Heterogeneous data processing method and system for energy big data

Abstract

The invention discloses a heterogeneous data processing method and system for energy big data, and the method comprises the steps: obtaining different types of energy data, inputting the different types of energy data into a pre-trained directed association graph model, and obtaining a cross-type dynamic contribution flow network node forest; a cross-type linkage report is obtained, and a basic directed positioning dynamic correction chain is constructed through report analysis positioning and causal analysis; and inputting the cross-type dynamic contribution flow network node forest and the basic directed positioning dynamic correction chain into a pre-trained positioning correction path model, and performing real-time positioning error correction adjustment on the cross-type linkage report to obtain an updated and adjusted cross-type linkage report. According to the invention, cross-type linkage analysis and report real-time error correction of the energy data are realized, and the accuracy and efficiency of energy data processing are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of energy data processing, and relates to a heterogeneous data processing method and system for energy big data. BACKGROUND

[0002] The current energy data center converges multi-source heterogeneous data, covering structured, semi-structured, unstructured and other forms, and there is a significant semantic gap between various energy subsystems, which leads to constraints such as data islandization, missing dynamic association and decision-making delay accumulation in cross-type data collaboration.

[0003] The existing energy data processing method has bottlenecks such as loss of heterogeneous data value, distortion of dynamic contribution evaluation, and lack of correction mechanism. The energy correlation law is fragmented, making it difficult to trace the wind and light abandoned electricity event, and the traditional factor analysis ignores the cross-type conduction of main-cooperative indicators, resulting in large correlation prediction errors between different types of energy data, so that the adjustment of the report can only rely on historical experience rules, and cannot dynamically correct the multi-energy chain imbalance, which is difficult to meet the accurate and linkage self-correction needs of the energy prediction result, and is difficult to provide accurate and timely analysis results for energy decision-making. SUMMARY

[0004] To solve the problems in the prior art, the application provides a heterogeneous data processing method and system for energy big data, which realizes cross-type linkage analysis and report real-time correction of energy data, and improves the accuracy and efficiency of energy data processing.

[0005] The application adopts the following technical solutions.

[0006] The application provides a heterogeneous data processing method for energy big data, which includes: Obtain different types of energy data and input a pre-trained directed association graph model to obtain a cross-type dynamic contribution flow network node forest; Obtain a cross-type linkage report and build a basic directed positioning dynamic correction chain through report analysis positioning and causal analysis; Input the cross-type dynamic contribution flow network node forest and the basic directed positioning dynamic correction chain into a pre-trained positioning correction path model to perform real-time positioning correction and adjustment on the cross-type linkage report, and obtain an updated and adjusted cross-type linkage report.

[0007] Preferably, the training process of the directed association graph model includes: Obtain each type of energy historical information and extract key variables to obtain an energy parameter time series set and an association change attribute set of each type of energy; performing multi-stage time-series factor analysis on the energy parameter time series set to obtain main index variables and corresponding collaborative index variable sequences, instantaneous main contribution degrees of the main index variables to the target result, instantaneous collaborative contribution degrees of the collaborative index variable sequences to the corresponding main index variables, and covariant correlation coefficients between the collaborative index variables; repeating the process of obtaining the instantaneous main contribution degrees and the instantaneous collaborative contribution degrees to obtain change paths of the instantaneous main contribution degrees and the instantaneous collaborative contribution degrees, and combining the covariant correlation coefficients between the collaborative index variables to construct a dynamic contribution flow network node tree corresponding to a single type of energy data; performing causal correlation analysis on the associated change attribute set to obtain joint contribution influence factors of the main index variables to the target result; based on the joint contribution influence factors and the dynamic contribution flow network node trees corresponding to all types of energy data, constructing a cross-type dynamic contribution flow network node forest; constructing a reverse comprehensive training function and performing reverse reasoning training by analyzing the process of obtaining the cross-type dynamic contribution flow network node forest, to obtain a trained directed association graph model.

[0008] Preferably, the obtaining of each type of energy historical information and the extraction of key variables, the obtaining of the energy parameter time series set and the associated change attribute set of each type of energy, comprises: obtaining each type of energy historical demand quantity change information, cost change information, and each type of energy associated change information and performing key variable extraction to obtain the energy parameter time series set and the associated change attribute set of each type of energy.

[0009] Preferably, the multi-stage time-series factor analysis on the energy parameter time series set to obtain main index variables and corresponding collaborative index variable sequences, instantaneous main contribution degrees of the main index variables to the target result, instantaneous collaborative contribution degrees of the collaborative index variable sequences to the corresponding main index variables, and covariant correlation coefficients between the collaborative index variables, comprises: based on the energy parameter time series set of each type of energy, obtaining a parameter loading matrix corresponding to each type of energy parameter continuous time point through a multi-stage time-series factor analysis model; performing statistical analysis on the parameter loading matrix corresponding to each time point to obtain each type of main index variable and corresponding collaborative index variable sequence, instantaneous main contribution degree of each main index variable to the target result, instantaneous collaborative contribution degree of each collaborative index variable sequence to the corresponding main index variable, and covariant correlation coefficient between the collaborative index variables.

[0010] Preferably, the repeated acquisition process of the instantaneous main contribution degree and the instantaneous synergistic contribution degree obtains the instantaneous main contribution degree change path and the instantaneous synergistic contribution degree change path, combines the covariant correlation coefficient between the synergistic index variables, and constructs a dynamic contribution flow network node tree corresponding to single-type energy data, including: The acquisition process of the instantaneous main contribution degree and the instantaneous synergistic contribution degree is repeated to obtain the instantaneous main contribution degree change path of each main index variable to the target result and the instantaneous synergistic contribution degree change path of each synergistic index variable to the corresponding main index variable at a continuous time point; The target result is taken as a target node of a first layer, the main index variable is taken as a main contribution node of a second layer, and the synergistic index variable sequence corresponding to each main index variable is taken as a synergistic contribution node set of a third layer; Based on the instantaneous main contribution degree change path of each main contribution node corresponding to each type of energy data to the target node, the instantaneous synergistic contribution degree change path of each synergistic contribution node to the corresponding main index variable, and the covariant correlation coefficient between the same synergistic contribution node set, a dynamic contribution flow network node tree corresponding to single-type energy data at a continuous time point is constructed through a graph neural network combined with a graph database.

[0011] Preferably, the causal correlation analysis on the associated change attribute set obtains a joint contribution influence factor of the main index variable to the target result, including: Based on the associated change attribute set of all types of energy parameters at each time point, the main joint contribution degree of all main index variables corresponding to each target result to the same target result and the associated influence relationship and the directed associated influence direction between all main index variables corresponding to each target result, and the directed causal correlation coefficient between the synergistic index variable sequences corresponding to different main index variables are obtained through a causal correlation analysis algorithm within a single time; The joint contribution influence factor of each main index variable to the same target result at a single time point is obtained based on the instantaneous main contribution degree of a single main index variable to the target result divided by the main joint contribution degree of all main index variables corresponding to each target result to the same target result within a single time.

[0012] Preferably, the dynamic contribution flow network node forest across types is constructed based on the joint contribution influence factor and the dynamic contribution flow network node tree corresponding to all types of energy data, including: The main contribution relationship connection between each main contribution node and the corresponding target node is constructed based on the joint contribution influence factor of each main index variable to the same target result at a single time point. Based on the correlation influence relationship and the directional correlation influence direction between all main indicator variables, the directional correlation connection between the main contribution nodes is constructed through a blockchain algorithm, and the cross-tree cooperative variable correlation connection is constructed based on the directional causality correlation coefficient between the sequences of different main indicator variables corresponding to the cooperative indicator variables; The dynamic contribution flow network node tree corresponding to all types of energy data is timestamped at the continuous time point, and after alignment, the topological space algorithm is combined to construct a cross-type dynamic contribution flow network node forest.

[0013] Preferably, the training process of the positioning correction path model comprises: Based on the cross-type dynamic contribution flow network node forest, the reverse comprehensive training function is constructed based on the corresponding instantaneous contribution degree mean square error, the instantaneous main contribution degree change path loss, the main indicator correlation relationship cross entropy loss, the cross-tree cooperative variable causality relationship loss, the continuous timestamp alignment loss, the complexity penalty of the number of main contribution nodes and corresponding cooperative contribution nodes, and the constraint penalty function of the cross-type dynamic contribution flow network node forest corresponding to the correlation strength less than the preset correlation coefficient. Based on the reverse comprehensive training function combined with the cross-type dynamic contribution flow network node forest, the trained directional correlation graph model is obtained through the reverse reasoning training from the third layer of cooperative contribution nodes to the first layer of target nodes by combining the random forest with the simulation algorithm.

[0014] Preferably, the training process of the positioning correction path model comprises: Obtain the key target result sequence in the table and the relationship of the key target result synchronous change, the position of the key target result in the table and its contribution impact factor through report analysis positioning, and obtain the key target result and the corresponding causality correlation strength through causality analysis. Based on the position of the key target result in the table, the key target result and the corresponding causality correlation strength, a training optimized directional positioning dynamic correction chain is constructed, and is mapped and similarity analyzed with the cross-type dynamic contribution flow network node forest to obtain a similarity consistency error. By analyzing the similarity consistency error, the corresponding parameter relationship between the training optimized directional positioning dynamic correction chain and the cross-type dynamic contribution flow network node forest, a forward comprehensive loss function is constructed and forward training is performed to obtain a trained positioning correction path model, and the directional correlation graph model is trained in real time and online based on the loss result of the forward training.

