A method, system, device and medium for attributing power marketing indicators based on causal reasoning

Abstract

The application relates to a power marketing index attribution method, system, device and medium based on causal reasoning. The method comprises the following steps: acquiring a power marketing multi-source heterogeneous data set, obtaining a comprehensive feature vector through multi-view feature construction and fusion; inputting the comprehensive feature vector into a decoupling variational autoencoder to separate latent confounding factors and generate a proxy variable sequence; carrying out causal graph structure learning and structure equation model calibration to obtain a structure equation model; based on the structure equation model, stripping the confounding effect of customer satisfaction to obtain a net satisfaction sequence; and through Shapley value variance decomposition of the attribution weight of the power marketing index, outputting the power marketing index attribution result. The method can effectively strip the interference of latent confounding factors, realize fair and quantitative attribution of the contribution degree of the power marketing index, improve the accuracy and interpretability of the attribution result, and provide quantitative decision support for optimal allocation of power marketing resources.

Description

Technical Field

[0001] This invention belongs to the field of data causal analysis, and in particular relates to a method, system, device and medium for attributing electricity marketing indicators based on causal reasoning. Background Technology

[0002] With the deepening of digital transformation in power marketing, power companies have accumulated massive amounts of multi-source, heterogeneous operational data, covering marketing service investment, fault repair work orders, customer satisfaction evaluations, smart meter load curves, customer interaction texts, and social media feedback. How to accurately quantify the true driving effect of various marketing service initiatives on customer satisfaction from this data, and thus provide scientific support for resource allocation decisions, has become a key technical problem that power marketing management urgently needs to solve.

[0003] Existing research in the field of causal inference shows that ignoring potential confounding variables can lead to serious biases in causal effect estimation. This can range from overestimating the effectiveness of management measures to arriving at completely opposite attribution conclusions, directly causing misallocation of marketing resources. To address confounding issues, advanced econometric methods such as instrumental variable methods, differences-in-differences (DID) methods, and regression discontinuities have been proposed. However, these methods typically rely on strict exclusionary assumptions, parallel trend assumptions, or precise discontinuity assumptions, and often require specialized experimental designs. In the context of electricity marketing, marketing activities are routine operational actions rather than randomized experiments. Instrumental variables can hardly simultaneously satisfy the requirements of exogeneity and strong correlation, and differences-in-differences methods struggle to construct clean treatment and control groups. This makes it difficult for traditional econometric methods to reliably isolate the interference of confounding effects.

[0004] A more challenging issue is that weather, macroeconomic fluctuations, and regional social events are typical potential confounding factors in electricity marketing scenarios. These factors directly influence users' electricity consumption behavior and satisfaction ratings, yet they cannot be directly and accurately measured at the micro-level of distribution stations. Current technologies, when faced with such unobservable confounding factors, either ignore their existence, leading to attribution bias, or attempt to find instrumental variables but struggle to meet stringent assumptions. No solution has yet been proposed that can automatically identify proxy variables from multi-source heterogeneous data and effectively isolate their influence when confounding factors are not directly observable. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, system, device, and medium for attributing electricity marketing indicators based on causal reasoning that can achieve confounding effect removal under conditions where no confounding factors are observed, in order to address the above-mentioned technical problems.

[0006] Firstly, this application provides an attribution method for electricity marketing indicators based on causal reasoning, including:

[0007] Data on marketing service input, fault repair time, and customer satisfaction in power marketing, as well as a multimodal operation feature set of power marketing, are obtained to form a multi-source heterogeneous dataset of power marketing. Multi-view features are constructed and fused on the multi-source heterogeneous dataset of power marketing to obtain a comprehensive feature vector.

[0008] The integrated feature vector is input into the decoupled variational autoencoder to separate potential confounding factors and generate surrogate variables, resulting in a surrogate variable sequence.

[0009] Based on the sequence of proxy variables, marketing service investment data, fault repair time data, and customer satisfaction data, causal graph structure learning and structural equation model calibration with proxy variables are performed to obtain the structural equation model.

[0010] Based on structural equation modeling, confounding effects are removed from customer satisfaction data to obtain the net satisfaction series.

[0011] Based on the net satisfaction series, the attribution weights of marketing service input data and fault repair time data are analyzed by Shapley value variance decomposition to obtain the attribution results of electricity marketing indicators.

[0012] Secondly, this application also provides an attribution system for electricity marketing indicators based on causal reasoning, the system comprising:

[0013] The model building module is used to acquire real-time status data of the factory equipment cluster and build a three-dimensional digital twin model of the factory equipment cluster based on the real-time status data.

[0014] The comprehensive feature construction module is used to acquire marketing service input data, fault repair time data, and customer satisfaction data of power marketing, as well as the multimodal operation feature set of power marketing, to form a multi-source heterogeneous dataset of power marketing. The module also performs multi-view feature construction and fusion on the multi-source heterogeneous dataset of power marketing to obtain a comprehensive feature vector.

[0015] The proxy variable generation module is used to input the comprehensive feature vector into the decoupled variational autoencoder to perform potential confounding factor separation and proxy variable generation, and obtain a proxy variable sequence.

[0016] The causal calibration module is used to perform causal graph structure learning and structural equation model calibration based on proxy variable sequences, marketing service input data, fault repair time data, and customer satisfaction data, to obtain a structural equation model.

[0017] The confounding removal module is used to remove confounding effects from customer satisfaction data based on structural equation modeling to obtain a net satisfaction sequence.

[0018] The attribution analysis module is used to perform attribution weight analysis on marketing service input data and fault repair time data based on the net satisfaction series and through Shapley value variance decomposition to obtain the attribution results of electricity marketing indicators.

[0019] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described attribution method for electricity marketing indicators based on causal reasoning.

[0020] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described attribution method for electricity marketing indicators based on causal reasoning.

[0021] The aforementioned attribution method, system, equipment, and medium for electricity marketing indicators based on causal reasoning construct a multi-source heterogeneous dataset for electricity marketing by acquiring marketing service input data, fault repair time data, customer satisfaction data, and multimodal operational feature sets. This data is then processed through multi-view feature construction and fusion to obtain a comprehensive feature vector. This technical approach enables multi-dimensional feature extraction and cross-modal unified fusion representation of multi-source heterogeneous electricity marketing data, forming a standardized feature carrier covering all operational dimensions. It addresses the problems of traditional electricity marketing attribution analysis relying solely on a few core indicators, having a single data source, and failing to effectively integrate and utilize multimodal operational features, resulting in incomplete feature information coverage and a lack of comprehensive feature support for attribution analysis.

