A predictive control method, system and storage medium for a clean room ventilation system

Abstract

The present invention relates to the technical field of clean room ventilation systems, and provides a prediction control method, system and storage medium for clean room ventilation systems, the method comprising: obtaining M groups of controlled variables based on the distribution of M groups of controllable variables, forming M real data samples, and obtaining real data sets; constructing a synthetic data set based on a BIM simulation model; preprocessing the real data set and the synthetic data set to obtain a training set, a verification set and a test set; constructing and training a prediction model; optimizing and controlling the ventilation system; evaluating the performance of the ventilation system based on the energy saving rate; using a multi-task learning architecture and a hybrid expert model when constructing the prediction model, constructing an input layer based on a fully connected neural network, constructing a shared layer based on a multi-level extraction network, and outputting the predicted value of the controlled variable based on the output layer. The present invention can comprehensively predict the air volume, has a high pressure difference prediction accuracy, can simultaneously complete the pressure difference and air volume prediction of multi-zone clean rooms, has a fast response speed, and is highly efficient.

Description

Technical Field

[0001] The present invention relates to the technical field of clean room ventilation systems, and in particular to a prediction control method, system and computer-readable storage medium for clean room ventilation systems. Background Art

[0002] A clean room is a well-sealed space where parameters such as air cleanliness, temperature, humidity, pressure, and noise are controlled as needed. Clean room technology is crucial in industries such as pharmaceuticals, microelectronics, and bioengineering to ensure a sterile and dust-free environment during product manufacturing. However, the operating cost of a clean room is high, and the energy consumption of its ventilation system accounts for a large proportion of the total energy consumption.

[0003] With the increasing emphasis on improving energy efficiency, optimizing cleanroom ventilation systems has become a hot topic of research. Existing cleanroom ventilation systems usually use the traditional testing, adjusting and balancing (TAB) method for air volume distribution and pressure difference control. For example, in the initial commissioning stage of the cleanroom ventilation system, technicians spend a lot of time measuring air volume, adjusting valve openings, and calibrating fan frequency to achieve the air volume and pressure difference indicators required by the design. The method based on manual experience is time-consuming, inefficient, and difficult to respond quickly when faced with complex working conditions.

[0004] Although some studies have attempted to introduce machine learning technology to solve the above problems, they are often limited to air volume prediction for a single area and lack the prediction and processing of the mutual influence between multiple areas. They cannot simultaneously complete the pressure difference and air volume prediction of multi-area clean rooms. In addition, the air volume prediction is not comprehensive enough, the pressure difference prediction accuracy is not high, and the efficiency is low, which cannot meet the short-term and high-frequency simulation prediction needs. Summary of the invention

[0005] To this end, the technical problem to be solved by the present invention is to overcome the problems in the prior art that the clean room ventilation system cannot simultaneously complete the pressure difference and air volume prediction of multi-zone clean rooms, and the air volume prediction is not comprehensive enough, the pressure difference prediction accuracy is not high, and the efficiency is low. A predictive control method, system and computer-readable storage medium for a clean room ventilation system are provided, which have comprehensive air volume prediction, high pressure difference prediction accuracy, and can simultaneously complete the pressure difference and air volume prediction of multi-zone clean rooms, with fast response speed and high efficiency.

[0006] In a first aspect, in order to solve the above technical problems, the present invention provides a predictive control method for a clean room ventilation system, comprising:

[0007] Step S1: Set the ventilation system stabilization time and determine M groups of controllable variables;

[0008] Step S2: sending a group of controllable variables to obtain a group of controlled variables; after the ventilation system stabilization time meets the set requirements, collecting the actual values ​​of the group of controllable variables and the group of controlled variables to form a real data sample;

[0009] Step S3: Repeat step S2 until M groups of controlled variables are obtained from the M groups of controllable variables, forming M actual data samples; and obtain an actual data set based on the M actual data samples;

[0010] Step S4: constructing a synthetic data set of the controllable variables and the controlled variables based on the BIM simulation model;

[0011] Step S5: preprocessing the actual data set and the synthetic data set to obtain a training set, a validation set and a test set;

[0012] Step S6: constructing and training a prediction model based on the training set, the validation set and the test set;

[0013] Step S7: Optimizing and controlling the ventilation system based on the prediction model;

[0014] Step S8: Evaluate the performance of the ventilation system after optimization control based on the energy saving rate;

[0015] Among them, when constructing the prediction model, a multi-task learning architecture and a hybrid expert model are adopted, the input layer of the prediction model is constructed based on a fully connected neural network, the shared layer of the prediction model is constructed based on a multi-level extraction network, and the predicted value of the controlled variable is output based on the output layer of the prediction model.

