Big data-based regional power distribution network short-term load intelligent prediction system

Abstract

The present application relates to the technical field of big data analysis, in particular to a regional power distribution network short-term load intelligent prediction system based on big data, which comprises a frequency domain filtering and load mode extraction module, which extracts a high-frequency fluctuation load sequence and a low-frequency reference load sequence according to the power distribution transformer topology node position coordinates and the initial power sequence collected by the smart meter at high frequency. In the present application, the low-frequency reference load sequence is trend extrapolated, and then echoed with the aforementioned disturbance result, so that the long-term electricity consumption trend and the short-term external impact jointly enter the prediction process, avoiding weakening the instantaneous fluctuation by only relying on the overall historical trend, and avoiding ignoring the basic load inertia by only focusing on the instantaneous change. Thus, synergistic gains are formed in terms of layered expression of load changes, depiction of node linkage relationship, depth of meteorological disturbance mapping and integrity of future trend connection, which can improve the adaptability of the regional power distribution network short-term load prediction result to complex scenarios.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, and in particular to a regional power distribution network short-term load intelligent forecasting system based on big data. Background Technology

[0002] The regional distribution network short-term load intelligent prediction technology mainly involves distribution network operation status analysis, power load change modeling, time series data processing, and multi-source information fusion processing. It analyzes the change pattern of power load in a short time scale based on the distribution transformer topology node structure, high-frequency data collected by smart meters, meteorological environmental parameters, and regional power consumption behavior characteristics.

[0003] Existing technologies focus on distribution network operation status analysis, electricity load change modeling, time-series data processing, and multi-source information fusion. While they can analyze the changing patterns of power load over short timescales, in practice, they typically integrate collected load information, meteorological parameters, and regional electricity consumption characteristics into a single analytical framework for overall judgment. This lack of separation between different levels of load variation leads to the overlapping of continuous changes and short-term fluctuations in the same data representation. This results in easily mixed prediction data, frequent short-term fluctuations, sudden load changes at local nodes, and concentrated weather disturbances during operation, making it difficult to accurately identify the sources of load changes and thus affecting the stability of prediction results. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a regional power distribution network short-term load intelligent prediction system based on big data.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a regional distribution network short-term load intelligent forecasting system based on big data includes: The frequency domain filtering and load mode extraction module extracts and generates a high-frequency fluctuating load sequence and a low-frequency reference load sequence based on the location coordinates of the distribution transformer topology nodes and the initial power sequence collected by the smart meter at high frequency. The spatial load correlation calculation module concatenates the high-frequency fluctuation load sequences of the corresponding nodes sequentially and calculates the node collaborative distribution correlation degree; it then calculates the node collaborative distribution correlation degree with a preset probability distribution value to construct a spatial transfer probability feature matrix. The meteorological feature cross-mapping module constructs a historical meteorological feature vector based on the collected temperature and humidity values, wind speed abrupt change amplitude, and rainfall range. This historical meteorological feature vector is then fitted with the high-frequency fluctuation load sequence of the corresponding node to obtain the node's meteorological influence weight. Combining this with the future meteorological feature vector composed of meteorological forecast data for the target prediction period and the node's meteorological influence weight, the module obtains the initial load fluctuation increment for each node. Finally, by combining the spatial transition probability feature matrix, the module solves for the initial load fluctuation increment of each node to obtain the meteorological disturbance load increment value for each node. The regional load joint forecasting module performs trend extrapolation on the low-frequency reference load sequence of each node to obtain a regional distribution network load forecasting scheme.

[0006] Preferably, the steps for obtaining the high-frequency fluctuating load sequence and the low-frequency reference load sequence are as follows: Based on the location coordinates of the distribution transformer topology nodes and the initial power sequence collected by the smart meter at high frequency, the initial power sequence collected by the smart meter at high frequency is checked by node-by-node correspondence, sampling time-by-sampling time alignment, and sampling point-by-sampling point according to the node number, node arrangement order, and node positioning mark in the location coordinates of the distribution transformer topology nodes. Power records with mismatched node numbers are eliminated, and power records with misaligned sampling times are corrected to form the node coordinate mapping initial power sequence. Based on the node coordinates, the initial power sequence is mapped, the power value at each sampling time is read node by node, the frequency components of each sampling segment are expanded according to the sampling order, the frequency components with continuous amplitude changes in continuous sampling segments are extracted, the frequency components with abnormal phase jump variables in adjacent sampling segments are screened out, and the reverse reconstruction and power value backfilling are performed according to the original sampling time order to generate a denoised power time series. Based on the denoised power time series, the power rise inflection point, power fall inflection point, local peak point and local valley point are identified segment by segment. The envelope boundary is connected based on the local peak point and local valley point. The load components are split according to the distance between adjacent extreme points and the length of continuous oscillation period. The load components that change continuously across multiple sampling periods are extracted. The load components that fluctuate repeatedly in adjacent sampling periods are extracted to obtain the high-frequency fluctuating load sequence and the low-frequency reference load sequence.

[0007] Preferably, the step of obtaining the node collaborative distribution correlation degree is as follows: The high-frequency fluctuating load sequences of the corresponding nodes are spliced ​​together sequentially according to the distribution topology node numbers. The sampling time order and node affiliation mark of the high-frequency fluctuating load sequences of each node are checked. The load values ​​of each node are extracted according to the unified sampling time. The load values ​​of each node are sorted from small to large at each sampling time. The changes in the sorting position of each node, the number of position swaps between adjacent nodes, and the number of cross-position moves are recorded at each sampling time to form the node sequence swapping statistics. Based on the node sequence swapping statistics, the ranking position of each node is extracted at all sampling times. The number of times the ranking changes in the same direction, the number of times the ranking changes in opposite directions, and the distribution of the ranking position difference between any two nodes at each sampling time are counted. The rank correlation coefficient is calculated based on the ranking position change relationship. The synchronous occurrence records of each node in the first and last ranking intervals are further screened. The common occurrence frequency and continuous maintenance frequency of extreme ranking positions are counted to determine the extreme value of tail dependency. The node cooperative distribution correlation degree is obtained by combining the rank correlation coefficient and the extreme value of tail dependency.