[0015] Preferably, the historical cross-type linkage report information is obtained, and the key target result sequence in the table and the relationship of the synchronous change of the key target result are located by report analysis, the position of the key target result in the table and its contribution impact factor are obtained, and through causal analysis, the key target result with correlated change relationship and the corresponding causal correlation strength are obtained, including: The historical cross-type linkage report information is obtained, and the key target result sequence in the cross-type linkage report and the relationship of the synchronous change of the remaining key target result when each key target result produces change, the position of each key target result in the cross-type linkage report, and the corresponding contribution impact factor of each key target result are obtained through a preset report analysis positioning model; Based on the key target result sequence, the relationship of the synchronous change of the remaining key target result when each key target result produces change, and the corresponding contribution impact factor of each key target result, the key target result with correlated change relationship and the corresponding causal correlation strength are obtained through a causal analysis algorithm.

[0016] Preferably, the training and optimization of the directed positioning dynamic correction chain is constructed based on the position of the key target result in the table, the key target result with correlated change relationship and the corresponding causal correlation strength, and is mapped and similarity analyzed with the cross-type dynamic contribution flow network node forest to obtain a similarity consistency error, including: The training and optimization of the directed positioning dynamic correction chain is constructed based on the position of the key target result in the table, the key target result with correlated change relationship and the corresponding causal correlation strength, and is mapped and similarity analyzed with the cross-type dynamic contribution flow network node forest to obtain a similarity consistency error, including: Each key target result in the training and optimization of the directed positioning dynamic correction chain is one-to-one mapped with the target node at the same time in the cross-type dynamic contribution flow network node forest, and according to the corresponding contribution impact factor of each key target result and the main contribution node and the set of collaborative contribution nodes at the same time point in the cross-type dynamic contribution flow network node forest, the similarity coefficient between the contribution impact factor and the main contribution node and the collaborative contribution node and the similarity coefficient between the contribution impact factor and the collaborative contribution node are obtained through a similarity algorithm; The similarity mapping of the contribution impact factor and the main contribution node and the collaborative contribution node is established according to the obtained similarity coefficient, and a similarity consistency error is constructed.

[0017] Preferably, the forward comprehensive loss function is constructed by analyzing the similarity consistency error, the corresponding parameter relationship between the training and optimization of the directed positioning dynamic correction chain and the cross-type dynamic contribution flow network node forest, and the forward training is performed to obtain a trained positioning correction path model, and the directed association graph model is trained in real time and online based on the loss result of the forward training, including: The error between the key target result of the training optimized directional positioning dynamic correction chain and the reverse target result of the same target node obtained by the reverse reasoning of the cross-type dynamic contribution flow network node forest is combined with the consistent error, and a forward comprehensive loss function is constructed by combining the error between the target result of each key target result in the training optimized directional positioning dynamic correction chain after the synchronous variation of the remaining key target result and the reverse target result of the remaining key target result in the same target node obtained by the reverse reasoning of the cross-type dynamic contribution flow network node forest, and the delay of the synchronous variation of the remaining key target result. The forward training is performed from the first layer corresponding to the target node to the third layer corresponding to the cooperative contribution node set, and a trained positioning correction path model is obtained. The forward comprehensive loss result obtained in each step of the positioning correction path model training is fed back to the directional association graph model for synchronous reverse adjustment training, and a real-time synchronous online trained directional association graph model is obtained.

[0018] The second aspect of the present application provides a heterogeneous data processing system for energy big data, comprising: A directional association module is configured to obtain different types of energy data and input a pre-trained directional association graph model to obtain a cross-type dynamic contribution flow network node forest. A causal analysis module is configured to obtain a cross-type linkage report and construct a basic directional positioning dynamic correction chain through report analysis positioning and causal analysis. A positioning correction module is configured to input the cross-type dynamic contribution flow network node forest and the basic directional positioning dynamic correction chain into a pre-trained positioning correction path model, and perform real-time positioning error correction adjustment on the cross-type linkage report to obtain an updated and adjusted cross-type linkage report.

[0019] The third aspect of the present application provides a terminal comprising a processor and a storage medium; the storage medium is configured to store instructions; and the processor is configured to operate according to the instructions to perform the steps of the method.

[0020] The fourth aspect of the present application provides a computer readable storage medium having a computer program stored thereon, which is executed by a processor to implement the steps of the method.

[0021] Compared with the prior art, the present application has at least the following beneficial effects: The present application forms a cross-type dynamic contribution flow network node forest through a pre-trained directional association graph model and a positioning correction path model, realizes deep fusion and dynamic association analysis of multi-source heterogeneous energy data, can eliminate data islands, establishes a reliable association mapping of cross-type energy data, accurately captures the nonlinear time-varying law of multi-energy coupling, and realizes cross-type conduction analysis of main-cooperative index dynamic contribution. The application improves the learning ability of the directed association graph model on the energy data association rule by comprehensively training the function and the reverse reasoning training, realizes the real-time error correction of the cross-type linkage report by combining the positioning correction path model, and dynamically corrects the multi-energy chain imbalance; The application provides accurate and linkage self-correction ability analysis results for energy prediction by means of the online synchronous training mechanism, adapts to the dynamic change of the energy system, and improves the accuracy and timeliness of energy decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 A heterogeneous data processing method flow chart of energy big data according to the application; Figure 2 An implementation flow chart of the heterogeneous data processing method of energy big data according to the application; Figure 3 A directed association graph model construction architecture diagram according to the application; Figure 4 A positioning correction path model construction architecture diagram according to the application; Figure 5 A heterogeneous data processing system principle diagram of energy big data according to the application. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical scheme and advantages of the application clearer, the technical scheme of the application will be described clearly and completely below in combination with the drawings in the embodiments of the application. The embodiments described in the application are only a part of the embodiments of the application, not all the embodiments. All other embodiments obtained by those skilled in the art without creative labor based on the spirit of the application belong to the protection scope of the application.

[0024] Embodiment 1 of the application provides a heterogeneous data processing method of energy big data, a directed association graph model is constructed by acquiring different types of energy data and combining an association algorithm and a graph database, and then a cross-type dynamic contribution flow network node forest is obtained; meanwhile, a cross-type linkage report is acquired, key target results and causal association strength existing in the association change relationship are extracted therefrom, a directed positioning dynamic correction chain is constructed by combining a blockchain algorithm and position information; finally, a positioning correction path model is constructed based on the cross-type dynamic contribution flow network node forest, the directed positioning dynamic correction chain and a text positioning algorithm, real-time positioning error correction adjustment of the cross-type linkage report is realized, and an updated cross-type linkage report is obtained, which is used for real-time positioning correction of multi-source heterogeneous data in coal, oil, gas, electricity, heat and other energy fields, such as Figures 1-2 As shown in the figure, it comprises: S1: acquiring different types of energy data and inputting a pre-trained directed association graph model to obtain a cross-type dynamic contribution flow network node forest; More preferably, different types of energy data are acquired, and a directed association graph model constructed by association algorithms and graph databases is combined to obtain a cross-type dynamic contribution flow network node forest; The directed relational graph model is constructed and trained by combining similar energy data with a relational algorithm and a graph database, such as... Figure 3 As shown, the process includes: (1) Obtain information on various types of energy and extract key variables to obtain the time series set of energy parameters and the set of associated change attributes for each type of energy: It acquires historical demand change information, cost change information, and related change information of various types of energy, and extracts key variables through a preset variable analysis model to obtain the time series set of parameters and related change attribute set of each type of energy. It should be further explained that the specific process for obtaining the time series set of each type of energy parameter and the associated change attribute set in this embodiment includes: First, clean and preprocess heterogeneous data from multiple sources such as coal, oil, gas, electricity, and heat. For missing values, use time series interpolation or rule filling based on domain knowledge. For outliers, use isolated forest or statistical thresholding methods to detect and correct them. Second, wavelet transform is used to perform multi-resolution analysis on time series data to extract features at different time scales; Third, mutual information and random forest feature importance scores are used to screen variables that are strongly correlated with energy demand and cost; Fourth, the Granger causality test is used to identify the direction and strength of causal relationships between variables. For the set of associated change attributes, the dynamic conditional covariate correlation coefficient is used to capture the time-varying correlation between parameters. Combined with the Copula function, the nonlinear dependency structure is characterized, and finally, a time series set of energy parameters and a set of associated change attributes containing multi-dimensional attributes such as energy price fluctuations and supply and demand relationships are formed.