[0016] In one embodiment of the present invention, in the multi-stage extraction network, each stage of the extraction network includes a supply air expert module, a return air expert module, an exhaust air expert module and a shared expert module. e The return air expert module consists of N e The exhaust air expert module consists of N return air experts. e The shared expert module consists of N exhaust experts. se shared experts; each level of the extraction network is internally provided with a supply air door network, a return air door network, an exhaust air door network and a shared door network, which correspond to the supply air expert module, the return air expert module, the exhaust air expert module and the shared expert module respectively; the output of each level of the extraction network is used as the input of the next level of extraction network; wherein the first level of the extraction network directly receives the feature vector from the input layer input, and the output of the last level of the extraction network is used as the output of the multi-level extraction network.

[0017] In one embodiment of the present invention, the corresponding relationship between the air supply door network and the air supply expert module is:

[0018] ;

[0019] The corresponding relationship between the return air door network and the return air expert module is:

[0020] ;

[0021] The corresponding relationship between the exhaust door network and the exhaust expert module is:

[0022] ;

[0023] The corresponding relationship between the shared gate network and the shared expert module is:

[0024] ;

[0025] in, is the activation function for the multi-classification problem, represents the air supply door network in the j-th extraction network, represents the input of the air supply expert module in the j-th extraction network, represents the internal parameter matrix of the air supply gate network in the j-th extraction network, Represents the N in the j-th level extraction network e The air supply experts and N se The initial output after integration by the shared experts; represents the return air door network in the j-th extraction network, represents the input of the return air expert module in the j-th extraction network, represents the internal parameter matrix of the return air door network in the j-th extraction network, Represents the N in the j-th level extraction network e The return air experts and N se The initial output after integration by the shared experts; represents the exhaust gate network in the j-th extraction network, represents the input of the exhaust expert module in the j-th level extraction network, represents the internal parameter matrix of the exhaust gate network in the j-th extraction network, Represents the N in the j-th level extraction network e The ventilation experts and N se The initial output after integration by the shared experts; represents the shared gate network in the j-th level extraction network, represents the input of the shared expert module in the j-th level extraction network, represents the internal parameter matrix of the shared gate network in the j-th level extraction network, Represents the N in the j-th level extraction network e The air supply experts, N e The return air experts, N e The ventilation experts and N se The preliminary output after integration by the shared experts.

[0026] In one embodiment of the present invention, the method for optimizing and controlling the ventilation system based on the prediction model is: using the IoT cloud platform to input the pressure difference design values ​​and air volume design values ​​of multiple zones as the main control targets into the optimizer of the prediction model, and iteratively optimizing the controllable variables based on the optimizer to meet the constraints; in each iteration, the optimizer calls the prediction model, predicts the controlled variables based on the controllable variables, and then calculates the deviation between the predicted operating conditions and the designed operating conditions based on the objective function; on the premise of satisfying the constraints, the optimizer iteratively obtains the controllable variable values ​​corresponding to the minimization of the objective function; and the optimal controllable parameters are issued based on the controllable variable values.

[0027] In one embodiment of the present invention, the method for minimizing the objective function is:

[0028] ;

[0029] The constraints are:

[0030] ;

[0031] in, represents the minimization function, Z represents the objective function, Indicates the number of fans. Indicates the number of pressure differences, Represents the predicted pressure difference array, represents the design pressure difference array, represents the fan frequency, st represents the constraint condition, Indicates the pressure difference deviation in any area, Indicates the preset pressure difference deviation threshold, Indicates the designed air supply volume. Indicates the predicted air supply volume, Indicates the designed exhaust volume, represents the predicted exhaust volume, Indicates the opening of the terminal air valve.

[0032] In one embodiment of the present invention, the method for preprocessing the actual data set and the synthetic data set to obtain the training set, the validation set and the test set is: first exclude unusable samples, then convert the controllable variable inputs and the controlled variable outputs in the actual data set and the synthetic data set into standard normal distributions based on the z-score normalization method, and then randomly divide them into training sets, validation sets and test sets according to the proportions set independently.

[0033] In one embodiment of the present invention, when constructing and training a prediction model based on the training set, the validation set and the test set, a mean square error function is used as a loss function to calculate the difference between the predicted value and the actual value of the prediction model; the prediction model is pre-trained based on the synthetic data set, and then the prediction model is fine-tuned based on the training set of the actual data set; the prediction model is verified in real time based on the validation set of the actual data set; after all training is completed, the prediction model is tested based on the test set of the actual data set.

[0034] In one embodiment of the present invention, the M groups of controllable variables are determined based on a Latin hypercube sampling strategy; and the M groups of controllable variables are issued through an on-site IoT cloud platform and a PLC.

