Power data implicit inference source identification method and device, equipment and storage medium

Abstract

The application relates to the technical field of electric power data analysis, and in particular to an electric power data implicit inference source identification method and device, equipment and a storage medium. By configuring a learnable initial inference contribution coefficient for each candidate field, the model parameters of an inference network model and the inference contribution coefficient are trained in the same framework during a training process based on an electric power data set, the importance of each candidate field is quantified as an explicit parameter that can be directly regulated by an optimization process, and an identification mechanism oriented to a multi-field joint inference capability is realized. Further, multi-stage successive training is performed along an increasing sparse intensity sequence, the actual exit order of each candidate field is used as a hierarchical basis, the exit critical intensity of each field is obtained for hierarchical classification, the problem that a traditional method relies on artificial setting of a judgment position and can only output a single binary subset is solved, an identification result with a strong and weak hierarchy is output, and more accurate decision-making basis is provided for electric power data analysis.

Description

Technical Field

[0001] This application relates to the field of power data analysis technology, and in particular to a method, apparatus, device and storage medium for identifying implicit inference sources of power data. Background Technology

[0002] Operational and market data in power systems are jointly determined by physical constraints, dispatch strategies, and market clearing mechanisms. Due to the extensive statistical coupling relationships between fields, in practical data sharing and hierarchical management scenarios, certain fields, such as exchange power and system marginal electricity price, are typically classified as highly sensitive and restricted from external access due to their involvement in system operation security or market confidentiality. These fields are referred to as implicit data sources or implicit inference sources. However, third parties may still obtain other public fields, such as various types of generation output, load, and reserve capacity, and use machine learning models to uncover the dependencies between public fields and implicit inference sources, thereby inferring the implicit inference sources of these highly sensitive fields.

[0003] Traditional methods typically employ two strategies. The first involves establishing manual judgment rules based on names and business meanings, or calculating the correlation coefficient between publicly disclosed fields and target inferred fields. This approach suffers from strong subjectivity and fails to reflect the relationships between multiple fields. The second strategy utilizes a trained prediction model to progressively screen the features of publicly disclosed fields. However, this approach is not specifically designed for the hierarchical nature of joint inference relationships between fields in power data scenarios, making it difficult to meet the practical needs of refined classification of inference sources in hierarchical management. Summary of the Invention

[0004] In view of the above-mentioned shortcomings and deficiencies of the prior art, this application provides a method, device, equipment and storage medium for identifying implicit inference sources of power data. The main purpose is to solve the problems in the current traditional methods for identifying implicit inference sources, such as the inability of static statistical indicators to reflect the joint inference capability of multiple fields and the dependence of the identification results on human-preset rules.

[0005] To achieve the above objectives, the main technical solutions adopted in this application include: In a first aspect, embodiments of this application provide a method for identifying implicit inference sources of power data, including: Obtain a power dataset containing historical samples from multiple time points, wherein the power dataset includes power data for candidate fields and power data for target fields; Before inputting the power dataset into the initial inference network model, configure the initial inference contribution coefficients that have a mapping relationship with each candidate field; Using the target field as the prediction objective, the initial inference network model is trained using the power dataset. After each model parameter update, a normalization process is performed on the first hidden layer of the initial inference network model to suppress the shrinkage of the inference contribution coefficient effectiveness loss, resulting in an inference network model containing the inference contribution coefficients updated after training. The inference contribution coefficients updated after training are used to characterize the contribution of the corresponding candidate field in predicting the target field. The inference network model is subjected to multi-stage sequential training along a preset sparse intensity sequence arranged in ascending order to obtain the critical intensity corresponding to each candidate field. In the multi-stage sequential training, the inference contribution coefficient is selectively shrunk with the current sparse intensity as a constraint in each stage. At the end of each stage, the survival status of each candidate field is determined according to the inference contribution coefficient of each candidate field in the current stage. The inference contribution coefficient of the candidate field determined to be in a non-survival state is set to zero and the corresponding sparse intensity is recorded as the critical intensity. The candidate fields are hierarchically divided according to the critical intensity and output as the identification result of the latent inference source of the target field.

[0006] Optionally, the model parameters include inference network parameters, and the initial inference network includes an input layer and multiple fully connected hidden layers following the input layer. Using the target field as the prediction objective, the initial inference network model is trained using the power dataset. After each model parameter update, the first hidden layer of the initial inference network model undergoes normalization to suppress the shrinkage of the inference contribution coefficient's effectiveness loss, resulting in an inference network model containing the inference contribution coefficients after training and updating. This includes: using the target field in the power dataset as the training label, using each candidate field and its corresponding inference contribution coefficient as the input vector of the initial inference network model, and outputting the predicted value of each candidate field relative to the target field to train the inference network parameters of the initial inference network model; wherein, during the training of the inference network parameters of the initial inference network model, after each inference network parameter update, the column vectors corresponding to each input vector in the weight matrix of the first hidden layer of the inference network model undergo unit norm normalization.

[0007] Optionally, the inference network model is subjected to multi-stage sequential training along a preset sparse intensity sequence arranged in ascending order, including: obtaining a preset increasing sparse intensity sequence [λ1, λ2, ..., λs], where λ1 < λ2 < ... < λs, and s is a positive integer; using the inference network model as the training starting point for the first scanning stage, training is performed using a composite loss function until convergence; wherein, the composite loss function includes a sparse penalty term L1 that only applies to the inference contribution coefficient, and the penalty coefficient of the sparse penalty term L1 is λ1; from the second scanning stage onwards, the model parameters after the convergence of the previous stage are used as the starting point for sequential training, and the penalty coefficient in the composite loss function of each stage is respectively taken as the sparse intensity λs corresponding to the current stage.

[0008] Optionally, the composite loss function Including the inference error term and the sparse penalty term L1 that only applies to the inference contribution coefficient; it can be expressed by the following formula: ; Where, N tr The number of candidate field data in the power data set. The target field label value for the training sample of the k-th candidate field. c is the predicted value corresponding to the k-th candidate field. j For the corresponding inferred contribution coefficient, the first term is the inference error term in the form of mean square error, and the second term is the sparse penalty term L1 that only applies to the inferred contribution coefficient.

[0009] Optionally, performing multi-stage sequential training on the inference network model along a preset sparse intensity sequence arranged in ascending order further includes: at the end of each scanning stage, obtaining statistics on the inference contribution coefficients of each candidate field within the convergence window of the current scanning stage; determining whether each candidate field is in a live state based on the relationship between the statistics and a preset tolerance; fixing the inference contribution coefficients corresponding to candidate fields in a non-live state to zero and stopping updates in subsequent scanning stages, while recording the sparse intensity of the current stage as the critical intensity of the candidate fields in the current non-live state; and retaining the inference contribution coefficients of candidate fields in a live state for the next scanning stage.

[0010] Optionally, the statistics include the mean and standard deviation of the absolute values ​​of the inferred contribution coefficients within the convergence window, and the preset tolerances include a zero-tolerance and a stability tolerance. Determining whether each candidate field is in a live state based on the relationship between the statistics and the preset tolerances includes: calculating the mean and standard deviation of the absolute values ​​of the inferred contribution coefficients of the current candidate field within the convergence window of the current scanning phase; if the mean is less than or equal to the zero-tolerance and the standard deviation is less than or equal to the stability tolerance, the current candidate field is determined to be in a non-live state; otherwise, it is in a live state.

