filled into the second vector as the element of the mutated individual;

[0087] When the first random number is not less than the preset crossover probability or the second random number is not equal to the position index of the element in the second vector, the element at the corresponding position in the individual to be mutated is filled into the second vector as the element of the mutated individual.

[0088] In some embodiments, the fill value of each element in the second vector may be calculated according to the following formula:

[0089]

[0090] Among them, v i,d is the number corresponding to the i-th individual (i.e., X i )’s dth element of the second vector, v i,d corresponds to U i The d-th element ofi,d Corresponding to X i CR is the preset crossover probability, rand(0,1) is a random number between 0 and 1, i.e., the first random number, d rand is the second random number.

[0091] As an optional implementation manner, the individuals to be mutated and the mutated individuals are selected to obtain individuals in the population retained in the current iteration, including:

[0092] The time multiplied by the absolute error integral criterion is used as the fitness function to measure the individual to be mutated and the individual after mutation to obtain a first deviation of the individual to be mutated and a second deviation of the individual after mutation;

[0093] When the first deviation is greater than the second deviation, the mutated individual is selected as the individual in the population retained in the current iteration;

[0094] When the first deviation is not greater than the second deviation, the individual to be mutated is selected as the individual in the population retained in the current iteration.

[0095] In some embodiments, the fitness function may adopt the following function:

[0096]

[0097] Among them, J is the calculation result of the fitness function, t s is the simulation time, and e(t) is the system deviation. The ITAE fitness function is an evaluation index commonly used in control systems and optimization problems. It is mainly used to measure the error between the actual output and the expected output and integrate the error.

[0098] As an optional implementation manner, the individuals to be mutated and the mutated individuals are selected to obtain individuals in the population retained in the current iteration, including:

[0099] When the current iteration reaches the target number of iterations, or the population retained in the current iteration converges, the solution process ends;

[0100] When the target number of iterations is not reached in the current iteration or the population retained in the current iteration has not converged, the solution process continues.

[0101] As an optional implementation, Figure 2 As shown, the step of inputting the desired temperature into the PID controller and configuring the real-time PID configuration parameters into the PID controller to control the temperature of the refractory furnace includes:

[0102] The output data of the PID controller is input into the temperature control system and the Smith predictor controller to realize the temperature control of the refractory furnace.

[0103] In some embodiments, it can be seen from the previous research content that there is a certain delay in the temperature control system during the temperature control process. Next, the time lag term of the system will be eliminated, and it is proposed to combine the Smith predictive control with the PID controller, so that better control performance can be obtained when processing a system with time lag. The control system is composed of the Smith predictive controller, the differential evolution algorithm, the BP neural network and the PID control algorithm. By connecting the Smith predictive controller in reverse parallel to the temperature control system, the Smith predictive controller is used to compensate the control error, thereby offsetting the influence caused by the pure hysteresis characteristic, significantly reducing the overshoot, and accelerating the adjustment process, and improving the control quality. The differential evolution algorithm, the BP neural network and the PID control algorithm make the system parameter adjustment more real-time by taking the optimization results of the PID parameter changes under different real-time states as the three input parameters of the PID, thereby achieving the purpose of improving the dynamic performance of the system. The temperature controller based on the Smith predictive controller can not only eliminate the negative impact of the pure hysteresis term on the system control quality, but also does not affect the adjustment of the PID parameters by the BP neural network. In theory, it can achieve the improvement of control accuracy and stability while solving the problem of large temperature control inertia. By using the differential evolution algorithm to optimize the BP neural network and further optimizing the PID controller parameters through the BP neural network, the best parameter combination can be automatically searched to achieve better control effects. In addition, the introduction of the Smith predictor controller to solve the time lag problem of the temperature control system can increase the prediction ability and stability of the system. It provides a more efficient and accurate solution for temperature prediction and control. By optimizing the control parameters and introducing predictive control, a more accurate and flexible temperature control solution can be provided, thereby improving the performance and response speed of the system.