[0035] In a second aspect, in order to solve the above technical problems, the present invention also provides a prediction control system for a clean room ventilation system, comprising:

[0036] The data acquisition module sequentially executes: step S1: setting the ventilation system stabilization time and determining M groups of controllable variables; step S2: issuing a group of the controllable variables to obtain a group of controlled variables; after the ventilation system stabilization time meets the set requirements, the actual values ​​of the group of controllable variables and the group of controlled variables are collected to form a real-collected data sample; step S3: repeating step S2 until M groups of controlled variables are obtained from the M groups of controllable variables to form M real-collected data samples; obtaining a real-collected data set based on the M real-collected data samples; step S4: constructing a synthetic data set of the controllable variables and the controlled variables based on the BIM simulation model;

[0037] A model building and training module preprocesses the actual data set and the synthetic data set to obtain a training set, a validation set, and a test set; and builds and trains a prediction model based on the training set, the validation set, and the test set;

[0038] An optimization control module performs optimization control on the ventilation system based on the prediction model;

[0039] A performance evaluation module for evaluating the performance of the ventilation system after optimized control based on energy saving rate;

[0040] Among them, when constructing the prediction model, the model construction and training module adopts a multi-task learning architecture and a hybrid expert model, constructs the input layer of the prediction model based on a fully connected neural network, constructs the shared layer of the prediction model based on a multi-level extraction network, and outputs the predicted value of the controlled variable based on the output layer of the prediction model.

[0041] In a third aspect, in order to solve the above-mentioned technical problem, the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the predictive control method of a clean room ventilation system are implemented.

[0042] The above technical solution of the present invention has the following beneficial effects compared with the prior art:

[0043] The present invention provides a predictive control method, system and computer-readable storage medium for a clean room ventilation system. A multi-task learning architecture and a hybrid expert model are used to construct a predictive model, which can fully utilize the coupling relationship between different types of air volumes, and the air volume prediction is comprehensive, the pressure difference prediction accuracy is high, and the pressure difference and air volume prediction of multi-zone clean rooms can be completed simultaneously; the ventilation system is optimized and controlled based on the predictive model, with fast response speed and high efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments of the present invention in conjunction with the accompanying drawings.

[0045] Figure 1 A step diagram of a predictive control method for a clean room ventilation system in a preferred embodiment of the present invention;

[0046] Figure 2 A prediction model architecture diagram of a prediction control method for a clean room ventilation system in a preferred embodiment of the present invention;

[0047] Figure 3 It is a module diagram of a predictive control system of a clean room ventilation system in an embodiment of the present invention.

[0048] Description of the Figures in the Specification:

[0049] Input layer; Shared Layer; Tower Layer; Output layer;

[0050] First level Extraction Network First level extraction network; Extraction Network j-1th j-1th level extraction network; Extraction Network jth j-1th level extraction network; Extraction Network j+1th j+1th level extraction network; Final level Extraction Network last level extraction network;

[0051] Experts on SA air supply door experts; Experts on RA air return door experts; Experts on EA air exhaust door experts; Shared Experts shared door experts; Tower Room a area a in multiple areas; Tower SA air supply tasks; Tower OA fresh air tasks;

[0052] Fan frequency; Terminal air valve opening; g gate network function symbol; H tensor; Predicted pressure difference array; Predict air supply volume; Predict exhaust air volume;

[0053] : The abbreviation of , which means the air supply door network in the first-stage extraction network;

[0054] : The abbreviation of , which means the return air door network in the first-stage extraction network;

[0055] : The abbreviation of , which means the exhaust door network in the first-stage extraction network;

[0056] : , which represents the shared gate network in the first-level extraction network;

[0057] : The first, second, ..., Nth in the j-th level extraction network e Air supply door expert;

[0058] : The first, second, ..., Nth in the j-th level extraction network e A return air door expert;

[0059] : The first, second, ..., Nth in the j-th level extraction network e An exhaust door expert;

[0060] : The first, second, ..., Nth in the j-th level extraction network se A shared door expert;

[0061] : The abbreviation of N in the j-th level extraction network e Air supply experts and N se The initial output after integration by the shared experts;

[0062] : Input of the air supply expert module in the j-th level extraction network;

[0063] : The abbreviation of , which represents the air supply door network in the j-th stage extraction network. DETAILED DESCRIPTION

[0064] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0065] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments.

[0066] Existing clean room ventilation systems usually use the traditional TAB method for air volume distribution and pressure difference control, but it is time-consuming, inefficient, and difficult to respond quickly when faced with complex working conditions. Some studies have tried to introduce machine learning technology to solve the above problems, but they are often limited to air volume prediction for a single area, lacking the prediction processing of the mutual influence between multiple areas, and cannot simultaneously complete the pressure difference and air volume prediction of multi-area clean rooms. In addition, the air volume prediction is not comprehensive enough, the pressure difference prediction accuracy is not high, and the efficiency is low, which cannot meet the short-term and high-frequency simulation prediction needs.

