A power consumption trend perception method based on power consumption behavior feature correlation analysis

Abstract

The application discloses a power consumption trend perception method based on power consumption behavior characteristic correlation analysis and belongs to the technical field of power consumption analysis. The method comprises the following steps: collecting total incoming line high-frequency current and voltage time sequence data, detecting and segmenting power consumption equipment operation event fragments and assigning type labels; counting event type adjacent appearance frequencies to generate a transition probability matrix, and constructing a directed weighted correlation network with events as nodes and probabilities as weights; calculating average clustering coefficients and structure hole constraint coefficients in a sliding window to form a structure characteristic sequence; monitoring characteristic mutations, and determining behavior mode variation when the clustering coefficient drops and the constraint coefficient rises simultaneously and exceeds a threshold value; and determining a power consumption trend inflection point and direction according to a variation time and a transition probability change gradient. The application realizes automatic capture and trend prediction of power consumption behavior mode variation.

Description

Technical Field

[0001] This invention relates to the field of electricity consumption analysis technology, specifically a method for sensing electricity consumption trends based on correlation analysis of electricity consumption behavior characteristics. Background Technology

[0002] Electricity consumption trend perception is a crucial foundation for electricity demand-side management and energy efficiency analysis. Existing methods for analyzing electricity consumption trends typically rely on statistical modeling of time series data of total power or electricity consumption. For example, they may use moving averages, exponential smoothing, or autoregressive models to fit historical data and detect trend reversals through residual changes. These methods focus on extracting trend signals from macro-level load data but fail to delve into the structured relationships within electricity consumption behavior. When users experience slow and non-transient adjustments to the combination and sequence of their electrical equipment usage, total data often fails to reflect trend changes early on, resulting in a lag in trend perception.

[0003] Another common approach is to acquire start-up and shutdown events of each electrical device based on device-level monitoring or non-intrusive load monitoring, statistically analyze event frequency or runtime, and infer electricity consumption trends based on changes in event frequency. While this method incorporates device operation information, when dealing with the temporal dependencies between multiple device operations, it typically focuses only on the number of state transitions of a single device, ignoring the transition patterns and changes in correlation strength between different device operation events. The fundamental changes in user electricity consumption behavior often manifest first in the reorganization of the correlation structure between device operations, rather than simply an increase or decrease in event frequency. Therefore, existing solutions lack the sensitivity to identify variations in electricity consumption behavior patterns and struggle to capture early signs of trend inflection points.

[0004] How to characterize the mutual transfer relationships between device operation events in electricity consumption trend sensing, and detect the inherent variations in electricity consumption behavior patterns from the perspective of the stability and compactness of this transfer structure, is a problem that needs to be solved. Furthermore, after detecting behavioral pattern variations, how to comprehensively utilize the dynamic evolution direction of the transfer probability after the variation moment to determine the inflection point nature and direction of electricity consumption trends is another problem that needs to be solved. Summary of the Invention

[0005] This paper presents a method for sensing electricity consumption trends based on the correlation analysis of electricity consumption behavior characteristics. By constructing a correlation network of transition probabilities between equipment operation events, the method captures the variation of the internal structure of electricity consumption behavior, and determines the inflection point and trend direction of electricity consumption trend based on the gradient of the change in transition probability after the variation, so as to enhance the early indication role of the subtle structural changes in electricity consumption behavior patterns on trends.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a method for sensing electricity consumption trends based on correlation analysis of electricity consumption behavior characteristics, comprising: High-frequency sampling current and voltage data of the main incoming line within the power consumption area are collected to form a time-series dataset. Preferably, the instantaneous voltage and current values ​​of the main incoming line are collected synchronously at a rate of no less than 256 points per cycle. The sequences of instantaneous voltage and current values ​​are subjected to digital low-pass filtering to remove high-frequency noise above the Nyquist frequency. The filtered current sequence is then phase-delay compensated to ensure that its fundamental phase is consistent with the fundamental voltage phase. The compensated voltage and current sequences are then stored in time-stamp alignment to form a time-series dataset.

[0007] This invention detects and segments multiple operational event segments of electrical equipment from a time-series dataset, and assigns an event type label to each segment. As a technical solution, the time-series dataset is subjected to differential operations to extract a sequence of active power abrupt changes. Voltage and current waveform segments of preset cycles before and after each abrupt change are extracted. Fast Fourier Transform is performed on the waveform segments to extract harmonic amplitude spectra, generating a load imprint vector. The load imprint vector is then compared with template vectors in a pre-stored equipment load imprint template library using cosine similarity. Abrupt changes with a maximum cosine similarity greater than a preset threshold are marked as electrical equipment start-up or stop events. The event type is determined based on the matched template vector, thereby segmenting the corresponding operational event segments of the electrical equipment and assigning event type labels. Preferably, the load imprint vector is composed of the amplitudes of the fundamental wave and the 2nd to 15th harmonics, normalized using the L2 norm. When calculating cosine similarity, if multiple template vectors have cosine similarities greater than the preset threshold, the event type corresponding to the template vector with the highest cosine similarity is selected as the matching result.

[0008] Based on the chronological order of events, the conditional frequency of adjacent occurrences of different event types is counted to generate a transition probability matrix between event types. The event sequence composed of event type labels is traversed. For the current event type as a predecessor event type, the occurrence frequency of its direct successor event type is counted to form an adjacency frequency table. The adjacency frequency from the predecessor event type to a certain successor event type is divided by the total occurrence frequency of the predecessor event type in the event sequence to obtain the transition probability in that direction. The transition probabilities between all event type pairs are arranged with the predecessor event type as the row and the successor event type as the column to construct the event type transition probability matrix.