[0025] (2) Perform multi-stage time-series factor analysis on the energy parameter time series set to obtain the main indicator variable and the corresponding co-indicator variable sequence, the instantaneous main contribution of the main indicator variable to the target result, the instantaneous co-contribution of the co-indicator variable sequence to the corresponding main indicator variable, and the covariate correlation coefficient between the co-indicator variables: Based on the time series set of each type of energy parameter, the parameter load matrix corresponding to the continuous time points of each type of energy parameter is obtained through a multi-stage time series factor analysis model. The implementation process of the multi-stage time series factor analysis model in the embodiment includes: performing time continuity test and outlier processing on the energy parameter time series set of each type of energy to obtain standardized energy parameter time series; dividing the standardized energy parameter time series into continuous time windows according to a preset time interval, performing Kaiser criterion test and factor rotation processing on the energy parameter time series in each time window to determine the number of common factors in the corresponding time window and the correlation between each common factor and the energy parameter; based on the correlation between the common factors in each time window and the energy parameter, calculating the loading coefficient of the energy parameter to the common factor in each time window by the maximum likelihood estimation method, arranging the loading coefficient in each time window according to the corresponding relationship between the energy parameter and the common factor to obtain the initial parameter loading matrix corresponding to each time window; performing smoothness correction on the initial parameter loading matrices of adjacent time windows to eliminate the sudden change of the loading coefficient caused by time window division, and finally obtaining the parameter loading matrix corresponding to each type of energy parameter at continuous time points.

[0026] For each time point corresponding parameter loading matrix, statistical analysis is performed to obtain each type of main indicator variable and corresponding collaborative indicator variable sequence, the instantaneous main contribution degree of each main indicator variable to the target result, the instantaneous collaborative contribution degree of each collaborative indicator variable sequence to the corresponding main indicator variable, and the covariate correlation coefficient between the collaborative indicator variables.

[0027] It should be further pointed out that the specific process of obtaining each type of main indicator variable and corresponding collaborative indicator variable sequence, the instantaneous main contribution degree of each main indicator variable to the target result, the instantaneous collaborative contribution degree of each collaborative indicator variable sequence to the corresponding main indicator variable, and the covariate correlation coefficient between the collaborative indicator variables in the embodiment includes: First, based on the original multi-source heterogeneous data, the factor rotation is performed on the parameter loading matrix by the maximum likelihood estimation method, the number of principal components is determined by combining the eigenvalue greater than 1 criterion and the scree plot test, and the variables with absolute value greater than 0.8 of the loading coefficient are selected as the main indicator variables, and the remaining variables constitute the collaborative indicator variable sequence; Second, based on the main indicator variable and the target result data, the standardized regression coefficient is calculated by the partial least squares regression, the coefficient significance is verified by the Bootstrap resampling method, and the instantaneous main contribution degree of the main indicator variable to the target result (i.e. the standardized regression coefficient) is obtained; Third, based on the collaborative indicator variable and the main indicator variable data, the path coefficient is calculated by constructing a structural equation model, the RMSEA (root mean square error approximation) and the CFI (comparative fit index) are ensured to meet the corresponding set threshold by combining multi-group confirmatory factor analysis, and the instantaneous collaborative contribution degree of the collaborative indicator variable to the main indicator variable (i.e. the path coefficient) is obtained; Fourth, based on the sequence of coordination index variables, the time-varying correlation matrix is calculated by the cosine correlation algorithm, and the correlation pairs with a significance greater than 0.95 are screened by combining the Bayesian model average method to obtain the significant covariant correlation coefficient matrix between the coordination index variables. Fifth, based on the above instantaneous main contribution degree, instantaneous coordination contribution degree and covariant correlation coefficient matrix, a structured feature set containing the main-coordination index hierarchical relationship, contribution degree matrix and correlation network is constructed.

[0028] (3) Repeat the process of obtaining the instantaneous main contribution degree and the instantaneous coordination contribution degree to obtain the instantaneous main contribution degree change path and the instantaneous coordination contribution degree change path, and combine the covariant correlation coefficient between the coordination index variables to construct the dynamic contribution flow network node tree corresponding to the single type of energy data: Repeat the process of obtaining the instantaneous main contribution degree and the instantaneous coordination contribution degree to obtain the instantaneous main contribution degree change path and the instantaneous coordination contribution degree change path, and combine the covariant correlation coefficient between the coordination index variables to construct the dynamic contribution flow network node tree corresponding to the single type of energy data: It needs to be further explained that the specific process of obtaining the instantaneous main contribution degree change path of each main index variable to the target result and the instantaneous coordination contribution degree change path of each coordination index variable to the corresponding main index variable at consecutive time points includes: First, based on historical data, segment by setting a sliding time window with a window length of N time points and a step length of 1, repeat the instantaneous contribution degree calculation process for each window, and smooth the contribution degree value by using local weighted regression to obtain the continuous contribution degree sequence after noise suppression; Second, based on the instantaneous main contribution degree of each main index variable to the target result at consecutive time points, model by hidden Markov model and divide the state into high, medium and low three states, calculate the state duration and transition probability, and obtain the state transition mode of the main index contribution degree; It needs to be further explained that the high, medium and low three states are divided by the person skilled in the art according to historical information and the specific needs, which will not be repeated here. When constructing the dynamic contribution flow network node tree corresponding to the single type of energy data, the time sequence features need to be integrated based on the main contribution node corresponding to each type of energy data and the instantaneous main contribution degree change path to the target node. The state transition mode of the main index contribution degree (the duration and transition probability of the high, medium and low three states) is one of the core dynamic features of the instantaneous main contribution degree change path, and its specific use is: The time sequence feature input graph neural network as the main contribution node: when capturing the time sequence feature through the graph convolution network combined with the gated recurrent unit, the state transition pattern is converted into quantitative features (such as the transition probability from high to medium state, the average duration of a certain state), which together with the numerical change of the instantaneous main contribution degree constitute the dynamic attributes of the main contribution node, helping the model to more accurately learn the evolution law of the main indicator contribution degree at continuous time points, avoiding the feature loss caused by relying only on numerical fluctuations and ignoring state change trends; Optimizing the edge weight calculation of the dynamic contribution flow network node tree: when obtaining the dynamic weight matrix between nodes through the Transformer multi-head attention mechanism, the state transition pattern will be used as a reference for attention allocation. For example, when the main indicator contribution degree transitions from high to medium state with a high transition probability, the decay rate of the edge weight between the main contribution node and the target node will be adjusted accordingly, so that the edge weight not only reflects the instantaneous contribution degree, but also reflects the transition trend of the contribution degree state, enhancing the representation ability of the node tree for the dynamic change of the main indicator contribution degree; Ensuring the time sequence continuity of the multi-version graph structure index: when building a multi-version graph structure index with timestamp as the index key through the Neo4j graph database, the state transition pattern will be stored together with the node tree structure at the corresponding time point. For example, the high-to-low state transition record within a certain time window is associated with the node tree of that window, so that when querying or backtracking the change of the main indicator contribution degree, the key time node (such as the state mutation point) corresponding to the node tree version can be quickly located through the state transition pattern, ensuring that the node tree can fully present the evolution process of the main indicator contribution degree from high to medium to low state, rather than isolated time point structure.

[0029] Third, based on the instantaneous collaborative contribution degree of each collaborative indicator variable to the corresponding main indicator variable at continuous time points, the time sequence features are extracted through the variational autoencoder and the mutation points are screened according to the reconstruction error, obtaining the feature sequence marked with key change points; Fourth, based on long-period energy data of different types, the seasonal decomposition algorithm is used to separate the trend item and the seasonal item, and the Kalman filter is used to interpolate and complete the missing time points, obtaining the de-seasonalized contribution degree trend sequence; Fifth, based on the time sequence continuity of each segment path of the de-seasonalized contribution degree trend sequence, the dynamic time warping (DTW) algorithm is used to align the contribution degree changes across windows, obtaining the dynamic change path of the main indicator variable and the collaborative indicator variable's contribution degree that integrates the trend component, the periodic component, and the mutation point marker.

[0030] It needs to be further explained that the periodic component is obtained from the previous time series preprocessing step for the instantaneous main contribution degree of the main indicator variable. Specifically, the second step has modeled the instantaneous main contribution degree of each main indicator variable to the target result at continuous time points (i.e., the main contribution degree time series) through a hidden Markov model, and the de-seasoned contribution degree trend sequence in the fifth step needs to be generated based on the main contribution degree time series through de-seasoning preprocessing; in the de-seasoning process, the original main contribution degree time series will be decomposed into two parts: one part is the trend component after removing the periodic fluctuations (i.e., the main body of the de-seasoned contribution degree trend sequence), and the other part is the periodic component separated, which has a fixed time interval rule (such as seasonal period, time period, etc. Periodic fluctuation characteristics), the periodic component obtained by this decomposition is the periodic component corresponding to the fusion of the trend component, the periodic component and the mutation point marker. The acquisition process of the periodic component in this embodiment is: Based on the seasonal term separated from the long-period different types of energy data based on the seasonal decomposition algorithm in the fourth step, and the long-period energy data without missing time points containing the instantaneous main contribution degree of the main indicator variable to the target result, the instantaneous collaborative contribution degree time series of the collaborative indicator variable to the corresponding main indicator variable after the missing time points were interpolated and completed based on the long-period different types of energy data through Kalman filtering, the repeating time interval of the seasonal term fluctuation, the time difference between adjacent fluctuation peak and valley were determined by statistical analysis to determine the fixed period type reflected by the seasonal term. By taking the seasonal term with the explicit fixed period type as the periodic component, the periodic component that can be used to construct the main indicator variable and the collaborative indicator variable contribution degree dynamic change path of the fusion trend component, the periodic component and the mutation point marker is obtained.