[0022] By inputting the comprehensive feature vector into a decoupled variational autoencoder, and performing latent confounding factor separation and surrogate variable generation to obtain a surrogate variable sequence, this technique can achieve the decoupled identification of latent confounding factors in multi-source features and the standardized generation of computable surrogate variables. It can also achieve the technical effect of making unobserved confounding factors observable. This can solve the problem that traditional causal attribution methods cannot identify and remove latent confounding factors hidden in multi-source data, which leads to bias in causal effect estimation due to confounding interference and insufficient authenticity of attribution results.

[0023] By using causal graph structure learning and structural equation model calibration based on proxy variable sequences, marketing service input data, fault repair time data, and customer satisfaction data, a technical means is used to obtain a calibrated structural equation model. This can achieve structured identification of causal relationships among electricity marketing variables and accurate quantification of the strength of action paths. The technical effect of constructing a standardized causal analysis model incorporating heterogeneous control can solve the problems of traditional marketing indicator attribution which is based solely on correlation analysis to deduce influence relationships, fails to clarify the true causal logic and the strength of path effects between variables, and results in attribution conclusions lacking causal support and unable to distinguish between causal associations and spurious correlations.

[0024] This technique, which uses structural equation modeling to remove confounding effects from customer satisfaction data to obtain a net satisfaction sequence, and then uses Shapley variance decomposition to analyze the attribution weights of marketing service input data and fault repair time data to obtain the attribution results of electricity marketing indicators, can achieve the quantification of pure satisfaction after removing confounding effects and the fair and reasonable allocation of the contribution of each electricity marketing indicator. It outputs accurate attribution conclusions with causal basis, which can solve the problems of traditional satisfaction attribution not removing confounding factors, lacking fairness in contribution weight allocation and causal logic support, resulting in large attribution results and failing to effectively guide the optimal allocation of marketing resources. It can improve the accuracy, fairness and interpretability of electricity marketing indicator attribution, and provide a quantitative decision-making basis for the allocation of electricity marketing resources and service optimization. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart of an attribution method for electricity marketing indicators based on causal reasoning, according to the present invention.

[0027] Figure 2 This is a flowchart illustrating the generation of a net satisfaction sequence in one optional embodiment of the present invention;

[0028] Figure 3 This is a structural diagram of an attribution system for electricity marketing indicators based on causal reasoning, according to the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0030] In one embodiment, such as Figure 1 As shown, a causal reasoning-based attribution method for electricity marketing indicators is provided. This embodiment illustrates the application of this method to an attribution analysis terminal. It is understood that this method can also be applied to an attribution analysis server, and further to an attribution analysis system including both an attribution analysis terminal and an attribution analysis server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0031] Step S101: Obtain marketing service input data, fault repair time data, and customer satisfaction data for electricity marketing, as well as a multimodal operation feature set for electricity marketing, to form a multi-source heterogeneous dataset for electricity marketing. Then, construct and fuse multi-view features on the multi-source heterogeneous dataset for electricity marketing to obtain a comprehensive feature vector.

[0032] Optionally, the attribution analysis terminal can obtain marketing service input data, fault repair duration data, customer satisfaction data, and multi-modal operation feature sets of power marketing from the power marketing business system, fault repair work order system, customer satisfaction evaluation system, and multi-source operation data platform, respectively. The attribution analysis terminal can integrate the marketing service input data, fault repair duration data, customer satisfaction data, and multi-modal operation feature sets of power marketing to obtain a multi-source heterogeneous dataset of power marketing. The attribution analysis terminal can construct feature representations of corresponding views according to different data modalities, and then perform cross-modal fusion processing on the multi-view features to obtain a comprehensive feature vector.

[0033] Specifically, the multi-source heterogeneous dataset for power marketing can be a collection of power marketing data covering multiple modalities such as structured operation indicators, time-series load data, and unstructured text data; multi-view feature construction can be a process of using appropriate feature extraction methods for different modalities of data and generating corresponding view feature representations; the comprehensive feature vector can be a feature vector of a unified dimension obtained after the fusion of multi-view features.

[0034] Step S102: Input the comprehensive feature vector into the decoupled variational autoencoder to separate potential confounding factors and generate surrogate variables, thereby obtaining a surrogate variable sequence.

[0035] Optionally, the attribution analysis terminal can input the comprehensive feature vector into the decoupled variational autoencoder, and separate the mutually independent latent confounding factors from the comprehensive feature vector through unsupervised decoupled representation learning; the attribution analysis terminal can combine prior knowledge in the field of electricity marketing to identify the latent dimensions corresponding to the latent confounding factors and generate a sequence of proxy variables.

[0036] Specifically, a decoupled variational autoencoder can be a variational autoencoder model with decoupled representation capabilities, which can map high-dimensional input data to mutually independent latent space dimensions, and achieve automatic separation of different variation factors; latent confounding factors can be unobserved factors hidden in the observed data that simultaneously affect the processing variables and the outcome variables, and latent confounding factors can be the main source of bias in causal attribution; the proxy variable sequence can be an observable time series that characterizes the changing trend of latent confounding factors.

[0037] Step S103: Based on the proxy variable sequence, marketing service investment data, fault repair time data, and customer satisfaction data, perform causal graph structure learning and structural equation model calibration with fused proxy variables to obtain the structural equation model.

[0038] Optionally, the attribution analysis terminal can use the proxy variable sequence as a supplementary control variable, and input the supplementary control variable, marketing service investment data, fault repair time data, and customer satisfaction data into the causal discovery algorithm to learn the causal dependencies between variables and construct a causal graph structure; the attribution analysis terminal can construct a structural equation model based on the path dependencies of the causal graph structure.

[0039] Specifically, causal graph structure learning can be a process of automatically mining causal dependencies and directions of action between variables from observation data. Causal graph structure learning can output a directed graph structure that represents causal logic. Structural equation modeling can be a simultaneous equation model based on causal graphs. Structural equation modeling can be used to quantify the strength of causal effects between variables. Structural equation modeling can be used for confounding effect stripping and causal effect calculation.

