Multi-layer fuzzy nested PID temperature control method and system based on improved LSTM neural network
Through the improved LSTM neural network and multi-layer fuzzy nested PID control method, combined with SC-RIME optimization algorithm and horizontal crossover strategy, the problem of inaccurate temperature control in multi-input and multi-output systems is solved, and the existing temperature control method is realized. High precision, high speed and efficiency temperature control and safety improvement.
Patent Information
- Application Number
- CN202411192944.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-28
- Publication Date
- 2024-11-12
AI Technical Summary
The existing temperature control methods lack the establishment of the controlled object model, cannot adapt to switching in different working conditions, and lack control accuracy and safety, especially in multi-input and multi-output systems, there is a problem of inaccurate temperature control.
Adopting an improved LSTM neural network and multi-layer fuzzy nested PID control method, combined with SC-RIME optimization algorithm and horizontal crossover strategy, the adaptive model establishment and online switching of the controlled object are realized, and the working condition mode is judged through the decision unit and optimized. Kp, Ki, Kd parameters improve the accuracy and efficiency of temperature control.
High-precision, high-speed and efficiency temperature control under SISO, SIMO, MISO, and MIMO conditions are realized, which enhances the system's adaptability and safety, and ensures multi-stage fusion and gradient adjustment of temperature control.
Smart Images

Figure CN118938650A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of temperature control, and in particular to a multi-layer fuzzy nested PID temperature control method and system based on an improved LSTM neural network. Background Art
[0002] Temperature control is an important technology in modern science and technology, especially in the precision manufacturing industry. It is one of the important technologies to further improve manufacturing precision. It involves measurement technology, control theory, actuators, materials and thermal management technology. Its core is to maintain or change the temperature of the system through measurement, feedback and regulation to meet the desired needs.
[0003] In industrial production, temperature control technology can improve product quality, optimize production processes and efficiency, and reduce production costs; in energy and power, temperature control technology can reduce energy consumption, reduce carbon emissions, and improve the efficiency and life of new energy technologies; in medical rehabilitation, temperature control technology can improve the stability of equipment and the safety of patients, and create a reasonable temperature environment for cell culture and bio-enzyme reactions; in experimental verification, temperature control is used in testing superconducting materials, testing chemical reaction kinetics, and testing the thermal properties of materials and devices.
[0004] The Chinese invention patent with the publication number of "CN118244815A" and the name of "A Simple Intelligent Temperature Control System" discloses a simple intelligent temperature control system, which proposes to use the temperature detection module to monitor the ambient temperature in real time, the control module to adjust the temperature, the execution module to execute the instructions of the control module, the adaptive module to automatically adjust according to the change of the ambient temperature, the alarm module to issue an alarm when the preset temperature is reached, and the intelligent learning module to continuously optimize and adjust the control strategy. However, this temperature control method lacks the establishment of a controlled object model, and cannot adapt to the timely switching according to the control requirements under the single input single output (SISO), single input multiple output (SIMO), multiple input single output (MISO), and multiple input multiple output (MIMO) working conditions. It cannot perform self-learning on the data collected by the I / O port, and the scope of application is relatively single.
[0005] The Chinese invention patent with publication number "CN117906198A" and titled "A device for constant temperature control of floor heating and its operating method" discloses a device for constant temperature control of floor heating and its operating method, and proposes a three-part temperature control system including a switch relay, a cycle time control module and a normally open relay. When the indoor temperature reaches the set temperature, the cycle time control module starts to intermittently supply hot water to the floor heating coil to maintain the stability of the floor temperature. However, directly controlling the power switch through temperature measurement and set temperature will cause large lags and fluctuations in ambient temperature. Frequent power-on of high-power switches will also accelerate system aging and bring safety hazards.
[0006] The Chinese invention patent with publication number "CN118216804A" and titled "A water dispenser and its water outlet temperature control method" discloses a water dispenser and its water outlet temperature control method, and proposes a control method based on real-time monitoring of the water outlet temperature and high-temperature or low-temperature water compensation in combination with the target water outlet temperature to achieve the set target water outlet temperature. However, this method lacks accurate modeling of the controlled object, so the accuracy of the water outlet temperature cannot be guaranteed. In addition, the singleness of the control algorithm also limits its ability to meet the temperature requirements of multiple channels and multiple water outlets. Summary of the invention
[0007] In order to solve the problem that the temperature control method in the above-mentioned prior art lacks the establishment of a controlled object model, and cannot adapt to timely switching according to control requirements under single-input single-output (SISO), single-input multiple-output (SIMO), multiple-input single-output (MISO), and multiple-input multiple-output (MIMO) working conditions, the data collected by the I / O port cannot be self-learned, the scope of application is relatively single, and the lack of accurate modeling of the controlled object leads to the inability to guarantee the accuracy of the temperature of the controlled object; frequent high-power switch power-on will also accelerate system aging and bring safety hazards. The present invention proposes a multi-layer fuzzy nested PID temperature control method and system based on an improved LSTM neural network, which realizes the integration of temperature control algorithms for single-input single-output (SISO), single-input multiple-output (SIMO), multiple-input single-output (MISO), and multiple-input multiple-output (MIMO) systems, aiming to solve the adaptive switching of multiple working conditions in model-free temperature control and realize high-precision, high-speed, and efficient temperature control.
[0008] The present invention is implemented by the following technical scheme: comprising a detection module connected to the controlled object, the detection module is used to output the real-time collected temperature signal T and power signal P to the processor, and also comprising
[0009] Input module, used to input the desired temperature set Optimize particle position dimension D and heating power limit P max , particle optimization range [0, Z], particle population size M, set maximum number of iterations, set fitness standard value, total temperature control time A and single temperature control cycle time t are output to the corresponding module, Z is the maximum value of the particle optimization range, which is an integer greater than 0; the input module outputs the expected temperature set Output to the processor and weighted summation module, the heating power limit P maxOutput to the SC-RIME optimization algorithm module, output the total temperature control time A and the single temperature control cycle time t to the processor, PID controller and weighted summation module, and output the optimal particle position dimension D, particle optimization range [0-Z], particle population size M, the set maximum number of iterations, and the set fitness standard value to the SC-RIME optimization algorithm module;
[0010] The processor is used to input the total temperature control time A, the single temperature control cycle time t, the temperature signal T, the power signal P and the expected temperature set Perform data calculation and fitting to obtain the temperature set of the controlled object Power Collection Transfer function G(s), desired temperature set and temperature collection The error set Error change rate set And transfer function G(s), error set Temperature Collection and power collection Output to the PID controller, the temperature set Power Collection Error Set Error change rate set Output to the multi-layer fuzzy nesting module, the temperature set and power collection Output to the improved LSTM neural network module;
[0011] SC-RIME optimization algorithm module, used to compare K p , integral K i , differential K d Iterate and optimize the number of hidden layer units, learning rate, and training times to obtain the ratio K p优 , integral K i优 , differential K d优 And the optimal number of hidden layer units, learning rate, and number of training times, the ratio K p优 , integral K i优 , differential K d优 Output to the PID controller, and output the optimal number of hidden layer units, learning rate, and training times to the improved LSTM neural network module;
[0012] Multi-layer fuzzy nesting module is used to set the temperature according to the input Power Collection Error Set and error rate of change set Get the ratio K p优 , integral K i优 and differential Kd优 The corresponding amplification factor is output to the PID controller;
[0013] PID controller, used to adjust the input proportional K p优 , integral K i优 and differential K d优 And the ratio K p优 , integral K i优 and differential K d优 The corresponding amplification factor obtains the theoretical input power set And output to the improved LSTM neural network and weighted summation module;
[0014] Improved LSTM neural network module, used to store improved LSTM neural network, based on the optimal number of hidden layer units, learning rate, number of training times, theoretical input power set Temperature Collection Power Collection Get the predicted temperature set And output to the weighted summation module;
[0015] Weighted sum module, used to predict the temperature set according to the input Theoretical input power set and the desired temperature set Get the final input power set And output to the controlled object.
