A coal preparation calorific value prediction method and system based on end-edge-cloud collaboration

Abstract

The present invention proposes a coal preparation calorific value prediction method and system based on edge-cloud collaboration. The method includes: selecting relevant variables affecting calorific value as input and the clean coal calorific value as output, constructing a calorific value digital twin model including a linear model and nonlinear terms; performing parameter identification to obtain identification errors; constructing a nonlinear dynamic system v(k) based on the nonlinear terms and identification errors; and using a long short-term memory (LSTM) multi-layer neural network to construct an offline deep learning model for v(k), an edge-side online deep learning model, and a cloud-based deep learning correction model. The correction is corrected according to all variables updated in real time by a cloud database using a preset self-correction mechanism. The system constructs a calorific value endpoint prediction model based on the nonlinear dynamic system, inputs a density setpoint and related variable data, and outputs a predicted calorific value data. The beneficial effect is to achieve real-time calorific value prediction, assist in density decision-making and calorific value control in the coal preparation process, and improve the quality of clean coal.

Description

Technical Field

[0001] The present application relates to the technical field of coal preparation process parameter prediction, and in particular to a coal preparation calorific value prediction method and system based on end-edge-cloud collaboration, a storage device, and a computer-readable storage medium. Background Art

[0002] Clean coal, a vital national strategic resource, is widely used in power generation, building materials, industrial boilers, domestic heating, metallurgy, and other applications. Coal preparation plants are the primary means of producing clean coal. Calorific value is a key indicator of clean coal product quality, representing the amount of heat released when a unit mass (or volume) of fuel is completely burned. Calorific value data is crucial for density decision-making and calorific value control in coal preparation plants. The accuracy of these decisions and controls significantly impacts clean coal product quality and has a significant impact on the company's economic performance.

[0003] In the process of implementing this application, the applicant discovered that the related technology has at least the following problems:

[0004] Coal preparation plants currently use manual testing to obtain the calorific value of clean coal, with calorific value data obtained every two hours. During this period, calorific value information cannot be obtained, resulting in serious lag. Summary of the Invention

[0005] (1) Technical issues to be solved

[0006] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a coal preparation calorific value prediction method and system based on end-edge-cloud collaboration, which solves the technical problem that the calorific value of clean coal is obtained by manual testing, which has serious lag and is not conducive to density decision-making and calorific value control in the coal preparation process.

[0007] (2) Technical solution

[0008] In order to achieve the above objectives, the main technical solutions adopted by the present invention include:

[0009] In the first aspect, the embodiment of the present invention provides a coal preparation calorific value prediction method based on end-edge-cloud collaboration, including: selecting relevant variables that affect calorific value changes from a process perspective as model input, using clean coal calorific value as model output, and establishing a calorific value digital twin model y(k). The calorific value digital twin model y(k) includes a linear model and nonlinear terms; for linear models Perform parameter identification to obtain identification error, and construct a nonlinear dynamic system v(k) based on the nonlinear term and the identification error; use a long short-term memory multilayer neural network LSTM to construct an offline deep learning model of the nonlinear dynamic system v(k) According to the offline deep learning model Use long short-term memory multi-layer neural network LSTM to build an edge online deep learning model and cloud-based deep learning calibration models The relevant variable data of the coal preparation production process collected on the terminal side is transmitted to the cloud database; the cloud deep learning correction model is adjusted based on all the relevant variable data currently updated in real time in the cloud database Correction is performed; the corrected cloud-based deep learning correction model is used According to the preset self-correction mechanism, the side online deep learning model Correction is performed to complete the correction of the nonlinear dynamic system v(k) and obtain the estimated value of the nonlinear dynamic system Based on linear model and the estimated value of the nonlinear dynamic system Constructing an endpoint prediction model The density setting value and related variable data are input into the endpoint prediction model, and the endpoint prediction model outputs the calorific value data prediction value to achieve the calorific value prediction of the coal preparation process.

[0010] Embodiments of the present invention propose a coal preparation calorific value prediction method and system based on device-edge-cloud collaboration. Based on the coal preparation production process and a data-driven strategy, this method uses relevant variables that influence calorific value changes from a process perspective as model inputs and the clean coal calorific value as the model output. A calorific value digital twin model is constructed, and a linear model and a nonlinear dynamic system are determined based on the calorific value digital twin model. Next, an offline deep learning model of the nonlinear dynamic system is constructed, and an online deep learning model on the edge and a deep learning correction model on the cloud are constructed based on the offline deep learning model. Furthermore, data on relevant coal preparation variables is collected on the edge and transmitted to a cloud database. Parameters of the cloud deep learning model are updated by obtaining specified relevant variable data from the cloud database. The cloud deep learning correction model is used to calibrate the online deep learning model on the edge according to a preset self-calibration mechanism, completing the calibration of the nonlinear dynamic system and obtaining an estimated value for the nonlinear dynamic system. The sum of the estimated values ​​of the linear model and the nonlinear dynamic system is then used as the endpoint prediction model for the coal preparation calorific value production indicator. The sum of the system output of the estimated value of the nonlinear dynamic system and the model output of the linear model is used as the calorific value data prediction value, completing the calorific value prediction of the coal preparation process. The present invention combines process analysis with industrial big data, system identification with deep learning, offline training with online correction, establishes an endpoint prediction model for coal preparation calorific value production indicators, and proposes a coal preparation calorific value prediction method based on end-edge cloud collaboration to correct the nonlinear part in the endpoint prediction model of coal preparation calorific value production indicators, realizes modeling of industrial process operation indicators whose dynamic characteristics change with the production process, and real-time prediction of calorific value, which is conducive to assisting density decision-making and calorific value control in the coal preparation process, thereby improving the quality of clean coal products.

[0011] Optionally, the formula of the calorific value digital twin model is as follows:

[0012]

[0013] Among them, y(k) is the calorific value of clean coal, x u , u=[1,9] is the relevant variable, x1 is the set value of the density of the heavy medium; x2 is the density of the first heavy medium barrel; x3 is the density of the second heavy medium barrel; x4 is the liquid level of the first heavy medium barrel; x5 is the liquid level of the first heavy medium barrel; x6 is the pressure of the first cyclone; x7 is the pressure of the second cyclone; x8 is the opening of the first valve; x9 is the opening of the second valve, n v ,v=[1,9] is the variable lag time, is the parameter identification equation, is a nonlinear term, and a, b, c, d, e, f, g, h, i, and l are identification parameters.