[0040] Step S104: Based on the structural equation model, the customer satisfaction data is stripped of confounding effects to obtain the net satisfaction sequence.

[0041] Optionally, the attribution analysis terminal can calculate the contribution of confounding factors to customer satisfaction data based on structural equation modeling and counterfactual reasoning; the attribution analysis terminal can subtract the contribution of confounding factors from the original customer satisfaction observations to obtain the net satisfaction sequence.

[0042] Specifically, confounding effect stripping can be the process of removing the influence of confounding factors on the outcome variable and extracting the pure causal effect of the treatment variable on the outcome variable; the net satisfaction sequence can be the customer satisfaction sequence after removing the interference of confounding factors, and the net satisfaction sequence can be used to reflect the actual driving effect of electricity marketing service initiatives on customer satisfaction.

[0043] Step S105: Based on the net satisfaction sequence, the attribution weights of the marketing service input data and fault repair time data are analyzed through Shapley value variance decomposition to obtain the attribution results of the electricity marketing indicators.

[0044] Optionally, the attribution analysis terminal can use the net satisfaction sequence as the attribution object, perform fair decomposition based on Shapley value theory, and calculate the marginal contribution of each electricity marketing indicator; the attribution analysis terminal can normalize the marginal contribution to obtain the attribution weight of each electricity marketing indicator, and integrate to generate the attribution results of the electricity marketing indicators.

[0045] Specifically, Shapley value variance decomposition can be an attribution method based on the Shapley axiom of cooperative game theory; attribution weights can be the proportion of each electricity marketing indicator's contribution to changes in customer satisfaction, and attribution weights can be used to reflect the degree to which different electricity marketing indicators drive customer satisfaction; the attribution results of electricity marketing indicators can be a comprehensive conclusion of the attribution weights and contributions of each electricity marketing indicator, and the attribution results of electricity marketing indicators can be used to provide a quantitative decision-making basis for the optimal allocation of electricity marketing resources.

[0046] The aforementioned attribution method for electricity marketing indicators based on causal reasoning involves acquiring a multi-source heterogeneous dataset of electricity marketing data, constructing and fusing multi-view features to obtain a comprehensive feature vector, inputting the comprehensive feature vector into a decoupled variational autoencoder to separate potential confounding factors and generate a sequence of proxy variables, conducting causal graph structure learning and structural equation model calibration to obtain a structural equation model, removing confounding effects on customer satisfaction based on the structural equation model to obtain a net satisfaction sequence, and outputting the attribution results of electricity marketing indicators by decomposing the attribution weights of electricity marketing indicators using Shapley variance decomposition. This method effectively removes interference from potential confounding factors, achieves fair quantitative attribution of the contribution of electricity marketing indicators, improves the accuracy and interpretability of the attribution results, and provides quantitative decision support for the optimal allocation of electricity marketing resources.

[0047] In one embodiment, the decoupled variational autoencoder includes an encoder network, a decoder network, and a discriminator branch; the synthesized feature vector is input into the decoupled variational autoencoder to perform latent confounding factor separation and surrogate variable generation, resulting in a surrogate variable sequence, which may include:

[0048] Step S201: Input the comprehensive feature vector into the encoder network to perform feature extraction and distribution parameter calculation, output the mean and log-variance of the latent variable distribution, determine the latent variable distribution based on the mean and log-variance, obtain the latent variable vector from the latent variable distribution through reparameterized sampling, and input the latent variable vector into the decoder network to perform feature decoding and reconstruction, output the reconstructed comprehensive feature vector.

[0049] Optionally, the attribution analysis terminal can input the comprehensive feature vector into the encoder network, extract deep abstract features through multiple nonlinear transformations, and output the mean and log-variance corresponding to the latent variable distribution. The attribution analysis terminal can determine the normal distribution of the latent variables based on the mean and log-variance, and obtain the latent variable vector from the normal distribution through reparameterization. The attribution analysis terminal can input the latent variable vector into the decoder network, reconstruct the original input features through inverse nonlinear transformation, and output the reconstructed comprehensive feature vector.

[0050] Specifically, the encoder network can be a neural network module in a decoupled variational autoencoder responsible for mapping input features to the latent space; the latent variable distribution can be the probability distribution of the input data in the latent space, and the latent variable distribution can take the form of a normal distribution; reparameterization sampling can be a technique for distribution sampling by performing a linear transformation on standard normal noise; the decoder network can be a neural network module responsible for mapping latent variables back to the original feature space, and the decoder network can be used to reconstruct the input data.

[0051] Step S202: Calculate the difference between the integrated feature vector and the reconstructed integrated feature vector to obtain the reconstruction loss; construct a standard normal distribution with zero mean and variance of an identity matrix as the regularization target distribution, and calculate the KL divergence between the latent variable distribution and the regularization target distribution; perform joint probability modeling on each dimension of the latent variable vector to obtain the joint distribution, and multiply the marginal probability modeling on each dimension of the latent variable vector to obtain the marginal distribution product; calculate the difference between the joint distribution and the marginal distribution product to obtain the total correlation penalty term; based on the reconstruction loss, KL divergence, and total correlation penalty term, construct the total loss function, and update the model parameters of the decoupled variational autoencoder through the total loss function to obtain the trained decoupled variational autoencoder.

[0052] Optionally, the attribution analysis terminal can calculate the mean squared error between the integrated feature vector and the reconstructed integrated feature vector to obtain the reconstruction loss; the attribution analysis terminal can construct a standard normal distribution with zero mean and variance of an identity matrix to obtain the regularization target distribution; the attribution analysis terminal can calculate the KL divergence between the latent variable distribution and the regularization target distribution; the attribution analysis terminal can calculate the difference between the product of the joint distribution of the latent variable vector and the marginal distribution of each dimension to obtain the total correlation penalty term; the attribution analysis terminal can weightedly sum the reconstruction loss, KL divergence, and total correlation penalty term to construct the total loss function; the attribution analysis terminal can iteratively update the model parameters of the decoupled variational autoencoder using the gradient descent algorithm combined with the total loss function to obtain the trained decoupled variational autoencoder.