[0009] The transition probability matrix between event types is constructed as a directed weighted network of electricity consumption behavior, with event types as nodes and transition probabilities as edge weights. A threshold for association strength is set, and each transition probability value in the event type transition probability matrix is ​​traversed. For transition probability values ​​greater than the threshold, directed edges are established between the originating event type node and the arriving event type node, and the transition probability value is assigned as the edge weight. Edges corresponding to transition probability values ​​lower than the threshold, as well as self-transition edges of the nodes themselves, are discarded, resulting in the electricity consumption behavior association network.

[0010] The average clustering coefficient and structural hole constraint coefficient of the electricity consumption behavior association network are calculated within a sliding time window to form a structural feature sequence. Using a preset time window duration and sliding step size, subsequences within each window are extracted from the complete event type label sequence. For each subsequence within a window, the transition probabilities between event types are recalculated, and the electricity consumption behavior association network is constructed to obtain the window association network. For the window association network, the clustering coefficient of each node is calculated as the ratio of the actual number of connected edges to the possible number of connected edges between its neighbors. The average clustering coefficient is then calculated as the average of all node clustering coefficients. For any node in the window association network, the proportion of its connection weight with each neighbor node to the sum of all outgoing edge connection weights of that node is calculated. This proportion is squared, and the squared values ​​of all neighbor nodes are summed to obtain the structural hole constraint coefficient of that node. The average of all node structural hole constraint coefficients is then calculated as the structural hole constraint coefficient. The average clustering coefficient and structural hole constraint coefficient obtained from each window are arranged along the time axis to generate the structural feature sequence.

[0011] The system monitors abrupt changes in the structural feature sequence. When the decrease in the average clustering coefficient and the increase in the structural hole constraint coefficient both exceed dynamic thresholds, a change in electricity consumption behavior pattern is identified. For the average clustering coefficient sequence, the exponentially weighted moving average and exponentially weighted standard deviation are calculated for each time step within the preceding window. The lower dynamic threshold is obtained by subtracting a first multiple of the exponentially weighted standard deviation from the exponentially weighted moving average. For the structural hole constraint coefficient sequence, the exponentially weighted moving average and exponentially weighted standard deviation are calculated for each time step within the preceding window. The upper dynamic threshold is obtained by adding a second multiple of the exponentially weighted moving average to the exponentially weighted moving average. The system continuously monitors whether the average clustering coefficient at the current time is less than the lower dynamic threshold and whether the structural hole constraint coefficient at the current time is greater than the upper dynamic threshold. When both conditions are met, and the magnitude of the decrease in the average clustering coefficient below the lower dynamic threshold exceeds a preset minimum decrease, and the magnitude of the increase in the structural hole constraint coefficient above the upper dynamic threshold exceeds a preset minimum increase, the current time step is marked as an abrupt change, triggering the determination of a change in electricity consumption behavior pattern.

[0012] Based on the occurrence time of the electricity consumption behavior pattern variation and the gradient of subsequent transition probabilities, the inflection point and direction of the electricity consumption trend are determined. Starting from the occurrence time of the electricity consumption behavior pattern variation, the transition probability matrix between event types corresponding to multiple consecutive sliding time windows is extracted. For the transition probability matrix of each subsequent window, all matrix elements are summed to obtain the total transition activity. The gradient sequence of the total transition activity over time is calculated. If the gradient sequence remains positive within a preset observation interval and its absolute value gradually increases, the electricity consumption trend is determined to have an upward inflection point at the occurrence time. If the gradient sequence remains negative within the preset observation interval and its absolute value gradually increases, the electricity consumption trend is determined to have a downward inflection point at the occurrence time. If the fluctuation range of the gradient sequence is less than a set weak threshold, the electricity consumption trend is determined to have continued without an inflection point.

[0013] This invention analyzes high-frequency sampling data from the main power supply line, segments equipment operation events, and generates load imprint vectors. It then constructs a directed weighted electricity consumption behavior correlation network with event types as nodes. A sliding time window is used to dynamically track the network's average clustering coefficient and structural hole constraint coefficient, enabling sensitive detection of variations in electricity consumption behavior patterns. When the average clustering coefficient significantly decreases and the structural hole constraint coefficient significantly increases, it indicates a break in the normal correlation behavior or a change in the dependency relationship between electrical equipment, thus accurately capturing the moment of electricity consumption behavior pattern variation. Based on pattern variation, by calculating the total transition activity and its gradient within the transition probability matrix in subsequent windows, inflection points and directions of electricity consumption trends can be identified early, providing effective basis for electricity management, anomaly warning, and load forecasting. This method does not rely on a single threshold or single indicator, but integrates the multidimensional features of the temporal correlation structure of electricity consumption behavior, enhancing the ability to perceive subtle pattern drifts and abrupt trends, and exhibiting high robustness and applicability.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By statistically analyzing the sequential adjacency relationships of device operation events into a transition probability matrix between event types, and constructing a directed weighted network of electricity consumption behavior associations with event types as nodes and transition probabilities as edge weights, the strength of transition coupling and the overall association structure among multiple device operations can be characterized from a network topology perspective. Based on this, the average clustering coefficient and structural hole constraint coefficient of the network are continuously calculated within a sliding time window. The average clustering coefficient reflects the degree of clustering of transition paths around event type nodes and the tightness of local associations, while the structural hole constraint coefficient reflects the degree to which event transition paths are controlled by a few key nodes and the degree of structural redundancy. When the electricity consumption behavior pattern undergoes inherent variation, the original clustering pattern of event transitions is broken, and the connections within the association network tend to disperse or reorganize, leading to a significant decrease in the average clustering coefficient. Simultaneously, operation transition paths gradually become dependent on certain specific nodes, and transition selection becomes centralized, resulting in a significant increase in the structural hole constraint coefficient. By simultaneously monitoring changes in these two indicators, we can capture the typical structural changes in the electricity consumption behavior network as it transitions from a tightly interconnected, multi-pathway state to a loosely constrained state. This change is more sensitive to variations in electricity consumption behavior patterns than simply statistical changes in the number of events or total power, and can reveal the reorganization of behavior patterns before significant changes occur in total electricity consumption data.

