Machine learning modeling method, program product and equipment for non-analytic physical system

Abstract

The embodiment of the invention relates to a machine learning modeling method and device for a non-analytic physical system, a program product and equipment, and the method comprises the steps: obtaining a non-analytic model of the physical system, and determining a control variable and an observation variable in the non-analytic model; selecting a measurement function representing a control variable to an observation variable, measuring the control variable by using the measurement function to determine a measurement data matrix, and performing pattern extraction based on the measurement data matrix to obtain a pattern of a measurement space; under the condition that the mode of the measurement space is correct, basic data are obtained from operation data of a physical system, data prediction is carried out on the basic data based on the mode of the measurement space, and prediction data are generated; and determining a polynomial fitting model corresponding to the equipment belonging to the non-analytical model in the physical system by using the prediction data. According to the invention, the polynomial fitting model of the non-analytical model corresponding to each device in the physical system can be accurately constructed.

Description

TECHNICAL FIELD

[0001] Embodiments of the present disclosure relate to the technical field of physical system modeling, and in particular, to a machine learning modeling method for non-analytical physical systems, a computer program product, and an electronic device. BACKGROUND

[0002] In the modeling of physical systems such as fluid systems, thermal systems, and electromechanical systems, analytical modeling is generally relied on. However, for some devices or processes in the system, the traditional analytical modeling method is difficult to accurately express, and the modeling cost is too high. In addition, simply relying on a data-driven black box model has problems such as poor interpretability and large data requirements. SUMMARY

[0003] The purpose of the present disclosure is to provide a machine learning modeling method for non-analytical physical systems, a computer program product, and an electronic device, thereby at least partially overcoming the problem of low modeling accuracy and poor interpretability due to the limitations and defects of the related art.

[0004] According to an aspect of the present disclosure, a machine learning modeling method for non-analytical physical systems is provided, comprising:

[0005] Obtaining a non-analytical model of a physical system, and determining control variables and observation variables in the non-analytical model;

[0006] Selecting a measurement function that describes the relationship between the control variables and the observation variables, using the measurement function to measure the control variables to determine a measurement data matrix, and performing pattern extraction based on the measurement data matrix to obtain a pattern of a measurement space;

[0007] In the case where the pattern of the measurement space is correct, obtaining basic data from the operation data of the physical system, performing data prediction on the basic data based on the pattern of the measurement space to generate predicted data;

[0008] Using the predicted data, determining a polynomial fitting model corresponding to a device belonging to the non-analytical model in the physical system.

[0009] In an exemplary embodiment of the present disclosure, the pattern extraction based on the measurement data matrix to obtain the pattern of the measurement space comprises:

[0010] Extracting a first matrix and a second matrix from the measurement data matrix, representing the second matrix through the first matrix to obtain an intermediate matrix;

[0011] Performing singular value decomposition on the first matrix, and performing dimension reduction on the first matrix based on the decomposition matrix to obtain a dimension-reduced measurement data matrix;

[0012] merge the reduced dimension measurement data matrix and the intermediate matrix, and perform eigen decomposition to obtain a mode of a measurement space.

[0013] In an example embodiment of the present disclosure, the merging the reduced dimension measurement data matrix and the intermediate matrix, and performing eigen decomposition to obtain a mode of a measurement space comprises:

[0014] merging the reduced dimension measurement data matrix and the intermediate matrix to generate a merged matrix K;

[0015] reducing dimension of the merged matrix to obtain a reduced dimension matrix, and performing eigen decomposition on the reduced dimension matrix to determine a decomposition matrix;

[0016] substituting the reduced dimension matrix into the eigen decomposition matrix, and determining a reference expression in combination with a preset value;

[0017] determining the mode of the measurement space based on a first column of the first matrix, an eigenvalue and an eigenvector of the merged matrix representing any column of the measurement data matrix.

[0018] In an example embodiment of the present disclosure, the measurement function comprises a first undetermined parameter and a second undetermined parameter, the first undetermined parameter is a highest power of the measurement function, and the second undetermined parameter is a dimension of the mode extraction.

[0019] In an example embodiment of the present disclosure, the determination process of the first undetermined parameter and the second undetermined parameter comprises:

[0020] performing data prediction on a control variable according to the mode of the measurement space, determining a predicted value of an observation variable corresponding to the control variable, and determining an error between the predicted value and a true value of the observation variable;

[0021] determining a target function according to the error and the first undetermined parameter, and determining the first undetermined parameter and the second undetermined parameter when the target function satisfies a condition.

[0022] In an example embodiment of the present disclosure, the data prediction on the basis data based on the mode of the measurement space to generate predicted data comprises:

[0023] when the first undetermined parameter is greater than the second undetermined parameter, obtaining basis data; the basis data is operation data of each device in the physical system;

[0024] performing data prediction based on the mode of the measurement space and an initial value of the basis data to determine initial predicted data;

[0025] returning the initial prediction data to the original space, and re-determining the basis data, and performing data prediction based on the re-determined basis data and the mode of the measurement space in a loop until a stop condition is met, to determine the prediction data.

[0026] In an example embodiment of the present disclosure, the re-determining the basis data comprises:

[0027] part of the initial prediction data is taken as original space data, and the original space data is transposed to re-determine the basis data.

[0028] In an example embodiment of the present disclosure, the physical system is an energy system, and the control variable is a variable for a target device in the energy system, the target device at least including one of a pump, a heat exchanger, and a cooling tower.

[0029] According to an aspect of the present disclosure, there is provided a computer program product, the computer program being executed by a processor to implement the machine learning modeling method of the non-analytical physical system according to any one of the preceding aspects.

[0030] According to an aspect of the present disclosure, there is provided an electronic device comprising:

[0031] a processor;

[0032] a memory for storing executable instructions of the processor;

[0033] wherein the processor is configured to implement the machine learning modeling method of the non-analytical physical system according to any one of the preceding aspects via execution of the executable instructions.

[0034] In the technical solutions provided in the embodiments of the present disclosure, on the one hand, the control variable in the non-analytical model is measured by selecting a measurement function to obtain a measurement data matrix, and the mode of the measurement space is obtained according to the measurement data matrix, the basis data of each device in the physical system is predicted based on the mode of the measurement space to obtain prediction data, and then a polynomial fitting model of each device is determined according to the prediction data, which can accurately establish the model of the device and reduce the modeling cost. On the other hand, since the polynomial fitting model is constructed based on the prediction data determined based on the mode of the measurement space, the amount of data required for model construction is reduced, and the interpretability of model construction is improved.

[0035] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and are not limiting to the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0036] The accompanying drawings, which are incorporated herein and form a part of the specification, illustrate embodiments consistent with the present disclosure and, together with the description, further serve to explain the principles of the present disclosure. It is to be understood that the drawings are designed solely for purposes of illustration and are not intended to limit the scope of the disclosure in any way.