[0053] Specifically, the reconstruction loss can be a loss term that measures the difference between the original data and the reconstructed data. The reconstruction loss can be used to ensure that the core information of the original data is preserved in the latent space. The KL divergence can be an information-theoretic index that measures the degree of difference between two probability distributions. The KL divergence can be used to constrain the posterior distribution to be close to the prior distribution. The total correlation penalty term can be a loss term that measures the degree of dependence between the dimensions of the latent variables. The total loss function can be an optimization objective function obtained by integrating the reconstruction loss, KL divergence and the total correlation penalty term. The total loss function can be used to guide the training iteration of the decoupled variational autoencoder.

[0054] Step S203: After training, a discriminator branch consisting of a gradient inversion layer and a fully connected classifier is connected to the end of the encoder network of the decoupled variational autoencoder. The meteorological observation data in the feature set of multimodal operation of power marketing are discretized to obtain known domain labels. The known domain labels are used as supervision signals to perform adversarial training on the encoder network and the discriminator branch to obtain the adversarial-trained encoder network. The comprehensive feature vector is input into the adversarial-trained encoder network to perform feature extraction and distribution parameter calculation to obtain the final latent representation.

[0055] Optionally, the attribution analysis terminal can connect a discriminator branch consisting of a gradient inversion layer and a fully connected classifier to the output of the encoder network after training and decoupling the variational autoencoder; the attribution analysis terminal can extract meteorological observation data from the multimodal operation feature set of power marketing, and perform hierarchical discretization processing on the meteorological observation data to generate corresponding known domain labels; the attribution analysis terminal can perform adversarial training on the encoder network and the discriminator branch through adversarial training mode to obtain the adversarially trained encoder network; the attribution analysis terminal can input the comprehensive feature vector into the adversarially trained encoder network to perform feature extraction and distribution parameter calculation, and output the decoupled final latent representation.

[0056] Specifically, the discriminator branch can be a discriminator network consisting of a gradient inversion layer and a classifier. The discriminator branch can be used to guide the encoder network to extract domain-invariant features during adversarial training. The gradient inversion layer can be a network layer that inverts the gradient sign during backpropagation. The gradient inversion layer can be used to implement adversarial training between the encoder network and the discriminator branch. The known domain label can be a domain category label generated from observable meteorological data. The known domain label can be used as a supervision signal for adversarial training. The final latent representation can be a latent feature representation after adversarial decoupling. The dimensions of the final latent representation can be independent of each other and the confounding factor dimension can be identified.

[0057] Step S204: Predict the air conditioning load ratio of the smart meter load curves of the multimodal operation feature set of power marketing, obtain the air conditioning load ratio sequence of each substation area-day, and perform sliding window statistics on the air conditioning load ratio sequence in substation area-day units to obtain the non-intrusive air conditioning load decomposition sequence.

[0058] Optionally, the attribution analysis terminal can extract the smart meter load curve from the multimodal operation features of power marketing, and perform non-intrusive decomposition of the daily load curve in the smart meter load curve to predict the daily air conditioning load ratio of each distribution area, and obtain the air conditioning load ratio sequence of each distribution area per day; the attribution analysis terminal can perform sliding window smoothing statistics on the air conditioning load ratio sequence to obtain the non-intrusive air conditioning load decomposition sequence.

[0059] Specifically, "transformer area-day" can be a data collection granularity that uses the power supply area of ​​a single distribution transformer as the spatial unit and a single day as the time unit.

[0060] Step S205: Calculate the mutual information between each dimension in the final potential representation and the meteorological observation variables in the multimodal operation feature set of power marketing, as well as the maximum information coefficient between each dimension and the non-intrusive air conditioning load decomposition sequence. Dimensions that simultaneously satisfy the condition that the mutual information exceeds the first threshold and the maximum information coefficient exceeds the second threshold are identified as mixed dimensions.

[0061] Optionally, the attribution analysis terminal can calculate the mutual information between each dimension of the final potential representation and the meteorological observation variables; the attribution analysis terminal can calculate the maximum information coefficient between each dimension and the non-intrusive air conditioning load decomposition sequence; the attribution analysis terminal can determine the dimension that simultaneously satisfies the condition that the mutual information exceeds the first threshold and the maximum information coefficient exceeds the second threshold as a mixed dimension.

[0062] Specifically, mutual information can be an information-theoretic index that measures the degree of linear and nonlinear dependence between two random variables; the maximum information coefficient can be a statistic that measures the strength of the generalized association between two variables; and the confounding dimension can be the dimension corresponding to the potential confounding factors in the final latent representation.

[0063] Step S206: Extract the one-dimensional continuous time series corresponding to the mixed dimensions, and standardize and normalize the one-dimensional continuous time series to obtain the proxy variable sequence.

[0064] Optionally, the attribution analysis terminal can extract the numerical values ​​corresponding to the confounding dimensions from the final latent representation to obtain a one-dimensional continuous time series; the attribution analysis terminal can perform Z-score standardization and max-min normalization on the one-dimensional continuous time series to obtain a surrogate variable sequence.

[0065] Specifically, the proxy variable sequence can be standardized time series data that characterizes the dynamic changes of potential confounding factors.

[0066] In one embodiment, the expression for the total loss function can be:

[0067] ;

[0068] In the formula, For the total loss function, For the comprehensive feature vector, For the latent variable vector, For the distribution of latent variables, To reconstruct the likelihood function of the loss, To regularize the target distribution, For joint distribution, The first in the latent variable vector dimensional marginal distribution, The product of marginal distributions, and These are the weighting coefficients. Let KL divergence be denoted as KL divergence.

[0069] In one embodiment, based on the sequence of proxy variables, marketing service investment data, fault repair time data, and customer satisfaction data, causal graph structure learning and structural equation model calibration are performed to obtain a structural equation model, which may include:

[0070] Step S301: Aggregate the proxy variable series by month to obtain the monthly average proxy variable; aggregate the marketing service input data by month to obtain the monthly average marketing input variable; aggregate the fault repair time data by month to obtain the monthly average repair time variable; aggregate the customer satisfaction data by month to obtain the monthly average comprehensive satisfaction variable; obtain the observable control variables; integrate the monthly average proxy variable, monthly average marketing input variable, monthly average repair time variable, monthly average comprehensive satisfaction variable, and observable control variables to obtain the analysis variable set.

[0071] Optionally, the attribution analysis terminal can perform monthly average aggregation of the proxy variable series, marketing service investment data, fault repair time data, and customer satisfaction data to obtain monthly average proxy variables, monthly average marketing investment variables, monthly average repair time variables, and monthly average comprehensive satisfaction variables. The attribution analysis terminal can obtain observable control variables such as macroeconomic conditions and user scale, and combine the monthly average proxy variables, monthly marketing investment variables, monthly repair time variables, monthly average comprehensive satisfaction variables, and observable control variables to obtain the set of analytical variables.

