Multi-source data fusion statistical analysis method based on Bayesian network

Abstract

The invention relates to a multi-source data fusion statistical analysis method based on a Bayesian network. The method comprises the following steps: collecting electric power multi-source data samples of a long time sequence within a period of time; identifying main features of the multi-source data through an auto-encoder to obtain reconstructed multi-source data; constructing a Bayesian network model by using the reconstructed multi-source data, and determining causal probability association features among the multi-source data; identifying time correlation features of the multi-source data by constructing a time sequence model; coupling the Bayesian network model and the time sequence model to obtain a multi-source data association model; and obtaining multi-source data of a to-be-configured time period, inputting the multi-source data association model, identifying a multi-source data variable with abnormal data, and adjusting the abnormal data variable to optimize power configuration. According to the invention, the problem that the accuracy and rationality of power configuration are low due to the fact that relevance among multi-source data is difficult to identify in the prior art is solved.

Description

Technical Field

[0001] The invention belongs to the technical field of power data analysis and relates to a multi-source data fusion statistical analysis method based on a Bayesian network. Background Art

[0002] In the electric power sector, there are data from many sources, such as user electricity consumption data from smart meters (including electricity consumption and power consumption in different time periods), equipment operation status data monitored by power equipment sensors (such as temperature, pressure, vibration, etc.), power flow data in the power grid dispatching system, and meteorological data related to power load provided by the meteorological department (temperature, humidity, wind speed, etc.). These data come from different sources, and the format, accuracy, time scale, etc. may vary.

[0003] A single data source can often only provide partial information and cannot fully reflect the overall operating status of the power system. For example, it is difficult to accurately determine whether an equipment failure will occur based on power consumption data alone, and it is not possible to fully understand the impact of user power consumption behavior on the power grid based solely on equipment monitoring data. By integrating multi-source data, more comprehensive and valuable information can be mined to achieve more accurate analysis, monitoring and management of the power system.

[0004] However, the existing technology for the fusion of multi-source data is more about integrating multi-source data into a single vector representation. Although this method can reflect the overall operating status of the power system, it is difficult to understand the correlation between multi-source data, resulting in difficulty in accurately optimizing the allocation of power resources based on the mutual influence of different factors (such as the correlation between load conditions and power generation capacity, transmission losses, etc.). This may lead to situations such as unreasonable power generation plans (which may lead to excess or insufficient power in some periods of time), inaccurate and inefficient grid scheduling (unable to coordinate the balance of power supply and demand in different regions and links), etc., which will affect the accuracy and rationality of power configuration. Summary of the invention

[0005] In order to solve the above problems existing in the prior art, the present invention provides a multi-source data fusion statistical analysis method based on a Bayesian network, which aims to obtain multi-source data of a long time series of the power system; construct a Bayesian network model to quantify the causal probability association characteristics between the multi-source data; and simultaneously construct a time series model to identify the time association characteristics of the multi-source data; and finally couple the Bayesian network model and the time series model to optimize the power configuration of the power system.

[0006] The purpose of the present invention can be achieved through the following technical solutions:

[0007] The present application provides a multi-source data fusion statistical analysis method based on a Bayesian network, comprising the following steps:

[0008] S1. Collect long-term series of multi-source power data samples over a period of time; the multi-source data includes user power consumption data, power equipment monitoring data, meteorological data and power grid dispatching data;

[0009] S2, identifying the main features of multi-source data through the autoencoder to obtain reconstructed multi-source data;

[0010] S3, constructing a Bayesian network model using reconstructed multi-source data to determine the causal probability association characteristics between multi-source data;

[0011] S4, identify the time correlation characteristics of multi-source data by building a time series model;

[0012] S5, coupling the Bayesian network model and the time series model to obtain a multi-source data association model;

[0013] S6. Obtain multi-source data for the time period to be configured, input the multi-source data association model, identify multi-source data variables with abnormal data, and adjust the abnormal data variables to optimize power configuration.

[0014] Furthermore, the multi-source data unifies the types and formats of the multi-source data through data standardization and data encoding; the data standardization includes unifying the data structure and standardizing the data type, and the data encoding includes classification encoding and encoding conversion.

[0015] Furthermore, in step S2, identifying the main features of the multi-source data by the autoencoder to obtain the reconstructed multi-source data includes the following steps:

[0016] S21. Data preprocessing: perform data cleaning, missing value processing and outlier processing operations on multi-source data;

[0017] S22. Determine the autoencoder architecture: design a suitable autoencoder network architecture, wherein the network architecture includes an input layer, a hidden layer, and an output layer, and uses ReLU as an activation function;

[0018] S23, model training: the preprocessed multi-source data is divided into a training set and a test set, and the mean square error function is selected as the loss function to measure the difference between the autoencoder reconstructed data and the original data. By minimizing the loss function, the autoencoder continuously adjusts the network parameters during the training process to improve the similarity between the reconstructed data and the original data;

[0019] S24. Feature extraction and reconstructed data acquisition: After the autoencoder model training is completed, the original multi-source data is input into the trained encoder part, and the encoder compresses the original data and converts it into a low-dimensional feature vector; the feature vector output by the encoder is then input into the trained decoder part, and the decoder reconstructs the reconstructed multi-source data based on the feature vector.

[0020] Furthermore, in step S3, the Bayesian network model includes the following construction steps:

[0021] S31. Determine nodes and variables: Determine nodes in the Bayesian network according to different attributes and categories reflected by multi-source data, where each node corresponds to a specific multi-source variable;

[0022] S32. Determine the causal relationship and directed edges between nodes: Combine the professional knowledge of power system operation, actual operation experience and the potential relationship between multi-source data to determine the causal relationship between nodes, and combine the analysis and statistics of historical data to sort out the association between different multi-source data, verify and refine the determined causal relationship; for the node pairs determined to have a causal relationship, construct directed edges in the Bayesian network model;

[0023] S33, determining the prior probability distribution of the node: using the collected historical data, the prior probability distribution of each node is calculated by statistical methods, wherein for the nodes of discrete variables, the frequency of occurrence of each observation category in the total number of observations is calculated as its prior probability; for the nodes of continuous variables, the prior probability distribution is determined by using a probability density function;

[0024] S34, calculating conditional probability distribution: according to Bayesian theorem, using the determined node prior probability distribution and the joint probability obtained by analyzing historical data to calculate the conditional probability distribution between nodes;

[0025] S35. Complete the network model: assign the calculated conditional probability distribution to the directed edges between the corresponding nodes to complete the construction of the Bayesian network.

[0026] Furthermore, in step S4, the time series model is configured as a SARIMA model, including the following construction steps:

[0027] S41, extracting time-related features: using MK test to extract time-related features of multi-source data, the time-related features include trend features, seasonal features, periodic features and lag-related features;

[0028] S42, training the SARIMA model with multi-source data as input and time-related features as output;

[0029] S43. Using evaluation indicators to measure the model's prediction ability for time-related features, the evaluation indicators are accuracy, precision, recall rate and F1 value.