[0037] Figure 1 A flowchart schematically illustrates a machine learning modeling method for a non-analytical physical system in embodiments of the present disclosure.

[0038] Figure 2 A structural schematic diagram of an energy system in embodiments of the present disclosure is schematically shown.

[0039] Figure 3 A variable relationship schematic diagram of an energy system in embodiments of the present disclosure is schematically shown.

[0040] Figure 4 A training result schematic diagram of a measurement variable of a pump in embodiments of the present disclosure is schematically shown.

[0041] Figure 5 A training result schematic diagram of a measurement variable of a heat exchanger in embodiments of the present disclosure is schematically shown.

[0042] Figure 6 A training result of a measurement variable of a cooling tower in embodiments of the present disclosure is schematically shown.

[0043] Figure 7 A flowchart schematically illustrates a machine learning modeling method for a non-analytical physical system in embodiments of the present disclosure.

[0044] Figure 8 A block diagram schematically illustrates a machine learning modeling device for a non-analytical physical system in embodiments of the present disclosure.

[0045] Figure 9 A block diagram schematically illustrates an electronic device in embodiments of the present disclosure. DETAILED DESCRIPTION

[0046] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations can be implemented in any number of manners, and are not limited to the examples described herein; rather, examples are provided as exemplary embodiments of the present disclosure to provide a thorough and enabling disclosure of present implementations, and to fully convey the scope of the concepts of example implementations to those skilled in the art. Described features, structures, or characteristics can be combined in any suitable manner in one or more implementations. In the following description, numerous specific details are provided to give a thorough understanding of implementations of the present disclosure. One skilled in the relevant art will recognize, however, that the techniques of the present disclosure can be practiced without one or more of the specific details, or with other methods, components, devices, steps, etc. In other instances, well-known structures have not been described in detail so as not to obscure aspects of the present disclosure.

[0047] Furthermore, the accompanying drawings are only intended to show illustrative views of the present disclosure, and are not necessarily drawn to scale. The same reference numerals in the drawings represent the same or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the drawings are functional entities, which do not necessarily have to correspond to physically or logically independent entities. These functional entities can be implemented in the form of software, or in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0048] In the modeling process of physical systems such as fluid systems, thermal systems, electromechanical systems, etc., the conventional method is generally to establish an analytical model based on physical laws. However, for some devices or processes in the physical system, the internal mechanism is complex, the variable relationship is highly nonlinear, and the traditional analytical modeling method is difficult to accurately express, and the modeling cost is too high. In addition, a purely data-driven black box model is flexible, but has problems such as poor interpretability and large data requirements. Therefore, a modeling method that combines physical structure knowledge and data-driven ability is needed, which is suitable for modeling tasks of complex physical systems.

[0049] Therefore, in order to solve the above technical problems, a machine learning modeling method for non-analytical physical systems is provided in the embodiments of the present disclosure. The modeling algorithm provided by the method is a hybrid modeling method combining Koopman operator and machine learning algorithm, which can be applied to modeling tasks of complex physical systems, such as fluid systems, thermal systems, electromechanical systems, energy systems, etc. Here, the application in energy systems is taken as an example for description. Figure 1 FIG. 1 schematically shows a machine learning modeling method for non-analytical physical systems, which is described in detail in the following. Figure 1 The method mainly includes the following steps:

[0050] Step S110, a non-analytical model of a physical system is acquired, and a control variable and a measured variable in the non-analytical model are determined.

[0051] Step S120, a measurement function describing a relationship between the control variable and the measured variable is selected, measurement data matrixes are determined by measuring the control variable using the measurement function, and a mode of a measurement space is obtained based on the measurement data matrixes through mode extraction.

[0052] Step S130, in a case where the mode of the measurement space is correct, basic data are acquired from operation data of the physical system, the basic data are predicted based on the mode of the measurement space, and predicted data are generated.

[0053] Step S140, the predicted data are used to determine a polynomial fitting model corresponding to a device belonging to the non-analytical model in the physical system.

[0054] In the technical solution provided by the embodiments of the present disclosure, the influence of the control variable on the observed variables in the experimental data and the operation data of the device is analyzed first. Then, model extraction is performed to determine the minimum dimension and the maximum power number of each model, so as to perform data prediction with the least data. Finally, the analytical formula of the model is fitted. In the case where part of the variables cannot be analytically modeled, an explicit mathematical model with physical interpretability can be constructed by using the experimental data and the operation data, and effective modeling of the system behavior can be realized. The measurement data matrixes are obtained by measuring the control variable in the non-analytical model by selecting the measurement function, the mode of the measurement space is obtained according to the measurement data matrixes, the basic data of each device in the physical system are predicted based on the mode of the measurement space, the predicted data are obtained, and then the polynomial fitting model of each device is determined according to the predicted data. The model of the device can be accurately established, and the modeling cost is reduced. Since the polynomial fitting model is constructed based on the predicted data obtained based on the mode of the measurement space, the amount of data required for model construction is reduced, and the interpretability of model construction is improved.

[0055] Next, the machine learning modeling method of the non-analytical physical system in the embodiments of the present disclosure will be described in detail in combination with embodiments.

[0056] Step S110, a non-analytical model of a physical system is acquired, and a control variable and a measured variable in the non-analytical model are determined.

[0057] In the embodiments of the present disclosure, the physical system can be an energy system. The energy system can be a circulating water system of a refinery, which is a complex nonlinear system. The energy system can include main devices such as pumps, fans, heat exchangers, cooling towers, valves, and solar photovoltaic devices, as shown in FIG. 1. For reference, the energy system can include a plurality of devices, and each device can be a nonlinear system. Figure 2 Figure 2 ​As shown, the entire energy system may include a power grid, pumps, multiple control valves, cooling towers, multiple heat exchangers, and other equipment. The operation of the entire energy system includes the following processes: Pumps drive circulating water into the heat exchange network. Energy exchange occurs between the circulating water and the working fluid. The heated circulating water enters the cooling tower and is cooled by a fan. The cooled circulating water then enters a water tank.

[0058] One variable in the load of this electric heating system is the working fluid that needs to be cooled. Its input temperature and flow rate do not change over a short period, while the output temperature has an upper limit. Another variable in the load is the output temperature of the cooling tower, assumed to be approximately the same as the pool temperature. Variables in the directly controlled equipment of this electric heating system include: pumps and fans with variable frequency drives and adjustable speeds; and valve openings controlled by a controller.

[0059] The operation of this electric heating system involves the coupling of multiple energy sources, including electrical energy, thermal energy, cold energy, and water energy. Its state variables include temperature, water flow rate, air volume, pressure, and resistance. The measured parameters of the energy system can be used as state variables. All state variables used are listed in Table 1. Table 1 categorizes the relationship between state variables and equipment and indicates their category within the control system.

