Industrial flexible load state space model parameter prediction method and system

Abstract

The invention discloses an industrial flexible load state space model parameter prediction method and system, and the method comprises the steps: constructing a time sequence segmentation index storage library based on load state segmentation, calculating a heterogeneous dynamic index through phase-space reconstruction, constructing a time-frequency recurrence spectrum through time-frequency analysis, calculating a frequency domain recurrence index, and carrying out the prediction of the parameters of an industrial flexible load state space model. And adaptively selecting a feature enhancement mode of an iterative feature screening and adversarial generation network or a dynamic time-frequency decomposition and variational coding network based on the comprehensive feature coefficient, and finally training a deep learning model to realize parameter prediction. Through adaptive feature processing driven by dynamic complexity, the accuracy and reliability of industrial load parameter analysis are improved, and the operation efficiency and stability of industrial equipment are optimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial flexible load parameter analysis, in particular to an industrial flexible load state space model parameter prediction method and system. BACKGROUND

[0002] With the development of industrial automation technology, the state space model parameter analysis of industrial flexible load has become the core link of power system optimization and equipment efficiency improvement. The state space model provides key support for equipment control and energy management by analyzing the parameters of load data under different states (such as state matrix, i.e. state transition matrix, which describes the dynamic relationship between internal state variables of the system and reflects the natural evolution characteristics of the system without external input, and input matrix, i.e. input gain matrix, which describes the action strength of external input on the system state, and its element value represents the influence direction and size of each input on each state variable). However, the industrial load data has the following characteristics, which poses challenges to existing methods:

[0003] (1) Dynamic complexity: the load state (light load, rated load, overload, etc.) frequently switches, and the data contains time-frequency domain dynamic characteristics and spatial heterogeneous characteristics. Traditional time series storage methods cause different state data to be mixed, and error accumulation is significant.

[0004] (2) Feature extraction limitation: existing methods rely on fixed feature extraction mode and cannot adapt to the dynamic changes of data nonlinear complexity, for example, simple data scene time-frequency analysis and complex data scene GAN enhancement lack adaptive switching mechanism.

[0005] (3) Blindness of mode selection: the selection of feature enhancement mode (such as recursive feature elimination, variational encoding network VAE) lacks quantitative basis, resulting in insufficient processing of high dynamic data or over-processing of simple data.

[0006] The existing technology fails to effectively solve the technical bottlenecks of load state dynamic segmentation storage, feature complexity quantitative evaluation and adaptive mode selection, resulting in insufficient parameter analysis accuracy. SUMMARY

[0007] The purpose of the present application is to provide an industrial flexible load state space model parameter prediction method and system, which solves the problems of low efficiency and accuracy of state space model parameter prediction of complex industrial load data.

[0008] Technical scheme: The industrial flexible load state space model parameter prediction method provided by the present application comprises the following steps:

[0009] Obtain the time series load parameters of the industrial equipment under different flexible load states and store them in the corresponding grid cells;

[0010] reconstructs the phase space of the time series load parameters in each grid cell, calculates Lyapunov exponents and Shannon entropy for the reconstructed vectors, and calculates a heterogeneous dynamic index according to the Lyapunov exponents and the Shannon entropy;

[0011] performs time-frequency analysis on the time series load parameters in each grid cell, constructs a time-frequency recurrence atlas by calculating the similarity of energy spectrum density, calculates recurrence rate and diagonal line length distribution in the time-frequency recurrence atlas, and calculates a frequency domain recurrence index according to the recurrence rate and the diagonal line length distribution;

[0012] calculates a comprehensive feature coefficient according to the heterogeneous dynamic index and the frequency domain recurrence index, selects a corresponding feature enhancement method according to the size relationship between the comprehensive feature coefficient and a threshold value, and performs feature enhancement on the time series load parameters in each grid cell;

[0013] formats the time series load parameters in all grid cells after feature enhancement and inputs them into a trained deep learning model to obtain predicted state space model parameters.

[0014] Further, the obtaining of the time series load parameters of the industrial equipment under different flexible load states and the storage into corresponding grid cells comprise:

[0015] storing the load parameters of the industrial equipment into the grid cells according to the load parameters of the industrial equipment, storing the load parameters in a time period corresponding to the load state into a new grid cell when the load state changes, and forming segmented storage of the load parameters under the same load state.

[0016] Further, the reconstructing of the phase space of the time series load parameters in each grid cell, the calculating of Lyapunov exponents and Shannon entropy for the reconstructed vectors, and the calculating of a heterogeneous dynamic index according to the Lyapunov exponents and the Shannon entropy comprise the following steps:

[0017] reconstructing the phase space of the time series load parameters in each grid cell by using a delay embedding method, mapping the time series load parameters to a high-dimensional space by fixing a delay time and embedding dimension, and obtaining a load parameter vector;

[0018] calculating the divergence rate of adjacent points to obtain Lyapunov exponents and calculating the probability distribution to obtain Shannon entropy for the load parameter vector;

[0019] weighting and summing the Lyapunov exponents and the Shannon entropy to obtain a heterogeneous dynamic index.

[0020] Further, the Lyapunov exponents where M is the total number of data points in the time series load parameters; d i(t) is the distance of the i th data point at time t; Δt is the time interval for calculating the divergence rate of adjacent points

[0021] Shannon entropy H = -∑ i P(a i )logP(a i ), wherein P(a i ) is the probability distribution of the load parameter vector, a i represents the i th load parameter vector.

[0022] Further, the time-frequency analysis of the time series load parameters in each grid cell is performed, and a time-frequency recurrence map is constructed by calculating the similarity of the energy spectrum density; the recurrence rate and the diagonal line length distribution are calculated in the time-frequency recurrence map, and the frequency domain recurrence index is calculated according to the recurrence rate and the diagonal line length distribution, including the following steps:

[0023] The time-frequency analysis of the time series load parameters in each grid cell is performed, and the energy spectrum density under each time window and frequency is calculated;

[0024] The similarity of the energy spectrum density is calculated by the cosine similarity, and a time-frequency recurrence map is constructed;

[0025] For the time-frequency recurrence map, the density of the recurrence points is calculated to obtain the recurrence rate, the diagonal line length distribution is calculated, and the frequency of the diagonal line length is counted to obtain the determinacy;

[0026] The frequency domain recurrence index is obtained by multiplying the recurrence rate and the determinacy.