[0030] Furthermore, in step S43, the evaluation index is calculated as follows:

[0031]

[0032] In the formula, ACC represents accuracy; PPV represents precision; TRP represents recall; F1 represents F1 value; TP represents true positive examples; TN represents true negative examples; FP represents false positive examples; and FN represents false negative examples.

[0033] Furthermore, in step S5, the multi-source data association model includes the following construction steps:

[0034] S51, feature extraction and fusion: extract the causal probability association feature between multi-source data from the constructed Bayesian network model, specifically including the causal relationship represented by the directed edges between each node and the corresponding conditional probability distribution; extract the time association feature of multi-source data from the trained time series model; fuse the causal probability association feature with the time association feature;

[0035] S52. Model structure integration: Based on the directed acyclic graph structure of the Bayesian network model, the time-related features are integrated into it, and each node in the Bayesian network model is given attributes or dynamic change rules related to the time-related features;

[0036] S53, model parameter adjustment and optimization: by traversing different parameter combinations, observe the performance of the model on the test data, and find the parameter settings that make the evaluation indicators reach the optimal level or meet the preset requirements; during the parameter adjustment process, use verification data to verify the model and optimize model performance.

[0037] Furthermore, in step S52, the model structure integration can also be based on the time series structure of the time series model, and according to the causal probability association characteristics of the Bayesian network model, the data of different time periods can be targetedly grouped or weighted.

[0038] Further, in step S6, the step of acquiring multi-source data of the time period to be configured, inputting the multi-source data association model, identifying multi-source data variables with data anomalies, and adjusting the abnormal data variables to optimize the power configuration includes the following steps:

[0039] S61, data collection and preprocessing: for the time period to be configured, multi-source data is collected from various relevant links of the power system, and the collected multi-source data is preprocessed;

[0040] S62, input model and analysis: input the preprocessed multi-source data into the constructed multi-source data association model, and output the predicted causal probability association feature and time association feature;

[0041] S63, identifying abnormal data variables: according to the preset abnormal judgment standard, comparing the causal probability association feature characteristics and time association features of the multi-source data association model in theory with the prediction results, and determining the multi-source data variables with abnormalities;

[0042] S64. Adjust abnormal data variables to optimize power configuration: adjust abnormal data variables and re-enter the multi-source data association model to form a closed loop until abnormal data variables no longer appear.

[0043] Beneficial effects of the present invention:

[0044] By collecting long-time series of multi-source power data samples within a period of time; identifying the main features of multi-source data through autoencoders to obtain reconstructed multi-source data; using reconstructed multi-source data to build a Bayesian network model to determine the causal probability association features between multi-source data; identifying the time association features of multi-source data by building a time series model; coupling the Bayesian network model and the time series model to obtain a multi-source data association model; obtaining multi-source data for the time period to be configured, inputting the multi-source data association model, identifying multi-source data variables with data anomalies, and adjusting the abnormal data variables to optimize power configuration. The present invention solves the problem that the prior art is difficult to identify the correlation between multi-source data, resulting in low accuracy and rationality of power configuration. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to facilitate understanding by those skilled in the art, the present invention is further described below with reference to the accompanying drawings.

[0046] Figure 1 The present invention is a flowchart of a multi-source data fusion statistical analysis method based on Bayesian network.

[0047] Figure 2 A flowchart of constructing a Bayesian network model in one embodiment of the present invention.

[0048] Figure 3 A flowchart of a time series model construction in one embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation mode, structure, characteristics and effects of the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments.

[0050] See also Figure 1-Figure 3 The present application provides a multi-source data fusion statistical analysis method based on a Bayesian network, comprising the following steps:

[0051] S1. Collect long-term series of multi-source power data samples over a period of time; the multi-source data includes user power consumption data, power equipment monitoring data, meteorological data and power grid dispatching data;

[0052] In this embodiment, data is collected within a specific time period, which is not a short moment or a very short period of time, but has a certain length, such as several years or even longer. The purpose of this is to capture the operating performance of the power system under various conditions such as different seasons, different economic development stages, and different climatic conditions, so as to analyze its laws and characteristics more comprehensively and accurately. "Long time series" means that for each type of data, it is necessary to continuously collect and record at a certain time interval to form a data sequence arranged in chronological order. For example, it can be to record the power consumption once an hour, the temperature in the meteorological data once a day, and some key parameters in the power equipment monitoring data once a minute. Through this long time series of data, it is possible to observe the changing trends of various data indicators over time, periodic laws, seasonal fluctuations, and sudden abnormal changes, which are crucial for a deep understanding of the dynamic operation process of the power system.

[0053] The multi-source data includes user power consumption data, power equipment monitoring data, meteorological data and power grid dispatching data, among which:

[0054] User electricity consumption data: This part of the data mainly reflects the electricity consumption behavior and electricity consumption status of the end users of the power system. It can include information such as the amount of electricity consumption of each user or user groups in different regions at different time points, the size of the electricity consumption, and the peak and valley time distribution of electricity consumption. By analyzing user electricity consumption data, we can understand the electricity consumption habits of different types of users such as residents, commercial and industrial users. For example, residential users may have peak electricity consumption in the summer night mainly because of the use of refrigeration equipment such as air conditioners, while commercial users may have peak electricity consumption during the daytime on weekdays. These habits are of great reference value for predicting power load, reasonably arranging power generation and scheduling, etc.

[0055] Power equipment monitoring data: It is real-time monitoring information about the operating status of various key equipment in the power system (such as generator sets, transmission lines, substation equipment, etc.). Specifically, it includes various operating parameters of the equipment, such as the output power, temperature, speed, vibration frequency, etc. of the generator sets, the current, voltage, temperature, ice thickness, etc. of the transmission lines, the oil temperature, oil level, winding temperature, the opening and closing status of the circuit breaker, the number of operations, etc. of the substation equipment. These data can help to timely discover whether there are hidden dangers of equipment failure, whether it is in the normal operating range, whether the performance of the equipment has declined over time, etc., so as to take maintenance and repair measures in advance to ensure the stable operation of the power system.

[0056] Meteorological data: Meteorological conditions have many impacts on the power system, so it is also one of the important data sources. Meteorological data covers the records of various meteorological elements related to the power system at different times, such as temperature, humidity, wind speed, wind direction, rainfall, snowfall, etc. The temperature will affect the user's electricity demand (such as the increase in air conditioning electricity consumption at high temperatures), and will also affect the operating performance of power equipment (such as low temperature may affect the start-up and operating efficiency of equipment); humidity will affect the insulation performance of power equipment; wind speed is directly related to the power generation capacity of wind turbines in wind power access areas, and excessive wind speed may also cause wind deviation on transmission lines. Combining these meteorological data, we can better consider the impact of environmental factors on the power system and prepare countermeasures in advance.

