Power consumer behavior analysis method and system based on multi-source heterogeneous data

Abstract

The invention discloses a power consumer behavior analysis method and system based on multi-source heterogeneous data, and relates to the technical field of power, and the method comprises the steps: determining the multi-source heterogeneous data of a target power grid user side, and the multi-source heterogeneous data comprises a plurality of user side data sources; performing feature extraction on the data sources, and outputting respective feature distribution; analyzing the feature distribution based on similarity probability distribution, and determining a plurality of fusion similarity indexes; selecting a first data source feature distribution with the highest similarity, and performing equivalence processing on other data source feature distributions based on the first data source feature distribution to obtain a plurality of equivalent feature distributions; and training a power consumer behavior analysis model by using the equivalent feature distribution as a training data set, and further performing power demand prediction. The technical problems of low information utilization rate and analysis efficiency and poor analysis result accuracy are solved, and the technical effects of improving the information utilization rate and analysis efficiency and improving the analysis accuracy are achieved.

Description

Technical Field

[0001] The present invention relates to the field of electric power technology, and in particular to an electric power user behavior analysis method and system based on multi-source heterogeneous data. Background Art

[0002] With the development of smart grid technology, the analysis of power user behavior is of great significance for the optimization of power grid operation and energy management. However, power user data comes from various sources, including smart meters, household electrical appliances, and external data such as meteorological information. These data show multi-source heterogeneity. Existing technologies are difficult to effectively integrate information from different data sources, and there are technical problems such as low information utilization and analysis efficiency, and poor accuracy of analysis results. Summary of the invention

[0003] The present invention provides a method and system for analyzing power user behavior based on multi-source heterogeneous data to solve the technical problems of low information utilization and analysis efficiency and poor accuracy of analysis results in the prior art, and to achieve the technical effect of improving information utilization and analysis efficiency and improving analysis accuracy.

[0004] In a first aspect, the present invention provides a method for analyzing electric power user behavior based on multi-source heterogeneous data, wherein the method comprises: Determine multi-source heterogeneous data on the user side of the target power grid, wherein the multi-source heterogeneous data includes multiple user-side data sources; perform feature extraction on the multiple user-side data sources, and output multiple data source feature distributions, wherein each user-side data source corresponds to a data source feature distribution; analyze the multiple data source feature distributions based on similarity probability distribution, and determine multiple fusion similarity indicators; determine a first data source feature distribution with the multiple fusion similarity indicators, wherein the first data source feature is the data source feature distribution with the highest similarity indicator; perform equivalent processing on the multiple data source feature distributions according to the first data source feature distribution, and obtain multiple equivalent feature distributions; call the multiple equivalent feature distributions as training data sets, train a power user behavior analysis model, and perform power prediction based on the power user behavior analysis model.

[0005] In a second aspect, the present invention further provides a power user behavior analysis system based on multi-source heterogeneous data, wherein the system comprises: A data extraction module is used to determine multi-source heterogeneous data on the user side of a target power grid, wherein the multi-source heterogeneous data includes multiple user-side data sources.

[0006] A data source feature module is used to extract features from the multiple user-side data sources and output multiple data source feature distributions, wherein each user-side data source corresponds to a data source feature distribution.

[0007] A distribution analysis module is used to analyze the feature distributions of the multiple data sources based on similarity probability distribution to determine multiple fusion similarity indicators.

[0008] A feature distribution selection module is used to determine a first data source feature distribution based on the multiple fusion similarity indicators, wherein the first data source feature is a data source feature distribution with the highest similarity indicator.

[0009] An equivalent generation module is used to perform equivalent processing on the multiple data source feature distributions according to the first data source feature distribution to obtain multiple equivalent feature distributions.

[0010] An analysis and prediction module is used to call the multiple equivalent feature distributions as training data sets, train a power user behavior analysis model, and perform power prediction based on the power user behavior analysis model.

[0011] The electric power user behavior analysis method and system based on multi-source heterogeneous data disclosed in the present invention solve the technical problems of low information utilization and analysis efficiency and poor accuracy of analysis results, and achieve the technical effect of improving information utilization and analysis effectiveness and improving analysis accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a flow chart of the method for analyzing the behavior of power users based on multi-source heterogeneous data of the present invention; Figure 2 It is a structural schematic diagram of the power user behavior analysis system based on multi-source heterogeneous data of the present invention.