[0016] Furthermore, the SC-RIME optimization algorithm module is used to store the SC-RIME optimization algorithm, and the SC-RIME optimization algorithm is obtained by replacing the random number initialization strategy in the frost and ice optimization algorithm RIME with the SINE chaotic map initialization strategy. At the same time, the horizontal cross mechanism in the vertical and horizontal cross optimization algorithm CSO is added before the greedy selection mechanism of the frost and ice optimization algorithm RIME.
[0017] Furthermore, the multi-layer fuzzy nested module includes a first decision unit module, a first layer fuzzy module, a second decision unit module and a second layer fuzzy module; the first decision unit module is used to determine which of the four systems SISO, SIMO, MISO and MIMO the controlled object belongs to.
[0018] The present invention also provides a temperature control method based on the improved LSTM neural network multi-layer fuzzy nested PID temperature control system according to the present invention, comprising the following steps:
[0019] S1. The processor records the input temperature signal T and power signal P in time series to obtain a time series prediction data set, which includes the temperature set of the controlled object. and power collection The data in the time series prediction data set are fitted by the least square method to obtain the transfer function G(s) of the controlled object. and the desired temperature set Get the error set and the error rate of change set
[0020] S2, SC-RIME optimization algorithm module and PID controller cooperate to iterate and optimize to obtain the proportion K p优 , integral K i优 , differential K d优 ;
[0021] S3, multi-layer fuzzy nesting module according to the input error set Error change rate set Temperature Collection and power collection Get the ratio K p优 , integral K i优 and differential K d优 The corresponding amplification factor is output to the PID controller;
[0022] S4, PID controller according to the input ratio K p优 , integral K i优 and differential K d优 And the corresponding amplification factor to obtain the theoretical input power set And output to the improved LSTM neural network module and weighted summation module;
[0023] S5, the SC-RIME optimization algorithm module and the improved LSTM neural network module cooperate to obtain the optimal number of hidden layer units, learning rate, and training times and output them to the improved LSTM neural network module;
[0024] S6. Build an improved LSTM neural network dataset and set the input temperature and power collection Processing, the temperature collection and power collection According to 15 consecutive temperature Plus 15 consecutive power Plus a theoretical input power P i , there are 31 data in total that constitute a set of data;
[0025] S7, Improved LSTM neural network based on the optimal number of hidden layer units, learning rate, number of training times, and temperature set of input Power Collection And the data set to get the predicted temperature set And output to the weighted summation module;
[0026] S8, weighted sum module for the expected temperature set of the input Theoretical input power set and predicted temperature ensemble Perform weighted summation to obtain the final input power set
[0027] S9, the final input power set obtained in step S8 Input the controlled object, loop steps S1 to S9, perform cyclic temperature control, and terminate the temperature control program after reaching the set total temperature control time A.
[0028] Furthermore, the steps of iterative optimization in which the SC-RIME optimization algorithm module and the PID controller cooperate in step S2 are as follows:
[0029] S2.1, SINE chaotic mapping initialization strategy initializes the positions of all particles in the population to obtain the initialized particle population
[0030] The formula for the SINE chaos map initialization strategy is as follows:
[0031]
[0032] Where n = (1, 2, 3...), is the chaotic system parameter, with a value of 0.99, x n is a random number between [0,1], and formula (1) can be used to obtain x 1 、x 2 , ..., x D , where the value of the optimal particle position dimension D is 3, that is, according to formula (1), we get x 1 、x 2 、x 3 Let x 1 、x 2 、x 3 Multiply each by the maximum value Z of the particle search range to get the particle position (x 1 *Z,x 2 *Z,x 3 *Z), let K p =x 1 *Z, K i =x 2 *Z and K d =x 3 *Z gets several groups of proportions K p , integral K i and differential K d ;
[0033] S2.2, using the soft frost growth strategy and the hard frost puncture strategy to further search the initialized particle population to obtain a new particle population;
[0034] S2.3, the horizontal crossover mechanism further crosses the new particle population in step S2.2 to obtain a particle position, that is, a set of proportions K p , integral K i and differential K d And output to the PID controller;
[0035] S2.4, PID controller according to the input temperature control total time A, step S2.3 obtained a set of proportions K p , integral K i , differential K d , transfer function G(s), error set Temperature Collection and power collection Get the simulation power set of the time B when the set temperature T' is reached for the first time and the total temperature control time A Simulation Error Collection And output to the SC-RIME optimization algorithm module;
[0036] S2.5, the greedy selection mechanism updates the optimal particle position, and the input of the first time to reach the set temperature T'B, the total temperature control time A simulation power set Simulation Error Collection Substitute the following fitness evaluation condition function formula to obtain the fitness value f 1 ,
[0037]
[0038] Fitness value f 1 It is necessary to satisfy the fitness evaluation condition function formula at the same time, according to the fitness value f 1 The size is used to make a decision, that is, the fitness value f of the new position 1 Less than the fitness value f of the old position 1 , then update the particle position, otherwise keep the old particle position;
[0039] S2.6, determine whether the optimization conditions are met. If the number of iterations is greater than the set maximum number of iterations or the fitness value in step 2.5 is less than the set fitness standard value, then output the optimal particle position after optimization and obtain the ratio K p优 , integral K i优 and differential K d优 Otherwise, execute step S2.2 and repeat steps S2.2-S2.6 until the optimization condition is met. p优 , integral K i优 and differential Kd优 Output to the PID controller.