[0060] Table 1

[0061]

[0062]

[0063] To clearly illustrate the structure of the energy system, a block diagram is used to show the relationships between state variables, equipment, energy, and each state variable, such as... Figure 3 As shown. Reference Figure 3 As shown, the input side includes electrical energy, water energy, and cold energy. The output side includes electrical energy, water energy, cold energy, and heat energy. Electrical energy enters the pump and fan; water energy and cold energy enter the valves and are transmitted to multiple heat exchangers to generate cold energy output; water energy and cold energy enter the valves and are transmitted to the heat exchangers and then into the cooling tower to generate heat energy output.

[0064] The method in this embodiment of the disclosure primarily aims to control the five direct control variables (n) of the control device. P ,o P ,n F ,o H1 ,o H2 To meet the load requirements (To1, To2, tT) and reduce energy consumption at the source (E) P E F E PV E G ).

[0065] The model formula of a physical system can be generally expressed as formula (1). For a model without an analytical formula in formula (1), it is taken as a non-analytical model, which can be expressed in the form shown in formula (2).

[0066] x = f(x; μ) formula (1)

[0067]

[0068] where x (k) ∈x is a certain observation variable; is a control variable} is a control variable set of the model in formula (2), the number of which is N Ck .

[0069] Further, experiments are performed on the related equipment of the energy system to collect data. The method of data collection experiment is to linearly change one control variable in the non-analytical model in turn, and then record the values of all observation variables. At the same time, the previous operation data and the field debugging data are collected. The control variable can be the variable corresponding to each device in the energy system that can be operated or adjusted. The control variable corresponding to each device can be different, which is determined according to the control condition of each device.

[0070] For each control variable, N (i) time sampling is performed to obtain a tensor T with a dimension of For the change of the control variable from x (i) to the observation variable x (k) , the tensor T is organized into the form shown in formula (3). In formula (3), x (i) from different data sources is a column, and the dimension of which is

[0071]

[0072] where, is the initial value of the sample x (i) , and x (i) is the change amount of x (i) .

[0073] Next, the step of measurement space reconstruction based on the Koopman operator can be performed.

[0074] It is assumed that the change of the control variable from x (i) to the observation variable x (k) satisfies formula (4), and the discrete form thereof can be expressed by formula (5).

[0075]

[0076] where g() is x (k) relative to x(i) the dynamic characteristics.

[0077]

[0078] where G() is the discrete form of the dynamic characteristics, is a column of data of T.

[0079] Equation (5) is a discrete nonlinear expression, which is linearized by using the Koopman operator, and is expressed as a process shown in equation (6).

[0080]

[0081] where, is a set of measurement functions; K is the Koopman operator acting on the measurement function , which is an infinite-dimensional linear operator; represents the combination of the model.

[0082] The nonlinear model shown in equation (5) is converted into a high-dimensional linear model by using the Koopman operator, which can be expressed in the form shown in equation (7).

[0083]

[0084] In step S120, a measurement function describing the relationship between the control variable and the observation variable is selected, the measurement data matrix is determined by measuring the control variable using the measurement function, and the mode extraction is performed based on the measurement data matrix to obtain the mode of the measurement space.

[0085] In the embodiments of the present disclosure, the measurement function can be used to describe the relationship between the control variable and the measurement variable, and the measurement function can be a factor used to represent the change of the control variable to the measurement variable. The measurement function can be a function consistent with the design and operation process of the device, and the measurement function is used to measure the data. Specifically, the measurement function can be a measurement function set including a plurality of functions, and the measurement function set can be expressed as a function set of powers, wherein the functions are all power functions of the control variable. The highest power is a pending parameter. The measurement function can be expressed as equation (8):

[0086]

[0087] It should be noted that the measurement function can contain a first undetermined parameter p, which is used to change the highest power of the polynomial. The first undetermined parameter is used for final model fitting, and after the first undetermined parameter p is determined, the polynomial form of the fitted model can be accurately provided. The value of the first undetermined parameter p also affects the error of the predicted data and the data storage amount. At the same time, if the value of the first undetermined parameter p is too large, the power of the fitted model will also be too high, which will cause overfitting and oscillation problems. Therefore, in the process of determining this parameter, the minimum error is not the most suitable choice. Based on this, the minimum value of the first undetermined parameter p can be selected under the condition that the error is not too large, which can not only meet the requirements of error and data storage, but also can avoid overfitting and oscillation as much as possible.

[0088] Exemplarily, when performing Koopman operation, a measurement function is selected, and each element in the measurement function is a power function of the control variable. Based on this, the highest power can be selected as the first undetermined parameter. The first undetermined parameter is used for final model fitting, and the value of the first undetermined parameter p is determined to accurately provide the polynomial form of the fitted model.

[0089] After the measurement function is determined, the control variable can be measured by using the measurement function, that is, the measurement data matrix can be obtained, as shown in formula (9).

[0090]

[0091] Wherein, each column y i is a column of the measurement data matrix, and the dimension is

[0092] When the measurement data matrix is used for the ML algorithm, the first matrix and the second matrix can be extracted from the measurement data matrix. The first matrix can be composed of the values of the 1-mth columns in the measurement data matrix, as shown in formula (10); the second matrix can be composed of the values of the 2-m+1th columns in the measurement data matrix, as shown in formula (11). After obtaining the first matrix and the second matrix, pattern extraction can be performed by using the first matrix and the second matrix.

[0093]

[0094] After measurement, the second matrix can be represented by the first matrix to obtain an intermediate matrix. Based on this, formula (11) can be expressed in the form shown in formula (12), and the obtained intermediate matrix corresponds to a linear equation, and is a constant matrix to be solved.

[0095] K·Y1=Y2 Formula (12)

[0096] After obtaining the intermediate matrix, the mode extraction of the measurement space can be performed based on the intermediate matrix and the first matrix. The specific process of extracting the mode of the measurement space includes the following steps: singular value decomposition is performed on the first matrix, dimension reduction is performed on the first matrix based on the decomposition matrix to obtain a dimension-reduced measurement data matrix, and the dimension-reduced measurement data matrix is combined with the intermediate matrix, and eigenvalue decomposition is performed to obtain the mode of the measurement space.

[0097] Exemplarily, singular value decomposition is performed on the first matrix Y1, and the obtained decomposition matrix can be as shown in formula (13); the decomposition matrix can be represented as the product of three matrices.

[0098] Y1=U·S·V * Formula (13)

[0099] Wherein, U∈ n×n and V∈ m×m are unitary matrices, and S∈ n×m is a diagonal matrix.

[0100] When singular value decomposition is performed on the first matrix, the dimension of the model can be reduced, and the reduced dimension is taken as the second undetermined parameter r.

[0101] In the embodiments of the present disclosure, it is desirable that this parameter r is as small as possible, because r groups of data can be used for prediction. When the physical system is running, the more data, the more difficult it is to collect. A smaller second undetermined parameter is determined, and less data can reduce the difficulty of data collection.