[0011] Optionally, before inputting the power dataset into the initial inference network model, the method further includes: using the equal-width bin histogram estimation method to calculate the mutual information between each candidate field and the target field, and normalizing the mutual information according to information entropy to obtain normalized mutual information; and removing candidate fields whose normalized mutual information is lower than a preset initial screening threshold.

[0012] Secondly, embodiments of this application provide a device for identifying implicit inference sources of power data, comprising: The acquisition unit is configured to acquire a power dataset containing historical samples from multiple time points, wherein the power dataset includes power data for candidate fields and power data for target fields; The configuration unit is configured to configure initial inference contribution coefficients that have a mapping relationship with each candidate field before inputting the power dataset into the initial inference network model; The first training unit is configured to train the initial inference network model using the target field as the prediction objective, through the power dataset, and after each model parameter update, perform normalization processing on the first hidden layer of the initial inference network model to suppress the shrinkage of the inference contribution coefficient effectiveness loss, thereby obtaining an inference network model containing the inference contribution coefficients updated after training; the inference contribution coefficients updated after training are used to characterize the contribution of the corresponding candidate field in predicting the target field. The second training unit is configured to perform multi-stage sequential training on the inference network model along a preset sparse intensity sequence arranged in ascending order. In the multi-stage sequential training, the inference contribution coefficient is selectively shrunk at each stage with the current sparse intensity as a constraint. At the end of each stage, the survival status of each candidate field is determined based on the inference contribution coefficient of each candidate field in the current stage. The inference contribution coefficient of the candidate fields determined to be in a non-survival state is set to zero and the corresponding sparse intensity is recorded as the critical intensity to obtain the critical intensity corresponding to each candidate field. The output unit is configured to hierarchically divide the candidate fields according to the critical intensity and output the results as the identification results of the latent inference source of the target field.

[0013] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for identifying implicit inference sources of power data as described in the first aspect.

[0014] Fourthly, this application provides an electronic device, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the computer program to implement the power data implicit inference source identification method described in the first aspect.

[0015] By employing the above technical solution, this application provides a method for identifying implicit inference sources in power data. Firstly, it configures learnable initial inference contribution coefficients for each candidate field. Then, during the training process based on the power dataset, the model parameters and inference contribution coefficients of the inference network model are trained within the same framework. This quantifies the importance of each candidate field into explicit parameters that can be directly controlled by the optimization process, realizing an identification mechanism guided by multi-field joint inference capabilities. Furthermore, multi-stage sequential training is performed along an increasing sparse intensity sequence, using the actual exit order of each candidate field as the grading basis. This obtains the exit critical intensity of each field for grading, eliminating the need for preset pruning thresholds or retention numbers. This solves the problem of traditional feature selection methods relying on manually set judgment positions and only outputting a single binary subset. Finally, it outputs identification results with varying strength levels, providing a more refined decision-making basis for power data analysis. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a method for identifying implicit inference sources of power data, provided in an embodiment of this application; Figure 2 A flowchart illustrating another method for identifying implicit sources of power data provided in this application embodiment; Figure 3 A schematic diagram of an inference model structure with inference contribution coefficients provided in an embodiment of this application; Figure 4 This is a schematic diagram of a device for identifying implicit inference sources of power data, provided in an embodiment of this application. Detailed Implementation

[0017] To better understand the above technical solutions, exemplary embodiments of this application will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of this application are shown in the drawings, it should be understood that this application can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application can be understood more clearly and thoroughly, and that the scope of this application can be fully conveyed to those skilled in the art.

[0018] Based on the aforementioned background technology, traditional methods currently mainly fall into two categories. One is based on static rules and statistical indicators. These methods involve manually formulating judgment rules based on field names, business meanings, and operational experience, or calculating statistical indicators such as correlation coefficients and mutual information between candidate and target fields. The fields are then sorted by indicator size or filtered by thresholds. Both rules and indicators are statically given before modeling. In this approach, static rules and statistical indicators are not driven by inference capabilities. Since the rules and indicators are statically given before modeling and do not adjust with the inference task, the manual rules are limited by subjective experience. Correlation coefficients and mutual information only measure the degree of dependence between two variables and cannot reflect the recovery capability under the combined effect of multiple fields. Fields with weak individual correlations may play a significant role in joint inference, and these indicators lack objective judgment thresholds when directly determining the source of inference.

[0019] Another approach is feature selection based on machine learning. These methods borrow from model pruning, first training a prediction model from candidate fields to target fields, then selecting fields based on feature importance: sorting by scores such as tree model importance and network connection weight magnitude, and pruning fields below a set threshold or ranked low; or setting a binary mask with values ​​of 0 or 1 for each feature, retaining or removing the corresponding feature based on the mask's open / closed state; or using recursive feature elimination to remove the least important fields round by round until a preset number is retained. However, the problem is that the final result is usually still a subset of binary fields, and the binary selection position needs to be predetermined. Ultimately, only a single field subset at a certain decision position can be obtained, and a multi-level inference source sequence with clear decision criteria cannot be directly formed. Furthermore, the location of the binary selection depends on the manually set pruning threshold, the number of retained fields, or the number of elimination rounds, while the number of fields constituting the implicit inference source is unknown beforehand, and different parameter values ​​may change the final field set.

[0020] To address the shortcomings of traditional implicit inference source identification methods, such as static statistical indicators failing to reflect the joint inference capability of multiple fields, reliance on manually preset pruning thresholds or retention numbers, and the inability to output only a single field subset without a strong-weak hierarchy, this application proposes a method for identifying implicit inference sources in power data. This method can be applied to scenarios such as hierarchical management of power system data, data sharing security review, and power market data correlation analysis. The execution entity of this method can be a power data implicit inference source identification device or a corresponding processing system. By configuring learnable inference contribution coefficients for each candidate field and performing multi-stage sequential training along an increasing sparse intensity sequence, each field differentiates and exits sequentially according to its actual contribution strength to the predicted target field, ultimately outputting a hierarchical inference source sequence with strong-weak hierarchy. Specifically, any of the power data implicit inference source identification methods mentioned below can be executed during runtime. Figure 1 As shown, the method includes: S101, Obtain the power dataset containing historical samples from multiple time points.

[0021] The power dataset includes power data for candidate fields and power data for target fields. The dataset records the operating status and market data of the power system at different times, such as the output of each generating unit, system load, nodal prices, and exchange power. The dataset contains two types of fields: candidate fields, which are all power data fields that could potentially be used to infer the target field; and target fields, such as fields that are set to a high sensitivity level and restricted from external access—the power data fields whose potential sources of inference need to be identified. In a specific example, fields such as "generating output," "load," "electricity price," and "reserve demand," along with their corresponding values, are candidate fields, while "net actual exchange power" is the target field whose value needs to be predicted.

[0022] S102, Before inputting the power dataset into the initial inference network model, configure the initial inference contribution coefficients that have a mapping relationship with each candidate field.

[0023] In S102, the inference contribution coefficient is a learnable parameter variable, corresponding one-to-one with each candidate field. Initially, all inference contribution coefficients are set to 1, and their magnitude indicates the degree to which the corresponding candidate field is preserved when predicting the target field. If the inference contribution coefficient of a candidate field approaches zero, the input channel for that field is blocked, meaning that field does not participate in the prediction of the target field. If the coefficient is large, the information of that field is preserved and used for prediction. By configuring an independent learnable coefficient for each candidate field, field importance is explicitly extracted, becoming a value that can be directly controlled by the optimization process, rather than being difficult to observe as field importance is dispersed across multiple layers of weights in traditional neural networks.