[0040] Furthermore, the step S3 obtains the ratio K p优 , integral K i优 and differential K d优 The corresponding steps of the amplification factor are as follows,
[0041] S3.1, the first decision unit module is based on the input temperature set and power collection Judge the working condition of the controlled object. If it is a SISO system or a MISO system, directly input the error set and error rate of change set Input the second decision unit module and execute step 3.3; otherwise, determine the error set Is it within the range of [-λ℃, λ℃], and proceed to step 3.2;
[0042] S3.2. If the error set In the range of [-λ℃, λ℃], the error set and error rate of change set Output to the first layer fuzzy module for the first fine fuzzy control; if it is not within the range, the first coarse fuzzy control is performed; the total error set is obtained Total error rate of change collection And output to the second decision unit module to execute step 3.3;
[0043] S3.3, the second decision unit module determines the error set of the input Or the total error set Is it within the range of [-λ℃, λ℃]? If so, the input set is output to the second-layer fuzzy module for the second fine fuzzy control; otherwise, the second coarse fuzzy control is performed; the ratio K is obtained. p优 , integral K i优 and differential K d优 The corresponding magnification factor.
[0044] Furthermore, the steps of obtaining the optimal number of hidden layer units, learning rate, and training times in step S5 are as follows:
[0045] S5.1, SINE chaotic mapping initialization strategy initializes the positions of all particles in the population, and obtains the initialized particle population, several groups x 1 、x 2 、x 3 and the corresponding particle position (x 1 *Z,x 2 *Z,x 3 *Z), let the number of hidden layer units = (48*x1 +2) Round up; let learning rate = x 2 / 100; let the number of training times = (50*x 3 +10) to get several sets of hidden layer unit numbers, learning rates and training times;
[0046] S5.2, using the soft frost growth strategy and the hard frost puncture strategy to further search the initialized particle population to obtain a new particle population;
[0047] S5.3, the horizontal crossover mechanism further crosses and optimizes the new particle population in step S5.2 to obtain a particle position, that is, a set of hidden layer unit numbers, learning rates and training times, and outputs them to the improved LSTM neural network module;
[0048] S5.4, the improved LSTM neural network module is based on the input temperature set Power Collection Step S5.3 obtains a set of hidden layer unit numbers, learning rates, and training times, and obtains the number of validation set samples m and iterative prediction temperature T i , measured temperature T i 'And output to the SC-RIME optimization algorithm module;
[0049] S5.5, Greedy selection mechanism updates the optimal particle position, inputs the number of validation set samples m, iterative prediction temperature T i , measured temperature T i 'Substitute the following fitness evaluation condition function formula to obtain the fitness value f 2 ,
[0050]
[0051] According to the fitness value f 2 The size is used to make a decision, that is, the fitness value f of the new position 2 Less than the fitness value f of the old position 2 , then update the particle position, otherwise keep the old particle position;
[0052] S5.6, determine whether the optimization conditions are met. If the number of iterations is greater than the set maximum number of iterations or the fitness value in step 5.5 is less than the set fitness standard value, then output the optimal solution particle position after optimization to obtain the optimal number of hidden layer units, learning rate and number of training times; otherwise, execute step S5.2, and loop steps S5.2-S5.6 until the optimization conditions are met; output the optimal number of hidden layer units, learning rate and number of training times to the improved LSTM neural network module.
[0053] Furthermore, the horizontal crossover mechanism is performed as follows:
[0054] NX(i,d)=r 1 ×X(i,d)+(1-r 1 )×X(j,d)+c 1 ×(X(i,d)-X(j,d))
[0055] NX(j,d)=r 2 ×X(j,d)+(1-r 2 )×X(i,d)+c 2 ×(X(j,d)-X(i,d))
[0056] i,j∈N(1,M),d∈N(1,D)
[0057] where c 1 、c 2 is a random number between (-1,1); r 1 、r 2 is a random number between (0,1); D is the dimension of the variable; M is the size of the population particles; X(i,d) and X(j,d) are the d-th dimensions of X(i) and X(j); NX(i,d) and NX(j,d) are the d-th offspring generated by X(i,d) and X(j,d) through the horizontal crossover mechanism.
[0058] Furthermore, the value of λ is 5.
[0059] Further, in step S8, according to formula Get the final input power set Where α = (βη) / η max , β is the gain factor, η max is the maximum number of cycles, and η is the current number of cycles.
[0060] Compared with the prior art, the present invention has the following beneficial effects:
[0061] 1. The multi-layer fuzzy nested PID algorithm and the improved LSTM neural network in the present invention can adaptively identify and switch the working mode online when facing SISO, SIMO, MISO, and MIMO working conditions through the decision-making unit according to the composition of the input and output data.
[0062] 2. The method of establishing the model of the controlled object in the present invention is to first fuzzify and then to make it precise. Fuzziness is to make preliminary control parameters according to the basic properties of the controlled object in the early stage of control; precision is to continuously establish and update the model of the controlled object according to the real-time data collected by each channel as the temperature control process is cyclical. In this way, an accurate model of the controlled object is obtained, and the temperature control performance is improved.
[0063] 3. The present invention adopts chaos mapping and horizontal cross strategy in the frost optimization algorithm RIME to improve the SC-RIME optimization algorithm, which has the advantages of K p , K i and K d More convenient optimization capability for parameter values.
[0064] 4. The multi-layer fuzzy nested PID algorithm in the present invention starts with the error er and its change rate ec in fuzzy control, and constructs a series-parallel structure of fine fuzzy control and rough fuzzy control according to the range of er [-λ℃, λ℃]. It has better temperature control performance.
[0065] 5. The SC-RIME optimization algorithm is used to optimize the number of hidden layer units, learning rate and number of training times of the improved LSTM neural network to enhance the fitting ability of the network. A structure of 15 consecutive sets of temperature plus 15 consecutive powers plus the input power of the next temperature control cycle is proposed, with a total of 31 data sets to build a data set. This improves the accuracy of the improved LSTM neural network in time series temperature prediction.
[0066] 6. The present invention comprehensively considers the theoretical input power of the controlled object obtained by the multi-layer fuzzy nested PID algorithm and the output temperature predicted by the improved LSTM neural network, so as to obtain the final input power after reliable and optimized processing, and improve the control performance of the temperature control system.
[0067] 7. During the temperature control process, the present invention can achieve the integration and nesting of multiple stages of heating, constant temperature and cooling by setting the expected temperature, so that these stages can be organically combined together, and under the premise of ensuring the high efficiency and accuracy of temperature control, the temperature control needs of multiple temperature control stages are met.
[0068] 8. The temperature control gradient of the present invention is adjustable. According to the requirements of general temperature control and parts thermal testing, the temperature rise gradient, constant temperature time, temperature drop gradient and temperature control cycle of the controlled object can be set manually to meet the specific expected temperature control requirements. BRIEF DESCRIPTION OF THE DRAWINGS
[0069] Figure 1 This is a block diagram of the temperature control system in the present invention.