[0102] Because the amount of experimental data is large, and when the system is actually running, a small amount of data is generally needed to complete modal extraction, so dimension reduction operation can be performed on the measurement data matrix. Exemplarily, on the basis of the decomposition matrix, U' can be taken as the first r columns of U, V' can be taken as the first r columns of V, and S' can be taken as the r×r upper left corner matrix of S. In this way, the measurement data matrix is reduced to r dimensions to obtain a dimension-reduced measurement data matrix, and the dimension-reduced measurement data matrix is represented in the form of formula (14).

[0103] Y1=U'·S'·V' * Formula (14)

[0104] Wherein, U'∈ n×r ; V'∈ m×r ; S'∈ r×r .

[0105] After obtaining the dimension-reduced measurement data matrix, the dimension-reduced measurement data matrix can be combined with the intermediate matrix, and the combination here refers to substituting the dimension-reduced measurement data matrix into the intermediate matrix to realize matrix combination. Specifically, formula (14) is substituted into formula (12) to obtain a combined matrix represented by formula (15).

[0106] K = Y2 · V'(S') -1 (U') * Equation (15)

[0107] Further, dimension reduction is performed on the merging matrix to obtain a dimension-reduced matrix. That is, dimension reduction is performed on the matrix K to obtain a dimension-reduced matrix which is shown in Equation (16).

[0108]

[0109] On this basis, the dimension-reduced matrix is subjected to eigenvalue decomposition to obtain a decomposition matrix shown in Equation (17):

[0110]

[0111] where W is a characteristic vector matrix of , Λ is a characteristic value matrix of , and λ i is a characteristic value.

[0112] Further, the dimension-reduced matrix is substituted into the decomposition matrix, and a reference expression can be determined in combination with a preset value. The preset value can be Y2 · V · S -1 , and the combination of the preset value means that the preset value Y2 · V · S -1 is multiplied on both sides of the equation. Specifically, Equation (16) is substituted into Equation (17), and then the preset value Y2 · V · S -1 is multiplied on both sides of the equation, to obtain a reference expression represented by Equation (18). Where Φ is a characteristic vector of K, and the corresponding characteristic value matrix is Λ.

[0113]

[0114] Since K is linear, any column in the measurement data matrix Y can be represented by the first column of the first matrix Y1 in the measurement data matrix, the characteristic value and the characteristic vector of K. Therefore, Φ shown in Equation (19) can be regarded as a mode of the measurement space. The mode of the measurement space is represented by a mode matrix of the measurement space. The mode of the measurement space can also be referred to as a model of the measurement space, and the mode of the measurement space refers to data values used to predict all state data after the mode.

[0115]

[0116] where Φ is a mode matrix of the measurement space; and φ i is a mode in the mode matrix.

[0117] Since the expression of the pattern of the measurement space contains the first undetermined parameter and the second undetermined parameter, the values of the first undetermined parameter and the second undetermined parameter can be determined, and thus the pattern of the measurement space is given.

[0118] In some embodiments, the specific process of determining the first undetermined parameter and the second undetermined parameter can include:

[0119] According to the pattern of the measurement space, the control variables are data predicted, the predicted value of the observation variable is determined, and the error between the predicted value of the observation variable and the true value of the observation variable is determined; the objective function is determined according to the error and the condition that the first undetermined parameter needs to satisfy, and the first undetermined parameter and the second undetermined parameter are determined when the objective function satisfies the iteration end condition. The condition that the first undetermined parameter needs to satisfy can be that the first undetermined parameter cannot be too large, for example, it can be set that the first undetermined parameter is less than a parameter threshold. Based on this, an enumeration search algorithm can be used to change the values of the first undetermined parameter and the second undetermined parameter, and based on each set of first undetermined parameter and second undetermined parameter, the pattern of the measurement space is calculated according to the above method, and the control variables are data predicted according to the pattern of the measurement space, and the predicted value of the observation variable corresponding thereto is obtained. The error is calculated according to the difference between the predicted value and the true value of the observation variable. Further, the objective function is designed according to the two conditions that the error is less than an error threshold and the first undetermined parameter cannot be too large (for example, less than a parameter threshold), and the first undetermined parameter and the second undetermined parameter are determined when the objective function reaches a minimum value.

[0120] In step S130, in the case that the pattern of the measurement space is correct, the basic data is obtained from the operation data of the physical system, the basic data is data predicted based on the pattern of the measurement space, and the predicted data is generated.

[0121] In the embodiments of the present disclosure, after obtaining the pattern of the measurement space, it can be judged whether the pattern of the measurement space is accurate. Exemplarily, the first undetermined parameter and the second undetermined parameter can be used to determine whether the pattern of the measurement space is accurate. If the pattern of the measurement space is accurate, it can be used for data prediction.

[0122] After the first undetermined parameter and the second undetermined parameter are determined, the mode dimension r represented by the pattern of the measurement space and the highest power p of each control variable in the model formula can be determined. In order to realize data prediction, the basic data can be obtained from the operation data of the physical system. The basic data can be data used for data prediction and model fitting, which can be real operation data. Based on this, for the general model of the non-analytical physical system in formula (2), the mode dimension of the control variable x (i) is r i , and the dimension of the basic data is

[0123] After obtaining the basic data and all the patterns, in the case that the first pending parameter is greater than the second pending parameter, the dimension of the basic data is not enough to determine the coefficients of the fitting model. At this time, the basic data can be used to generate prediction data by data prediction, so as to obtain online data for subsequent work based on the prediction data. The specific process of data prediction can include the following steps: preprocessing the basic data, measuring the preprocessed basic data by a measurement function to obtain measurement space data, extracting the patterns of the measurement space from the measurement space data, performing data prediction based on the patterns of the measurement space and the initial value of the basic data to determine the initial prediction data, returning the initial prediction data to the original space and re-determining the basic data, and performing data prediction based on the re-determined basic data and the patterns of the measurement space in a cycle until a stop condition is met to determine the prediction data. The preprocessing of the above basic data can specifically include: processing the basic data into a multi-column form, for example, into a three-column form. The preprocessed basic data is measured by a measurement function to obtain measurement space data, which can be used as a measurement data matrix. Further, the first matrix and the second matrix can be extracted from the measurement space data, the second matrix is represented by the first matrix to obtain an intermediate matrix, the first matrix is singular value decomposed, and the first matrix is dimensionally reduced based on the decomposed matrix to obtain reduced dimension measurement space data, and the reduced dimension measurement space data and the intermediate matrix are merged and eigenvalue decomposed to extract the patterns of the measurement space obtained from the basic data from the measurement space data.

[0124] Further, the initial value of the basic data can be predicted by the patterns of the measurement space, and the prediction can be the first prediction, and the result of the first prediction can be the initial prediction data. For example, the data prediction can be realized by formula (20):

[0125]

[0126] wherein, represents the first column of the measurement space data. w k = ln(λ k ) / x (i) represents the normalization of the eigenvalue, represents the initial value.