[0067] To this end, embodiments of the present application provide a predictive control method, system, and computer-readable storage medium for a clean room ventilation system. Embodiment 1

[0068] This embodiment provides a predictive control method for a clean room ventilation system. Figure 1 As shown, including:

[0069] Step S1: Set the ventilation system stabilization time and determine M groups of controllable variables;

[0070] Step S2: sending a group of controllable variables to obtain a group of controlled variables; after the ventilation system stabilization time meets the set requirements, collecting the actual values ​​of the group of controllable variables and the group of controlled variables to form a real data sample;

[0071] Step S3: Repeat step S2 until M groups of controlled variables are obtained from the M groups of controllable variables, forming M actual data samples; and obtain an actual data set based on the M actual data samples;

[0072] Step S4: constructing a synthetic data set of the controllable variables and the controlled variables based on the BIM simulation model;

[0073] Step S5: preprocessing the actual data set and the synthetic data set to obtain a training set, a validation set and a test set;

[0074] Step S6: constructing and training a prediction model based on the training set, the validation set and the test set;

[0075] Step S7: Optimizing and controlling the ventilation system based on the prediction model;

[0076] Step S8: Evaluate the performance of the ventilation system after optimization control based on the energy saving rate;

[0077] Among them, when constructing the prediction model, a multi-task learning architecture and a hybrid expert model are adopted, the input layer of the prediction model is constructed based on a fully connected neural network, the shared layer of the prediction model is constructed based on a multi-level extraction network, and the predicted value of the controlled variable is output based on the output layer of the prediction model.

[0078] The present embodiment provides a predictive control method for a clean room ventilation system. (1) A multi-task learning architecture and a hybrid expert model are used to construct a predictive model, which can fully utilize the coupling relationship between different types of air volumes. The air volume prediction is comprehensive, the pressure difference prediction accuracy is high, and the pressure difference and air volume prediction of multiple clean rooms can be completed simultaneously. (2) The ventilation system is optimized and controlled based on the predictive model, with fast response speed and high efficiency, thereby reducing energy waste caused by excessive or insufficient ventilation.

[0079] Next, a prediction control method for a clean room ventilation system provided in this embodiment is described in detail:

[0080] 1. Working steps:

[0081] Step S1:

[0082] Specifically, the ventilation system stabilization time is set and M groups of controllable variables are determined;

[0083] Optionally, the M groups of controllable variables are determined based on a Latin hypercube sampling strategy; and the M groups of controllable variables are issued through an on-site IoT cloud platform and PLC.

[0084] Optionally, the controllable variables include controllable parameters of equipment operation: frequency of each fan and opening of each air valve.

[0085] Optionally, the controlled variables include system controlled parameters: pressure difference between zones and system air volume.

[0086] Optionally, the waiting time is set to 2 minutes.

[0087] Step S2:

[0088] Specifically, a group of controllable variables is sent down to obtain a group of controlled variables; after the ventilation system stabilization time meets the set requirements, the actual values ​​of the group of controllable variables and the group of controlled variables are collected to form a real data sample;

[0089] Step S3:

[0090] Specifically, step S2 is repeatedly performed until M groups of controlled variables are obtained from the M groups of controllable variables to form M actual data samples; and an actual data set is obtained based on the M actual data samples;

[0091] Step S4:

[0092] Specifically, constructing a synthetic data set of the controllable variables and the controlled variables based on the BIM simulation model;

[0093] Step S5:

[0094] Specifically, the actual data set D1 and the synthetic data set D2 are preprocessed to obtain a training set, a validation set, and a test set.

[0095] Optionally, unusable samples are excluded first, and then the controllable variable inputs and controlled variable outputs in the actual data set and the synthetic data set are respectively converted into standard normal distributions based on the z-score normalization method, and then randomly divided into the training set, the validation set and the test set according to a self-set ratio.

[0096] Optionally, the unavailable samples include incomplete samples.

[0097] Optionally, the training set, the validation set and the test set are randomly divided in a ratio of 8:1:1.

[0098] Step S6:

[0099] Specifically, a prediction model is constructed and trained based on the training set, the validation set and the test set.

[0100] For details, please refer to Figure 2 As shown, when constructing the prediction model, a multi-task learning architecture and a hybrid expert model are used, the input layer Input of the prediction model is constructed based on a fully connected neural network, the shared layer Shared Layer of the prediction model is constructed based on a multi-level extraction network, and the predicted value of the controlled variable is output based on the output layer Output of the prediction model.

[0101] Optionally, the prediction model uses Leaky ReLU as a basic activation function.

[0102] Optionally, in the multi-stage extraction network, each stage of the extraction network includes an air supply expert module E SA , Return Air Expert Module E RA 、Exhaust Expert Module E EA and shared expert module E S , the air supply expert module E SA By N e The return air expert module E RA By N e The exhaust air expert module E EA By N e The shared expert module E S By N se Shared experts, please refer to Figure 2 shown.

[0103] Specifically, each level of the extraction network is internally provided with an air supply door network, a return air door network, an exhaust door network and a shared door network, which correspond to the air supply expert module E SA , the return air expert module E RA , the exhaust expert module E EA and the shared expert module E S .

[0104] Specifically, the output of each level of extraction network serves as the input of the next level of extraction network; wherein the first level of extraction network directly receives the feature vector input from the input layer, and the output of the last level of extraction network serves as the output of the multi-level extraction network.