[0057] Grid dispatch data: This is data related to dispatch management during grid operation. It includes power generation plan data, i.e., the total planned power generation in different time periods and the allocation tasks of each power generation unit, etc. It reflects how the grid dispatch department arranges power generation resources according to the predicted power demand and the actual situation of the power system; it also includes flow data, i.e., the flow direction and distribution of voltage, current, power, etc. at each node in the grid over time. Flow data can reflect the real-time operation status of the grid, which is of great significance for discovering blocking points in the grid and optimizing the grid topology. By analyzing the grid dispatch data, we can understand the rationality and effectiveness of grid dispatch and whether it can meet the operation needs of the power system.

[0058] Furthermore, the multi-source data unifies the types and formats of the multi-source data through data standardization and data encoding.

[0059] In this embodiment, through data standardization and data encoding, the types and formats of multi-source data in the power system can be effectively unified, providing a good data foundation for subsequent work such as multi-source data fusion statistical analysis based on Bayesian networks, so that data from different sources can be more smoothly integrated, analyzed and applied, thereby better exploring the operating laws of the power system and realizing the optimal management of the power system.

[0060] Data standardization includes unifying data structures and standardizing data types:

[0061] For multi-source data in the power system, such as user power consumption data, power equipment monitoring data, meteorological data, and power grid dispatching data, a common and standardized data structure model will be formulated. For example, for user power consumption data, it may be required to be presented in a table format, which contains fixed column headers, such as "user number", "collection time", "power consumption (kWh)", "power consumption (kW)", etc. In this way, no matter which specific collection point or system the data comes from, it can be organized according to this unified structure, so that the data is consistent in form, which is convenient for subsequent processing and analysis. For power equipment monitoring data, a similar standardized structure will also be clarified. For example, with equipment as the unit, each row records the data corresponding to a monitoring time, and the column headers include "equipment number", "monitoring time", "equipment temperature (Celsius)", "equipment vibration frequency (Hz)" and other key parameters reflecting the operating status of the equipment. Through this unified data structure, the monitoring conditions of different equipment at different times can be clearly presented in an orderly manner, providing a neat and standardized basic data format for fusion analysis.

[0062] Clarify the exact data type that should be used for each data field. Taking electricity consumption data as an example, it is stipulated that the floating-point data type should be used uniformly, so that the numerical value of electricity consumption can be accurately represented, avoiding calculation errors or analysis deviations caused by inconsistent data types (for example, some places may use integers to record electricity consumption, which will lose information about the decimal part). For fields representing time, such as "collection time", a specific date and time format will be uniformly used, such as the ISO 8601 format, to ensure that the time information of different data sources can be accurately compared and sorted, which is crucial for analyzing the changes in data over time.

[0063] For some text-descriptive data, such as text descriptions of equipment status ("normal", "fault", "warning", etc.), their encoding methods will also be standardized, and digital codes may be used instead of text. For example, "0" represents normal, "1" represents fault, "2" represents warning, etc., and the text description is converted into a unified data type (integer) to facilitate quantitative operations and logical judgments during data analysis and processing.

[0064] Data coding includes classification coding and coding conversion:

[0065] Data from different sources are classified and coded so that the source and attributes of the data can be quickly identified during the data fusion process. For example, a specific coding category is assigned to user electricity data, such as "01", power equipment monitoring data is coded as "02", meteorological data is "03", and power grid dispatch data is "04". In this way, in the subsequent data processing process, when you see a piece of data with a specific code, you can immediately know the data category to which it belongs, which is convenient for targeted processing and fusion operations.

[0066] Specific items or parameters in the data can also be further encoded. For example, in the power equipment monitoring data, different codes can be given to different equipment types (such as generator sets, transmission lines, substation equipment, etc.), and there can also be corresponding coding systems for different monitoring parameters under the same equipment type (such as the temperature, speed, vibration frequency, etc. of the generator set). Through this detailed coding, data can be managed and used more accurately, improving the efficiency and accuracy of data fusion.

[0067] In some cases, different data sources may use different encoding methods for the same concept or parameter. For example, in the weather data, some weather stations may use "1" to represent sunny days and "2" to represent cloudy days, while other weather stations may use "01" to represent sunny days and "02" to represent cloudy days. In this case, it is necessary to perform code conversion to unify all weather condition codes into a standard coding method, such as uniformly using "01" to represent sunny days and "02" to represent cloudy days, to ensure that there will be no confusion or wrong judgments due to coding differences during data fusion.

[0068] Similarly, for the codes of some equipment operating status, if the equipment produced by different manufacturers uses different codes to represent the same status, they also need to be unified through code conversion. For example, transformers from different manufacturers may use different codes to represent the state of oil temperature being too high. By analyzing and formulating unified code conversion rules, the codes of all manufacturers' equipment regarding oil temperature being too high can be unified into a standard code, so that the real operating status of the equipment can be accurately identified during the data fusion and analysis process.

[0069] S2, identifying the main features of multi-source data through the autoencoder to obtain reconstructed multi-source data;

[0070] In this embodiment, the autoencoder is a neural network model for unsupervised learning, which is mainly used to learn the effective representation of input data. It consists of two major parts: an encoder and a decoder. The encoder receives the original input data, such as multi-source data in the power system (including data on user electricity consumption, equipment monitoring, meteorology, power grid dispatching, etc.), compresses and transforms it through a series of neural network layers and activation functions, maps the high-dimensional original data to a low-dimensional feature space, and extracts key features to form a low-dimensional feature vector. The decoder receives the feature vector, and then expands and transforms it through the corresponding neural network layer and activation function, trying to construct reconstructed data that is as similar as possible to the original input data, which can reflect the main characteristics of the original data as a whole, thereby realizing the function of automatically extracting representative features from the input data and reconstructing the data.

[0071] Furthermore, in step S2, identifying the main features of the multi-source data by the autoencoder to obtain the reconstructed multi-source data includes the following steps:

[0072] S21. Data preprocessing: Clean multi-source data (including user electricity consumption data, power equipment monitoring data, meteorological data, and power grid dispatching data, etc.). Check and process missing values ​​in the data. For example, for the missing parts in the electricity consumption data, they can be reasonably filled according to the user's historical electricity consumption rules (such as mean filling, adjacent time point interpolation, etc.); at the same time, remove obvious erroneous data, such as abnormal values ​​that exceed the reasonable range in the power equipment monitoring data (which may be caused by sensor failure, etc.), and judge and eliminate these abnormal data by setting reasonable thresholds to ensure the quality and reliability of the data.