[0027] Further, the determinacy wherein l represents the length of the diagonal line; f(l) is the frequency of the diagonal line length l, l min is the minimum value of the diagonal line length, l max is the maximum value of the diagonal line length, R(k, v) represents whether the time t k and t v are similar, and is 1 if similar, otherwise 0.

[0028] Further, the comprehensive feature coefficient is calculated according to the heterogeneous dynamic index and the frequency domain recurrence index, and the corresponding feature enhancement method is selected according to the size relationship between the comprehensive feature coefficient and its threshold value, including:

[0029] The heterogeneous dynamic index and the frequency domain recurrence index are weighted and summed to obtain a comprehensive feature coefficient, and if the comprehensive feature coefficient is not less than its threshold value, the iterative feature screening and the adversarial generation network enhancement mode are adopted; otherwise, the dynamic time-frequency decomposition and the variational coding network enhancement mode are adopted.

[0030] Further, the comprehensive feature coefficient Wherein, NFC and RPC are isomeric dynamic index and frequency domain recurrence index respectively; ω1 and ω2 are preset proportion coefficients of isomeric dynamic index and frequency domain recurrence index respectively, and both are greater than zero.

[0031] The industrial flexible load state space model parameter prediction system provided by the application comprises:

[0032] A data collection module is configured to acquire time series load parameters of an industrial device under different flexible load states and store the time series load parameters in corresponding grid cells.

[0033] A spatial isomerism analysis module is configured to perform phase space reconstruction on the time series load parameters in each grid cell, calculate Lyapunov exponents and Shannon entropies for the reconstructed vectors, and calculate an isomeric dynamic index based on the Lyapunov exponents and the Shannon entropies.

[0034] A frequency domain dynamic analysis module is configured to perform time-frequency analysis on the time series load parameters in each grid cell, construct a time-frequency recurrence map by calculating the similarity of energy spectrum densities, calculate a recurrence rate and a diagonal line length distribution in the time-frequency recurrence map, and calculate a frequency domain recurrence index based on the recurrence rate and the diagonal line length distribution.

[0035] An intelligent decision module is configured to calculate a comprehensive feature coefficient based on the isomeric dynamic index and the frequency domain recurrence index.

[0036] A feature enhancement module is configured to select a corresponding feature enhancement method based on the size relationship between the comprehensive feature coefficient and a threshold value of the comprehensive feature coefficient, and perform feature enhancement on the time series load parameters in each grid cell.

[0037] A parameter prediction module is configured to input the time series load parameters in all grid cells after feature enhancement into a trained deep learning model in a formatted manner, and obtain predicted state space model parameters.

[0038] The electronic device provided by the application comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is loaded into the processor to implement the industrial flexible load state space model parameter prediction method.

[0039] The computer readable storage medium provided by the application stores a computer program, wherein the computer program is executed by a processor to implement the industrial flexible load state space model parameter prediction method.

[0040] Advantages: Compared with the prior art, the advantages of the present application are: the present application realizes the whole process optimization from data organization to model training through load state segmentation storage, double exponential dynamic evaluation model and self-adaptive feature enhancement mode, significantly improves the reliability and flexibility of parameter analysis. Specifically, the present application collects and constructs a time sequence segmentation index repository based on load state segmentation, judges the load state by real-time load parameters, and stores the data to the grid unit when the state changes, realizes the efficient organization and management of data, ensures the independent analysis of data under different load states, avoids the complexity and error caused by mixed data; the data in each grid unit is reconstructed in phase space, the heterogeneous dynamic index is calculated, the complex dynamic behavior and spatial heterogeneous characteristics of load data are revealed, and the accurate evaluation of high dynamic and complex data is ensured; the energy spectrum density is calculated by time-frequency analysis and the time-frequency recurrence map is constructed, the recurrence rate and the diagonal line length distribution are calculated by cosine similarity, the frequency domain recurrence index is calculated, the dynamic complexity of data is fully evaluated, which is helpful to select the most suitable processing mode under different dynamic conditions; by comprehensively processing the heterogeneous dynamic index and the frequency domain recurrence index, the comprehensive feature coefficient is obtained, the iterative feature screening and the adversarial generative network enhancement mode or the dynamic time-frequency decomposition and the variational coding network enhancement mode are flexibly decided, the efficient feature enhancement and optimization processing of data are realized, and the accuracy and reliability of the state space model parameter analysis are significantly improved; for the grid unit data processed by the above method, the data is formatted and segmented, and the deep learning model is trained to predict and analyze the state space model parameters, providing an efficient and accurate parameter analysis method to improve the operation efficiency and stability of industrial flexible load under different states. By flexibly selecting the processing mode suitable for different feature data, the present application effectively deals with complex industrial load data, optimizes the state space model parameter analysis process, and improves the overall performance of the industrial system. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 The industrial flexible load state space model parameter prediction method of the present application is shown in the flow chart. DETAILED DESCRIPTION

[0042] The technical solutions of the present application will be further described below in conjunction with the drawings.

[0043] As shown in Figure 1 , the industrial flexible load state space model parameter prediction method comprises the following steps.

[0044] S1, collect and construct a time sequence segmentation index repository based on load state segmentation, judge the load state by real-time load parameters, and store the data to the grid unit when the state changes.

[0045] S2, phase space reconstruction is performed on the data in each grid cell, and based on Lyapunov exponent and Shannon entropy, a heterogeneous dynamic index is calculated to evaluate the nonlinear complexity of the data.

[0046] S3, time-frequency analysis is performed on the data in each grid cell, energy spectrum density is calculated, and a time-frequency recurrence map is constructed, recurrence rate and diagonal length distribution are calculated by cosine similarity, and a frequency domain recurrence index is obtained to evaluate the dynamic complexity of the data and assist in selecting the processing mode.

[0047] S4, the heterogeneous dynamic index and the frequency domain recurrence index are comprehensively processed to obtain a comprehensive feature coefficient, and decision is made to use iterative feature screening and generative adversarial network enhanced mode or dynamic time-frequency decomposition and variational coding network enhanced mode.

[0048] S5, for the grid cell data after the feature enhancement processing in step S4, the data is formatted and segmented, and a deep learning model is trained to predict and analyze the state space model parameters of new data.