[0013] Explanation of the accompanying drawings: data extraction module 11, data source feature module 12, distribution analysis module 13, feature distribution selection module 14, equivalent generation module 15, analysis and prediction module 16. DETAILED DESCRIPTION

[0014] The technical solution provided in the embodiments of the present invention is to solve the technical problems of low information utilization and analysis efficiency and poor accuracy of analysis results in the prior art. The overall idea adopted is as follows: First, determine the multi-source heterogeneous data on the user side of the target power grid. These data come from multiple user-side data sources, which may include smart meters, power equipment monitors, sensor data, etc. Then, extract features from the multiple user-side data sources, output the feature distribution of each data source, and form multiple data source feature distributions. The feature distribution of each data source describes the unique data characteristics of the data source. Then, analyze the feature distributions of multiple data sources based on similarity probability distribution, and determine multiple fusion similarity indicators through similarity calculation. The similarity indicator is used to measure the similarity of feature distributions between different data sources. Based on these fusion similarity indicators, determine the first data source feature distribution, that is, the data source feature distribution with the highest similarity index. This distribution serves as a reference for the feature distributions of other data sources. Equilibrium processing is performed on the feature distributions of multiple data sources according to the first data source feature distribution to generate multiple equivalent feature distributions, ensuring that the features of all data sources are comparable and unified in the same feature space. Finally, call the multiple equivalent feature distributions as training data sets to train the power user behavior analysis model. The model predicts the user's power usage behavior based on the data of the equivalent feature distribution, and then predicts and optimizes power demand.

[0015] The above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods of the specification to better understand the above technical solution. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments of the present invention. It should be understood that the present invention is not limited to the example embodiments used only to explain the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In addition, it should be noted that, for the convenience of description, only the parts related to the present invention are shown in the drawings, rather than all of them. Embodiment 1

[0016] Figure 1 The figure is a flow chart of a method for analyzing the behavior of power users based on multi-source heterogeneous data according to the present invention, wherein the method comprises: Determine multi-source heterogeneous data on the user side of the target power grid, wherein the multi-source heterogeneous data includes multiple user-side data sources.

[0017] Specifically, the data on the target power grid user side includes data from different devices and different data attributes, and the data format of each data attribute is also different. Among them, multiple user-side data sources correspond to multiple different data acquisition devices or data transmission devices.

[0018] The optional data attributes of multi-source heterogeneous data may include: power consumption data, i.e., power consumption information of each user collected from smart meters, including time series power consumption data, peak and valley power price data, etc.; load monitoring data, collecting load data generated by user-side load monitoring equipment, and recording load demand fluctuations in different time periods; distributed energy data, if distributed energy (such as solar panels, wind power generation, etc.) is installed on the user side, then its power generation, usage and grid-connected data are obtained; equipment status data, i.e., operating data of equipment in the user's home or factory, including the switch status of electrical equipment, the health status of equipment, etc.; environmental monitoring data, collecting environmental information data on the user side, such as indoor and outdoor temperature, humidity, light intensity, etc., environmental factors that affect power demand; power grid interaction data, involving interactive information between users and the power grid, such as load response, demand-side management data, and information such as power consumption plans and power purchase agreements in which users participate; power factor data, which is a parameter for measuring the efficiency of power utilization, represents the ratio between actual power (active power) and apparent power, and is used to evaluate the efficiency of users.

[0019] Optionally, the data format depends on the device and path for data collection and acquisition. User-side data acquired through different collection devices and different transmission encoding protocols have different data formats.

[0020] Specifically, the multi-source heterogeneous data on the user side can be structured data (such as power consumption, equipment status) or unstructured data (such as interaction records between the power grid and users, log data). For example, power consumption and load data usually exist in the form of time series, while equipment status or fault information is often event-driven data.

[0021] In addition, the user attributes on the target power grid user side are also diversified. For example, they can be divided into residential users, industrial users and commercial users according to user attributes; they can be divided into high-load users, renewable energy users, etc. according to their electricity consumption characteristics. Different types of users will have different electricity consumption behaviors and data requirements. Through the above steps, multi-source heterogeneous data on the target power grid user side can be determined, which can achieve more accurate analysis of user electricity consumption behavior.

[0022] Feature extraction is performed on the multiple user-side data sources, and multiple data source feature distributions are output, wherein each user-side data source corresponds to a data source feature distribution.

[0023] Specifically, feature extraction is performed on multiple user-side data sources and corresponding data source feature distributions are generated to extract representative information from each data source. The data source feature distribution is used to characterize the characteristics and data generation patterns of the multiple data sources. Exemplarily, feature extraction is performed on multiple data source feature distributions, and feature indicator items include: data volume, data granularity, data format, data attributes, data category, data accuracy, etc.