[0072] Specifically, the monthly average proxy variable can be the proxy variable aggregated on a monthly average basis; the observable control variable can be the external influencing factor that can be directly observed; and the set of analytical variables can be the entire set of variables used for causal analysis.

[0073] Step S302: Perform the first stage of the PCMCI causal discovery algorithm on the set of analyzed variables, and perform conditional independence testing. Use the kernel test based on Gaussian process as the testing method, delete the undirected edges between the variable pairs that are judged to be conditionally independent, and obtain the causal skeleton graph.

[0074] Optionally, the attribution analysis terminal can run the first stage of the PCMCI algorithm, using a kernel conditional independence test method based on Gaussian processes to test the conditional independence between variables in the variable set one by one; the attribution analysis terminal can delete the connection edges between variable pairs that are determined to be conditionally independent to obtain a causal skeleton graph.

[0075] Specifically, the PCMCI causal discovery algorithm can be a causal discovery method that combines the PC algorithm and the conditional independence test. The PCMCI causal discovery algorithm can be used to effectively handle the identification of causal structures in high-dimensional variables; the causal skeleton graph can be a causal graph structure after deleting the undirected edges between variable pairs that are determined to be conditionally independent.

[0076] Step S303: In the causal skeleton graph, delete the directed edges from the monthly average comprehensive satisfaction variable to the monthly average marketing investment variable and the monthly average emergency repair time variable, delete the directed edges from the monthly average marketing investment variable and the monthly emergency repair time variable to the monthly average proxy variable, and execute the second stage of the PCMCI causal discovery algorithm to perform direction orientation, determine the direction of the remaining edges in the causal skeleton graph, and obtain the enhanced causal graph.

[0077] Optionally, the attribution analysis terminal can impose constraints based on domain prior knowledge, deleting directed edges from the monthly average comprehensive satisfaction variable to the monthly average marketing investment variable and the monthly average emergency repair time variable, as well as directed edges from the monthly average marketing investment variable and the monthly emergency repair time variable to the monthly average proxy variable, to obtain a corrected causal skeleton graph; the attribution analysis terminal can then execute the second stage of the PCMCI algorithm to determine the direction of the remaining edges in the corrected causal skeleton graph, thus obtaining an enhanced causal graph.

[0078] Specifically, an enhanced causal graph can be a causal directed graph that incorporates domain prior constraints and completes edge orientation. Enhanced causal graphs can be used to reflect the causal logical direction between variables.

[0079] Step S304: Based on the enhanced causal graph, the monthly average proxy variable and the observable control variable are used as exogenous variables. A first linear regression equation is established for the monthly average marketing input variable, with the monthly average proxy variable and the observable control variable as predictors. A second linear regression equation is established for the monthly average emergency repair time variable, with the monthly average proxy variable and the observable control variable as predictors. A third linear regression equation is established for the monthly average comprehensive satisfaction variable, with the monthly average marketing input variable, the monthly emergency repair time variable, the monthly average proxy variable, and the observable control variable as predictors. The first, second, and third linear regression equations are combined to form a structural equation model.

[0080] Alternatively, the equation system of the linear structural equation model can be expressed as follows:

[0081] ;

[0082] In the formula, As a variable for average monthly marketing expenditure, The average monthly repair time is a variable. The variable is the average monthly overall satisfaction level. This is the monthly average proxy variable. For observable control variables, and For model parameters, and This is the error term.

[0083] Specifically, structural equation modeling can be a causal quantitative model composed of multiple linear regression equations. Structural equation modeling can be used to simultaneously characterize the effects of confounding factors on treatment variables and the influence paths of all variables on outcome variables.

[0084] In one embodiment, such as Figure 2 As shown, based on structural equation modeling, confounding effects are removed from customer satisfaction data to obtain the net satisfaction series, which may include:

[0085] Step S401: For each data record in the set of analytical variables, calculate the mean of each proxy variable sequence, fix the value of each proxy variable sequence to the mean, and keep the marketing service input data and fault repair time data unchanged. Solve the third linear regression equation to obtain the counterfactual comprehensive satisfaction prediction value of each data record.

[0086] Optionally, the attribution analysis terminal can traverse each data record in the set of analysis variables and calculate the mean of the proxy variable under each data record; the attribution analysis terminal can fix the value of the proxy variable sequence to this mean, while keeping the marketing service investment data and fault repair time data unchanged, and substitute them into the third linear regression equation for prediction to obtain the counterfactual comprehensive satisfaction prediction value corresponding to each data record.

[0087] Specifically, the counterfactual composite satisfaction forecast can be the recorded satisfaction forecast when confounding factors are assumed to be at an average level, used to calculate the individual contribution of confounding factors.

[0088] Step S402: Take the monthly average comprehensive satisfaction variable corresponding to each data record as the comprehensive satisfaction observation value, and calculate the difference between each comprehensive satisfaction observation value and each counterfactual comprehensive satisfaction prediction value to obtain the confounding effect contribution of each data record.

[0089] Optionally, the attribution analysis terminal can use the monthly average comprehensive satisfaction variable corresponding to each data record as the comprehensive satisfaction observation value, and calculate the difference between the comprehensive satisfaction observation value and the corresponding counterfactual comprehensive satisfaction prediction value to obtain the confounding effect contribution of each data record.

[0090] Specifically, the confounding effect contribution can be the magnitude of the impact on customer satisfaction when confounding factors deviate from the average level. The confounding effect contribution can be used to reflect the degree of interference of confounding factors on the overall satisfaction observation.

[0091] Step S403: Subtract the contribution of each confounding effect from each overall satisfaction observation to obtain the net satisfaction sequence.

[0092] Optionally, the attribution analysis terminal can subtract the corresponding confounding effect contribution from the overall satisfaction observation of each data record, arrange the calculation results of all data records in chronological order, and obtain the net satisfaction sequence.

[0093] Specifically, the net satisfaction series can be a customer satisfaction series after removing confounding factors, and the net satisfaction series can be used to reflect the actual effectiveness of electricity marketing efforts.