[0070] Figure 2 Flow chart of the temperature control method of the present invention.
[0071] Figure 3 SC-RIME optimization algorithm for K p优 , K i优 , K d优 Flowchart for optimization. DETAILED DESCRIPTION
[0072] Advantages and features of the present invention will be illustrated and explained by the following non-limiting description of preferred embodiments thereof, which are given by way of example only with reference to the accompanying drawings.
[0073] like Figure 1 As shown, the present invention provides a multi-layer fuzzy nested PID temperature control system based on an improved LSTM neural network, comprising a detection module connected to a controlled object, the detection module outputs a temperature signal T and a power signal P of the controlled object collected in real time to a processor, and an input module outputs an input desired temperature set Optimize particle position dimension D and heating power limit P max , particle optimization range [0, Z], particle population size M, set maximum number of iterations, set fitness standard value, total temperature control time A and single temperature control cycle time t are output to the corresponding module, where Z is the maximum value of the particle optimization range, which is an integer greater than 0; the input module outputs the expected temperature set Output to the processor and weighted summation module, the heating power limit P max Output to the SC-RIME optimization algorithm module, output the total temperature control time A and the single temperature control cycle time t to the processor, PID controller and weighted summation module, and output the optimal particle position dimension D, particle optimization range [0-Z], particle population size M, the set maximum number of iterations, and fitness standard value to the SC-RIME optimization algorithm module.
[0074] The processor inputs the total temperature control time A, the single temperature control cycle time t, the temperature signal T, the power signal P and the expected temperature set Perform data calculation and fitting to obtain the temperature set of the controlled object Power Collection Transfer function G(s), desired temperature set and temperature collection The error set Error change rate set And transfer function G(s), error set Temperature Collection and power collection Output to the PID controller, the temperature set Power Collection Error Set Error change rate set Output to the multi-layer fuzzy nesting module, the temperature set and power collection Output to the improved LSTM neural network module, the temperature set Power Collection Composed time series forecasting dataset.
[0075] The SC-RIME optimization algorithm module is based on the input optimal particle position dimension D and heating power limit P max , particle optimization range [0, Z], particle population size M, set maximum number of iterations and set fitness standard value, and get the proportion K in the iteration process p , integral K i , differential K d And the number of hidden layer units, learning rate, and number of training times, the ratio K p , integral K i , differential K d Output to the PID controller, output the number of hidden layer units, learning rate, and training times to the improved LSTM neural network, and the simulation power set of the time B for the first reaching the set temperature T' and the total temperature control time A fed back by the PID controller Simulation Error Collection As well as the number of validation set samples m fed back by the improved LSTM neural network, the iterative prediction temperature T i , measured temperature T i 'Perform iterative optimization to obtain the ratio K p优 , integral K i优 , differential K d优 And the optimal number of hidden layer units, learning rate, number of training times, and the ratio K p优 , integral K i优 and differential K d优 Output to the PID controller, and output the optimal number of hidden layer units, learning rate, and training times to the improved LSTM neural network module. The SC-RIME optimization algorithm module is used to store the SC-RIME optimization algorithm, and the SC-RIME optimization algorithm is replaced by the SINE chaotic map initialization strategy by the random number initialization strategy in the frost and ice optimization algorithm RIME. At the same time, the horizontal crossover mechanism in the vertical and horizontal crossover optimization algorithm CSO is added before the greedy selection mechanism of the frost and ice optimization algorithm RIME.
[0076] The PID controller is based on the input total temperature control time A, single temperature control cycle time t, and ratio K p , integral K i , differential K d , transfer function G(s), error set Temperature Collection and power collection Get the simulation power set of the time B when the set temperature T' is reached for the first time and the total temperature control time A Simulation Error Collection And feedback to the SC-RIME optimization algorithm module, according to the input ratio K p优 , integral K i优and differential K d优 And the ratio K p优 , integral K i优 and differential K d优 The corresponding amplification factor obtains the theoretical input power set And output to the improved LSTM neural network and weighted summation module.
[0077] The multi-layer fuzzy nesting module includes a first decision unit module, a first layer fuzzy module, a second decision unit module and a second layer fuzzy module. The first decision unit module is based on the input temperature set. Power Collection Determine the system mode of the controlled object. If it is a SISO or MISO system, directly set the input error Error change rate set Output to the second decision unit module, otherwise the input error set Error change rate set Output to the first layer fuzzy module; the first layer fuzzy module is based on the input error set Error change rate set Get the total error set Total error rate of change collection And output to the second decision unit module, the second decision unit module outputs the input set to the second layer fuzzy module to obtain the ratio K p优 , integral K i优 and differential K d优 The corresponding amplification factor is output to the PID controller.
[0078] The improved LSTM neural network module is used to store the improved LSTM neural network according to the input temperature set Power Collection The number of hidden layer units, learning rate, and training times are used to obtain the number of validation set samples m and the iterative prediction temperature T i , measured temperature T i 'And feedback to the SC-RIME optimization algorithm module, according to the input optimal number of hidden layer units, learning rate, number of training times, theoretical input power set Temperature Collection Power Collection Get the predicted temperature set And output to the weighted summation module.
[0079] The weighted sum module calculates the predicted temperature set based on the input Theoretical input power set and the desired temperature set Get the final input power set And output to the controlled object.
[0080] The desired temperature set It is determined according to the material and temperature control requirements of the controlled object. It can be a fixed value or a changing value with a gradient increase or decrease.
[0081] like Figure 3 As shown, the present invention provides a temperature control method based on an improved LSTM neural network multi-layer fuzzy nested PID temperature control system, comprising the following steps:
[0082] Step 1: The detection module collects the temperature and power data of the controlled object and outputs the temperature signal T and the power signal P to the processor. The processor records the temperature signal T and the power signal P in time series to obtain a time series prediction data set, which includes the temperature set of the controlled object. and power collection The data in the time series prediction data set are fitted by the least square method to obtain the transfer function G(s) of the controlled object. According to the formula: Error set Get the error set According to the formula: Error change rate set (Single temperature control cycle time t).
[0083] In the set temperature control single cycle time t, the input module collects the temperature and power of the controlled object in real time, records them according to the time series, and obtains a time series prediction data set, which includes the temperature set of the controlled object and power collection The least squares method is used to fit the transfer function G(s) of the controlled object to the time series prediction data set. The least squares method is set in the system identification component of Matlab to obtain the transfer function G(s) of the controlled object in this cycle. This process is carried out continuously according to the artificially set cycle, and the transfer function G(s) is iteratively updated. The transfer function G(s) formula is as follows:
[0084]
[0085] The detection module includes several temperature sensors and power sensors. The processor groups the signals of each sensor separately, and the number of groups corresponds to the number of sensors.