[0024] Specifically, the characteristic distribution of the data source includes the characteristic performance of the data provided by the data source in multiple dimensions. For example, the characteristics of the power consumption data source are extracted, including the characteristic distribution in multiple dimensions such as total power consumption, peak load, load curve shape, and power consumption pattern; the characteristics of the load monitoring data source are extracted, including the characteristic distribution in multiple dimensions such as load fluctuation rate, equipment power distribution, and load peak and valley characteristics; feature extraction is performed on the distributed energy data source to obtain the characteristic distribution in dimensions including power generation, grid-connected power, and power generation volatility; feature extraction is performed on the equipment status data source, and the obtained data source feature distribution includes the characteristic distribution in dimensions such as equipment usage frequency, failure rate, and average operating time.

[0025] Furthermore, the obtained multiple features are distributed and represented in a high-dimensional space to generate a multidimensional feature space, which is output as a data source feature distribution of the corresponding user-side data source. By extracting features from multiple user-side data sources and generating feature distributions, the characteristics and data generation patterns of the data sources can be structured, thereby facilitating the understanding of multiple user-side data sources.

[0026] The feature distributions of the multiple data sources are analyzed based on the similarity probability distribution to determine multiple fusion similarity indicators.

[0027] In some embodiments, the multiple data source feature distributions are analyzed based on the similarity probability distribution to determine multiple fused similarity indicators, including: identifying the multiple data source feature distributions separately based on the similarity probability distribution, analyzing N-1 similarity indicators between any data source feature distribution and N-1 data source feature distributions one by one, where N is the number of data sources; integrating the N-1 similarity indicators as fused similarity indicators of the feature distributions of each data source, and outputting the multiple fused similarity indicators.

[0028] Specifically, for each data source feature distribution, a similarity algorithm (such as cosine similarity, Euclidean distance, KL divergence, etc.) is used to calculate the similarity between it and other data source feature distributions, and the similarity results are standardized to generate a similarity probability distribution to measure the similarity between the two. The similarity index refers to the similarity between feature index items of the same category.

[0029] For example, there are data source A and data source B, and the similarity between them is S(A, B). The similarity can be calculated by some metric (such as cosine similarity or Euclidean distance). The similarity value S(A, B) is between [0, 1], 1 means complete similarity, and 0 means complete dissimilarity. The similarity probability distribution is calculated by the similarity probability distribution formula: ; in: Represents the probability of similarity between A and B; is the similarity value between objects A and B; It is the sum of the similarities between object A and i possible comparison objects, used for normalization.

[0030] Specifically, for each data source, N-1 similarity probabilities between it and other N-1 data sources (i.e., the N-1 similarity indicators of the data source) are calculated as the fused similarity indicator of the data source feature distribution of the data source, and then multiple fused similarity indicators of multiple data sources are output.

[0031] Through the above method steps, combined with the similarity probability distribution formula, the similarity between data generation patterns among multiple data sources is quantitatively evaluated, and different similarities have good comparability, which is convenient for the subsequent selection of data feature distribution.

[0032] A first data source feature distribution is determined based on the multiple fusion similarity indices, wherein the first data source feature is a data source feature distribution with the highest similarity index.

[0033] In some embodiments, determining the first data source feature distribution based on the multiple fusion similarity indicators includes: The sizes of the multiple fusion similarity indicators are analyzed to determine a first data source feature distribution and a second data source feature distribution, wherein the second data source feature distribution is a data source feature distribution with a second highest similarity indicator; a similarity indicator between the first data source feature distribution and the second data source feature distribution is calculated; if the similarity indicator is greater than or equal to a preset similarity threshold, the multiple data source feature distributions are equivalently processed with the first data source feature distribution; if the similarity indicator is less than the preset similarity threshold, the multiple data source feature distributions are equivalently processed with the first data source feature distribution and the second data source feature distribution.

[0034] Specifically, analyze the fusion similarity index and determine the first and second data source feature distributions: First, analyze multiple fusion similarity indexes, select the data source with the highest similarity index as the first data source feature distribution, and the data source with the second highest similarity index as the second data source feature distribution. The above two data sources will serve as reference points for subsequent equivalence processing and are the first two data sources with the greatest commonality with other data sources.

[0035] Specifically, based on the selected similarity algorithm (such as Euclidean distance, cosine similarity, etc.), the similarity index of the first and second data sources is calculated to determine the degree of difference between the two.