[0008] Preferably, the steps for obtaining the spatial transition probability feature matrix are as follows: The correlation degree of each group of nodes is multiplied according to the preset probability distribution of each item. According to the connection order from the node's starting position to the node's ending position, each product item is continuously integrated and the integration result is extracted to obtain the numerical item of the integration result of each node. Then, according to the row position and column position of the distribution topology node position, each numerical item of the integration result is filled in to obtain the spatial transition probability feature matrix.

[0009] Preferably, the step of obtaining the initial load fluctuation increment is as follows: Based on the collected temperature and humidity values, wind speed fluctuation range, and rainfall range, the temperature and humidity values, wind speed fluctuation range, and rainfall range are mapped node by node according to a unified sampling time. The temperature and humidity values, wind speed fluctuation range, and rainfall range are combined in a fixed order at the same node and time to form a historical meteorological feature vector. Simultaneously, the temperature and humidity forecast values, wind speed fluctuation forecast range, and rainfall forecast range for the target forecast period are acquired and combined in the same order to form a future meteorological feature vector. Extract the temperature and humidity values, wind speed fluctuation amplitude, and rainfall range corresponding to each sampling time in the historical meteorological feature vector. Read the high-frequency fluctuating load sequence of the corresponding node at the same sampling time. Match the historical meteorological feature vector and the high-frequency fluctuating load sequence item by item according to the sampling time. Statistically analyze the matching of each meteorological component with the load fluctuation direction, load fluctuation amplitude, and load fluctuation duration. Assign the influence weight of each meteorological component according to the degree of matching. Map the influence weight to the future meteorological feature vector. Summarize the influence values ​​of each forecast meteorological component on the future load fluctuation at each forecast time to obtain the initial load fluctuation increment.

[0010] Preferably, the step of obtaining the incremental value of the meteorological disturbance load is as follows: Read the matrix position values ​​of each node in the spatial transfer probability feature matrix, map the initial load fluctuation increment to the spatial transfer probability feature matrix item by item according to the node connection order, calculate the transmission value of the initial load fluctuation increment of each node to adjacent nodes and associated nodes, accumulate the diffusion results in each transmission direction, extract the average value of the diffusion results within the sliding time window according to the continuous prediction time, and obtain the meteorological interference load increment value.

[0011] Preferably, the step of obtaining the short-term predicted total load value is as follows: A time series forecasting algorithm is used to extrapolate the evolution trend of the low-frequency reference load sequence of each node to obtain the future low-frequency reference forecast sequence of each node in the target forecast period. Based on the equipment capacity setpoint of each distribution transformer topology node, the equipment capacity setpoint of each distribution transformer topology node, the incremental value of the meteorological interference load, and the future low-frequency reference forecast sequence are read one by one to form a node load superposition correspondence. Based on the node load superposition correspondence, according to the prediction time sequence of each distribution transformer topology node, the incremental value of meteorological interference load is extracted time by time, and the load value in the future low-frequency reference prediction sequence is extracted time by time. The incremental value of meteorological interference load at each prediction time is superimposed on the load value in the future low-frequency reference prediction sequence of the same distribution transformer topology node and the same prediction time. The node load results after superposition at each prediction time are recorded to obtain the total short-term predicted load value of each node.

[0012] Preferably, the steps for obtaining the regional distribution network load forecasting scheme are as follows: Based on the short-term predicted total load value of each node, the overload setting value of the equipment corresponding to each distribution transformer topology node is read one by one. The short-term predicted total load value and the equipment overload setting value of each node are compared item by item according to the distribution transformer topology node number. Node records that have not exceeded the equipment overload setting value are extracted, and node records that have reached or exceeded the equipment overload setting value are extracted. The over-limit status, corresponding prediction time and corresponding load value of each distribution transformer topology node are summarized to obtain the regional distribution network load prediction scheme.

