A method for modeling shock loads by fusing dual modal decomposition and physical information

Abstract

This invention discloses a method for modeling impact loads by integrating dual-mode decomposition and physical information, relating to the field of load forecasting technology. The method includes constructing a training set of historical time-series data of impact loads and acquiring physical information related to the formation mechanism of the impact loads; performing a first-order mode decomposition on the historical time-series data, and analyzing and recombining the results of the first-order mode decomposition based on a complexity evaluation index; this invention improves the accuracy of impact feature extraction through dual-mode decomposition and classification and recombination based on complexity indices; employing differentiated modeling to adapt to the multi-scale variation characteristics of the load, enhancing the model's fitting and generalization capabilities; constructing a penalty quantity based on the fusion results of each model output and physical information, and updating the high-frequency component model parameters based on the penalty quantity to make the model output more closely match the load operation mechanism, reducing deviations and lags during operating condition switching and load abrupt changes, and improving the ability to characterize the transient characteristics of impact loads.

Description

Technical Field

[0001] This invention relates to the field of load forecasting technology, and in particular to a method for modeling impact loads that integrates dual-mode decomposition and physical information. Background Technology

[0002] In industrial production scenarios, equipment such as electric arc furnaces, rolling mills, and welding machines constitute typical impact load sources. These loads are characterized by severe power fluctuations and significant impact and intermittent characteristics. High-precision modeling of these loads is a key technical support for industrial enterprises to achieve refined energy scheduling and improve the operational stability of the power supply system. It has important application value for ensuring industrial power supply safety and efficient energy use.

[0003] Existing methods for modeling impact loads mainly rely on single time-series models and shallow machine learning models. Some methods use single-mode decomposition to process load sequences and rely on a single model to complete the overall load fitting. They do not build adaptive modeling methods for different components, and the modeling process is mainly driven by pure data without incorporating information on the physical mechanism of load operation into the model training.

[0004] Existing methods using single-mode decomposition struggle to distinguish between strong impact features and high-frequency random noise, easily leading to mode aliasing and reduced impact feature extraction accuracy. Single models cannot adapt to multi-scale load variations, and purely data-driven approaches lack physical constraints. Furthermore, they are prone to fitting biases and response lags during frequent load changes and sudden load shifts, failing to accurately characterize the instantaneous transient characteristics of the load. Therefore, existing technologies suffer from technical problems such as easy mode aliasing, mismatched modeling methods, and a lack of physical mechanisms. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an impact load modeling method that integrates dual modal decomposition and physical information, thereby solving the technical problems of easy aliasing of modal decomposition, mismatch of modeling methods, and lack of physical mechanisms in existing technologies.

[0006] The technical means employed in this invention are as follows:

[0007] In a first aspect, embodiments of the present invention provide a method for modeling impact loads that integrates dual-modal decomposition and physical information, including: Construct a training set of historical time-series data on impact loads to obtain physical information related to the formation mechanism of the impact loads; The historical time series data is subjected to a first mode decomposition. The results of the first mode decomposition are analyzed and reorganized based on the complexity evaluation index to obtain trend components, oscillation components and high-frequency components of the first mode decomposition. The high-frequency components of the first mode decomposition are subjected to second mode decomposition to obtain the high-frequency components of the second mode decomposition. Differential modeling is performed on the trend component, the oscillation component, and the high-frequency component of the second-order mode decomposition to obtain the trend component model, the oscillation component model, and the high-frequency component model. The physical information is introduced and a joint loss function is constructed during the training process of the trend component model, the oscillation component model, and the high-frequency component model. Based on the deviation between the fusion results of the trend component model, the oscillation component model, and the high-frequency component model and the physical information, a penalty amount is constructed. The penalty amount is then incorporated as a physical correction term into the backpropagation process of the joint loss function to update the parameters of the high-frequency component model, thereby obtaining the corrected output of the high-frequency component model. The output of the trend component model, the output of the oscillation component model, and the corrected output of the high-frequency component model are fused to obtain the model calculation value of the impact load.

[0008] Furthermore, the physical information includes time-series variables related to the operating status of the equipment.

[0009] Furthermore, the step of performing a mode decomposition on the historical time series data, and analyzing and reorganizing the results of the mode decomposition based on a complexity evaluation index to obtain trend components, oscillation components, and high-frequency components, includes: performing a mode decomposition on the historical time series data using an adaptive noise complete set empirical mode decomposition method to obtain multiple mode components, calculating the fuzzy entropy of the multiple mode components, and classifying and reorganizing the multiple mode components according to the magnitude of the fuzzy entropy to obtain trend components, oscillation components, and high-frequency components.