[0094] In one embodiment, based on the net satisfaction sequence, attribution weight analysis is performed on marketing service input data and fault repair time data through Shapley value variance decomposition to obtain the attribution results of electricity marketing indicators, which may include:

[0095] Step S501: Using the net satisfaction sequence as the dependent variable and the monthly average marketing investment variable, the monthly average emergency repair time variable, and the observable control variable as independent variables, a regression model is constructed, and the determination coefficient of the regression model is calculated to obtain the determination coefficient of the whole model.

[0096] Optionally, the attribution analysis terminal can use the net satisfaction series as the dependent variable and the monthly average marketing investment variable, the monthly average emergency repair time variable, and all observable control variables as independent variables to fit a regression model; the attribution analysis terminal can calculate the determination coefficient of the regression model to obtain the overall model determination coefficient.

[0097] Specifically, the full model determination coefficient can be the determination coefficient of a regression model that includes all independent variables. The full model determination coefficient can be used to measure the extent to which the regression model explains the total variation in net satisfaction.

[0098] Step S502: Combine the monthly average marketing service input variable and the monthly average emergency repair time variable into a feature set. Iterate through all subsets of the feature set to construct a reduced model that does not include the variables in each subset. Calculate the determination coefficient of the reduced model to obtain the determination coefficient of the reduced model. Calculate the difference between the determination coefficient of the full model and the determination coefficient of the reduced model to obtain the marginal explained variance contribution of each subset.

[0099] Optionally, the attribution analysis terminal can form a feature set by combining the monthly average marketing investment variable and the monthly average emergency repair time variable; the attribution analysis terminal can enumerate all possible subsets of the feature set, construct a reduced model for each subset excluding the variables in the subset, and calculate the determination coefficient of the reduced model to obtain the determination coefficient of the reduced model; the attribution analysis terminal can calculate the difference between the determination coefficient of the full model and the determination coefficient of the reduced model to obtain the marginal explained variance contribution of each subset.

[0100] Specifically, the feature set can be a set of electricity marketing indicator variables to be assigned attribution weights; the reduced model can be a regression model constructed after removing some features, and the reduced model can be used to calculate the decrease in explanatory power caused by removing features; the marginal explained variance contribution can be the increase in the explanatory variance of the model after adding variables, and the marginal explained variance contribution can be used to reflect the explanatory power of the variable for the variation of the results.

[0101] Step S503: Based on Shapley's axiom, the marginal explained variance contributions of marketing service input data and fault repair time data on all subsets are weighted and averaged to obtain the attribution weights of the monthly average marketing service input variable and the monthly average repair time variable. The marketing service input data and fault repair time data are then associated with their corresponding attribution weights to generate the attribution results of electricity marketing indicators.

[0102] Alternatively, the formula for calculating attribution weights can be:

[0103] ;

[0104] In the formula, For variables Attribution weights For feature set, for excluding variables a subset of For subset The number of variables in the middle. The total number of variables in the feature set. For subsets The coefficient of determination of the reduced model constructed from the variables in the model. For subsets Variables and variables The coefficient of determination of the entire regression model. This explains the marginal contribution to variance.

[0105] Specifically, attribution weights can be the percentage contribution of each variable to the variation in satisfaction, calculated based on the Shapley value.

[0106] In one embodiment, constructing and fusing multi-view features from a multi-source heterogeneous dataset of electricity marketing data to obtain a comprehensive feature vector may include:

[0107] Step S601: Extract multi-scale time-series features from the smart meter load curves in the multi-modal operation feature set of electricity marketing to obtain load pattern coding.

[0108] Optionally, the attribution analysis terminal can extract the smart meter load curve from the multimodal operation features of power marketing, and perform multi-scale time series transformation on the smart meter load curve to extract load morphology features at different time resolutions and obtain load morphology codes.

[0109] Specifically, load pattern coding can be a low-dimensional vector representation of the load curve of a smart meter obtained by feature extraction.

[0110] Step S602: Perform domain-adaptive semantic encoding on customer interaction text and social media text in the multimodal operation feature set of power marketing, extract sentence vector embeddings, and perform context-aware aggregation on the sentence vector embeddings to obtain a global text semantic representation.

[0111] Optionally, the attribution analysis terminal can use a language model in the power sector to semantically encode customer interaction text and social media text to obtain sentence vector embeddings; the attribution analysis terminal can use an attention aggregation mechanism to weighted aggregate all sentence vectors to obtain a global text semantic representation.

[0112] Specifically, sentence vector embedding can be a semantic vector representation corresponding to a single text; global text semantic representation can be a global text feature vector obtained by aggregating the semantics of all texts, and global text semantic representation can be used to reflect the overall mood and feedback tendency of customers.

[0113] Step S603: Perform non-linear feature compression on the structured fields in the marketing service investment data, fault repair time data, and customer satisfaction data to obtain a structured compressed representation.

[0114] Optionally, the attribution analysis terminal can perform nonlinear transformation and feature crossing on the structured indicators in marketing service investment data, fault repair time data, and customer satisfaction data, and then perform dimensional compression through a fully connected network to obtain a structured compressed representation.

[0115] Specifically, the structured compressed representation can be a feature vector obtained by nonlinearly compressing electricity marketing indicators, while retaining the core information of the structured data.

[0116] Step S604: Weighted fusion of load morphology encoding, global text semantic representation, and structured compressed representation is performed to obtain a comprehensive feature vector.

[0117] Optionally, the attribution analysis terminal can use an adaptive weight fusion strategy to perform weighted concatenation and fusion transformation of load morphology encoding, global text semantic representation and structured compressed representation to obtain a comprehensive feature vector.

[0118] Specifically, the comprehensive feature vector can be a unified feature representation obtained after fusing features from multiple views. The comprehensive feature vector can integrate information from three types of data: time series, text, and structured data.

[0119] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0120] Based on the same inventive concept, this application also provides a causal reasoning-based electricity marketing indicator attribution system for implementing the aforementioned causal reasoning-based electricity marketing indicator attribution method. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more embodiments of the causal reasoning-based electricity marketing indicator attribution system provided below can be found in the limitations of the causal reasoning-based electricity marketing indicator attribution method described above, and will not be repeated here.