[0015] After detecting the occurrence time of the variation in electricity consumption behavior patterns using the feature sequence of the associated network structure, the transition probability matrix corresponding to multiple consecutive sliding time windows after the variation time is extracted. The total transition activity is obtained by summing the matrix elements of each window, and then the gradient sequence of the total transition activity over time is calculated. This gradient reflects the dynamic development direction of the overall activity of mutual transfer between power equipment operation events after the variation occurs. If the gradient sequence remains positive and the absolute value gradually increases within the preset observation interval, it indicates that the transfer between different types of events has entered an enhanced cycle driven by the potential energy released by the variation in the previous stage, and the electricity consumption activity tends to be more intensive and efficient. Thus, it is determined that the electricity consumption trend has reached an upward inflection point at the time of the variation. If the gradient sequence remains negative and the absolute value gradually increases, it indicates that the transfer activity between events continues to decline, the correlation of electricity consumption behavior tends to weaken, and the electricity consumption trend has reached a downward inflection point. By using the gradient of the transition probability change after mutation to determine the trend direction and inflection point nature, we can avoid making misjudgments based solely on the instantaneous fluctuations of a single indicator at the time of mutation. By extracting trend criteria from the evolutionary dynamics of the behavioral association network after mutation, we can obtain more robust trend direction interpretation results. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of a method for sensing electricity consumption trends based on the correlation analysis of electricity consumption behavior characteristics; Figure 2 This is a flowchart of the process for constructing the transition probability matrix between event types; Figure 3 It is a curve showing the change of network structure characteristics related to electricity consumption behavior over time. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 This invention provides a method for sensing electricity consumption trends based on the correlation analysis of electricity consumption behavior characteristics, comprising: collecting high-frequency sampling current and voltage data of the total incoming lines within an electricity consumption area to form a time-series dataset; detecting and segmenting multiple equipment operation event segments from the time-series dataset, and assigning event type labels to each equipment operation event segment; based on the time sequence of events, statistically analyzing the frequency of adjacent occurrences of different event types to generate a transition probability matrix between event types; constructing the transition probability matrix between event types into a directed weighted electricity consumption behavior correlation network with event types as nodes and transition probabilities as edge weights; calculating the average clustering coefficient and structural hole constraint coefficient of the electricity consumption behavior correlation network within a sliding time window to form a structural feature sequence; monitoring the abrupt change in the structural feature sequence, and determining an electricity consumption behavior pattern mutation when the decrease in the average clustering coefficient and the increase in the structural hole constraint coefficient simultaneously exceed a dynamic threshold; and determining the inflection point and trend direction of the electricity consumption trend based on the occurrence time of the electricity consumption behavior pattern mutation and the subsequent gradient of the transition probability.

[0020] Example 1 In practice, high-frequency sampling current and voltage data of the main incoming line within the power consumption area are collected to form a time-series dataset. The instantaneous voltage and current values ​​of the main incoming line are synchronously collected at a rate of no less than 256 points per cycle. For a power system with a power frequency of 50 Hz, a sampling rate of 256 points per cycle corresponds to a sampling frequency of 12.8 kHz. The synchronous collection of instantaneous voltage and current values ​​is achieved by triggering an analog-to-digital converter using the same sampling clock, ensuring that the instantaneous voltage and current value sequences are time-aligned.

[0021] The sequences of instantaneous voltage and current values ​​are subjected to digital low-pass filtering. The cutoff frequency of the digital low-pass filter is set to half the sampling frequency, i.e., the Nyquist frequency, to filter out high-frequency noise above the Nyquist frequency. The digital low-pass filter adopts a finite impulse response filter structure, with a Hanning window as the window function. The filter order is set according to the requirements of passband ripple and stopband attenuation. The passband cutoff frequency is lower than the Nyquist frequency, the stopband start frequency is the Nyquist frequency, and the stopband attenuation is not less than 60 dB.

[0022] Phase delay compensation is performed on the low-pass filtered current sequence. First, the fundamental phase angles of the filtered voltage and current sequences are extracted. The initial phase angle of the fundamental wave for a given data segment is obtained using a Fast Fourier Transform (FFT), and the phase difference between the initial phase angles of the voltage and current fundamental waves is calculated. This phase difference is converted into a time delay corresponding to the number of sampling points. If the current fundamental phase lags behind the voltage fundamental phase, the corresponding number of sampling points are deleted from the beginning of the current sequence. If the current fundamental phase leads the voltage fundamental phase, the corresponding number of zero-value sampling points are added before the beginning of the current sequence, ensuring that the fundamental phase of the compensated current sequence is consistent with the fundamental phase of the voltage.