[0010] Furthermore, the step of performing secondary mode decomposition on the high-frequency components includes: performing secondary mode decomposition on the high-frequency components using variational mode decomposition.

[0011] Furthermore, the differentiated modeling of the trend component, the oscillation component, and the high-frequency component includes: modeling the trend component and the oscillation component using an ensemble learning regression model, and modeling the high-frequency component using a neural network model.

[0012] Furthermore, the step of introducing the physical information and constructing a joint loss function during the training process of the trend component model, the oscillation component model, and the high-frequency component model includes: introducing the physical information and constructing a joint loss function based on the data fitting loss term and the physical information loss term during the training process of the trend component model, the oscillation component model, and the high-frequency component model.

[0013] Furthermore, the formula for calculating the penalty amount is as follows:

[0014] in, express The penalty amount corresponding to each moment; express The weighting coefficient corresponding to each time point; express The deviation corresponding to the time; The adjustment coefficient representing the deviation derivative term; Indicates the rate of change of deviation; This represents the i-th sampling time.

[0015] Furthermore, the calculation formula for the model value of the impact load is as follows:

[0016] in, express The calculated values ​​of the impact load model at any given time; express The trend component model output at time step; express The output of the oscillation component model at time step; express The model correction output of the high-frequency component of the second-order mode decomposition at time m; M represents the number of high-frequency components in the second-order mode decomposition. This represents the i-th sampling time.

[0017] Secondly, embodiments of the present invention also provide an impact load modeling system that integrates dual-modal decomposition and physical information, comprising: The data acquisition module is used to construct a training set of historical time-series data of impact loads and acquire physical information related to the formation mechanism of the impact loads; The first mode decomposition module is used to perform first mode decomposition on the historical time series data, and analyze and reorganize the first mode decomposition results based on the complexity evaluation index to obtain trend components, oscillation components and high-frequency components of the first mode decomposition. The secondary mode decomposition module is used to perform secondary mode decomposition on the high-frequency components of the primary mode decomposition to obtain the secondary mode decomposition high-frequency components. The differential modeling module is used to perform differential modeling on the trend component, the oscillation component, and the high-frequency component of the second-mode decomposition to obtain the trend component model, the oscillation component model, and the high-frequency component model. The physical information is introduced and a joint loss function is constructed during the training process of the trend component model, the oscillation component model, and the high-frequency component model. The model correction module is used to construct a penalty amount based on the deviation between the fusion result of the trend component model, the oscillation component model and the high-frequency component model and the physical information, and to incorporate the penalty amount as a physical correction term into the backpropagation process of the joint loss function to update the parameters of the high-frequency component model and obtain the corrected output of the high-frequency component model. The result fusion module is used to fuse the output of the trend component model, the output of the oscillation component model, and the corrected output of the high-frequency component model to obtain the model calculation value of the impact load.

[0018] Compared with the prior art, the present invention has the following advantages: This invention effectively reduces mode aliasing and improves the accuracy of impact feature extraction by using dual mode decomposition and complexity index classification and recombination. It adopts differentiated modeling to adapt to the multi-scale variation characteristics of the load, enhancing the model fitting and generalization capabilities. Based on the deviation between the fusion results of each model output and the physical information, a penalty amount is constructed and incorporated as a physical correction term into the backpropagation process of the joint loss function to update the parameters of the high-frequency component model. This makes the model output more consistent with the load operation mechanism, reduces the deviation and lag during operating condition switching and load abrupt changes, improves the ability to characterize the transient characteristics of impact loads, and helps to improve the scheduling and operation stability of industrial power supply systems.

[0019] Based on the above reasons, this invention can be widely applied in fields such as load forecasting. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the impact load modeling method that integrates dual-modal decomposition and physical information according to the present invention. Figure 2 This is a diagram showing the result of a first mode decomposition of the original sequence of the rolling mill electrical load in an embodiment of the present invention; Figure 3 This is a diagram showing the fuzzy entropy results calculated from the intrinsic mode function components obtained by a single decomposition in an embodiment of the present invention. Figure 4 This is a diagram showing the results of analyzing and recombining the modal components based on fuzzy entropy in an embodiment of the present invention. Figure 5This is a diagram showing the modeling result of the rolling mill electrical load after integrating dual-mode decomposition and physical information in an embodiment of the present invention. Detailed Implementation

[0022] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0023] It should be noted that the terms "comprising" and "having" and any variations thereof in this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or device.