[0014] The above formula is used to establish a calorific value digital twin model that includes a linear parameter identification part and a nonlinear part based on LSTM learning. Then, by performing parameter identification on the linear parameter identification part and correcting the nonlinear part, an endpoint prediction model that can predict the calorific value of coal preparation in real time can be obtained, thereby realizing the modeling of industrial process operation indicators whose dynamic characteristics change with the production process.

[0015] Optionally, for linear models Perform parameter identification to obtain identification error, and construct a nonlinear dynamic system v(k) based on the nonlinear term and the identification error, including: determining the linear model according to the calorific value digital twin model y(k) Will Expressed as Y(k)=X(k-1)θ, the parameter θ is identified based on the objective function J1 of the least squares algorithm to obtain the estimated value of the parameter θ Where, Y(k) is the input vector, Y(k) = [y(k)y(k+1)y(k+2)…y(k+N)] T , X(k-1) is the input matrix,

[0016]

[0017] Will As the identification error, according to the identification error With nonlinear terms The nonlinear dynamic system v(k) is constructed as follows:

[0018] The linear model is identified using the least squares algorithm to obtain estimated values ​​of the identification parameters. Since the output of the linear model is a calorific value that does not take into account changes in the dynamic characteristics of the calorific value caused by changes in the sorting and transportation processes, changes in coal particle length and impurity composition, there is an error between this calorific value and the actual calorific value. The error caused by the parameter estimation is used as the identification error, and a nonlinear dynamic system is constructed based on the identification error and the nonlinear term. At this time, the sum of the nonlinear dynamic system and the linear model is the calorific value digital twin model.

[0019] Optionally, a long short-term memory multilayer neural network LSTM is used to construct an offline deep learning model of the nonlinear dynamic system v(k) The method comprises: obtaining relevant quantity data and calorific value test data as sample data; determining the number of hidden layer nodes h, the number of network layers L, and the number of network neurons n of the long short-term memory multilayer neural network LSTM based on the sample data, and obtaining the offline deep learning model.

[0020] By obtaining relevant variable data in the coal preparation process and calorific value test data, sample data is formed, and then the network architecture of the offline deep learning model is determined based on the sample data, namely the number of hidden layer nodes h, the number of network layers L, and the number of network neurons n.

[0021] Optionally, according to the offline deep learning model Use long short-term memory multi-layer neural network LSTM to build an edge online deep learning model and cloud-based deep learning calibration models Including: Based on the offline deep learning model of long short-term memory multi-layer neural network LSTM The side-online deep learning model is constructed using the number of hidden layer nodes h, the number of network layers L, and the number of network neurons n. and cloud-based deep learning calibration models

[0022] The side-online deep learning model is constructed using the same number of hidden layer nodes h, network layers L, and network neurons n as the offline deep learning model. and cloud-based deep learning calibration models To facilitate online deep learning models based on the side and cloud-based deep learning calibration models Correction of nonlinear dynamic systems.

[0023] Optionally, the cloud-based deep learning model is calibrated based on all relevant variable data currently updated in real time in the cloud database. Perform calibration, including: calibrating the cloud-based deep learning model based on all relevant variable data currently updated in real time in the cloud database All layer weights and bias parameters are corrected.

[0024] All relevant variable data currently updated in real time in the cloud database constitute N S (k) groups of real-time updated production process data ψ(k), ψ(k-1),…, ψ(k-n+1)(k=1,...,N S ), real-time correction of all layer weights and biases of the cloud-based deep learning correction model. As the coal preparation process proceeds, the number of data N S (k) continues to increase. The cloud-based deep learning calibration model continuously performs calibration training on all relevant process and test data in the continuously updated cloud database, and promptly updates all layer weights and bias parameters of the cloud-based deep learning calibration model.

[0025] Optionally, use the calibrated cloud-based deep learning calibration model According to the preset self-correction mechanism, the side online deep learning model Correction is performed to complete the correction of the nonlinear dynamic system v(k) and obtain the estimated value of the nonlinear dynamic system Includes: Calculate and calibrate the cloud-based deep learning calibration model Model accuracy evaluation index RMSE1(k) and side-by-side online deep learning model The model accuracy evaluation index RMSE2(k); RMSE1(k) and RMSE2(k) are compared with the preset self-correction mechanism to generate comparison results; when the comparison result indicates the correction model, the deep learning correction model after correction is fixed The number of hidden layer nodes h, the number of network layers L, and the number of network neurons n are calculated based on the calibrated cloud-based deep learning model. Replace the side-by-side online deep learning model with all layer weights and bias parameters All layer weights and bias parameters to complete the side-by-side online deep learning model The correction of the nonlinear dynamic system v(k) is completed, and the estimated value of the nonlinear dynamic system is obtained.

[0026] Based on the experimental results, the upper bound of the error of the model output of the nonlinear dynamic system (RMSE = δ1) is set. Based on the upper bound of the error, the preset self-correction mechanism is obtained. Then, the model accuracy evaluation index of the deep learning correction model is calculated according to the formula:

[0027]

[0028] Where k is the sampling period (k=1,2,...), Calibrate model outputs for deep learning.

[0029] The model accuracy evaluation index of the cloud-based deep learning model is calculated using the same method, denoted as RMSE2(k). The model accuracy evaluation indexes of the side-online deep learning model and the cloud-based deep learning model in the calorific value endpoint prediction model are tested separately. The model accuracy evaluation indexes of the side-online deep learning model and the cloud-based deep learning correction model are compared with the preset self-correction mechanism to generate a comparison result. When the model accuracy evaluation index of the side-online deep learning model in the calorific value endpoint prediction model (RMSE1(k) ≥ 20) and the model accuracy evaluation index of the cloud-based deep learning correction model (RMSE2(k) < 20) are both above, a comparison result indicating model correction is generated. All layer weights and bias parameters of the cloud-based deep learning model are then downloaded and replaced with those of the side-online deep learning model to complete the correction of the side-online deep learning model.