[0023] The compensated voltage and current sequences are stored aligned with the sampling timestamps, which are generated by the count value of the sampling clock with a precision of microseconds. The aligned instantaneous voltage and current values ​​constitute a time series dataset.

[0024] Multiple equipment operation event segments are detected and segmented from the time-series dataset, and an event type label is assigned to each segment. Differential operations are performed on the time-series dataset to extract the active power mutation sequence. The instantaneous value of active power is obtained by multiplying the instantaneous voltage and current values ​​at the corresponding timestamp. The active power instantaneous value sequence is subjected to active power integral value difference operations between adjacent power frequency cycles; the instantaneous active power values ​​within a complete power frequency cycle are accumulated and divided by the number of sampling points to obtain the average active power value for that cycle. The difference between the average active power values ​​of adjacent cycles is calculated, and the time point where the absolute value of the difference exceeds a preset power mutation threshold is marked as an active power mutation point. All marked active power mutation points are arranged in chronological order to form an active power mutation point sequence. The preset power mutation threshold is determined by multiplying the standard deviation of power fluctuations during the previous stable operating period of the time-series data by a set multiple, with the set multiple ranging from 3 to 5 times.

[0025] Centered on each active power surge point, voltage and current waveform segments are extracted for each preset cycle before and after the surge point. The preset cycle number is set to ensure that the extracted waveform segments encompass the complete equipment switching transient process; the preset cycle number is 10 power frequency cycles. A Fast Fourier Transform (FFT) is performed on each extracted current waveform segment. The data length of the FFT is the preset cycle number multiplied by the number of sampling points per cycle. The fundamental amplitude and the amplitudes of the 2nd to 15th harmonics are extracted from the FFT results to form an original vector containing 16 elements. The original vector is then normalized using the L2 norm to obtain the load imprint vector. The normalization process is expressed by the following formula:

[0026] in, Represents the first in the load imprint vector One normalized amplitude component, The fundamental amplitude, For the first element in the original vector One harmonic amplitude component, when hour For the fundamental amplitude, when hour This is the amplitude of the second harmonic, and so on. hour The amplitude of the 15th harmonic; denominator Let L2 be the norm of the original vector.

[0027] The cosine similarity of the load imprint vector generated for each current waveform segment is calculated one by one with the template vectors in the pre-stored equipment load imprint template library. The pre-stored equipment load imprint template library stores the load imprint template vectors corresponding to known electrical equipment during start-up or shutdown operations, as well as the electrical equipment operation event type bound to each template vector. When calculating the cosine similarity, the inner product of the load imprint vector and the template vector is first calculated, and then divided by the product of the L2 norm of the load imprint vector and the L2 norm of the template vector. If the cosine similarity of multiple template vectors is greater than the preset similarity threshold, the event type corresponding to the template vector with the highest cosine similarity is selected as the matching result; if only one template vector has a cosine similarity greater than the preset similarity threshold, the event type corresponding to that template vector is directly used as the matching result. The preset similarity threshold is no less than 0.85.

[0028] Active power abrupt changes where the maximum cosine similarity exceeds a preset similarity threshold are marked as electrical equipment operation events. Event type labels are assigned to these abrupt changes based on the event type determined by the matched template vector. For abrupt changes marked as start or stop events, corresponding electrical equipment operation event segments are extracted from the event type label sequence, using the time of the abrupt change as the dividing boundary. Each electrical equipment operation event segment includes transient waveform segments before and after the event and the event type label assigned to that event.

[0029] Example 2 In practice, an event sequence consisting of event type labels is obtained, where each element in the event sequence is an event type label, and the event type labels are arranged in chronological order of the events occurring.

[0030] Traverse the event sequence. For each event type appearing in the sequence, designate each event type as a predecessor event type. For each predecessor event type, locate its position in the event sequence and extract the next event type label immediately following the predecessor event type as the direct successor event type. Count the number of occurrences of the corresponding direct successor event type for each occurrence of the predecessor event type, recording the frequency using the direct successor event type as the index, forming an adjacency frequency table. The adjacency frequency table is a two-dimensional table structure, with rows corresponding to predecessor event types and columns corresponding to successor event types. Each element in the table records the adjacency frequency from the corresponding predecessor event type to the corresponding successor event type.

[0031] For any pair of predecessor and successor event types, the transition probability in the direction from the predecessor to the successor event type is obtained by dividing the adjacency frequency of the predecessor event type recorded in the adjacency frequency table by the total number of occurrences of the predecessor event type in the event sequence. The total number of occurrences of the predecessor event type in the event sequence is obtained by counting the occurrences of the predecessor event type label in the event sequence. The formula for calculating the transition probability is expressed as:

[0032] in, Indicates the type of preceding event To subsequent event types The transition probability, Indicates the type of preceding event The type of subsequent event appears immediately after it appears. The frequency of adjacency, Indicates the type of preceding event Total number of occurrences in the event sequence.

[0033] After calculating the pairwise transition probabilities for all event types, all transition probabilities are arranged with preceding event types as rows and succeeding event types as columns, constructing an event type transition probability matrix. The number of rows in the event type transition probability matrix equals the total number of event types appearing in the event sequence, and the number of columns also equals the total number of event types appearing in the event sequence. The [missing information - likely a specific element or structure] in the event type transition probability matrix... Line number The element value of the column is the predecessor event type. To subsequent event types transition probability .