[0013] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, based on the location coordinates of the distribution transformer topology nodes and the initial power sequence collected by the smart meter at high frequency, the different frequency components in the original power change are first separated, and then the high-frequency fluctuating load sequence and the low-frequency reference load sequence are further extracted. This allows short-term disturbance information and continuous change information in the load to enter the subsequent calculation chain, making the source of load change clearer and the subsequent judgment more targeted. On this basis, the high-frequency fluctuating load sequences of the corresponding nodes are sequentially spliced ​​together, and the synchronous fluctuation relationship, the order change relationship, and the tail concentration change relationship between nodes are continuously calculated. Then, the node cooperative distribution correlation degree and the probability distribution preset quantity are combined to form a spatial transfer probability feature matrix, so that the linkage propagation characteristics between different nodes in the region are transformed from single-point observation to overall correlation expression. The traction relationship of local disturbances on surrounding nodes is quantified, and the regional load fluctuation is no longer limited to static judgment at the level of isolated nodes. Furthermore, temperature and humidity values, wind speed change amplitude, and rainfall range values ​​are introduced to form a historical meteorological feature vector, and the historical meteorological feature vector is fitted and calculated with the high-frequency fluctuating load sequence of the corresponding node. By obtaining the node meteorological impact weights and combining them with the future meteorological feature vector composed of meteorological forecast data for the target prediction period, an initial load fluctuation increment is formed. Then, by combining the spatial transfer probability feature matrix, the node-to-node extension solution is performed, so that meteorological disturbances are not only reflected as external influences at a single point in time, but are transformed into load disturbance results that combine node differences, future time-period directionality, and spatial transmission. This improves the ability to identify sudden weather, local weather differences, and disturbance diffusion phenomena in adjacent areas. At the same time, the low-frequency benchmark load sequence is extrapolated to reflect the trend, and then it is combined with the aforementioned disturbance results. This allows long-term electricity consumption trends and short-term external shocks to enter the prediction process together, avoiding the weakening of instantaneous fluctuations by judging solely based on overall historical trends, and avoiding the neglect of basic load inertia by focusing only on instantaneous changes. This results in synergistic gains in the hierarchical expression of load changes, the characterization of node linkage relationships, the depth of meteorological disturbance mapping, and the completeness of future trend connection. This can improve the adaptability of the short-term load prediction results of the regional distribution network to complex scenarios, the ability to identify abnormal fluctuations in advance, and the accuracy of characterizing node-level load differences. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the principle of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] Please see Figure 1 This invention provides a technical solution: a regional distribution network short-term load intelligent forecasting system based on big data, comprising: The frequency domain filtering and load mode extraction module extracts and generates a high-frequency fluctuating load sequence and a low-frequency reference load sequence based on the location coordinates of the distribution transformer topology nodes and the initial power sequence collected by the smart meter at high frequency. The spatial load correlation calculation module concatenates the high-frequency fluctuating load sequences of the corresponding nodes sequentially and calculates the node collaborative distribution correlation degree; it then calculates the node collaborative distribution correlation degree with the preset probability distribution to construct a spatial transfer probability feature matrix. The meteorological feature cross-mapping module constructs a historical meteorological feature vector based on the collected temperature and humidity values, wind speed abrupt change amplitude, and rainfall range. It then fits the historical meteorological feature vector with the high-frequency fluctuation load sequence of the corresponding node to obtain the meteorological influence weight of the node. Combining the future meteorological feature vector composed of meteorological forecast data for the target prediction period with the node meteorological influence weight, it obtains the initial load fluctuation increment of a single node. Finally, by combining the spatial transition probability feature matrix, it solves for the initial load fluctuation increment of each node to obtain the meteorological disturbance load increment value of each node. The regional load joint forecasting module extrapolates the low-frequency reference load sequence of each node to obtain a regional distribution network load forecasting scheme.

[0017] The steps for obtaining the high-frequency fluctuating load sequence and the low-frequency reference load sequence are as follows: Based on the location coordinates of the distribution transformer topology nodes and the initial power sequence collected by the smart meter at high frequency, the initial power sequence collected by the smart meter at high frequency is checked by node-by-node correspondence, sampling time-by-sampling time alignment, and sampling point-by-sampling point according to the node number, node arrangement order, and node positioning mark in the location coordinates of the distribution transformer topology nodes. Power records with mismatched node numbers are eliminated, and power records with misaligned sampling times are corrected to form the node coordinate mapping initial power sequence. Based on the node coordinate mapping of the initial power sequence, the power value of each sampling time is read node by node, the frequency components of each sampling segment are expanded according to the sampling order, the frequency components with continuous amplitude changes in continuous sampling segments are extracted, the frequency components with abnormal phase jump variables in adjacent sampling segments are screened out, and the reverse reconstruction and power value backfilling are performed according to the original sampling time order to generate a denoised power time series. Based on the denoised power time series, the power rise inflection point, power fall inflection point, local peak point and local valley point are identified segment by segment. The envelope boundary is connected according to the local peak point and local valley point. The load components are split according to the distance between adjacent extreme points and the length of continuous oscillation period. The load components that change continuously across multiple sampling periods are extracted, and the load components that fluctuate repeatedly in adjacent sampling periods are extracted to obtain the high-frequency fluctuating load sequence and the low-frequency reference load sequence.

[0018] Specifically, based on the location coordinates of the distribution transformer topology nodes and the initial power sequence collected by the smart meters at high frequency, the unique node number, geographical coordinates (latitude and longitude), and upstream and downstream connection relationships of each node in the power grid topology are first parsed and extracted from the topology node location coordinates. Based on this, a fast lookup mapping table of node numbers and unique hardware identifiers (IDs) of smart meters is constructed in memory. Then, power records in the initial power sequence are processed one by one. The mapping table is consulted to verify the attribution of the meter ID in each record. If a meter ID does not have a corresponding node number in the mapping table, the record is determined as an invalid record with a mismatched node number and is directly discarded. Next, to address the issue of inconsistent sampling times among the meters, a globally unified standard time grid is set, with its time interval determined according to the sampling frequency. For example, for a sampling frequency of once every 15 seconds, the time grid points are 0 seconds, 15 seconds, 30 seconds, and 45 seconds per minute. The original timestamps of all power records are merged and aligned to the nearest standard time grid point, while a time alignment tolerance is set. The threshold is obtained by analyzing the distribution of sampling time intervals exceeding 95% in the historical dataset, for example, set to 2 seconds. If the absolute value of the difference between the original timestamp of a record and the most recent standard time point exceeds 2 seconds, it is determined to be a sampling time misalignment, and its power value is corrected by linear interpolation. Specifically, the estimated power value at the current standard time point is calculated based on the two most recent successfully aligned effective power values ​​of the node before and after the misalignment time point. Finally, all sampling points are numerically verified. Based on the rated capacity of the equipment nameplate of each distribution transformer node, its effective operating range of power value is set. For example, for a transformer with a rated capacity of 500kVA, its effective power range can be set to 0kW to 600kW (considering 1.2 times the short-term overload capacity). Any power value outside this range, or a sampling point recorded as a null value or non-numerical (NaN), is marked as an anomaly. For segments with no more than 3 consecutive anomalies, the average value of the 5 effective sampling points before and after the anomaly point is used to fill the gap. After completing all the above processing, the initial power sequence of node coordinate mapping is formed.