[0030] Optionally, based on a linear model and the estimated value of the nonlinear dynamic system Constructing an endpoint prediction model The method includes: using an unknown nonlinear function f(·) to represent a nonlinear dynamic system v(k): v(k) = f(ψ(k-1), ψ(k-2), ..., ψ(kn)), where f(·) is a nonlinear function of unknown variation, n is the order of the nonlinear term, and ψ(k-1) is an input data vector. in is obtained by performing zero-order hold interpolation on the discrete thermal value data. is obtained by linear interpolation of discrete heat value data; the heat value digital twin model y(k) is expressed as: v(k) to obtain the dynamic model of calorific value y(k+1) at time k+1: Establishing a prediction model for nonlinear dynamic systems Establishing a calorific value endpoint prediction model

[0031] A prediction model is established based on the nonlinear dynamic system, and then a calorific value endpoint prediction model is established based on the prediction model and the linear model, so as to make real-time predictions on calorific value data based on the calorific value endpoint prediction model.

[0032] In a second aspect, an embodiment of the present invention provides a coal preparation calorific value prediction system based on device-edge-cloud collaboration, the system comprising:

[0033] The first building block is used to select relevant variables that affect the calorific value change from a process perspective as model input, and use the clean coal calorific value as the model output to establish a calorific value digital twin model y(k). The calorific value digital twin model y(k) includes a linear model and nonlinear terms;

[0034] The second building block is used to Perform parameter identification to obtain an identification error, so as to construct a nonlinear dynamic system v(k) according to the nonlinear term and the identification error;

[0035] The third building block is used to construct an offline deep learning model of the nonlinear dynamic system v(k) using a long short-term memory multi-layer neural network LSTM

[0036] The fourth building block is used to build a Use long short-term memory multi-layer neural network LSTM to build an edge online deep learning model and cloud-based deep learning models

[0037] The data acquisition module is used to transmit the relevant variable data of the coal preparation production process collected on the terminal side to the cloud database;

[0038] The first calibration module is used to calibrate the cloud-based deep learning model based on all relevant variable data currently updated in real time in the cloud database. Correction;

[0039] The second correction module is used to use the corrected cloud-based deep learning model According to the preset self-correction mechanism, the side online deep learning model Correction is performed to complete the correction of the nonlinear dynamic system v(k) and obtain the estimated value of the nonlinear dynamic system

[0040] The fifth building block is used for linear model-based and the estimated value of the nonlinear dynamic system Constructing an endpoint prediction model

[0041] The calorific value prediction module is used to input the density setting value and related variable data into the endpoint prediction model, and the endpoint prediction model outputs the calorific value data prediction value to realize the calorific value prediction of the coal preparation process.

[0042] The coal preparation calorific value prediction system based on end-edge-cloud collaboration provided by the technical solution of the present invention is used to implement the steps of the coal preparation calorific value prediction method based on end-edge-cloud collaboration provided by the first aspect of the present invention. Therefore, the coal preparation calorific value prediction system based on end-edge-cloud collaboration has all the technical effects of the coal preparation calorific value prediction method based on end-edge-cloud collaboration, which will not be repeated here.

[0043] In a third aspect, an embodiment of the present invention provides a storage device having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps of any one of the methods described in the first aspect.

[0044] In a fourth aspect, an embodiment of the present invention provides a computer storage medium, comprising a storage device, a processor, and a computer program stored on the storage device and executable on the processor, wherein when the processor executes the program, the steps of any one of the methods described in the first aspect are implemented.

[0045] (3) Beneficial effects

[0046] The beneficial effects of the present invention are as follows: a coal preparation calorific value prediction method and system based on end-edge-cloud collaboration, based on the coal preparation production process and a data-driven strategy, uses relevant variables that affect calorific value changes from a process perspective as model inputs and the calorific value of clean coal as the model output. A calorific value digital twin model is constructed, and a linear model and a nonlinear dynamic system are determined based on the calorific value digital twin model. Next, an offline deep learning model of the nonlinear dynamic system is constructed, and an end-side online deep learning model and a cloud-based deep learning model are constructed based on the offline deep learning model. Furthermore, coal preparation-related variable data is collected on the end-side and transmitted to a cloud-based database. Parameters of the cloud-based deep learning model are updated by obtaining specified relevant variable data from the cloud-based database. The end-side online deep learning model is then corrected using the cloud-based deep learning correction model according to a preset self-correction mechanism to complete the correction of the nonlinear dynamic system and obtain an estimated value of the nonlinear dynamic system. The sum of the estimated values ​​of the linear model and the nonlinear dynamic system is then used as the endpoint prediction model for the coal preparation calorific value production indicator. The sum of the system output of the estimated value of the nonlinear dynamic system and the model output of the linear model is used as the calorific value data prediction value to complete the calorific value prediction of the coal preparation process. Compared with related technologies, the present invention combines process analysis with industrial big data, system identification with deep learning, offline training with online correction, establishes an endpoint prediction model for coal preparation calorific value production indicators, and proposes a coal preparation calorific value prediction method based on end-edge cloud collaboration to correct the nonlinear part in the endpoint prediction model of coal preparation calorific value production indicators, realize the modeling of industrial process operation indicators whose dynamic characteristics change with the production process, and the real-time prediction of calorific value, which is conducive to assisting the density decision and calorific value control of the coal preparation process, thereby improving the quality of clean coal products. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A schematic diagram of a process for predicting the calorific value of coal preparation based on device-edge-cloud collaboration provided by an embodiment of the present invention;

[0048] Figure 2 A schematic diagram showing the linear identification results of the coal preparation calorific value prediction method based on device-edge-cloud collaboration provided in an embodiment of the present application is shown;

[0049] Figure 3 A schematic diagram of parameter determination of a coal preparation calorific value prediction method based on device-edge-cloud collaboration provided by an embodiment of the present application is shown;

[0050] Figure 4 A schematic diagram of parameter determination of a coal preparation calorific value prediction method based on device-edge-cloud collaboration provided by an embodiment of the present application is shown;

[0051] Figure 5 A schematic diagram of parameter determination of a coal preparation calorific value prediction method based on device-edge-cloud collaboration provided by an embodiment of the present application is shown;

[0052] Figure 6 A schematic diagram of the network structure of a coal preparation calorific value prediction method based on device-edge-cloud collaboration provided in an embodiment of the present application is shown;

[0053] Figure 7 A schematic diagram showing test results of a coal preparation calorific value prediction method based on device-edge-cloud collaboration provided by an embodiment of the present application is shown;

[0054] Figure 8 A schematic diagram showing test results of a coal preparation calorific value prediction method based on device-edge-cloud collaboration provided by an embodiment of the present application is shown;

[0055] Figure 9 A schematic diagram showing test results of a coal preparation calorific value prediction method based on device-edge-cloud collaboration provided by an embodiment of the present application is shown;

[0056] Figure 10 A schematic diagram showing test results of a coal preparation calorific value prediction method based on device-edge-cloud collaboration provided by an embodiment of the present application is shown;

[0057] Figure 11 A schematic diagram showing test results of a coal preparation calorific value prediction method based on device-edge-cloud collaboration provided by an embodiment of the present application is shown;

[0058] Figure 12 A block diagram of a coal preparation calorific value prediction system based on end-edge-cloud collaboration provided in an embodiment of the present application is shown.