[0031] The target result is taken as the target node of the first layer, the main indicator variable is taken as the main contribution node of the second layer, and the sequence of collaborative indicator variables corresponding to each main indicator variable is taken as the collaborative contribution node set of the third layer. Based on the instantaneous main contribution degree change path of each type of energy data corresponding to the main contribution node to the target node, the instantaneous collaborative contribution degree change path of each collaborative contribution node to the corresponding main indicator variable, and the covariate correlation coefficient between the same collaborative contribution node set, a dynamic contribution flow network node tree corresponding to a single type of energy data at continuous time points is constructed through a graph neural network combined with a graph database.

[0032] It needs to be further explained that the specific implementation process of obtaining the dynamic contribution flow network node tree corresponding to a single type of energy data at continuous time points in this embodiment includes: First, based on the target result, the main indicator variable and the collaborative indicator variable, three-layer node sets are constructed. The dynamic weight matrix between nodes is obtained by inputting the main contribution node, the instantaneous main contribution degree change path of the target node, the instantaneous collaborative contribution degree change path of the collaborative contribution node and the correlation coefficient of the covariant as edge features into the multi-head attention mechanism of the Transformer. Second, based on the weight matrix and the initial features of the nodes, the time sequence features are captured by combining the graph convolution network with the gated recurrent unit of the time dimension, and then the neighborhood information is weighted and aggregated by the attention mechanism to obtain the updated node features. Third, based on the node features and edge weights at consecutive time points, the graph structure is aligned by the dynamic time warping algorithm, and the topological structure of the directed acyclic graph is obtained by combining the topological sorting algorithm to eliminate circular dependencies. Fourth, based on the directed acyclic graph, the nodes and edges are stored in the Neo4j graph database, and the multi-version graph structure index is constructed with the timestamp as the index key to finally obtain the single-type energy dynamic contribution flow network node tree containing three-layer node hierarchy, dynamic edge weight and time sequence evolution features.

[0033] (4) Causal association analysis is performed on the set of associated change attributes to obtain the joint contribution influence factor of the main indicator variable on the target result: Based on the set of associated change attributes of all types of energy parameters at each time point, the main joint contribution degree of all main indicator variables corresponding to each target result to the same target result and the associated influence relationship and directed association influence direction between all main indicator variables corresponding to each target result, as well as the directed causal correlation coefficient between the sequences of collaborative indicator variables corresponding to different main indicator variables, are obtained by the causal association analysis algorithm. Based on the instantaneous main contribution degree of a single main indicator variable to the target result divided by the main joint contribution degree of all main indicator variables corresponding to each target result to the same target result within a single time, the joint contribution influence factor of each main indicator variable to the same target result at a single time point is obtained.

[0034] It should be further explained that the specific process of obtaining the joint contribution influence factor of each main indicator variable to the same target result at a single time point in this embodiment includes: First, based on the multi-energy main indicator variables at a single time point, the direct contribution degree of each main indicator variable to the target result is obtained by joint modeling through the structural equation model and solving the standardized path coefficient by maximum likelihood estimation. Second, based on the main index variable set, the conditional dependence relationship is inferred by Bayesian network, the D-separation criterion is used to verify the directional association influence direction and the correlation strength is quantified by transfer entropy, and the directional association relationship matrix between the main indexes is obtained. The conditional dependence relationship refers to the association influence relationship and the directional association influence direction between all main index variables corresponding to each target result. Specifically, in the foregoing, the joint contribution influence factor of the main index variable to the target result is constructed, and the analysis is carried out around all main index variables corresponding to each target result in a single time. The main index variable set here refers to all main index variables corresponding to the target result. Through the conditional dependence relationship inferred by the Bayesian network, the mutual influence correlation characteristics (i.e. the association influence relationship) between the main index variables in the set are analyzed under the constraint of the target result, and then the D-separation criterion is used to verify the directionality of the correlation (i.e. the directional association influence direction). The directional association relationship matrix between the main indexes formed by subsequently quantifying the correlation strength by transfer entropy is a quantitative representation of the association influence relationship and the directional association influence direction between all main index variables corresponding to each target result. The two are completely consistent in the analysis object and the correlation dimension.

[0035] Third, based on the collaborative index variable sequence, the Granger causality test is combined with the time-varying parameter vector autoregressive model, the Markov chain Monte Carlo sampling is used to estimate the causal correlation coefficient, and the directional causal correlation coefficient matrix across the collaborative indexes is obtained. Fourth, based on the direct contribution degree of each main index variable and the interaction effect between the main index variables, the Shapley value decomposition method is used to distribute the total contribution degree of the target result to each main index variable, and the main joint contribution degree considering the interaction effect is obtained. Fifth, based on the individual contribution degree of a single main index variable and the main joint contribution degree, the ratio is calculated and the bias is corrected by the bootstrap method, and the standardized joint contribution influence factor set containing the relative importance proportion, the correlation direction and the causal strength of the main index is obtained.

[0036] (5) Based on the joint contribution influence factor and the dynamic contribution flow network node tree corresponding to all types of energy data, a cross-type dynamic contribution flow network node forest is constructed: Based on the joint contribution influence factor of each main index variable to the same target result at a single time point, the main contribution relationship connection between each main contribution node and the corresponding target node is constructed. Based on the association influence relationship and the directional association influence direction between all main index variables, the directional association connection between the main contribution nodes is constructed by the blockchain algorithm, and the cross-tree covariant association connection is constructed based on the directional causal correlation coefficient between the collaborative index variable sequences corresponding to different main index variables. The dynamic contribution flow network node tree corresponding to all types of energy data is time-stamped and aligned at the continuous time point with the constructed main contribution relationship connection, directional association connection and cross-tree covariant association connection. After alignment, the topology space algorithm is combined to obtain a cross-type dynamic contribution flow network node forest.

[0037] It should be further explained that the detailed acquisition process of the cross-type dynamic contribution flow network node forest of the embodiment includes: First, based on the joint contribution influence factor of each main indicator variable at a single time point, the directional edge weight between the target node and the main contribution node is obtained by mapping and normalizing the directional edge weight between the target node and the main contribution node, and the main contribution relationship connection is obtained. Second, based on the causal relationship between the main contribution nodes inferred by the Bayesian network, the association direction is recorded by the alliance blockchain framework and verified by the smart contract, and the association strength history information is stored by the Merkle tree, and the directional association connection between the main contribution nodes is obtained. Third, based on the directional causal association coefficient output by the time-varying parameter vector autoregressive model, the time series of different energy types are aligned by the dynamic time warping algorithm, and the cross-tree covariant association connection with consistent time sequence is obtained. Fourth, based on the node tree structure of each energy type, the time offset pattern is learned by the sliding window matching algorithm combined with the bidirectional long short-term memory network, the attention mechanism is introduced to dynamically adjust the alignment weight, and the time-stamped multi-source node tree set is obtained. Fifth, based on the time-stamped node tree, the nodes are mapped to a high-dimensional space by the topology space algorithm combined with the graph embedding algorithm (such as Node2Vec), the k-nearest neighbor algorithm is used to obtain the cross-tree node similarity and combined with hierarchical clustering, and the cross-type node forest with multi-scale topological structure is obtained. Sixth, based on the multi-scale topological structure of the cross-type node forest and the time sequence alignment dynamic characteristics, the contribution degree change in the time dimension and the topological association in the space dimension are integrated by the spatio-temporal graph convolution network, a three-dimensional node forest containing energy type, time sequence and causal relationship is obtained, and the contribution flow visualization and dynamic analysis of cross-type energy data are realized.

[0038] (6) The reverse comprehensive training function is constructed by analyzing the acquisition process of the cross-type dynamic contribution flow network node forest and performing reverse reasoning training, and a trained directional association graph model is obtained. The reverse comprehensive training function is constructed based on the corresponding instantaneous contribution degree mean square error, the instantaneous main contribution degree change path loss, the main index correlation relationship cross entropy loss, the cross-tree covariant causal relationship loss, the continuous timestamp alignment loss, the main contribution node and the corresponding collaborative contribution node number complexity penalty, and the constraint penalty function of the cross-type dynamic contribution flow network node forest corresponding to the correlation strength less than the preset correlation coefficient. The correlation strength refers to the correlation strength of the corresponding connection between the nodes in the cross-type dynamic contribution flow network node forest. The first is the main contribution relationship strength between the main contribution node and the corresponding target node, which is determined based on the joint contribution influence factor of each main index variable to the same target result at a single time point, reflecting the contribution correlation degree of the main index variable to the target result. The second is the directed correlation strength between the main contribution nodes, which is determined based on the correlation influence relationship and the directed correlation influence direction between the main index variables of all types of energy, reflecting the mutual influence strength between the main index variables of different types of energy. The third is the cross-tree covariant correlation strength, which is determined based on the directed causal correlation coefficient between the corresponding collaborative index variable sequences of different main index variables, reflecting the cross-type causal correlation degree between the collaborative index variables of different types of energy. The above three types of correlation strength jointly constitute the strength quantization index of various connection relationships in the cross-type dynamic contribution flow network node forest. When the strength of a certain correlation is less than the preset correlation coefficient, the corresponding constraint penalty is triggered to optimize the effectiveness of the network structure.

