Electrical load analysis method and device based on big data, and medium

Abstract

The invention discloses an electrical load analysis method and device based on big data, and a medium, and relates to the technical field of electric power big data analysis, and the method comprises the steps: collecting a power grid operation multi-source data set, and carrying out the preprocessing; node-level time sequence feature analysis is carried out on the preprocessed power grid operation multi-source data set, demand change features and environment association features are extracted, node management feature vectors are generated, nodes and connecting edges are created based on the organization structure and operation association relation of the power grid, and a dynamic space-time management graph is constructed; and based on the dynamic space-time management diagram, performing influence propagation path backtracking and association reasoning analysis on the abnormal management event sequence, identifying a root management event in space-time association with the abnormal management event, and generating an abnormal operation management report. According to the invention, through anomaly tracing based on the dynamic space-time management diagram, evolution from surface anomaly identification to deep root diagnosis is realized, and accurate identification of an anomaly root and optimization guidance of operation and maintenance response are achieved.

Description

Technical Field

[0001] This invention relates to the field of power big data analysis technology, and in particular to a method, equipment and medium for analyzing electricity load based on big data. Background Technology

[0002] Against the backdrop of the rapid development of smart grids and the energy internet, the field of power load analysis is facing challenges such as a surge in data volume and significant multi-source heterogeneity. Existing methods mainly rely on statistical time series analysis (such as seasonal differential autoregressive moving average models) and traditional machine learning algorithms (such as support vector regression, gradient boosting decision trees, etc.) to construct single-point or multi-point load forecasting models after cleaning, denoising, and normalizing historical load data, and provide day-ahead or short-term load forecasting curves for the power grid dispatch center. Existing methods provide data support for the basic operation of the power grid by mining the temporal patterns of the load itself.

[0003] However, existing methods have room for improvement in dealing with the fusion of high-dimensional heterogeneous data and anomaly tracing; the integration of multi-source data such as meteorological, topological and event data is mostly limited to feature splicing, and fails to establish a unified spatiotemporal framework to explore the deep coupling relationship between data, affecting the completeness of model input information; the anomaly analysis process lacks the ability to dynamically track the fault propagation path, making it difficult to locate the root cause from the spatiotemporal dimension, which restricts the accuracy of operation and maintenance decisions. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a big data-based electricity load analysis method to solve the problems of insufficient multi-source data fusion and weak anomaly tracing capabilities.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a big data-based method for analyzing electricity load, comprising: collecting and preprocessing multi-source datasets of power grid operation; performing node-level time-series feature analysis on the preprocessed multi-source datasets of power grid operation, extracting demand change features and environmental correlation features, generating node management feature vectors, and creating nodes and connecting edges based on the organizational structure and operational correlation of the power grid to construct a dynamic spatiotemporal management graph; inputting the dynamic spatiotemporal management graph into a spatiotemporal graph neural network model, and deriving the expected load demand of each node in the future time period through node feature aggregation and cross-node feature propagation calculation to generate a power grid load operation plan; collecting actual operation data at the corresponding time of execution of the power grid load operation plan, comparing and analyzing it with the power grid load operation plan, performing anomaly detection and judgment according to deviation judgment rules and deviation duration conditions, and generating an anomaly management event sequence; and performing influence propagation path backtracking and correlation reasoning analysis on the anomaly management event sequence based on the dynamic spatiotemporal management graph to identify the root management events spatiotemporally related to the anomaly management events and generate an anomaly operation management report.

[0008] As a preferred embodiment of the big data-based electricity load analysis method of the present invention, the multi-source dataset of power grid operation includes historical and real-time electricity load data, power grid topology data, meteorological data of the operating environment, and planned operation event data;

[0009] The preprocessing includes data cleaning, spatiotemporal alignment, and feature standardization.

[0010] As a preferred embodiment of the big data-based electricity load analysis method of the present invention, the steps for extracting demand change characteristics and environmental correlation characteristics to generate node management feature vectors are as follows:

[0011] Using power grid nodes as units, historical and real-time electricity load data are reassembled into a time series dataset based on node identifiers to form a node load time series dataset.

[0012] Based on the node load time series dataset, the demand change characteristics of each power grid node are extracted by the time series pattern mining algorithm to generate the node's intrinsic demand feature vector.

[0013] The node's intrinsic demand feature vector is cross-referenced and fused with the meteorological data of the operating environment and the planned operational event data to generate a node management feature vector.

[0014] As a preferred embodiment of the big data-based electricity load analysis method of the present invention, the steps for creating nodes and connecting edges based on the organizational structure and operational relationship of the power grid, and constructing a dynamic spatiotemporal management graph are as follows:

[0015] Using node management feature vectors as node attributes, and combining them with power grid topology data to map the physical connection relationships between power grid nodes, the association weights are calculated to generate a static management association graph.

[0016] In the static management association graph, the changes in the node management feature vector within a continuous time window are updated to generate a dynamic spatiotemporal management graph.

[0017] As a preferred embodiment of the big data-based electricity load analysis method of the present invention, the steps of inputting the dynamic spatiotemporal management map into the spatiotemporal graph neural network model, and calculating the expected load demand of each node in the future time period through node feature aggregation and cross-node feature propagation to generate a power grid load operation plan are as follows.

[0018] The dynamic spatiotemporal management graph is input into the spatiotemporal graph neural network model. The graph convolutional layer performs multi-order neighborhood node feature convolution aggregation and attention weight allocation on the node management feature vector along the connection edge to generate a node operation status profile.