[0105] Specifically, the air supply expert module E SA , the return air expert module E RA and the exhaust expert module E EA There are three types of specific expert modules, each of which is only responsible for extracting the corresponding airflow characteristics in the clean room ventilation system; the shared expert module E SThe coupling interaction information between different airflow characteristics is extracted.

[0106] Specifically, the number of stages of the multi-stage extraction network is determined during the training process of the prediction model.

[0107] Specifically, the corresponding relationship between the air supply door network and the air supply expert module is:

[0108] ;

[0109] in, is the activation function for multi-class classification problems, represents the air supply door network in the j-th extraction network, represents the input of the air supply expert module in the j-th extraction network, represents the internal parameter matrix of the air supply gate network in the j-th extraction network, Represents the N in the j-th level extraction network e The air supply experts and N se The preliminary output after integration by the shared experts.

[0110] Specifically, the corresponding relationship between the return air door network and the return air expert module is:

[0111] ;

[0112] in, represents the return air door network in the j-th extraction network, represents the input of the return air expert module in the j-th extraction network, represents the internal parameter matrix of the return air door network in the j-th extraction network, Represents the N in the j-th level extraction network e The return air experts and N se The preliminary output after integration by the shared experts.

[0113] Specifically, the corresponding relationship between the exhaust door network and the exhaust expert module is:

[0114] ;

[0115] in, represents the exhaust gate network in the j-th extraction network, represents the input of the exhaust expert module in the j-th level extraction network, represents the internal parameter matrix of the exhaust gate network in the j-th extraction network, Represents the N in the j-th level extraction network e The ventilation experts and Nse The preliminary output after integration by the shared experts.

[0116] Specifically, the corresponding relationship between the shared gate network and the shared expert module is:

[0117] ;

[0118] in, represents the shared gate network in the j-th level extraction network, represents the input of the shared expert module in the j-th level extraction network, represents the internal parameter matrix of the shared gate network in the j-th level extraction network, Represents the N in the j-th level extraction network e The air supply experts, N e The return air experts, N e The ventilation experts and N se The preliminary output after integration by the shared experts.

[0119] Specifically, the , , and stated Can be expressed as:

[0120]

[0121] Where T stands for transpose.

[0122] Optionally, the characteristic vector includes the frequency of each fan and the opening of each terminal air valve.

[0123] For details, please refer to Figure 2 As shown, after the multi-stage extraction network gradually separates the airflow characteristics, the data output by the last stage extraction network is spliced ​​and combined into a tensor H and passed into the task tower layer.

[0124] Optionally, in the task tower layer Tower Layer, each task has an independent branch tower, such as the independent branch tower Tower Room a of area a among multiple areas; the task tower layer Tower Layer adopts a multi-layer fully connected network to re-learn the deep features extracted by the shared layer Shared Layer, and finally outputs the pressure difference of each area, the system supply air volume, the system return air volume, the system exhaust air volume, and the system fresh air volume; wherein, the fresh air volume does not have a corresponding specific expert, and the fresh air volume is obtained by comprehensive calculation of other types of air volumes.

[0125] Optionally, a mean squared error function (MSE) is used as the loss function L2 to calculate the difference between the predicted value and the actual value of the prediction model; the prediction model is pre-trained based on the synthetic data set D2, and then the prediction model is fine-tuned based on the training set of the actual data set; the prediction model is verified in real time based on the verification set of the actual data set; after all training is completed, the prediction model is tested based on the test set of the actual data set.

[0126] Step S7:

[0127] Specifically, the ventilation system is optimized and controlled based on the prediction model.

[0128] Optionally, the prediction model optimizes the control of the ventilation system based on a genetic algorithm with enhanced elite retention.

[0129] Specifically, the IoT cloud platform is used to input the pressure difference design values ​​and air volume design values ​​of multiple areas as the main control objectives into the optimizer of the prediction model, and the controllable variables are iteratively optimized based on the optimizer to meet the constraints; in each iteration, the optimizer calls the prediction model, predicts the controlled variables based on the controllable variables, and then calculates the deviation between the predicted operating conditions and the designed operating conditions based on the objective function Z; on the premise of satisfying the constraints, the optimizer obtains the controllable variable values ​​corresponding to the minimization of the objective function through iteration; and the optimal controllable parameters are issued based on the controllable variable values.

[0130] Optionally, the method for minimizing the objective function is:

[0131] ;

[0132] in, represents the minimization function, Z represents the objective function, Indicates the number of fans. Indicates the number of pressure differences, Represents the predicted pressure difference array, represents the design pressure difference array, Indicates the fan frequency.

[0133] Optionally, the constraint condition is:

[0134] ;

[0135] Among them, st represents the constraints, Indicates the pressure difference deviation in any area, Indicates the preset pressure difference deviation threshold, Indicates the designed air supply volume. Indicates the predicted air supply volume, Indicates the designed exhaust volume, represents the predicted exhaust volume, Indicates the opening of the terminal air valve.