[0024] The electrical load of a rolling mill is only a typical example of an impact load. The technical solution of this invention is also applicable to other impact loads with similar characteristics, including but not limited to electric arc furnaces and welding machines. Those skilled in the art can make adaptive adjustments to the parameters of this invention according to different types of impact loads, and all such adjustments should fall within the protection scope of this invention.

[0025] The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0026] Please see Figure 1 , Figure 1 This is a flowchart illustrating the impact load modeling method that integrates dual-modal decomposition and physical information according to the present invention.

[0027] This application provides a method for modeling impact loads that integrates dual-modal decomposition and physical information, including the following steps: Step 101: Construct a historical time-series data training set for impact loads to obtain physical information related to the formation mechanism of impact loads. This involves preprocessing the historical impact load data (including missing value imputation and outlier handling) to construct a regular historical time-series data training set; obtaining time-series data of physical information related to the formation mechanism of impact loads to construct a physical information dataset; the historical time-series data training set and the physical information dataset are strictly aligned in the time dimension, meaning that for any given moment, there exists corresponding historical time-series data and physical information to reflect the physical characteristics of the load data at any given moment.

[0028] In some embodiments, the impact load includes, but is not limited to, the rolling mill electrical load and the electric arc furnace electrical load, and the physical information includes time-series variables related to the equipment operating status. The specific representation quantities of the physical information differ for different impact loads. For example, when the modeling object is the rolling mill electrical load, the physical information may include rolling force and main motor speed; when the modeling object is the electric arc furnace electrical load, the physical information may include electrode current, arc voltage, and electrode insertion depth.

[0029] The expression for the training set of historical time-series data on impact loads is:

[0030] in, A training set representing historical time-series data of shock loads; This represents the k-th time series sample obtained by dividing the original shock load sequence through a sliding time window; k represents the sample index; and N represents the total number of time series samples contained in the training set.

[0031] The expression for the physical information dataset is:

[0032] in, A dataset representing physical information related to the formation mechanism of impact loads; This represents the physical information time series data sample corresponding to the k-th time series sample of the impact load; k represents the sample index; N represents the total number of time series samples.

[0033] Step 102: Perform a first-order mode decomposition on the historical time-series data. Analyze and reorganize the first-order mode decomposition results based on a complexity evaluation index to obtain trend components, oscillation components, and high-frequency components of the first-order mode decomposition. This achieves multi-scale feature separation of historical time-series data of impact loads, improves the distinguishability of different dynamic behaviors of impact loads, and can clearly characterize the long-term trend of load changes, periodic oscillations, and instantaneous impact characteristics.

[0034] In some embodiments, a first mode decomposition is performed on historical time-series data, and the results of the first mode decomposition are analyzed and recombined based on a complexity evaluation index to obtain trend components, oscillation components, and high-frequency components. This includes: performing a first mode decomposition on historical time-series data using an adaptive noise complete set empirical mode decomposition method to obtain multiple mode components; effectively suppressing mode aliasing, enabling the decomposed mode components to reflect load change characteristics at different frequency scales; calculating the fuzzy entropy of multiple mode components, and classifying and recombining the multiple mode components according to the magnitude of the fuzzy entropy to obtain trend components, oscillation components, and high-frequency components; and reasonably dividing mode components of different complexities so that each recombined component has a clear physical meaning and clear dynamic behavior characteristics.

[0035] Specifically, the adaptive noise complete set empirical mode decomposition (CEEMDAN) method is used to perform a first-order mode decomposition on the historical time series data, so that the original signal is represented as the sum of several intrinsic mode functions and residual terms. The expression for the first-order mode decomposition is:

[0036] in, The original historical time-series data signal represents the impact load; J represents the number of intrinsic mode functions (IMFs); j represents the modal component index; This represents the j-th eigenmode function component; This represents the residual term.

[0037] The fuzzy entropy is calculated for each intrinsic mode function component to characterize its complexity. Fuzzy entropy can characterize the randomness and irregularity of a time series. The larger the value, the higher the complexity of the series.

[0038] The specific calculation process of fuzzy entropy is as follows: reconstruct the phase space of each intrinsic mode function component, and let... Then, the original modal component sequence with length (number of sampling points) L is... Construct an s-dimensional vector sequence:

[0039] in, This represents the v-th s-dimensional phase space vector after reconstruction; represents the intrinsic mode function component value corresponding to the v-th sampling point; s is the embedding dimension of the phase space reconstruction; represents the average value of the s-dimensional vector; v represents the sampling point index; L represents the total number of sampling points in the original modal component sequence.

[0040] s The formula for calculating the average value of a dimensional vector is:

[0041] in, represents the average value of the s-dimensional vector; s represents the embedding dimension of the phase space reconstruction; v represents the sampling point index of the modal component sequence; Represents the variable to be summed; Represents the first modal component in the modal component sequence The component values ​​corresponding to each sampling point.