[0019] The node operation status profile is input into the temporal network layer of the spatiotemporal graph neural network model. The dynamic evolution of the operation status of each power grid node over time is captured through a gated recurrent mechanism, and a node temporal operation encoding vector is generated.

[0020] By performing nonlinear regression calculations on the node time-series operation encoding vectors through the fully connected layer of the spatiotemporal graph neural network model, the expected load demand of each power grid node in the future period is mapped, and the power grid load operation plan is generated by integrating the results.

[0021] As a preferred embodiment of the big data-based electricity load analysis method of the present invention, the steps of collecting actual operating data at the corresponding time of the execution of the power grid load operation plan and comparing and analyzing it with the power grid load operation plan are as follows:

[0022] At each execution point in the power grid load operation plan, the actual power load data of each power grid node is collected to generate a real-time operation performance dataset.

[0023] The real-time operational performance dataset is compared with the projected load demand in the power grid load operation plan, and the absolute deviation and relative deviation rate are calculated to generate an operational performance deviation dataset.

[0024] As a preferred embodiment of the big data-based electricity load analysis method of the present invention, the steps of performing anomaly detection and judgment based on deviation judgment rules and deviation duration conditions to generate anomaly management event sequences are as follows:

[0025] Based on the deviation judgment rules, the operational performance deviation dataset is scanned to identify the primary abnormal signal set;

[0026] The persistence and stability of the primary abnormal signal set are determined by applying the deviation duration condition. Abnormal signals that meet the persistence requirement are identified as operational abnormal events and integrated to generate an abnormal management event sequence.

[0027] As a preferred embodiment of the big data-based electricity load analysis method of the present invention, the steps for generating the abnormal operation management report are as follows:

[0028] Extract the power grid node and occurrence time corresponding to each abnormal management event in the abnormal management event sequence, delineate the associated spatiotemporal impact range in the dynamic spatiotemporal management diagram, and form a set of spatiotemporal impact ranges of abnormalities;

[0029] Based on the connection edge relationships in the dynamic spatiotemporal management graph, the spatiotemporal range set of anomaly impacts is traced in reverse to identify the set of potential impact propagation paths;

[0030] By matching and analyzing the potential impact propagation path set with the dynamic spatiotemporal management diagram, the root cause with the highest spatiotemporal coupling with the abnormal management event is identified and determined to be the root cause management event.

[0031] Integrate the sequence of abnormal management events, the set of potential impact propagation paths, and root cause management events to generate an abnormal operation management report.

[0032] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the big data-based power load analysis method described in the first aspect of the present invention.

[0033] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the big data-based power load analysis method described in the first aspect of the present invention.

[0034] The beneficial effects of this invention are as follows: by constructing a dynamic spatiotemporal management diagram, a holographic depiction of the power grid's operating status is achieved, enabling big data analysis to be upgraded from isolated point analysis to networked correlation analysis, uncovering the spatiotemporal propagation law of load, and improving the completeness of the input information of the spatiotemporal diagram neural network model; through anomaly tracing based on the dynamic spatiotemporal management diagram, the evolution from surface anomaly identification to deep root cause diagnosis is realized, and accurate identification of anomaly root causes and optimized guidance for operation and maintenance response are achieved. Attached Figure Description

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of a big data-based electricity load analysis method.

[0037] Figure 2 A flowchart for constructing a dynamic spatiotemporal management graph.

[0038] Figure 3 The flowchart for processing the spatiotemporal graph neural network model.

[0039] Figure 4 This is a flowchart for anomaly tracing analysis.

[0040] Figure 5 This is a data chart comparing the Top-K hit rates of dynamic spatiotemporal management graphs and static management correlation graphs.

[0041] Figure 6 A chart comparing the event consistency improvement rate under different operating regions. Detailed Implementation

[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0043] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0044] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0045] Reference Figures 1-6 This is one embodiment of the present invention, which provides a method for analyzing electricity load based on big data, including the following steps:

[0046] S1. Collect multi-source datasets of power grid operation and perform preprocessing.

[0047] S1.1: The multi-source dataset for power grid operation includes historical and real-time electricity load data, power grid topology data, meteorological data of the operating environment, and planned operation event data;

[0048] It should be noted that historical and real-time electricity load data refers to operational data reflecting the actual electricity demand levels of each node in the power grid at different time scales. This data is used to characterize the basic state and evolution of power grid load over time, including the power consumption values, electricity consumption statistics, and corresponding sampling time information of each power grid node at different time points. Power grid topology data refers to structured data describing the physical connections and hierarchical relationships between nodes in the power grid. This data reflects the basic paths of load and management impact propagation between nodes during power grid operation, including power grid node information, connection relationships between nodes, connection attributes of lines or substations, and the node's position within the power grid organizational structure. The hierarchical relationship in the data is as follows: meteorological data refers to data reflecting the external natural environment of the power grid operating area, used to characterize the degree of influence of environmental factors on changes in power grid load demand. It is obtained through meteorological monitoring stations, meteorological service interfaces or third-party meteorological data platforms, and matched with power grid nodes according to geographical location and time dimensions, including temperature, humidity, rainfall, wind speed and solar radiation intensity; planned operation event data refers to data reflecting known or expected operation management activities and business events of the power grid within a specific time period, used to characterize the impact of non-random factors on power grid load and operating status, including equipment maintenance plans, load adjustment arrangements, demand response events and major electricity consumption activities.

[0049] S1.2: Preprocessing includes data cleaning, spatiotemporal alignment, and feature standardization.