[0024] S103, with the target field as the prediction objective, the initial inference network model is trained using the power dataset. After each model parameter update, the first hidden layer of the initial inference network model is normalized to suppress the shrinkage of the inference contribution coefficient effectiveness loss, resulting in an inference network model containing the inference contribution coefficients after training and updating.

[0025] Using the target field as the prediction objective, the electricity dataset obtained in step S101 is used to train the initial inference network model. Specifically, the electricity data for each candidate field is multiplied by its corresponding inference contribution coefficient, and then input into the initial inference network model. The output is the predicted value for the target field, and the true value of the target field is used as a supervision signal to train the model. During this training process, on the one hand, the inference contribution coefficient corresponding to each candidate field is continuously optimized as a learnable parameter of the model, and on the other hand, the model parameters of the inference network model are also continuously trained.

[0026] The initial inference network model uses a multi-layer fully connected neural network with an input layer of dimension q, followed by several fully connected hidden layers. Each hidden layer uses the ReLU activation function, and the output layer has a dimension of 1 and no activation function. The role of the inference network model is to learn the mapping relationship between the candidate field and the target field from the input of the candidate field. The training objective is to make the predicted value as close as possible to the true value. After training and updating, the inference contribution coefficient is used to characterize the degree of contribution of the corresponding candidate field in predicting the target field.

[0027] Furthermore, during the training of the inference network model in S103, after each model parameter update, a normalization process is performed on the first hidden layer of the initial inference network model to suppress the shrinkage effectiveness loss of the inference contribution coefficient. Since the model training process trains both the inference contribution coefficient and the model parameters of the inference network model, a coupling relationship exists between the inference contribution coefficient and the weights of the first hidden layer of the inference network model. That is, when the inference contribution coefficient is reduced by sparsity penalty, the corresponding input weight column of the first hidden layer may recover the signal strength of that field by proportionally amplifying it, thus effectively offsetting the shrinkage effect of the inference contribution coefficient. Therefore, after each model parameter update, performing a normalization process on the first hidden layer of the inference network model to suppress the shrinkage effectiveness loss of the inference contribution coefficient limits the total signal amplification factor of each input field through the first hidden layer. This prevents the shrinkage of the inference contribution coefficient from being offset by the reverse compensation of the network weights, thereby ensuring that the inference contribution coefficient, after training and updating, can truly and reliably represent the contribution of the corresponding candidate field in predicting the target field.

[0028] S104, perform multi-stage sequential training on the inference network model along the preset sparse intensity sequence arranged in ascending order to obtain the critical intensity corresponding to each candidate field.

[0029] In the multi-stage sequential training, the inferred contribution coefficient is selectively shrunk with the current sparsity intensity as a constraint in each scanning stage. At the end of each scanning stage, the survival status of each candidate field is determined based on the inferred contribution coefficient of each candidate field in the current stage. The inferred contribution coefficient of the candidate field determined to be in a non-survival state is set to zero and the corresponding sparsity intensity is recorded as the critical intensity.

[0030] In S104, firstly, a sparse intensity sequence is preset in ascending order. This sequence is a set of values ​​arranged from smallest to largest, such as [λ1, λ2, ..., λs], where λ1 < λ2 < ... < λs, and s is a positive integer. Each value in the sequence represents the sparsity penalty applied to the inference contribution coefficient in a scanning phase. This sparse intensity sequence acts as a progressively tightening screening criterion: the smaller the value, the looser the screening, and most fields can survive; the larger the value, the stricter the screening, and only fields that contribute strongly to the prediction target field can survive, while fields that contribute weakly are eliminated. The ascending order ensures that the elimination process for each field progresses gradually from weak to strong, rather than being compressed all at once.

[0031] Furthermore, multi-stage sequential training is performed based on this sequence. Starting from the first scan stage, the sparsity intensity of the current stage is used as the penalty coefficient of the sparsity penalty term in the loss function to train the inference network model. During training, the L1 sparsity penalty term applies a contraction effect to each inference contribution coefficient, but the ability of each coefficient to resist contraction depends on its actual contribution to the predicted target field: fields with larger contributions have coefficients that resist contraction more effectively under the constraint of the inference error term, maintaining a larger amplitude; fields with smaller contributions lack the constraint of the error term, and their coefficients tend to zero more quickly under the penalty effect. From the second scan stage onwards, training continues using the model parameters converged from the previous stage as the starting point.

[0032] Furthermore, at the end of each stage, the inference contribution coefficients of each candidate field within the convergence window of the current stage are obtained, and this is used to determine the survival status of each candidate field. Survival status refers to whether the field retains an actual inference contribution under the current sparsity intensity. For example, a survival status indicates that the field still retains an actual inference contribution under the current sparsity intensity and is not considered a field to be eliminated; a non-survival status indicates that the inference contribution coefficient of the field has steadily shrunk to near zero and can exit the inference process. The sparsity intensity at which the field exits is then used as the critical intensity for that field, thus obtaining the critical intensity corresponding to each candidate field. Therefore, through the above method, each candidate field obtains a corresponding critical intensity value, which expresses the elimination process of each field from weak to strong, that is, its correlation with the target field.

[0033] S105: The candidate fields are hierarchically divided according to the critical intensity and output as the identification result of the latent inference source of the target field.

[0034] In this embodiment, by configuring a learnable initial inference contribution coefficient for each candidate field, the model parameters and inference contribution coefficients of the inference network model are trained within the same framework during the training process based on the power dataset. This quantifies the importance of each candidate field into explicit parameters that can be directly controlled by the optimization process, realizing a recognition mechanism guided by multi-field joint inference capabilities. Furthermore, multi-stage sequential training is performed along an increasing sparse intensity sequence, using the actual exit order of each candidate field as the grading basis. This obtains the exit critical intensity of each field for grading, eliminating the need for preset pruning thresholds or retention numbers. This solves the problem that traditional feature selection methods rely on manually setting the decision position and can only output a single binary subset. The final output is a recognition result with strong and weak hierarchies, providing a more refined decision-making basis for the analysis of power data.

[0035] Optionally, the model parameters include inference network parameters, and the initial inference network includes an input layer and multiple fully connected hidden layers following the input layer.

[0036] In S103, the target field is used as the prediction objective. The initial inference network model is trained using the power dataset. After each model parameter update, the first hidden layer of the initial inference network model is normalized to suppress the shrinkage of the inference contribution coefficient effectiveness loss, resulting in an inference network model containing the inference contribution coefficients after training and updating. This includes: using the target field in the power dataset as the training label, using each candidate field and its corresponding inference contribution coefficient as the input vector of the initial inference network model, and outputting the predicted value of each candidate field relative to the target field to train the inference network parameters of the initial inference network model; wherein, during the training of the inference network parameters of the initial inference network model, after each inference network parameter update, the column vectors corresponding to each input vector in the weight matrix of the first hidden layer of the inference network model are normalized using the unit norm.

[0037] In this embodiment, the inference network parameters are the model parameters of the inference network model mentioned above. Specifically, they may include the weight matrix and bias vector of each fully connected layer, which are continuously updated through backpropagation during training. Each candidate field and its corresponding inference contribution coefficient are used as the input vector of the initial inference network model. Specifically, the power data of each candidate field can be multiplied by its corresponding inference contribution coefficient and then fed into the inference network as the input vector. This outputs the predicted value for the target field, and the prediction error is calculated using the true value of the target field as the training label. Backpropagation is then used to simultaneously optimize both the inference contribution coefficient and the inference network parameters.