[0034] Example 3 In practical implementation, after obtaining the transition probability matrix between event types, a correlation strength threshold is set. The correlation strength threshold is determined as follows: extract all non-zero transition probability values ​​from the event type transition probability matrix, sort the non-zero transition probability values ​​in ascending order of numerical value, and take the non-zero transition probability value at the 75th percentile as the correlation strength threshold. The basis for selecting the 75th percentile as the correlation strength threshold is that, after filtering by this threshold, the top 25% of the strongest transition correlations in the event type transition probability matrix are retained, while a large number of weak correlations are eliminated, so that the electricity consumption behavior correlation network can highlight high-frequency and stable electricity consumption behavior conversion relationships.

[0035] Iterate through each transition probability value in the transition probability matrix between event types. For each transition probability value, determine whether it is greater than the association strength threshold. If the transition probability value is greater than the association strength threshold, further determine whether the originating event type and the arriving event type corresponding to the transition probability value are the same. If the originating event type and the arriving event type are different, establish a directed edge between the originating event type node and the arriving event type node, with the direction of the directed edge pointing from the originating event type node to the arriving event type node, and assign the transition probability value as the edge weight of the directed edge.

[0036] If the transition probability value is less than or equal to the association strength threshold, the edge corresponding to the transition probability value is discarded and no directed edge is established.

[0037] If the transition probability value is greater than the association strength threshold, but the starting event type and the arriving event type are the same, that is, the transition probability value is located on the diagonal of the transition probability matrix between event types, representing the node's own self-transition edge, the edge corresponding to the transition probability value is also discarded, and no self-loop edge is established on the node itself.

[0038] After traversing and judging all transition probability values, the electricity consumption behavior association network is formed by all the retained directed edges and corresponding nodes. Each node in the electricity consumption behavior association network corresponds to an event type, and each directed edge corresponds to a valid transition association from the starting event type to the arriving event type. The edge weight is the corresponding transition probability value.

[0039] Example 4 In specific implementation, please refer to Figure 2 Set a preset time window duration and sliding step size. The preset time window duration is 168 hours to cover the complete weekly cycle of electricity consumption behavior patterns; the sliding step size is 24 hours to ensure that adjacent windows maintain continuity in time and capture changes in daytime electricity consumption behavior.

[0040] Extract subsequences from the complete event type label sequence within each window. Using the start time of the event type label sequence as the start time of the first time window, extract a continuous segment of event type labels with a length equal to the preset time window duration, forming the subsequence within the first window. Increase the start time of the first time window by one sliding step to create the start time of the second time window. Extract a continuous segment of event type labels with a length equal to the preset time window duration, forming the subsequence within the second window. Continue sliding the time windows and extracting subsequences in this manner until the end time of the time window exceeds the end time of the event type label sequence.

[0041] For each subsequence within a window, the transition probabilities between event types are recalculated, and an electricity consumption behavior association network is constructed to obtain the window association network. The method for calculating the transition probabilities between event types is as follows: traverse the event types appearing in the subsequences within the window. For each event type as a predecessor event type, count the occurrences of the direct successor event type of the predecessor event type. Divide the occurrence count from the predecessor event type to a certain successor event type by the total occurrence count of the predecessor event type in the subsequence to obtain the transition probability from the predecessor event type to that successor event type. Arrange all transition probabilities into an event type transition probability matrix. The method for constructing the electricity consumption behavior association network is as follows: set an association strength threshold, traverse each transition probability value in the event type transition probability matrix, establish directed edges from the originating event type node to the arriving event type node for non-self-transition probability values ​​greater than the association strength threshold, assign the transition probability value as the edge weight of the directed edge, and discard the remaining edges to obtain the window association network.

[0042] For the window association network, calculate the clustering coefficient for each node. For the node numbered in the window association network... The node will connect all nodes in the network that are connected to the window via directed edges. Other connected nodes are determined as nodes. The set of neighboring nodes, denoted as . Count the total number of directed edges that actually exist between all nodes in the neighbor node set within the window association network, and divide twice the total number of directed edges by... and reduce The product of these terms yields the node. The clustering coefficients are calculated by taking the arithmetic mean of the clustering coefficients of all nodes in the window association network, thus obtaining the average clustering coefficient of the window association network.

[0043] For any node in the window association network Determine the node Out-edge neighbor set Outward neighbor gathering Each element in For a slave node The neighboring node pointed to by the directed edge from which the node originates. To the neighbor The edge weight of a directed edge is denoted as . edge weight The value is the transition probability value. Calculate the node. To the neighbor Connection weight of nodes The proportion of the total weight of all outgoing edges connected ,Proportion The following formula is used to derive:

[0044] in, Represents a node To the neighbor The weight of the directed edge; Represents a node Gather at the neighboring area. any neighbor The weight of the directed edge. To traverse the set of outgoing edges and neighbors The index variable of each neighboring node; denominator Represents a node The sum of the weights of the directed edges to all its outgoing neighbors.

[0045] Calculate the proportion The square value of the outgoing edge neighbor set Sum the squared values ​​of the proportions of all outgoing neighbors to obtain the node. Structural hole constraint coefficient ,Right now The structural hole constraint coefficients of the window association network are obtained by calculating the arithmetic mean of the structural hole constraint coefficients of all nodes in the network.

