A power distribution network power supply capacity intelligent prediction method and system

Abstract

The application is suitable for the power supply prediction technical field, and provides an intelligent power supply capacity prediction method for a power distribution network, comprising: acquiring power supply data; decomposing and processing the power supply data by using a time series decomposition algorithm to obtain short-term fluctuation components and long-term trend components; smoothing processing the short-term fluctuation components by using an adaptive filter to obtain smooth time series features; performing feature extraction processing on the smooth time series features by using a convolutional neural network to obtain a multi-dimensional feature vector; extracting preset key features from the multi-dimensional feature vector, performing data fusion processing on the preset key features by using a weighted aggregation algorithm, and determining feature weights corresponding to the preset key features; generating an initial prediction distribution according to the feature weights and the preset key features; and performing power supply prediction according to the initial prediction distribution to generate a power supply prediction result, so that accurate prediction of the power supply capacity of the power distribution network can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power supply prediction, and in particular to an intelligent power supply capacity prediction method and system for a power distribution network. BACKGROUND

[0002] Power supply capacity prediction for a power distribution network refers to a technology for estimating the maximum power load that can be safely and stably supplied by the power distribution network in a future time period by analyzing historical operation data, real-time state, and external influencing factors of the power distribution network. The core goal is to ensure that the power grid meets user electricity demand while avoiding overload or power supply shortage, thereby improving power supply reliability and energy utilization efficiency.

[0003] In traditional schemes, a traditional model method is mainly used, such as establishing a linear or nonlinear mathematical model to evaluate power supply capacity. This method has a large difference from actual load distribution, which may cause the evaluation result to deviate from the actual demand. Alternatively, a traditional optimization algorithm (such as a genetic algorithm or a particle swarm algorithm) is used for prediction, but this method is prone to local optimization, affecting the prediction accuracy. Therefore, how to accurately predict the power supply capacity of the power distribution network has become a technical problem to be solved. SUMMARY

[0004] The embodiments of the present application provide an intelligent power supply capacity prediction method, device, and equipment for a power distribution network and a storage medium, which can solve the technical problem of how to accurately predict the power supply capacity of the power distribution network.

[0005] In a first aspect, the embodiments of the present application provide an intelligent power supply capacity prediction method for a power distribution network, which comprises:

[0006] Obtaining power supply data, the power supply data comprising at least one of the following: substation operation data, feeder load data, and user electricity consumption data;

[0007] Decomposing and processing the power supply data using a time series decomposition algorithm to obtain a short-term fluctuation component and a long-term trend component, the short-term fluctuation component being used to represent short-term changes in the power supply data, and the long-term trend component being used to represent long-term changes in the power supply data;

[0008] Smoothing the short-term fluctuation component using an adaptive filter to obtain a smoothed time series feature, wherein the adaptive filter adjusts the corresponding filter parameters according to the long-term trend component;

[0009] Extracting features from the smoothed time series feature using a convolutional neural network to obtain a multi-dimensional feature vector, the multi-dimensional feature vector being used to represent the association between the substation operation data, the feeder load data, and the user electricity consumption data;

[0010] The preset key features are extracted from the multi-dimensional feature vector, and a weighted aggregation algorithm is used for data fusion processing of the preset key features to determine the feature weight corresponding to the preset key features;

[0011] An initial prediction distribution is generated according to the feature weight and the preset key feature;

[0012] Power supply prediction is performed according to the initial prediction distribution, and a power supply prediction result is generated.

[0013] In one embodiment, the smooth time series is divided into a first time period and a second time period according to the data change rate, and the first time period is smaller than the second time period;

[0014] The convolutional neural network includes at least two convolutional layers, and the convolutional neural network is used for feature extraction processing of the smooth time series features to obtain a multi-dimensional feature vector, including:

[0015] The first convolutional layer is used for feature extraction of the smooth time series corresponding to the first time period to obtain a first feature vector;

[0016] The second convolutional layer is used for feature extraction of the smooth time series corresponding to the second time period to obtain a second feature vector, and the size of the convolution kernel of the first convolutional layer is smaller than the size of the second convolution kernel;

[0017] The multi-dimensional feature vector is obtained according to the first feature vector and the second feature vector.

[0018] In one embodiment, the short-term fluctuation component is smoothed by using an adaptive filter to obtain a smooth time series feature, including:

[0019] The long-term trend direction and the data change rate are determined according to the long-term trend component;

[0020] The filter window size and the filter weight coefficient of the adaptive filter are determined according to the long-term trend direction and the data change rate;

[0021] The short-term fluctuation component is smoothed according to the filter window size and the filter weight coefficient to obtain the smooth time series feature.