[0038] During this process, after each update of the inference network parameters, unit norm normalization is performed on the column vectors corresponding to each input vector in the weight matrix of the first hidden layer of the inference network model. This specifically constrains the weight columns of the first hidden layer that are directly connected to each input field: since each input field corresponds to only one column in the weight matrix of the first hidden layer, the unit norm constraint on this column vector limits the total signal amplification factor of that input field after passing through the first hidden layer. This prevents the first hidden layer from recovering the signal strength by proportionally amplifying the weight column corresponding to the input field when the inference contribution coefficient is reduced by the sparsity penalty, thus blocking the path of the network to compensate for coefficient shrinkage by increasing the weight scale. Since only the first hidden layer has a direct and one-to-one connection with the original input fields, applying column vector norm constraints to this layer can effectively eliminate scaling coupling at the input channel level, ensuring that the shrinkage of the inference contribution coefficient is real and not canceled out by subsequent weights in the network.

[0039] Optionally, in S104, multi-stage sequential training is performed on the inference network model along a preset sparse intensity sequence arranged in ascending order, including: obtaining a preset increasing sparse intensity sequence [λ1, λ2, ..., λs], where λ1 < λ2 < ... < λs, and s is a positive integer; using the inference network model as the training starting point for the first scan stage, training is performed using a composite loss function until convergence; wherein, the composite loss function includes a sparse penalty term L1 that only applies to the inference contribution coefficient, and the penalty coefficient of the sparse penalty term L1 is λ1; from the second scan stage onwards, the model parameters after the convergence of the previous stage are used as the starting point for sequential training, and the penalty coefficient in the composite loss function of each stage is taken as the sparse intensity λs corresponding to the current stage.

[0040] This embodiment specifically describes the process of performing multi-stage sequential training on the inference network model along an increasing sparsity intensity sequence to obtain the exit critical intensity for each candidate field. In this process, the inference network model obtained after training in S103, containing the updated inference contribution coefficients, is used as the starting point for the first scanning stage. A composite loss function is used for training until convergence. This composite loss function includes an inference error term and a sparsity penalty term L1 that only applies to the inference contribution coefficients, where the penalty coefficient of the sparsity penalty term L1 is λ1. From the second scanning stage onwards, the model parameters after convergence in the previous stage are used as the starting point for sequential training. The penalty coefficient in the composite loss function of each stage is taken as the corresponding sparsity intensity λs. Through this multi-stage sequential training method with increasing sparsity intensity, the strength of the sparsity penalty gradually increases, and the inference contribution coefficients of each field gradually differentiate according to their actual contribution to the predicted target field: that is, fields with stronger contributions can resist larger sparsity penalties and remain non-zero, while fields with weaker contributions are shrunk to zero at lower sparsity intensities. The training is carried out by hot start between each stage, so that the adjacent stages do not need to relearn the complete inference relationship, which reduces the overall training cost and ensures the continuous evolution of the model state in each stage.

[0041] Specifically, the composite loss function Including the inference error term and the sparse penalty term L1 that only applies to the inference contribution coefficient; it can be expressed by the following formula: ; Where, N tr The number of candidate field data in the power data set. The target field label value for the training sample of the k-th candidate field. c is the predicted value corresponding to the k-th candidate field. j For the corresponding inferred contribution coefficient, the first term is the inference error term in the form of mean square error, and the second term is the sparse penalty term L1 that only applies to the inferred contribution coefficient.

[0042] In this embodiment, the first term in the formula is the inference error term in the form of mean squared error, which measures the model's prediction accuracy for the target field. Its function is to encourage the model to predict the target field from candidate fields as accurately as possible, thereby enabling fields that are truly useful for prediction to obtain larger inference contribution coefficients to resist shrinkage. The second term is a sparse penalty term that only acts on the inference contribution coefficients, which is used to penalize the absolute value of each inference contribution coefficient, pushing the coefficients of fields with weaker contributions towards zero. During training, the inference error term is used to maintain a large amplitude for the coefficients corresponding to fields with high relevance, while the sparse penalty term only pushes irrelevant coefficients towards zero, thereby continuously differentiating the coefficients.

[0043] Optionally, performing multi-stage sequential training on the inference network model along a preset sparse intensity sequence arranged in ascending order further includes: at the end of each scanning stage, obtaining the statistics of the inference contribution coefficients of each candidate field within the convergence window of the current scanning stage; determining whether each candidate field is in a live state based on the relationship between the statistics and the preset tolerance; fixing the inference contribution coefficients corresponding to candidate fields in a non-live state to zero and stopping updates in subsequent scanning stages, while recording the sparse intensity of the current stage as the critical intensity of the candidate fields in the current non-live state; and retaining the inference contribution coefficients of candidate fields in a live state in the next scanning stage.

[0044] This embodiment describes how to determine the survival status of candidate fields. At the end of each scanning phase, the statistics of the inferred contribution coefficients of each candidate field within the convergence window of the current scanning phase are obtained. Here, the convergence window refers to the time window formed by the last W rounds of training in the current scanning phase, where W is a positive integer and can be preset. This is used to avoid the influence of single-point reading fluctuations at the end of training on the determination results. Then, based on the relationship between the statistics and the preset tolerance, it is determined whether each candidate field is in a survival state. If the inferred contribution coefficient of a certain field has been stably maintained at a level close to zero within the convergence window, the field is determined to be in a non-survival state, its inferred contribution coefficient is fixed to zero, and updates are stopped in subsequent scanning phases. At the same time, the sparsity intensity of the current phase is recorded as the critical intensity of the field. If the inferred contribution coefficient of a certain field has not yet returned to zero or is still unstable, it is determined to be in a survival state, and its inferred contribution coefficient is retained to continue participating in training in the next scanning phase. Through the above determination and freeze exit mechanism, the survival set of each phase shrinks in a hierarchical nested manner. Fields that have exited will not re-enter subsequent phases, ensuring the stability and comparability of the exit critical intensity of each field.

[0045] Furthermore, the statistics include the mean and standard deviation of the absolute values ​​of the inferred contribution coefficients within the convergence window, and the preset tolerances include the zero-return tolerance and the stability tolerance. Based on the relationship between the statistics and the preset tolerances, it is determined whether each candidate field is in a live state, including: calculating the mean and standard deviation of the absolute values ​​of the inferred contribution coefficients of the current candidate field within the convergence window of the current scanning stage; if the mean is less than or equal to the zero-return tolerance and the standard deviation is less than or equal to the stability tolerance, the current candidate field is determined to be in a non-live state; otherwise, it is in a live state.

[0046] In this embodiment, the survival status determination uses a dual criterion of mean and standard deviation. Specifically, the statistics include the mean and standard deviation of the absolute values ​​of the inferred contribution coefficients within the convergence window, and the preset tolerances include a zero-reach tolerance and a stability tolerance. When the mean of the absolute values ​​of the magnitudes of the inferred contribution coefficients of a field is not higher than the zero-reach tolerance, and its standard deviation is not higher than the stability tolerance, the current candidate field is determined to be in a non-survival state. This indicates that the field has stably shrunk to zero, fixing its inferred contribution coefficient to zero and ceasing updates in subsequent scanning phases. The sparsity intensity of the scanning phase in which the field is determined to have stably shrunk to zero is recorded as the exit threshold intensity for the field. Otherwise, the field is determined to be in a survival state and retained for the next stage.