[0050] It should be noted that data cleaning refers to identifying and removing outliers (such as instantaneous spikes caused by sensor malfunctions) that significantly exceed reasonable ranges by setting threshold ranges based on statistical distributions (such as the 3σ rule) for historical and real-time electricity load data, meteorological data of the operating environment, and planned operational event data. For missing values, a linear interpolation algorithm based on time series is used to eliminate the interference of noise and erroneous data on subsequent spatiotemporal neural network models, thereby improving the reliability of the analysis results. Spatiotemporal alignment refers to assigning a unified timestamp (such as synchronizing to the hour or minute) to all data from heterogeneous sources in the multi-source data set of power grid operation. The system uses clock and spatial labels (such as those associated with specific substations or power supply areas) to interpolate or aggregate data of different frequencies (such as minute-level load and hourly-level weather) to the same time granularity. It also matches the meteorological data of the operating environment to the corresponding power grid nodes through a gridding method, ensuring that data from different sources are comparable and fusionable in time and space. Feature standardization refers to scaling various numerical feature data (such as load values, temperature, and wind speed) in the multi-source dataset of power grid operation to the [0,1] interval through the Min-Max normalization method, eliminating the impact caused by excessive differences in the units and value ranges between different features.

[0051] S2. Perform node-level time-series feature analysis on the preprocessed multi-source dataset of power grid operation, extract demand change features and environmental correlation features, generate node management feature vectors, and create nodes and connection edges based on the organizational structure and operational correlation of the power grid to construct a dynamic spatiotemporal management graph.

[0052] S2.1: Using power grid nodes as units, historical and real-time power load data are reassembled into a time series dataset based on node identifiers to form a node load time series dataset.

[0053] Furthermore, taking power grid nodes as units, historical and real-time power load data are classified according to power grid node identifiers. All load data belonging to the same power grid node identifier are sorted in ascending order by timestamp. The continuity of timestamps is checked on the sorted load data, and missing timestamp positions are filled using linear interpolation to form a continuous and complete load value sequence arranged in chronological order under each power grid node identifier. The load value sequences corresponding to all power grid node identifiers are integrated to form a node load time series dataset.

[0054] S2.2: Based on the node load time series dataset, the demand change characteristics of each power grid node are extracted through the time series pattern mining algorithm to generate the node's intrinsic demand feature vector.

[0055] Furthermore, based on the node load time-series dataset, a seasonal trend decomposition algorithm is applied to the load value sequence of each power grid node to separate the load value sequence into trend component, periodic component, and residual component. For the trend component, a system of linear equations is constructed using the time point sequence as the independent variable and the trend component value as the dependent variable. The regression coefficients are obtained by solving the system of linear equations using the least squares method. The regression coefficients of the time variable are the slope parameters reflecting the long-term monotonically increasing or decreasing trend of the load, and their expression is:

[0056] ;

[0057] in, It is a power grid node The load trend slope parameter is used to characterize the long-term changing trend of electricity demand at power grid nodes over time. It is a power grid node timestamp Corresponding trend component values; It is a power grid node The trend component mean; It's a timestamp. The corresponding time variable; It is the mean of the time variable; It is a power grid node index.

[0058] For periodic components, a fast Fourier transform algorithm is applied to the periodic component sequence to transform the time-domain sequence to the frequency domain, obtaining the corresponding spectrum. In the spectrum, the main frequency components with energy exceeding a preset energy threshold are identified, and the complex modulus of each main frequency component is calculated as the amplitude parameter, and the argument is calculated as the phase parameter. For residual components, the standard deviation is calculated to quantify the fluctuation intensity. The slope parameter, amplitude parameter, phase parameter, and fluctuation intensity are concatenated in a fixed order to form a feature array characterizing the electricity consumption behavior of power grid nodes. After normalizing the feature array, a standardized node intrinsic demand feature vector is generated.

[0059] It should be noted that the seasonal trend decomposition algorithm is a statistical method that decomposes time series data into trend components, periodic components, and residual components. The principle is to iteratively extract long-term trends using a locally weighted regression smoothing method, and then separate seasonal fluctuations using a periodic sliding window, ultimately obtaining three components reflecting the inherent regularity of the data. In this embodiment, the seasonal trend decomposition algorithm is used to decompose the load value sequence of each power grid node, accurately separating the trend component reflecting the long-term direction of electricity consumption changes, the periodic component reflecting periodic regularity, and the residual component representing random fluctuations. This provides a quantitative indicator with clear physical meaning for constructing the node's inherent demand feature vector. The energy threshold setting process is as follows: the periodic component sequence is subjected to a Fast Fourier Transform... In the obtained spectrum, the energy values ​​of all frequency components are arranged in descending order. The first difference between adjacent energy values ​​(i.e., the steepness of the descent) is calculated, and the energy value corresponding to the first position where the difference value is less than the mean of the overall difference sequence is set as the energy threshold. An exemplary value range is 1% to 5% of the total spectrum energy. If it is less than 1% (too lenient), a large amount of background noise is easily misjudged as a valid signal, significantly increasing the dimension and noise of the node's intrinsic demand feature vector. If it is more than 5% (too strict), it is easy to miss periodic components with small energy but crucial for accurate modeling (such as semi-diurnal harmonics), resulting in an incomplete node intrinsic demand feature vector, which affects the ability of the subsequent spatiotemporal graph neural network model to capture complex load waveforms (such as bimodal curves).

[0060] S2.3: Perform feature cross-referencing and fusion of the node's intrinsic demand feature vector with the meteorological data of the operating environment and the planned operation event data to generate the node management feature vector.