[0042] Calculate the fuzzy similarity between each s-dimensional phase space vector within the same modal component, and obtain the average similarity functions for embedding dimensions s and s+1, respectively. and (r), thus calculating the fuzzy entropy. The formula for calculating fuzzy entropy is:

[0043] in, represents the fuzzy entropy value of the sequence; s represents the embedding dimension of the phase space reconstruction; r represents the tolerance threshold for similarity calculation; L represents the total number of sampling points of the original modal component sequence; This represents the average similarity function when the embedding dimension is s; (r) represents the average similarity function when the embedding dimension is s+1.

[0044] Based on the magnitude of fuzzy entropy, the modal components are classified and reorganized. Low-complexity components are classified as trend components, medium-complexity components as oscillation components, and high-complexity components as high-frequency components. Among them, high-frequency components contain significant impact characteristics and random noise components.

[0045] Specifically, a low-complexity threshold and a high-complexity threshold are set for fuzzy entropy. Modal components with fuzzy entropy values ​​below the low-complexity threshold are classified as low-complexity components, serving as trend components. Modal components with fuzzy entropy values ​​between the low-complexity threshold and the high-complexity threshold are classified as medium-complexity components, serving as oscillation components. Modal components with fuzzy entropy values ​​above the high-complexity threshold are classified as high-complexity components, serving as high-frequency components. Based on the complexity differences of each modal component, precise classification of different dynamic characteristic components can be achieved, enabling trend components, oscillation components, and high-frequency components to clearly reflect long-term load changes, periodic fluctuations, and instantaneous impact behaviors, respectively, thereby improving the physical interpretability and feature identification of each component.

[0046] Step 103: Perform secondary mode decomposition on the high-frequency components of the primary mode decomposition to obtain the secondary mode decomposition high-frequency components. This achieves further refined feature analysis of the high-frequency components of the primary mode decomposition, effectively separating the impact components from random noise, and improving the proportion and recognition of effective impact features in the high-frequency components.

[0047] In some embodiments, performing secondary mode decomposition on the high-frequency components includes: performing secondary mode decomposition on the high-frequency components using variational mode decomposition; for high-frequency components containing severe impact characteristics, performing secondary decomposition using variational mode decomposition (VMD) to represent them as several sub-mode components with finite bandwidth; and achieving effective separation of impact components from random noise by constraining the spectral concentration of each sub-mode.

[0048] The formula for calculating the objective function of Variational Mode Decomposition (VMD) is as follows:

[0049] Where M represents the number of high-frequency components in the second-order mode decomposition; This represents taking the partial derivative with respect to time t; This represents the unit impulse function; j represents the imaginary unit. This represents the m-th submode component after high-frequency component decomposition; This is the convolution operator; This represents the center frequency of the m-th submodal component; This represents the square of the L2 norm.

[0050] The constraint formulas for Variational Mode Decomposition (VMD) are as follows:

[0051] in, This represents the m-th submode component after high-frequency component decomposition; M represents the high-frequency component load sequence to be decomposed; M represents the number of high-frequency components in the second-order mode decomposition.

[0052] Step 104: Differentiate modeling of trend component, oscillation component and high-frequency component of second mode decomposition to obtain trend component model, oscillation component model and high-frequency component model. In the training process of trend component model, oscillation component model and high-frequency component model, physical information is introduced and a joint loss function is constructed.

[0053] Based on the dynamic characteristics of each component, trend component models, oscillation component models, and high-frequency component models are constructed respectively. During the training process of each model, physical information related to the formation mechanism of impact loads is introduced. A joint loss function is constructed based on the prediction error and physical constraint terms of each component model. A penalty is constructed based on the deviation between the fused results of the trend component model, the oscillation component model, and the high-frequency component model and the physical information. This penalty is incorporated as a physical correction term into the backpropagation process of the joint loss function to update the parameters of the high-frequency component model. Through differentiated modeling and the introduction of physical information, the trend component model, the oscillation component model, and the high-frequency component model can be adapted to the variation law of the corresponding component, while ensuring that the prediction results of the models conform to the physical constraints of the load, thereby improving the ability of each component model to characterize the behavior of impact loads and the reliability of prediction.

[0054] In some embodiments, differentiated modeling of trend components, oscillation components, and high-frequency components includes: modeling trend components and oscillation components using an ensemble learning regression model, and modeling high-frequency components using a neural network model.