[0046] The average clustering coefficients and structural hole constraint coefficients calculated for each time window are arranged in chronological order according to the start time of each time window, generating a structural feature sequence composed of the average clustering coefficient sequence and the structural hole constraint coefficient sequence.

[0047] See Figure 3 In the figure, the horizontal axis represents the sliding time window number, and the vertical axis represents the network structure coefficient value calculated within the corresponding time window. The blue solid line represents the average clustering coefficient sequence, and the red dashed line represents the structural hole constraint coefficient sequence. The data in the figure covers multiple continuous sliding time windows, with a time window length of 168 hours and a sliding step size of 24 hours, which can reflect the weekly cycle characteristics and diurnal variation trend of electricity consumption behavior.

[0048] The average clustering coefficient curve shows that the overall value fluctuates between 0.55 and 0.71, exhibiting obvious periodic fluctuations with alternating peaks and troughs. This reflects the significant periodic fluctuations in the degree of clustering of triangular connections between nodes in the electricity consumption behavior network over time. The structural hole constraint coefficient sequence fluctuates between 0.34 and 0.46, also showing a periodic trend, but the magnitude of the change is gentler than that of the average clustering coefficient. The overall trend shows a certain negative correlation with the average clustering coefficient; that is, when the average clustering coefficient increases, the structural hole constraint coefficient decreases, and vice versa.

[0049] Specifically observing the curve changes, the average clustering coefficient reaches local peaks at approximately time windows 50, 210, and 370, while the structural hole constraint coefficient shows local troughs near these points, indicating that the local clustering density of the electricity consumption behavior network is enhanced and the structural hole constraint is weakened, suggesting that the correlation between electricity consumption event types is closer and the behavior pattern is more stable. At approximately time windows 130, 290, and 450, the average clustering coefficient drops to the bottom, and the structural hole constraint coefficient reaches its peak, indicating that the triangular connections between nodes in the network are reduced, the structural hole constraint is strengthened, the network structure is more dispersed, and there is a potential risk of variation in electricity consumption behavior patterns.

[0050] Example 5 In the specific implementation, after obtaining the structural feature sequence composed of the average clustering coefficient sequence and the structural hole constraint coefficient sequence, a look-forward window length is set for the average clustering coefficient sequence. The look-forward window length is 168 time windows, corresponding to 168 sliding steps, to cover the one-week fluctuation cycle of electricity consumption behavior characteristics. For the average clustering coefficient at time n in the average clustering coefficient sequence, 168 consecutive average clustering coefficient values ​​within the look-forward window length before time n are extracted, and the exponentially weighted moving average of the 168 average clustering coefficient values ​​is calculated. The calculation of the exponentially weighted moving average uses a decay factor, which is the quotient of 2 divided by the look-forward window length plus 1, i.e., 2 divided by 169, and the decay factor is approximately equal to 0.0118. When calculating the exponentially weighted moving average, the average clustering coefficient closest to time n is assigned the largest weight, and the weight decreases exponentially with time backtracking according to the decay factor. At the same time, the exponentially weighted standard deviation is calculated using the same 168 average clustering coefficient values ​​and the same decay factor. The exponentially weighted standard deviation is the square root of the exponentially weighted average of the squared differences between the average cluster coefficient values ​​and the exponentially weighted moving average.

[0051] The lower dynamic threshold at time n is obtained by subtracting the first multiple of the exponentially weighted standard deviation from the exponentially weighted moving average at time n. The first multiple is set to 3. The basis for setting the first multiple to 3 is that Chebyshev's inequality states that under any distribution, at least 89% of the data points fall within the range of the mean plus or minus 3 times the standard deviation. Data points outside this range can be regarded as anomalous mutation points.

[0052] For the structural hole constraint coefficient sequence, a lookahead window of 168 time windows is also set. For the structural hole constraint coefficient at time n in the sequence, 168 consecutive structural hole constraint coefficient values ​​within the lookahead window before time n are extracted. The exponentially weighted moving average of these 168 values ​​is calculated, with a decay factor of 2 divided by 169. Simultaneously, the exponentially weighted standard deviation is calculated using the same 168 values ​​and the same decay factor.

[0053] The upper dynamic threshold at time n is obtained by adding the exponentially weighted moving average of the structural hole constraint coefficient at time n to the second exponentially weighted standard deviation. The second multiplier is set to 3, which is also based on the Chebyshev inequality principle to ensure the statistical significance of the abrupt change.

[0054] Continuous monitoring is performed at each time point in the structural feature sequence. The monitoring process is as follows: It is determined whether the average clustering coefficient at the current time n is less than the lower dynamic threshold at the current time n, and simultaneously whether the structural hole constraint coefficient at the current time n is greater than the upper dynamic threshold at the current time n. When both conditions are met simultaneously, the amplitude of the average clustering coefficient being lower than the lower dynamic threshold and the structural hole constraint coefficient being greater than the upper dynamic threshold are further calculated. The amplitude is the difference between the lower dynamic threshold and the average clustering coefficient, and the amplitude of the structural hole constraint coefficient being higher than the upper dynamic threshold is also calculated. The amplitude of the average clustering coefficient being lower than the lower dynamic threshold is compared with a preset minimum decrease, which is 0.5 times the exponentially weighted standard deviation of the average clustering coefficient sequence within the preceding window length. The amplitude of the structural hole constraint coefficient being higher than the upper dynamic threshold is compared with a preset minimum increase, which is also 0.5 times the exponentially weighted standard deviation of the structural hole constraint coefficient sequence within the preceding window length. When the average clustering coefficient is lower than the lower dynamic threshold by an amplitude exceeding the preset minimum decrease, and the structural hole constraint coefficient is higher than the upper dynamic threshold by an amplitude exceeding the preset minimum increase, the current time n is marked as the abrupt change time, triggering the determination of the change in electricity consumption behavior pattern.