[0061] Furthermore, the node's intrinsic demand feature vector is first concatenated with the temperature, humidity, and wind speed values ​​from the meteorological data of the operating environment in the column direction to form a load-meteorological hybrid feature vector. The event type fields (such as "planned maintenance" and "power supply guarantee for large-scale events") in the planned operation event data are converted into binary event feature vectors through one-hot encoding, and then concatenated with the load-meteorological hybrid feature vector in the column direction in the second round to generate a high-dimensional hybrid feature matrix. The high-dimensional hybrid feature matrix is ​​then multiplied with a learnable linear transformation matrix to obtain an intermediate feature matrix. The number of rows of the learnable linear transformation matrix is ​​equal to the preset target dimension of the node management feature vector (usually 64-256 dimensions; a higher dimension of 128-256 dimensions is preferred when the number of power grid nodes is large or the feature relationships are complex; a lower dimension of 64-128 dimensions is preferred when computational resources are limited or the amount of data is small), and the number of columns is equal to the number of columns of the high-dimensional hybrid feature matrix. The ReLU activation function is applied to each element of the intermediate feature matrix for nonlinear transformation, and the node management feature vector is output.

[0062] S2.4: Using node management feature vectors as node attributes, combine them with power grid topology data to map the physical connection relationships between power grid nodes, calculate the association weights, and generate a static management association graph.

[0063] Furthermore, using the node management feature vector as the attribute vector of each power grid node in the graph, the physical connection pairs between power grid nodes are determined one by one from the node connection relationships recorded in the power grid topology data, and each pair of power grid nodes with a physical connection relationship is mapped as a node connection edge. For each node connection edge, the node management feature vectors corresponding to the power grid nodes at both ends of the node connection edge are extracted, and the ratio between the inner product of the two vectors and the product of the Euclidean norm of the two vectors is constructed. The quotient between the inner product of the two vectors and the product of the Euclidean norm of the two vectors is used as the feature similarity, and a uniform scaling process is performed on the feature similarity. The scaled feature similarity is used as the association weight of the node connection edge. All power grid nodes and node connection edges are integrated to generate a static management association graph.

[0064] S2.5: In the static management association graph, perform association updates on the changes of node management feature vectors within a continuous time window to generate a dynamic spatiotemporal management graph.

[0065] Furthermore, in the static management correlation graph, the set of node management feature vectors corresponding to each time slice within a continuous time window is obtained. For each power grid node, the mean and variance of its node management feature vector within the time window are calculated as temporal change features. Based on the temporal change features of the power grid nodes at both ends of each pair of connecting edges, the Pearson correlation coefficient of their changing trends is calculated as the dynamic correlation strength. The correlation weight of each node connecting edge in the static management correlation graph is weighted and summed with the corresponding dynamic correlation strength to obtain the dynamic correlation weight considering spatiotemporal changes. The dynamic correlation weight is used to replace the original weight in the static management correlation graph. At the same time, the node management feature vectors of each time slice within the time window are integrated into a temporal feature tensor in chronological order to generate a dynamic spatiotemporal management graph containing spatiotemporal dynamic information.

[0066] Figure 5 This paper presents a comparison of the Top-K hit rate trends obtained from two structures—dynamic spatiotemporal management graphs and static management correlation graphs—under the same multi-source data conditions of power grid operation. The upper part is a full-range overview curve, used to show the overall hit rate changes of the two graph structures in the anomaly root cause localization task as the Top-K value gradually increases. The lower part provides a local magnification of the selected interval to present the performance differences within the selected interval. Figure 5It is evident that, within most Top-K intervals, the hit rate based on the dynamic spatiotemporal management graph is generally higher than that based on the static management association graph. This indicates that by constructing a dynamic spatiotemporal management graph and uniformly modeling the spatiotemporal associations of node management feature vectors and connecting edges between nodes, the spatiotemporal propagation law of load can be more fully depicted, thereby improving the accuracy and stability of anomaly root cause localization. This verifies the effectiveness of the dynamic spatiotemporal management graph at the structural expression level.

[0067] Figure 6 This paper compares the event consistency improvement rate of a dynamic spatiotemporal management graph constructed based on node management feature vectors and a structure formed based on node intrinsic demand feature vectors under different operating conditions. The paper statistically analyzes the event consistency improvement rate for multiple operating regions, labels the corresponding values, and also presents the overall average improvement rate differences. Figure 6 It is evident that, within each operating region, the event consistency improvement rate based on the dynamic spatiotemporal management graph is higher than that of the comparative structure, and the improvement rate remains stable under multi-region conditions. This indicates that the dynamic spatiotemporal management graph, by integrating load change characteristics and environmental correlation characteristics, strengthens the spatiotemporal correlation expression capability between nodes and effectively enhances the consistency characterization capability of event propagation logic. This verifies the structural applicability and generalization capability of the present invention in complex multi-region operating environments.

[0068] S3. Input the dynamic spatiotemporal management map into the spatiotemporal graph neural network model, and deduce the expected load demand of each node in the future period through node feature aggregation and cross-node feature propagation calculation, and generate the power grid load operation plan.