[0059] [Description of Reference Numerals]

[0060] 1200: Coal preparation calorific value prediction system based on device-edge-cloud collaboration;

[0061] 1201: first building block;

[0062] 1202: second building block;

[0063] 1203: the third building block;

[0064] 1204: fourth building block;

[0065] 1205: data acquisition module;

[0066] 1206: first correction module;

[0067] 1207: second correction module;

[0068] 1208: fifth building block;

[0069] 1209: Calorific value prediction module. DETAILED DESCRIPTION

[0070] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation methods in conjunction with the accompanying drawings.

[0071] Reference Figure 1 This embodiment provides a coal preparation calorific value prediction method based on device-edge-cloud collaboration, including:

[0072] S102, select the relevant variables that affect the calorific value change from the process perspective as the model input, use the clean coal calorific value as the model output, and establish the calorific value digital twin model y(k). The calorific value digital twin model y(k) includes the linear model and nonlinear terms;

[0073] S104, for linear models Perform parameter identification to obtain identification error, and then construct the nonlinear dynamic system v(k) based on the nonlinear term and identification error;

[0074] S106, using long short-term memory multi-layer neural network LSTM to build an offline deep learning model for nonlinear dynamic system v(k)

[0075] S108, based on offline deep learning model Use long short-term memory multi-layer neural network LSTM to build an edge online deep learning model and cloud-based deep learning calibration models

[0076] S110, transmitting the relevant variable data of the coal preparation production process collected on the terminal side to the cloud database;

[0077] S112, calibrating the cloud deep learning model based on all relevant variable data currently updated in real time in the cloud database Correction;

[0078] S114, using the calibrated cloud-based deep learning calibration model According to the preset self-correction mechanism, the side online deep learning model Correction is performed to complete the correction of the nonlinear dynamic system v(k) and obtain the estimated value of the nonlinear dynamic system

[0079] S116, based on linear model and the estimated value of the nonlinear dynamic system Constructing an endpoint prediction model

[0080] S118, inputting density setting value and related variable data into the endpoint prediction model, and the endpoint prediction model outputs the calorific value data prediction value to realize the calorific value prediction of the coal preparation process.

[0081] Embodiments of the present invention propose a coal preparation calorific value prediction method and system based on device-edge-cloud collaboration. Based on the coal preparation production process and a data-driven strategy, this method uses relevant variables that influence calorific value changes from a process perspective as model inputs and the calorific value of clean coal as the model output. A calorific value digital twin model is constructed, and a linear model and a nonlinear dynamic system are determined based on the calorific value digital twin model. Next, an offline deep learning model of the nonlinear dynamic system is constructed. Based on the offline deep learning model, an online deep learning model on the edge and a deep learning correction model on the cloud are constructed. Furthermore, data on relevant coal preparation variables is collected on the edge and transmitted to a cloud database. Specified relevant variable data is retrieved from the cloud database to update the parameters of the cloud deep learning model. The cloud deep learning correction model is used to calibrate the online deep learning model on the edge based on a preset self-calibration mechanism to calibrate the nonlinear dynamic system and obtain an estimated value for the nonlinear dynamic system. The sum of the estimated values ​​of the linear model and the nonlinear dynamic system is then used as the endpoint prediction model for the coal preparation calorific value production indicator. The sum of the estimated value of the nonlinear dynamic system and the model output of the linear model is used as the calorific value data prediction value to complete the calorific value prediction of the coal preparation process. The present invention combines process analysis with industrial big data, system identification with deep learning, offline training with online correction, establishes an endpoint prediction model for coal preparation calorific value production indicators, and proposes a coal preparation calorific value prediction method based on end-edge cloud collaboration to correct the nonlinear part in the endpoint prediction model of coal preparation calorific value production indicators, realizes modeling of industrial process operation indicators whose dynamic characteristics change with the production process, and real-time prediction of calorific value, which is conducive to assisting density decision-making and calorific value control in the coal preparation process, thereby improving the quality of clean coal products.

[0082] Optionally, the formula for the calorific value digital twin model is as follows:

[0083]

[0084] Among them, y(k) is the calorific value of clean coal, x u , u=[1,9] is the relevant variable, x1 is the set value of the density of the heavy medium; x2 is the density of the first heavy medium barrel; x3 is the density of the second heavy medium barrel; x4 is the liquid level of the first heavy medium barrel; x5 is the liquid level of the first heavy medium barrel; x6 is the pressure of the first cyclone; x7 is the pressure of the second cyclone; x8 is the opening of the first valve; x9 is the opening of the second valve, n v ,v=[1,9] is the variable lag time, is the parameter identification equation, is a nonlinear term, and a, b, c, d, e, f, g, h, i, and l are identification parameters.

[0085] The above formula is used to establish a calorific value digital twin model that includes a linear parameter identification part and a nonlinear part based on LSTM learning. Then, by performing parameter identification on the linear parameter identification part and correcting the nonlinear part, an endpoint prediction model that can predict the calorific value of coal preparation in real time can be obtained, thereby realizing the modeling of industrial process operation indicators whose dynamic characteristics change with the production process.

[0086] Optionally, for linear models Perform parameter identification and obtain identification error to construct nonlinear dynamic system v(k) based on nonlinear terms and identification error, including: determining linear model based on calorific value digital twin model y(k) Will Expressed as Y(k)=X(k-1)θ, the parameter θ is identified based on the objective function J1 of the least squares algorithm to obtain the estimated value of the parameter θ Where, Y(k) is the input vector, Y(k) = [y(k)y(k+1)y(k+2)…y(k+N)] T , X(k-1) is the input matrix, Will As the identification error, according to the identification error With nonlinear terms The nonlinear dynamic system v(k) is constructed as follows:

[0087] The linear parameter identification was performed using the least squares algorithm with the 162,048 sets of processed data, and the parameter estimates obtained were for: The calorific value obtained by the linear part of the least squares identification model is compared with the actual calorific value image. Figure 2 As shown in the figure, the error RMSE is 80.7423. The error caused by the parameter estimation value is used as the identification error. Based on the identification error and the nonlinear term, a nonlinear dynamic system is constructed. At this time, the sum of the nonlinear dynamic system and the linear model is the thermal value digital twin model.