[0121] In one exemplary embodiment, such as Figure 3 As shown, an attribution system 500 for electricity marketing metrics based on causal reasoning is provided, including:

[0122] The comprehensive feature construction module 501 can be used to obtain marketing service input data, fault repair time data and customer satisfaction data of power marketing, as well as the multimodal operation feature set of power marketing, to form a multi-source heterogeneous dataset of power marketing, and to construct and fuse multi-view features of the multi-source heterogeneous dataset of power marketing to obtain a comprehensive feature vector.

[0123] The proxy variable generation module 502 can be used to input the comprehensive feature vector into the decoupled variational autoencoder to perform potential confounding factor separation and proxy variable generation, and obtain a proxy variable sequence.

[0124] The causal calibration module 503 can be used to perform causal graph structure learning and structural equation model calibration based on proxy variable sequences, marketing service investment data, fault repair time data and customer satisfaction data, to obtain a structural equation model.

[0125] The confounding removal module 504 can be used to remove confounding effects from customer satisfaction data based on structural equation modeling to obtain a net satisfaction sequence.

[0126] The attribution analysis module 505 can be used to perform attribution weight analysis on marketing service input data and fault repair time data based on the net satisfaction sequence and through Shapley value variance decomposition to obtain the attribution results of electricity marketing indicators.

[0127] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the causal reasoning-based attribution method for electricity marketing metrics as described above.

[0128] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0129] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0130] The above embodiments merely illustrate several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the embodiments of this application, and these all fall within the protection scope of the embodiments of this application.

Claims

1. An attribution method for electricity marketing indicators based on causal reasoning, characterized in that, The method includes: Data on marketing service investment, fault repair time, and customer satisfaction in power marketing, as well as a multimodal operation feature set of power marketing, are acquired to form a multi-source heterogeneous dataset of power marketing. Multi-view features are constructed and fused on the multi-source heterogeneous dataset of power marketing to obtain a comprehensive feature vector. The comprehensive feature vector is input into a decoupled variational autoencoder to separate potential confounding factors and generate proxy variables, thereby obtaining a proxy variable sequence. Based on the proxy variable sequence, the marketing service investment data, the fault repair time data, and the customer satisfaction data, causal graph structure learning and structural equation model calibration are performed to obtain the structural equation model. Based on the structural equation model, confounding effects are removed from the customer satisfaction data to obtain the net satisfaction sequence. Based on the net satisfaction sequence, the attribution weights of the marketing service input data and the fault repair time data are analyzed by Shapley value variance decomposition to obtain the attribution results of the electricity marketing indicators.

2. The method according to claim 1, characterized in that, The decoupled variational autoencoder includes an encoder network, a decoder network, and a discriminator branch; the process of inputting the synthesized feature vector into the decoupled variational autoencoder for latent confounding factor separation and surrogate variable generation yields a surrogate variable sequence, including: The comprehensive feature vector is input into the encoder network for feature extraction and distribution parameter calculation, outputting the mean and log-variance of the latent variable distribution. The latent variable distribution is determined based on the mean and log-variance, and the latent variable vector is obtained by reparameterized sampling from the latent variable distribution. The latent variable vector is then input into the decoder network for feature decoding and reconstruction, outputting the reconstructed comprehensive feature vector. The difference between the comprehensive feature vector and the reconstructed comprehensive feature vector is calculated to obtain the reconstruction loss. A standard normal distribution with zero mean and variance of an identity matrix is ​​constructed as the regularization target distribution, and the KL divergence between the latent variable distribution and the regularization target distribution is calculated. Joint probability modeling is performed on each dimension of the latent variable vector to obtain the joint distribution. Marginal probability modeling is performed on each dimension of the latent variable vector, and the results are multiplied to obtain the marginal distribution product. The difference between the joint distribution and the marginal distribution product is calculated to obtain the total correlation penalty term. Based on the reconstruction loss, the KL divergence, and the total correlation penalty term, a total loss function is constructed, and the model parameters of the decoupled variational autoencoder are updated using the total loss function to obtain the trained decoupled variational autoencoder. The discriminator branch, consisting of a gradient inversion layer and a fully connected classifier, is connected to the end of the encoder network of the decoupled variational autoencoder after training. The meteorological observation data in the multimodal operation feature set of power marketing is discretized to obtain known domain labels. The known domain labels are used as supervision signals to perform adversarial training on the encoder network and the discriminator branch to obtain the adversarially trained encoder network. The comprehensive feature vector is input into the adversarially trained encoder network for feature extraction and distribution parameter calculation to obtain the final latent representation. The air conditioning load ratio is predicted by the smart meter load curve of the multimodal operation feature set of the power marketing, and the air conditioning load ratio sequence of each distribution area is obtained. The air conditioning load ratio sequence is statistically analyzed by sliding window with the distribution area as the unit to obtain a non-intrusive air conditioning load decomposition sequence. The distribution area as the unit is the data collection granularity with the power supply area of ​​a single distribution transformer as the spatial unit and the single day as the time unit. The mutual information between each dimension in the final potential representation and the meteorological observation variables in the multimodal operation feature set of power marketing is calculated, as well as the maximum information coefficient between each dimension and the non-intrusive air conditioning load decomposition sequence. The dimension that simultaneously satisfies the mutual information exceeding the first threshold and the maximum information coefficient exceeding the second threshold is identified as a hybrid dimension. The one-dimensional continuous time series corresponding to the mixed dimension is extracted, and the one-dimensional continuous time series is standardized and normalized to obtain the proxy variable sequence.

3. The method according to claim 2, characterized in that, The expression for the total loss function is: ； In the formula, Let the total loss function be... The comprehensive feature vector, Let be the latent variable vector. For the distribution of the latent variables, Let be the likelihood function of the reconstruction loss. For the regularization target distribution, For the joint distribution, For the th in the latent variable vector dimensional marginal distribution, The product of the edge distributions, and These are the weighting coefficients. Let KL be the divergence.