[0069] It should be noted that the spatiotemporal graph neural network model is pre-trained as follows: Historical power grid operation multi-source datasets with labeled historical load values ​​are collected as training data. Each sample consists of a dynamic spatiotemporal management map and the corresponding actual load value. The samples are randomly divided into training and validation sets according to a fixed ratio (e.g., 7:3). During the training phase, the dynamic spatiotemporal management map from the training set is input into the spatiotemporal graph neural network model, and processed sequentially through graph convolutional layers, temporal network layers, and fully connected layers. The fully connected layer outputs a continuous scalar value as the predicted load demand, which is compared with the corresponding actual load value. The result is calculated using the mean squared error loss function. Calculate the difference between the two to obtain the loss value; based on the loss value, calculate the gradient of the parameters of each layer of the spatiotemporal graph neural network model through the backpropagation algorithm, and use the stochastic gradient descent algorithm (with an example learning rate of 0.001) to update all parameters of the spatiotemporal graph neural network model according to the gradient; repeat the above process for multiple rounds on the entire training set, and evaluate the performance of the current spatiotemporal graph neural network model on the validation set after each training round, calculate the average loss on the validation set, and stop training when the average loss on the validation set has not decreased in multiple consecutive training rounds (such as 10 rounds), and obtain the trained spatiotemporal graph neural network model.

[0070] S3.1: Input the dynamic spatiotemporal management graph into the spatiotemporal graph neural network model, and perform multi-level neighborhood node feature convolution aggregation and attention weight allocation on the node management feature vector along the connection edge through graph convolutional layer to generate a node operation status profile.

[0071] Furthermore, in the graph convolutional layer of the spatiotemporal graph neural network model, for each central node in the dynamic spatiotemporal management graph, its multi-order neighbor node set is identified through adjacency matrix exponentiation; linear transformations are performed on the node management feature vectors of the central node and its various order neighbor nodes; unnormalized attention scores are generated by calculating the dot product of the transformed feature vector of the central node and the transformed feature vectors of each neighbor node; the unnormalized attention scores are input into the softmax function for normalization to obtain the attention weights of each neighbor node relative to the central node; the transformed feature vectors of the neighbor nodes are weighted and summed according to their corresponding attention weights, and combined with the transformed feature vector of the central node to generate a node operation status profile.

[0072] It should be noted that adjacency matrix exponentiation refers to the mathematical operation of multiplying the adjacency matrix, which describes the connection relationship between nodes in a graph structure, multiple times. When calculating the kth power of the adjacency matrix, the position of each non-zero element in the resulting matrix indicates that there is a connected path of length k between corresponding two nodes in the original graph structure. In dynamic spatiotemporal management graphs, by calculating different powers of the adjacency matrix (such as first power and second power), the set of all reachable neighboring nodes of each central node within different topological distances (i.e., multi-order neighborhoods) can be identified.

[0073] S3.2: Input the node operation status profile into the temporal network layer of the spatiotemporal graph neural network model, and capture the dynamic evolution of the operation status of each power grid node over time through a gated loop mechanism to generate node temporal operation encoding vectors.

[0074] Furthermore, the node operation status profiles are organized into a sequence according to time steps and input into the temporal network layer of the spatiotemporal graph neural network model. The gated recurrent mechanism in the temporal network layer receives the corresponding node operation status profile at each time step. It generates update gate and reset gate gating signals by linearly transforming the node operation status profile of the current time step with the hidden state of the previous time step and inputting it into the sigmoid function. The reset gate gating signal is used to modulate the hidden state of the previous time step and fuse it with the current input. Then, a candidate hidden state is generated by using the tanh activation function. The update gate gating signal is used to perform a weighted sum of the hidden state of the previous time step and the current candidate hidden state to obtain the hidden state of the current time step. The above process is iterated in time step order, and the hidden state of the last time step is output as the node temporal operation encoding vector.

[0075] S3.3: By performing nonlinear regression calculation on the node time-series operation encoding vector through the fully connected layer of the spatiotemporal graph neural network model, the expected load demand of each power grid node in the future period is mapped out, and the power grid load operation plan is generated by integrating the results.

[0076] Furthermore, the node time-series operation coding vector is input into the fully connected layer of the spatiotemporal graph neural network model. Linear transformation and nonlinear activation function operations are performed on the node time-series operation coding vector in sequence, so that the node time-series operation coding vector is mapped to the preliminary prediction vector of the load value for each future time period. The preliminary prediction vector is de-standardized and mapped back to the original load dimension to obtain the specific expected load demand value of each power grid node at each time point in the future time period. According to the time order and node affiliation, all expected load demand values ​​are organized into a triplet sequence containing timestamp, node identifier and load value to generate the power grid load operation plan.

[0077] S4. At the corresponding time of the execution of the power grid load operation plan, collect actual operating data and compare and analyze it with the power grid load operation plan. Perform anomaly detection and judgment according to the deviation judgment rules and deviation duration conditions, and generate anomaly management event sequence.

[0078] S4.1: At each execution point in the power grid load operation plan, collect the actual power load data of each power grid node and generate a real-time operation performance dataset.

[0079] Furthermore, the timestamp and node identifier set corresponding to each execution time point are read from the power grid load operation plan. The active power, reactive power and current and voltage measurements of each timestamp are collected synchronously by the measurement devices deployed at each power grid node. The data are matched and aligned according to the power grid node identifier and timestamp, and invalid data records caused by communication interruption are removed. The valid data records are arranged in ascending order of timestamp and associated with the corresponding power grid node identifier to form a real-time operation performance dataset.

[0080] S4.2: Compare the real-time operational performance dataset with the projected load demand in the power grid load operation plan, calculate the absolute deviation and relative deviation rate, and generate an operational performance deviation dataset.