[0104] As an optional implementation, assigning the hidden layer weights, bias initial values, and hyperparameters obtained by solving the problem to the BP neural network, and training the BP neural network to obtain the trained BP neural network includes:

[0105] Assigning the hidden layer weights, initial bias values ​​and hyperparameters obtained by solving to the BP neural network to obtain an initialized BP neural network;

[0106] The training set of BP neural network is constructed by taking the expected temperature, actual temperature, temperature deviation and system control variables of the refractory furnace as input and the corresponding PID configuration parameters as true values;

[0107] The training set of the BP neural network is used to train the initialized BP neural network to obtain a trained neural network.

[0108] As an optional implementation, the inputting of the expected temperature, actual temperature, temperature deviation and system control variables collected in real time into the trained BP neural network to obtain real-time PID configuration parameters includes:

[0109] The expected temperature, actual temperature, temperature deviation and system control variables collected in real time are input into the trained BP neural network, and the proportional coefficient, integral coefficient and differential coefficient output by the neural network are obtained as real-time PID configuration parameters.

[0110] As an optional implementation, after inputting the desired temperature into the PID controller and configuring the real-time PID configuration parameters into the PID controller to control the temperature of the refractory furnace, the method further includes:

[0111] The expected temperature, actual temperature, temperature deviation and system control variables are collected in real time and input into the trained BP neural network, so that the trained BP neural network can update the real-time PID configuration parameters.

[0112] The application scenarios of various embodiments of the present invention are described below with exemplary implementations. Figure 3 As shown, the details are as follows:

[0113] During the test of the temperature control system, the initial setting is first performed, including accurately setting the temperature in the furnace and adjusting the gas nozzles to ensure that they are in a state suitable for ignition. Then check whether all the necessary safety and operating conditions before the system is ignited have been met, such as whether the gas pressure and combustion air flow are within a reasonable range. After verifying that these conditions are met, the system will ignite the gas and then immediately check whether the ignition is successful. If the ignition fails, the system will reset, check and adjust the initial settings again, and then try to re-ignite. If the ignition is successful, the refractory furnace is monitored in real time, the temperature in the furnace is detected, and it is continuously compared with the preset target temperature. If the measured temperature does not reach the predetermined target, the control system will adjust the gas flow or the air door opening to increase or decrease the size of the flame, thereby adjusting the temperature in the furnace. This adjustment process is a dynamic and continuous adjustment performed by the refractory furnace temperature control method provided by each embodiment of the present invention, thereby ensuring that the furnace temperature can accurately follow the set temperature curve during the entire test. The control system will continuously loop this monitoring and adjustment process until the conditions for the end of the test are detected, such as the time reaching a predetermined duration or other set termination conditions. Once it is confirmed that the test should be ended, the control system will execute a safe shutdown procedure, slowly reducing the gas supply and gradually lowering the temperature in the furnace until the furnace is safely shut down, ultimately completing the entire fire resistance test process.

[0114] Those skilled in the art can understand that all or part of the processes in the above method embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes in the above method embodiments. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0115] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A refractory furnace temperature control method, characterized in that: The method comprises: Obtain a BP neural network constructed with the expected temperature, actual temperature, temperature deviation and system control variables of the refractory furnace as input and PID configuration parameters as output; A differential evolution algorithm is used to solve the hidden layer weights, initial bias values ​​and hyperparameters of the BP neural network; wherein, during the solving process, the scaling ratio of the mutation process in the differential evolution algorithm changes linearly with a negative correlation with the number of iterations; Assigning the hidden layer weights, bias initial values ​​and hyperparameters obtained by solving the problem to the BP neural network, and training the BP neural network to obtain a trained BP neural network; The expected temperature, actual temperature, temperature deviation and system control variables collected in real time are input into the trained BP neural network to obtain real-time PID configuration parameters; The desired temperature is input into the PID controller, and the real-time PID configuration parameters are configured into the PID controller to control the temperature of the refractory furnace.

2. The method according to claim 1, characterized in that The method of using the differential evolution algorithm to solve the hidden layer weights, bias initial values ​​and hyperparameters of the BP neural network includes: Obtaining a first ratio of the current number of iterations to the target number of iterations, and calculating the remaining proportion outside the first ratio; Get the first difference between the maximum and minimum values ​​of the scaling factor; Obtaining the product of the first difference and the remaining proportion to obtain a value to be superimposed; The sum of the minimum value of the scaling ratio and the value to be superimposed is taken as the scaling ratio required for the current iteration.