[0039] It should be further explained that the instantaneous contribution degree mean square error of the embodiment is constructed based on the square sum of the difference between the partial least squares regression (PLS) prediction value and the actual contribution degree, which is used to quantify the prediction accuracy of the main-collaborative index contribution degree and avoid the estimation bias of the key variable contribution degree. The instantaneous main contribution degree change path loss is a weighted sum of the dynamic time warping distance and the long short-term memory network hidden state difference, which is used to constrain the smoothness of the contribution degree path at continuous time points and capture the time sequence evolution law of energy data. The main index correlation relationship cross entropy loss is constructed by the cross entropy of the correlation probability distribution inferred by the Bayesian network and the actual causal relationship label, which is used to optimize the recognition accuracy of the directed correlation between the main index variables and ensure the logical rationality of the energy transmission path. The cross-tree covariant causal relationship loss is a Huber loss of the transfer entropy estimate value and the Granger causality test result, which is used to balance the estimation bias and outlier robustness of the causal strength between the collaborative index variables. The continuous timestamp alignment loss is constructed based on the smooth L1 loss of the time offset predicted by the bidirectional long short-term memory network and the actual alignment error, which is used to ensure the consistency of the time dimension of the multi-energy data. The complexity penalty term is constructed by L1 regularization of the number of main contribution nodes and collaborative contribution nodes and sparse constraint of the edge weight of the graph network, to avoid model overfitting and improve the generalization ability of cross-type energy data. The correlation strength constraint penalty is an exponential penalty applied to the correlation edges below the preset threshold, to force to retain strong correlation relationships and simplify the node forest topology to improve computational efficiency.

[0040] Based on the reverse comprehensive training function combined with the cross-type dynamic contribution flow network node forest (i.e., training the cross-type dynamic contribution flow network node forest based on the reverse comprehensive training function), through the random forest combined with the simulation algorithm, the reverse reasoning training from the third layer of the collaborative contribution node set to the target node of the first layer is performed, to obtain the trained directed correlation graph model.

[0041] This process realizes the unified representation and deep coupling of multi-source heterogeneous energy data such as coal, oil, gas, electricity and heat through multi-level dynamic graph modeling and spatio-temporal fusion mechanism; in particular, based on the three-layer node tree structure (target node-main contribution node-collaborative contribution node set) of the graph neural network, combined with dynamic time warping and gated recurrent unit, the contribution path evolution law of main-collaborative indicators is effectively captured; through the joint contribution influence factor of the Bayesian network and the Shapley value decomposition calculation, the interaction effect between multi-energy main indicator variables is innovatively quantified, breaking through the limitation of the non-comparable contribution degree in traditional single-type energy analysis; secondly, the correlation relationship storage mechanism enabled by the blockchain verifies the causal relationship direction and the Merkle tree historical version traceability, to build an unalterable directed correlation connection, ensuring the credibility of the cross-type conduction path. In the topology space embedding stage, the graph embedding algorithm and k-nearest neighbor clustering fusion map the multi-dimensional energy nodes to a high-dimensional continuous space, solving the problem of spatial heterogeneity of heterogeneous energy systems; thirdly, the three-dimensional modeling capability of the spatio-temporal graph convolution network realizes the visualization tracking of the fluctuation conduction path, providing a technical foundation for the cross-system tracing of complex events such as wind and light electricity abandonment.

[0042] S2: Obtain a cross-type linkage report and build a basic directed positioning dynamic correction chain through report analysis positioning and causal analysis; Further preferably, a cross-type linkage report is obtained, and based on the key target results and corresponding causal correlation strengths of the cross-type linkage report obtained through entity relationship extraction algorithm, a basic directed positioning dynamic correction chain is constructed in combination with the blockchain algorithm and the position of each key target result in the report; The basic directed positioning dynamic correction chain is constructed by combining the directed correlation curve with the blockchain algorithm, and one implementation in the embodiment is as follows: (1) Based on the cross-type linkage report, the text content is extracted by the OCR technology and the key target results are located by the NLP entity recognition algorithm to obtain a set of key target results containing position coordinates; For example, the key target results in this embodiment are key parameters and corresponding result data in the report; for example, the current power price, the coal price corresponding to the current power price, or the coal demand; (2) Based on the set of key target results, the association and change relationship between the key target results are identified by the graph neural network algorithm combined with the semantic analysis technology, the causal association strength is calculated by the gradient boosting tree, and the pair of key target results with the association and change relationship and the corresponding causal strength matrix are obtained; (3) Based on the causal strength matrix, each pair of key target results and the causal association strength is converted into a unique hash value by the hash algorithm of the blockchain, the validity of the association and change relationship is automatically verified by the smart contract, and the set of association relationships stored on the chain is obtained; (4) Based on the layout analysis result of the cross-type linkage report, the physical coordinates of the key target results in the report are converted into topological structure vectors by the relative position coding technology, the hierarchical relationship is identified by the spatial clustering algorithm, and the position coding vector of the key target results is obtained; (5) Based on the set of association relationships stored on the chain and the position coding vector, the causal association network between the key target results is constructed by the directed acyclic graph algorithm, the time series are aligned by the dynamic time warping algorithm, and the directed positioning dynamic network is obtained; based on the directed positioning dynamic network, the network integrity is verified by the zero-knowledge proof algorithm under the premise of protecting data privacy, the multi-party consensus is realized by the threshold signature algorithm, and finally the basic directed positioning dynamic correction chain containing causal association, position coding and time dimension is obtained.

[0043] It should be noted that the cross-type linkage report (report linked by different energies and different information) in this embodiment is generated by using existing report software, such as FineReport, AnyReport, and other professional report tools, which will not be described in detail here.

[0044] S3: inputting the cross-type dynamic contribution flow network node forest and the basic directed positioning dynamic correction chain into the pre-trained positioning correction path model to perform real-time positioning error correction adjustment on the cross-type linkage report, and obtaining the updated and adjusted cross-type linkage report.

[0045] Further preferably, based on the cross-type linkage report, the cross-type linkage report is adjusted in real time by the positioning correction path model constructed by the cross-type dynamic contribution flow network node forest, the directed positioning dynamic correction chain and the text positioning algorithm, and the updated and adjusted cross-type linkage report is obtained.

[0046] The positioning correction path model is constructed by a cross-type linkage report combined with a directed positioning dynamic correction chain, as shown in Figure 4 The process includes: 1) Obtain cross-type linkage report information and locate through report analysis, obtain the key target result sequence in the table and the relationship of the synchronous change of the key target result, the position of the key target result in the table and its contribution impact factor, and through causal analysis, obtain the key target result and the corresponding causal correlation strength of the existence of the correlation change relationship: Obtain historical cross-type linkage report information, through a preset report analysis positioning model (based on Faster R-CNN and OCR / NLP fusion), obtain the key target result sequence in the cross-type linkage report and the relationship of the synchronous change of the remaining key target result when each key target result produces a change, the position of each key target result in the cross-type linkage report, and the corresponding contribution impact factor of each key target result; The specific implementation process includes: First, based on the report image data, the table area is detected through Faster R-CNN and the text content is extracted by combining the OCR algorithm, the key target result is located by using the NLP named entity recognition algorithm, and the key target result set containing the position coordinates is obtained; Second, based on the time series data of the key target result, the dependence strength is evaluated by calculating the time series mutual information matrix, the change causal direction is identified by combining the Granger causality test, and the directed linkage graph is constructed by using the Bayesian network algorithm, to obtain the linkage relationship matrix between the key target results; Third, based on the report layout analysis result, the cell coordinates of the table are mapped to the pixel space and the hierarchical structure is represented by using the relative position coding, to obtain the position coding vector of the key target result; Based on the key target result and the associated index data, the feature importance is calculated by gradient boosting tree, the causal path graph is constructed by combining the structural equation model and the path coefficient is standardized, to obtain the direct contribution degree of each index; Fourth, based on the direct contribution degree calculation result, the interactive effect contribution is allocated by the Shapley value decomposition method, the SEM parameter estimation is regularized by LASSO, and the strength is optimized by cross-validation, to obtain the contribution impact factor for processing high-dimensional sparse data; Fifth, based on the nonlinear relationship characteristics, the function mapping between indexes is fitted by Gaussian process regression (GPR), and the complex dependence is captured by combining kernel function selection (such as RBF kernel), to obtain the contribution impact factor tensor containing nonlinear effects; Sixth, based on the linkage relationship matrix, the position coding vector and the contribution impact factor tensor, a structured feature set containing three-dimensional information of indexes, time and target results is constructed, to support the accurate calculation of subsequent positioning correction paths.

[0047] Based on the key target result sequence and the contribution impact factor corresponding to each key target result, the key target results with correlated variation relationship and the corresponding causal correlation strength are obtained through a causal analysis algorithm.