[0019] Based on the initial power sequence mapped from the node coordinates, the power time series of each node is denoised independently. First, the complete sequence data is divided into multiple equal-length, overlapping data segments. For example, each data segment is set to 2048 sampling points, and the overlap between adjacent data segments is set to 1024 sampling points. A Hanning window function is applied to each data segment to reduce spectral leakage. Then, a Fast Fourier Transform (FFT) is used to transform it from the time domain to the frequency domain, obtaining the amplitude and phase information of each frequency component within that data segment. To extract frequency components with consistent amplitude variations, the coefficient of variation (i.e., standard deviation divided by the average) of the amplitude sequence at each frequency point across all consecutive data segments needs to be calculated. The average value is used to determine the amplitude consistency threshold. This threshold is obtained by analyzing the distribution of the coefficient of variation of frequency components under stable operating conditions in a large amount of historical data. The 90th percentile of this distribution is taken as the threshold, for example, 0.35. If the coefficient of variation of amplitude at a certain frequency point is greater than 0.35, it is considered that the frequency component is caused by random noise or non-periodic disturbance, and its amplitude is unstable. Therefore, the amplitude value of the frequency component in the current data segment is set to zero. Next, in order to filter out frequency components with abnormal phase jumps, the phase difference of each frequency component between adjacent data segments is calculated, and the theoretical phase shift determined by the frequency value and the data segment translation time is subtracted to obtain the phase jump deviation. A phase stability threshold is then set, for example... The value of radians is determined based on statistical analysis of the phase jitter of an ideal sinusoidal signal superimposed with a small amount of noise. If the absolute value of the phase jump deviation of a frequency component exceeds this threshold, its phase information is considered unreliable, and its amplitude is also set to zero. After filtering out unstable frequency components, the modified frequency domain representation of all data segments is converted back to the time domain through inverse fast Fourier transform (IFFT). Finally, the overlapping addition method is used to smoothly stitch together all the processed time domain data segments to reconstruct a complete power time series that has been filtered out of high-frequency noise and aperiodic interference, thus generating a denoised power time series.

[0020] Based on the denoised power time series, the sequence data of each node is sampled and analyzed segment by segment. The inflection points of power change are identified by calculating the first difference of the sequence. When the first difference value changes from positive to negative, the corresponding point is identified as a local peak point; when the first difference value changes from negative to positive, it is identified as a local valley point. After identifying all local peak and valley points, a cubic spline interpolation algorithm is applied to both sets of local peak and valley points. This algorithm generates a smooth upper envelope passing through all peak points and a smooth lower envelope passing through all valley points. These two envelopes define the dynamic range of load fluctuations. Subsequently, a central trend line is constructed by calculating the average value of the upper and lower envelopes at each time step. This process is expressed by the formula: ,in, It is a moment The load value of the center trend line, Is the upper envelope at time The load value, Is the lower envelope at time The load value, with this central trend line representing the load component that changes continuously across multiple sampling periods (e.g., hours to a day), is thus extracted as the low-frequency reference load sequence. Next, the value of the original denoised power time series at each moment is subtracted from the value of the corresponding low-frequency reference load sequence. ,in To denoise the power time series at time... The power values, and the resulting difference sequence This refers to the load component that fluctuates repeatedly within adjacent sampling periods, i.e., the high-frequency fluctuating load sequence. To ensure the effectiveness of the split, a period length judgment threshold is set, such as 15 minutes. By performing spectral analysis on the two obtained sequences, it is confirmed that the energy of the low-frequency reference load sequence is mainly concentrated in the low-frequency band with a period of more than 15 minutes, while the energy of the high-frequency fluctuating load sequence is mainly distributed in the high-frequency band with a period of less than 15 minutes. If it meets this distribution characteristic, the split result is finally confirmed, and the high-frequency fluctuating load sequence and the low-frequency reference load sequence are obtained.

[0021] The steps to obtain the correlation degree of node collaborative distribution are as follows: The high-frequency fluctuating load sequences of the corresponding nodes are spliced ​​together in sequence according to the distribution topology node number. The sampling time order and node affiliation mark of the high-frequency fluctuating load sequences of each node are checked. The load values ​​of each node are extracted according to the unified sampling time. The load values ​​of each node are sorted from small to large at each sampling time. The changes in the sorting position of each node, the number of position swaps between adjacent nodes, and the number of cross-position moves are recorded at each sampling time to form the node sequence swapping statistics. Based on the node sequence swapping statistics, the ranking position of each node is extracted at all sampling times. The number of times the ranking changes in the same direction, the number of times the ranking changes in opposite directions, and the distribution of the ranking position difference between any two nodes at each sampling time are counted. The rank correlation coefficient is calculated based on the ranking position change relationship. The synchronous occurrence records of each node in the first and last ranking intervals are further screened. The common occurrence frequency and continuous maintenance frequency of extreme ranking positions are counted to determine the extreme value of tail dependency. The node cooperative distribution correlation degree is obtained by combining the rank correlation coefficient and the extreme value of tail dependency.

[0022] Specifically, the high-frequency fluctuating load sequences of all corresponding nodes are logically concatenated according to the distribution topology node numbers to form a time-aligned data matrix. Each row represents a unified sampling time, and each column represents a specific node. The values ​​in the matrix are the load values ​​of that node at that time. Then, each row of this data matrix is ​​traversed, i.e., each sampling time, and the load values ​​of all nodes at that time are sorted in ascending order from smallest to largest. The ranking of each node at that time is recorded. For example, at time... If there are 5 nodes A, B, C, D, and E with load values ​​of 10, 30, 5, 25, and 15 respectively, then the sorted positions are C(1), A(2), E(3), D(4), and B(5). Record the positions at all times to form a time series of the position of each node. Next, compare two adjacent times (… and The position sequence of the nodes is used to count the changes in position. The exchange of positions between adjacent nodes refers to the occurrence of a position change at time 0. Two nodes with adjacent positions (e.g., positions k and k+1) at time... The positions of the nodes are exactly interchanged. The total number of times this happens during the entire sampling period is counted. The number of cross-position moves refers to the absolute value of the position change of a node being greater than a preset move threshold. This threshold is set to 5% of the total number of nodes. For example, if there are 100 nodes, the threshold is 5. If the position of a node changes from 20 to 26, it is counted as one cross-position move. The number of adjacent position interchanges and the number of cross-position moves of all nodes are accumulated to form the node sequence swap statistics.