[0088] Among them, the variable lag time n1 = 500s, n2 = 525s, n3 = 700s, n4 = 605s, n5 = 430s, n6 = 485s, n7 = 940s, n8 = 875s, and n9 = 665s.

[0089] Optionally, a long short-term memory multilayer neural network LSTM is used to construct an offline deep learning model of the nonlinear dynamic system v(k), including: obtaining relevant quantity data and calorific value test data as sample data; determining the number of hidden layer nodes h, the number of network layers L, and the number of network neurons n of the long short-term memory multilayer neural network LSTM based on the sample data to obtain an offline deep learning model.

[0090] By obtaining relevant variable data in the coal preparation process and calorific value test data, sample data is formed, and then the network architecture of the offline deep learning model is determined based on the sample data, namely the number of hidden layer nodes h, the number of network layers L, and the number of network neurons n.

[0091] Specifically, N sets of input and output data are used to form a large data sample. The number of network layers is selected as 1. The order n of the input data vector is selected as the number of neurons in each layer. The data vectors ψ(k-1), ψ(k-2), ..., ψ(kn) are used as input data, and the data vectors v(k), v(k-1), ..., v(k-n+1) are used as output data. A training algorithm is used to make the label data v(k) consistent with the actual output of the offline deep learning model. The error Δv(k) is as small as possible. The objective function of the training algorithm is as follows:

[0092]

[0093] in and are the weights and biases of the fully connected layer, and h1(k) is the output of the nth neuron.

[0094] Among them, h1(k) T =o1(k)⊙tanh(C1(k)), The output gate o1(k) and the state gate C1(k) are o1(k)=σ(W1 o x1(k) T+b1 o ), In the output gate o1(k), The input to the nth neuron is x1(k)=[x1(k-n1),...,x9(k-n9),v(k-1)] T , in the state gate C1(k), the forget gate f1=σ(W1 f x1(k) T +b1 f ), input gate i1=σ(W1 in x1(k) T +b1 in ), state candidate value is a matrix of h×(h+3), b1 f , b1 in , b1 c is an h×1 column vector.

[0095] The gradient descent algorithm is used to train the connection weights and bias parameters, and the output gate connection weight W1 o and weight b1 o The training algorithm is

[0096] in, α is the learning rate. The same training algorithm is then used to determine the other connection weights and biases of the model.

[0097] Next, in order to identify the structure of the deep learning model, the root mean square error of Δv(k) is introduced:

[0098]

[0099] Specifically, 22,000 sets of actual process data are used as training sets, and 8,169 sets of data are used as test set data to calculate the RMSE index of the output error of the nonlinear dynamic system model. First, the number of LSTM neurons n is determined. Let n = 1, 2, 3, 4, ... and calculate the root mean square error RMSE of the model output error Δv(k). When the RMSE of Δv(k) is minimized, the value of the number of neurons n is determined. Then fix the number of neurons n, let the number of nodes h = 1, 2, 3, 4, ... and calculate the root mean square error of the model output error Δv(k). When the RMSE of Δv(k) is minimized, the value of the number of hidden layer nodes h is determined. Finally, fix n and h of the deep learning model, and calculate the root mean square error of the model output error Δv(k) by increasing the number of network layers L = 1, 2, 3, 4, .... When the RMSE of Δv(k) is minimized, the value of the number of network layers L is determined. Finally, as Figure 3 、 Figure 4 and Figure 5As shown, it can be concluded that the number of neurons n is 30, the number of hidden layer nodes of a single neuron h is 188, and the number of network layers L is 3. The network architecture of the offline deep learning model is determined as follows: Figure 6 shown.

[0100] Optionally, based on an offline deep learning model Use long short-term memory multi-layer neural network LSTM to build an edge online deep learning model and cloud-based deep learning calibration models Including: Based on the long short-term memory multi-layer neural network LSTM based on the offline deep learning model The side-by-side online deep learning model is constructed using the number of hidden layer nodes h, the number of network layers L, and the number of network neurons n. and cloud-based deep learning calibration models

[0101] The side-online deep learning model is constructed using the same number of hidden layer nodes h, network layers L, and network neurons n as the offline deep learning model. and cloud-based deep learning calibration models To facilitate online deep learning models on the side and cloud-based deep learning calibration models Correction of nonlinear dynamic systems.

[0102] Optionally, the cloud-based deep learning model is calibrated based on all relevant variable data currently updated in real time in the cloud database. Perform calibration, including: calibrating the cloud-based deep learning model based on all relevant variable data currently updated in real time in the cloud database All layer weights and bias parameters are corrected.

[0103] All relevant variable data currently updated in real time in the cloud database constitute N S (k) groups of real-time updated production process data ψ(k), ψ(k-1), ..., ψ(k-n+1) (k=1, ..., N S ), real-time correction of all layer weights and biases of the cloud-based deep learning correction model. As the coal preparation process proceeds, the number of data N S (k) continues to increase. The cloud-based deep learning calibration model continuously performs calibration training on all relevant process and test data in the continuously updated cloud database, and promptly updates all layer weights and bias parameters of the cloud-based deep learning calibration model.

[0104] It should be noted that the objective function of the training algorithm is Among them, N S (k) represents the number of input and output data pairs used by the cloud-based deep learning correction model. In fact, NS (k), k represents the kth sampling moment, the sampling period is 15 seconds, N S (k) It increases continuously as the coal preparation process progresses, increasing by 1 every 15 seconds.

[0105] Optionally, use the calibrated cloud-based deep learning calibration model According to the preset self-correction mechanism, the side online deep learning model Perform correction to complete the correction of the nonlinear dynamic system v(k) and obtain the estimated value of the nonlinear dynamic system, including: calculating the corrected cloud-based deep learning correction model Model accuracy evaluation index RMSE1(k) and side-by-side online deep learning model The model accuracy evaluation index RMSE2(k); RMSE1(k) and RMSE2(k) are compared with the preset self-correction mechanism to generate comparison results; when the comparison result indicates the correction model, the corrected cloud deep learning correction model Replace the side-by-side online deep learning model with all layer weights and bias parameters All layer weights and bias parameters to complete the side-by-side online deep learning model The correction of the nonlinear dynamic system v(k) is completed, and the estimated value of the nonlinear dynamic system is obtained.