[0022] In one embodiment, the short-term fluctuation component is smoothed according to the filter window size and the filter weight coefficient to obtain the smooth time series feature, including:

[0023] The short-term data change rate is calculated according to the short-term fluctuation component, the filter window size of the adaptive filter is optimized according to the short-term data change rate to obtain an optimized filter window size; or,

[0024] The fluctuation time period is determined according to the short-term fluctuation component, and the filter window size of the adaptive filter is optimized according to the fluctuation time period, so as to obtain an optimized filter window size.

[0025] The short-term fluctuation component is smoothed according to the optimized filter window size, so as to obtain a smoothed time sequence feature.

[0026] In an embodiment, the power supply data is decomposed by using a time sequence decomposition algorithm to obtain the short-term fluctuation component and the long-term trend component, which includes:

[0027] The change range of the substation operation data is determined according to the short-term fluctuation component, so as to obtain a stability distribution.

[0028] The electricity peak-valley difference of the user electricity consumption data is determined according to the long-term trend component, so as to obtain a peak-valley difference regularity distribution.

[0029] The DBSCAN clustering algorithm is used to perform abnormal fluctuation elimination processing on the stability distribution and the peak-valley difference regularity distribution, so as to obtain a normal fluctuation range.

[0030] The normal short-term fluctuation component and the normal long-term trend component are determined according to the normal fluctuation range.

[0031] In an embodiment, after the initial prediction distribution is generated according to the feature weight and the preset key feature, the method further includes:

[0032] The initial prediction distribution is compared with a preset distribution threshold value, and if the initial prediction distribution is greater than the preset distribution threshold value, the feature weight is adjusted by using a gradient optimization algorithm to obtain an optimized feature weight.

[0033] The optimized prediction distribution is generated according to the optimized feature weight and the preset key feature.

[0034] The power supply prediction is performed according to the optimized prediction distribution, so as to generate a power supply prediction result.

[0035] In an embodiment, the power supply prediction is performed according to the optimized prediction distribution, so as to generate a power supply prediction result, which includes:

[0036] The user electricity consumption density of different power supply regions is calculated according to the optimized prediction distribution.

[0037] The grid division is performed on different power supply regions according to the user electricity consumption density, so as to generate a set of sampling points, wherein each grid represents a sampling point, and the size of the grid is in a negative correlation relationship with the user electricity consumption density.

[0038] The inverse distance weighted interpolation algorithm is used to calculate the predicted load value of an unknown sampling point by weighting the load value of a known sampling point, so as to obtain the power supply prediction result.

[0039] In one embodiment, the method further comprises:

[0040] constructing a spatial gradient matrix according to the load values of the known sampling points and the predicted load values of the unknown sampling points;

[0041] performing cluster analysis on the spatial gradient matrix to obtain a plurality of classification regions;

[0042] determining a prediction priority of the plurality of classification regions according to the fluctuation sensitivity of the plurality of classification regions;

[0043] sequentially performing power supply prediction according to the prediction priority of the plurality of classification regions to generate power supply prediction results corresponding to the plurality of classification regions.

[0044] In one embodiment, the method further comprises:

[0045] comparing the power supply prediction results corresponding to the plurality of classification regions with actual power supply values corresponding to the plurality of classification regions to determine an update time range corresponding to the plurality of classification regions;

[0046] performing smoothing processing on the smoothed time series feature in the update time range to obtain an updated smoothed time series;

[0047] inputting the updated smoothed time series into a convolutional neural network for feature extraction processing to obtain an updated multi-dimensional feature vector.

[0048] In a second aspect, embodiments of the present application provide an intelligent power supply capacity prediction system for a power distribution network. The system has the function of implementing the method in the first aspect or any possible implementation manner thereof. Specifically, the system includes units for implementing the method in the first aspect or any possible implementation manner thereof.

[0049] In one embodiment, the system comprises:

[0050] an acquisition unit configured to acquire power supply data, the power supply data including at least one of the following: substation operation data, feeder load data, and user power consumption data;

[0051] a decomposition unit configured to perform decomposition processing on the power supply data using a time series decomposition algorithm to obtain a short-term fluctuation component and a long-term trend component, the short-term fluctuation component being used to represent short-term changes in the power supply data, and the long-term trend component being used to represent long-term changes in the power supply data;

[0052] a smoothing unit configured to perform smoothing processing on the short-term fluctuation component using an adaptive filter to obtain a smoothed time series feature, wherein the adaptive filter adjusts corresponding filter parameters according to the long-term trend component;

[0053] The convolution unit is used to extract the features of the smoothed time series using a convolutional neural network to obtain a multidimensional feature vector. The multidimensional feature vector is used to represent the correlation between the substation operation data, the feeder load data, and the user power consumption data.