[0055] The regression model for ensemble learning can be an extreme gradient boosting (XGBoost) model, and the neural network model can be a long short-term memory (LSTM) network model. Using the extreme gradient boosting model to model the trend and oscillatory components can effectively capture the nonlinear relationships and periodic change patterns present within them, demonstrating a strong ability to fit the overall trend and fluctuation patterns of the data, thus improving the stability of the trend and oscillatory component predictions. Using the LSTM network model to model the high-frequency components of the quadratic mode decomposition can effectively learn the temporal dependencies of instantaneous impact characteristics in the high-frequency components, exhibiting a strong ability to capture complex and irregular dynamic changes, thus helping to improve the prediction accuracy of the impact component in the high-frequency components.

[0056] The process of constructing the joint loss function is as follows: the outputs of the trend component model, oscillation component model, and high-frequency component model are fused to obtain the initial model calculation values ​​for the impact load. The formula for calculating the initial model calculation values ​​is:

[0057] in, express Initial model calculation values ​​for momentary impact loads; express The calculated values ​​of the time-trend components using the XGBoost model; express The calculated value of the oscillation component at any given time using the XGBoost model; express The total calculated value of the LSTM model for high-frequency components at each time step; This represents the i-th sampling time.

[0058] t i The formula for calculating the total calculated value of the LSTM model for the high-frequency components at time step 1 is:

[0059] in, express The total calculated value of the LSTM model for high-frequency components at each time step; express The LSTM model calculation value of the m-th submode component of the high-frequency component at time step M; M represents the number of high-frequency components in the second-order mode decomposition. This represents the i-th sampling time.

[0060] During the training of the XGBoost model for the trend component, the XGBoost model for the oscillation component, and the LSTM model for the high-frequency component, physical information related to the formation mechanism of impact loads is introduced to construct a joint loss function.

[0061] The formula for calculating the joint loss function is:

[0062] in, Represents the joint loss function; This represents the data fitting loss term; Represents the physical information loss term; and This represents the weighting coefficient, the magnitude of which is positively correlated with the degree of abrupt change in the current operating condition.

[0063] The formula for calculating the data fitting loss term is:

[0064] in, H represents the data fitting loss term; H represents the total sample length of the time series. express Initial model calculation values ​​for momentary impact loads; express Measured values ​​of impact load at any given moment; This represents the i-th sampling time.

[0065] The formula for calculating the physical information loss term is:

[0066] in, The physical information loss term is represented by H; the total length of the time series samples is represented by H. express Initial model calculation values ​​for momentary impact loads; Indicates in At any given moment, a set of physical information composed of time-series variables related to the operating status of the equipment; This represents a function constructed based on physical laws, which takes physical information as input and obtains the theoretical benchmark reference value for impact load.

[0067] For example, when the impact load is the electrical load of the rolling mill, the physical information may include the rolling force and the main motor speed.

[0068] The physical information set of the rolling mill electrical load is represented as follows:

[0069] in, A set of physical information representing the electrical load of a rolling mill; express The total rolling force at any given time is the sum of the rolling force on the drive side and the rolling force on the working side. express The rotational speed of the main drive motor of the rolling mill at any given time.

[0070] The functional expression for the rolling mill electrical load, based on the electromechanical energy conversion law, is as follows:

[0071] in, This represents a function constructed based on the electrical energy conversion law of a rolling mill. A set of physical information representing the electrical load of a rolling mill; Indicates the electromechanical energy conversion coefficient; Indicates the lever arm coefficient; express The total rolling force at any given time is the sum of the rolling force on the drive side and the rolling force on the working side. J represents the lever arm length of the rolling deformation zone; J represents the equivalent moment of inertia of the rolling mill drive system. This indicates the rotational speed of the main drive motor of the rolling mill; This represents the i-th sampling time.

[0072] The formula for calculating the physical information loss term of the rolling mill electrical load is as follows:

[0073] in, The physical information loss item representing the electrical load of the rolling mill; This represents the total sample length of the rolling mill electrical load time series; express Initial model calculation values ​​of the mill electrical load at any given time; Indicates the electromechanical energy conversion coefficient; Indicates the lever arm coefficient; express The total rolling force at any given time is the sum of the rolling force on the drive side and the rolling force on the working side. J represents the lever arm length of the rolling deformation zone; J represents the equivalent moment of inertia of the rolling mill drive system. This indicates the rotational speed of the main drive motor of the rolling mill; This represents the i-th sampling time.