[0036] Specifically, the preset similarity threshold is a standard for distinguishing whether the feature distributions of two data sources are similar enough. If the similarity index is greater than or equal to the preset threshold, it means that the feature distributions of the two data sources are highly similar, and a simpler equivalent processing method can be used. If the similarity index is less than the preset threshold, more complex processing is required to fully consider the feature differences between the two.

[0037] Among them, if the similarity between the feature distributions of the first and second data sources is high (that is, the similarity index is greater than or equal to the preset similarity threshold), the feature distribution of the first data source is directly used to perform equivalent processing on the feature distributions of all data sources, that is, the feature distributions of other data sources are adjusted according to the characteristics of the first data source, so that their characteristics are closer to the first data source.

[0038] If the similarity between the feature distributions of the first and second data sources is low (i.e., the similarity index is less than the preset similarity threshold), the feature distributions of the first and second data sources are used together to perform equivalent processing on the feature distributions of other data sources. That is, by comprehensively considering the feature differences and commonalities between the two, the features of the other data sources after equivalent processing better match the feature distributions of the first and second data sources.

[0039] Exemplarily, the distribution of data source features is equivalently processed, including defining operation paths (methods) for specific equivalent operations such as data dimension conversion, data format conversion, data granularity alignment, and data mode conversion on the data features of multiple data sources. Among them, data dimension conversion includes dimensionality reduction and dimensionality increase. Data format conversion is used to convert non-numeric data into numerical data or to normalize or standardize numerical data. Data granularity alignment refers to converting data to the same granularity or scale, involving interpolation fitting of data. Data mode conversion is to convert data from one mode to another, such as converting time series data into static data, converting unstructured data into structured data, and converting event-driven data into serialized data.

[0040] In some implementations, performing equivalent processing on the plurality of data source feature distributions using the first data source feature distribution includes: An equivalent processing model is constructed, wherein the equivalent processing model includes a dimensionality reduction equivalent processing channel and a high-dimensional equivalent processing channel; the multiple data source feature distributions are divided according to the first data source feature distribution, and a first category of data source feature distribution and a second category of data source feature distribution are output; wherein the dimensionality reduction equivalent processing channel is used to perform feature dimensionality reduction equivalent processing on the first category of data source feature distribution, and the high-dimensional equivalent processing channel is used to perform feature high-dimensional equivalent processing on the first category of data source feature distribution.

[0041] Among them, data dimension conversion involves dimensionality reduction or dimensionality increase. Among them, dimensionality reduction is achieved through the dimensionality reduction equivalent processing channel in the equivalent processing model. The dimensionality reduction equivalent processing channel can be constructed using data dimensionality reduction methods such as principal component analysis (PCA), linear discriminant analysis (LDA), multidimensional scaling (MDS), local linear embedding (LLE), independent component analysis (ICA), etc., to eliminate redundant features, reduce the complexity of data, and retain the most important information.

[0042] Among them, dimensionality increase is achieved through a high-dimensional equivalent processing channel in the equivalent processing model. The high-dimensional equivalent processing channel can use methods such as convolution, pooling, and self-multiplication to increase the dimension of the data and extract more features.

[0043] In particular, for signal data that has both time domain and frequency domain, a high-dimensional equivalent processing channel can be constructed through signal analysis and processing methods such as wavelet transform, Fourier transform, short-time Fourier transform, etc., to extract the characteristics of the signal for use as high-dimensional data. In this case, the above-mentioned signal analysis and processing methods can be regarded as data dimensionality increase methods.

[0044] Specifically, the data sources are divided based on the first data source feature distribution, where the first type of data source feature distribution refers to a feature distribution with richer information dimensions than the first data source feature distribution. In other words, the first type of data source feature distribution contains all the feature dimensions and corresponding dimensional data required to form the first data source feature distribution, and can be equivalent through dimensionality reduction. The second type of data source feature distribution refers to a data source with more streamlined information dimensions than the first data source feature distribution, which cannot be equivalently processed by simple dimensionality reduction, and requires a more complex equivalent processing strategy to expand the dimension.

[0045] Furthermore, the feature distribution results of the data sources that have undergone dimensionality reduction and high-dimensional equivalence processing are unified and integrated to form a complete set of equivalent processed data sources. This dataset not only retains important low-dimensional features, but also can handle the complexity of high-dimensional features, providing optimized feature data for subsequent data fusion and analysis.