[0073] S22. Autoencoder model construction:

[0074] Determine the network architecture: Design a suitable autoencoder network architecture based on the characteristics and analysis requirements of multi-source data. Generally speaking, an autoencoder consists of an input layer, several hidden layers (including hidden layers of the encoder part and hidden layers of the decoder part), and an output layer. For multi-source power data, the number of hidden layers, the number of neurons in each layer, etc. need to be determined based on the dimension and complexity of the data. For example, if the multi-source data has a high dimension and complex features, it may be necessary to set up more hidden layers (such as 3 to 5 layers), and the number of neurons in each layer can be gradually decreased (in the encoder part) and then gradually increased (in the decoder part) to achieve effective data compression and feature extraction as well as subsequent reconstruction functions.

[0075] Select activation function: Select a suitable activation function for the neurons of the autoencoder. In the encoder part, commonly used activation functions include ReLU (Rectified Linear Unit), which can effectively handle nonlinear data relationships and perform nonlinear transformations on the input data, so that the encoder can better extract complex features in the data. In the decoder part, you can also select a suitable activation function as needed, such as Sigmoid function (when the reconstructed data needs to be in the interval [0,1]) or ReLU, etc., to ensure that the decoder can accurately reconstruct data similar to the original data based on the feature vector output by the encoder.

[0076] S23, model training:

[0077] Divide the preprocessed multi-source data into two parts according to a certain ratio (e.g. 80% as training set and 20% as test set). The training set is used to train the autoencoder model so that it can learn the feature representation of multi-source data; the test set is used to evaluate the performance of the model during and after training to verify whether the model can accurately extract features and reconstruct data.

[0078] Define the loss function: Choose a suitable loss function to measure the difference between the autoencoder reconstructed data and the original data. Common loss functions include the mean square error (MSE) function, which calculates the average of the sum of squares of the differences between the reconstructed data and the corresponding elements of the original data. By minimizing this loss function, the autoencoder continuously adjusts the network parameters during the training process to improve the similarity between the reconstructed data and the original data.

[0079] Set training parameters: Determine the relevant parameters for model training, such as learning rate, training rounds (epochs), etc. The learning rate determines the step size of each update of the model parameters. A suitable learning rate can ensure the stability and efficiency of model training (if the learning rate is too large, it may cause the training process to skip the optimal solution, and if it is too small, the training speed will be too slow); the training rounds indicate the number of times the model conducts a complete training of the entire training set. Generally, it is necessary to conduct multiple experiments according to the specific situation to determine the appropriate training rounds to ensure that the model can fully learn the characteristics of multi-source data.

[0080] Start training: Use the training set to train the autoencoder model. According to the set loss function and training parameters, the network parameters of the autoencoder are continuously adjusted through the back propagation algorithm, so that the loss function value continues to decrease, that is, the similarity between the reconstructed data and the original data continues to increase. During the training process, you can regularly check the changes in the loss function values ​​on the training set and the test set to understand the training progress and performance of the model.

[0081] S24, feature extraction and reconstruction data acquisition:

[0082] Feature extraction: After the autoencoder model is trained, the original multi-source data is input into the trained encoder part, which compresses the original data and converts it into a low-dimensional feature vector. This feature vector contains the main features of the multi-source data. For example, for the power system data, it is possible to extract features related to the trend of power load changes, key features of equipment operating status, and key features of the impact of meteorological conditions on the power system.

[0083] Reconstructed data acquisition: The feature vector output by the encoder is input into the trained decoder part, and the decoder will reconstruct the reconstructed data similar to the original input data based on this feature vector. In this way, reconstructed multi-source data is obtained, which is reconstructed based on the recognition and extraction of the main features of the original multi-source data by the autoencoder, and can be used as a new data representation for subsequent analysis and processing. At the same time, the effect of the autoencoder on feature extraction and reconstruction can be evaluated by calculating the similarity between the reconstructed data and the original data (such as using indicators such as mean square error) to determine whether it meets the requirements of subsequent analysis.

[0084] S3, constructing a Bayesian network model using reconstructed multi-source data to determine the causal probability association characteristics between multi-source data;

[0085] In this embodiment, the Bayesian network model is a valuable probabilistic graphical model that presents the complex relationship between variables in the structure of a directed acyclic graph (DAG). It is widely used in many fields, especially in dealing with situations such as power systems that involve multi-source data and complex relationships. The model is mainly composed of two parts: nodes and directed edges. The nodes symbolize various random variables. Take the power system as an example. There will be "power consumption nodes" representing the power consumption of users, "equipment temperature nodes" and "equipment vibration frequency nodes" that reflect the operating status of power equipment, "temperature nodes" and "humidity nodes" that reflect meteorological factors, and "current distribution nodes" and "power generation plan nodes" related to power grid scheduling. Each node has its own unique probability distribution, and its value will change according to the actual situation and its relationship with other nodes.

[0086] The directed edge is the key link connecting these nodes, which accurately expresses the causal relationship or conditional dependency between variables. For example, in the context of the power system, the "temperature node" points to the "power consumption node", which clearly shows that changes in temperature will affect the user's power consumption, that is, there is a causal flow from temperature to power consumption. Moreover, this relationship is conditional. When the specific temperature conditions are known, the probability distribution of power consumption will be affected accordingly. Through the Bayesian network model, not only can the causal relationship between different variables be intuitively and clearly understood, but more importantly, these relationships can be quantified with the help of conditional probability distribution. For example, if the values ​​of certain parameters in the meteorological conditions are known, the model can be used to accurately calculate the probability of specific changes in other related variables such as power load. In this way, it can deeply explore the inherent connection of the data, provide a strong, accurate and reliable decision-making basis for many key links such as power system operation analysis, fault diagnosis, and load forecasting, and effectively promote the scientific management and efficient operation of the power system.

[0087] Furthermore, in step S3, the Bayesian network model includes the following construction steps:

[0088] S31. Determine nodes and variables:

[0089] Determine the node type: Determine the nodes in the Bayesian network based on the different attributes and categories reflected by the data; each node corresponds to a specific variable, which will serve as the basic elements for building network relationships.

[0090] S32. Determine the causal relationship and directed edges between nodes:

[0091] Based on professional knowledge and experience: Combine professional knowledge of power system operation, actual operation experience and understanding of the potential relationship between various data to determine the possible causal relationship between nodes. For example, according to common sense, rising temperatures will increase the power consumption of residents' air conditioners, so it can be determined that there is a causal relationship between the "temperature node" and the "user power consumption node", and the causal flow is from the "temperature node" to the "user power consumption node".

[0092] Sorting out data associations: Through the analysis and statistics of historical data, further sorting out the associations between different data, verifying and refining the determined causal relationships. For example, looking at the actual data of user electricity consumption under different temperature conditions in the past, to more accurately grasp the degree and manner of the impact of the "temperature node" on the "user electricity consumption node", so as to more reasonably determine the directed edges between the two and the approximate range of the probability distribution of related conditions.