[0049] The operating parameters of industrial equipment under different load conditions have significant differences, and the mixed storage of data under different conditions will reduce the complexity and accuracy of subsequent analysis; the existing method based on time series data storage often mixes data under different load conditions, which requires a lot of time and computing resources to separate and process data under each condition when analyzing data, which not only increases the complexity of data processing, but also easily introduces errors and reduces the accuracy of analysis; therefore, segmenting the data according to the load state can help improve the efficiency and accuracy of data processing. By classifying data under the same load condition into the same grid cell, not only the continuity of the data can be maintained, but also a clearer basis for subsequent feature extraction and pattern recognition can be provided; in addition, this segmented storage method facilitates quick retrieval and analysis of data under a specific load condition, supports more detailed and targeted research, and improves the overall performance and quality of industrial load data analysis. Therefore, implementing data segmentation storage based on load state is a key step in optimizing industrial load data processing and analysis, and provides a solid foundation for accurate state space model parameter analysis and industrial system optimization.

[0050] Step S1 includes the following contents:

[0051] S1.1, First, define different load states of the industrial equipment, which may include light load, rated load, overload, start-stop, etc. The definition of load state is based on the current, voltage, power and other parameters of the industrial equipment; for example, the light load state can be defined as the stage when the device power is less than 30% of the rated power, and the current and voltage fluctuate within a certain range; the rated load state is the stage when the device operates stably at rated power; through clear parameter standards and thresholds, the current load state of the device can be judged in real time, laying the foundation for data segmented storage.

[0052] S1.2, Initialize the load data storage structure, create an empty grid cell list to store related data under each load state; when traversing the time series data, segment the data according to the change of load state, specifically, first create an empty grid cell and use it as the current storage unit; then, starting from the start time of the data, read the load parameters at each time point, and determine the current load state according to these load parameters; when detecting a change in load state, store the data of the current grid cell in the database and create a new grid cell to store the data under the new load state.

[0053] S1.3, The data segmented storage process starts with initializing the grid cell list, creating an empty grid cell list and defining the current load state and the current grid cell; when traversing the time series data, get the load parameters at each time point, determine the current load state according to these load parameters, if the current load state is the same as the previous load state, add the data to the current grid cell; if the current load state is different from the previous load state, store the current grid cell in the grid cell list, create a new grid cell and update the current state to the new load state.

[0054] S1.4, After traversing the time series data, store the last current grid cell in the grid cell list, at this time, all data of the same continuous load state has been segmented and stored in different grid cells; next, store all grid cells in the grid cell list in the database, each grid cell contains all data under a load state, this storage method can improve the organization and management efficiency of data, and facilitate subsequent data analysis and feature extraction.

[0055] When collecting industrial load data and constructing a database, the data is stored in segments according to the load state; different load states of industrial equipment are defined, and the current load state is judged in real time according to the load parameters; in the time series data, whenever the load state changes, the data in the current load state time period is stored in a grid cell, and then the subsequent load state is analyzed, and the data in the next load state time period is stored in the next grid cell; in this way, a data grid storage form based on continuous same load state is constructed, which is convenient for subsequent data analysis and feature extraction.

[0056] In the process of analyzing the parameters of the industrial flexible load state space model, the operating parameters of the industrial equipment under different load states have significant differences. Therefore, the data is stored in segments according to the load state, and the data in each grid cell is analyzed separately, which can significantly improve the accuracy and effectiveness of data processing.

[0057] The traditional method of storing time series data stores all data together in chronological order without considering the differences in load states. This method has several shortcomings: first, the data under different load states is mixed together, increasing the difficulty of data processing and leading to poor feature extraction and pattern recognition results; second, when analyzing the parameters of the state space model, noise and interference in the mixed data can reduce the accuracy and reliability of the analysis; finally, this storage method is not convenient for quickly retrieving and analyzing data under a specific load state, limiting in-depth research and optimization of the operating performance of industrial equipment.

[0058] By storing data in segments according to the load state and separately analyzing and evaluating the data in each grid cell, the data characteristics under each load state can be more accurately captured. Specifically, the non-linear complexity and dynamic characteristics of the data in each grid cell are evaluated separately, which can accurately identify the complexity and trend of the data and select the most appropriate preprocessing mode, thereby improving the efficiency and accuracy of data processing.

[0059] This method has multiple benefits: first, it avoids the interference caused by mixing data under different load states, ensuring that the data under each load state has high coherence and consistency; second, by accurately evaluating the data characteristics of each grid cell, the most suitable preprocessing mode can be selected, further improving data quality and analysis accuracy; finally, this method is convenient for quickly retrieving and analyzing data under a specific load state, which helps in-depth research and optimization strategies for the operating performance of industrial equipment.

[0060] Step S2 includes the following:

[0061] Calculating and analyzing heterogeneous dynamic indices is essential when analyzing each grid cell individually, as industrial load data exhibits significant spatial heterogeneity under different load states. These characteristics are crucial for accurate state-space model parameter analysis. Calculating heterogeneous dynamic indices captures complex dynamic behaviors and potential patterns within the data, aiding in the identification and understanding of industrial equipment's operational performance and behavior under specific conditions. Existing methods based on time-series data storage and analysis cannot effectively separate and process these complex spatial heterogeneities, leading to reduced accuracy and reliability in state-space model parameter analysis. However, by analyzing and evaluating heterogeneous dynamic indices separately within each grid cell, the data for each load state receives the most suitable preprocessing and analysis, improving the accuracy and effectiveness of data processing and providing high-quality data support for subsequent state-space model parameter analysis. This approach not only avoids interference from mixed data of different load states but also allows for more precise selection and application of the most suitable preprocessing mode, further enhancing the depth of research on industrial equipment operation optimization and system stability.

[0062] The logic for obtaining the heterogeneous dynamic index is as follows:

[0063] A1. For the data within each network unit, the phase space is reconstructed using the delayed embedding method. A fixed delay time and embedding dimension are selected to map the time series data to a high-dimensional space.

[0064] A(t)=[a(t), a(t+τ), a(t+2τ),..., a(t+(N-1)τ)]

[0065] Where A(t) represents the reconstructed phase space vector at time t, which is used to embed one-dimensional time series data into a high-dimensional space to better capture its dynamic characteristics and complex behavior. Specifically, it is a multi-dimensional vector containing data from the current time point and several time points before and after it; a(t) represents the data value at different time points t, which is the original one-dimensional time series data; τ is the delay time, which is the time interval used to reconstruct the phase space. It is used to delay the original time series data for a period of time before sampling, which helps to capture the dynamic characteristics of the time series; N is the embedding dimension, which is the dimension of the reconstructed phase space and indicates how many sampling points with delay time are used when reconstructing the phase space.