[0046] In some implementations, dividing the plurality of data source feature distributions according to the first data source feature distribution further includes: Obtain a first feature strength of the first data source feature distribution; obtain multiple feature strengths of the multiple data source feature distributions; output the data source feature distributions that are greater than or equal to the first feature strength among the multiple feature strengths as first-category data source feature distributions, and output the data source feature distributions that are less than the first feature strength among the multiple feature strengths as second-category data source feature distributions.

[0047] Specifically, first, the feature strength of the feature distribution of the first data source is analyzed to determine its comprehensive performance in terms of data dimension, granularity, data volume, information density, etc., and the output is the first feature strength. The value of the first feature strength is a reference benchmark for dividing the feature distribution of other data sources. Then, based on the same analysis and evaluation method, the feature strength analysis of the feature distributions of multiple data sources is performed separately to obtain the feature strength value of each data source.

[0048] Specifically, the first type of data source feature distribution is a feature distribution whose feature intensity is greater than or equal to the feature intensity of the first data source feature distribution. The dimensional information of these data source feature distributions is relatively complete and contains sufficiently rich feature dimensions. Therefore, redundant features can be compressed through dimensionality reduction equivalent processing to achieve the effect of feature balancing and simplification. The second type of data source feature distribution is a feature distribution whose feature intensity is less than the feature intensity of the first data source feature distribution. The dimensional information of these data source feature distributions is relatively concise, and some information is missing or less, so it cannot be equivalent through simple dimensionality reduction processing. More complex equivalent processing strategies such as dimensional expansion and feature enhancement are required to complete the missing information and effectively align it with other data source feature distributions.

[0049] By analyzing the feature strength through the above-mentioned method steps, the richness of the feature distribution of the data source can be effectively distinguished, and then the feature distribution of multiple data sources can be divided to further optimize the data processing and equivalence strategy.

[0050] The multiple data source feature distributions are equivalently processed according to the first data source feature distribution to obtain multiple equivalent feature distributions.

[0051] The multiple equivalent feature distributions are called as training data sets to train a power user behavior analysis model, and power prediction is performed according to the power user behavior analysis model.

[0052] In some embodiments, training a power user behavior analysis model includes: The first data source feature distribution is used to perform equivalent processing on the multiple data source feature distributions, and multiple first equivalent feature distributions are output; the second data source feature distribution is used to perform equivalent processing on the multiple data source feature distributions, and multiple second equivalent feature distributions are output; the multiple first equivalent feature distributions and the multiple second equivalent feature distributions are used as training data sets to train the power user behavior analysis model.

[0053] Specifically, to train the power user behavior analysis model, first, based on the feature distribution of the first data source, perform equivalent processing on the feature distribution of multiple data sources, and output multiple first equivalent feature distributions. This equivalent processing is used for data sources with rich features, and simplifies high-dimensional features through dimensionality reduction processing while maintaining the integrity of key information. Then, based on the feature distribution of the second data source, perform equivalent processing on the feature distribution of multiple data sources, and output multiple second equivalent feature distributions. This equivalent processing is used for data sources with relatively concise information, and fills in information gaps through technologies such as extended dimensions and feature enhancement, so that its feature distribution is consistent with other data sources.

[0054] Specifically, the generated first equivalent feature distribution and the second equivalent feature distribution are used together as training data sets. These data sets contain features of multiple data sources after equivalent processing, have higher versatility and consistency, and are convenient for model learning.

[0055] Furthermore, these equivalent feature distribution data sets are used to train the power user behavior analysis model. The power user behavior analysis model is a machine learning model constructed by supervised learning methods, exemplarily including decision trees, random forests, neural networks, etc. The power user behavior analysis model uses historical power user behavior data to identify the power consumption patterns of different users in a specific time period, and performs pattern recognition and behavior modeling. Through the power user behavior analysis model, it is possible to identify the power consumption habits, changing trends, abnormal behaviors, etc. of specific users, provide more accurate power demand forecasts for power companies, and optimize power dispatch and load management.

[0056] In some embodiments, calling the multiple equivalent feature distributions as training data sets to train a power user behavior analysis model includes: According to the influence relationship between the multiple user-side data sources and the accuracy of power prediction, multiple penalty factors are established; based on the multiple penalty factors, model parameters of the power user behavior analysis model are defined to obtain a defined power user behavior analysis initial model; the multiple equivalent feature distributions are called as training data sets to train the power user behavior analysis initial model to obtain the power user behavior analysis model.