[0086] The processor transfers the function G(s), the error set Temperature Collection and power collection Output to the PID controller, the temperature set Power Collection Error Set Error change rate set Output to the multi-layer fuzzy nesting module, the temperature set and power collection Output to the improved LSTM neural network module.
[0087] like Figure 2 As shown, it is a flowchart of iterative optimization of the SC-RIME optimization algorithm module in the present invention.
[0088] Step 2: The SC-RIME optimization algorithm module and the PID controller cooperate to iterate and optimize to obtain the ratio K p优 , integral K i优 , differential K d优 .
[0089] Step 2.1: The SINE chaotic mapping initialization strategy initializes the positions of all particles in the population to obtain an initialized particle population.
[0090] In the population initialization process, the SINE chaotic map initialization strategy is used to replace the random number initialization strategy in the traditional frost ice optimization algorithm RIME. This improvement can significantly improve the randomness and diversity of the algorithm's initial population, thereby effectively improving the performance of the algorithm. The formula for the SINE chaotic map initialization strategy is as follows:
[0091]
[0092] Where n = (1, 2, 3...), is the chaotic system parameter, and its value is 0.99. n is a random number between [0,1], and formula (2) can be used to obtain x 1 、x 2 , ..., x D In this step, the value of the optimal particle position dimension D is 3, that is, according to formula (2), x 1 、x 2 、x 3 Let x 1 、x 2 、x 3 Multiply each by the maximum value Z of the particle search range to get the particle position (x 1 *Z,x 2 *Z,x 3 *Z), let K p =x 1 *Z, K i =x 2 *Z and K d =x 3 *Z gets several groups of proportions K p , integral K i and differential K dAt the same time, the positions of all particles in the population are initialized to obtain the initialized particle population.
[0093] Step 2.2, using the soft frost growth strategy and the hard frost puncture strategy to further search the initialized particle population to obtain a new particle population;
[0094] Step 2.3: Subsequently, the new particle population in step S2.2 is further cross-optimized using the horizontal crossover mechanism to obtain a particle position, i.e., a set of proportions K p , integral K i and differential K d And output to the PID controller.
[0095] The horizontal crossover mechanism formula is as follows:
[0096]
[0097] Where c 1 、c 2 is a random number between (-1,1); r 1 、r 2 is a random number between (0,1); D is the dimension of the variable; M is the size of the population particles; X(i,d) and X(j,d) are the d-th dimensions of X(i) and X(j); NX(i,d) and NX(j,d) are the d-th dimensional offspring generated by X(i,d) and X(j,d) through the horizontal crossover mechanism. The horizontal crossover operation mainly generates offspring with high probability within the hypercube, and at the same time jumps out of the space contained by the parent generation with low probability to explore the outer edge of the cube, effectively reducing the blind spots in the search process.
[0098] Step 2.4: PID controller is based on the conventional PID controller principle and the input total temperature control time A, and a set of ratios K obtained in step 2.3. p , integral K i , differential K d , transfer function G(s), error set Temperature Collection and power collection Get the simulation power set of the time B when the set temperature T' is reached for the first time and the total temperature control time A Simulation Error Collection And output to the SC-RIME optimization algorithm module;
[0099] Step 2.5, then update the optimal particle position according to the greedy selection mechanism, and input the simulation power set of the first time B to reach the set temperature T' and the total temperature control time A Simulation Error Collection Substitute the following fitness evaluation condition function formula to obtain the fitness value f1 ,
[0100]
[0101] Fitness value f 1 It is necessary to satisfy the fitness evaluation condition function formula at the same time, according to the fitness value f 1 The size is used to make a decision, that is, the fitness value f of the new position 1 Less than the fitness value f of the old position 1 , then update the particle position, otherwise keep the old particle position.
[0102] Step 2.6: Finally, determine whether the optimization conditions are met. If the number of iterations is greater than the set maximum number of iterations or the fitness value f in step 2.5, 1 If the value is less than the set fitness standard value, the optimal particle position after optimization is output, and the ratio K is obtained. p优 , integral K i优 and differential K d优 ; Otherwise, execute step 2.2, and repeat steps 2.2-2.6 until the optimization condition is met. p优 , integral K i优 and differential K d优 Output to the PID controller.
[0103] Heating power limit P in step 2 max The preferred value is 200W; the particle optimization range is preferably [0,500]; the particle population size M is preferably 50; and the maximum number of iterations is preferably 50.
[0104] Step 3: The multi-layer fuzzy nesting module is based on the input error set Error change rate set Temperature Collection and power collection Get the ratio K p优 , integral K i优 and differential K d优 The corresponding amplification factor is output to the PID controller.
[0105] Step 3.1: First, the first decision-making unit module determines the temperature set according to the input and power collection Judge the working condition of the controlled object. If it is a SISO system or a MISO system, directly input the error set and the error rate of change set Input the second decision unit module and execute step 3.3; otherwise, determine the error set If the temperature is within the range of [-λ℃, λ℃], proceed to step 3.2.
[0106] Step 3.2: If the error set In the range of [-λ℃, λ℃], the error set and the error rate of change set Output to the first layer fuzzy module for the first fine fuzzy control; if it is not within the range, the error set and the error rate of change set Output to the first layer fuzzy module for the first rough fuzzy control, the error set and the error rate of change set The sets obtained through the first fine fuzzy control or the first rough fuzzy control are all total error sets. Total error rate of change set And output to the second decision unit module to execute step 3.3.
[0107] Step 3.3: The second decision unit module determines the input error set Or the total error set Is it within the range of [-λ℃, λ℃]? If so, the input set is output to the second-layer fuzzy module for the second fine fuzzy control; otherwise, the input set is output to the second-layer fuzzy module for the second coarse fuzzy control. The values obtained by the second fine fuzzy control or the second coarse fuzzy control are both K p优 , K i优 and K d优 The corresponding magnification factor.
[0108] The present invention can adaptively switch between the following four major working conditions, namely, SISO system, SIMO system, MISO system and MIMO system. In the SISO system, the power of a single heating point corresponds to the temperature of a single temperature measuring point; in the SIMO system, the power of a single heating point corresponds to the temperature of multiple temperature measuring points; in the MISO system, the power of multiple heating points corresponds to the temperature of a single temperature measuring point; in the MIMO system, the power of multiple heating points corresponds to the temperature of multiple temperature measuring points, and can be divided into input-output symmetrical working conditions, that is, the number of inputs and outputs is equal, including SISO system and MIMO system; and input-output asymmetrical working conditions, that is, the number of inputs and outputs is not equal, including SIMO system and MISO system.