[0074] Step 105: Based on the deviation between the fusion results of the trend component model, oscillation component model, and high-frequency component model outputs and the physical information, a penalty amount is constructed. This penalty amount is then incorporated as a physical correction term into the backpropagation process of the joint loss function to update the parameters of the high-frequency component model, resulting in the corrected output of the high-frequency component model. The physical information dataset is not used as a model input feature; instead, it is used to construct a physical information loss term and a penalty amount. This is applied to the parameter update of the high-frequency component model through gradient backpropagation, thereby correcting the model calculation results to conform to the actual energy conversion relationship and dynamic change characteristics of impulsive electrical loads.

[0075] In some embodiments, physical information is introduced and a joint loss function is constructed during the training process of the trend component model, oscillation component model, and high-frequency component model, including: introducing physical information and constructing a joint loss function based on the data fitting loss term and the physical information loss term during the training process of the trend component model, oscillation component model, and high-frequency component model.

[0076] By introducing a penalty based on physical information to correct the parameters of the high-frequency component model, the output of the high-frequency component model can better fit the physical change law of the impact load, improve the reliability of the fusion prediction results of the trend component model, oscillation component model and high-frequency component model, and reduce the prediction bias caused by data-driven modeling alone.

[0077] The process of constructing the penalty term is as follows: Based on the electromechanical energy conversion relationship during the formation of impact loads, the degree of deviation between the model output and physical laws is calculated, physical deviations are constructed, and penalty terms are generated. The physical relationship expression is as follows:

[0078] in, express Initial model calculation values ​​for momentary impact loads; Indicates in At any given moment, a set of physical information composed of time-series variables related to the operating status of the equipment; This represents a function constructed based on physical laws, which takes physical information as input and obtains the theoretical benchmark reference value for impact load.

[0079] definition The physical deviation at time is:

[0080] in, express The physical deviation value at any given time; express Initial model calculation values ​​for momentary impact loads; Indicates in At any given moment, a set of physical information composed of time-series variables related to the operating status of the equipment; This represents a function constructed based on physical laws, which takes physical information as input and obtains the theoretical benchmark reference value for impact load.

[0081] Based on the degree of physical deviation and its changing trend, construct The penalty amount at each moment, and the formula for calculating the penalty amount is:

[0082] in, express The penalty amount corresponding to each moment; express The weighting coefficient corresponding to each time point; express The deviation corresponding to the time; The adjustment coefficient representing the deviation derivative term; Indicates the rate of change of deviation; This represents the i-th sampling time.

[0083] The formula for calculating the weighting coefficient at time step is:

[0084] in, express Weighting coefficients at different times; The adjustment coefficient representing the weighting coefficient; This represents the rate of change of the initial calculated values ​​of the model.

[0085] During the backpropagation of the joint loss function, the formula for updating the high-frequency component model parameters is as follows:

[0086] in, This represents the new model parameters after penalty; This represents the old model parameters before backpropagation update; This represents the learning rate, which controls the size of each weight update step. H represents the joint loss function; H represents the total length of the time series samples; P represents the penalty amount.

[0087] Step 106: Fuse the outputs of the trend component model, the oscillation component model, and the corrected output of the high-frequency component model to obtain the model calculation value for the impact load. By fusing the outputs of the trend component model, the oscillation component model, and the corrected high-frequency component model, the ability of each component model to characterize different change patterns can be comprehensively utilized, resulting in a more comprehensive model calculation value that better reflects the actual change characteristics of the impact load. This helps to improve the reliability and accuracy of the overall prediction results.

[0088] In some embodiments, the formula for calculating the model value of impact load is as follows:

[0089] in, express The calculated values ​​of the impact load model at any given time; express The trend component model output at time step; express The output of the oscillation component model at time step; express The model correction output of the high-frequency component of the second-order mode decomposition at time m; M represents the number of high-frequency components in the second-order mode decomposition. This represents the i-th sampling time.

[0090] This invention also provides a system for modeling impact loads that integrates dual-mode decomposition and physical information, comprising: a data acquisition module for constructing a training set of historical time-series data of impact loads and acquiring physical information related to the formation mechanism of impact loads; a primary mode decomposition module for performing primary mode decomposition on the historical time-series data, analyzing and reorganizing the results of the primary mode decomposition based on a complexity evaluation index to obtain trend components, oscillation components, and high-frequency components of the primary mode decomposition; a secondary mode decomposition module for performing secondary mode decomposition on the high-frequency components of the primary mode decomposition to obtain high-frequency components of the secondary mode decomposition; and a differentiated modeling module for further processing the trend components, oscillation components, and high-frequency components of the secondary mode decomposition. Differential modeling is performed to obtain trend component model, oscillation component model, and high-frequency component model. Physical information is introduced and a joint loss function is constructed during the training of the trend component model, oscillation component model, and high-frequency component model. The model correction module is used to construct a penalty amount based on the deviation between the fusion result of the trend component model, oscillation component model, and high-frequency component model output and the physical information. The penalty amount is incorporated as a physical correction term into the backpropagation process of the joint loss function to update the parameters of the high-frequency component model and obtain the corrected output of the high-frequency component model. The result fusion module is used to fuse the output of the trend component model, the output of the oscillation component model, and the corrected output of the high-frequency component model to obtain the model calculation value of the impact load.