[0057] Specifically, according to the influence relationship between multiple data sources on the user side and the accuracy of power prediction, a corresponding penalty factor is established for each data source feature distribution. The above penalty factors are used to measure the contribution and deviation of different data source feature distributions to the power prediction model. A smaller penalty factor is assigned to features that have a greater impact on prediction accuracy, while a larger penalty factor is assigned to features that have a negative impact on model prediction, ensuring that user-side data sources with a greater degree of impact are given priority.

[0058] Exemplarily, the types of penalty factors include: data quality-related penalty factors (feature distributions with poor data quality, such as features with a lot of noise and missing values, will be given higher penalties), data redundancy-related penalty factors (for feature distributions with redundant information, the penalty factor is used to suppress their excessive contribution to model training), and data consistency-related penalty factors (if the consistency between feature distributions of data sources is poor, the penalty factor will increase, thereby limiting its negative impact on the model).

[0059] Furthermore, multiple penalty factors are used to adjust the key parameters of the model, such as the regularization coefficient and learning rate, to avoid overfitting of the model to certain data source characteristics and enhance its generalization ability on different data sources, thus obtaining the initial version of the power user behavior analysis model.

[0060] Specifically, parameter settings include: assigning weights to different features so that high-quality features contribute more to model training, while low-quality features that are subject to penalty factors contribute less. Adjusting the regularization strength through penalty factors to prevent the model from being too complex or overly dependent on certain features. Introducing penalty factors into the loss function so that it not only considers the error, but also comprehensively considers the contribution and impact of data source features.

[0061] Specifically, the multiple first equivalent feature distributions and second equivalent feature distributions generated previously are called as training data sets to train the initial model for power user behavior analysis. Through training, the model can learn user behavior characteristics, and through multiple rounds of iterations and cross-validation, the hyperparameters of the model are further optimized to ensure stable performance on different data source characteristics.

[0062] Optionally, the model is fine-tuned for specific electricity user groups, regions or time periods to improve the personalization capabilities of electricity forecasting, including enhanced training of the electricity user behavior analysis model based on a specific training data set, or adjustment of penalty factors, weights, learning rates and other parameters corresponding to specific electricity user groups, regions or time periods.

[0063] Through the above method, we can utilize the different feature distributions in multi-source heterogeneous data and combine the penalty factor to reasonably constrain the model, so that the power user behavior analysis model has stronger prediction and data processing capabilities, and improve the accuracy and efficiency of power forecasting.

[0064] In summary, the power user behavior analysis method based on multi-source heterogeneous data provided by the present invention has the following technical effects: By determining the multi-source heterogeneous data on the user side of the target power grid, these data include data sources from multiple user sides; extracting features from each user side data source, and outputting the corresponding data source feature distribution, wherein each data source has a corresponding feature distribution; calculating the fusion similarity index between multiple data source feature distributions by analyzing the similarity probability distribution; determining the first data source feature distribution, that is, the data source feature distribution with the highest similarity index according to the fusion similarity index; based on the first data source feature distribution, performing equivalent processing on other data source feature distributions to generate multiple equivalent feature distributions; using these equivalent feature distributions as training data sets, training the power user behavior analysis model, and then predicting power demand according to the model. Thus, the technical effect of improving information utilization and analysis efficiency and improving analysis accuracy is achieved. Embodiment 2

[0065] Figure 2 Schematic diagram of the structure of the power user behavior analysis system based on multi-source heterogeneous data of the present invention. Figure 1 The flow chart of the method for analyzing the behavior of power users based on multi-source heterogeneous data of the present invention can be shown as follows: Figure 2 The structure shown is implemented.

[0066] Based on the same concept as the power user behavior analysis method based on multi-source heterogeneous data in the above embodiment, the present invention also provides a power user behavior analysis system based on multi-source heterogeneous data, including: The data extraction module 11 is used to determine multi-source heterogeneous data on the user side of the target power grid, wherein the multi-source heterogeneous data includes multiple user-side data sources.

[0067] The data source feature module 12 is used to extract features from the multiple user-side data sources and output multiple data source feature distributions, wherein each user-side data source corresponds to a data source feature distribution.

[0068] The distribution analysis module 13 is used to analyze the feature distributions of the plurality of data sources based on similarity probability distribution, and determine a plurality of fusion similarity indicators.

[0069] The feature distribution selection module 14 is used to determine a first data source feature distribution based on the multiple fusion similarity indicators, wherein the first data source feature is a data source feature distribution with the highest similarity indicator.