[0081] Furthermore, based on the real-time operational performance dataset, and according to the grid node identifier and execution time point, the matching expected load demand value is found in the grid load operation plan; the difference between the active power and the expected load demand value is calculated as the absolute deviation; the ratio of the absolute deviation to the expected load demand value is calculated and converted into a percentage as the relative deviation rate; the absolute deviation and relative deviation rate obtained by each grid node at each execution time point are associated and summarized with the corresponding node identifier and timestamp to form an operational performance deviation dataset.

[0082] S4.3: Based on the deviation judgment rules, scan the operational performance deviation dataset to identify the primary abnormal signal set.

[0083] Furthermore, based on the deviation judgment rules, each data record in the operational performance deviation dataset is sequentially matched with conditions. When the absolute deviation value of a data record exceeds the preset absolute deviation threshold or the relative deviation rate value exceeds the preset relative deviation rate threshold, the data record is marked as abnormal. All data records marked as abnormal are aggregated according to the power grid node identifier and the occurrence time to generate a primary abnormal signal set containing the abnormal node, abnormal timestamp, actual deviation value (i.e., absolute deviation value), and the type of abnormal condition triggered (such as absolute deviation exceeding the limit, relative deviation rate exceeding the limit, and combined exceeding the limit).

[0084] It should be noted that the deviation judgment rule is a set of preset logical conditions used to determine whether operational performance deviation data constitutes an anomaly. This includes an absolute deviation threshold set for absolute deviation values, a relative deviation rate threshold set for relative deviation rate values, and a composite judgment logic that determines an anomaly when either the absolute deviation or relative deviation rate exceeds its corresponding threshold. The absolute deviation threshold is set by analyzing the statistical distribution characteristics of historical load data, using the mean of the historical absolute deviation sequence combined with twice the standard deviation as the threshold boundary. An exemplary value range is 50kW to 500kW. Values ​​above 500kW will ignore many meaningful operational anomalies, leading to missed reports. Values ​​below 50kW will misjudge normal random fluctuations as anomalies, resulting in frequent false alarms. The relative deviation rate threshold is set based on the allowable fluctuation range of load forecasting errors, using the 90% to 98% quantile of the historical relative deviation rate sequence as the threshold boundary. An exemplary value range is 15% to 30%. Values ​​above 30% will tolerate excessive forecasting deviations, losing monitoring significance. Values ​​below 15% will generate numerous invalid alarms for minor fluctuations with no substantial impact.

[0085] S4.4: Apply deviation duration conditions to the primary abnormal signal set to determine its persistence and stability. Abnormal signals that meet the persistence requirements are identified as operational abnormal events and integrated to generate an abnormal management event sequence.

[0086] Furthermore, the primary abnormal signal set is grouped according to the grid node identifier and abnormal condition type, and the abnormal signals within each group are sorted by timestamp; abnormal signals with consecutive timestamps in the sorted abnormal signal sequence are detected, and consecutive abnormal signals (i.e., the time interval between adjacent abnormal signals does not exceed 5 minutes) are grouped into the same consecutive abnormal signal segment; the difference between the start and end timestamps of each consecutive abnormal signal segment is calculated to obtain the signal segment duration; the signal segment duration is compared with the deviation duration condition, and consecutive abnormal signal segments with a signal segment duration greater than or equal to the deviation duration condition are selected and marked as operational abnormal events; all operational abnormal events are sorted by start time to generate an abnormal management event sequence.

[0087] It should be noted that the deviation duration condition is set based on the statistical distribution of the duration of historical abnormal events. An exemplary value range is 10 to 30 minutes. If it is higher than 30 minutes, it will lead to the omission of short-term abnormalities (such as short-term fluctuations caused by equipment switching). If it is lower than 10 minutes, it will misjudge the instantaneous jitter caused by normal equipment start-up and shutdown or measurement noise as an operational abnormal event, resulting in a significant increase in the false alarm rate.

[0088] S5. Based on the dynamic spatiotemporal management diagram, perform impact propagation path backtracking and correlation reasoning analysis on the sequence of abnormal management events, identify the root management events that are spatiotemporally related to the abnormal management events, and generate an abnormal operation management report.

[0089] S5.1: Extract the power grid node and occurrence time corresponding to each abnormal management event in the abnormal management event sequence, delineate the associated spatiotemporal impact range in the dynamic spatiotemporal management diagram, and form a set of spatiotemporal ranges of abnormal impact.

[0090] Furthermore, the power grid node and occurrence time of each abnormal management event in the abnormal management event sequence are extracted. The corresponding power grid node is located in the dynamic spatiotemporal management graph. Taking this power grid node as the center, all associated power grid nodes within a specified number of hops (e.g., 2 hops) are traversed outward based on the connection edge relationship in the dynamic spatiotemporal management graph to form a spatial influence range. At the same time, a fixed time window (e.g., 15 minutes before and after) is extended forward and backward based on the abnormal occurrence time to form a temporal influence range. The spatial influence range and the temporal influence range are combined to form the spatiotemporal influence range associated with the abnormal management event. The above process is repeated for all abnormal management events, and the set of spatiotemporal influence ranges formed is the set of spatiotemporal influence ranges of the abnormality.

[0091] S5.2: Based on the connection edge relationship in the dynamic spatiotemporal management graph, reverse tracing is performed on the spatiotemporal range set of anomaly impacts to identify the set of potential impact propagation paths.