[0136] Optionally, an AdamW optimizer is used as the optimizer to train the prediction model; the AdamW optimizer combines the advantages of an adaptive moment estimation (Adaptive Moment Estimation, referred to as Adam) optimizer and the regularization effect of the loss function L2, which helps the prediction model to converge quickly.

[0137] Step S8:

[0138] Specifically, the performance of the ventilation system after optimized control is evaluated based on the energy saving rate.

[0139] Specifically, the fan power P and the fan frequency is proportional to the cube of .

[0140] Furthermore, by reducing the average operating frequency of all fans, energy-saving operation can be achieved on the basis of meeting the pressure gradient and ventilation frequency requirements. Relative to reference frequency The percentage of reduction is used as the evaluation control energy saving rate Direct indicators:

[0141] ;

[0142] 2. Examples:

[0143] For example, a multi-zone clean room has 6 independent zones, namely a, b, c, d, e, and f, and the key components of its ventilation system include 3 fans, 7 supply air valves, 4 return air valves, and 4 exhaust valves.

[0144] Specifically, the fan frequency The value range is 0~50Hz, and its minimum controllable degree is 0.1Hz; the terminal air valve opening The value range is 0~90°, and its minimum controllable degree is 5°.

[0145] Real-time data set D1: Step S1: Set the ventilation system stabilization time to 2 minutes, and use the Latin hypercube sampling strategy to formulate 436 groups of controllable variable (3 fan frequencies + 15 terminal air valve openings) samples; Step S2: Send a group of the controllable variables to obtain a group of controlled variables, wait for 2 minutes until the ventilation system is stable, and then collect the actual values ​​of the group of controllable variables and the group of controlled variables to form a real-time data sample; Step S3: Repeat step S2 until 436 groups of controllable variables are obtained from the 436 groups of controllable variables to form 436 real-time data samples, and obtain the real-time data set D1 based on the 436 real-time data samples;

[0146] Synthetic data set D2: The BIM simulation model is established based on field data, with a minimum controllable degree of 1 Hz for the fan frequency and a minimum controllable degree of 1° for the terminal air valve opening; 400,000 sets of controllable variables are formulated using the Latin hypercube sampling strategy, and their corresponding controlled variables are calculated through simulation.

[0147] Furthermore, the sample data in D1 and D2 are preprocessed: after excluding unusable samples such as incomplete data, the controllable variable inputs and controlled variable outputs in D1 and D2 are converted into standard normal distributions using the z-score normalization method.

[0148] Specifically, D1 is randomly divided into training set A1 containing 403 training samples, validation set A2 containing 12 validation samples, and test set A3 containing 21 test samples; D2 is randomly divided into training set B1 containing 319,890 training samples, validation set B2 containing 39,986 validation samples, and test set B3 containing 39,987 test samples.

[0149] Furthermore, a prediction model is constructed and trained based on the training set A1, the validation set A2, the test set A3, the training set B1, the validation set B2 and the test set B3. Specifically, the input feature vector includes the three fan frequencies and the 15 terminal air valve openings, and the prediction model is constructed using the multi-task learning architecture and the hybrid expert; the prediction model uses Leaky ReLU as the basic activation function; the air supply expert module E SA , the return air expert module E RA and the exhaust expert module E EA A set of fully connected networks of equal width are used to construct the multi-stage extraction network, and the total number of stages is J; the depth and width of the independent task branch tower are D T and W T; Adopt the AdamW optimizer as the optimizer of the prediction model, and adopt MSE as the loss function L2; Use the synthetic data set D2 to pre-train the prediction model, and then fine-tune the prediction model on the training set A1; e , D e , W e , J, N se , D se , W se , D T , W T , training learning rate I r These parameters, such as BS and batch size, are automatically tuned.

[0150] Specifically, during the fine-tuning training process, the performance of the prediction model is verified in real time using the data of the verification set A2, so as to ensure the correct training of the prediction model.

[0151] Furthermore, after all the training is completed, the performance of the trained prediction model is tested using the data of the test set A3, which can ensure that the trained prediction model has good prediction accuracy and generalization ability.

[0152] Furthermore, the prediction model optimizes the control of the ventilation system based on the genetic algorithm with enhanced elite retention.

[0153] Furthermore, the performance of the ventilation system after optimized control is evaluated based on the energy saving rate ξ.