[0023] Based on the node sequence swapping statistics and the time series of each node's position, Spearman correlation is first used to calculate the correlation between any two nodes (e.g., node i and node j) based on their position time series. This method does not depend on the specific distribution of the data, but only focuses on the changing trend of the ranking position, thus obtaining a rank correlation coefficient value representing the synchronicity of the load fluctuation trends of the two nodes. Then, to capture the correlation under extreme load conditions, the first and last ranking intervals are defined. The range of these two intervals is determined based on the probability of extreme events occurring in historical data. For example, the first 10% of the ranking positions are defined as the first ranking interval (high load area), and the last 10% are defined as the last ranking interval (low load area). Next, all sampling times are screened, and the node... The number of times node i and node j appear synchronously in the first interval of the sorting sequence, and the number of times they appear synchronously in the last interval of the sorting sequence, are calculated as the proportions of these two counts to the total number of sampling times. These are used as the upper tail dependency and lower tail dependency, respectively. Simultaneously, the longest duration for which they remain continuously in either extreme interval is counted as the continuous holding frequency. Then, a weighted fusion method is used to determine the final extreme value of the tail dependency. For example, the extreme value of the tail dependency = 0.7 * (upper tail dependency + lower tail dependency) + 0.3 * (normalized continuous holding frequency), where the weights 0.7 and 0.3 are empirically set values. Finally, the rank correlation coefficient and the calculated extreme value of the tail dependency are linearly combined to obtain the node cooperative distribution correlation degree, which is calculated as: Node cooperative distribution correlation degree = *Rank correlation coefficient + (1- *Extrema of tail dependency, where the weights Settings are configured based on the experience of distribution network operation experts, for example... It can be set to 0.6 to balance the impact of overall trend similarity and extreme event synchronicity on the correlation degree, thus obtaining the node collaborative distribution correlation degree.

[0024] The steps to obtain the spatial transition probability feature matrix are as follows: The correlation degree of each group of nodes is multiplied according to the preset probability distribution of each item. According to the connection order from the starting position to the ending position of the node, each product item is continuously integrated and the integration result is extracted to obtain the numerical item of the integration result of each node. Then, according to the row position and column position of the distribution topology node, each numerical item of the integration result is filled in to obtain the spatial transition probability feature matrix.

[0025] Specifically, the correlation degree of each group of nodes is associated with a preset probability distribution value, which is then multiplied. This preset probability distribution value is pre-calculated based on the physical topology of the distribution network and reflects the inherent possibility of influence transmission between nodes. Specifically, for any two nodes i and j, their preset probability distribution values ​​are inversely proportional to the electrical distance between them (the sum of line impedances). For example, the preset probability distribution value... ,in The equivalent impedance from node i to node j is given, and the preset values ​​for all nodes are normalized. The calculated node cooperative distribution correlation degree is then used. With the corresponding probability distribution preset quantity Multiplying them yields an initial correlation strength value. Then, for each row in the matrix (representing a starting node i), calculate the sum of all elements in that row, i.e. Where N is the total number of nodes in the network, then, using each element in that row... Divide by the sum of the rows To obtain the transition probability ,this This is the "numerical term of the integration result," which represents the probability of the influence being transferred from node i to node j. Finally, all the calculated transfer probabilities are combined according to the order of the distribution topology node numbers. Fill the matrix into an N×N matrix, where the element in the i-th row and j-th column is the... The spatial transition probability feature matrix is ​​obtained.

[0026] The steps for obtaining the initial load fluctuation increment are as follows: Based on the collected temperature and humidity values, wind speed fluctuation range, and rainfall range, the temperature and humidity values, wind speed fluctuation range, and rainfall range are mapped node by node according to a unified sampling time. The temperature and humidity values, wind speed fluctuation range, and rainfall range are combined in a fixed order at the same node and time to form a historical meteorological feature vector. Simultaneously, the temperature and humidity forecast values, wind speed fluctuation forecast range, and rainfall forecast range for the target forecast period are acquired and combined in the same order to form a future meteorological feature vector. Extract the temperature and humidity values, wind speed fluctuation amplitude, and rainfall range corresponding to each sampling time from the historical meteorological feature vector. Read the high-frequency fluctuating load sequence of the corresponding node at the same sampling time. Match the historical meteorological feature vector and high-frequency fluctuating load sequence item by item according to the sampling time. Statistically analyze the matching of each meteorological component with the load fluctuation direction, load fluctuation amplitude, and load fluctuation duration. Assign the influence weight of each meteorological component according to the degree of matching. Map the influence weight to the future meteorological feature vector. Summarize the influence values ​​of each forecast meteorological component on the future load fluctuation at each forecast time to obtain the initial load fluctuation increment.

[0027] Specifically, based on the collected temperature and humidity values, wind speed fluctuations, and rainfall ranges, these meteorological data are first spatially matched with the distribution network topology nodes. For each node, the inverse distance weighting (IDW) method is used to obtain data from the three nearest meteorological stations, and a weighted average is calculated to obtain the estimated meteorological value for that node's location. For example, the estimated temperature for node A... ,in It is the temperature of the i-th weather station, with weights. , This is the distance from node A to the i-th weather station. Then, the meteorological data of all nodes at all historical moments are aligned and combined. For example, for node A at time t, it is combined into a four-dimensional vector in a fixed order of [temperature, humidity, wind speed change amplitude, rainfall range], such as [28.5℃, 75%, 3m / s, 5mm]. The vectors of all moments are stacked to form the historical meteorological feature vector sequence of the node. The wind speed change amplitude and rainfall range are calculated based on the original hourly meteorological data. For example, the wind speed change amplitude is the absolute value of the difference between the wind speed of the current hour and the wind speed of the previous hour, and the rainfall range is the difference between the maximum rainfall intensity and the minimum rainfall intensity of the hour. Simultaneously, the hourly temperature and humidity forecast values, wind speed forecast values, and rainfall forecast values ​​for the target prediction period (e.g., the next 24 hours) are obtained from a third-party meteorological service interface, and the wind speed change forecast amplitude and rainfall range are obtained using the same calculation method. Then, they are combined in the same order and format as the historical data to form the future meteorological feature vector, forming the historical meteorological feature vector.