[0109] The first decision unit module can be based on the input temperature set and power collection The data is composed of automatic switching to the corresponding working mode. and power collection When there is only a single set of data in the system, it is determined to be a SISO system; when the power set There is only a single set of data in the temperature set When there are multiple sets of data in the system, it is determined to be a SIMO system; when the power set There are multiple sets of data, temperature sets When there is only a single set of data in the system, it is determined to be a MISO system; when the power set There are multiple sets of data, temperature sets When there are multiple sets of data in the system, it is determined to be a MIMO system. At the same time, the first decision unit module can also Whether it is within the set value range determines the subsequent fuzzy control method. If the error is within the range of [-λ℃, λ℃], the first fine fuzzy control in the first layer fuzzy module is performed. If the error is not within the range, the first coarse fuzzy control in the first layer fuzzy module is performed. The total error set is obtained. and the total error rate of change The second decision unit module determines the total error set Is it within the range of [-λ℃, λ℃]? If so, the second fine fuzzy control in the second layer fuzzy module is performed; otherwise, the second coarse fuzzy control in the second layer fuzzy module is performed to obtain the ratio K p优 , integral K i优 and differential K d优 The corresponding amplification factor is output to the PID controller. Where λ is the temperature value set for fine fuzzy control, preferably 5.
[0110] Step 4: PID controller is based on the input ratio K p优 , integral K i优 and differential K d优 And the ratio K p优 , integral K i优 and differential K d优 The corresponding amplification factor obtains the theoretical input power set And output to the improved LSTM neural network module and weighted summation module.
[0111] The PID controller sets the ratio K p优 , integral K i优 and differential K d优 Corresponding magnification factor and ratio K p优 , integral K i优 and differential K d优 Multiply them together to get the ratio K p总 , integral K i总 and differential K d总 , and obtain the theoretical input power set
[0112] Step 5: The SC-RIME optimization algorithm module and the improved LSTM neural network module cooperate to obtain the optimal number of hidden layer units, learning rate, and training times, and output them to the improved LSTM neural network module.
[0113] This step is performed simultaneously with step 2.
[0114] Step 5.1: SINE chaotic mapping initialization strategy initializes the positions of all particles in the population to obtain an initialized particle population.
[0115] The formula for the SINE chaos map initialization strategy is as follows:
[0116]
[0117] Where n = (1, 2, 3...), is the chaotic system parameter, and its value is 0.99. n is a random number between [0,1], and from formula (5) we can get x 1 、x 2 , ..., x D In this step, the value of the optimal particle position dimension D is 3, that is, according to formula (5), x 1 、x 2 、x 3 Let x 1 、x 2 、x 3 Multiply each by the maximum value Z of the particle search range to get the particle position (x 1 *Z,x 2 *Z,x 3 *Z), let the number of hidden layer units = (48*x 1 +2) is rounded, that is, the number of hidden layer units is an integer between [2,50]; let the learning rate = x 2 / 100; set the number of training times
[0118] =(50*x 3 +10), that is, the number of training times is an integer between [10,60], and several groups of hidden layer unit numbers, learning rates and training times are obtained. At the same time, the positions of all particles in the population are initialized to obtain the initialized particle population, and the particle positions correspond to each group of hidden layer unit numbers, learning rates and training times.
[0119] Step 5.2, using the soft frost growth strategy and the hard frost puncture strategy to further search the initialized particle population to obtain a new particle population;
[0120] Step 5.3: Subsequently, the new particle population in step 5.2 is further cross-optimized using the horizontal crossover mechanism to obtain a particle position, i.e., a set of hidden layer unit numbers, learning rates, and training times, and output them to the improved LSTM neural network module;
[0121] Step 5.4: Improved LSTM neural network module based on the input temperature set Power Collection The number of hidden layer units, learning rate, and training times obtained in step 5.3 are used to obtain the number of validation set samples m and the iterative prediction temperature T i , measured temperature T i 'And output to the SC-RIME optimization algorithm module;
[0122] The improved LSTM neural network module uses the Adam algorithm to train the internal network parameters, and the activation function in the hidden layer is set to the Relu function. The storage set samples in the improved LSTM neural network module are divided into a training set and a validation set at a ratio of 8:2. The improved LSTM neural network module is based on the input temperature set. Power Collection The improved LSTM neural network is trained with a set of hidden layer unit numbers, learning rates, training times and training sets obtained in step 5.4. The trained improved LSTM neural network obtains the validation set sample number m, iterative prediction temperature T according to the validation set. i , measured temperature T i 'And output to the SC-RIME optimization algorithm module.
[0123] Step 5.5, then update the optimal particle position according to the greedy selection mechanism, input the number of validation set samples m, iterative prediction temperature T i , measured temperature T i 'Substitute the following fitness evaluation condition function formula to obtain the fitness value f 2 ,
[0124]
[0125] According to the fitness value f 2 The size is used to make a decision, that is, the fitness value f of the new position 2 Less than the fitness value f of the old position 2 , then update the particle position, otherwise keep the old particle position.
[0126] Step 5.6: Finally, determine whether the optimization conditions are met. If the number of iterations is greater than the set maximum number of iterations or the fitness value f in step 5.5, 2 If the value is less than the set fitness standard value, the optimal particle position after optimization is output, and the optimal number of hidden layer units, learning rate and number of training times are obtained; otherwise, step 5.2 is executed, and steps 5.2-5.6 are looped until the optimization conditions are met. The optimal number of hidden layer units, learning rate and number of training times are output to the improved LSTM neural network module.
[0127] In step 5, the particle population size M is preferably 10; the maximum number of iterations is preferably 50.
[0128] Step 6: Build an improved LSTM neural network dataset and set the input temperature and power collection To process, first, the temperature set and power collection According to 15 consecutive temperature Plus 15 consecutive power Plus a theoretical input power P i , there are 31 data in total that constitute a set of data.
[0129] Step 7: Improve the LSTM neural network based on the optimal number of hidden layer units, learning rate, number of training times, and temperature set of the input Power Collection And the data set to get the predicted temperature set And output to the weighted summation module.
[0130] Step 8: Weighted sum module calculates the expected temperature set for the input Theoretical input power set and predicted temperature ensemble Perform weighted summation to obtain the final input power set
[0131] The expected temperature set Theoretical input power set and predicted temperature ensemble Get the final input power set The formula is as follows:
[0132]
[0133] In formula (7), α = (βη) / η max , where β is the gain factor, η max is the maximum number of cycles, i.e. η max =A / t, η is the current cycle number. The present invention comprehensively considers the multi-layer fuzzy nesting algorithm and the improved LSTM neural network to obtain a reliable and accurate final input power
[0134] Step 9: Set the final input power obtained in step 8 Enter the temperature control system, cycle steps 1 to 9, and perform cyclic temperature control until the set total temperature control time A is reached and then the temperature control program is terminated.