[0154] Specifically, the predictive control method for a clean room ventilation system provided in this embodiment only requires 3% of the manual debugging time (about 17.21 seconds) to find and deploy a more optimal operating point, which greatly improves the debugging efficiency; by accurately controlling the system supply air volume, system return air volume and system exhaust air volume, and reasonably arranging the ventilation times and pressure difference gradients in each area, unnecessary ventilation redundancy can be reduced, and the overall energy consumption and clean room operating costs can be reduced. Embodiment 2

[0155] This embodiment provides a prediction control system for a clean room ventilation system. Figure 3 As shown, including:

[0156] The data acquisition module sequentially executes: step S1: setting the ventilation system stabilization time and determining M groups of controllable variables; step S2: issuing a group of the controllable variables to obtain a group of controlled variables; after the ventilation system stabilization time meets the set requirements, the actual values ​​of the group of controllable variables and the group of controlled variables are collected to form a real-collected data sample; step S3: repeating step S2 until M groups of controlled variables are obtained from the M groups of controllable variables to form M real-collected data samples; obtaining a real-collected data set based on the M real-collected data samples; step S4: constructing a synthetic data set of the controllable variables and the controlled variables based on the BIM simulation model;

[0157] A model building and training module preprocesses the actual data set and the synthetic data set to obtain a training set, a validation set, and a test set; and builds and trains a prediction model based on the training set, the validation set, and the test set;

[0158] An optimization control module performs optimization control on the ventilation system based on the prediction model;

[0159] A performance evaluation module for evaluating the performance of the ventilation system after optimized control based on energy saving rate;

[0160] Among them, when constructing the prediction model, the model construction and training module adopts a multi-task learning architecture and a hybrid expert model, constructs the input layer of the prediction model based on a fully connected neural network, constructs the shared layer of the prediction model based on a multi-level extraction network, and outputs the predicted value of the controlled variable based on the output layer of the prediction model.

[0161] For an introduction to a predictive control system for a clean room ventilation system provided in this embodiment, please refer to the first embodiment, and this embodiment will not be described in detail here.

[0162] This embodiment provides a predictive control system for a clean room ventilation system, which has the same beneficial effects as the predictive control method for a clean room ventilation system described above. Embodiment 3

[0163] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of a predictive control method for a clean room ventilation system as described in the first embodiment are implemented.

[0164] For an introduction to a computer-readable storage medium provided in this embodiment, please refer to Embodiment 1, and this embodiment will not be described in detail here.

[0165] The computer-readable storage medium provided in this embodiment has the same beneficial effects as the predictive control method for a clean room ventilation system.

[0166] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0167] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0168] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0169] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0170] Obviously, the above embodiments are merely examples for clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived from these are still within the protection scope of the invention.

Claims

1. A predictive control method for a clean room ventilation system, characterized in that: include: Step S1: Set the ventilation system stabilization time and determine M groups of controllable variables; Step S2: sending a group of controllable variables to obtain a group of controlled variables; after the ventilation system stabilization time meets the set requirements, collecting the actual values ​​of the group of controllable variables and the group of controlled variables to form a real data sample; Step S3: Repeat step S2 until M groups of controlled variables are obtained from the M groups of controllable variables, forming M actual data samples; and obtain an actual data set based on the M actual data samples; Step S4: constructing a synthetic data set of the controllable variables and the controlled variables based on the BIM simulation model; Step S5: preprocessing the actual data set and the synthetic data set to obtain a training set, a validation set and a test set; Step S6: constructing and training a prediction model based on the training set, the validation set and the test set; Step S7: Optimizing and controlling the ventilation system based on the prediction model; Step S8: Evaluate the performance of the ventilation system after optimization control based on the energy saving rate; Wherein, when constructing the prediction model, a multi-task learning architecture and a hybrid expert model are adopted, the input layer of the prediction model is constructed based on a fully connected neural network, the shared layer of the prediction model is constructed based on a multi-level extraction network, and the predicted value of the controlled variable is output based on the output layer of the prediction model; The hybrid expert model includes a supply air expert module, a return air expert module, an exhaust air expert module and a shared expert module, as well as a supply air door network, a return air door network, an exhaust air door network and a shared door network; In the multi-stage extraction network, each stage of the extraction network includes the air supply expert module, the return air expert module, the exhaust air expert module and the shared expert module. The air supply expert module is composed of N e The return air expert module consists of N e The exhaust air expert module consists of N return air experts. e The shared expert module consists of N exhaust experts. se shared experts; each level of extraction network is internally provided with the supply air door network, the return air door network, the exhaust air door network and the shared door network, which correspond to the supply air expert module, the return air expert module, the exhaust air expert module and the shared expert module respectively; the output of each level of extraction network is used as the input of the next level of extraction network; wherein the first level extraction network directly receives the feature vector from the input layer input, and the output of the last level extraction network is used as the output of the multi-level extraction network.