[0127] Further, the prediction data of the above measurement space is returned to the original space. Specifically, in the measurement function shown in formula (8), the first measurement function is θ1(x) = x, which is not equal to x 0 , so the first N rows of the prediction data are x (i)The data in the original space. Therefore, the first N rows in the initial predicted data are returned to the original space as the original space data, and the original space data is transposed to re-determine the basis data. Wherein, N can be represented as

[0128] After re-determining the basis data, the re-determined basis data is used as the basis data for the second prediction. The data prediction can be performed according to the re-determined basis data and the mode of the measurement space to obtain the second predicted data. The second predicted data is further used as the basis data, and the above data prediction process is repeatedly executed until the stop condition is met, so as to determine the final predicted data, and the finally obtained predicted data can be used as the online data for model fitting. Wherein, when the number of cycles reaches the number threshold, it can be considered that the stop condition is met.

[0129] Since the data prediction can be realized by the mode of the measurement space, the highest power and the minimum dimension of the mode are specified in the mode of the measurement space, so as to perform the data prediction with the least data, and further to reduce the data amount of the predicted data for constructing the model.

[0130] In step S140, the polynomial fitting model corresponding to the device belonging to the non-analytical model in the physical system is determined by using the predicted data.

[0131] In the embodiments of the present disclosure, for the non-analytical model of the physical system, the polynomial fitting model can be obtained by using a large amount of predicted data, wherein the highest power of each variable in the polynomial fitting model is determined by the highest power p represented by the first undetermined parameter in the measurement function. At the same time, the p with the smallest error in the regression function is selected as the highest power of each variable in the polynomial.

[0132] Exemplarily, since the constant term is not included in the measurement function, the highest power of the polynomial fitting model can be p+1, which is the highest power of the measurement function plus 1.

[0133] Through the above steps, the polynomial fitting model constructed by the non-analytical model described in formula (2) can be as shown in formula (21). Wherein, the coefficient fitting of the polynomial fitting model can adopt the least square method.

[0134]

[0135] In the embodiments of the present disclosure, when the physical system is an energy system, the devices contained therein can include pumps, fans, heat exchangers, cooling towers, valves, solar photovoltaic, etc. main equipment, for each device, its corresponding physical model can be determined.

[0136] Exemplarily, the model formula related to the pump and its outlet valve includes: the electric power (E P ), the rotation speed (nP ) and the outlet valve opening (o P ) can be written in the form of equation (26), which is a non-analytical model.

[0137]

[0138] In addition, the model of the outlet flow (q P ), the control variable rotation speed (n P ) and the outlet valve opening (o P ) can be written in the form of equation (27), which is a non-analytical model.

[0139]

[0140] The model formula related to the pump and its outlet valve can also include the model of the outlet flow (H P ), the rotation speed (n P ) and the outlet valve opening (o P ), which can be written in the form of equation (28), which is also a non-analytical model.

[0141]

[0142] The pump converts electrical energy into water energy, and its rotation speed is controlled by the frequency converter, thereby adjusting the rotation speed of the motor. In this energy system, the rated rotation speed of the pump is 3000 RPM. From the variable relationship, the pump and its outlet valve realize the coupling of variables such as electrical energy, flow, and head.

[0143] The model formula related to the heat exchanger includes the model of the working medium output temperature (To n ), the circulating water input temperature (ti) and the circulating water flow (q Hn ), and the indirect control can be written in the form of equation (29), which is a non-analytical model.

[0144]

[0145] The model formula related to the heat exchanger can also include the model of the circulating water output temperature (to n ), the circulating water input temperature (ti) and the circulating water flow (q Hn ), which can be written in the form of equation (30), which is also a non-analytical model.

[0146]

[0147] wherein Ti n and Q n are parameters that will change during the operation of the system.

[0148] The heat exchanger exchanges the heat energy in the working medium with the cold energy in the circulating water. The tube-shell heat exchanger used in EH has the characteristics of large heat transfer area per unit device volume, good heat transfer effect, and solid structure. From the variable relationship, the heat exchanger is a complex system, which couples variables such as flow, temperature, resistance, and material.

[0149] The fan-related model formula can include the model of electric power (E F ) and rotational speed (n F ), i.e., the direct control model shown in equation (31). Since the fan has a detailed manufacturer's manual, this model has an analytical formula, as shown in equation (32).

[0150]

[0151] In the formula, n FN is the rated rotational speed, and E FN is the rated electric power.

[0152] The fan-related model formula can also include the model of air volume (G) and rotational speed (n F ), i.e., the direct control model shown in equation (33), and its analytical formula is shown in equation (34).

[0153] G = f G (n F ) equation (33)

[0154]

[0155] In the formula, G N is the rated air volume.

[0156] The fan is a device that converts electrical energy into cold energy. The air volume provided by the fan cools the circulating water in the cooling tower. The fan used in our EH has a rated rotational speed of 3000 RPM, a rated power of 1600000 W, and a rated air volume of 236.1 m3 / s.

[0157] The construction process of the cooling tower-related model formula is as follows: In order to facilitate modeling, in the cooling tower study, an intermediate variable, i.e., the gas-water ratio (GWr), is usually introduced, which can be calculated by equation (35).

[0158]

[0159] Where y a is the air volume density, as shown in equation (36), and other parameters are shown in equations (37) and (40), respectively.

[0160]

[0161] Where Pa represents the atmospheric pressure, which is measured as 104,000 Pa at the local value.

[0162] P = 98066.5 · 10 E Formula (38)

[0163]

[0164] wherein a = 0.0141966, b = 3142.305, c = 8.2, d = 0.0024802.

[0165] T = t + 273.16 Formula (40)

[0166] wherein T is the temperature in Kelvin.

[0167] The difference between the lower tower temperature (tT) and the upper tower temperature (to) is recorded as the temperature difference (Dt), as shown in Formula (41).

[0168] Dt = to - tT Formula (41)

[0169] Thus, the model of the gas-water ratio (GWr), the lower tower temperature (tT), and the temperature difference (Dt) can be written in the form of Formula (42), which is a non-analytical model.

[0170] GWr = f GWr (tT, Dt; ts, tg) Formula (42)

[0171] The cooling tower is provided with air volume by the above-mentioned fan, and the heat energy of circulating water is discharged into the air to complete the conversion of cold and heat energy.

[0172] There is a mechanical draft cooling tower in the power plant, with a height of 9 m. Considering the safety of the system, there is a 1 m margin, so the water head borne by the cooling tower is 10 m.

[0173] The construction process of the pipe network related model formula includes: the pipe network flow satisfies the mass conservation law, and the analytical formula as shown in Formula (43) can be obtained, wherein the control valve model satisfies the analytical equations as shown in Formula (44) and Formula (45).

[0174]

[0175]

[0176]

[0177] wherein k n represents the coefficient related to the branch control valve and the experimentally measured resistance k1 = 3.393e-6 and k2 = 1.433e-5, and H Net is the pipe network pressure loss.