[0054] A fusion unit is used to extract preset key features from the multidimensional feature vector, perform data fusion processing on the preset key features using a weighted aggregation algorithm, and determine feature weights corresponding to the preset key features;

[0055] A calculation unit, used to generate an initial prediction distribution based on feature weights and preset key features;

[0056] The prediction unit is used to perform power supply prediction based on the initial prediction distribution and generate a power supply prediction result.

[0057] In a third aspect, an embodiment of the present application provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the computer device implements any one of the methods of the first aspect described above.

[0058] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a computer device, the computer device implements the method of any one of the implementation methods of the above-mentioned first aspect.

[0059] In a fifth aspect, an embodiment of the present application provides a computer program product, which, when executed on a computer device, enables the computer device to execute any one of the methods of the first aspect.

[0060] Compared with the prior art, the embodiments of the present application have the following beneficial effects: obtaining comprehensive power supply data, using a time series decomposition algorithm to decompose the power supply data into short-term fluctuation components and long-term trend components, the decomposed data can more clearly show the change patterns at different time scales, and help subsequent models capture the potential relationship between features; using adaptive filtering to dynamically filter out short-term noise, retain long-term change trends, improve the data signal-to-noise ratio, and prevent the model from overfitting due to noise; using a convolutional neural network (CNN) for feature extraction, it can mine complex nonlinear relationships and deeply learn the spatiotemporal correlation between substations, feeders, and user electricity consumption; performing key feature extraction and weighted aggregation can screen features that have a significant impact on power supply predictions, and quantify their contributions through weight allocation, avoiding irrelevant features from interfering with predictions and focusing on core factors; comprehensively generating a probability distribution based on weighted features, outputting a final prediction value, and realizing accurate prediction of the power supply capacity of the distribution network. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1is a flowchart of an intelligent power supply capacity prediction method of a power distribution network provided by an embodiment of the present application.

[0062] Figure 2 is a structure diagram of a convolutional neural network provided by an embodiment of the present application.

[0063] Figure 3 is a structure diagram of an intelligent power supply capacity prediction system of a power distribution network provided by an embodiment of the present application.

[0064] Figure 4 is a structure diagram of a computer device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0065] In order to further illustrate the technical solutions of the present application, specific embodiments are described below.

[0066] Figure 1 is a flowchart of an intelligent power supply capacity prediction method of a power distribution network provided by an embodiment of the present application.

[0067] As shown in Figure 1 , the above method comprises the following steps S101-S107.

[0068] S101, obtaining power supply data.

[0069] The power supply data comprises at least one of the following: substation operation data, feeder load data, and user power consumption data.

[0070] The substation operation data refers to various monitoring and operation data generated by the substation during operation, including but not limited to voltage data, current data, power data, and equipment operation state data (such as the opening and closing states of circuit breakers and disconnectors). These data reflect the operating conditions of the substation equipment.

[0071] The feeder load data refers to the load conditions of each feeder from the substation outlet to the user end during operation, including but not limited to load rate (the ratio of actual load to rated load of the feeder), current peak value (the maximum current value of the feeder within a certain time), active power and reactive power, etc.

[0072] The user power consumption data refers to various power consumption information of power users during power use, including but not limited to power consumption, power consumption time data, user load curve (the change of user power load at different time points), user category data (such as industrial users, commercial users, residential users, etc. different categories), and user spatial location.

[0073] As an example but not limitation, multi-source data is obtained from substation operation data, feeder load and user electricity consumption, heterogeneous data types are identified by a classifier, and a classification result is obtained. According to the classification result, data cleaning is performed on the heterogeneous data to eliminate format differences, and cleaned data is obtained. Through standardization processing, the cleaned data is converted into a unified data set to determine data consistency. If the substation operation data and the feeder load have abnormalities in the unified data set, the abnormal values are separated by a clustering algorithm to obtain a normal data set.

[0074] For example, substation operation data may include voltage, current and power factor, feeder load involves line load rate and peak current, and user electricity consumption data covers electricity consumption and electricity consumption period. These data sources are diverse and have different formats, such as substation operation data may be a structured table collected in real time, feeder load may be time series data, and user electricity consumption may be scattered log records. Through the classifier to identify the heterogeneous data types, it can be understood as using machine learning algorithms such as decision trees or support vector machines to classify these data into structured, semi-structured and unstructured categories.

[0075] For example, voltage data is classified as structured data, and user electricity consumption logs may be unstructured text, and the classification result provides a clear basis for subsequent processing.