[0135] In addition to the above embodiments, the present application may also have other implementation modes. Any technical solutions formed by equivalent replacement or equivalent transformation shall fall within the protection scope required by the present application.
Claims
1. A multi-layer fuzzy nested PID temperature control system based on an improved LSTM neural network, comprising a detection module connected to a controlled object, wherein the detection module is used to output a temperature signal T and a power signal P collected in real time to a processor, Features: Also includes Input module, used to input the desired temperature set Optimize particle position dimension D and heating power limit P max , particle optimization range [0, Z], particle population size M, set maximum number of iterations, set fitness standard value, total temperature control time A and single temperature control cycle time t are output to the corresponding module, Z is the maximum value of the particle optimization range, which is an integer greater than 0; the input module outputs the expected temperature set Output to the processor and weighted summation module, the heating power limit P max Output to the SC-RIME optimization algorithm module, output the total temperature control time A and the single temperature control cycle time t to the processor, PID controller and weighted summation module, and output the optimal particle position dimension D, particle optimization range [0-Z], particle population size M, the set maximum number of iterations, and the set fitness standard value to the SC-RIME optimization algorithm module; The processor is used to input the total temperature control time A, the single temperature control cycle time t, the temperature signal T, the power signal P and the expected temperature set Perform data calculation and fitting to obtain the temperature set of the controlled object Power Collection Transfer function G(s), desired temperature set and temperature collection The error set Error change rate set And transfer function G(s), error set Temperature Collection and power collection Output to the PID controller, the temperature set Power Collection Error Set Error change rate set Output to the multi-layer fuzzy nesting module, the temperature set and power collection Output to the improved LSTM neural network module; SC-RIME optimization algorithm module, used to compare K p , integral K i , differential K d Iterate and optimize the number of hidden layer units, learning rate, and training times to obtain the ratio K p优 , integral K i优 , differential K d优 And the optimal number of hidden layer units, learning rate, and number of training times, the ratio K p优 , integral K i优 , differential K d优 Output to the PID controller, and output the optimal number of hidden layer units, learning rate, and training times to the improved LSTM neural network module; Multi-layer fuzzy nesting module is used to set the temperature according to the input Power Collection Error Set and error rate of change set Get the ratio K p优 , integral K i优 and differential K d优 The corresponding amplification factor is output to the PID controller; PID controller, used to adjust the input proportional K p优 , integral K i优 and differential K d优 And the ratio K p优 , integral K i优 and differential K d优 The corresponding amplification factor obtains the theoretical input power set And output to the improved LSTM neural network and weighted summation module; Improved LSTM neural network module, used to store improved LSTM neural network, based on the optimal number of hidden layer units, learning rate, number of training times, theoretical input power set Temperature Collection Power Collection Get the predicted temperature set And output to the weighted summation module; Weighted sum module, used to predict the temperature set according to the input Theoretical input power set and the desired temperature set Get the final input power set And output to the controlled object.
2. According to the multi-layer fuzzy nested PID temperature control system based on the improved LSTM neural network described in claim 1, Features: The SC-RIME optimization algorithm module is used to store the SC-RIME optimization algorithm. The SC-RIME optimization algorithm is obtained by replacing the random number initialization strategy in the frost and ice optimization algorithm RIME with the SINE chaotic map initialization strategy. At the same time, the horizontal cross mechanism in the vertical and horizontal cross optimization algorithm CSO is added before the greedy selection mechanism of the frost and ice optimization algorithm RIME.
3. According to the multi-layer fuzzy nested PID temperature control system based on the improved LSTM neural network described in claim 2, Features: The multi-layer fuzzy nesting module includes a first decision unit module, a first layer fuzzy module, a second decision unit module and a second layer fuzzy module; the first decision unit module is used to determine which of the four systems of SISO, SIMO, MISO and MIMO the controlled object belongs to.
4. A temperature control method based on the improved LSTM neural network multi-layer fuzzy nested PID temperature control system according to claim 3, It is characterized in that The following steps are included: S1. The processor records the input temperature signal T and power signal P in time series to obtain a time series prediction data set, which includes the temperature set of the controlled object. and power collection The data in the time series prediction data set are fitted by the least square method to obtain the transfer function G(s) of the controlled object. and the desired temperature set Get the error set and error rate of change set S2, SC-RIME optimization algorithm module and PID controller cooperate to iterate and optimize to obtain the proportion K p优 , integral K i优 , differential K d优 ; S3, multi-layer fuzzy nesting module according to the input error set Error change rate set Temperature Collection and power collection Get the ratio K p优 , integral K i优 and differential K d优 The corresponding amplification factor is output to the PID controller; S4, PID controller according to the input ratio K p优 , integral K i优 and differential K d优 And the ratio K p优 , integral K i优 and differential K d优 The corresponding amplification factor obtains the theoretical input power set And output to the improved LSTM neural network module and weighted summation module; S5, the SC-RIME optimization algorithm module and the improved LSTM neural network module cooperate to obtain the optimal number of hidden layer units, learning rate, and training times and output them to the improved LSTM neural network module; S6. Build an improved LSTM neural network dataset and set the input temperature and power collection Processing, the temperature collection and power collection According to 15 consecutive temperature Plus 15 consecutive power Plus a theoretical input power P i , there are 31 data in total that constitute a set of data; S7, Improved LSTM neural network based on the optimal number of hidden layer units, learning rate, number of training times, and temperature set of input Power Collection And the data set to get the predicted temperature set And output to the weighted summation module; S8, weighted sum module for the expected temperature set of the input Theoretical input power set and predicted temperature ensemble Perform weighted summation to obtain the final input power set S9, the final input power set obtained in step S8 Input the controlled object, loop steps S1 to S9, perform cyclic temperature control, and terminate the temperature control program after reaching the total temperature control time A.