[0055] After identifying variations in electricity consumption behavior patterns, starting from the marked abrupt change, the event type transition probability matrix of the window association network corresponding to multiple consecutive sliding time windows following the abrupt change is extracted. The number of extractions is set to 24 sliding time windows, corresponding to 24 sliding step sizes. For each subsequent window's event type transition probability matrix, the values ​​of all matrix elements are summed to obtain the total transition activity for that window. The total transition activity reflects the overall strength of event transitions in the electricity consumption behavior association network within the corresponding time window.

[0056] The total transition activity of each of the 24 windows is arranged in chronological order, and the gradient sequence of the total transition activity sequence is calculated. The gradient is calculated as follows: for the total transition activity from window t to window t+1, the total transition activity of window t+1 is subtracted from the total transition activity of window t, and the difference is taken as the gradient value of time step t. This process is repeated to obtain a gradient sequence consisting of 23 gradient values.

[0057] After obtaining the gradient sequence, a preset observation interval is set, which is the interval from the 4th gradient value after the abrupt change to the 20th gradient value. All gradient values ​​within the preset observation interval are analyzed: if all gradient values ​​within the preset observation interval are positive, and the absolute value of the gradient values ​​gradually increases within the preset observation interval, then the electricity consumption trend is determined to have reached an upward inflection point at the abrupt change; if all gradient values ​​within the preset observation interval are negative, and the absolute value of the gradient values ​​gradually increases within the preset observation interval, then the electricity consumption trend is determined to have reached a downward inflection point at the abrupt change.

[0058] A weak threshold is set, which is 0.2 times the arithmetic mean of the absolute values ​​of all gradient values ​​in the gradient change sequence. The fluctuation range of the gradient values ​​within a preset observation interval is compared with the set weak threshold. The fluctuation range is the absolute value of the difference between the maximum and minimum values ​​of the gradient values ​​within the preset observation interval. If the fluctuation range is less than the set weak threshold, it is determined that the electricity consumption trend continues its original trend after the abrupt change and no inflection point has occurred.

[0059] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for sensing electricity consumption trends based on correlation analysis of electricity consumption behavior characteristics, characterized in that, include: Collect high-frequency sampling current and voltage data of the total incoming lines within the power consumption area to form a time-series dataset; Multiple electrical equipment operation event segments are detected and segmented from the time-series dataset, and an event type label is assigned to each electrical equipment operation event segment; Based on the chronological order of events, the frequency of consecutive occurrences of different event types is counted to generate a transition probability matrix between event types. The transition probability matrix between the event types is constructed as a directed weighted network of electricity consumption behavior associations, with event types as nodes and transition probabilities as edge weights. The average clustering coefficient and structural hole constraint coefficient of the electricity consumption behavior association network are calculated within a sliding time window to form a structural feature sequence; Monitor the abrupt change in the structural feature sequence. When the decrease in the average clustering coefficient and the increase in the structural hole constraint coefficient both exceed the dynamic threshold, it is determined that a change in electricity consumption behavior pattern has occurred. Based on the timing of the changes in the electricity consumption behavior pattern and the gradient of the subsequent transition probability, the inflection point and direction of the electricity consumption trend are determined.

2. The method for sensing electricity consumption trends based on correlation analysis of electricity consumption behavior characteristics according to claim 1, characterized in that, The process of detecting and segmenting multiple electrical equipment operation event fragments from the time-series dataset, and assigning an event type label to each electrical equipment operation event fragment, includes: Perform a difference operation on the time series dataset to extract the sequence of abrupt changes in active power; The voltage and current waveform segments of the preset cycles before and after each abrupt change point are extracted. The waveform segment is subjected to Fast Fourier Transform to extract the harmonic amplitude spectrum and generate a load imprint vector; The cosine similarity between the load imprint vector and the template vector in the pre-stored equipment load imprint template library is calculated one by one. The mutation points where the maximum cosine similarity is greater than a preset threshold are marked as electrical equipment start-up or stop events. The event type is determined based on the matched template vector, thereby segmenting the corresponding electrical equipment operation event fragments and assigning event type labels.

3. The method for sensing electricity consumption trends based on correlation analysis of electricity consumption behavior characteristics according to claim 1, characterized in that, The step of calculating the frequency of consecutive occurrences of different event types based on the chronological order of events, and generating a transition probability matrix between event types, includes: Traverse the event sequence composed of the event type labels, and for the current event type as the predecessor event type, count the occurrences of its direct successor event type to form an adjacency frequency table; The transition probability in that direction is obtained by dividing the adjacency frequency of the preceding event type to a certain successor event type by the total number of occurrences of the preceding event type in the event sequence. Arrange the transition probabilities between all event type pairs by rowing the preceding event type and columning the succeeding event type, and construct the transition probability matrix between the event types.