[0093] Constructing directed edges: For node pairs that are determined to have a causal relationship, construct directed edges for them in the Bayesian network model. The direction of the directed edge clearly indicates the flow direction of the causal relationship, that is, the cause node (parent node) points to the result node (child node). Continuing with the above example, a directed edge is established from the "temperature node" to the "user electricity consumption node" between the "temperature node" and the "user electricity consumption node".

[0094] S33. Determine the prior probability distribution of the node:

[0095] Calculate prior probability: Use the collected historical data to calculate the prior probability distribution of each node through statistical methods. For nodes with discrete variables, such as the "weather condition node" (assuming it is divided into several conditions such as sunny, cloudy, rainy, etc.), calculate the frequency of each weather condition in the total number of observations as its prior probability; for nodes with continuous variables, such as the "user electricity consumption node", a probability density function (such as normal distribution) can be used to fit the data to determine its prior probability distribution.

[0096] S34. Calculate the conditional probability distribution:

[0097] Based on joint probability and prior probability: According to Bayes' theorem, the conditional probability distribution between nodes is calculated using the determined node prior probability distribution and the joint probability obtained through the analysis of historical data (i.e. the probability of a certain value combination appearing at multiple nodes at the same time). For example, given the prior probability distribution of the "temperature node" and the joint probability of a specific value combination appearing at the "temperature node" and the "user power consumption node", the conditional probability of the "user power consumption node" appearing at the corresponding value can be calculated given a certain value of the "temperature node".

[0098] S35. Complete the network model: Assign the calculated conditional probability distribution to the directed edges between the corresponding nodes to make the Bayesian network model more complete. By continuously calculating and improving the conditional probability distribution between each node, the causal probability association characteristics between multi-source data can be more accurately reflected, thereby providing strong support for subsequent analysis, prediction and decision-making based on the model. After the model is completed, the test set can be used to verify the model's predictive ability to optimize the model performance.

[0099] S4, identify the time correlation characteristics of multi-source data by building a time series model;

[0100] In this embodiment, the time series model is a statistical model that is specifically used to conduct in-depth analysis of data arranged in chronological order. In the power system, multiple sources of data are involved, such as user power consumption data, power equipment monitoring data, meteorological data, and power grid dispatching data. These data accumulate over time and influence each other. The time series model is analyzed based on this.

[0101] The first is the choice of model. Common models include autoregressive moving average model (ARMA), autoregressive difference moving average model (ARIMA), seasonal autoregressive moving average model (SARMA), seasonal autoregressive difference moving average model (SARIMA), etc. The specific model to be selected should be determined based on the characteristics of multi-source data and the specific purpose of the analysis. For example, if the data has obvious seasonal fluctuations, such as electricity consumption changes in different seasons due to the use of cooling or heating equipment, then the SARIMA model is more suitable.

[0102] After selecting the model, it is necessary to accurately estimate the model parameters through statistical methods such as maximum likelihood estimation and least squares method to find the parameter values ​​that allow the model to best fit the data.

[0103] The core advantage of this model is that it can identify multiple time-related features of multi-source data. It can clearly present the trend characteristics of data over a longer time range, such as the trend of increasing electricity consumption of residents year by year with economic development; it can accurately capture seasonal characteristics, such as the law of peak electricity consumption due to frequent use of air conditioners in summer; it can also discover other periodic characteristics, such as the difference in industrial electricity consumption on weekdays and weekends with a weekly cycle; it can also identify lagging correlation characteristics, that is, after a data source changes, it will take some time to affect the data of other data sources. In short, the time series model provides an extremely important basis and strong support for in-depth understanding of the dynamic change law of multi-source data in the power system, optimizing the operation and management of the power system, and improving the prediction ability.

[0104] Furthermore, in step S4, the time series model is configured as a SARIMA model, including the following construction steps:

[0105] S41, extracting time-related features: using MK test to extract time-related features of multi-source data, the time-related features include trend features, seasonal features, periodic features and lag-related features;

[0106] The MK test (Mann-Kendall Test) is a non-parametric statistical test method, which is often used to analyze the trend and seasonality of time series data. Its advantage is that it does not require specific assumptions about the distribution of data, can effectively process various types of data, and is more applicable when analyzing the time correlation characteristics of multi-source data in power systems.

[0107] Trend feature extraction:

[0108] For multi-source data (such as user electricity consumption data, power equipment monitoring data, meteorological data, and power grid dispatching data, etc.), the MK test can determine trend characteristics by analyzing the sequential changes of data points in a longer time series. For example, in user electricity consumption data, if the cumulative value of electricity consumption shows a continuous upward or downward trend over time, the MK test can identify whether the overall trend is an upward trend or a downward trend, or determine whether it is a relatively stable state.

[0109] Seasonality feature extraction:

[0110] When it comes to data with seasonal fluctuations, the MK test can extract seasonal characteristics by observing the repetitive patterns of data in different seasons or fixed time periods. Taking meteorological data as an example, it can find seasonal characteristics such as the generally high temperature in summer and generally low temperature in winter every year, as well as the obvious difference in rainfall in the rainy season and the dry season, thereby determining the seasonal variation pattern of the data in the time series.

[0111] Periodic feature extraction:

[0112] In addition to obvious seasonal cycles, multi-source data may also have other cyclical characteristics. The MK test identifies these cyclical characteristics by analyzing the repeated changes in data at different time intervals. For example, industrial electricity consumption data may have a weekly change pattern, that is, higher electricity consumption on weekdays and lower electricity consumption on weekends. The MK test can capture this other cyclical change pattern that is not based on an annual cycle, reflecting the cyclical fluctuations of data in the time series.

[0113] Lag-correlation feature extraction:

[0114] In the power system, there may be a time-lag correlation between data from different data sources. The MK test can extract the time-lag correlation features by analyzing the response of data from other data sources in the subsequent time period after the data from one data source changes. For example, the change in rainfall in meteorological data may affect the water level of the river a few days later, and then affect the power generation of the hydropower plant. The MK test can identify the time-lag correlation between the change in rainfall and the change in power generation, and determine the time-lag influence pattern between data from different data sources.

[0115] S42, training the SARIMA model with multi-source data as input and time-related features as output;

[0116] Data preparation and input:

[0117] First, the multi-source data of the power system after cleaning and preprocessing (such as removing missing values, outliers, unifying data formats, etc.) is used as the input of the SARIMA model. These multi-source data contain rich information and are the basic materials for the model to learn time-related features.

[0118] Set target output:

[0119] The time-related features (including trend features, seasonal features, cyclical features, and lagged-related features) extracted by the MK test are used as the target output of model training. This means that during the training process, the SARIMA model should strive to make its output as close as possible to these known time-related features, so as to learn the time-related laws between data.