[0178] The pipe network also satisfies the law of conservation of heat, and thus another analytical model as shown in equation (46) can be obtained.

[0179]

[0180] The pipe network of the system consists of one main trunk and two parallel branches.

[0181] The model formula related to the pool can be expressed as:

[0182] tT= ti equation (47)

[0183] The pool collects the circulating water cooled by the cooling tower and supplies it to the electric heating system. It is assumed that the pool only serves as a connection with the circulating water, and the changes in the temperature under the tower and the inlet temperature of the circulating water are not considered.

[0184] The model formula related to the total pressure balance of the electric heating system can be as follows:

[0185] H P -H Net ≥H T equation (48)

[0186] The electric heating system should meet the total pressure requirement; that is, the head provided by the pump should meet the pressure requirement of the circulating water tower after the pressure loss of the pipe network.

[0187] After the model of each device is determined, the non-analytical model can be screened from the model of each device. Further, the mode of the measurement space is determined based on the non-analytical model, and the parameters of the control variables of each device are determined using a machine learning model algorithm.

[0188] Exemplarily, the screened non-analytical model includes the model of the pump, the model of the heat exchanger, and the model of the cooling tower.

[0189] Figure 4 The training results of the measurement variables of the pump are schematically shown, wherein the measurement variables of the pump include E P , q P , H P , and the control variables are n P , o P . Figure 5 The training results of the measurement variables of the heat exchanger are schematically shown. The measurement variables of the heat exchanger include To n , to n , and the control variables include ti, q Hn . Figure 6 The training results of the measurement variables of the cooling tower are schematically shown. The measurement variables of the cooling tower include GWr, and the control variables include tT, Dt. Referring to Figures 4-6The training results of the measurement variables of the equipment corresponding to the non-analytical model shown in the middle of the figure represent the size of the prediction error when the mode extraction dimension r and the highest power p of the measurement function are different, that is, the value of the generation in the machine learning algorithm under different parameters. Figures 4-6 The results of learning determination of each measurement variable are represented in each wireframe in the middle of the figure. The machine learning results of each measurement variable are shown in Table 2, which shows the numerical values of the first undetermined parameter and the second undetermined parameter, the dimension of the basic data, and the power of the polynomial fitting model.

[0190] Table 2

[0191]

[0192] The fitting model parameters obtained by the learning results can be used to fit the coefficients of each control variable by least squares method, and then the polynomial fitting model can be determined by combining the control variables, the coefficients and the highest power and other parameters, so that the polynomial fitting model corresponding to the equipment with a non-analytical model in the energy system can be obtained.

[0193] When determining the polynomial fitting model by the prediction data, the prediction data can be fitted by using the polynomial regression model, and the error between the prediction data and the reference data corresponding to the prediction data can be calculated, and the error can be the mean square error. The reference data refers to the real data corresponding to the prediction data. When the error is the smallest, the parameter corresponding to the smallest error can be determined as the model parameter of the polynomial regression model. If the model is under-fitted, the order of the polynomial may need to be increased; if it is over-fitted, the order may need to be reduced or regularization may be introduced. The obtained polynomial fitting model can be used to predict the performance or output value of the equipment in the future time period. At the same time, the polynomial fitting model can also be used to analyze the trend or abnormal situation in the operation of the equipment.

[0194] The polynomial fitting model can be represented as shown in formula (49)-formula (56).

[0195]

[0196] Figure 7 The flowchart for determining the polynomial fitting model based on the machine learning model is shown in the middle of the figure, and the reference Figure 7As shown in the figure, offline data can be obtained, which refers to basic data, which can be the running data of each device in the physical system. Further, the basic data can be preprocessed in the first stage. The preprocessing here can be to process the basic data into a multi-column form, for example, into a three-column form to obtain standard offline data. Further, the second stage of Koopman operation can be performed, the preprocessed basic data is measured based on the measurement function, the measurement space data is obtained, and the first undetermined parameter p is set in the measurement function, which is the highest power of the control variable. The standard offline data is processed into a high-dimensional space. In the third stage of mode extraction, the dimension of the model can be reduced, and the reduced dimension is taken as the second undetermined parameter r, so as to extract the mode of the measurement space according to the first undetermined parameter and the second undetermined parameter. In the fourth stage, the values of the first undetermined parameter and the second undetermined parameter can be adjusted, and the mode of the measurement space is calculated based on the adjusted first undetermined parameter and the second undetermined parameter, and the control variable is predicted according to the mode of the measurement space, to obtain the predicted value of the observation variable corresponding thereto. The error is calculated according to the difference between the predicted value and the true value of the observation variable. Further, the target function is designed according to the two conditions that the error is less than the error threshold and the first undetermined parameter cannot be too large (for example, less than the parameter threshold), and when the target function reaches the minimum value, the values of the first undetermined parameter and the second undetermined parameter can be determined.

[0197] After obtaining the first undetermined parameter and the second undetermined parameter, the model function can be determined, and the fifth stage is further performed to determine the polynomial fitting model of each device with a non-analytic model according to the model function.

[0198] In the embodiments of the present disclosure, a modeling method suitable for a complex physical system is proposed, which can construct an explicit mathematical model with physical interpretability through experimental data and running data in the case of non-analytic modeling of part of the variables, and effectively model the behavior of the physical system. By analyzing the influence of the control variable on the observation variable in the experimental data and the running data of the device, the minimum dimension and the maximum power of each model are determined, so as to perform data prediction with the least data. Finally, the analytical formula of the model is fitted.

[0199] After obtaining the polynomial fitting model of the device with a non-analytic model in the physical system, for these devices, the observation variable corresponding to the control variable of the device can be predicted according to the polynomial fitting model, and the control and modification of the control variable can be realized based on the observation variable, which can accurately determine the control direction and control degree of the control variable, accurately control the state of the device, and improve the accuracy and convenience of device control.

[0200] The technical solution provided in the embodiments of the present disclosure can measure the control variables in the non-analytical model by selecting a measurement function to obtain a measurement data matrix, obtain the mode of the measurement space according to the measurement data matrix, predict the basic data of each device in the physical system based on the mode of the measurement space to obtain predicted data, and then determine the polynomial fitting model of each device according to the predicted data, thereby accurately establishing the model of the device and reducing the modeling cost. In addition, since the polynomial fitting model is constructed based on the predicted data determined based on the mode of the measurement space, the amount of data required for model construction is reduced, the interpretability of model construction is improved, and the accuracy of device control is improved.

[0201] In the embodiments of the present disclosure, a machine learning modeling device for a non-analytical physical system is also provided, which is shown in Figure 8 The machine learning modeling device 800 for a non-analytical physical system includes:

[0202] A non-analytical model determination module 801 is configured to obtain a non-analytical model of a physical system and determine the control variables and observation variables in the non-analytical model.