[0076] In one possible implementation, when data cleaning is performed on heterogeneous data, eliminating format differences is key. Assuming that in the substation operation data record, the voltage unit is "kV" or "V", it needs to be uniformly converted to "kV" during cleaning, such as converting 220V to 0.22kV. Feeder load data may have missing values, such as a certain period of load rate being empty, which can be filled by the average value of adjacent periods. There may be duplicate records in user electricity consumption data, such as a user's electricity consumption being recorded twice on a certain day, which needs to be cleaned. The cleaned data has a consistent format, which is convenient for further processing, which can effectively improve the data quality and avoid analysis deviation caused by format problems.

[0077] Specifically, the cleaned data is converted into a unified data set through standardization processing to ensure data consistency.

[0078] For example, the voltage is standardized to the range of 0 to 1, assuming that the original voltage range is 0.22kV to 110kV, a value such as 11kV can be calculated by linear normalization processing, the result is (11-0.22) / (110-0.22)≈0.1. If the load rate of the feeder load is in percentage form, such as 50%, it can be directly converted to 0.5. If the user electricity consumption has different units, such as some in kWh and some in Wh, it needs to be unified to kWh. This standardization not only eliminates the dimensional differences, but also facilitates the application of subsequent algorithms, improving the compatibility of data processing.

[0079] It should be noted that if the substation operation data and the feeder load are abnormal in the unified data set, the abnormal values are separated by the clustering algorithm.

[0080] Preferably, the K-means clustering algorithm can be used.

[0081] For example, assuming that in the voltage data of a certain substation, most values are around 10kV, and a sudden change is 50kV, which is far beyond the normal range. The feeder load rate usually fluctuates between 0.3 and 0.7, but a value reaches 0.95, which is obviously abnormal. After clustering, these abnormal values are separated, and the normal data set retains the typical values of 10kV and 0.3 to 0.7. Separating abnormal values helps to identify potential faults such as transformer overload or line short circuit, thereby improving prediction reliability.

[0082] S102, decompose the power supply data by using a time series decomposition algorithm to obtain short-term fluctuation components and long-term trend components.

[0083] The short-term fluctuation component is used to represent the short-term change of the power supply data; the long-term trend component is used to represent the long-term change of the power supply data.

[0084] The time series decomposition algorithm is an analysis method that decomposes time series data into multiple components. Through time series decomposition, short-term changes, long-term trends and periodic patterns in time series can be more accurately captured, thereby improving the accuracy of prediction.

[0085] The time series decomposition method is used to process the unified data set, separate the short-term fluctuation and the long-term trend, and obtain the fluctuation component and the trend component.

[0086] The voltage stability is analyzed by the short-term fluctuation component to determine the change range of the substation operation data, and the stability distribution is obtained.

[0087] According to the long-term trend component, the peak-valley difference is extracted, the periodic characteristics of the user end data are obtained, and the regularity of the peak-valley difference is judged.

[0088] For example, the short-term fluctuation reflects the instantaneous change of voltage or load, while the long-term trend reveals the overall trend of electricity demand.

[0089] In one possible implementation, the decomposition can be realized by moving average method. Assuming that the voltage data of a certain substation contains 1440 points collected daily, the short-term fluctuation may be a minute-level mutation, such as the voltage jumping from 10kV to 10.5kV, while the trend component shows the regularity of the slow decline of the voltage within a week. When analyzing the voltage stability by the fluctuation component, the frequent change range of the voltage can be focused on.

[0090] Specifically, the voltage fluctuation of a certain substation within a day is between 9.8 kV and 10.2 kV. By statistically analyzing the frequency and amplitude of these fluctuations, a stability distribution can be obtained, such as 90% of the time the voltage is maintained within 10 kV ± 0.1 kV. This distribution can intuitively reflect the smoothness of the substation operation.

[0091] It should be noted that if the fluctuation range suddenly expands to 9.5 kV to 10.8 kV on a certain day, it may indicate equipment aging or external interference. When extracting the peak-valley difference of electricity consumption based on the trend component, the focus is on capturing periodic characteristics.

[0092] For example, after decomposing the electricity consumption data of a certain area user, the trend component shows that the electricity consumption rises from 50 kWh at 8 am to a peak of 200 kWh, and falls to 80 kWh at 6 pm, with a peak-valley difference of 150 kWh. This regularity can be verified by multi-day data, and if the peak-valley difference fluctuates between 140 kWh and 160 kWh every Monday, it indicates that the periodicity is stable.

[0093] S103, smoothing the short-term fluctuation component using an adaptive filter to obtain a smoothed time series feature.

[0094] Among them, the adaptive filter adjusts the corresponding filter parameters according to the long-term trend component.

[0095] Adaptive filtering is a processing method that can dynamically adjust parameters according to data characteristics, especially suitable for processing noise in short-term fluctuation components while preserving key change directions.