[0091] The same or similar parts among the various embodiments in this specification can be referred to mutually, and will not be repeated here.

[0092] Figures 2 to 5 This is a drawing for an electrical load for a rolling mill according to an embodiment of the present invention, wherein the electrical load for the rolling mill is an example of an impact load.

[0093] Please see Figure 2 , Figure 2This diagram shows the result of a first mode decomposition of the original mill electrical load sequence in this embodiment of the invention. It illustrates the 13 intrinsic mode function (IMF) components and one residual term obtained from the original load signal decomposition. This decomposition significantly improves the suppression of mode aliasing. Specifically, each IMF exhibits a transition from high to low frequency over time. Through this decomposition, the complex multi-scale characteristics of the mill electrical load are effectively extracted; for example, components in the high-frequency range accurately capture severe impacts and random noise. This lays a reliable data foundation for subsequent component recombination based on complexity indices and secondary decomposition of high-frequency components.

[0094] Please see Figure 3 , Figure 3 This diagram shows the fuzzy entropy results calculated from the intrinsic mode function components obtained from a single decomposition in this embodiment of the invention. It characterizes the complexity and randomness of each mode component. The fuzzy entropy values ​​of IMF1-IMF4 are significantly high (≥0.76), exhibiting strong fluctuations and high frequency, belonging to the same reconstructed component, i.e., high-frequency components. The fuzzy entropy values ​​of IMF5-IMF7 are in the middle range (0.48-0.10), showing certain oscillations and periodicity, belonging to the same reconstructed component, i.e., oscillating components. The fuzzy entropy value of IMF8 is 0.03, and the fuzzy entropy values ​​of IMF9-IMF13 are 0, exhibiting stability and trend, belonging to the same reconstructed component, i.e., trend components. This fuzzy entropy result provides a scientific and quantitative basis for reconstructing the components into high-frequency, oscillating, and trend components.

[0095] Please see Figure 4 , Figure 4 This is the result diagram after analyzing and recombining each modal component based on fuzzy entropy in an embodiment of the present invention; the modal components are divided into trend components, oscillation components and high-frequency components. Compared with trend components and oscillation components, high-frequency components still have large load abrupt changes and irregularities.

[0096] Please see Figure 5 , Figure 5 The figure shows the modeling results of the rolling mill electrical load after integrating dual-modal decomposition and physical information in this embodiment of the invention. It compares the calculated values ​​with the measured values, reflecting the fitting effect of this method in the process of impact changes. As can be seen from the figure, this method exhibits a good fitting effect in complex impact load modeling. Quantitative analysis shows that the model's coefficient of determination (R²) is... 2 The mean absolute percentage error (MAPE) was 0.964, and the mean absolute percentage error (MAPE) was 7.86%. This indicates that the method not only captures the macroscopic evolution trend of the load sequence, but also accurately reproduces most of the impact characteristics.

[0097] This invention effectively reduces mode aliasing and improves the accuracy of impact feature extraction by using dual mode decomposition and complexity index classification and recombination. It adopts differentiated modeling to adapt to the multi-scale variation characteristics of the load, enhancing the model fitting and generalization capabilities. Based on the deviation between the fusion results of each model output and the physical information, a penalty amount is constructed and incorporated as a physical correction term into the backpropagation process of the joint loss function to update the parameters of the high-frequency component model. This makes the model output more consistent with the load operation mechanism, reduces the deviation and lag during operating condition switching and load abrupt changes, improves the ability to characterize the transient characteristics of impact loads, and helps to improve the scheduling and operation stability of industrial power supply systems.