[0047] Optionally, before inputting the power dataset into the initial inference network model, the method further includes: using the equal-width bin histogram estimation method to calculate the mutual information between each candidate field and the target field, and normalizing the mutual information according to the information entropy to obtain normalized mutual information; and removing candidate fields whose normalized mutual information is lower than a preset initial screening threshold.

[0048] In this embodiment, before proceeding to S102, preliminary screening of candidate fields can be performed to improve computational efficiency and reduce data volume. Specifically, the equal-width binning histogram estimation method is used to calculate the mutual information between each candidate field and the target field, and the mutual information is normalized according to information entropy to obtain normalized mutual information. Normalized mutual information can measure the general statistical dependency between a single candidate field and the target field, and its value is between 0 and 1. Subsequently, candidate fields with normalized mutual information lower than a preset preliminary screening threshold are removed, and the remaining candidate fields constitute the preliminary screening candidate field set, which is used for subsequent inference contribution coefficient configuration and model training, thereby improving the overall computational efficiency of the method.

[0049] Furthermore, the overall solution will be described below through a more specific implementation method. For example... Figure 2 As shown, another method for identifying implicit inference sources in power data is illustrated, including S1, S2, S3, and S4. Specifically, S1: Obtain a power dataset containing historical samples from multiple time points. For each candidate field in the candidate field set, calculate its normalized mutual information with the target field. Remove fields whose normalized mutual information is lower than the initial screening threshold and construct the initial screening candidate field set F.

[0050] Specifically, S1 includes: S11, let the complete set of candidate fields be... The target field is y, where x i Let x represent the i-th candidate field, and p be the total number of candidate fields. For each candidate field x... i The mutual information I(x) between the equal-width binned histogram and the target field y is estimated and calculated.i ;y), the specific formula is as follows: ; Where X is the candidate field x i The range of values ​​is divided into equal widths. The set of intervals obtained after dividing the target field y into two intervals is Y, which is the set of intervals obtained after dividing the target field y into two intervals in the same way. B is the preset number of bins, p(a,b) is the joint frequency of a sample falling into both interval a and interval b, and p(a) and p(b) are the corresponding edge frequencies.

[0051] S12: Normalize the mutual information according to the information entropy to obtain normalized mutual information.

[0052] ; in, , for field x i Information entropy, H(y) is the information entropy obtained from the target field y in the same way, normalized mutual information. The value is between 0 and 1.

[0053] S13: Based on the pre-set initial screening threshold, retain candidate fields whose normalized mutual information is greater than the initial screening threshold to form the initial screening candidate field set. Where q is the number of candidate fields to be retained. In one feasible implementation, the initial screening threshold is a conservative, small value, used only to remove fields that clearly lack statistical dependence on the target field, and is not used as a basis for determining the source of inference.

[0054] S2: At the input of the deep neural network, each field in the initial candidate field set is assigned a learnable inference contribution coefficient that corresponds one-to-one with it. Element-wise weighting is performed on the input, and the weighted input vector is input into the inference network to output the predicted value of the target field, thus forming an inference model with inference contribution coefficients. The structure of this inference model with inference contribution coefficients is as follows: Figure 3 As shown.

[0055] Specifically, S2 includes: S21: At the input end of the deep neural network, for each field in the initial candidate field set... Configure a learnable inference contribution coefficient c that corresponds one-to-one with it. j Perform element-wise weighted processing on the input. The weighted input vector is obtained. All inferred contribution coefficients constitute the inferred contribution coefficient vector. All coefficients are initially set to 1.

[0056] S22: Constructing the inference network model f θSpecifically, it is a multi-layer fully connected neural network with an input layer of dimension q, followed by several fully connected hidden layers. Each hidden layer uses the ReLU activation function, and the output layer has a dimension of 1 and no activation function. The weighted input vector is fed into the inference network, and the output is the predicted value of the target field. θ represents the model parameters of the inference network model. After each model parameter update, unit norm normalization is performed on the column vectors corresponding to each input field in the weight matrix of the first hidden layer of the inference network. This suppresses the phenomenon that the shrinkage of the inference contribution coefficient is directly offset by the scaling of the input weight columns of the first hidden layer, thus making the scaling of the inference contribution coefficient to the field input channel more explicit. The following formula is an example of a normalization method: ; in, Let be the column vector corresponding to the j-th input in the weight matrix of the first hidden layer, and δ be a minimal positive number to prevent the normalization denominator from being zero. The inference contribution coefficient c... j For a learnable variable that takes continuous values, its magnitude representation field The degree to which the input channel is preserved during the target field recovery process.

[0057] S3: Perform multi-stage sequential training on the inference model with inferred contribution coefficients along a preset sparse intensity sequence from small to large. Each scan stage uses a composite loss function consisting of an inference error term and a sparse penalty term that only affects the inferred contribution coefficients. The penalty intensity is taken as the sparse intensity of the current stage. The first scan stage starts training from the initial model. Subsequent scan stages use the model parameters after convergence in the previous stage as the starting point for hot-start training, so that the inferred contribution coefficients of each field are differentiated according to their strength of contribution to the recovery of the target field while taking into account the accuracy of the target field recovery. At the end of each scan stage, the survival status of each field is determined based on the statistical value of the amplitude of the inferred contribution coefficients of each field within the convergence window of that stage and the preset tolerance. Zero tolerance and stability tolerance are set. Zero tolerance is used to determine whether the inferred contribution coefficients are close to the zero point, and stability tolerance is used to determine whether the coefficient fluctuations within the convergence window have reached a stable state.

[0058] Specifically, step S3 includes: S31: A pre-defined increasing sparse intensity sequence [λ1, λ2, ..., λs], where λ1 < λ2 < ... < λs, and λs is the sparse intensity of the s-th scan stage, and s is the total number of scan stages; the composite loss function of the s-th scan stage can be expressed as: ; Where, N tr The number of candidate field data in the power data set. The target field label value for the training sample of the k-th candidate field. c is the predicted value corresponding to the k-th candidate field. j For the corresponding inferred contribution coefficient, the first term is the inference error term in the form of mean square error, and the second term is the sparse penalty term L1 that only applies to the inferred contribution coefficient.

[0059] S32: The first scan phase starts with the initial model, using the Adam optimizer with a learning rate of [missing information]. Batch size train When convergence is reached; from the 2nd to the sth scan phase, training is continued by hot-starting with all model parameters from the previous phase after convergence. wheel.

[0060] Step S33: At the end of each scanning phase, use the last scan of that phase as the starting point. The rounds form a convergence window, and for each field, the mean of the absolute values ​​of the inference contribution coefficients within the convergence window is calculated. with standard deviation : ; in, For the first The set of training epochs included in the phase convergence window. To be the length of the convergence window, For the first Field at the end of the round of training Inferred contribution coefficient; setting zero tolerance With stability tolerance Both values ​​are much smaller than the effective inference contribution coefficient by an order of magnitude. and Then determine the field In the If the phase stabilizes and shrinks to zero, its inferred contribution coefficient is fixed at zero and will not participate in parameter updates in subsequent scan phases; otherwise, the field is considered alive. The survival set consists of all survival fields in each stage. The exit from the freeze ensures that the survival sets at each stage are nested and contracted step by step.