2. The predictive control method for a clean room ventilation system according to claim 1, characterized in that: The corresponding relationship between the air supply door network and the air supply expert module is: ； The corresponding relationship between the return air door network and the return air expert module is: ； The corresponding relationship between the exhaust door network and the exhaust expert module is: ； The corresponding relationship between the shared gate network and the shared expert module is: ； in, is the activation function for the multi-classification problem, represents the air supply door network in the j-th extraction network, represents the input of the air supply expert module in the j-th extraction network, represents the internal parameter matrix of the air supply gate network in the j-th extraction network, Represents the N in the j-th level extraction network e The air supply experts and N se The initial output after integration by the shared experts; represents the return air door network in the j-th extraction network, represents the input of the return air expert module in the j-th extraction network, represents the internal parameter matrix of the return air door network in the j-th extraction network, Represents the N in the j-th level extraction network e The return air experts and N se The initial output after integration by the shared experts; represents the exhaust gate network in the j-th extraction network, represents the input of the exhaust expert module in the j-th level extraction network, represents the internal parameter matrix of the exhaust gate network in the j-th extraction network, Represents the N in the j-th level extraction network e The ventilation experts and N se The initial output after integration by the shared experts; represents the shared gate network in the j-th level extraction network, represents the input of the shared expert module in the j-th level extraction network, represents the internal parameter matrix of the shared gate network in the j-th level extraction network, Represents the N in the j-th level extraction network e The air supply experts, N e The return air experts, N e The ventilation experts and N se The preliminary output after integration by the shared experts.

3. The predictive control method for a clean room ventilation system according to claim 1, characterized in that: The method for optimizing the control of the ventilation system based on the prediction model is: Using the IoT cloud platform, the pressure difference design values ​​and air volume design values ​​of multiple regions are input into the optimizer of the prediction model as the main control targets, and the controllable variables are iteratively optimized based on the optimizer to meet the constraint conditions; In each iteration, the optimizer calls the prediction model, predicts the controlled variables based on the controllable variables, and then calculates the deviation between the predicted operating conditions and the designed operating conditions based on the objective function; on the premise of satisfying the constraints, the optimizer obtains the corresponding controllable variable values ​​when the objective function is minimized through iteration; and the optimal controllable parameters are issued based on the controllable variable values.

4. The predictive control method for a clean room ventilation system according to claim 3, characterized in that: The method for minimizing the objective function is: ； The constraints are: ； in, represents the minimization function, Z represents the objective function, Indicates the number of fans. Indicates the number of pressure differences, Represents the predicted pressure difference array, represents the design pressure difference array, represents the fan frequency, st represents the constraint condition, Indicates the pressure difference deviation in any area, Indicates the preset pressure difference deviation threshold, Indicates the designed air supply volume. Indicates the predicted air supply volume, Indicates the designed exhaust volume, represents the predicted exhaust volume, Indicates the opening of the terminal air valve.

5. The predictive control method for a clean room ventilation system according to claim 1, characterized in that: The method of preprocessing the actual data set and the synthetic data set to obtain the training set, the validation set and the test set is: First, unusable samples are excluded, and then the controllable variable inputs and controlled variable outputs in the actual data set and the synthetic data set are converted into standard normal distributions based on the z-score normalization method, and then randomly divided into training sets, validation sets, and test sets according to the proportions set by themselves.

6. The predictive control method for a clean room ventilation system according to claim 1, characterized in that: When constructing and training a prediction model based on the training set, the validation set and the test set, a mean square error function is used as a loss function to calculate the difference between the predicted value and the actual value of the prediction model; the prediction model is pre-trained based on the synthetic data set, and then the prediction model is fine-tuned based on the training set of the actual data set; the prediction model is verified in real time based on the validation set of the actual data set; After all the training is completed, the prediction model is tested based on the test set of the actual data set.

7. The predictive control method for a clean room ventilation system according to claim 1, characterized in that: The M groups of controllable variables are determined based on a Latin hypercube sampling strategy; and the M groups of controllable variables are issued through an on-site IoT cloud platform and a PLC.

8. A predictive control system for a clean room ventilation system, characterized in that: A predictive control method for a clean room ventilation system according to any one of claims 1 to 7, comprising: The data acquisition module executes in sequence: step S1: setting the ventilation system stabilization time and determining M groups of controllable variables; step S2: issuing a group of the controllable variables to obtain a group of controlled variables; after the ventilation system stabilization time meets the set requirements, the actual values ​​of the group of controllable variables and the group of controlled variables are collected to form a real-collected data sample; step S3: repeating step S2 until M groups of controlled variables are obtained from the M groups of controllable variables to form M real-collected data samples; obtaining a real-collected data set based on the M real-collected data samples; step S4: constructing a synthetic data set of the controllable variables and the controlled variables based on the BIM simulation model; the model construction and training module preprocesses the real-collected data set and the synthetic data set to obtain a training set, a validation set and a test set; constructing and training a prediction model based on the training set, the validation set and the test set; An optimization control module performs optimization control on the ventilation system based on the prediction model; A performance evaluation module for evaluating the performance of the ventilation system after optimized control based on energy saving rate; Among them, when constructing the prediction model, the model construction and training module adopts a multi-task learning architecture and a hybrid expert model, constructs the input layer of the prediction model based on a fully connected neural network, constructs the shared layer of the prediction model based on a multi-level extraction network, and outputs the predicted value of the controlled variable based on the output layer of the prediction model.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the predictive control method for a clean room ventilation system according to any one of claims 1 to 7 are implemented.

Images