[0092] Furthermore, based on the connection edge relationships in the dynamic spatiotemporal management graph, starting from the power grid node corresponding to each abnormal event in the spatiotemporal range of the abnormal impact, the propagation direction is determined according to the time sequence of whether the slope of the node management feature vector changes exceeds a preset slope threshold. Specifically, for the power grid nodes at both ends of each node connection edge in the dynamic spatiotemporal management graph, the change slope sequence of the node management feature vector corresponding to each time slice within the time window of the abnormal occurrence is calculated. The timestamp of the first time the slope of the node management feature vector changes exceeds the slope threshold is determined. The power grid node with the earlier timestamp of the first time exceeding the slope threshold is defined as the upstream node, and the timetamp of the first time exceeding the slope threshold is defined as the upstream node. The power grid node with the later instability is defined as the downstream node, and the node connection edges are assigned directed attributes accordingly. Starting from the power grid node corresponding to the abnormal management event, a breadth-first search is performed along the upstream direction of the power grid node connection edges. During the search, the search is expanded hop by hop in units of nodes, and the complete path from the current abnormal node to each upstream node is recorded. The order of appearance and connection relationship of the nodes in the path are completely saved. When the search range exceeds the spatial boundary defined by the set of spatiotemporal range of abnormal impact, the expansion of the current path is terminated. The obtained paths are deduplicated, duplicate paths are merged and loops are eliminated to form a set of potential impact propagation paths containing all possible propagation directions.

[0093] It should be noted that the slope threshold is set by extracting the slope of the node management feature vector change of each power grid node within the time period without anomaly markers, constructing a historical slope sample set, and calculating the mean and standard deviation of the historical slope sample set. An exemplary value range is the range of values ​​corresponding to the mean plus one to three times the standard deviation.

[0094] It should be noted that the specific method for eliminating loops is as follows: During the path recording process of reverse breadth-first search, a set of path node access records is maintained synchronously. When expanding from the current node to a downstream node, it is checked whether the downstream node already exists in the node sequence of the current path. If it already exists, it indicates that continuing to expand will form a loop. At this time, the expansion of the current branch of the current path is immediately terminated. Reverse breadth-first search is an algorithm that traverses the graph in reverse from the target node. Starting from the abnormal node specified in the set of abnormal impact spatiotemporal range, it traverses the upstream adjacent nodes layer by layer along the reverse direction of the connecting edge in the dynamic spatiotemporal management graph. During the traversal, the order of the nodes to be visited is maintained through a queue structure, and the shortest path distance and path sequence from each node to the starting point are recorded.

[0095] S5.3: Match and perform causal relationship analysis on the set of potential impact propagation paths with the dynamic spatiotemporal management diagram to identify the root cause with the highest spatiotemporal coupling degree with the abnormal management event and determine it as the root cause management event.

[0096] Furthermore, for each path node sequence in the potential impact propagation path set, the node management feature vector of the corresponding path node in the dynamic spatiotemporal management diagram is extracted within a specific time window before and after the occurrence of the abnormal event (e.g., 30 minutes before the occurrence of the abnormal event to 15 minutes after the occurrence of the abnormal event), and the slope of the change of the node management feature vector over time is calculated to form a path node feature change slope sequence; simultaneously, the slope of the change of the node management feature vector of the power grid node corresponding to the abnormal event is extracted within the same time period to form an abnormal event feature change slope sequence; the Pearson correlation coefficient between the path node feature change slope sequence and the abnormal event feature change slope sequence is calculated, and the expression is:

[0097] ;

[0098] in, It is the Pearson correlation coefficient, which represents the degree of linear correlation between the slope sequence of path node feature changes and the slope sequence of abnormal event feature changes; It is the first of the path node feature change slope sequence Each data point corresponds to the slope of the characteristic change of a certain path node in the dynamic spatiotemporal management diagram at a specific time point; It is the first of the abnormal event characteristic change slope sequence Each data point corresponds to the characteristic change slope of the power grid node corresponding to the abnormal management event at the same time point; It is the mean of the slope sequence of path node feature changes; It is the mean of the slope sequence of the abnormal event characteristics; It represents the number of data points, corresponding to the total number of sampling points within the time window before and after the occurrence of the abnormal event; It is a data point index.

[0099] Retrieve planned operation event data associated with the dynamic spatiotemporal management graph and check whether there are planned event records for path nodes before and after abnormal times; select the path node with the highest spatiotemporal correlation coefficient and the existence of planned event records as the root cause with the highest spatiotemporal coupling with the abnormal management event, and determine it as the root cause management event.

[0100] S5.4: Integrate the sequence of abnormal management events, the set of potential impact propagation paths, and root cause management events to generate an abnormal operation management report.

[0101] Furthermore, the power grid node identifier, occurrence time, deviation value, and anomaly type of each anomaly management event in the anomaly management event sequence are integrated to form an anomaly event summary; the propagation path node sequence and propagation direction corresponding to the potential impact propagation path set are associated to form an anomaly propagation analysis; the event type, associated nodes, and judgment criteria of the root cause management event are associated to form an anomaly root cause location; the anomaly event summary, anomaly propagation analysis, and anomaly root cause location are structured and organized, and the report generation time and unique identifier are added to output an anomaly operation management report.

[0102] This embodiment also provides a computer device applicable to the electricity load analysis method based on big data, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the electricity load analysis method based on big data as proposed in the above embodiment.

[0103] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0104] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the big data-based power load analysis method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0105] In summary, this invention achieves a holographic depiction of the power grid's operating status by constructing a dynamic spatiotemporal management graph, upgrading big data analysis from isolated point analysis to networked correlation analysis, uncovering the spatiotemporal propagation patterns of load, and improving the completeness of the input information of the spatiotemporal graph neural network model; through anomaly tracing based on the dynamic spatiotemporal management graph, it realizes the evolution from surface anomaly identification to deep root cause diagnosis, and achieves accurate identification of anomaly root causes and optimized guidance for operation and maintenance responses.