[0066] The delayed embedding method transforms one-dimensional time series data into multi-dimensional vectors. The purpose of this process is to embed the time series data into a high-dimensional space in order to reveal the dynamic characteristics and potential patterns of the data.

[0067] Phase space reconstruction is the foundation for subsequent steps. By converting one-dimensional data into multi-dimensional data, it provides the necessary input for subsequent calculations of the Lyapunov exponent and Shannon entropy.

[0068] A2, calculate the divergence rate of adjacent points, derive the Lyapunov exponent, measure the nonlinear dynamic behavior of the system.

[0069]

[0070] wherein a represents the Lyapunov exponent, used to measure the complexity of the data; M is the total number of data points; d i (t) is the distance of the i-th data point at time t; Δt is the time interval, used to calculate the divergence rate of adjacent points, usually Δt = τ.

[0071] Lyapunov exponent calculation depends on the reconstructed phase space vector, by analyzing the divergence of these vectors over time, the chaoticity and spatial heterogeneity of the system can be quantified.

[0072] A3, calculate the probability distribution P(a i ) of the reconstructed phase space vector, and calculate the Shannon entropy H based on the probability distribution, which is used to measure the complexity of the data.

[0073]

[0074] The probability distribution of the load parameter vector refers to the probability distribution of the high-dimensional load parameter vector (reflecting the time series characteristics of the industrial load state) in the phase space obtained by phase space reconstruction. The frequency of each vector (or the region where the vector is located) in the high-dimensional phase space is counted, and then the probability distribution of each vector (or the region where the vector is located) is obtained, that is, the probability of each phase space vector is determined by counting the ratio of the number of times each phase space vector appears to the total number of vectors, and finally the probability distribution of the load parameter vector is formed, which is used to quantify the possibility of different load state vectors. In essence, it reflects the distribution characteristics of industrial load data in high-dimensional space, and can be used to evaluate the uncertainty, complexity and dynamic behavior of the data. The more dispersed the probability distribution, the higher the uncertainty and complexity of the data; the more concentrated the distribution, the more stable and regular the dynamic characteristics of the data; the calculation of Shannon entropy also depends on the reconstructed phase space vector, and by analyzing the probability distribution of these vectors, the complexity of the data can be evaluated.

[0075] A4, calculate the entropy values of Lyapunov exponent and Shannon entropy, then determine the weights of each according to the entropy values of the two, and then use the respective weights to weight the sum of the entropy values of Lyapunov exponent and Shannon entropy, and obtain the heterogeneity dynamic index.

[0076] The heterogeneous dynamic index is used to represent and embody the complex dynamic behavior of industrial load data under different load states and the complexity of the data. The larger the heterogeneous dynamic index, the more complex the dynamic characteristics and the higher the uncertainty of the industrial load data under the state, which may involve more nonlinear relationships and chaotic behavior, which means that more complex and sophisticated analysis and processing methods are needed during the state space model parameter analysis process. On the contrary, a smaller heterogeneous dynamic index indicates that the dynamic behavior of the industrial load data is relatively simple, the complexity of the data is low, and there are fewer nonlinear relationships and uncertainties, which is suitable for using a relatively simple preprocessing mode. By evaluating and comparing the heterogeneous dynamic indices in different grid cells, the most suitable preprocessing mode can be selected, thereby improving the accuracy and reliability of the state space model parameter analysis and ultimately optimizing the operation efficiency and stability of the industrial equipment.

[0077] When analyzing each grid cell separately, it is necessary to calculate and analyze the frequency domain recurrence index because industrial load data has significant time-frequency dynamic characteristics under different load states, which are crucial for accurate state space model parameter analysis. By calculating the frequency domain recurrence index, the recurrence patterns and dynamic complexity of the data in the time-frequency domain can be captured, which helps to identify and understand the operation performance and behavior of the industrial equipment under specific conditions. Existing methods based on time series data storage and analysis cannot effectively separate and process these complex time-frequency dynamic characteristics, resulting in reduced accuracy and reliability during state space model parameter analysis. By separately analyzing and evaluating the frequency domain recurrence index for data in each grid cell, the most suitable preprocessing and analysis for each load state data can be ensured, improving the accuracy and effectiveness of data processing and providing high-quality data support for subsequent state space model parameter analysis. This method not only avoids interference caused by mixing data from different load states, but also allows for more precise selection and application of the most suitable preprocessing mode, further enhancing the depth of research on industrial equipment operation optimization and system stability.

[0078] Step S3 includes the following:

[0079] The logic for obtaining the frequency domain recurrence index is as follows:

[0080] B1. Preprocess the original data in each grid cell to eliminate noise and unnecessary fluctuations, making the data smoother and more stable.

[0081] Use a low-pass filter to remove high-frequency noise in the data, ensuring that the main features of the data are not disturbed by noise during subsequent analysis.

[0082] Standardize the data to have a mean of 0 and a variance of 1, thereby eliminating the effects of dimensions and facilitating the comparison and processing of different features.

[0083] B2, time-frequency analysis is performed on the preprocessed time series data using short-time Fourier transform (STFT) to obtain time-frequency domain representation, and the energy spectral density at each time window and frequency is calculated to reveal the characteristics of data changes at different frequencies and times, providing more abundant information for the construction of time-frequency recurrence atlas.

[0084] Short-time Fourier transform is performed on the time series data of each grid cell to obtain time-frequency domain representation.

[0085] For each time window and frequency, the energy spectral density is calculated.

[0086] B3, by calculating the similarity between the time-frequency energy spectral density, construct time-frequency recurrence atlas, display the recurrence mode of data in time-frequency domain, the specific method includes using cosine similarity to calculate the similarity between different time windows, and selecting appropriate similarity threshold according to the characteristics of data to construct time-frequency recurrence atlas.

[0087] The cosine similarity is used to calculate the energy spectral density similarity between different time windows, and the formula is:

[0088]

[0089] Where S(k,v) represents the similarity between time t k and t v , P(t k ,f) and P(t v ,f) represent the energy distribution vector of the signal at time t k and t v , respectively, and the frequency is f, ||P(t k ,f)|| and ||P(t v ,f)|| are the norms of the vectors, k and v are time window indices, and k≠v.