[0203] A mode extraction module 802 is configured to select a measurement function describing the relationship between the control variables and the observation variables, measure the control variables by using the measurement function to determine a measurement data matrix, and extract a mode based on the measurement data matrix to obtain the mode of the measurement space.

[0204] A predicted data generation module 803 is configured to obtain basic data from the running data of the physical system in the case where the mode of the measurement space is correct, predict the basic data based on the mode of the measurement space to generate predicted data.

[0205] A model construction module 804 is configured to determine a polynomial fitting model corresponding to the device belonging to the non-analytical model in the physical system by using the predicted data.

[0206] In an exemplary embodiment of the present disclosure, the mode extraction based on the measurement data matrix to obtain the mode of the measurement space includes:

[0207] A first matrix and a second matrix are extracted from the measurement data matrix, and the second matrix is represented by the first matrix to obtain an intermediate matrix.

[0208] The first matrix is singular value decomposed, and the first matrix is reduced in dimension based on the decomposed matrix to obtain a reduced dimension measurement data matrix.

[0209] The reduced dimension measurement data matrix and the intermediate matrix are merged, and eigenvalue decomposition is performed to obtain the mode of the measurement space.

[0210] In an example embodiment of the present disclosure, the merging the reduced dimension measurement data matrix with the intermediate matrix and performing eigen decomposition to obtain the mode of the measurement space comprises:

[0211] The merging the reduced dimension measurement data matrix with the intermediate matrix generates a merged matrix K;

[0212] The dimension reduction of the merged matrix generates a reduced dimension matrix, and the eigen decomposition of the reduced dimension matrix determines a decomposition matrix;

[0213] The reduced dimension matrix is substituted into the eigen decomposition matrix, and a reference expression is determined in combination with a preset value;

[0214] Based on the first column of the first matrix, the eigenvalue and the eigenvector of the merged matrix representing any column of the measurement data matrix, the mode of the measurement space is determined.

[0215] In an example embodiment of the present disclosure, the measurement function comprises a first undetermined parameter and a second undetermined parameter, the first undetermined parameter is the highest power of the measurement function, and the second undetermined parameter is the dimension of the mode extraction.

[0216] In an example embodiment of the present disclosure, the determination process of the first undetermined parameter and the second undetermined parameter comprises:

[0217] According to the mode of the measurement space, data prediction is performed on the control variable, the predicted value of the observation variable corresponding to the control variable is determined, and the error between the predicted value and the true value of the observation variable is determined.

[0218] According to the error and the first undetermined parameter, a target function is determined, and when the target function satisfies a condition, the first undetermined parameter and the second undetermined parameter are determined.

[0219] In an example embodiment of the present disclosure, the data prediction on the basis data based on the mode of the measurement space generates predicted data, comprising:

[0220] When the first undetermined parameter is greater than the second undetermined parameter, the basis data is obtained; the basis data is the running data of each device in the physical system;

[0221] Based on the mode of the measurement space and the initial value of the basis data, data prediction is performed to determine initial predicted data;

[0222] The initial predicted data is returned to the original space, and the basis data is re-determined, and the data prediction is circularly performed based on the re-determined basis data and the mode of the measurement space until a stop condition is satisfied, so as to determine the predicted data.

[0223] In an example embodiment of the present disclosure, the re-determined basis data comprises:

[0224] Part of the initial prediction data is taken as original space data, and the original space data is transposed to re-determine the basis data.

[0225] In an example embodiment of the present disclosure, the physical system is an energy system, and the control variable is a variable for a target device in the energy system, the target device at least including one of a pump, a heat exchanger, and a cooling tower.

[0226] It should be noted that the specific details of each module in the machine learning modeling device for the non-analytical physical system have been described in detail in the corresponding machine learning modeling method for the non-analytical physical system, which will not be repeated here.

[0227] It should be noted that although several modules or units of the device for action execution are mentioned in the foregoing detailed description, such division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided into several modules or units.

[0228] In addition, although the steps of the method in the present disclosure are described in a specific order in the accompanying drawings, this does not require or imply that the steps must be performed in this specific order, or that all the steps shown must be performed to achieve the desired result. In addition or alternatively, some steps can be omitted, a plurality of steps can be combined into one step, and / or one step can be divided into a plurality of steps, etc.

[0229] In an example embodiment of the present disclosure, an electronic device capable of implementing the above method is also provided.

[0230] Those skilled in the art can understand that various aspects of the present disclosure can be implemented as a system, a method or a program product. Therefore, various aspects of the present disclosure can be embodied as a complete hardware embodiment, a complete software embodiment (including firmware, microcode, etc.), or an embodiment combining hardware and software aspects, which can be collectively referred to as "circuitry", "module" or "system" here.

[0231] The electronic device 900 according to this embodiment of the present disclosure will be described below with reference to Figure 9 Figure 9 The displayed electronic device 900 is only an example and should not impose any limitation on the functions and use range of the embodiments of the present disclosure.

[0232] ​like Figure 9 As shown, the electronic device 900 is manifested in the form of a general-purpose computing device. The components of the electronic device 900 may include, but are not limited to: at least one processing unit 910, at least one storage unit 920, a bus 930 connecting different system components (including storage unit 920 and processing unit 910), and a display unit 940.

[0233] The storage unit stores program code that can be executed by the processing unit 910, causing the processing unit 910 to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of this disclosure. For example, the processing unit 910 can perform actions such as... Figure 1 The steps are shown in the figure.

[0234] Storage unit 920 may include readable media in the form of volatile storage units, such as random access memory (RAM) 9201 and / or cache memory 9202, and may further include read-only memory (ROM) 9203.

[0235] Storage unit 920 may also include a program / utility 9204 having a set (at least one) program module 9205, such program module 9205 including but not limited to: operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0236] Bus 930 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0237] Electronic device 900 can also communicate with one or more external devices 1000 (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with the electronic device 900, and / or with any device that enables the electronic device 900 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed through input / output (I / O) interface 950. Furthermore, electronic device 900 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 960. Figure 9As shown, network adapter 960 communicates with the other components of electronic device 900 via bus 930. It should be understood that although not shown, other hardware and / or software components could be used in conjunction with electronic device 900. These include, but are not limited to, microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data archival storage systems, etc.

[0238] It should be noted that some embodiments of the present disclosure further provide a computer program product. The computer program product includes a computer program. When the computer program is executed by a processor, the above-described method is implemented.

[0239] In an embodiment, the computer program product can be a tangible product containing the computer program, such as a computer-readable storage medium storing the computer program. The readable storage medium can be a storage medium based on electric, magnetic, optical, electromagnetic, infrared, etc. signals, including but not limited to random access memory (RAM), read-only memory (ROM), magnetic tape, floppy disk, flash memory (Flash), mechanical hard disk (HDD), solid state disk (SSD), etc. For example, the computer program product can be implemented as a non-volatile storage medium storing the computer program, such as read-only memory, Nand flash memory, etc. In an embodiment, the computer program product can be an intangible product containing the computer program. For example, the computer program product can be implemented as a virtual digital product, such as an executable file, an installation package, etc. digital file storing the computer program.