[0070] The equivalent generation module 15 is used to perform equivalent processing on the multiple data source feature distributions according to the first data source feature distribution to obtain multiple equivalent feature distributions.

[0071] The analysis and prediction module 16 is used to call the multiple equivalent feature distributions as training data sets, train a power user behavior analysis model, and perform power prediction according to the power user behavior analysis model.

[0072] Wherein, the distribution analysis module 13 includes: The similarity and feature distribution identification unit is used to identify the feature distributions of the multiple data sources respectively based on the similarity probability distribution, and analyze N-1 similarity indicators between any data source feature distribution and N-1 data source feature distributions one by one, where N is the number of data sources.

[0073] The indicator integration unit is used to integrate the N-1 similarity indicators as fused similarity indicators of feature distributions of each data source, and output the multiple fused similarity indicators.

[0074] In some embodiments, the feature distribution selection module 14 includes: The selection unit is used to analyze the magnitudes of the multiple fusion similarity indicators to determine a first data source feature distribution and a second data source feature distribution, wherein the second data source feature distribution is a data source feature distribution with a second highest similarity indicator.

[0075] The similarity index calculation unit is used to calculate the similarity index between the first data source feature distribution and the second data source feature distribution.

[0076] The first equivalent processing execution unit is configured to perform equivalent processing on the plurality of data source feature distributions using the first data source feature distribution if the similarity index is greater than or equal to a preset similarity threshold.

[0077] The second equivalent processing execution unit is used to perform equivalent processing on the multiple data source feature distributions using the first data source feature distribution and the second data source feature distribution if the similarity index is less than the preset similarity threshold.

[0078] In some implementations, the first equivalent processing execution unit in the feature distribution selection module 14 includes: The equivalent processing model construction unit is used to construct an equivalent processing model, wherein the equivalent processing model includes a dimensionality reduction equivalent processing channel and a high-dimensional equivalent processing channel.

[0079] The data source feature distribution division unit is used to divide the multiple data source feature distributions according to the first data source feature distribution, and output a first type of data source feature distribution and a second type of data source feature distribution.

[0080] Among them, the dimensionality reduction equivalent processing channel is used to perform feature dimensionality reduction equivalent processing on the feature distribution of the first type of data source, and the high-dimensional equivalent processing channel is used to perform feature high-dimensional equivalent processing on the feature distribution of the first type of data source.

[0081] In some implementations, the data source feature distribution division unit in the feature distribution selection module 14 includes: A feature strength acquisition unit is used to acquire a first feature strength of the feature distribution of the first data source.

[0082] The feature strength comparison and classification unit is used to obtain multiple feature strengths of the feature distributions of the multiple data sources.

[0083] The classification output unit is used to output the data source feature distribution whose feature strengths are greater than or equal to the first feature strength among the multiple feature strengths as a first type of data source feature distribution, and to output the data source feature distribution whose feature strengths are less than the first feature strength among the multiple feature strengths as a second type of data source feature distribution.

[0084] In some implementations, the analysis and prediction module 16 includes: The equivalent processing execution unit is used to perform equivalent processing on the multiple data source feature distributions with the first data source feature distribution, and output multiple first equivalent feature distributions.

[0085] The equivalent processing execution unit is used to perform equivalent processing on the multiple data source feature distributions using the second data source feature distribution, and output multiple second equivalent feature distributions.

[0086] The behavior analysis model training unit is used to train the power user behavior analysis model by using the multiple first equivalent feature distributions and the multiple second equivalent feature distributions as training data sets.

[0087] In some embodiments, the analysis and prediction module 16 further includes: The penalty factor establishment and model parameter definition unit is used to establish multiple penalty factors according to the influence relationship between the multiple user-side data sources and the power prediction accuracy.

[0088] The model parameter definition unit defines the model parameters of the power user behavior analysis model based on the multiple penalty factors to obtain a defined initial model for the power user behavior analysis.

[0089] The model training unit calls the multiple equivalent feature distributions as training data sets to train the power user behavior analysis initial model to obtain the power user behavior analysis model.

[0090] It should be understood that the embodiments mentioned in this specification focus on their differences from other embodiments. The specific embodiments in the aforementioned embodiment one are also applicable to the power user behavior analysis system based on multi-source heterogeneous data described in embodiment two. For the sake of brevity of the specification, they will not be further elaborated here.