[0090] According to the characteristics of data, select appropriate similarity threshold, construct time-frequency recurrence atlas, the formula is:

[0091] R(k,v)=μ[ε-S(k,v)]

[0092] Where R(k,v) represents whether time t k and t v are similar, 1 if similar, otherwise 0, μ[] is Heaviside step function, and ε is the preset similarity threshold.

[0093] B4, calculate the recurrence rate (RR), that is, the density of recurrence points in the time-frequency recurrence atlas, reflect the recurrence frequency of system state, by statistics time-frequency recurrence atlas value is 1 point number and calculate the recurrence rate formula to realize.

[0094] B5, calculate the diagonal length distribution (DET) reflecting the duration and certainty of system state by counting the frequency of diagonal length in the statistical time-frequency recurrence map and calculating the certainty:

[0095]

[0096] wherein, l represents the length of the diagonal; f(l) is the frequency of the diagonal length l, l min is the minimum value of the diagonal length, l max is the maximum value of the diagonal length.

[0097] B6, calculate the frequency domain recurrence index (RPC) which is the product of the recurrence rate and the certainty by integrating the recurrence rate and the diagonal length distribution, to comprehensively evaluate the dynamic complexity of the data, and provide a reliable evaluation index for mode selection, so as to optimize the state space model parameter analysis and the industrial equipment operation performance.

[0098] The frequency domain recurrence index is used to represent and embody the recurrence mode and dynamic complexity of the industrial load data in the time-frequency domain. The greater the frequency domain recurrence index, the higher the recurrence frequency and certainty of the data, which reflects that the dynamic behavior of the data is more complex and has more long-term dependence. On the contrary, the smaller the frequency domain recurrence index, the lower the recurrence frequency and certainty of the data, which reflects that the dynamic behavior of the data is relatively simple and has strong short-term dependence. In this scenario, the frequency domain recurrence index can assist in selecting the most suitable processing mode: when the frequency domain recurrence index is large, it indicates that the industrial load data is complex and has high dynamic characteristics, and mode one, i.e., the iterative feature screening and GAN enhancement mode, is suitable for dealing with complex nonlinear relationships and dynamic characteristics; when the frequency domain recurrence index is small, it indicates that the industrial load data is relatively simple and has weak dynamic characteristics, and mode two, i.e., the dynamic time-frequency decomposition and VAE enhancement mode, is suitable for dealing with relatively simple data characteristics more efficiently, so as to optimize the state space model parameter analysis and the industrial equipment operation performance.

[0099] Step S4 includes the following contents:

[0100] The heterogeneous dynamic index and the frequency domain recurrence index are weighted and summed to calculate a comprehensive feature coefficient. For example, the following calculation formula can be used:

[0101]

[0102] wherein, DCAC represents the comprehensive feature coefficient; NFC and RPC are the heterogeneous dynamic index and the frequency domain recurrence index, respectively; ω1 and ω2 are the preset proportion coefficients of the heterogeneous dynamic index and the frequency domain recurrence index, respectively, and both are greater than zero; the logarithmic transformation is used to normalize the magnitude difference between the heterogeneous dynamic index and the frequency domain recurrence index, so as to facilitate threshold comparison.

[0103] The comprehensive feature coefficient is used to represent and embody the comprehensive dynamic characteristics and complexity of the industrial load data, combining the heterogeneous dynamic index and the frequency domain recurrence index, and the larger comprehensive feature coefficient indicates that the industrial load data has higher nonlinear dynamic behavior and complex recurrence mode, reflecting that the data has strong complexity and dynamic characteristics, and is suitable for using the iterative feature screening and GAN (Generative Adversarial Network) enhancement mode to process complex data features; on the contrary, the smaller comprehensive feature coefficient indicates that the nonlinear dynamic behavior and recurrence mode of the industrial load data are relatively simple, and the dynamic characteristics are relatively low, and it is suitable to use the dynamic time-frequency decomposition and VAE (Variational Auto-Encoder) enhancement mode to more efficiently process the simpler data features, so as to optimize the state space model parameter analysis and the operation performance of the industrial equipment.

[0104] The comprehensive feature coefficient and the degree threshold value are compared, if the comprehensive feature coefficient is greater than or equal to the degree threshold value, it indicates that the industrial load data in the grid cell has higher spatial heterogeneity characteristics and complex dynamic behavior, the data has higher recurrence frequency and certainty, the dynamic characteristics of the system are complex and have long-term dependence, in this case, it is suitable to use the iterative feature screening and GAN (Generative Adversarial Network) enhancement mode to process the data, to deal with complex nonlinear relationships and dynamic characteristics, and to ensure the accuracy and reliability of the state space model parameter analysis.

[0105] On the contrary, if the comprehensive feature coefficient is less than the degree threshold value, it indicates that the spatial heterogeneity characteristics and dynamic complexity of the industrial load data in the grid cell are relatively low, the data has weak recurrence frequency and certainty, the dynamic behavior of the system is relatively simple and has strong short-term dependence, in this case, it is suitable to use the dynamic time-frequency decomposition and VAE (Variational Auto-Encoder) enhancement mode to process the data, to more efficiently extract features and optimize the state space model parameter analysis, so as to improve the operation efficiency and stability of the industrial equipment.

[0106] The iterative feature screening and GAN (Generative Adversarial Network) enhancement mode is a technology that recursively eliminates redundant and irrelevant features through an adaptive method, and uses the GAN to enhance and generate more representative feature data, in the industrial flexible load state space model parameter analysis method, this mode can effectively process data with complex and high dynamic characteristics, recursively eliminate redundant features through an adaptive iterative algorithm, retain key dimensions based on a feature importance threshold, combine the generator-discriminator architecture of the GAN, generate enhanced features consistent with the distribution of the original data, and is suitable for complex data scenarios with high heterogeneous dynamic index and frequency domain recurrence index.

[0107] Dynamic time-frequency decomposition and VAE enhanced mode is a feature enhancement method combining time-frequency analysis technology and variational encoding network (VAE); time-frequency analysis is used to decompose and analyze the characteristics of data at different times and frequencies, while VAE is used to extract and enhance the spatial heterogeneous characteristics of data; in the industrial flexible load state space model parameter analysis method, this mode can efficiently process relatively simple and low dynamic data, through short-time Fourier transform for dynamic time-frequency decomposition of data, combined with variational encoding network to extract low-dimensional feature representation, to realize feature optimization of simple data, specifically, first, the data is analyzed in frequency domain by sliding time window to generate energy spectrum density matrix, then the nonlinear mapping ability of variational encoding network is used to compress high-dimensional time-frequency features to latent space to extract representative low-dimensional features, which is suitable for low dynamic complexity load data.