4. The method according to claim 1, characterized in that, The process involves learning the causal graph structure of the fused proxy variables and calibrating the structural equation model based on the proxy variable sequence, the marketing service investment data, the fault repair time data, and the customer satisfaction data, to obtain the structural equation model, including: The proxy variable sequence is aggregated monthly to obtain the monthly average proxy variable; the marketing service investment data is aggregated monthly to obtain the monthly average marketing investment variable; the fault repair time data is aggregated monthly to obtain the monthly average repair time variable; and the customer satisfaction data is aggregated monthly to obtain the monthly average comprehensive satisfaction variable. Observable control variables are then obtained, and the monthly average proxy variable, the monthly average marketing investment variable, the monthly average repair time variable, the monthly average comprehensive satisfaction variable, and the observable control variables are integrated to obtain the analysis variable set. The first stage of the PCMCI causal discovery algorithm is performed on the set of analytical variables. Conditional independence is tested, and a kernel test based on Gaussian process is used as the test method. Undirected edges between variable pairs that are determined to be conditionally independent are deleted to obtain a causal skeleton graph. In the causal skeleton graph, the directed edges from the monthly average comprehensive satisfaction variable to the monthly average marketing investment variable and the monthly average emergency repair time variable are deleted, and the directed edges from the monthly average marketing investment variable and the monthly emergency repair time variable to the monthly average proxy variable are also deleted. Then, the second stage of the PCMCI causal discovery algorithm is executed to perform direction orientation and determine the direction of the remaining edges in the causal skeleton graph to obtain an enhanced causal graph. Based on the enhanced causal graph, the monthly average proxy variable and the observable control variable are used as exogenous variables. A first linear regression equation is established for the monthly average marketing investment variable, with the monthly average proxy variable and the observable control variable as predictors. A second linear regression equation is established for the monthly average emergency repair time variable, with the monthly average proxy variable and the observable control variable as predictors. A third linear regression equation is established for the monthly average overall satisfaction variable, with the monthly average marketing investment variable, the monthly average emergency repair time variable, the monthly average proxy variable, and the observable control variable as predictors. The first linear regression equation, the second linear regression equation, and the third linear regression equation are combined to form the structural equation model. The equations of the linear structural equation model are expressed as follows: ； In the formula, Let the monthly average marketing expenditure variable be... Let the average monthly repair time be the variable. The monthly average overall satisfaction variable is... For the monthly average proxy variable, For the observable control variable, and For model parameters, and This is the error term.

5. The method according to claim 4, characterized in that, Based on the structural equation model, the customer satisfaction data is stripped of confounding effects to obtain a net satisfaction sequence, including: For each data record in the set of analytical variables, the mean of each proxy variable sequence is calculated, the value of each proxy variable sequence is fixed to the mean, and the marketing service investment data and the fault repair time data are kept unchanged. The third linear regression equation is solved to obtain the counterfactual comprehensive satisfaction prediction value of each data record. The monthly average comprehensive satisfaction variable corresponding to each of the data records is taken as the comprehensive satisfaction observation value, and the difference between each comprehensive satisfaction observation value and each counterfactual comprehensive satisfaction prediction value is calculated to obtain the confounding effect contribution of each of the data records. The net satisfaction sequence is obtained by subtracting the contribution of each confounding effect from each of the overall satisfaction observations.

6. The method according to claim 4, characterized in that, Based on the net satisfaction sequence, the marketing service input data and the fault repair time data are analyzed by Shapley variance decomposition to obtain the attribution results of the electricity marketing indicators, including: Using the net satisfaction sequence as the dependent variable and the monthly average marketing investment variable, the monthly average emergency repair time variable, and the observable control variable as independent variables, a regression model is constructed, and the determination coefficient of the regression model is calculated to obtain the overall model determination coefficient. The monthly average marketing service input variable and the monthly average emergency repair time variable are combined into a feature set. All subsets of the feature set are traversed to construct a reduced model that does not include the variables in each subset. The determination coefficient of the reduced model is calculated to obtain the determination coefficient of the reduced model. The difference between the determination coefficient of the full model and the determination coefficient of the reduced model is calculated to obtain the marginal explained variance contribution of each subset. Based on Shapley's axiom, the marginal explained variance contributions of the marketing service input data and the fault repair time data are weighted and averaged across all subsets to obtain the attribution weights of the monthly average marketing service input variable and the monthly average repair time variable. The marketing service input data and the fault repair time data are then associated with their corresponding attribution weights to generate the attribution results of the electricity marketing indicators. The formula for calculating the attribution weights is as follows: ； In the formula, For variables The attribution weights, For the set of features, for The variable not mentioned The subset mentioned above, For the subset The number of variables in the middle. The total number of variables in the feature set. For subsets The coefficient of determination of the reduced model constructed from the variables in the model. For subsets The variables in the middle and the variables mentioned The coefficient of determination of the entire model of the constructed regression model. This is the marginal explained variance contribution.

7. The method according to claim 2, characterized in that, The process of constructing and fusing multi-view features from the multi-source heterogeneous dataset of electricity marketing to obtain a comprehensive feature vector includes: Multi-scale time-series feature extraction is performed on the smart meter load curves in the multi-modal operation feature set of the electricity marketing to obtain load pattern coding; Domain-adaptive semantic encoding is performed on customer interaction text and social media text in the multimodal operation feature set of power marketing to extract sentence vector embeddings, and context-aware aggregation is performed on the sentence vector embeddings to obtain a global text semantic representation; The structured fields in the marketing service investment data, the fault repair time data, and the customer satisfaction data are subjected to non-linear feature compression to obtain a structured compressed representation. The comprehensive feature vector is obtained by weighted fusion of the load morphology encoding, the global text semantic representation, and the structured compressed representation.

8. An attribution system for electricity marketing indicators based on causal reasoning, characterized in that, The system includes: The comprehensive feature construction module is used to acquire marketing service input data, fault repair time data, and customer satisfaction data of power marketing, as well as the multimodal operation feature set of power marketing, to form a multi-source heterogeneous dataset of power marketing. The module then performs multi-view feature construction and fusion on the multi-source heterogeneous dataset of power marketing to obtain a comprehensive feature vector. The proxy variable generation module is used to input the comprehensive feature vector into the decoupled variational autoencoder to perform potential confounding factor separation and proxy variable generation, thereby obtaining a proxy variable sequence. The causal calibration module is used to perform causal graph structure learning and structural equation model calibration based on the proxy variable sequence, the marketing service investment data, the fault repair time data and the customer satisfaction data, to obtain the structural equation model. The confounding removal module is used to remove confounding effects from the customer satisfaction data based on the structural equation model to obtain a net satisfaction sequence. The attribution analysis module is used to perform attribution weight analysis on the marketing service input data and the fault repair time data based on the net satisfaction sequence through Shapley value variance decomposition, so as to obtain the attribution results of the electricity marketing indicators.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images