[0106] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for analyzing electricity load based on big data, characterized in that: include, Collect and preprocess multi-source datasets of power grid operation data. Node-level time-series feature analysis is performed on the preprocessed multi-source dataset of power grid operation to extract demand change features and environmental correlation features, generate node management feature vectors, and create nodes and connection edges based on the organizational structure and operational correlation of the power grid to construct a dynamic spatiotemporal management graph. The dynamic spatiotemporal management map is input into the spatiotemporal graph neural network model. Through node feature aggregation and cross-node feature propagation calculation, the expected load demand of each node in the future period is deduced, and a power grid load operation plan is generated. At the corresponding time of the execution of the power grid load operation plan, actual operation data is collected and compared with the power grid load operation plan. Anomaly detection and judgment are performed according to the deviation judgment rules and deviation duration conditions, and an anomaly management event sequence is generated. Based on the dynamic spatiotemporal management diagram, the impact propagation path of the abnormal management event sequence is traced back and the correlation reasoning analysis is performed to identify the root management events that are spatiotemporally related to the abnormal management events and generate an abnormal operation management report.

2. The electricity load analysis method based on big data as described in claim 1, characterized in that: The multi-source dataset for power grid operation includes historical and real-time electricity load data, power grid topology data, meteorological data of the operating environment, and planned operation event data; The preprocessing includes data cleaning, spatiotemporal alignment, and feature standardization.

3. The electricity load analysis method based on big data as described in claim 1, characterized in that: The steps for extracting demand change features and environmental correlation features to generate node management feature vectors are as follows. Using power grid nodes as units, historical and real-time electricity load data are reassembled into a time series dataset based on node identifiers to form a node load time series dataset. Based on the node load time series dataset, the demand change characteristics of each power grid node are extracted by the time series pattern mining algorithm to generate the node's intrinsic demand feature vector. The node's intrinsic demand feature vector is cross-referenced and fused with the meteorological data of the operating environment and the planned operational event data to generate a node management feature vector.

4. The electricity load analysis method based on big data as described in claim 3, characterized in that: The steps for creating nodes and connecting edges based on the organizational structure and operational relationships of the power grid, and constructing a dynamic spatiotemporal management graph, are as follows: Using node management feature vectors as node attributes, and combining them with power grid topology data to map the physical connection relationships between power grid nodes, the association weights are calculated to generate a static management association graph. In the static management association graph, the changes in the node management feature vector within a continuous time window are updated to generate a dynamic spatiotemporal management graph.

5. The electricity load analysis method based on big data as described in claim 1, characterized in that: The process of inputting the dynamic spatiotemporal management map into the spatiotemporal graph neural network model, and calculating through node feature aggregation and cross-node feature propagation to deduce the expected load demand of each node in future time periods, thereby generating a power grid load operation plan, is as follows: The dynamic spatiotemporal management graph is input into the spatiotemporal graph neural network model. The graph convolutional layer performs multi-order neighborhood node feature convolution aggregation and attention weight allocation on the node management feature vector along the connection edge to generate a node operation status profile. The node operation status profile is input into the temporal network layer of the spatiotemporal graph neural network model. The dynamic evolution of the operation status of each power grid node over time is captured through a gated recurrent mechanism, and a node temporal operation encoding vector is generated. By performing nonlinear regression calculations on the node time-series operation encoding vectors through the fully connected layer of the spatiotemporal graph neural network model, the expected load demand of each power grid node in the future period is mapped, and the power grid load operation plan is generated by integrating the results.

6. The electricity load analysis method based on big data as described in claim 1, characterized in that: The steps for collecting actual operating data at the corresponding time points during the execution of the power grid load operation plan and comparing and analyzing it with the power grid load operation plan are as follows: At each execution point in the power grid load operation plan, the actual power load data of each power grid node is collected to generate a real-time operation performance dataset. The real-time operational performance dataset is compared with the projected load demand in the power grid load operation plan, and the absolute deviation and relative deviation rate are calculated to generate an operational performance deviation dataset.

7. The electricity load analysis method based on big data as described in claim 6, characterized in that: The steps for performing anomaly detection and judgment based on deviation judgment rules and deviation duration conditions, and generating anomaly management event sequences are as follows. Based on the deviation judgment rules, the operational performance deviation dataset is scanned to identify the primary abnormal signal set; The persistence and stability of the primary abnormal signal set are determined by applying the deviation duration condition. Abnormal signals that meet the persistence requirement are identified as operational abnormal events and integrated to generate an abnormal management event sequence.

8. The electricity load analysis method based on big data as described in claim 1, characterized in that: The steps for generating an abnormal operation management report are as follows: Extract the power grid node and occurrence time corresponding to each abnormal management event in the abnormal management event sequence, delineate the associated spatiotemporal impact range in the dynamic spatiotemporal management diagram, and form a set of spatiotemporal impact ranges of abnormalities; Based on the connection edge relationships in the dynamic spatiotemporal management graph, the spatiotemporal range set of anomaly impacts is traced in reverse to identify the set of potential impact propagation paths; By matching and analyzing the potential impact propagation path set with the dynamic spatiotemporal management diagram, the root cause with the highest spatiotemporal coupling with the abnormal management event is identified and determined to be the root cause management event. Integrate the sequence of abnormal management events, the set of potential impact propagation paths, and root cause management events to generate an abnormal operation management report.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the power load analysis method based on big data as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the big data-based power load analysis method as described in any one of claims 1 to 8.

Images