[0120] Training process:

[0121] During the training phase, the SARIMA model will gradually optimize the performance of the model by adjusting its own parameters (such as seasonal cycle, autoregressive order, difference order, and moving average order, etc.) according to the input multi-source data and the set target output. It will constantly try different parameter combinations to find the parameter setting that can make the model output closest to the target output (that is, the extracted time-related features), thereby achieving effective learning and fitting of the time-related features of multi-source data.

[0122] S43. Using evaluation indicators to measure the model's prediction ability for time-related features, the evaluation indicators are accuracy, precision, recall rate and F1 value.

[0123] Accuracy:

[0124] Accuracy refers to the proportion of correct predictions by the model to the total prediction results. When evaluating the predictive ability of the SARIMA model for time-related features, it measures the match between the time-related features predicted by the model and the actual time-related features. For example, if the number of times the trend characteristics and seasonal characteristics of user electricity consumption data predicted by the model match the actual situation accounts for a high proportion of the total prediction times, it means that the model has a high accuracy, that is, the model can more accurately predict the time-related features of multi-source data.

[0125] Precision:

[0126] Precision mainly focuses on the proportion of results predicted by the model as positive (i.e., predicted the existence of a certain time-related feature) that are truly positive (i.e., the time-related feature actually exists). For example, when predicting whether user electricity consumption data has seasonal characteristics, if the proportion of cases where the model predicts that there are seasonal characteristics is high, it means that the model has high precision. It reflects the reliability of the model's prediction results, that is, the model will not easily misjudge the existence of a certain time-related feature when it does not exist.

[0127] Recall:

[0128] The recall rate measures the proportion of time-related features that the model can correctly predict when there is a certain time-related feature. Taking periodic features as an example, if there is a certain periodic change pattern in the actual power equipment monitoring data, the recall rate indicates the proportion of such periodic features that the model can accurately identify. It reflects the model's ability to capture various time-related features, that is, the model will not miss the actual time-related features.

[0129] F1 Score:

[0130] The F1 value is an evaluation indicator that comprehensively considers precision and recall. It is obtained by calculating the harmonic mean of precision and recall. The F1 value can more comprehensively evaluate the model's ability to predict time-related features because it takes into account both the reliability of the model's prediction results (precision) and the ability to capture actual features (recall). When the F1 value is high, it means that the model performs well in predicting time-related features, and can accurately predict while not missing actual features.

[0131] Furthermore, in step S43, the evaluation index is calculated as follows:

[0132]

[0133] In the formula, ACC represents accuracy; PPV represents precision; TRP represents recall; F1 represents F1 value; TP represents true positive examples; TN represents true negative examples; FP represents false positive examples; and FN represents false negative examples.

[0134] S5, coupling the Bayesian network model and the time series model to obtain a multi-source data association model;

[0135] In this embodiment, the Bayesian network model is good at capturing the causal probability association characteristics between data, and can clearly present the causal relationship between different variables (such as user power consumption data, power equipment monitoring data, meteorological data, power grid dispatching data, etc.) and quantify this relationship through conditional probability distribution. The time series model, especially when configured as a SARIMA model, focuses on identifying the time association characteristics of data, such as trend characteristics, seasonal characteristics, periodic characteristics, and lag association characteristics. The purpose of coupling these two models is to combine their respective advantages and build a more comprehensive and powerful multi-source data association model, so as to more deeply and accurately mine the complex internal connections between multi-source data in the power system, thereby providing more powerful support for many aspects such as operation analysis, fault diagnosis, and load forecasting of the power system.

[0136] Furthermore, in step S5, the multi-source data association model includes the following construction steps:

[0137] S51, feature extraction and fusion:

[0138] Bayesian network feature extraction: Extract causal probability association features between multi-source data from the constructed Bayesian network model. Specifically, it includes the causal relationship represented by the directed edges between nodes and the corresponding conditional probability distribution. For example, in the context of the power system, extract the causal influence of the "temperature node" on the "power consumption node" and the probability distribution of the "power consumption node" under different temperature conditions.

[0139] Time series feature extraction: Extract the time-related features of data from the trained time series model (such as the SARIMA model). It includes trend features (such as the upward or downward trend of electricity consumption over time), seasonal features (such as the peak of electricity consumption in summer occurs in the afternoon and evening), periodic features (such as the change pattern of industrial electricity consumption in a weekly cycle) and lagged correlation features (such as the impact of rainfall changes on power generation after a few days).

[0140] Feature fusion: The causal probability association features of the Bayesian network model extracted above are fused with the time association features of the time series model. Two different types of features are combined in a reasonable way to form a more comprehensive and integrated feature description. For example, combining the causal impact of the "temperature node" on the "electricity consumption node" and the seasonal characteristics of electricity consumption, it can be described as follows: rising temperatures will lead to an increase in electricity consumption in summer, and this increase is more obvious in the afternoon and evening. At the same time, the specific probability of this increase can be clarified based on the Bayesian conditional probability distribution. Through such fusion, the model can understand and describe the relationship between data from multiple perspectives.

[0141] S52, Model structure integration:

[0142] Based on the Bayesian network structure: Based on the directed acyclic graph (DAG) structure of the Bayesian network model, consider how to integrate the time-related features of the time series model into it. For each node in the Bayesian network model, additional attributes or dynamic change rules are given to the node based on the relevant time-related features extracted by the time series model. For example, for the "electricity consumption node", in addition to the original causal relationship with other nodes, the trend characteristics and seasonal characteristics of electricity consumption identified by the time series model are combined to make it a more dynamic and comprehensive node. This integration can be achieved by adding subsidiary variables or modifying the probability distribution of the node.

[0143] Based on the time series structure: or based on the time series structure of the time series model, according to the causal probability association characteristics of the Bayesian network model, the data of different time periods are targetedly grouped or weighted. For example, if the "temperature node" has a causal effect on the "electricity consumption node", when analyzing the time series of electricity consumption data, the electricity consumption data can be grouped according to different values ​​of the temperature, and then the time association characteristics of each group of data can be studied separately, which can better reflect the causal probability association characteristics between the data. In this way, the causal relationship of the Bayesian network model is integrated into the analysis process of the time series model.

[0144] S53, Model parameter adjustment and optimization:

[0145] Determine the evaluation indicators: Select appropriate evaluation indicators to measure the performance of the multi-source data association model. You can consider indicators such as mean square error (MSE), mean absolute error (MAE), accuracy, precision, recall, and F1 value. These indicators can reflect the model's ability to fit the data, predict, and capture features from different aspects.

[0146] Parameter adjustment: According to the selected evaluation indicators, adjust the parameters of the coupled multi-source data association model. By constantly trying different parameter combinations and observing the performance of the model on the test data, find the parameter settings that make the evaluation indicators reach the optimal level or meet certain requirements. For example, adjust the parameters related to node attribute fusion, grouping or weighted processing parameters, etc. to improve the performance and accuracy of the model.