[0106] Based on the experimental results, an upper bound (RMSE = δ1) was set for the model output of the nonlinear dynamic system. Based on this upper bound, a predefined self-correction mechanism was derived. The RMSE, an evaluation metric for the self-correction mechanism, was calculated using 8,169 data sets from the test set. The calculated RMSE was 20.

[0107] Then calculate the model accuracy evaluation index of the deep learning correction model according to the formula:

[0108]

[0109] Where k is the sampling period (k=1, 2, ...), is the output value of the side online deep learning model.

[0110] The model accuracy evaluation index of the cloud-based deep learning correction model is calculated using the same method, denoted as RMSE2(k). The model accuracy evaluation indexes of the side-online deep learning model and the cloud-based deep learning correction model in the calorific value endpoint prediction model are respectively tested. The model accuracy evaluation indexes of the side-online deep learning model and the cloud-based deep learning correction model are compared with the preset self-correction mechanism to generate a comparison result. When the model accuracy evaluation index of the side-online deep learning model in the calorific value endpoint prediction model (RMSE1(k) ≥ 20) and the model accuracy evaluation index of the cloud-based deep learning correction model (RMSE2(k) < 20) are both equal, a comparison result indicating model correction is generated. All layer weights and bias parameters of the cloud-based deep learning correction model are then downloaded and replaced with those of the side-online deep learning model to complete the correction of the deep learning correction model.

[0111] In actual applications, the end-side is a data acquisition and transmission system used to collect process data and calorific value test data. The cloud-side is a cloud-based computing device, such as an NVIDIA server, which is used to calibrate all layer weights and bias parameters of the cloud-based deep learning calibration model based on real-time production process data updated in the cloud database. The edge-side is an edge computing device, such as a Dongtu Industrial server, which is used to predict real-time calorific value data using the calorific value endpoint prediction model given the current density setpoint and other relevant variable data. When the self-calibration mechanism conditions are met, all layer weights and biases of the cloud-based deep learning calibration model are downloaded and replaced with all layer weights and biases of the edge-side online deep learning model. It should be noted that in actual applications, the output of the nonlinear dynamic system after calibration is reflected in the output of the deep learning calibration model in the calorific value endpoint prediction model. Therefore, the output value of the linear model in the calorific value endpoint prediction model is finally added to the output value of the deep learning calibration model in the calorific value endpoint prediction model to output the current calorific value.

[0112] Optionally, based on a linear model and the estimated value of the nonlinear dynamic system Constructing an endpoint prediction model The method includes: using an unknown nonlinear function f(·) to represent a nonlinear dynamic system v(k): v(k)=f(ψ(k-1), ψ(k-2), ..., ψ(kn)), where f(·) is a nonlinear function of unknown variation, n is the order of the nonlinear term, ψ(k-1) is an input data vector, in is obtained by performing zero-order hold interpolation on the discrete thermal value data. y(k) is obtained by linear interpolation of the discrete heat value data; the heat value digital twin model y(k) is expressed as: To obtain the dynamic model of calorific value y(k+1) at time k+1: Establishing a prediction model for nonlinear dynamic systems Establishing a calorific value endpoint prediction model

[0113] A prediction model is established based on the nonlinear dynamic system, and then a calorific value endpoint prediction model is established based on the prediction model and the linear model, so as to make real-time predictions on calorific value data based on the calorific value endpoint prediction model.

[0114] In order to evaluate the accuracy of the prediction model, the following model is introduced: according to the fluctuation range of the calorific value target value required on site, the proportion of the calorific value predicted by the model falling within the range of 30 degrees above and below the actual calorific value is calculated.

[0115] Figure 7 The accuracy test results of the cloud-based deep learning calibration model were calculated, and the RMSE was 16.924; Figure 8 The calorific value endpoint prediction model accuracy test results show a calculated model prediction accuracy of 94.6129%. These test results show that the cloud-based deep learning calibration model ensures high accuracy for the calorific value endpoint prediction model. Therefore, when the accuracy of the sideline online deep learning model for the calorific value endpoint prediction model falls short of requirements, a self-calibration mechanism is implemented. All layer weights and biases of the cloud-based calibration model are downloaded and replaced with the sideline online deep learning model.

[0116] The calorific value endpoint prediction model is simulated in real time based on training data of different orders of magnitude. The real-time simulation results of the cloud-based deep learning correction model are 80,000 sets of training data and 147,000 sets of training data respectively. Figure 9 and Figure 7 As shown in the figure, the prediction error RMSE of the cloud-based deep learning correction model obtained with 80,000 sets of training data is 23.635, and the prediction error RMSE of the cloud-based deep learning correction model obtained with 147,000 sets of training data is 16.924, with an error reduction of 6.711. Figure 10 It can be seen that with the increase of the amount of training data, the root mean square error of the output error of the cloud-based deep learning correction model shows a downward trend. The test results of the calorific value endpoint prediction model using 80,000 sets of training data for real-time simulation and 147,000 sets of training data for real-time simulation are respectively Figure 11 and Figure 8 As shown. Figure 11 and Figure 8As can be seen, the calorific value endpoint prediction model using 147,000 sets of training data for real-time simulation achieves significantly better accuracy than the model using 80,000 sets of training data. The prediction accuracy of the calorific value endpoint prediction model using 147,000 sets of training data and 80,000 sets of training data for real-time simulation is 94.6129% and 84.7096%, respectively, representing a 9.9033% improvement in model accuracy. Therefore, as the amount of training data increases, the model learns more dynamic characteristics. Furthermore, through device-edge-cloud collaborative correction, the model accuracy continues to improve, meeting prediction requirements.

[0117] Optionally, the method further includes: sending the predicted calorific value data to a display terminal, and displaying the predicted calorific value data through the display terminal.