[0061] S34: Exit critical strength for each field record: ; in, For fields The scan phase number that was determined to have reached stable contraction and zeroing. That is, the minimum sparsity intensity required for the field to exit the inference process; for fields that survive after the scan is completed, denoted as .

[0062] S4: Classify fields according to exit critical strength. The higher the exit critical strength, the higher the inference contribution level. Fields that survive after scanning constitute the highest inference contribution level. Output the graded inference source sequence and the exit critical strength of each field to realize the graded identification of implicit inference sources of power data.

[0063] Specifically, S4 includes: S41: Fields with the same exit critical strength will be grouped into the same inference contribution level. The larger the value, the higher the inference contribution level. Fields that survive after the scan constitute the highest inference contribution level, resulting in a hierarchical inference source sequence, where the highest-level field group is... In the The field group for stage exit is The field group that exits in stage 1 is .

[0064] S42: Fields within the same inference contribution level are arranged from largest to smallest according to the mean magnitude of the coefficients within the convergence window of the last survival stage.

[0065] S43: Output the hierarchical inference source sequence of the target field, the exit critical strength of each field, and the survival set of each stage, as the hierarchical field-level identification result of the implicit inference relationship of the target field.

[0066] Furthermore, using a more specific example, we will use hourly operating and market data from the US PJM electricity market in 2025 as an example for illustration. The sample, aligned hourly, contains 8760 records, and the data table has 70 columns, including one timestamp and one target field. Net actual exchange power (net_actual_interchange_mw) is selected as the target field. The candidate fields comprise a total of 68 dimensions, covering categories such as power generation output, load, exchange capacity, market price, and reserve and ancillary services based on fuel type. The samples were randomly divided into training and test sets using a fixed seed at a ratio of 0.8 and 0.2, respectively. Both the target and target fields were standardized using the statistics from the training set.

[0067] The following steps are performed: the normalized mutual information between each of the 68 candidate fields and the target field is calculated and initially screened. The number of bins B estimated by the equal-width binning histogram is set to 32. The initial screening threshold is conservatively set to 0.06. Only 13 fields that clearly lack statistical dependence on the target field are removed, and 55 candidate fields are retained to form the initial screening candidate field set.

[0068] The process follows step S2 above: Construct an inference model with inference contribution coefficients. The input consists of 55 reserved fields, each configured with an inference contribution coefficient, and the initial values ​​are all set to 1. The input layer of the inference network has a dimension of 55, followed by three fully connected hidden layers with 64, 32, and 16 neurons respectively. Each hidden layer uses the ReLU activation function, and the output layer has a dimension of 1 and no activation function is applied. After each parameter update, unit norm normalization is performed on each input column of the weight matrix of the first hidden layer.

[0069] Process according to step S3 above: sparse intensity sequence is taken. to There are 7 stages in total, with values ​​of 0.0005, 0.001, 0.002, 0.004, 0.008, 0.016, and 0.032 respectively, increasing by a factor of 2 for each adjacent stage; the first stage starts training from the initial model. Rounds 2 through 7 of the warm-up training program Rounds; using the Adam optimizer, learning rate Take 0.00065, batch size Set to 50; convergence window length Set to 20, zero tolerance. Take 0.02, stability tolerance The value is set to 0.01. The number of surviving fields at the end of stages 1 to 7 are 55, 49, 46, 38, 25, 16, and 10 respectively. The coefficients of fields determined to have shrunk to zero are immediately set to zero and frozen. The coefficients of determination of the test set at the end of each stage after the exit channel is closed are... The values ​​were 0.9693, 0.9721, 0.9742, 0.9746, 0.9370, 0.9549, and 0.8756, respectively. The recovery accuracy fluctuated at the end of each stage as fields exited. In some subsequent stages, the recovery accuracy improved after a warm-start training. However, at the end of the scan, as the number of surviving fields decreased significantly, the overall recovery accuracy declined, indicating that the exit channel had a real impact on the model's recovery capability. Example of exit points: No fields exited in stage 1; six fields, including day-ahead and real-time integrated node electricity prices and day-ahead load forecasts, exited first in stage 2, with an exit threshold intensity of 0.001; ten fields survived until the scan was completed, with an exit threshold intensity greater than 0.032, representing the strongest contributing fields.

[0070] Processing is performed according to step S4 above: output is based on exit critical intensity levels, resulting in 7 non-empty inference contribution levels. The highest level consists of the 10 fields that remain alive after the scan is completed, ranked from largest to smallest by the mean amplitude of the final stage convergence window coefficient: hourly estimated load (0.147), natural gas power generation output (0.100), coal power generation output (0.069), photovoltaic power generation output (0.038), nuclear power generation output (0.034), hydropower generation output (0.028), wind power generation output (0.028), multi-fuel unit output (0.022), gross planned exchange power (0.022), and day-ahead primary and reserve demand (0.021); the second highest level is... The classification is divided into six fields: power output of oil-fired power plants with a critical exit intensity of 0.032, power output of other renewable energy sources, gross actual exchange power, gross unplanned exchange power, total grid loss, and day-ahead 30-minute reserve self-planning. Following these are nine fields for a critical exit intensity of 0.016, thirteen fields for 0.008, eight fields for 0.004, and three fields for 0.002. The lowest level has six fields for a critical exit intensity of 0.001, including day-ahead and real-time integrated node prices, day-ahead and latest load forecasts, day-ahead primary and reserve clearing prices, and wind power generation ratio. The highest level fields mainly include those closely related to power balance, such as system load, power generation output by fuel type, and planned exchange power, as well as day-ahead primary and reserve demand. Price-related fields and some day-ahead and latest load forecast fields are mainly located in lower levels. The overall classification results are consistent with the operating mechanism where net actual exchange power is affected by load, power generation output, and exchange plans.

[0071] The classification results were validated by independently modeling the survival set and its complement at each stage using a standard fully connected network trained for a fixed 100 epochs (with the same structure as the inference network, but without inference contribution coefficients and weight column normalization). The test set for modeling the survival set at each stage was also provided. The values ​​are as follows: field 55 0.9675, field 49 0.9651, field 46 0.9667, field 38 0.9664, field 25 0.9692, field 16 0.9639, and field 10 0.9530; corresponding to the complement modeling. The values ​​are as follows: 6 fields 0.4309, 9 fields 0.5290, 17 fields 0.7173, 30 fields 0.7773, 39 fields 0.7906, and 45 fields 0.8510. The 25 fields with an exit threshold strength greater than 0.008, less than half of which maintained recovery accuracy comparable to the full candidate fields, saw the first two levels (16 fields) still achieve 0.9639, while the highest level (10 fields) achieved 0.9530, significantly higher than the 0.8510 of its 45-field complement. The difference in recovery accuracy along the hierarchical sequence characterizes the information carrying capacity of each level of field group, and this validation was verified using independently retrained models, indicating that the hierarchical results reflect the actual inference ability of each field group.

[0072] Based on the specific steps in the above embodiments, it can be understood that this embodiment has at least the following technical effects: (1) Normalized mutual information is used as the initial screening index, which can measure the general statistical dependence that the linear correlation coefficient cannot reflect; the initial screening threshold is a conservative small value, which only removes obviously irrelevant fields to reduce the input scale of subsequent path scanning, and does not undertake the function of inferring the source.

[0073] (2) Configure a learnable inference contribution coefficient with a one-to-one corresponding continuous value for each original field at the input end, so that the recognition result can be interpreted to the specific data field; perform unit norm normalization on the column vectors corresponding to each input field in the weight matrix of the first hidden layer of the inference network to suppress the phenomenon that the inference contribution coefficient shrinks directly by increasing the scale of the input weight column of the first hidden layer, so that the inference contribution coefficient has a clearer representation of the scaling degree of the field input channel.