[0048] 2) Based on the position of the key target result in the table, the key target results with correlated variation relationship and the corresponding causal correlation strength, a training optimization directed positioning dynamic correction chain is constructed and mapped with the cross-type dynamic contribution flow network node forest for similarity analysis to obtain a similar consistency error: Based on the key target results with correlated variation relationship and the corresponding causal correlation strength, a directed positioning dynamic correction chain is constructed by combining the blockchain algorithm with the position of each key target result in the report; it needs to be further explained that the specific construction process of the directed positioning dynamic correction chain in this embodiment includes: First, based on the linkage relationship matrix of the key target result sequence, the key target result pairs with a causal correlation strength greater than the threshold value are selected to form a directed edge set; it needs to be further explained that the linkage relationship matrix in this embodiment is generated by Granger causality test and Bayesian network; Second, the causal relationship is notarized through the alliance blockchain architecture, the directionality of the edge is verified through the smart contract, the correlation strength history value at each timestamp is stored using the Merkle tree to ensure traceability; Third, the pixel coordinates obtained by analyzing the report layout are converted into topological graph node encodings, a multi-dimensional position vector is generated using a relative position embedding algorithm, the neighborhood position features are aggregated through a graph attention network, and the spatial correlation constraint is strengthened; wherein the causal correlation strength in this embodiment is updated in real time through a dynamic conditional correlation model, and the time sequence features are extracted by combining a variational autoencoder for abnormal detection and weight adjustment of the causal relationship of the sudden variation.

[0049] Fourth, in the chain structure construction stage, this embodiment takes the key target result as the node and the causal relationship as the edge, generates a directed acyclic chain through a topological sorting algorithm, introduces a time-dimension gated recurrent unit to capture the time delay of causal conduction in the chain, adopts a PBFT consensus algorithm to ensure the consistency of the cross-node chain structure, and realizes the dynamic update of the directed positioning dynamic correction chain through a smart contract. Fifth, when a new correlated variation relationship is detected, the chain structure is recalculated, the Louvain community discovery algorithm is used to optimize the hierarchical division of the chain, and finally a directed positioning dynamic correction chain containing spatial and temporal position features, dynamic causal weights and time sequence conduction delay is formed to realize real-time positioning and error correction of the abnormality of the cross-type linkage report.

[0050] It should be noted that the construction of the directed positioning dynamic correction chain in S2 and S3 belongs to the progressive support relationship of basic construction-training optimization. In S2, the input layer basic construction needs to focus on the key information extraction and basic association modeling of the report itself, and the directed positioning dynamic correction chain is constructed based on real-time cross-type linkage report, through OCR text extraction, NLP entity recognition positioning key target result, combining graph neural network analysis causal association strength, and block chain evidence association relationship construction. The core purpose is to extract the position, causal association and time sequence characteristics of the key target result from the report, form the core input data of the positioning correction path model, and not combine the cross-type dynamic contribution flow network node forest. In S3, the model layer optimization construction needs to optimize the accuracy and spatio-temporal adaptability of the dynamic contribution relationship optimization chain of the node forest, and the directed positioning dynamic correction chain is constructed in the training scene of the positioning correction path model based on the linkage relationship matrix of historical cross-type linkage report, further combining the mapping analysis (similarity algorithm, error calculation) of the cross-type dynamic contribution flow network node forest, through topological sorting optimization hierarchical structure, PBFT consensus algorithm to guarantee consistency, and gated recurrent unit to capture time delay. The core purpose is to provide accurate carriers for model calculation similarity consistency error and construction of forward comprehensive loss function, serving model training and error correction.

[0051] Map each key target result in the training optimization directed positioning dynamic correction chain to the target node in the cross-type dynamic contribution flow network node forest at the same time, and according to the contribution influence factor corresponding to each key target result and the main contribution node and collaborative contribution node set at the same time point in the cross-type dynamic contribution flow network node forest, obtain the similarity coefficient between the contribution influence factor and the main contribution node and the collaborative contribution node and the similarity coefficient between the contribution influence factor and the collaborative contribution node through a similarity algorithm; According to the obtained similarity coefficient, establish the similarity mapping of the contribution influence factor and the main contribution node and the collaborative contribution node, and construct the similarity consistency error.

[0052] 3) Construct a forward comprehensive loss function by analyzing the similarity consistency error, the corresponding parameter relationship between the training optimization directed positioning dynamic correction chain and the cross-type dynamic contribution flow network node forest, and perform forward training to obtain a trained positioning correction path model, and based on the loss result of forward training, real-time synchronous online training of the directed association graph model: The keyword matching algorithm based on the directed positioning dynamic correction chain is used to correct the positioning algorithm, and the error between the positioning key target result of the training and optimization directed positioning dynamic correction chain and the reverse target result of the same target node obtained by the reverse reasoning of the cross-type dynamic contribution flow network node forest, the error between the target result of each key target result in the training and optimization directed positioning dynamic correction chain after the synchronous change of the remaining key target result and the reverse target result of the remaining key target result in the same target node obtained by the reverse reasoning of the cross-type dynamic contribution flow network node forest, and the forward comprehensive loss function constructed by the delay of the synchronous change of the remaining key target result, are combined to perform forward training from the first layer corresponding to the target node to the third layer corresponding to the collaborative contribution node set, and obtain the trained positioning correction path model. It should be further explained that one implementation mode of the embodiment is specifically: Based on the directed positioning dynamic correction chain, the semantic mapping of the key target result and the cross-type dynamic contribution flow network node forest is established by using the keyword matching algorithm (such as BM25) to obtain the initial matching feature vector. Based on the feature vector, the similarity consistency error is calculated by using the contrast loss function to obtain the similarity loss value, and the semantic alignment of the contribution influence factor and the main-collaborative node feature is forced to obtain the similarity loss value. Based on the key target result positioned by the directed positioning dynamic correction chain, the reverse target result of the same target node is obtained by the reverse reasoning of the cross-type dynamic contribution flow network node forest, the smooth L1 error between the two is calculated, and the reverse error value of the target node is obtained. Based on the target result after the synchronous change of the remaining key target result corresponding to each key target result in the directed positioning dynamic correction chain, the reverse target result under the same target node is obtained by the reverse reasoning of the cross-type dynamic contribution flow network node forest, the Huber loss between the two is calculated, and the reverse error value of the collaborative node is obtained. Based on the time sequence of the synchronous change of the remaining key target result, the actual conduction delay is estimated by using the hidden Markov model, the mean square error of the model prediction delay is calculated, and the change delay loss value is obtained. Based on the similarity loss value, the reverse error value of the target node, the reverse error value of the collaborative node, and the change delay loss value, the forward comprehensive loss function is constructed by linear combination of the adaptive weight matrix dynamically adjusted by gradient back propagation. Based on the forward comprehensive loss function, the positioning correction path model is iteratively trained by forward propagation from the first layer corresponding to the target node to the third layer corresponding to the collaborative contribution node set, combined with the random forest and the simulated annealing algorithm, and the trained positioning correction path model is obtained.

[0053] It needs to be further explained that the similarity consistency error in the embodiment is based on the cosine similarity (similarity coefficient) of the contribution influence factor and the main-collaborative contribution node feature, which is constructed by using the contrast loss to force similar mapping, to avoid feature deviation caused by semantic gap and ensure semantic alignment of report indicators and node forest; Specifically: based on the mapping relationship formed by one-to-one mapping each key target result in the directional positioning dynamic correction chain with the target node in the cross-type dynamic contribution flow network node forest at the same time, according to the similarity coefficient between the contribution influence factor corresponding to each key target result and the main contribution node and the similarity coefficient between the contribution influence factor and the collaborative contribution node set in the cross-type dynamic contribution flow network node forest at the same time point, and the similarity mapping of the contribution influence factor and the main contribution node and the collaborative contribution node based on the above similarity coefficient, the deviation degree of the similarity coefficient in the similarity mapping is quantified, and the similarity difference is calculated combined with the contrast loss function, to obtain the similarity consistency error.

[0054] The target node reverse error is constructed by the smooth L1 loss of the target result positioned by the directional positioning dynamic correction chain and the node forest reverse reasoning result, which is used to combine the dynamic time warping distance constraint to capture the cross-modal conduction deviation of energy data; The collaborative node reverse error is constructed by the structural equation model path coefficient difference loss term, which applies Huber loss to the deviation of the target result after the collaborative contribution node changes and the node forest reasoning result, to balance the robustness and accuracy of outliers; The change delay loss is constructed based on the mean square error of the state transition time of the hidden Markov model and the actual conduction delay, and the embodiment introduces an attention mechanism to dynamically weight the delay deviation of different time scales, to optimize the time sequence response capability of multi-energy data linkage.

[0055] The forward comprehensive loss result obtained in each step of the positioning correction path model training is fed back to the training process (6) of the directional association graph model, combined with the online training framework, to perform synchronous reverse adjustment training and obtain the real-time synchronous online training directional association graph model.