[0118] Second, as Figure 12 As shown, an embodiment of the present invention provides a coal preparation calorific value prediction system 1200 based on end-edge-cloud collaboration, which includes: a first construction module 1201, a second construction module 1202, a third construction module 1203, a fourth construction module 1204, a data acquisition module 1205, a first correction module 1206, a second correction module 1207, a fifth construction module 1208, and a calorific value prediction module 1209. Among them, the first construction module is used to select relevant variables that affect the calorific value change from a process perspective as model input, and use the clean coal calorific value as the model output to establish a calorific value digital twin model y(k). The calorific value digital twin model y(k) includes a linear model and nonlinear terms; the second building block is used to Perform parameter identification to obtain the identification error, and then construct the nonlinear dynamic system v(k) based on the nonlinear term and the identification error; the third construction module is used to construct an offline deep learning model of the nonlinear dynamic system v(k) using a long short-term memory multilayer neural network LSTM The fourth building block is used to Use long short-term memory multi-layer neural network LSTM to build an edge online deep learning model and cloud-based deep learning calibration models The data acquisition module is used to transmit the relevant variable data of the coal preparation production process collected on the end side to the cloud database; the first correction module is used to correct the cloud deep learning model according to all the relevant variable data currently updated in real time in the cloud database. The second correction module is used to use the corrected cloud-based deep learning correction model According to the preset self-correction mechanism, the side online deep learning model Correction is performed to complete the correction of the nonlinear dynamic system v(k) and obtain the estimated value of the nonlinear dynamic system The fifth building block is used based on the linear model and the estimated value of the nonlinear dynamic system Constructing an endpoint prediction model The calorific value prediction module is used to input density setpoints and related variable data into the endpoint prediction model, which then outputs the predicted calorific value data to achieve calorific value prediction for the coal preparation process. The coal preparation calorific value prediction system based on end-edge-cloud collaboration, provided according to the technical solution of the present invention, implements the steps of the coal preparation calorific value prediction method based on end-edge-cloud collaboration provided in the first aspect of the present invention. Therefore, this coal preparation calorific value prediction system based on end-edge-cloud collaboration has all the technical effects of the coal preparation calorific value prediction method based on end-edge-cloud collaboration, and will not be further elaborated here.

[0119] In a third aspect, an embodiment of the present invention provides a storage device having a computer program stored thereon, wherein the program, when executed by a processor, implements the steps of any one of the methods described in the first aspect.

[0120] In a fourth aspect, an embodiment of the present invention provides a computer storage medium, comprising a storage device, a processor, and a computer program stored on the storage device and executable on the processor, wherein when the processor executes the program, the steps of any one of the methods described in the first aspect are implemented.

[0121] In addition, this embodiment also provides a schematic diagram of the deployment of an endpoint prediction system for coal preparation calorific value production indicators based on end-edge-cloud collaboration. In this embodiment, an end-edge-cloud architecture is adopted, in which the end side acquires production process data in real time or periodically, and transmits the acquired data to the edge database for storage through a switch on the one hand, and sends it to the cloud storage platform (i.e., industrial cloud server) for real-time storage on the other hand via wireless transmission, so as to realize the "cloud" of industrial site data for subsequent reading and viewing. At the same time, the cloud side uses the powerful computing power of the industrial cloud server to train the cloud-side deep learning correction model based on the stored production data, and communicates with the edge computing device. When the edge calorific value endpoint prediction model does not reach the target accuracy, the model trained on the cloud side is sent to the edge side. The edge computing device makes a real-time prediction of the production indicators based on the updated edge model, and sends the predicted data to the edge terminal for display. Based on the proposed system architecture, corresponding devices are deployed on the end-edge-cloud side to execute the endpoint prediction method of coal preparation calorific value production indicators based on end-edge-cloud collaboration in the above embodiment.

[0122] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0123] Obviously, those skilled in the art may make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if such modifications and variations fall within the scope of the claims and their equivalents, the present invention shall also include such modifications and variations.

[0124] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0125] In the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," "connect," "fixed," etc. should be understood broadly. For example, they may refer to fixed connection, detachable connection, or integration; mechanical connection or electrical connection; direct connection or indirect connection through an intermediate medium; and internal communication between two components or interaction between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0126] In the present invention, unless otherwise expressly specified or limited, when a first feature is "above" or "below" a second feature, it may mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediate medium. Furthermore, when a first feature is "above," "above," or "above" a second feature, it may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is at a higher level than the second feature. When a first feature is "below," "below," or "below" a second feature, it may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is at a lower level than the second feature.

[0127] In the description of this specification, the terms "one embodiment", "some embodiments", "embodiments", "examples", "specific examples" or "some examples" refer to the specific features, structures, materials or characteristics described in conjunction with the embodiment or example and included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and features of different embodiments or examples, unless they are mutually inconsistent.

[0128] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may alter, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A coal preparation calorific value prediction method based on end-edge-cloud collaboration, characterized in that: include: The relevant variables that affect the calorific value change from the process perspective are selected as the model input, and the clean coal calorific value is used as the model output to establish the calorific value digital twin model y(k). The calorific value digital twin model y(k) includes the linear model and nonlinear terms; For linear models Perform parameter identification to obtain an identification error, so as to construct a nonlinear dynamic system v(k) according to the nonlinear term and the identification error; The offline deep learning model of the nonlinear dynamic system v(k) is constructed using a long short-term memory multilayer neural network LSTM. According to the offline deep learning model Use long short-term memory multi-layer neural network LSTM to build an edge online deep learning model and cloud-based deep learning calibration models Transmit relevant variable data of the coal preparation production process collected on the terminal side to the cloud database; The cloud-based deep learning model is calibrated based on all relevant variable data currently updated in real time in the cloud database. Correction; Use the calibrated cloud-based deep learning calibration model According to the preset self-correction mechanism, the side online deep learning model Correction is performed to complete the correction of the nonlinear dynamic system v(k) and obtain the estimated value of the nonlinear dynamic system Based on linear model and the estimated value of the nonlinear dynamic system Constructing an endpoint prediction model Inputting a density setting value and related variable data into the endpoint prediction model, the endpoint prediction model outputs a calorific value data prediction value to achieve calorific value prediction of the coal preparation process; The formula of the calorific value digital twin model is as follows: Among them, y(k) is the calorific value of clean coal, x u , u=[1,9] is the relevant variable, x1 is the set value of the density of the heavy medium; x2 is the density of the first heavy medium barrel; x3 is the density of the second heavy medium barrel; x4 is the liquid level of the first heavy medium barrel; x5 is the liquid level of the first heavy medium barrel; x6 is the pressure of the first cyclone; x7 is the pressure of the second cyclone; x8 is the opening of the first valve; x9 is the opening of the second valve, n v , v = [1, 9] is the variable lag time, k is the sampling time, is the parameter identification equation, is a nonlinear term, a, b, c, d, e, f, g, h, i, l are identification parameters; For linear models Perform parameter identification to obtain an identification error, so as to construct a nonlinear dynamic system v(k) based on the nonlinear term and the identification error, including: Determine the linear model based on the calorific value digital twin model y(k) Will Expressed as Y(k)=X(k-1)θ, the parameter θ is identified based on the objective function J1 of the least squares algorithm to obtain the estimated value of the parameter θ Where, Y(k) is the input vector, Y(k) = [y(k) y(k+1) y(k+2) … y(k+N)] T , X(k-1) is the input matrix, Will As the identification error, according to the identification error With nonlinear terms The nonlinear dynamic system v(k) is constructed as follows:

2. The method according to claim 1, characterized in that The offline deep learning model of the nonlinear dynamic system v(k) is constructed using a long short-term memory multilayer neural network LSTM. include: Obtain relevant variable data and calorific value test data as sample data; Based on the sample data, the number of hidden layer nodes h, the number of network layers L, and the number of network neurons n of the long short-term memory multilayer neural network LSTM are determined to obtain the offline deep learning model 3. The method according to claim 2, characterized in that According to the offline deep learning model Use long short-term memory multi-layer neural network LSTM to build an edge online deep learning model and cloud-based deep learning calibration models include: Based on the offline deep learning model based on the long short-term memory multi-layer neural network LSTM The side-online deep learning model is constructed using the number of hidden layer nodes h, the number of network layers L, and the number of network neurons n. and cloud-based deep learning calibration models 4. The method according to claim 1, wherein The cloud-based deep learning model is calibrated based on all relevant variable data currently updated in real time in the cloud database. Perform calibration, including: The cloud-based deep learning model is calibrated based on all relevant variable data currently updated in real time in the cloud database. All layer weights and bias parameters are corrected.

5. The method according to claim 2, characterized in that Use the calibrated cloud-based deep learning calibration model According to the preset self-correction mechanism, the side online deep learning model Correction is performed to complete the correction of the nonlinear dynamic system v(k) and obtain the estimated value of the nonlinear dynamic system include: Calculate the corrected cloud-based deep learning calibration model Model accuracy evaluation index RMSE1(k) and side-by-side online deep learning model Model accuracy evaluation index RMSE2(k); Compare RMSE1(k) and RMSE2(k) with the preset self-correction mechanism to generate comparison results; In the case where the comparison result indicates the correction model, the deep learning correction model after the correction is fixed The number of hidden layer nodes h, the number of network layers L, and the number of network neurons n are used to calibrate the cloud-based deep learning model. Replace the side-by-side online deep learning model with all layer weights and bias parameters All layer weights and bias parameters to complete the side-by-side online deep learning model The correction of the nonlinear dynamic system v(k) is completed, and the estimated value of the nonlinear dynamic system is obtained.

6. The method according to any one of claims 1 to 5, characterized in that Based on linear model and the estimated value of the nonlinear dynamic system Constructing an endpoint prediction model include: The nonlinear dynamic system v(k) is represented by an unknown nonlinear function f(·): v(k) = f(ψ(k-1), ψ(k-2), ..., ψ(kn)), where f(·) is the unknown nonlinear function, n is the order of the nonlinear term, and ψ(k-1) is the input data vector. in is obtained by performing zero-order hold interpolation on the discrete thermal value data. y(k) is obtained by linear interpolation of the discrete heat value data; The calorific value digital twin model y(k) is expressed as: To obtain the dynamic model of calorific value y(k+1) at time k+1: Establishing a prediction model for nonlinear dynamic systems Establishing a calorific value endpoint prediction model 7. A coal preparation calorific value prediction system based on end-edge-cloud collaboration, characterized in that: include: The first building block is used to select relevant variables that affect the calorific value change from a process perspective as model input, and use the clean coal calorific value as the model output to establish a calorific value digital twin model y(k). The calorific value digital twin model y(k) includes a linear model and nonlinear terms; The second building block is used to Perform parameter identification to obtain an identification error, so as to construct a nonlinear dynamic system v(k) according to the nonlinear term and the identification error; The third building block is used to construct an offline deep learning model of the nonlinear dynamic system v(k) using a long short-term memory multi-layer neural network LSTM The fourth building block is used to build a Use long short-term memory multi-layer neural network LSTM to build an edge online deep learning model and cloud-based deep learning calibration models The data acquisition module is used to transmit the relevant variable data of the coal preparation production process collected on the terminal side to the cloud database; The first correction module is used to correct the cloud deep learning model based on all relevant variable data currently updated in real time in the cloud database Correction; The second correction module is used to use the corrected cloud-based deep learning correction model According to the preset self-correction mechanism, the side online deep learning model Correction is performed to complete the correction of the nonlinear dynamic system v(k) and obtain the estimated value of the nonlinear dynamic system The fifth building block is used for linear model-based and the estimated value of the nonlinear dynamic system Constructing an endpoint prediction model The calorific value prediction module is used to input the density setting value and related variable data into the endpoint prediction model, and the endpoint prediction model outputs the calorific value data prediction value to realize the calorific value prediction of the coal preparation process. The formula of the calorific value digital twin model is as follows: Among them, y(k) is the calorific value of clean coal, x u , u=[1,9] is the relevant variable, x1 is the set value of the density of the heavy medium; x2 is the density of the first heavy medium barrel; x3 is the density of the second heavy medium barrel; x4 is the liquid level of the first heavy medium barrel; x5 is the liquid level of the first heavy medium barrel; x6 is the pressure of the first cyclone; x7 is the pressure of the second cyclone; x8 is the opening of the first valve; x9 is the opening of the second valve, n v ,v=[1,9] is the variable lag time, k is the sampling time, is the parameter identification equation, is a nonlinear term, a, b, c, d, e, f, g, h, i, l are identification parameters; For linear models Perform parameter identification to obtain an identification error, so as to construct a nonlinear dynamic system v(k) based on the nonlinear term and the identification error, including: Determine the linear model based on the calorific value digital twin model y(k) Will Expressed as Y(k)=X(k-1)θ, the parameter θ is identified based on the objective function J1 of the least squares algorithm to obtain the estimated value of the parameter θ Where, Y(k) is the input vector, Y(k) = [y(k) y(k+1) y(k+2) … y(k+N)] T , X(k-1) is the input matrix, Will As the identification error, according to the identification error With nonlinear terms The nonlinear dynamic system v(k) is constructed as follows:

8. A storage device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images