[0108] These two modes have a significant impact on the processing process in the industrial flexible load state space model parameter analysis method; the iterative feature screening and GAN feature enhancement mode improves the processing capability of high dynamic and complex data through complex data preprocessing and feature enhancement, and enhances the accuracy and robustness of state space model parameter analysis; the dynamic time-frequency decomposition and VAE feature enhancement mode improves the efficiency and stability of processing simple data through efficient data decomposition and feature extraction, and optimizes the process and results of state space model parameter analysis. The choice of these two modes is based on the size of the comprehensive feature coefficient to adapt to the needs of different data characteristics, so as to optimize the operation performance and stability of industrial equipment.

[0109] Step S5 includes the following:

[0110] S5.1, format the processed data into a deep learning model input format, generate label data, and divide the data set into training set, validation set and test set; arrange the grid cell data after feature enhancement into an input format that can be accepted by the deep learning model.

[0111] Arrange the processed data into a two-dimensional array, with each row representing a sample and each column representing a feature.

[0112] Generate corresponding label data according to actual needs, which is usually the state space model parameters of industrial flexible load (such as state matrix A, input matrix B, etc.).

[0113] Divide the data set into training set, validation set and test set, with a common ratio of 70% training set, 15% validation set and 15% test set.

[0114] S5.2, construct a deep learning model architecture containing input layer, convolutional layer, LSTM layer, fully connected layer and output layer to effectively extract and learn spatial and temporal features in data.

[0115] Define the input layer with the number of input nodes equal to the number of features.

[0116] Design multiple convolutional layers (e.g., Conv1D) to extract spatial features from the data. Each convolutional layer is followed by a max pooling layer (MaxPooling1D) to reduce the feature dimension.

[0117] Add LSTM layers to capture the time series features of the data. The LSTM layers can be stacked to increase the learning capacity of the model.

[0118] Add a dense layer (Dense) after the LSTM layers to further extract high-level features.

[0119] Define the output layer with the number of output nodes equal to the number of state space model parameters, using a linear activation function.

[0120] S5.3, Select Mean Squared Error as the loss function, use Adam optimizer, and set evaluation metrics to compile the deep learning model.

[0121] Select Mean Squared Error (MSE) as the loss function for regression problems.

[0122] Select Adam optimizer and set hyperparameters such as learning rate.

[0123] Select Mean Squared Error (MSE) as the evaluation metric.

[0124] S5.4, Train the model using training data, optimize network parameters through forward propagation, error calculation, and backpropagation, and evaluate and adjust on validation data to prevent overfitting.

[0125] Input the training set into the deep learning model, perform forward propagation, error calculation, and backpropagation, and adjust network weights and biases.

[0126] After each training cycle, evaluate the model performance using the validation set, monitor the loss value and evaluation metrics, and prevent overfitting.

[0127] Adjust the model's hyperparameters (such as learning rate, batch size, etc.) based on the performance of the validation set to optimize the model's performance.

[0128] S5.5, Evaluate the model's generalization ability using test data, verify the model's performance by calculating the loss value and evaluation metrics, and fine-tune the model based on the results.

[0129] S5.6, Preprocess and enhance the features of new data, input them into the trained model to predict state space model parameters, and analyze the prediction results to evaluate the actual application effect.

[0130] The new data is preprocessed and feature enhanced in the same way as the training data.

[0131] The preprocessed new data is input into the trained deep learning model, and the predicted state space model parameters are output.

[0132] The prediction results are analyzed to evaluate the actual application effect of the model, and corrections and optimizations are made as needed.

[0133] The present application collects and constructs a time series segmentation index repository based on load state segmentation, uses real-time load parameters to determine the load state, and stores data in grid cells when the state changes, achieving efficient organization and management of data, ensuring independent analysis of data under different load states, and avoiding complexity and errors caused by mixed data. The data in each grid cell is reconstructed in phase space, and the heterogeneous dynamic index is calculated to reveal the complex dynamic behavior and spatial heterogeneity of the load data, ensuring accurate evaluation of high dynamic and complex data. Through time-frequency analysis, the energy spectrum density is calculated and the time-frequency recurrence map is constructed, the recurrence rate and the diagonal line length distribution are calculated using cosine similarity, and the frequency domain recurrence index is calculated to comprehensively evaluate the dynamic complexity of the data, which helps to select the most suitable processing mode under different dynamic conditions. By integrating the heterogeneous dynamic index and the frequency domain recurrence index, the comprehensive feature coefficient is obtained, and the iterative feature screening and adversarial generative network enhancement mode or the dynamic time-frequency decomposition and variational coding network enhancement mode is flexibly selected to realize efficient feature enhancement and optimization processing of data, significantly improving the accuracy and reliability of state space model parameter analysis. For the grid cell data processed by the above method, the data is formatted and segmented, and a deep learning model is trained to predict and analyze the state space model parameters, providing an efficient and accurate parameter analysis method to improve the operating efficiency and stability of industrial flexible load under different conditions. By flexibly selecting a processing mode that adapts to different feature data, the present application effectively deals with complex industrial load data, optimizes the state space model parameter analysis process, and improves the overall performance of the industrial system.

[0134] To verify the effectiveness of the method, the state space model parameters (state matrix A, input matrix B) are predicted and verified using the motor load data of an industrial production line.

[0135] Step running early: In a three-phase asynchronous motor, define the state matrix variable: X(t) = [w, i a ], w is the motor speed (rad / s), representing the mechanical dynamic characteristics, i a is the armature current (A), representing the electrical dynamic characteristics; the input matrix variable: terminal voltage u (v), as an external control input.

[0136] The state space model parameters to be predicted are: state matrix A: describes the coupling relationship between speed and current and self-attenuation characteristics; input matrix B: describes the influence degree of voltage input on speed and current.

[0137] The first step is data collection and segmentation.

[0138] Data source: collect the current, voltage and speed time series data of a motor under light load (power < 30% rated power) and rated load (power 80%-100% rated power), sampling frequency 100Hz, total duration 2 hours;

[0139] Segmented storage: light load state data is stored in grid unit C1 (duration 40 minutes, including startup-light load steady state stage); rated load state data is stored in grid unit C2 (duration 80 minutes, including load switching-rated steady state stage).

[0140] The second step is feature calculation and mode selection.