[0147] Optimization verification: During the parameter adjustment process, the model is continuously verified with verification data to ensure that the performance of the model is continuously optimized. Through repeated experiments and verifications, the multi-source data association model can better adapt to the characteristics of multi-source data in the power system, accurately describe the complex relationship between data, and provide reliable support for subsequent analysis, prediction and decision-making.

[0148] S6. Obtain multi-source data for the time period to be configured, input the multi-source data association model, identify multi-source data variables with abnormal data, and adjust the abnormal data variables to optimize power configuration.

[0149] In this embodiment, multi-source data of the power system in the time period to be configured is obtained, covering aspects such as user power consumption, equipment monitoring, meteorology and power grid dispatching. After collection, it is pre-processed, such as cleaning, standardization, and order arrangement, so that it is regular and reliable. Then it is input into the multi-source data association model constructed by coupling the Bayesian network and the time series model. The model can comprehensively analyze and use the causal probability association characteristics determined by the Bayesian network and the time association characteristics identified by the time series model to fully grasp the data situation. Based on the model analysis results, combined with the characteristics of the power system, determine the abnormal judgment criteria, such as variable values ​​exceeding the range, inconsistent change trends, etc. According to this standard, multi-source data variables with data abnormalities are identified, such as abnormal temperature rise of equipment, unreasonable fluctuations in user power consumption, etc. Finally, adjustment strategies are formulated for these abnormal variables, such as guiding users to use electricity reasonably for abnormal power consumption, maintenance and repair for abnormal equipment, and adjusting the power grid dispatching plan for abnormal meteorology. And implement adjustment operations, optimize the power configuration in the time period to be configured through these measures, improve the stability, reliability and operation efficiency of the power system, and ensure its reasonable and orderly operation.

[0150] Further, in step S6, the step of acquiring multi-source data of the time period to be configured, inputting the multi-source data association model, identifying multi-source data variables with data anomalies, and adjusting the abnormal data variables to optimize the power configuration includes the following steps:

[0151] S61. Data collection and preprocessing:

[0152] Comprehensive data collection: For the time period to be configured, multi-source data is comprehensively collected from all relevant links of the power system. This includes obtaining detailed power consumption data from the user side, such as power consumption and power consumption in different time periods; collecting equipment monitoring data from the power equipment side, such as equipment temperature, vibration frequency, operating status and other information; obtaining meteorological data within the time period from the meteorological department, such as temperature, humidity, rainfall, wind speed, etc.; and obtaining power grid dispatching data from the power grid dispatching center, such as power flow distribution, power generation plan, load forecast, etc. Ensure that the data sources are extensive and comprehensive, and can fully reflect the overall status of the power system during the time period to be configured.

[0153] Data preprocessing operation: Perform a series of preprocessing on the collected multi-source data. First, perform data cleaning and process the missing values ​​in the data. Appropriate filling methods can be used according to the characteristics of the data. For example, for the missing part of the electricity consumption data, if the data fluctuations in the surrounding time period are not large, the mean value can be used for filling; for obviously abnormal values, they can be eliminated by setting a reasonable threshold. Then perform data standardization to unify data of different magnitudes and formats into a standard form that is easy to analyze. For example, the Z-score standardization method is used to make the data have similar numerical ranges and distribution characteristics. Finally, ensure that the data is accurately arranged in chronological order to prepare for subsequent input models.

[0154] S62. Input model and analysis:

[0155] Input multi-source data association model: The pre-processed multi-source data of the time period to be configured is provided as input to the constructed multi-source data association model. This model combines the advantages of the Bayesian network model in capturing causal probability association features and the time series model in identifying time association features, and can perform in-depth analysis of the input data.

[0156] Comprehensive model analysis: After receiving the input data, the multi-source data association model conducts comprehensive analysis based on its own structure and parameter settings. On the one hand, the function of the Bayesian network part is used to analyze the causal relationship between different variables and the probability of such a relationship, such as judging the causal impact of temperature changes on electricity consumption and the probability of electricity consumption changes under different temperature conditions. On the other hand, with the help of the function of the time series part, the time correlation characteristics of the data are identified, such as seasonal changes, periodic changes, and lagged correlation characteristics of electricity consumption data, so as to accurately grasp the overall situation of multi-source data in the time period to be configured.

[0157] S63. Identify abnormal data variables:

[0158] Determine the abnormality judgment standard: Combine the operation principle of the power system, historical data and past experience to formulate the judgment standard of data abnormality. These standards cover many aspects. For example, if the temperature of a certain power equipment exceeds its normal operating temperature range by a certain margin, it can be judged as abnormal equipment temperature data; if the deviation of the user's power consumption exceeds a certain percentage compared with the historical average power consumption in the same time period, it is considered as abnormal power consumption data; if the rainfall in the meteorological data is greatly deviated from the historical rainfall in the same period in the area, it is also considered as abnormal meteorological data, etc.

[0159] Identify abnormal variables according to standards: Based on the established abnormal judgment standards, carefully screen and identify those multi-source data variables that meet the abnormal standards in the results of the input multi-source data association model analysis. For example, it may be identified that the operating temperature of a certain power equipment is continuously too high during the configuration period, exceeding the normal range, and it is determined to be an abnormal equipment temperature variable; or the power consumption of users in a certain area during this period has dropped significantly compared with the same period in history, deviating from the normal situation, and it is confirmed as an abnormal power consumption variable. By accurately identifying these abnormal variables, specific goals can be clarified for subsequent adjustment and optimization work.

[0160] S64. Adjust abnormal data variables to optimize power configuration:

[0161] Formulate targeted adjustment strategies: Develop targeted adjustment strategies for different types of abnormal data variables identified. For abnormal user power consumption data, such as excessive power consumption, consider guiding users to use power reasonably by appropriately raising electricity prices and carrying out energy-saving publicity. If power consumption is too low, analyze the reasons. If it is due to equipment failure affecting user power consumption, arrange maintenance in time. If it is caused by the user's own behavior habits, encourage users to increase power consumption through preferential policies. For abnormal power equipment monitoring data, such as excessive equipment temperature, immediately arrange equipment maintenance and overhaul work, and replace relevant components when necessary to ensure the normal operation of the equipment. For abnormal meteorological data, such as extreme weather conditions, adjust the power grid dispatch plan according to the degree of impact of the weather on the power system, increase standby power generation capacity, optimize power flow distribution, etc. For abnormal power grid dispatch data, such as excessive deviation between load forecast and actual load, re-evaluate and optimize power generation plan, adjust power flow distribution and other measures.