[0056] The process realizes the accurate positioning and self-correction closed loop of multi-source heterogeneous data in the energy field for the first time by constructing a bidirectional mapping mechanism of directed positioning dynamic correction chain and cross-type dynamic contribution flow network node forest. Especially, first of all, based on the report analysis technology of Faster R-CNN and OCR / NLP fusion, the cell coordinates of physical report are converted into topological space encoding, combined with the linkage relationship matrix generated by Granger causality test and Bayesian network, the dual binding of report location and energy semantics is established at pixel level precision. This spatial-semantic joint positioning mechanism effectively solves the disconnection between location information and business logic in traditional report analysis; secondly, through the PBFT consensus algorithm to ensure the alliance chain architecture, the distributed evidence of causal correlation strength is combined with the multi-version tracing ability of Merkle tree to ensure the non-tamperability of the transmission path, among them, the time delay features captured by the gated recurrent unit and the hierarchical optimization algorithm of Louvain community discovery cooperate to build a chain structure with spatio-temporal adaptive ability; when detecting the gas price mutation triggering the electricity price linkage, the chain can automatically locate the target area of the electricity report and complete the transmission path tracing in a short time, which establishes a technical foundation for real-time intervention of emergency events; thirdly, the cosine similarity calculation of contribution influence factor and cross-type dynamic contribution flow network node forest features solves the term gap across energy systems, such as the quantitative equivalence of "peak valley difference" in electricity and "thermal inertia" in heat. This mapping not only eliminates feature deviation, but also verifies the equivalence of path coefficients through structural equation modeling, realizing the unified expression of report indicators and forest nodes at the mathematical logic level. Fourth, the smooth L1 loss in the reverse error of the target node combined with dynamic time warping constrains the transmission consistency of wind power output fluctuations in the power, heat, and gas systems; the variable delay loss optimizes the cascade response speed of "carbon price change, coal-fired power cost, industrial electricity price" through hidden Markov model state transition and attention weighting mechanism; while the Huber loss guarantees the robustness of the prediction of collaborative contribution nodes in extreme scenarios; the finally formed closed loop training architecture restructures the energy data analysis paradigm, especially, the positioning correction path model establishes a dynamic correction mechanism for the linkage report through the cooperative optimization of four-dimensional comprehensive loss function (including similarity consistency, target matching, variable matching, and time delay constraint). The time offset learning and attention weight adjustment based on bidirectional long short-term memory network significantly improves the response speed of multi-energy chain imbalance under emergency conditions.The closed-loop architecture of forward training and reverse feedback enables the directed association graph model to have a continuous evolution capability and to be self-adaptive to market fluctuations; finally, the embodiment based on the space-time contribution flow modeling unifies the dynamic representation paradigm of energy data, so that the heterogeneous system has a computable semantic interworking capability; based on the causal blockchain mechanism, the verification difficulty of cross-type conduction paths is solved, and reliable analysis is provided for complex correlations such as carbon price, coal power and power cost cascading effect; based on the closed-loop correction system, the self-iterative optimization of the analysis result is realized, and the paradigm transition of energy decision from static report to dynamic deduction is promoted.

[0057] Embodiment 2 of the present application provides a heterogeneous data processing system for energy big data, as shown in Figure 5 The embodiment of the present application comprises: A directed association module is configured to obtain different types of energy data and input a pre-trained directed association graph model to obtain a cross-type dynamic contribution flow network node forest, thereby realizing database construction. A causal analysis module is configured to obtain a cross-type linkage report and construct a basic directed positioning dynamic correction chain through report analysis positioning and causal analysis. A positioning correction module is configured to input the cross-type dynamic contribution flow network node forest and the basic directed positioning dynamic correction chain into a pre-trained positioning correction path model, to perform real-time positioning error correction adjustment on the cross-type linkage report, and obtain an updated and adjusted cross-type linkage report.

[0058] Embodiment 3 of the present application provides a terminal comprising a processor and a storage medium; the storage medium is configured to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method.

[0059] Embodiment 4 of the present application provides a computer readable storage medium having a computer program stored thereon, which is executed by a processor to implement the steps of the method.

[0060] Compared with the prior art, the present application has at least the following beneficial effects: The present application forms a cross-type dynamic contribution flow network node forest by constructing a directed association graph model and a positioning correction path model, realizes deep fusion and dynamic association analysis of multi-source heterogeneous energy data, can eliminate data islands, establishes a reliable association mapping of cross-type energy data, accurately captures the nonlinear time-varying law of multi-energy coupling, and realizes cross-type conduction analysis of main-cooperative index dynamic contribution; The present application improves the learning ability of the model for energy data association rules by comprehensive training function and reverse reasoning training, realizes real-time error correction of cross-type linkage report in combination with the positioning correction path model, and dynamically corrects multi-energy chain imbalance; The application provides accurate and self-correcting analysis results for energy prediction through an online synchronous training mechanism, and improves the accuracy and timeliness of energy decision-making.

[0061] The present disclosure can be a system, a method, and / or a computer program product. The computer program product can include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present disclosure.

[0062] The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium can be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium include the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch-cards or punched tape, a magneto-optical or other optical medium, and / or any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission media (e.g., light pulses passing through a fiber-optic cable), or electrical signals transmitted through a wire.

[0063] Computer readable program instructions described herein can be downloaded to respective computing / processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and / or a wireless network. The network can comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers and / or edge servers. A network adapter card or network interface in each computing / processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing / processing device.

[0064] Computer readable program instructions for carrying out operations of the present disclosure can be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state-setting data, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++ or the like, and conventional procedural programming languages such as the "C" programming language or similar programming languages. The computer readable program instructions can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate array (FPGA), or programmable logic array (PLA) can execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of the present disclosure.

[0065] Finally, it should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application, rather than limiting the present application. Although the present application has been described in detail with reference to the above-mentioned embodiments, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and any modification or equivalent replacement without departing from the spirit and scope of the present application should be covered in the protection scope of the claims of the present application.

Claims

1. A heterogeneous data processing method for energy big data, characterized in that, The method comprises: acquiring different types of energy data and inputting a pre-trained directed association graph model to obtain a cross-type dynamic contribution flow network node forest; acquiring cross-type linkage reports and constructing a basic directed positioning dynamic correction chain through report analysis positioning and causal analysis; inputting the cross-type dynamic contribution flow network node forest and the basic directed positioning dynamic correction chain into a pre-trained positioning correction path model to perform real-time positioning error correction adjustment on the cross-type linkage report to obtain an updated and adjusted cross-type linkage report.

2. The heterogeneous data processing method of energy big data according to claim 1, characterized in that: the training process of the directed association graph model comprises: acquiring historical information of each type of energy and extracting key variables to obtain an energy parameter time sequence set and an associated change attribute set of each type of energy; performing multi-stage time series factor analysis on the energy parameter time sequence set to obtain main index variables and corresponding collaborative index variable sequences, instantaneous main contribution degrees of the main index variables to the target result, instantaneous collaborative contribution degrees of the collaborative index variable sequences to the corresponding main index variables, and covariant correlation coefficients between the collaborative index variables; repeating the acquisition process of the instantaneous main contribution degrees and the instantaneous collaborative contribution degrees to obtain change paths of the instantaneous main contribution degrees and the instantaneous collaborative contribution degrees, and combining the covariant correlation coefficients between the collaborative index variables to construct a dynamic contribution flow network node tree corresponding to the single-type energy data; performing causal association analysis on the associated change attribute set to obtain joint contribution influence factors of the main index variables to the target result; based on the joint contribution influence factors and the dynamic contribution flow network node trees corresponding to all types of energy data, a cross-type dynamic contribution flow network node forest is constructed; a reverse comprehensive training function is constructed by analyzing the acquisition process of the cross-type dynamic contribution flow network node forest and reverse reasoning training is performed to obtain a trained directed association graph model.

3. The heterogeneous data processing method of energy big data according to claim 2, characterized in that: the acquisition of historical information of each type of energy and the extraction of key variables to obtain an energy parameter time sequence set and an associated change attribute set of each type of energy comprises: acquiring historical demand quantity change information, cost change information and associated change information of each type of energy and extracting key variables to obtain an energy parameter time sequence set and an associated change attribute set of each type of energy.

4. The heterogeneous data processing method of energy big data according to claim 2, characterized in that: the multi-stage time series factor analysis on the energy parameter time sequence set to obtain main index variables and corresponding collaborative index variable sequences, instantaneous main contribution degrees of the main index variables to the target result, instantaneous collaborative contribution degrees of the collaborative index variable sequences to the corresponding main index variables, and covariant correlation coefficients between the collaborative index variables comprises: based on the energy parameter time sequence set of each type of energy, a parameter loading matrix corresponding to each type of energy parameter continuous time point is obtained through a multi-stage time series factor analysis model; Statistical analysis is performed on the parameter load matrix corresponding to each time point to obtain each type of main indicator variable and the corresponding sequence of collaborative indicator variables, the instantaneous main contribution degree of each main indicator variable to the target result, the instantaneous collaborative contribution degree of each sequence of collaborative indicator variables to the corresponding main indicator variable, and the covariant correlation coefficient between the collaborative indicator variables.