[0074] (3) The sparsity penalty is shifted from the internal connection weights of the inference network to the inference contribution coefficients. The inference error term is used to maintain the recovery ability of the target field. The sparsity penalty term promotes the shrinkage of the coefficients of low contribution fields, so that the inference contribution coefficients of each field are differentiated according to their actual role in reducing the inference error. The hot start and continuous training means that the adjacent stages do not need to relearn the complete inference relationship, which reduces the training cost of path scanning. The convergence window mean and standard deviation dual criteria reduce the impact of single-point reading fluctuations on the judgment results at the end of training. The coefficients of the fields that are judged to shrink to zero are fixed at zero and stop updating, ensuring that the survival set is nested level by level and preventing the exited fields from re-entering the subsequent scanning stage.

[0075] (4) Unlike traditional feature selection methods that directly use binary masks or convert continuous importance scores into a single subset of retained fields at preset thresholds, rankings, retention numbers, or elimination rounds, this embodiment forms multiple inferred contribution levels based on the exit order of each field in the incremental sparse intensity scan. The exit critical intensity constitutes a comparable relative contribution intensity measure under unified scanning conditions, expanding the identification result from a simple binary selection of retention and elimination to a hierarchical inferred source sequence with strong and weak levels; at the same time, there is no need to pre-specify the final number of retained fields or a fixed judgment position for dividing high and low contribution fields.

[0076] Furthermore, as Figure 1 and Figure 2 The specific implementation of the method shown in this embodiment provides a device for identifying implicit inference sources of power data, such as... Figure 4 As shown, the device includes: The acquisition unit 401 is configured to acquire a power dataset containing historical samples from multiple time periods, wherein the power dataset includes power data for candidate fields and power data for target fields. Configuration unit 402 is configured to configure initial inference contribution coefficients that have a mapping relationship with each candidate field before inputting the power dataset into the initial inference network model; The first training unit 403 is configured to train the initial inference network model using the target field as the prediction objective, through the power dataset, and after each model parameter update, perform normalization processing on the first hidden layer of the initial inference network model to suppress the shrinkage of the inference contribution coefficient effectiveness loss, thereby obtaining an inference network model containing the inference contribution coefficients updated after training; the inference contribution coefficients updated after training are used to characterize the contribution of the corresponding candidate field in predicting the target field. The second training unit 404 is configured to perform multi-stage sequential training on the inference network model along a preset sparse intensity sequence arranged in ascending order. In the multi-stage sequential training, the inference contribution coefficient is selectively shrunk with the current sparse intensity as a constraint in each stage. At the end of each stage, the survival status of each candidate field is determined according to the inference contribution coefficient of each candidate field in the current stage. The inference contribution coefficient of the candidate field determined to be in a non-survival state is set to zero and the corresponding sparse intensity is recorded as the critical intensity to obtain the critical intensity corresponding to each candidate field. Output unit 405 is configured to hierarchically divide the candidate fields according to the critical intensity and output the result as the identification result of the latent inference source of the target field. In a specific application scenario, the first training unit 403 is further configured to use the target field in the power dataset as the training label, use each candidate field and its corresponding inference contribution coefficient as the input vector of the initial inference network model, and output the predicted value of each candidate field relative to the target field in order to train the inference network parameters of the initial inference network model. Specifically, during the training of the inference network parameters of the initial inference network model, after each update of the inference network parameters, the column vectors corresponding to each input vector in the weight matrix of the first hidden layer of the inference network model are subjected to unit norm normalization.

[0077] In a specific application scenario, the second training unit 404 is further configured to acquire a preset incremental sparse intensity sequence [λ1, λ2, ..., λs], where λ1 < λ2 < ... < λs, and s is a positive integer; using the inference network model as the training starting point for the first scanning stage, training is performed using a composite loss function until convergence; wherein, the composite loss function includes a sparse penalty term L1 that only applies to the inference contribution coefficient, and the penalty coefficient of the sparse penalty term L1 is λ1; from the second scanning stage onwards, the model parameters after the convergence of the previous stage are used as the starting point for continuous training, and the penalty coefficient in the composite loss function of each stage is respectively taken as the sparse intensity λs corresponding to the current stage.

[0078] The composite loss function Including the inference error term and the sparse penalty term L1 that only applies to the inference contribution coefficient; it can be expressed by the following formula: ; Where, N tr The number of candidate field data in the power data set. The target field label value for the training sample of the k-th candidate field. c is the predicted value corresponding to the k-th candidate field. j For the corresponding inferred contribution coefficient, the first term is the inference error term in the form of mean square error, and the second term is the sparse penalty term L1 that only applies to the inferred contribution coefficient.

[0079] In specific application scenarios, the second training unit 404 is further configured to, at the end of each scanning phase, obtain the statistics of the inferred contribution coefficients of each candidate field within the convergence window of the current scanning phase; determine whether each candidate field is in a live state based on the relationship between the statistics and the preset tolerance; fix the inferred contribution coefficients of candidate fields in a non-live state to zero and stop updating them in subsequent scanning phases, while recording the sparsity intensity of the current phase as the critical intensity of the candidate fields in a non-live state; and retain the inferred contribution coefficients of candidate fields in a live state in the next scanning phase.

[0080] In specific application scenarios, the second training unit 404 is further configured to calculate the mean and standard deviation of the absolute values ​​of the inference contribution coefficients of the current candidate field within the convergence window of the current scanning phase; if the mean is less than or equal to the zero tolerance and the standard deviation is less than or equal to the stability tolerance, the current candidate field is determined to be in a non-survival state; otherwise, it is in a survival state.

[0081] In specific application scenarios, the configuration unit 402 is further configured to use the equal-width binning histogram estimation method to calculate the mutual information between each candidate field and the target field, and to normalize the mutual information according to the information entropy to obtain normalized mutual information; and to remove candidate fields whose normalized mutual information is lower than a preset initial screening threshold.

[0082] It should be noted that other corresponding descriptions of the functional units involved in the implicit inference source identification device for power data provided in this embodiment can be found in [reference]. Figure 1 and Figure 2 The corresponding description in [the document] will not be repeated here.

[0083] Based on the above, Figure 1 and Figure 2 Accordingly, this embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. Figure 1 and Figure 2 The method shown.

[0084] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.

[0085] Based on the above, Figure 1 and Figure 2 The method shown, and Figure 4 To achieve the above objectives, this application also provides an electronic device, which can be configured on a computer side, etc. The device includes a storage medium and a processor; the storage medium is used to store a computer program; the processor is used to execute the computer program to achieve the above-described objectives. Figure 1 and Figure 2 The method shown.

[0086] Optionally, the aforementioned physical devices may also include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, input units such as a keyboard, etc., and optional user interfaces may also include USB interfaces, card reader interfaces, etc. The network interface may optionally include standard wired interfaces, wireless interfaces (such as Wi-Fi interfaces), etc.

[0087] Those skilled in the art will understand that the physical device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or have different component arrangements.

[0088] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.