[0141] Heterogeneous dynamic index (NFC) calculation:

[0142] C1 data: Lyapunov exponent α = 0.12, Shannon entropy H = 0.85, weighted sum NFC = 0.48;

[0143] C2 data: α = 0.35, H = 1.23, NFC = 0.79;

[0144] Frequency domain recurrence index (RPC) calculation:

[0145] C1 data: recurrence rate RR = 0.32, determinism DET = 0.65, RPC = 0.208;

[0146] C2 data: RR = 0.58, DET = 0.82, RPC = 0.476;

[0147] Comprehensive feature coefficient (DCAC) calculation:

[0148] ω1 = ω2 = 0.5 is preset, and the calculation is:

[0149] C1: DCAC = ln[(0.5*0.48+0.5*0.208) / (0.5+0.5)+1] = 0.53;

[0150] C2: DCAC = ln[(0.5*0.79+0.5*0.476) / (0.5+0.5)+1] = 0.89;

[0151] Mode selection:

[0152] C1 (DCAC=0.53 < threshold 0.7): dynamic time-frequency decomposition and variational autoencoder network enhanced mode;

[0153] C2 (DCAC=0.89 >= threshold 0.7): iterative feature selection and generative adversarial network enhanced mode.

[0154] Third step, feature enhancement and model training.

[0155] C1 data processing: dynamic time-frequency decomposition, short-time Fourier transform (window length 10 seconds) to generate energy spectrum density matrix; variational autoencoder network: compress high-dimensional time-frequency features to 50-dimensional latent space, generate enhanced feature vector;

[0156] C2 data processing: iterative feature selection, based on random forest algorithm to iteratively delete 20% redundant features, and retain 80-dimensional key features; generative adversarial network, generate 5000 enhanced samples consistent with the original data distribution;

[0157] Model architecture: input layer, 100-dimensional features (light load) / 80-dimensional features (rated load); product layer, 1D convolution kernel x2, extract spatial features; LSTM layer, 128 units, capture time sequence features; output layer, 2-dimensional (state matrix A, input matrix B), linear activation;

[0158] Training configuration: loss function, mean square error (MSE); optimizer, Adam (learning rate 0.001); training period, 50 epochs, batch size 32.

[0159] Fourth step, parameter prediction and verification.

[0160] Test data: collect real-time data of the motor in overload state (power 120% rated power), input the segmented data into the trained model;

[0161] Prediction results: state matrix A predicted value: [[-0.85, 0.32], [0.15, -0.91]], true value [[-0.82, 0.35], [0.18, -0.89]], MSE=0.005; input matrix B predicted value: [0.45, 0.62], true value [0.48, 0.60], MSE=0.003;

[0162] The method of the present application can accurately predict the state space model parameters under different load conditions, and verifies the effectiveness and robustness of the method.

[0163] The above formulas are all dimensionless values calculated, the formula is obtained by collecting a large amount of data to simulate the recent real situation, and the preset parameters in the formula are set by the person skilled in the art according to the actual situation; in the true value and the predicted value, positive and negative numbers represent the attenuation or feedback mechanism of the system, such as positive feedback and negative feedback, for example, in the motor load model, the negative number represents damping or energy consumption; for the state matrix A, the element is negative, which is common in a stable system, because the negative eigenvalue indicates that the system state will decay over time, which meets the stability requirements of the actual physical system, for example, the speed or current change rate of the motor may show a decay trend; the elements of the input matrix B are positive, which usually represent the direction of the influence of the input on the state, and the positive number represents positive excitation, which meets the physical meaning of the actual control input;

[0164] Matrix state A:

[0165] The diagonal elements are all negative, indicating that the system state (such as speed, current) has a self-decay characteristic, which meets the natural shutdown process of the motor when there is no input; the positive non-diagonal elements represent the positive coupling between state variables (such as the influence of voltage on current);

[0166] Input matrix B: true B = [0.48 0.60], predicted B = [0.45 0.62];

[0167] The positive number represents the positive control action of the external input (such as the voltage command) on the state variable, and the numerical difference is within a reasonable error range.

[0168] The industrial flexible load state space model parameter prediction system described in the application comprises:

[0169] A data collection module is used to obtain time series load parameters of industrial equipment under different flexible load states and store them in corresponding grid cells;

[0170] A spatial heterogeneity analysis module is used to reconstruct the phase space of the time series load parameters in each grid cell, calculate the Lyapunov exponent and Shannon entropy for the reconstructed vector, and calculate the heterogeneity dynamic index according to the Lyapunov exponent and Shannon entropy;

[0171] A frequency domain dynamic analysis module is used to perform time-frequency analysis on the time series load parameters in each grid cell, construct a time-frequency recurrence atlas by calculating the similarity of energy spectrum density, calculate the recurrence rate and diagonal line length distribution in the time-frequency recurrence atlas, and calculate the frequency domain recurrence index according to the recurrence rate and diagonal line length distribution;

[0172] An intelligent decision-making module is used to calculate a comprehensive feature coefficient according to the heterogeneity dynamic index and the frequency domain recurrence index;

[0173] The feature enhancement module is used for selecting a corresponding feature enhancement method according to the size relationship between the comprehensive feature coefficient and the threshold value, and performing feature enhancement on the time sequence load parameter in each grid unit.

[0174] The parameter prediction module is used for inputting the time sequence load parameter in all grid units after being formatted into the trained deep learning model to obtain the predicted state space model parameter.

[0175] The electronic device comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, and the computer program realizes the industrial flexible load state space model parameter prediction method when being loaded to the processor.

[0176] The computer readable storage medium stores the computer program, and the computer program realizes the industrial flexible load state space model parameter prediction method when being executed by the processor.

[0177] The computer readable storage medium can include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, flash memory or any other medium that can be used to store program codes in the form of instructions or data structures and can be accessed by a computer.

[0178] The processor is used for executing the computer program stored in the memory to realize each step in the method related to the above-mentioned embodiments.