[0096] The adaptive filter takes time series data as input and performs smoothing processing on the short-term fluctuation component by dynamically adjusting the filter coefficients. It can automatically weaken the influence of noise and irregular fluctuations, making the output signal smoother and easier to analyze.

[0097] During the smoothing process, the filter identifies the location and change trend of the feeder load peak based on the electricity consumption regularity reflected by the substation voltage stability and the user's electricity peak-valley difference. For example, if there is a significant upward trend in the feeder load in a certain time period, the adaptive filter will smooth the short-term fluctuations while ensuring that the direction of this upward trend is preserved. This way, the key change characteristics of the feeder load can be accurately captured.

[0098] By smoothing the short-term fluctuation component using an adaptive filter, the change direction of the feeder load peak is preserved, and the smoothed time series feature is obtained.

[0099] For example, in the feeder load data, short-term fluctuations may include transient jumps caused by switch operations or sudden demand, and adaptive filtering can smooth these noises and highlight the trend of the load peak.

[0100] In one possible implementation, assuming that the load data collected by a certain feeder is 1000 points per day, there can be minute-level spikes in the original data, such as a sudden increase from 500 kW to 550 kW and a rapid drop. The adaptive filter can reduce the burr effect of such spikes by adjusting the dynamic window, and retain the true change direction from 500 kW to 520 kW.

[0101] Specifically, the implementation of the adaptive filter can be based on the principle of weighted moving average, and automatically adjust the weight according to the fluctuation degree of historical data.

[0102] For example, the load of a certain feeder increases from 480 kW to 510 kW at 8 am, with 10 kW of random jitter in between. After the filter detects the high-frequency characteristics of these jitters, it reduces the weight, and finally outputs a smoothed sequence showing the process of a smooth transition from 480 kW to 510 kW.

[0103] It can be understood that this smoothing retains the trend of the peak rise, facilitating the analysis of the change law of the load demand.

[0104] In one embodiment, the peak value of the feeder load can vary due to weather changes, for example, high temperature on a certain day causes the air conditioning load to increase, and the data jumps from 600 kW to 650 kW for a short time and then stabilizes at 630 kW. The adaptive filter smoothes to a sequence from 600 kW to 630 kW gradually by identifying the transience of the jump, clearly presenting the peak direction.

[0105] Preferably, this method can also adapt to the fluctuation characteristics of different time periods, such as smaller load fluctuations at night, and the filter window is automatically reduced to ensure that the details are not excessively smoothed.

[0106] In one embodiment, the implementation of the adaptive filter can be based on the fluctuation amplitude of historical data to adjust the weight. The load of a certain feeder is stable at about 460 kW at night, with only 5 kW of small fluctuations, and the filter window is automatically reduced to retain the details; when the load jumps to 530 kW during the daytime peak, the window is enlarged to smooth the 20 kW spikes. This flexibility ensures the adaptability of data at different time periods.

[0107] It should be noted that the key to retaining the change direction lies in the adaptability of the filter.

[0108] For example, the load of a certain feeder increases from 450 kW to a peak of 520 kW and then falls to 470 kW within a day, with 20 kW of noise in between. After the adaptive filter is smoothed, the output sequence reflects the overall shape of the rise to 520 kW and then the fall, rather than rigidly flattening the data. This feature helps to track the time point and amplitude of the peak occurrence.

[0109] In one possible implementation, the smoothed time series features can be used for subsequent analysis.

[0110] For example, the peak value of a certain feeder changes by 510 kW, 515 kW, and 508 kW for three consecutive days, and the smoothed sequence shows a consistent daily upward trend, indicating that the load growth has a periodic regularity.

[0111] It can be understood that this filtering method can avoid misjudgment of peak demand due to noise.

[0112] In one embodiment, the smoothed data can also be compared with historical records. If the peak value of a certain day suddenly increases to 540 kW, which is beyond the normal range, it may indicate potential abnormalities and further investigation is required.

[0113] Preferably, the adaptive filtering can also be combined with other methods to enhance the effect.

[0114] For example, after smoothing short-term fluctuations, if the retained peak value change direction is consistent with the long-term trend, such as the load steadily rising from 450 kW to 520 kW within a week, the reliability of the data can be verified.

[0115] S104, using a convolutional neural network to perform feature extraction processing on the smoothed time series features to obtain a multi-dimensional feature vector.

[0116] The multi-dimensional feature vector is used to represent the association between substation operation data, feeder load data, and user electricity consumption data.

[0117] The convolutional neural network is a deep learning model that is good at processing data with grid structure, such as time series. It automatically extracts features in the data through convolutional layers.

[0118] As an example but not limitation, the smoothed features are processed by the convolutional neural network to extract the dynamic association between voltage stability and load peak, obtaining a spatiotemporal correlation matrix.