4. The method for sensing electricity consumption trends based on correlation analysis of electricity consumption behavior characteristics according to claim 1, characterized in that, The step of constructing the transition probability matrix between the event types into a directed weighted network of electricity consumption behavior associations, with event types as nodes and transition probabilities as edge weights, includes: Set an association strength threshold, traverse each transition probability value in the transition probability matrix between the event types, establish a directed edge between the starting event type node and the arriving event type node corresponding to the transition probability value that is greater than the association strength threshold, and assign the transition probability value as the edge weight. By discarding edges corresponding to transition probability values ​​below the association strength threshold and the self-transition edges of the nodes themselves, the electricity consumption behavior association network is obtained.

5. The method for sensing electricity consumption trends based on correlation analysis of electricity consumption behavior characteristics according to claim 1, characterized in that, The average clustering coefficient and structural hole constraint coefficient of the electricity consumption behavior association network are calculated within a sliding time window to form a structural feature sequence, including: Using a preset time window duration and sliding step size, extract subsequences within each window from the complete event type label sequence; For each subsequence within a window, the transition probabilities between event types are recalculated and the electricity consumption behavior association network is constructed to obtain the window association network; For the window association network, the clustering coefficient of each node is calculated as the ratio of the actual number of connected edges between its neighboring nodes to the possible number of connected edges. Then, the average clustering coefficient of all nodes is calculated to obtain the average clustering coefficient. For any node in the window association network, calculate the proportion of its connection weight with each neighbor node to the total of all outgoing edge connection weights of the node, square the proportion, and sum the squared values ​​of all neighbor nodes to obtain the structural hole constraint coefficient of the node. Then calculate the average value of the structural hole constraint coefficients of all nodes as the structural hole constraint coefficient; The average clustering coefficients and structural hole constraint coefficients obtained from each window are arranged along the time axis to generate the structural feature sequence.

6. The method for sensing electricity consumption trends based on correlation analysis of electricity consumption behavior characteristics according to claim 1, characterized in that, When the monitoring of abrupt changes in the structural feature sequence detects that the decrease in the average clustering coefficient and the increase in the structural hole constraint coefficient simultaneously exceed a dynamic threshold, a change in electricity consumption behavior pattern is determined to have occurred, including: For the sequence of average clustering coefficients, calculate the exponentially weighted moving average and exponentially weighted standard deviation within the preceding window length at each time step, and use the exponentially weighted moving average minus the first multiple of the exponentially weighted standard deviation as the lower dynamic threshold. For the sequence of structural hole constraint coefficients, calculate the exponentially weighted moving average and exponentially weighted standard deviation within the preceding window length at each time step, and add a second multiple of the exponentially weighted standard deviation to the exponentially weighted moving average as the upper dynamic threshold. Continuously monitor whether the average clustering coefficient at the current moment is less than the lower dynamic threshold and whether the structural hole constraint coefficient at the current moment is greater than the upper dynamic threshold; When both conditions are met, and the magnitude of the average clustering coefficient being lower than the lower dynamic threshold exceeds a preset minimum decrease, and the magnitude of the structural hole constraint coefficient being higher than the upper dynamic threshold exceeds a preset minimum increase, the current moment is marked as a sudden change moment, triggering the determination of electricity consumption behavior pattern variation.

7. The method for sensing electricity consumption trends based on correlation analysis of electricity consumption behavior characteristics according to claim 1, characterized in that, The step of determining the inflection point and trend direction of electricity consumption trends based on the occurrence time of the changes in electricity consumption behavior patterns and the gradient of subsequent transition probabilities includes: Starting from the moment when the electricity consumption behavior pattern variation occurs, extract the transition probability matrix between the event types corresponding to multiple consecutive sliding time windows thereafter; For each subsequent window's transition probability matrix, sum all matrix elements to obtain the total transition activity; Calculate the gradient sequence of the total transfer activity over time; If the change gradient sequence remains positive within the preset observation interval and its absolute value gradually increases, it is determined that the electricity consumption trend has reached an upward inflection point at the time of occurrence. If the change gradient sequence remains negative within the preset observation interval and its absolute value gradually increases, then it is determined that the electricity consumption trend has reached a downward inflection point at the time of occurrence. If the fluctuation range of the change gradient sequence is less than the set weak threshold, it is determined that the electricity consumption trend continues without an inflection point.

8. The method for sensing electricity consumption trends based on correlation analysis of electricity consumption behavior characteristics according to claim 1, characterized in that, The high-frequency sampling current and voltage data of the total incoming lines within the power consumption area are collected to form a time-series dataset, including: The instantaneous voltage and current values ​​of the main incoming line are synchronously collected at a rate of no less than 256 points per cycle. The sequences of instantaneous voltage and current values ​​are subjected to digital low-pass filtering to remove high-frequency noise above the Nyquist frequency; Phase delay compensation is applied to the filtered current sequence to make its fundamental phase consistent with the fundamental phase of the voltage. The compensated voltage and current sequences are stored in alignment with timestamps to form the time-series dataset.

9. The method for sensing electricity consumption trends based on correlation analysis of electricity consumption behavior characteristics according to claim 2, characterized in that, The load imprint vector is formed by normalizing the amplitudes of the fundamental wave and the 2nd to 15th harmonics using the L2 norm.

10. A method for sensing electricity consumption trends based on correlation analysis of electricity consumption behavior characteristics according to claim 2, characterized in that, When calculating the cosine similarity between the load imprint vector and the template vectors in the pre-stored equipment load imprint template library one by one, if there are multiple template vectors whose cosine similarity is greater than a preset threshold, then the event type corresponding to the template vector with the largest cosine similarity is selected as the matching result.

Images