[0028] The temperature and humidity values, wind speed fluctuation amplitude, and rainfall range corresponding to each sampling time are extracted from the historical meteorological feature vector. The high-frequency fluctuating load sequence of the corresponding node at the same sampling time is also read. To quantify the impact of each meteorological component on load fluctuation, a multiple linear regression model is used for fitting. A separate model is built for each node, and the model form is as follows: ,in, It is the high-frequency fluctuating load value at time t. These represent the magnitude of the sudden changes in temperature, humidity, wind speed, and the range of rainfall at time t. It is the intercept term. These are the regression coefficients of each meteorological component, i.e., the weights of their effects to be determined. This is the error term. The values ​​of these influence weights are estimated using historical data (e.g., synchronous data from the past year) through the least squares method. For example, for a given node, the calculated influence weights are [temperature: 0.5, humidity: 0.1, sudden wind speed change: 0.3, rainfall range: 0.2]. These weight values ​​reflect the average change in high-frequency fluctuating load for each meteorological component changing by one unit. Next, these calculated influence weights are directly mapped to the future meteorological feature vector. For each future forecast time, the predicted meteorological component value at that time (e.g., temperature forecast of 25℃) is multiplied by the corresponding influence weight (e.g., 0.5), and then all product terms are summed. ,in It is the weather forecast value at the predicted time in the future, calculated from... This is the initial load fluctuation increment caused by meteorological factors at the predicted time for that node, thus obtaining the initial load fluctuation increment.

[0029] The steps for obtaining the incremental value of meteorological disturbance load are as follows: Read the matrix position values ​​of each node in the spatial transfer probability feature matrix, map the initial load fluctuation increment to the spatial transfer probability feature matrix item by item according to the node connection order, calculate the transmission value of the initial load fluctuation increment of each node to adjacent nodes and related nodes, accumulate the diffusion results in each transmission direction, extract the average value of the diffusion results within the sliding time window according to the continuous prediction time, and obtain the meteorological interference load increment value.

[0030] Specifically, the previously calculated spatial transition probability feature matrix is ​​read, and each element of this matrix... This represents the probability that the load fluctuation of node i will affect node j. Simultaneously, the initial load fluctuation increments of all nodes at each prediction time are obtained, forming an initial increment vector. Each element represents the initial load fluctuation increment value of the corresponding node. To calculate the diffusion effect of load fluctuations in the power grid, the initial increment vector is iteratively multiplied with the spatial transfer probability feature matrix to simulate the transmission process of the influence. The load increment vector after the first transmission is... After the second transmission, it is And so on, the increment vector after k iterations is The number of iterations k is set based on the average convergence time of power grid disturbance propagation obtained from historical data analysis. For example, setting k=5 means that the influence of up to 5 levels of adjacent nodes is considered. The incremental vectors obtained from the 1st to the kth iteration are accumulated to obtain the total incremental vector considering the multi-level diffusion effect. Each element of the vector represents the total cumulative impact on the corresponding node throughout the diffusion process. Subsequently, in order to smooth short-term random fluctuations, a sliding time window mean filter is applied to the total incremental time series of each node. The size of the sliding time window is set according to the prediction step size and load fluctuation characteristics. For example, for a prediction step size of 15 minutes, a sliding window with a length of 3 (i.e., 45 minutes) can be selected. The average value of the diffusion results within the window is calculated, and this average value is used as the final meteorological disturbance load increment value of the node at the prediction time, thus obtaining the meteorological disturbance load increment value.

[0031] The steps to obtain the short-term forecast total load value are as follows: The evolution trend of the low-frequency reference load sequence of each node is extrapolated using a time series prediction algorithm to obtain the future low-frequency reference prediction sequence of each node in the target prediction period. Based on the equipment capacity setpoint of each distribution transformer topology node, the equipment capacity setpoint, meteorological interference load increment value, and future low-frequency reference prediction sequence of each distribution transformer topology node are read one by one to form the node load superposition correspondence. Based on the corresponding relationship of node load superposition, according to the prediction time sequence of each distribution transformer topology node, the incremental value of meteorological interference load is extracted time by time, and the load value in the future low-frequency reference prediction sequence is extracted time by time. The incremental value of meteorological interference load at each prediction time is superimposed on the load value in the future low-frequency reference prediction sequence of the same distribution transformer topology node and the same prediction time. The node load results after superposition at each prediction time are recorded to obtain the total short-term predicted load value of each node.

[0032] Specifically, the Prophet time series forecasting algorithm is used to independently extrapolate the evolution trend of the low-frequency reference load sequence for each node. This algorithm can automatically handle holiday effects and periodic changes, and fit the historical low-frequency reference load sequence (e.g., data from the past year) for each node. The model learns the load change pattern by decomposing the time series into trend, seasonal, and holiday terms, and predicts the load value for the target period (e.g., the next 24 hours) in advance, thereby obtaining the future low-frequency reference forecast sequence for each node in the target forecast period. Then, a data structure is established to store all relevant forecast information for each distribution transformer topology node. The equipment capacity setting (e.g., 500kVA) on the equipment nameplate of each node is read one by one, as well as the meteorological interference load increment value sequence calculated in the previous step, which is perfectly aligned with the future low-frequency reference forecast sequence in time. The equipment capacity setting, the future low-frequency reference forecast sequence, and the meteorological interference load increment value sequence of the same node are stored together. For example, an object containing node ID, equipment capacity, an array of future low-frequency reference forecast sequences, and an array of meteorological interference load increment value sequences is used to organize them, forming a node load superposition correspondence.