[0119] The spatiotemporal correlation matrix is used to describe the correlation of data in time and space dimensions, reflecting the mutual influence and correlation degree between different time points and different spatial positions.

[0120] The spatiotemporal correlation matrix obtained from the dynamic association between voltage stability and load peak shows the correlation between voltage stability and load peak at different time points and different spatial positions (such as different substations or feeders).

[0121] The time dimension covers multiple time points, such as hours, days, etc., for observing the changes of voltage and load at different times.

[0122] The spatial dimension covers multiple spatial locations, such as different substations, feeder lines, or user areas, for analyzing the voltage and load relationship between different locations.

[0123] The correlation is used to measure the strength and direction of the linear relationship between two variables (voltage stability and load peak). The correlation coefficient is usually between -1 and 1, 1 indicates a complete positive correlation, -1 indicates a complete negative correlation, and 0 indicates no linear correlation.

[0124] For example, the feeder data of a certain substation contains the phenomenon of synchronous occurrence of voltage micro-decrease and load peak, and the convolutional neural network captures these spatio-temporal patterns through multi-layer convolution operation.

[0125] For example, the load increases from 450kW to 520kW between 8am and 10am, while the voltage micro-decreases from 220V to 218V. The network scans through the convolution kernel to extract the potential relationship between the two, generating the corresponding spatio-temporal correlation matrix. It can be understood that different time series data extraction will result in different spatio-temporal correlation matrices, which will not be listed here.

[0126] It can be understood that this matrix reflects the change rule in the time dimension and the feeder distribution characteristics in the spatial dimension, laying the foundation for subsequent analysis.

[0127] In one implementation, for the spatio-temporal correlation matrix, a principal component analysis method is used for dimension reduction to obtain a multi-dimensional feature vector.

[0128] Specifically, assuming that the matrix contains correlation data of 1440 time points and 10 feeders, the principal component analysis compresses the high-dimensional data into a few key feature vectors by identifying the main change direction.

[0129] For example, the load peak is concentrated in the 510kW to 520kW interval, and the voltage fluctuation range is within 2V. The principal component analysis extracts two vectors corresponding to the peak period and voltage offset.

[0130] S105, from the multi-dimensional feature vector, extract the preset key feature, use the weighted aggregation algorithm to fuse the data of the preset key feature, and determine the feature weight corresponding to the preset key feature.

[0131] The preset key feature includes but is not limited to voltage stability, load peak, and peak-valley difference.

[0132] As an example but not limited to, the above-mentioned preset key features are extracted by multi-dimensional feature extraction, and the weighted aggregation method is used to fuse the data to obtain the preliminary aggregation result. According to the preliminary aggregation result, the feature weight (weight coefficient) corresponding to the preset key feature is calculated.

[0133] For example, in the feeder data of a certain substation, the voltage stability is represented by a slight fluctuation of 220V to 218V, the load peak is about 510kW, and the peak-valley difference of electricity consumption is represented by the difference between 460kW at night and 530kW during the day. When these data are fused using the weighted aggregation method, initial weights can be assigned according to the importance of each feature, for example, the weight of voltage stability is set to 0.4, the weight of load peak is 0.3, and the weight of peak-valley difference is 0.3, to obtain a comprehensive preliminary aggregation result. The core of this method is to balance the influence of data in different dimensions and reflect the overall characteristics of the system.

[0134] Suppose the aggregation result shows that the load peak has a greater impact on the system, then the weight is dynamically adjusted to 0.45, and the weight of voltage stability is reduced to 0.35. This adjustment is based on the statistical characteristics of historical data, for example, when the load peak suddenly increases to 550kW on a certain day, the voltage stability is significantly affected, and the weight coefficient needs to reflect this correlation.

[0135] S106, generating an initial prediction distribution according to the feature weights and the preset key features.

[0136] Suppose the preset key features and their corresponding weights have been determined through S105.

[0137] For example, the preset key features include substation voltage stability (feature A, weight w1=0.4), feeder load peak (feature B, weight w2=0.5), and user-side electricity consumption peak-valley difference (feature C, weight w3=0.1).

[0138] The numerical values of the key features are obtained, for example, the numerical value of the substation voltage stability (feature A) at the current time is 0.85 (normalized value); the numerical value of the feeder load peak (feature B) is 520kW; and the numerical value of the user-side electricity consumption peak-valley difference (feature C) is 120kW.

[0139] According to the feature weights and the numerical values of the key features, an initial prediction distribution is generated using weighted summation or other fusion methods. For example, the initial prediction value = w1xA+w2xB+w3xC. Substituting the data in the above example can obtain an initial prediction value of 0.4x0.85+0.5x520+0.1x120=272.34.