[0162] Implement adjustment operations: According to the formulated adjustment strategy, actual adjustment operations are performed on the identified abnormal data variables. During the implementation process, pay close attention to the effect of the adjustment and feedback the adjustment information in a timely manner so as to further optimize the adjustment strategy according to the actual situation. Through the effective adjustment of abnormal data variables, the purpose of optimizing power configuration is achieved, and the stability, reliability and operating efficiency of the power system in the period to be configured are improved.

[0163] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A multi-source data fusion statistical analysis method based on Bayesian network, characterized by: The following steps are involved: S1. Collect long-term series of multi-source power data samples over a period of time; the multi-source data includes user power consumption data, power equipment monitoring data, meteorological data and power grid dispatching data; S2, identifying the main features of multi-source data through the autoencoder to obtain reconstructed multi-source data; S3, constructing a Bayesian network model using reconstructed multi-source data to determine the causal probability association characteristics between multi-source data; S4, identify the time correlation characteristics of multi-source data by building a time series model; S5, coupling the Bayesian network model and the time series model to obtain a multi-source data association model; S6. Obtain multi-source data for the time period to be configured, input the multi-source data association model, identify multi-source data variables with abnormal data, and adjust the abnormal data variables to optimize power configuration.

2. The multi-source data fusion statistical analysis method based on Bayesian network according to claim 1 is characterized by: The multi-source data unifies the types and formats of the multi-source data through data standardization and data encoding; the data standardization includes unifying the data structure and standardizing the data type, and the data encoding includes classification encoding and encoding conversion.

3. The multi-source data fusion statistical analysis method based on Bayesian network according to claim 1 is characterized by: In step S2, the main features of the multi-source data are identified by the autoencoder to obtain the reconstructed multi-source data, including the following steps: S21. Data preprocessing: perform data cleaning, missing value processing and outlier processing operations on multi-source data; S22. Determine the autoencoder architecture: design a suitable autoencoder network architecture, wherein the network architecture includes an input layer, a hidden layer, and an output layer, and uses ReLU as an activation function; S23, model training: the preprocessed multi-source data is divided into a training set and a test set, and the mean square error function is selected as the loss function to measure the difference between the autoencoder reconstructed data and the original data. By minimizing the loss function, the autoencoder continuously adjusts the network parameters during the training process to improve the similarity between the reconstructed data and the original data; S24. Feature extraction and reconstructed data acquisition: After the autoencoder model training is completed, the original multi-source data is input into the trained encoder part, and the encoder compresses the original data and converts it into a low-dimensional feature vector; the feature vector output by the encoder is then input into the trained decoder part, and the decoder reconstructs the reconstructed multi-source data based on the feature vector.

4. The multi-source data fusion statistical analysis method based on Bayesian network according to claim 1 is characterized by: In step S3, the Bayesian network model includes the following construction steps: S31. Determine nodes and variables: Determine nodes in the Bayesian network according to different attributes and categories reflected by multi-source data, where each node corresponds to a specific multi-source variable; S32. Determine the causal relationship and directed edges between nodes: Combine the professional knowledge of power system operation, actual operation experience and the potential relationship between multi-source data to determine the causal relationship between nodes, and combine the analysis and statistics of historical data to sort out the association between different multi-source data, verify and refine the determined causal relationship; for the node pairs determined to have a causal relationship, construct directed edges in the Bayesian network model; S33, determining the prior probability distribution of the node: using the collected historical data, the prior probability distribution of each node is calculated by statistical methods, wherein for the nodes of discrete variables, the frequency of occurrence of each observation category in the total number of observations is calculated as its prior probability; for the nodes of continuous variables, the prior probability distribution is determined by using a probability density function; S34, calculating conditional probability distribution: according to Bayesian theorem, using the determined node prior probability distribution and the joint probability obtained by analyzing historical data to calculate the conditional probability distribution between nodes; S35. Complete the network model: assign the calculated conditional probability distribution to the directed edges between the corresponding nodes to complete the construction of the Bayesian network.

5. The multi-source data fusion statistical analysis method based on Bayesian network according to claim 1 is characterized by: In step S4, the time series model is configured as a SARIMA model, including the following construction steps: S41, extracting time-related features: using MK test to extract time-related features of multi-source data, the time-related features include trend features, seasonal features, periodic features and lag-related features; S42, training the SARIMA model with multi-source data as input and time-related features as output; S43. Using evaluation indicators to measure the model's prediction ability for time-related features, the evaluation indicators are accuracy, precision, recall rate and F1 value.

6. The multi-source data fusion statistical analysis method based on Bayesian network according to claim 5 is characterized by: In step S43, the evaluation index is calculated as follows: In the formula, ACC represents accuracy; PPV represents precision; TRP represents recall; F1 represents F1 value; TP represents true positive examples; TN represents true negative examples; FP represents false positive examples; and FN represents false negative examples.

7. The multi-source data fusion statistical analysis method based on Bayesian network according to claim 1 is characterized by: In step S5, the multi-source data association model includes the following construction steps: S51. Feature extraction and fusion: Extract causal probability association features between multi-source data from the constructed Bayesian network model. Specifically, it includes the causal relationship represented by the directed edges between nodes and the corresponding conditional probability distribution; extracting the time correlation features of multi-source data from the trained time series model; fusing the causal probability correlation features with the time correlation features; S52. Model structure integration: Based on the directed acyclic graph structure of the Bayesian network model, the time-related features are integrated into it, and each node in the Bayesian network model is given attributes or dynamic change rules related to the time-related features; S53, model parameter adjustment and optimization: by traversing different parameter combinations, observe the performance of the model on the test data, and find the parameter settings that make the evaluation indicators reach the optimal level or meet the preset requirements; during the parameter adjustment process, use verification data to verify the model and optimize model performance.

8. The multi-source data fusion statistical analysis method based on Bayesian network according to claim 7 is characterized by: In step S52, the model structure integration is also based on the time series structure of the time series model, and according to the causal probability association characteristics of the Bayesian network model, the data of different time periods are targetedly grouped or weighted.

9. The multi-source data fusion statistical analysis method based on Bayesian network according to claim 1 is characterized by: In step S6, the step of acquiring multi-source data of the time period to be configured, inputting the multi-source data association model, identifying multi-source data variables with abnormal data, and adjusting the abnormal data variables to optimize the power configuration includes the following steps: S61, data collection and preprocessing: for the time period to be configured, multi-source data is collected from various relevant links of the power system, and the collected multi-source data is preprocessed; S62, input model and analysis: input the preprocessed multi-source data into the constructed multi-source data association model, and output the predicted causal probability association feature and time association feature; S63, identifying abnormal data variables: according to the preset abnormal judgment standard, comparing the causal probability association feature characteristics and time association features of the multi-source data association model in theory with the prediction results, and determining the multi-source data variables with abnormalities; S64. Adjust abnormal data variables to optimize power configuration: adjust abnormal data variables and re-enter the multi-source data association model to form a closed loop until abnormal data variables no longer appear.

Images