[0089] Through the above description of the implementation methods, those skilled in the art can clearly understand that this application can be implemented using software plus necessary general-purpose hardware platforms, or it can be implemented in hardware. By applying the scheme of this embodiment, a learnable initial inference contribution coefficient is configured for each candidate field. Then, during the training process based on the power dataset, the model parameters and inference contribution coefficients of the inference network model are trained under the same framework, so that the importance of each candidate field is quantified into explicit parameters that can be directly controlled by the optimization process, realizing a recognition mechanism guided by multi-field joint inference capability. Furthermore, multi-stage sequential training is performed along an increasing sparse intensity sequence, using the actual exit order of each candidate field as the grading basis, thereby obtaining the exit critical intensity of each field for grading. There is no need to preset the pruning threshold or the number of retentions, solving the problem that traditional feature selection methods rely on manually setting the judgment position and can only output a single binary subset. Finally, the recognition results with strong and weak levels are output, providing a more refined decision basis for the analysis of power data.

[0090] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0091] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0092] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for identifying implicit inference sources in power data, characterized in that, include: Obtain a power dataset containing historical samples from multiple time points, wherein the power dataset includes power data for candidate fields and power data for target fields; Before inputting the power dataset into the initial inference network model, configure the initial inference contribution coefficients that have a mapping relationship with each candidate field; Using the target field as the prediction objective, the initial inference network model is trained using the power dataset. After each model parameter update, a normalization process is performed on the first hidden layer of the initial inference network model to suppress the shrinkage of the inference contribution coefficient effectiveness loss, resulting in an inference network model containing the inference contribution coefficients updated after training. The inference contribution coefficients updated after training are used to characterize the contribution of the corresponding candidate field in predicting the target field. The inference network model is subjected to multi-stage sequential training along a preset sparse intensity sequence arranged in ascending order to obtain the critical intensity corresponding to each candidate field. In the multi-stage sequential training, the inference contribution coefficient is selectively shrunk with the current sparse intensity as a constraint in each stage. At the end of each stage, the survival status of each candidate field is determined according to the inference contribution coefficient of each candidate field in the current stage. The inference contribution coefficient of the candidate field determined to be in a non-survival state is set to zero and the corresponding sparse intensity is recorded as the critical intensity. The candidate fields are hierarchically divided according to the critical intensity and output as the identification result of the latent inference source of the target field.

2. The method according to claim 1, characterized in that, The model parameters include inference network parameters, and the initial inference network includes an input layer and multiple fully connected hidden layers following the input layer. Using the target field as the prediction objective, the initial inference network model is trained using the power dataset. After each model parameter update, a normalization process is performed on the first hidden layer of the initial inference network model to suppress the shrinkage of the inference contribution coefficient effectiveness loss, resulting in an inference network model containing the trained and updated inference contribution coefficients, including: The target field in the power dataset is used as the training label, and each candidate field and its corresponding inference contribution coefficient are used as the input vector of the initial inference network model. The predicted value of each candidate field relative to the target field is output to train the inference network parameters of the initial inference network model. Specifically, during the training of the inference network parameters of the initial inference network model, after each update of the inference network parameters, the column vectors corresponding to each input vector in the weight matrix of the first hidden layer of the inference network model are subjected to unit norm normalization.

3. The method according to claim 1, characterized in that, The inference network model is subjected to multi-stage sequential training along a preset sparse intensity sequence arranged in ascending order, including: Obtain a preset increasing sparse intensity sequence [λ1, λ2, ..., λs], where λ1 < λ2 < ... < λs, and s is a positive integer; The inference network model is used as the starting point for training in the first scanning phase, and a composite loss function is used for training until convergence; wherein, the composite loss function includes a sparse penalty term L1 that only applies to the inference contribution coefficient, and the penalty coefficient of the sparse penalty term L1 is λ1. Starting from the second scanning phase, the model parameters after the previous phase training convergence are used as the starting point for continuous training, and the penalty coefficient in the composite loss function of each phase is respectively taken as the sparsity intensity λs corresponding to the current phase.

4. The method according to claim 3, characterized in that, The composite loss function Including the inference error term and the sparse penalty term L1 that only applies to the inference contribution coefficient; it can be expressed by the following formula: ； Where, N tr The number of candidate field data in the power data set. The target field label value for the training sample of the k-th candidate field. c is the predicted value corresponding to the k-th candidate field. j For the corresponding inferred contribution coefficient, the first term is the inference error term in the form of mean square error, and the second term is the sparse penalty term L1 that only applies to the inferred contribution coefficient.

5. The method according to claim 3, characterized in that, Performing multi-stage sequential training on the inference network model along a preset sparse intensity sequence arranged in ascending order further includes: At the end of each scanning phase, obtain the statistics of the inferred contribution coefficients of each candidate field within the convergence window of the current scanning phase; Based on the relationship between the statistics and the preset tolerance, it is determined whether each candidate field is in a live state; The inference contribution coefficients corresponding to candidate fields in non-surviving states are fixed at zero and stopped being updated in subsequent scanning stages. At the same time, the sparsity intensity of the current stage is recorded as the critical intensity of the candidate fields in the current non-surviving state. The inference contribution coefficients of candidate fields for survival status are retained in the next scanning stage.

6. The method according to claim 5, characterized in that, The statistics include the mean and standard deviation of the absolute values ​​of the inferred contribution coefficients within the convergence window, and the preset tolerances include zero tolerance and stability tolerance. The step of determining whether each candidate field is in a live state based on the relationship between the statistics and the preset tolerance includes: Calculate the mean and standard deviation of the absolute values ​​of the inferred contribution coefficients of the current candidate fields within the convergence window of the current scanning phase; If the mean is less than or equal to the zero tolerance and the standard deviation is less than or equal to the stability tolerance, the current candidate field is determined to be in a non-live state; otherwise, it is in a live state.

7. The method according to claim 1, characterized in that, Before inputting the power dataset into the initial inference network model, the method further includes: The equal-width binning histogram estimation method is used to calculate the mutual information between each candidate field and the target field, and the mutual information is normalized according to the information entropy to obtain normalized mutual information. Candidate fields whose normalized mutual information is lower than a preset initial screening threshold are removed.

8. A device for identifying implicit inference sources in power data, characterized in that, include: The acquisition unit is configured to acquire a power dataset containing historical samples from multiple time points, wherein the power dataset includes power data for candidate fields and power data for target fields; The configuration unit is configured to configure initial inference contribution coefficients that have a mapping relationship with each candidate field before inputting the power dataset into the initial inference network model; The first training unit is configured to train the initial inference network model using the target field as the prediction objective, through the power dataset, and after each model parameter update, perform normalization processing on the first hidden layer of the initial inference network model to suppress the shrinkage of the inference contribution coefficient effectiveness loss, thereby obtaining an inference network model containing the inference contribution coefficients updated after training; the inference contribution coefficients updated after training are used to characterize the contribution of the corresponding candidate field in predicting the target field. The second training unit is configured to perform multi-stage sequential training on the inference network model along a preset sparse intensity sequence arranged in ascending order. In the multi-stage sequential training, the inference contribution coefficient is selectively shrunk at each stage with the current sparse intensity as a constraint. At the end of each stage, the survival status of each candidate field is determined based on the inference contribution coefficient of each candidate field in the current stage. The inference contribution coefficient of the candidate fields determined to be in a non-survival state is set to zero and the corresponding sparse intensity is recorded as the critical intensity to obtain the critical intensity corresponding to each candidate field. The output unit is configured to hierarchically divide the candidate fields according to the critical intensity and output the results as the identification results of the latent inference source of the target field.

9. An electronic device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of 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 method of any one of claims 1 to 7.

Images