[0140] This initial prediction value can be represented as a point estimate, or can be further extended to a distribution. For example, considering the uncertainty and fluctuation range of the prediction value, suppose the fluctuation range of the prediction value is ±5%, then the initial prediction distribution can be represented as an interval [272.34x0.95, 272.34x1.05], i.e. [258.723, 285.957].

[0141] S107, generating a power supply prediction result according to the initial prediction distribution.

[0142] According to the initial prediction distribution, a final power supply prediction result is generated in combination with specific application scenarios and requirements. For example, the middle value of the prediction interval can be taken as the prediction result, or the entire prediction interval is given as the result, and more detailed prediction information can also be provided in combination with statistical methods such as probability density functions.

[0143] For example, the middle value of the prediction interval can be taken as the prediction result, that is, 272.34, indicating that the power supply capacity at a future time or time period is expected to be 272.34 units (the specific unit is determined according to the actual data, such as kW). The entire prediction interval [258.723, 285.957] can also be given, indicating that the power supply capacity has a 90% probability of falling within this interval (assuming a confidence level of 90%).

[0144] In one embodiment, the smooth time series is divided into a first time period and a second time period according to the data change rate, the first time period is smaller than the second time period; the convolutional neural network includes at least two convolutional layers, and the convolutional neural network is used for feature extraction processing of the smooth time series features to obtain a multi-dimensional feature vector, including: using a first convolutional layer to perform feature extraction on the smooth time series corresponding to the first time period to obtain a first feature vector; using a second convolutional layer to perform feature extraction on the smooth time series corresponding to the second time period to obtain a second feature vector, the size of the convolution kernel of the first convolutional layer is smaller than the size of the second convolution kernel; and the multi-dimensional feature vector is obtained according to the first feature vector and the second feature vector.

[0145] The basic principles of the convolutional neural network will be introduced below. Figure 2

[0146] Figure 2 is a structure diagram of a convolutional neural network provided by an embodiment of the present application.

[0147] As shown in Figure 2 , the convolutional neural network (CNN) can include an input layer, a convolutional layer, a pooling layer, and an output layer.

[0148] It can be understood that when the convolutional neural network has multiple convolutional layers, the initial convolutional layer often extracts more general features (low-level features); as the depth of the convolutional neural network increases, the features extracted by the later convolutional layers become more and more complex.

[0149] A pooling layer often needs to be periodically introduced after a convolutional layer, which can be one convolutional layer followed by one pooling layer, or multiple convolutional layers followed by one or more pooling layers, and the main function is to reduce the data volume.

[0150] ​After the processing of the convolutional / pooling layers, in order to generate the final output information, a fully connected layer is usually needed to realize data output. Among them, the output layer often also has a preset loss function, which is used to calculate the prediction error. When the prediction error meets the preset accuracy, the convolutional neural network represents the completion of training and can be deployed for use.

[0151] In an implementation manner, during the processing of the convolutional neural network, the key time period can be further focused by adjusting the size of the convolution kernel. For example, for the feature of rapid rise of morning load, a shorter convolution kernel is selected to extract local patterns, and a longer kernel is used to analyze the overall change for the whole day trend.

[0152] As an example but not limitation, according to the data change rate of the smoothed time series, the time series is divided into different time periods. For example, it can be divided into a short-time fast-changing period (such as the morning load rapid rise period) and a long-time slow-changing period (such as the overall trend of the whole day).

[0153] The convolutional neural network includes at least two convolutional layers, each of which uses a different size of convolution kernel to extract features of different scales.

[0154] Among them, the first convolutional layer is used to extract features from the smoothed time series of the short time period (such as the morning load rapid rise period), and the first feature vector is obtained; the second convolutional layer is used to extract features from the smoothed time series of the long time period (such as the overall trend of the whole day), and the second feature vector is obtained.

[0155] For example, the size of the convolution kernel of the first convolutional layer is 3*3, which is used to extract local features in the short time period; the size of the convolution kernel of the second convolutional layer is 5*5, which is used to extract overall features in the long time period.

[0156] The first feature vector and the second feature vector are combined to form a multi-dimensional feature vector, which is used to comprehensively describe the features of the time series.

[0157] Suppose there is a smoothed time series containing hourly load data within a day, which shows obvious morning load rapid rise characteristics and overall load change trends, and a convolutional neural network including two convolutional layers is used to process these data.

[0158] According to the data change rate, the time series is divided into a first time period (such as 6am to 9am in the morning, the load rapid rise period) and a second time period (such as 24 hours a day, the overall load change trend).

[0159] The first convolutional layer is used to extract features from the smoothed time series of the first time period (6am to 9am in the morning), and the first feature vector is obtained; the second convolutional layer is used to extract features from the smoothed time series of the second time period (24 hours a day), and the second feature vector is obtained.