[0091] It should be understood that the embodiments disclosed in the present invention and the above description can enable those skilled in the art to use the present invention to implement the present invention. At the same time, the present invention is not limited to the above-mentioned embodiments. It should be understood that those skilled in the art can still modify the technical solutions recorded in the above-mentioned embodiments, or replace some of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention.

Claims

1. A method for analyzing power user behavior based on multi-source heterogeneous data, characterized in that: The method comprises: Determine multi-source heterogeneous data on the user side of the target power grid, wherein the multi-source heterogeneous data includes multiple user-side data sources; Extracting features from the multiple user-side data sources, and outputting multiple data source feature distributions, wherein each user-side data source corresponds to a data source feature distribution; Analyze the feature distributions of the multiple data sources based on similarity probability distribution to determine multiple fusion similarity indicators; Determine a first data source feature distribution using the multiple fusion similarity indices, wherein the first data source feature is a data source feature distribution having the highest similarity index; Performing equivalent processing on the plurality of data source feature distributions according to the first data source feature distribution to obtain a plurality of equivalent feature distributions; The multiple equivalent feature distributions are called as training data sets to train a power user behavior analysis model, and power prediction is performed according to the power user behavior analysis model.

2. The method according to claim 1, characterized in that Analyzing the feature distributions of the multiple data sources based on similarity probability distribution to determine multiple fusion similarity indicators, the method includes: Based on the similarity probability distribution, the plurality of data source feature distributions are respectively identified, and N-1 similarity indicators between any data source feature distribution and N-1 data source feature distributions are analyzed one by one, where N is the number of data sources; The N-1 similarity indices are integrated as fused similarity indices of feature distribution of each data source, and the multiple fused similarity indices are output.

3. The method according to claim 1, characterized in that Determining the first data source feature distribution using the multiple fusion similarity indicators, the method further includes: Analyze the magnitudes of the multiple fusion similarity indices to determine a first data source feature distribution and a second data source feature distribution, wherein the second data source feature distribution is a data source feature distribution having a second highest similarity index; Calculate a similarity index between the first data source feature distribution and the second data source feature distribution; If the similarity index is greater than or equal to a preset similarity threshold, performing equivalent processing on the plurality of data source feature distributions using the first data source feature distribution; If the similarity index is less than the preset similarity threshold, the first data source feature distribution and the second data source feature distribution are used to perform equivalent processing on the multiple data source feature distributions.

4. The method according to claim 3, characterized in that The method of performing equivalent processing on the plurality of data source feature distributions by using the first data source feature distribution includes: Constructing an equivalent processing model, wherein the equivalent processing model includes a dimensionality reduction equivalent processing channel and a high-dimensional equivalent processing channel; Dividing the plurality of data source feature distributions according to the first data source feature distribution, and outputting a first type of data source feature distribution and a second type of data source feature distribution; Among them, the dimensionality reduction equivalent processing channel is used to perform feature dimensionality reduction equivalent processing on the feature distribution of the first type of data source, and the high-dimensional equivalent processing channel is used to perform feature high-dimensional equivalent processing on the feature distribution of the first type of data source.

5. The method according to claim 4, characterized in that Dividing the plurality of data source feature distributions according to the first data source feature distribution includes: Obtaining a first feature intensity of the feature distribution of the first data source; Obtaining multiple feature intensities of the feature distributions of the multiple data sources; The data source feature distributions of the multiple feature intensities that are greater than or equal to the first feature intensities are output as first-category data source feature distributions, and the data source feature distributions of the multiple feature intensities that are less than the first feature intensities are output as second-category data source feature distributions.

6. The method according to claim 3, characterized in that Training a power user behavior analysis model, including: Performing equivalent processing on the plurality of data source feature distributions using the first data source feature distribution, and outputting a plurality of first equivalent feature distributions; performing equivalent processing on the plurality of data source feature distributions using the second data source feature distribution, and outputting a plurality of second equivalent feature distributions; The plurality of first equivalent feature distributions and the plurality of second equivalent feature distributions are used as training data sets to train a power user behavior analysis model.

7. The method according to claim 1, characterized in that The plurality of equivalent feature distributions are called as training data sets to train a power user behavior analysis model, the method comprising: Establishing a plurality of penalty factors according to the influence relationship between the plurality of user-side data sources and the accuracy of power prediction; Defining model parameters of the power user behavior analysis model based on the multiple penalty factors to obtain a defined initial model for power user behavior analysis; The multiple equivalent feature distributions are called as training data sets to train the initial model for power user behavior analysis to obtain the power user behavior analysis model.

Images