Claims

1. An industrial flexible load state space model parameter prediction method, characterized by, Includes the following steps: Acquire the time-series load parameters of industrial equipment under different flexible load conditions and store them in the corresponding grid cells; Phase space reconstruction is performed on the temporal load parameters in each grid cell. The Lyapunov exponent and Shannon entropy are calculated for the reconstructed vectors. The heterogeneous dynamics exponent is calculated based on the Lyapunov exponent and Shannon entropy. Time-frequency analysis is performed on the temporal load parameters in each grid cell, and a time-frequency recurrence map is constructed by calculating the similarity of energy spectral density; the recurrence rate and diagonal length distribution are calculated in the time-frequency recurrence map, and the frequency domain recurrence index is calculated based on the recurrence rate and diagonal length distribution; The composite feature coefficients are calculated based on the heterogeneous dynamic index and the frequency domain recurrence index. The corresponding feature enhancement method is selected based on the relationship between the composite feature coefficients and their thresholds, and the time-series load parameters in each grid cell are enhanced. The temporal load parameters in all feature-enhanced grid cells are formatted and input into the trained deep learning model to obtain the predicted state-space model parameters.

2. The industrial flexible load state space model parameter prediction method according to claim 1, characterized in that, The process of acquiring time-series load parameters of industrial equipment under different flexible load states and storing them in the corresponding grid cells includes: The current load status is determined based on the load parameters of the industrial equipment. The load parameters of the industrial equipment are stored in grid cells. When the load status changes, the load parameters for the corresponding time period are stored in a new grid cell, forming a segmented storage of load parameters for continuous load status.

3. The industrial flexible load state space model parameter prediction method of claim 1, wherein, The step of reconstructing the phase space of the temporal load parameters in each grid cell, calculating the Lyapunov exponent and Shannon entropy for the reconstructed vector, and calculating the heterogeneous dynamics exponent based on the Lyapunov exponent and Shannon entropy includes the following steps: For the time-series load parameters in each grid cell, the phase space is reconstructed using a delay embedding method. By fixing the delay time and embedding dimension, the time-series load parameters are mapped to a high-dimensional space to obtain the load parameter vector. For the load parameter vector, the Lyapunov exponent is obtained by calculating the divergence rate of adjacent points, and the Shannon entropy is obtained by calculating the probability distribution. The heterogeneous dynamics index is obtained by weighted summation of the Lyapunov exponent and the Shannon entropy.

4. The method for predicting parameters of an industrial flexible load state-space model according to claim 3, characterized in that, Lyapunov exponent where M is the total number of data points in the time series load parameter; d i (t) is the distance of the i-th data point at time t; Δt is the time interval used to calculate the divergence rate of adjacent points Shannon entropy H = -∑ i P(a i ) log P(a i ), where P(a i ) is the probability distribution of the load parameter vector, a i denotes the i-th load parameter vector.

5. The industrial flexible load state space model parameter prediction method of claim 1, wherein, The process of performing time-frequency analysis on the temporal load parameters in each grid cell, constructing a time-frequency recurrence map by calculating the similarity of energy spectral density, and calculating the recurrence rate and diagonal length distribution in the time-frequency recurrence map, and calculating the frequency domain recurrence index based on the recurrence rate and diagonal length distribution includes the following steps: Time-frequency analysis was performed on the temporal load parameters in each grid cell to calculate the energy spectral density at each time window and frequency. The similarity of energy spectral density is calculated by cosine similarity to construct a time-frequency recurrence map; For time-frequency recurrence patterns, the recurrence rate is obtained by calculating the density of recurrence points, and the determinism is obtained by calculating the diagonal length distribution and statistically analyzing the frequency of the diagonal length. The frequency domain recurrence index is obtained by multiplying the recursion rate and the determinism.

6. The industrial flexible load state space model parameter prediction method according to claim 5, characterized in that, deterministic where l denotes the length of the diagonal; f(l) is the frequency of the diagonal length l, l min is the minimum value of the diagonal length, l max is the maximum value of the diagonal length, R(k, v) indicates whether the time t k and t v is similar, and 1 otherwise.

7. The industrial flexible load state space model parameter prediction method of claim 1, wherein, The method comprises the following steps: The comprehensive feature coefficient is calculated according to the heterogeneous dynamic index and the frequency domain recurrence index, and a corresponding feature enhancement method is selected according to the size relationship between the comprehensive feature coefficient and its threshold value.

8. The industrial flexible load state space model parameter prediction method according to claim 1 or 7, characterized by, Comprehensive characteristic coefficient wherein NFC and RPC are the isomerization dynamic index and the frequency domain recurrence index, respectively; ω1 and ω2 are preset proportionality coefficients of the isomerization dynamic index and the frequency domain recurrence index, respectively, and are both greater than zero.

9. An industrial flexible load state space model parameter prediction system characterized by, The heterogeneous dynamic index and the frequency domain recurrence index are weighted and summed to obtain a comprehensive feature coefficient, if the comprehensive feature coefficient is not less than its threshold value, then the iterative feature screening and the generative adversarial network enhancement mode are adopted; otherwise, the dynamic time-frequency decomposition and the variational coding network enhancement mode are adopted. The method comprises the following steps: The data collection module is used for acquiring the time sequence load parameters of the industrial equipment under different flexible load states and storing the time sequence load parameters into corresponding grid cells; The spatial heterogeneity analysis module is used for phase space reconstruction on the time sequence load parameters in each grid cell, calculating the Lyapunov exponent and the Shannon entropy for the reconstructed vectors, and calculating the heterogeneous dynamic index according to the Lyapunov exponent and the Shannon entropy; The frequency domain dynamic analysis module is used for time-frequency analysis on the time sequence load parameters in each grid cell, constructing a time-frequency recurrence atlas by calculating the similarity of energy spectrum density, calculating the recurrence rate and the diagonal line length distribution in the time-frequency recurrence atlas, and calculating the frequency domain recurrence index according to the recurrence rate and the diagonal line length distribution; The intelligent decision module is used for calculating a comprehensive feature coefficient according to the heterogeneous dynamic index and the frequency domain recurrence index; The feature enhancement module is used for selecting a corresponding feature enhancement method according to the size relationship between the comprehensive feature coefficient and its threshold value, and performing feature enhancement on the time sequence load parameters in each grid cell; 10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The parameter prediction module is used for inputting the time sequence load parameters in all grid cells after feature enhancement into a trained deep learning model in a formatted manner, and obtaining predicted state space model parameters.

11. A computer-readable storage medium storing a computer program, wherein the computer program comprises the following steps of: receiving a request for a resource from a client; determining whether the client is authorized to access the resource; and if the client is authorized to access the resource, providing the resource to the client. The computer program is loaded into the processor to realize the industrial flexible load state space model parameter prediction method according to any one of claims 1-8. The computer program is executed by the processor to realize the industrial flexible load state space model parameter prediction method according to any one of claims 1-8.

Images