[0240] The code of the computer program can be written in one or more programming languages. Programming languages such as C, Java, C++, etc. The program code can be executed entirely on a user computing device, or partially on a user computing device, or as a separate software package, or partially on a user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device can be connected to the user computing device through any kind of network, such as a local area network (LAN), a wide area network (WAN), etc., or can be connected to an external computing device (for example, through an Internet connection provided by an operator).

[0241] The computer program can be carried or transmitted by electronic, magnetic, optical, electromagnetic, infrared, etc. signals. The electronic device can convert the signal carrying the computer program into a digital signal, and then run the computer program. When the computer program is running on the electronic device, its code is used to make the electronic device execute (more specifically, can make the processor of the electronic device execute) the method steps of various exemplary embodiments of the present disclosure.

[0242] Through the above description of the embodiments, those skilled in the art can easily understand that the example embodiments described herein can be implemented by software, or by software in combination with necessary hardware. Therefore, the technical solutions according to the embodiments of the present disclosure can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash disk, a mobile hard disk, or the like) or a network, and includes a number of instructions to make a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) execute the methods according to the embodiments of the present disclosure.

[0243] In addition, the above-described drawings are only schematic illustrations of the processes included in the method according to the exemplary embodiments of the present disclosure, and are not intended for limiting purposes. It is easy to understand that the processes shown in the above-described drawings do not indicate or limit the time sequence of the processes. In addition, it is also easy to understand that the processes can be executed synchronously or asynchronously, for example, in a plurality of modules.

[0244] It should be noted that, although several modules or units of the device for action execution are mentioned in the above detailed description, such a division is not mandatory. Indeed, according to the embodiments of the present disclosure, the features and functionalities of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functionalities of one module or unit described above can be further divided into embodied by a plurality of modules or units.

[0245] Other embodiments of the present disclosure will be apparent to those skilled in the art upon consideration of the specification and the practice of the present disclosure. The present application is intended to cover any variations, uses, or adaptations of the present disclosure following the general principles thereof and including such departures from the present disclosure that come within known, accepted, or customary practice in the art to which the present disclosure pertains. The specification and examples are to be regarded as exemplary only, and the true scope and spirit of the present disclosure are indicated by the claims.

[0246] It should be understood that the present disclosure is not limited to the precise structures described and shown in the accompanying drawings, and that various modifications and changes can be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A machine learning modeling method for non-analytical physical systems, characterized in that, include: Obtain a non-analytical model of the physical system and determine the control variables and observation variables in the non-analytical model; A measurement function representing the relationship between control variables and observed variables is selected. The control variables are measured using the measurement function to determine the measurement data matrix. Pattern extraction is then performed based on the measurement data matrix to obtain the pattern of the measurement space. If the pattern of the measurement space is correct, basic data is obtained from the operating data of the physical system, and data prediction is performed on the basic data based on the pattern of the measurement space to generate predicted data. Using the predicted data, the polynomial fitting model corresponding to the device in the physical system that belongs to the non-analytical model is determined.

2. The machine learning modeling method for non-analytical physical systems according to claim 1, characterized in that, The process of extracting patterns from the measurement data matrix to obtain patterns in the measurement space includes: Extract the first matrix and the second matrix from the measurement data matrix, and represent the second matrix using the first matrix to obtain the intermediate matrix; Singular value decomposition is performed on the first matrix, and the dimension reduction of the first matrix is ​​performed based on the decomposition matrix to obtain a dimension-reduced measurement data matrix; The reduced-dimensional measurement data matrix is ​​merged with the intermediate matrix, and feature decomposition is performed to obtain the pattern of the measurement space.

3. The machine learning modeling method for non-analytical physical systems according to claim 2, characterized in that, The step of merging the dimensionality-reduced measurement data matrix with the intermediate matrix and performing feature decomposition to obtain the pattern of the measurement space includes: The dimensionality-reduced measurement data matrix is ​​merged with the intermediate matrix to generate a merged matrix K; The dimensionality of the merged matrix is ​​reduced to obtain a dimensionality-reduced matrix, and the dimensionality-reduced matrix is ​​then subjected to eigenvalue decomposition to determine the decomposition matrix. Substitute the dimensionality reduction matrix into the eigenvalue decomposition matrix and combine it with preset values ​​to determine the reference expression; The pattern of the measurement space is determined based on the first column of the first matrix, the eigenvalues ​​of the merged matrix, and the eigenvectors representing any column of the measurement data matrix.

4. The machine learning modeling method for non-analytical physical systems according to claim 1, characterized in that, The measurement function includes a first undetermined parameter and a second undetermined parameter. The first undetermined parameter is the highest power of the measurement function, and the second undetermined parameter is the dimension of the modality extraction.

5. The machine learning modeling method for non-analytical physical systems according to claim 4, characterized in that, The process of determining the first undetermined parameter and the second undetermined parameter includes: Based on the pattern of the measurement space, data prediction is performed on the control variables to determine the predicted values ​​of the observed variables corresponding to the control variables, and the error between the predicted values ​​and the actual values ​​of the observed variables is determined. The objective function is determined based on the error and the first undetermined parameter, and the first undetermined parameter and the second undetermined parameter are determined when the objective function satisfies the conditions.

6. The machine learning modeling method for non-analytical physical systems according to claim 1, characterized in that, The process of predicting the basic data based on the pattern of the measurement space to generate predicted data includes: When the first undetermined parameter is greater than the second undetermined parameter, basic data is obtained; the basic data is the operating data of each device in the physical system. Data prediction is performed based on the pattern of the measurement space and the initial values ​​of the basic data to determine the initial prediction data; The initial prediction data is returned to the original space, and the basic data is redefined. Based on the redefined basic data and the pattern of the measurement space, data prediction is performed cyclically until the stopping condition is met, so as to determine the prediction data.

7. The machine learning modeling method for non-analytical physical systems according to claim 1, characterized in that, The re-determination of basic data includes: A portion of the initial prediction data is used as the original spatial data, and the original spatial data is transposed to redetermine the underlying data.

8. The machine learning modeling method for non-analytical physical systems according to claim 1, characterized in that, The physical system is an energy system, and the control variables are variables for target devices in the energy system. The target devices include at least one of pumps, heat exchangers, and cooling towers.

9. A computer program product, characterized in that, When the computer program is executed by the processor, it implements the machine learning modeling method for the non-analytical physical system as described in any one of claims 1-8.

10. An electronic device, characterized in that, include: processor; Memory for storing the executable instructions of the processor; The processor is configured to implement the machine learning modeling method for the non-analytical physical system according to any one of claims 1-8 by executing the executable instructions.

Images