3. The method according to claim 2, characterized in that The method of using the differential evolution algorithm to solve the hidden layer weights, bias initial values ​​and hyperparameters of the BP neural network includes: Among the multiple individuals of the population retained in the previous iteration, select the individuals to be mutated and perform mutation processing to obtain a mutation vector; According to a preset crossover probability, the individual to be mutated and the mutation vector are crossed to obtain a mutated individual; The individuals to be mutated and the mutated individuals are selected to obtain individuals in the population retained in the current iteration.

4. The method according to claim 3, characterized in that The step of selecting an individual to be mutated from among the multiple individuals of the population retained in the previous iteration and performing mutation processing to obtain a mutation vector includes: Among the multiple individuals of the population retained in the previous iteration, the individual to be mutated is selected and the first individual, the second individual and the third individual corresponding to the individual to be mutated are randomly selected; wherein each individual at least includes the potential solution vector formed by the hidden layer weights, the initial value of the bias and the hyperparameters of the BP neural network; Multiplying the difference vector between the first individual and the second individual by the scaling factor required for the current iteration to obtain a first vector; The third individual is added to the first vector to obtain the mutation vector.

5. The method according to claim 3, characterized in that: The step of crossing the individual to be mutated and the mutation vector according to a preset crossover probability to obtain a mutated individual includes: Construct a second vector corresponding to the mutated individual; Obtain a first random number corresponding to each element of the second vector; wherein the first random number is not greater than 100%; Obtain a second random number corresponding to each element of the second vector; wherein the second random number is not greater than the number of elements in the first vector; When the first random number is less than a preset crossover probability or the second random number is equal to the position index of the element in the second vector, the element at the corresponding position in the mutation vector is filled into the second vector as the element of the mutated individual; When the first random number is not less than the preset crossover probability or the second random number is not equal to the position index of the element in the second vector, the element at the corresponding position in the individual to be mutated is filled into the second vector as the element of the mutated individual.

6. The method according to claim 3, characterized in that The individuals to be mutated and the mutated individuals are selected to obtain individuals in the population retained in the current iteration, including: Using the time multiplied by the absolute error integral criterion as the fitness function, the individual to be mutated and the individual after mutation are measured to obtain a first deviation of the individual to be mutated and a second deviation of the individual after mutation; When the first deviation is greater than the second deviation, selecting the mutated individual as an individual in the population retained in the current iteration; When the first deviation is not greater than the second deviation, the individual to be mutated is selected as the individual in the population retained in the current iteration.

7. The method according to claim 3, characterized in that The individuals to be mutated and the mutated individuals are selected to obtain individuals in the population retained in the current iteration, including: When the current iteration reaches the target number of iterations, or the population retained in the current iteration converges, the solution process is terminated; When the current iteration does not reach the target number of iterations, or the population retained in the current iteration does not converge, the solution process continues.

8. The method according to claim 1, characterized in that The step of inputting the desired temperature into the PID controller and configuring the real-time PID configuration parameters into the PID controller to control the temperature of the refractory furnace includes: The output data of the PID controller is input into the temperature control system and the Smith predictor controller to realize the temperature control of the refractory furnace.

9. The method according to claim 1, characterized in that: The step of assigning the hidden layer weights, bias initial values, and hyperparameters obtained by solving the problem to the BP neural network, and training the BP neural network to obtain a trained BP neural network includes: Assigning the hidden layer weights, initial bias values ​​and hyperparameters obtained by solving to the BP neural network to obtain an initialized BP neural network; Taking the expected temperature, actual temperature, temperature deviation and system control variables of the refractory furnace as input and the corresponding PID configuration parameters as true values, a training set of the BP neural network is constructed; The initialized BP neural network is trained using the training set of the BP neural network to obtain the trained neural network.

10. The method according to claim 1, characterized in that The method of inputting the real-time collected expected temperature, actual temperature, temperature deviation and system control variables into the trained BP neural network to obtain real-time PID configuration parameters includes: The expected temperature, actual temperature, temperature deviation and system control variable collected in real time are input into the trained BP neural network, and the proportional coefficient, integral coefficient and differential coefficient output by the neural network are obtained as the real-time PID configuration parameters.

Images