[0033] Based on the corresponding relationship of node load superposition, each distribution transformer topology node is calculated independently, and the calculations are performed in chronological order of the predicted times, starting from the time point. Start by processing each time point up to the last time point of the target prediction period. At every prediction time First, the incremental value of meteorological disturbance load at that moment is extracted from the node load superposition correspondence of that node. Simultaneously, the load values ​​in the future low-frequency reference prediction sequence at that moment are extracted. Then, add the two values ​​together using the following formula: ,in That is, at time The short-term forecast total load value, this superposition process, such as load fluctuations caused by weather, is added to the basic load trend, and the calculated value for each forecast time point is... The node load results after superposition Record the data in chronological order to form a complete predicted load time series. Repeat this process for all nodes to obtain a set of short-term predicted total load values ​​for each time period within the target prediction period, covering all nodes of the distribution network in the entire region. This yields the short-term predicted total load value for each node.

[0034] The steps for obtaining a regional power distribution network load forecasting scheme are as follows: Based on the short-term predicted total load value of each node, the overload setting value of the equipment corresponding to each distribution transformer topology node is read one by one. The short-term predicted total load value and the equipment overload setting value of each node are compared item by item according to the distribution transformer topology node number. The node records that have not exceeded the equipment overload setting value are extracted, and the node records that have reached or exceeded the equipment overload setting value are extracted. The over-limit status, corresponding prediction time and corresponding load value of each distribution transformer topology node are summarized to obtain the regional distribution network load prediction scheme.

[0035] Specifically, based on the short-term predicted total load value of each node, the overload setting value corresponding to each distribution transformer topology node is first read from the pre-stored equipment file. This overload setting value is set according to the equipment's technical specifications and operating procedures. For example, for long-term operation, the overload setting value is usually 100% of its rated capacity, while for short-term operation (e.g., 15 minutes), it may be allowed to reach 120%. Here, 120% is used as the equipment overload setting value. For example, for a 500kVA transformer, its equipment overload setting value is 600kVA. Then, the process iterates through the distribution transformer topology node numbers, comparing the time series of the short-term predicted total load value of each node at each predicted time. If at a certain predicted time... The short-term predicted total load value of the node If the load is less than the equipment overload setting, the node's status at that moment is marked as "normal," and the record (containing the node number, prediction time, and prediction load value) is added to the set of node records that have not exceeded the equipment overload setting. Conversely, if the load is greater than the equipment overload setting, the node's status at that moment is marked as "normal," and the record is added to the set of node records that have not exceeded the equipment overload setting. If the load value is greater than or equal to the equipment overload setting, the node's status at that moment is marked as "overloaded" and the record is added to the node record set that reaches or exceeds the equipment overload setting. After comparing all predicted times for all nodes, all node records that reach or exceed the equipment overload setting are summarized to generate a list containing the number of each overloaded node, the specific predicted time of the overload, and the corresponding load value. This list constitutes the final early warning information, resulting in the regional distribution network load prediction scheme.

Claims

1. A regional distribution network short-term load intelligent forecasting system based on big data, characterized in that, The system includes: The frequency domain filtering and load mode extraction module extracts and generates a high-frequency fluctuating load sequence and a low-frequency reference load sequence based on the location coordinates of the distribution transformer topology nodes and the initial power sequence collected by the smart meter at high frequency. The spatial load correlation calculation module concatenates the high-frequency fluctuation load sequences of the corresponding nodes sequentially and calculates the node collaborative distribution correlation degree; it then calculates the node collaborative distribution correlation degree with a preset probability distribution value to construct a spatial transfer probability feature matrix. The meteorological feature cross-mapping module constructs a historical meteorological feature vector based on the collected temperature and humidity values, wind speed abrupt change amplitude, and rainfall range. This historical meteorological feature vector is then fitted with the high-frequency fluctuation load sequence of the corresponding node to obtain the node's meteorological influence weight. Combining this with the future meteorological feature vector composed of meteorological forecast data for the target prediction period and the node's meteorological influence weight, the module obtains the initial load fluctuation increment for each node. Finally, by combining the spatial transition probability feature matrix, the module solves for the initial load fluctuation increment of each node to obtain the meteorological disturbance load increment value for each node. The regional load joint forecasting module performs trend extrapolation on the low-frequency reference load sequence of each node to obtain a regional distribution network load forecasting scheme.

2. The regional distribution network short-term load intelligent forecasting system based on big data according to claim 1, characterized in that, The steps for obtaining the high-frequency fluctuation load sequence and the low-frequency reference load sequence are as follows: Based on the location coordinates of the distribution transformer topology nodes and the initial power sequence collected by the smart meter at high frequency, the initial power sequence collected by the smart meter at high frequency is checked by node-by-node correspondence, sampling time-by-sampling time alignment, and sampling point-by-sampling point according to the node number, node arrangement order, and node positioning mark in the location coordinates of the distribution transformer topology nodes. Power records with mismatched node numbers are eliminated, and power records with misaligned sampling times are corrected to form the node coordinate mapping initial power sequence. Based on the node coordinates, the initial power sequence is mapped, the power value at each sampling time is read node by node, the frequency components of each sampling segment are expanded according to the sampling order, the frequency components with continuous amplitude changes in continuous sampling segments are extracted, the frequency components with abnormal phase jump variables in adjacent sampling segments are screened out, and the reverse reconstruction and power value backfilling are performed according to the original sampling time order to generate a denoised power time series. Based on the denoised power time series, the power rise inflection point, power fall inflection point, local peak point and local valley point are identified segment by segment. The envelope boundary is connected based on the local peak point and local valley point. The load components are split according to the distance between adjacent extreme points and the length of continuous oscillation period. The load components that change continuously across multiple sampling periods are extracted. The load components that fluctuate repeatedly in adjacent sampling periods are extracted to obtain the high-frequency fluctuating load sequence and the low-frequency reference load sequence.