5. The heterogeneous data processing method of energy big data according to claim 2, characterized in that: the repeated process of obtaining the instantaneous main contribution degree and the instantaneous collaborative contribution degree obtains the change path of the instantaneous main contribution degree and the change path of the instantaneous collaborative contribution degree, and in combination with the covariant correlation coefficient between the collaborative indicator variables, a dynamic contribution flow network node tree corresponding to a single type of energy data is constructed, including: repeating the process of obtaining the instantaneous main contribution degree and the instantaneous collaborative contribution degree to obtain the change path of the instantaneous main contribution degree of each main indicator variable to the target result and the change path of the instantaneous collaborative contribution degree of each collaborative indicator variable to the corresponding main indicator variable at consecutive time points; taking the target result as the target node of the first layer, the main indicator variable as the main contribution node of the second layer, and the sequence of collaborative indicator variables corresponding to each main indicator variable as the collaborative contribution node set of the third layer; based on the instantaneous main contribution degree change path of each type of energy data corresponding to the main contribution node to the target node, the change path of the instantaneous collaborative contribution degree of each collaborative contribution node to the corresponding main indicator variable, and the covariant correlation coefficient between the same collaborative contribution node set, a dynamic contribution flow network node tree corresponding to a single type of energy data at consecutive time points is constructed through a graph neural network combined with a graph database.

6. The heterogeneous data processing method of energy big data according to claim 2, characterized in that: the causal association analysis of the set of associated change attributes to obtain the joint contribution influence factor of the main indicator variable to the target result includes: based on the set of associated change attributes of all types of energy parameters at each time point, the main joint contribution degree of all main indicator variables corresponding to each target result to the same target result and the associated influence relationship and the directed associated influence direction between all main indicator variables corresponding to each target result, and the directed causal correlation coefficient between the sequence of collaborative indicator variables corresponding to different main indicator variables are obtained through a causal association analysis algorithm within a single time; based on the instantaneous main contribution degree of a single main indicator variable to the target result divided by the main joint contribution degree of all main indicator variables corresponding to each target result to the same target result within a single time, the joint contribution influence factor of each main indicator variable to the same target result at a single time point is obtained.

7. The heterogeneous data processing method of energy big data according to claim 6, characterized in that: the construction of the cross-type dynamic contribution flow network node forest based on the joint contribution influence factor and the dynamic contribution flow network node tree corresponding to all types of energy data includes: based on the joint contribution influence factor of each main indicator variable to the same target result at a single time point, the main contribution relationship connection between each main contribution node and the corresponding target node is constructed. Based on the correlation influence relationship and the directional correlation influence direction between all main indicator variables, the directional correlation connection between the main contribution nodes is constructed through the blockchain algorithm, and the cross-tree cooperative variable correlation connection is constructed based on the directional causal correlation coefficient between the corresponding cooperative indicator variable sequences of different main indicator variables; The dynamic contribution flow network node forest of all types of energy data is timestamped at the continuous time point after the main contribution relationship connection, the directional correlation connection and the cross-tree cooperative variable correlation connection are constructed, and the cross-type dynamic contribution flow network node forest is constructed by combining the topological space algorithm after alignment.

8. The heterogeneous data processing method of energy big data according to claim 5, characterized in that: the reverse comprehensive training function is constructed and the reverse reasoning training is performed based on the acquisition process of the cross-type dynamic contribution flow network node forest, and a trained directional correlation graph model is obtained, comprising: a reverse comprehensive training function is constructed based on the instantaneous contribution degree mean square error, the instantaneous main contribution degree change path loss, the main indicator correlation relationship cross entropy loss, the cross-tree cooperative variable causal relationship loss, the continuous timestamp alignment loss, the complexity penalty of the number of main contribution nodes and corresponding cooperative contribution nodes, and the constraint penalty function of the cross-type dynamic contribution flow network node forest corresponding to the correlation coefficient less than the preset correlation coefficient; a trained directional correlation graph model is obtained by combining the cross-type dynamic contribution flow network node forest through the random forest combining simulation algorithm, and the reverse reasoning training is performed from the third layer of cooperative contribution node set to the first layer of target node.

9. The heterogeneous data processing method of energy big data according to claim 1, characterized in that: the training process of the positioning correction path model comprises: acquiring historical cross-type linkage report information and positioning through report analysis to obtain the key target result sequence in the table, the relationship of the key target result synchronous change, the position of the key target result in the table and its contribution impact factor, and through causal analysis, the key target results with correlation change relationship and the corresponding causal correlation strength are obtained; a training optimization directional positioning dynamic correction chain is constructed based on the position of the key target result in the table, the key target results with correlation change relationship and the corresponding causal correlation strength, and is mapped and similarity analyzed with the cross-type dynamic contribution flow network node forest to obtain the similarity consistency error; a forward comprehensive loss function is constructed by analyzing the corresponding parameter relationship between the similarity consistency error, the training optimization directional positioning dynamic correction chain and the cross-type dynamic contribution flow network node forest, and the forward training is performed to obtain a trained positioning correction path model, and the directional correlation graph model is trained in real time and online based on the loss result of the forward training.

10. The heterogeneous data processing method of energy big data according to claim 9, characterized in that: The acquisition history cross-type linkage report information is positioned through report analysis, key target result sequences in the table and the relationship of synchronous changes of key target results, the position of key target results in the table and the contribution impact factor are obtained, and through causal analysis, the key target results with correlated change relationship and the corresponding causal correlation strength are obtained, including: Acquiring historical cross-type linkage report information, through a preset report analysis positioning model, obtaining key target result sequences in the cross-type linkage report and the relationship of synchronous changes of remaining key target results when each key target result produces changes, the position of each key target result in the cross-type linkage report, and the corresponding contribution impact factor of each key target result; Based on the key target result sequence, the relationship of synchronous changes of remaining key target results when each key target result produces changes, and the corresponding contribution impact factor of each key target result, the key target results with correlated change relationship and the corresponding causal correlation strength are obtained through a causal analysis algorithm.

11. The heterogeneous data processing method of energy big data according to claim 9, characterized in that: Based on the position of key target results in the table, the key target results with correlated change relationship and the corresponding causal correlation strength, a training optimized directional positioning dynamic correction chain is constructed, and is mapped and similarity analyzed with the cross-type dynamic contribution flow network node forest to obtain a similarity consistency error, including: Based on the key target results with correlated change relationship and the corresponding causal correlation strength, and the position of each key target result in the report, a training optimized directional positioning dynamic correction chain is constructed by combining a blockchain algorithm; Each key target result in the training optimized directional positioning dynamic correction chain is one-to-one mapped with a target node in the cross-type dynamic contribution flow network node forest at the same time, and according to the corresponding contribution impact factor of each key target result and the main contribution node and the set of collaborative contribution nodes in the cross-type dynamic contribution flow network node forest at the same time point, the similarity coefficient between the contribution impact factor and the main contribution node and the collaborative contribution node and the similarity coefficient between the contribution impact factor and the collaborative contribution node are obtained through a similarity algorithm; The similarity mapping of the contribution impact factor and the main contribution node and the collaborative contribution node is established according to the obtained similarity coefficient, and a similarity consistency error is constructed.

12. The heterogeneous data processing method of energy big data according to claim 9, characterized in that: Through analyzing the similarity consistency error, the corresponding parameter relationship between the training optimized directional positioning dynamic correction chain and the cross-type dynamic contribution flow network node forest, a forward comprehensive loss function is constructed and forward training is performed to obtain a trained positioning correction path model, and a directional association graph model is trained in real time and online based on the loss result of forward training, including: The error between the key target result of the training optimized directional positioning dynamic correction chain positioning and the reverse target result of the same target node obtained by the reverse reasoning of the cross-type dynamic contribution flow network node forest is combined with the similar consistent error, and the error between the target result of each key target result in the training optimized directional positioning dynamic correction chain after the synchronous variation of the remaining key target result and the reverse target result of the remaining key target result in the same target node obtained by the reverse reasoning of the cross-type dynamic contribution flow network node forest is combined with the error between the target result of the remaining key target result after the synchronous variation of the remaining key target result and the reverse target result of the remaining key target result in the same target node obtained by the reverse reasoning of the cross-type dynamic contribution flow network node forest, and the forward comprehensive loss function constructed by the delay of the synchronous variation of the remaining key target result is trained from the first layer corresponding to the target node to the third layer corresponding to the cooperative contribution node set, and a trained positioning correction path model is obtained. The forward comprehensive loss result obtained in each step of the positioning correction path model training is fed back to the directional association graph model for synchronous reverse adjustment training, and a real-time synchronous online trained directional association graph model is obtained.

13. A heterogeneous data processing system for energy big data, which runs the method of any one of claims 1-12. The system comprises: a directional association module configured to obtain different types of energy data and input a pre-trained directional association graph model to obtain a cross-type dynamic contribution flow network node forest; a causal analysis module configured to obtain a cross-type linkage report and construct a basic directional positioning dynamic correction chain through report analysis positioning and causal analysis; a positioning correction module configured to input the cross-type dynamic contribution flow network node forest and the basic directional positioning dynamic correction chain into a pre-trained positioning correction path model, and perform real-time positioning error correction adjustment on the cross-type linkage report to obtain an updated and adjusted cross-type linkage report. 14.A terminal comprising a processor and a storage medium; characterized in that: the storage medium is configured to store instructions; the processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-12.

15. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps of the method according to any one of claims 1-12.

Images