5. According to claim 4, a multi-layer fuzzy nested PID temperature control method based on an improved LSTM neural network, Features: The steps of iterative optimization in step S2 in which the SC-RIME optimization algorithm module and the PID controller cooperate to perform the following steps: S2.1, SINE chaotic mapping initialization strategy initializes the positions of all particles in the population to obtain the initialized particle population. The formula for the SINE chaos map initialization strategy is as follows: Where n = (1, 2, 3...), is the chaotic system parameter, with a value of 0.99, x n is a random number between [0,1], and from formula (1) we can get x 1 、x 2 , ..., x D , where the value of the optimal particle position dimension D is 3, that is, according to formula (1), we get x 1 、x 2 、x 3 Let x 1 、x 2 、x 3 Multiply each by the maximum value Z of the particle search range to get the particle position (x 1 *Z,x 2 *Z,x 3 *Z), let K p =x 1 *Z, K i =x 2 *Z and K d =x 3 *Z gets several groups of proportions K p , integral K i and differential K d ; S2.2, using the soft frost growth strategy and the hard frost puncture strategy to further search the initialized particle population to obtain a new particle population; S2.3, the horizontal crossover mechanism further crosses the new particle population in step S2.2 to obtain a particle position, that is, a set of proportions K p , integral K i and differential K d And output to the PID controller; S2.4, PID controller according to the input temperature control total time A, step S2.3 obtained a set of proportions K p , integral K i , differential K d , transfer function G(s), error set Temperature Collection and power collection Get the simulation power set of the time B when the set temperature T' is reached for the first time and the total temperature control time A Simulation Error Collection And output to the SC-RIME optimization algorithm module; S2.5, the greedy selection mechanism updates the optimal particle position, and the input of the first time to reach the set temperature T'B, the total temperature control time A simulation power set Simulation Error Collection Substitute the following fitness evaluation condition function formula to obtain the fitness value f 1 , IF er≥0 Else Fitness value f 1 It is necessary to satisfy the fitness evaluation condition function formula at the same time, according to the fitness value f 1 The size is used to make a decision, that is, the fitness value f of the new position 1 Less than the fitness value f of the old position 1 , then update the particle position, otherwise keep the old particle position; S2.6, determine whether the optimization conditions are met. If the number of iterations is greater than the set maximum number of iterations or the fitness value in step 2.5 is less than the set fitness standard value, then output the optimal particle position after optimization and obtain the ratio K p优 , integral K i优 and differential K d优 ; Otherwise, execute step S2.2 and loop through steps S2.2-S2.
6. Until the optimization condition is met, the ratio K p优 , integral K i优 and differential K d优 Output to the PID controller.
6. According to claim 4, a multi-layer fuzzy nested PID temperature control method based on an improved LSTM neural network, Features: The step S3 obtains the ratio K p优 , integral K i优 and differential K d优 The corresponding steps of the amplification factor are as follows, S3.1, the first decision unit module is based on the input temperature set and power collection Judge the working condition of the controlled object. If it is a SISO system or a MISO system, directly input the error set and error rate of change set Input the second decision unit module and execute step 3.3; otherwise, determine the error set Is it within the range of [-λ℃, λ℃], and proceed to step 3.2; S3.
2. If the error set In the range of [-λ℃, λ℃], the error set and error rate of change set Output to the first layer fuzzy module for the first fine fuzzy control; if it is not within the range, the first coarse fuzzy control is performed; Get the total error set Total error rate of change collection And output to the second decision unit module to execute step 3.3; S3.3, the second decision unit module determines the error set of the input Or the total error set Is it within the range of [-λ℃, λ℃]? If so, the input set is output to the second-layer fuzzy module for the second fine fuzzy control; otherwise, the second coarse fuzzy control is performed; the ratio K is obtained. p优 , integral K i优 and differential K d优 The corresponding magnification factor.
7. According to claim 4, a multi-layer fuzzy nested PID temperature control method based on an improved LSTM neural network, Features: The steps of obtaining the optimal number of hidden layer units, learning rate, and training times in step S5 are as follows: S5.1, SINE chaotic mapping initialization strategy initializes the positions of all particles in the population, and obtains the initialized particle population, several groups x 1 、x 2 、x 3 and the corresponding particle position (x 1 *Z,x 2 *Z,x 3 *Z), let the number of hidden layer units = (48*x 1 +2) Round up; let learning rate = x 2 / 100; let the number of training times = (50*x 3 +10) to get several sets of hidden layer unit numbers, learning rates and training times; S5.2, using the soft frost growth strategy and the hard frost puncture strategy to further search the initialized particle population to obtain a new particle population; S5.3, the horizontal crossover mechanism further crosses and optimizes the new particle population in step S5.2 to obtain a particle position, i.e., a set of hidden layer unit numbers, learning rates, and training times, and outputs them to the improved LSTM neural network module; S5.4, the improved LSTM neural network module is based on the input temperature set Power Collection Step S5.3 obtains a set of hidden layer unit numbers, learning rates, and training times, and obtains the number of validation set samples m and iterative prediction temperature T i , measured temperature T i 'And output to the SC-RIME optimization algorithm module; S5.5, Greedy selection mechanism updates the optimal particle position, inputs the number of validation set samples m, iterative prediction temperature T i , measured temperature T i 'Substitute the following fitness evaluation condition function formula to obtain the fitness value f 2 , According to the fitness value f 2 The size is used to make a decision, that is, the fitness value f of the new position 2 Less than the fitness value f of the old position 2 , then update the particle position, otherwise keep the old particle position; S5.6, determine whether the optimization conditions are met. If the number of iterations is greater than the set maximum number of iterations or the fitness value in step 5.5 is less than the set fitness standard value, then output the optimal particle position after optimization, and obtain the optimal number of hidden layer units, learning rate and number of training times; Otherwise, execute step S5.2, and loop steps S5.2-S5.6 until the optimization condition is met; output the optimal number of hidden layer units, learning rate and number of training times to the improved LSTM neural network module.
8. A temperature control method based on an improved LSTM neural network multi-layer fuzzy nested PID according to claim 5 or 7, Features: The horizontal crossover mechanism is carried out as follows, NX(i,d)=r 1 ×X(i,d)+(1-r 1 )×X(j,d)+c 1 ×(X(i,d)-X(j,d)) NX(j,d)=r 2 ×X(j,d)+(1-r 2 )×X(i,d)+c 2 ×(X(j,d)-X(i,d)) i,j∈N(1,M),d∈N(1,D) where c 1 、c 2 is a random number between (-1,1); r 1 、r 2 is a random number between (0,1); D is the dimension of the variable; M is the size of the population particles; X(i,d) and X(j,d) are the d-th dimensions of X(i) and X(j); NX(i,d) and NX(j,d) are the d-th offspring generated by X(i,d) and X(j,d) through the horizontal crossover mechanism.
9. According to claim 6, a multi-layer fuzzy nested PID temperature control method based on an improved LSTM neural network, Features: The value of λ is 5.
10. According to claim 4, a multi-layer fuzzy nested PID temperature control method based on an improved LSTM neural network, Features: In step S8, according to the formula Get the final input power set Where α = (βη) / η max , β is the gain factor, η max is the maximum number of cycles, and η is the current number of cycles.
Citation Information
Patent Citations
Device for floor heating floor constant temperature control and operation method thereof
CN117906198A
Water dispenser and outlet water temperature control method thereof
CN118216804A
Simple intelligent temperature control system
CN118244815A
Cited By
Temperature control intelligent optimization method and system for adhesive tape production
CN119396221A
Temperature control system and method based on bimodal and GRU neural network fusion
CN121326057A
Temperature control method and system based on flexible production line spraying and drying
CN121455262A