3. The regional distribution network short-term load intelligent forecasting system based on big data according to claim 1, characterized in that, The steps for obtaining the node collaborative distribution correlation degree are as follows: The high-frequency fluctuating load sequences of the corresponding nodes are spliced ​​together sequentially according to the distribution topology node numbers. The sampling time order and node affiliation mark of the high-frequency fluctuating load sequences of each node are checked. The load values ​​of each node are extracted according to the unified sampling time. The load values ​​of each node are sorted from small to large at each sampling time. The changes in the sorting position of each node, the number of position swaps between adjacent nodes, and the number of cross-position moves are recorded at each sampling time to form the node sequence swapping statistics. Based on the node sequence swapping statistics, the ranking position of each node is extracted at all sampling times. The number of times the ranking changes in the same direction, the number of times the ranking changes in opposite directions, and the distribution of the ranking position difference between any two nodes at each sampling time are counted. The rank correlation coefficient is calculated based on the ranking position change relationship. The synchronous occurrence records of each node in the first and last ranking intervals are further screened. The common occurrence frequency and continuous maintenance frequency of extreme ranking positions are counted to determine the extreme value of tail dependency. The node cooperative distribution correlation degree is obtained by combining the rank correlation coefficient and the extreme value of tail dependency.

4. The regional distribution network short-term load intelligent forecasting system based on big data according to claim 1, characterized in that, The steps for obtaining the spatial transition probability feature matrix are as follows: The correlation degree of each group of nodes is multiplied according to the preset probability distribution of each item. According to the connection order from the node's starting position to the node's ending position, each product item is continuously integrated and the integration result is extracted to obtain the numerical item of the integration result of each node. Then, according to the row position and column position of the distribution topology node position, each numerical item of the integration result is filled in to obtain the spatial transition probability feature matrix.

5. The regional distribution network short-term load intelligent forecasting system based on big data according to claim 1, characterized in that, The steps for obtaining the initial load fluctuation increment are as follows: Based on the collected temperature and humidity values, wind speed fluctuation range, and rainfall range, the temperature and humidity values, wind speed fluctuation range, and rainfall range are mapped node by node according to a unified sampling time. The temperature and humidity values, wind speed fluctuation range, and rainfall range are combined in a fixed order at the same node and time to form a historical meteorological feature vector. Simultaneously, the temperature and humidity forecast values, wind speed fluctuation forecast range, and rainfall forecast range for the target forecast period are acquired and combined in the same order to form a future meteorological feature vector. Extract the temperature and humidity values, wind speed fluctuation amplitude, and rainfall range corresponding to each sampling time in the historical meteorological feature vector. Read the high-frequency fluctuating load sequence of the corresponding node at the same sampling time. Match the historical meteorological feature vector and the high-frequency fluctuating load sequence item by item according to the sampling time. Statistically analyze the matching of each meteorological component with the load fluctuation direction, load fluctuation amplitude, and load fluctuation duration. Assign the influence weight of each meteorological component according to the degree of matching. Map the influence weight to the future meteorological feature vector. Summarize the influence values ​​of each forecast meteorological component on the future load fluctuation at each forecast time to obtain the initial load fluctuation increment.

6. The regional distribution network short-term load intelligent forecasting system based on big data according to claim 1, characterized in that, The steps for obtaining the incremental value of the meteorological disturbance load are as follows: Read the matrix position values ​​of each node in the spatial transfer probability feature matrix, map the initial load fluctuation increment to the spatial transfer probability feature matrix item by item according to the node connection order, calculate the transmission value of the initial load fluctuation increment of each node to adjacent nodes and associated nodes, accumulate the diffusion results in each transmission direction, extract the average value of the diffusion results within the sliding time window according to the continuous prediction time, and obtain the meteorological interference load increment value.

7. The regional distribution network short-term load intelligent forecasting system based on big data according to claim 1, characterized in that, The steps for obtaining the short-term predicted total load value are as follows: A time series forecasting algorithm is used to extrapolate the evolution trend of the low-frequency reference load sequence of each node to obtain the future low-frequency reference forecast sequence of each node in the target forecast period. Based on the equipment capacity setpoint of each distribution transformer topology node, the equipment capacity setpoint of each distribution transformer topology node, the incremental value of the meteorological interference load, and the future low-frequency reference forecast sequence are read one by one to form a node load superposition correspondence. Based on the node load superposition correspondence, according to the prediction time sequence of each distribution transformer topology node, the incremental value of meteorological interference load is extracted time by time, and the load value in the future low-frequency reference prediction sequence is extracted time by time. The incremental value of meteorological interference load at each prediction time is superimposed on the load value in the future low-frequency reference prediction sequence of the same distribution transformer topology node and the same prediction time. The node load results after superposition at each prediction time are recorded to obtain the total short-term predicted load value of each node.

8. The regional distribution network short-term load intelligent forecasting system based on big data according to claim 1, characterized in that, The steps for obtaining the regional distribution network load forecasting scheme are as follows: Based on the short-term predicted total load value of each node, the overload setting value of the equipment corresponding to each distribution transformer topology node is read one by one. The short-term predicted total load value and the equipment overload setting value of each node are compared item by item according to the distribution transformer topology node number. Node records that have not exceeded the equipment overload setting value are extracted, and node records that have reached or exceeded the equipment overload setting value are extracted. The over-limit status, corresponding prediction time and corresponding load value of each distribution transformer topology node are summarized to obtain the regional distribution network load prediction scheme.

Images