[0098] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for modeling impact loads that integrates dual-mode decomposition and physical information, characterized in that, include: Construct a training set of historical time-series data on impact loads to obtain physical information related to the formation mechanism of the impact loads; The historical time series data is subjected to a first mode decomposition. The results of the first mode decomposition are analyzed and reorganized based on the complexity evaluation index to obtain trend components, oscillation components and high-frequency components of the first mode decomposition. The high-frequency components of the first mode decomposition are subjected to second mode decomposition to obtain the high-frequency components of the second mode decomposition. Differential modeling is performed on the trend component, the oscillation component, and the high-frequency component of the second-order mode decomposition to obtain the trend component model, the oscillation component model, and the high-frequency component model. The physical information is introduced and a joint loss function is constructed during the training process of the trend component model, the oscillation component model, and the high-frequency component model. Based on the deviation between the fusion results of the trend component model, the oscillation component model, and the high-frequency component model and the physical information, a penalty amount is constructed. The penalty amount is then incorporated as a physical correction term into the backpropagation process of the joint loss function to update the parameters of the high-frequency component model, thereby obtaining the corrected output of the high-frequency component model. The output of the trend component model, the output of the oscillation component model, and the corrected output of the high-frequency component model are fused to obtain the model calculation value of the impact load.

2. The method for modeling impact loads that integrates dual-mode decomposition and physical information according to claim 1, characterized in that, The physical information includes time-series variables related to the operating status of the equipment.

3. The method for modeling impact loads that integrates dual-mode decomposition and physical information according to claim 1, characterized in that, The historical time-series data undergoes a first-order mode decomposition. The results of this first-order mode decomposition are analyzed and reorganized based on a complexity evaluation index to obtain trend components, oscillation components, and high-frequency components, including: The historical time series data is decomposed once using the adaptive noise complete set empirical mode decomposition method to obtain multiple mode components. The fuzzy entropy of the multiple mode components is calculated. The multiple mode components are classified and reorganized according to the magnitude of the fuzzy entropy to obtain trend components, oscillation components and high-frequency components.

4. The method for modeling impact loads that integrates dual-mode decomposition and physical information according to claim 1, characterized in that, The step of performing secondary mode decomposition on the high-frequency components includes: performing secondary mode decomposition on the high-frequency components using variational mode decomposition.

5. The method for modeling impact loads that integrates dual-mode decomposition and physical information according to claim 1, characterized in that, The differentiated modeling of the trend component, the oscillation component, and the high-frequency component includes: modeling the trend component and the oscillation component using an ensemble learning regression model, and modeling the high-frequency component using a neural network model.

6. The method for modeling impact loads that integrates dual-mode decomposition and physical information according to claim 1, characterized in that, The step of introducing the physical information and constructing a joint loss function during the training process of the trend component model, the oscillation component model, and the high-frequency component model includes: introducing the physical information and constructing a joint loss function based on the data fitting loss term and the physical information loss term during the training process of the trend component model, the oscillation component model, and the high-frequency component model.

7. The method for modeling impact loads that integrates dual-mode decomposition and physical information according to claim 1, characterized in that, The formula for calculating the penalty amount is: in, express The penalty amount corresponding to each moment; express The weighting coefficient corresponding to each time point; express The deviation corresponding to the time; The adjustment coefficient representing the deviation derivative term; Indicates the rate of change of deviation; This represents the i-th sampling time.

8. The method for modeling impact loads that integrates dual-mode decomposition and physical information according to claim 1, characterized in that, The formula for calculating the model value of the impact load is as follows: in, express The calculated values ​​of the impact load model at any given time; express The trend component model output at time step; express The output of the oscillation component model at time step; express The model correction output of the high-frequency component of the second-order mode decomposition at time m; M represents the number of high-frequency components in the second-order mode decomposition. This represents the i-th sampling time.

9. A system for modeling impact loads that integrates dual-mode decomposition and physical information, characterized in that, include: The data acquisition module is used to construct a training set of historical time-series data of impact loads and acquire physical information related to the formation mechanism of the impact loads; The first mode decomposition module is used to perform first mode decomposition on the historical time series data, and analyze and reorganize the first mode decomposition results based on the complexity evaluation index to obtain trend components, oscillation components and high-frequency components of the first mode decomposition. The secondary mode decomposition module is used to perform secondary mode decomposition on the high-frequency components of the primary mode decomposition to obtain the secondary mode decomposition high-frequency components. The differential modeling module is used to perform differential modeling on the trend component, the oscillation component, and the high-frequency component of the second-mode decomposition to obtain the trend component model, the oscillation component model, and the high-frequency component model. The physical information is introduced and a joint loss function is constructed during the training process of the trend component model, the oscillation component model, and the high-frequency component model. The model correction module is used to construct a penalty amount based on the deviation between the fusion result of the trend component model, the oscillation component model and the high-frequency component model and the physical information, and to incorporate the penalty amount as a physical correction term into the backpropagation process of the joint loss function to update the parameters of the high-frequency component model and obtain the corrected output of the high-frequency component model. The result fusion module is used to fuse the output of the trend component model, the output of the oscillation component model, and the corrected output of the high-frequency component model to obtain the model calculation value of the impact load.

Images