[0160] For example, the first feature vector can contain 16 features, the second feature vector can contain 32 features, and the combined multi-dimensional feature vector contains 48 features. This multi-dimensional feature vector comprehensively describes the characteristics of the time series in short and long time periods, which can be used for subsequent analysis and prediction.

[0161] Compared with the traditional scheme, different time periods are divided according to the data change rate, and different convolution layers are set for different time periods to extract features, that is, a small convolution kernel is designed to focus on rapid changes in a short time period, and local features such as instantaneous fluctuations of morning load are accurately captured; a large convolution kernel is designed to cover a longer time window, and overall features such as slow change trends of all-day load are effectively extracted, multi-scale feature fusion is realized, data characteristics are more comprehensively reflected, the model can more sensitively capture nonlinear relationships in complex data, the understanding and analysis ability of the data is improved, and the prediction performance is enhanced. The traditional method often cannot balance different scale features, resulting in poor prediction accuracy.

[0162] In one embodiment, a smoothing filter is used to smooth the short-term fluctuation component to obtain a smoothed time series feature, including: determining a long-term trend direction and a data change rate according to the long-term trend component; determining a filter window size and a filter weight coefficient of the adaptive filter according to the long-term trend direction and the data change rate; and smoothing the short-term fluctuation component according to the filter window size and the filter weight coefficient to obtain the smoothed time series feature.

[0163] The filter window defines the number of input data considered. A larger window can provide stronger smoothing effect, but will reduce the sensitivity to signal details; a smaller window retains more details, but the smoothing effect is weaker.

[0164] In the weighted filtering method, each data has a weight coefficient, and the greater the weight, the greater the impact of the data on the filtering output.

[0165] As an example but not limitation, by calculating and analyzing the long-term trend component, the long-term trend direction (such as rising, falling or stable) and the data change rate (such as the magnitude of data change per unit time) of the data are determined. According to the long-term trend direction and the data change rate, the filter window size and the filter weight coefficient of the adaptive filter are dynamically determined.

[0166] For example, dynamically adjusting the filter window size according to the data change rate and the long-term trend direction can be expressed as: W = W base × e k1⋅v ×(1+ k 2⋅ d ).

[0167] wherein, W represents the adjusted filter window size, W base is the base filter window size, k 1 and k 2 are adjustment coefficients for controlling the data variation rate v and the long-term trend direction d the degree of influence on the filter window size, d which can be represented by the trend slope, such as the slope obtained by linear regression, a positive slope indicating an upward trend and a negative slope indicating a downward trend.

[0168] For example, if the long-term trend shows that the data variation rate is fast, the filter window size can be appropriately reduced and the weight coefficient can be increased to better track the rapidly changing signal; on the contrary, if the data variation rate is slow, the filter window size can be appropriately increased and the weight coefficient can be reduced to enhance the smoothing effect on the signal.

[0169] Based on the determined filter window size and weight coefficient, the short-term fluctuation component is smoothed. This is usually achieved by methods such as weighted moving average, in which the weight of each data point is dynamically adjusted according to its distance or correlation with the current data point, so as to obtain the smoothed time series feature.

[0170] This filtering method can better preserve the key features in the short-term fluctuation component, such as the rapid change of load peak value, etc. This is because the filter window size and weight coefficient can be dynamically adjusted according to the long-term trend direction and data variation rate, so as to fully consider the characteristics of the data in the smoothing process.

[0171] By adaptively adjusting the filter parameters, noise and irregular fluctuations can be effectively removed, improving the quality of the smoothed data, making the data more smooth and stable, and closer to the real signal change trend.

[0172] The smoothed time series feature after adaptive filtering can more accurately reflect the true situation of the data, providing more reliable input for subsequent feature extraction, prediction and decision-making analysis, and helping to improve the reliability of prediction.

[0173] In an embodiment, the short-term fluctuation component is smoothed according to a filter window size and a filter weight coefficient to obtain a smoothed time series feature, including: calculating a short-term data change rate according to the short-term fluctuation component, optimizing the filter window size of the adaptive filter according to the short-term data change rate to obtain an optimized filter window size; or determining a fluctuation time period according to the short-term fluctuation component, optimizing the filter window size of the adaptive filter according to the fluctuation time period to obtain an optimized filter window size; and smoothing the short-term fluctuation component according to the optimized filter window size to obtain the smoothed time series feature.

[0174] For example, the short-term data change rate is obtained by numerically differentiating the short-term fluctuation component or calculating the difference between adjacent data points.

[0175] The fluctuation time period is identified in the short-term fluctuation component, for example, by setting a threshold value. When the short-term fluctuation component exceeds the threshold value, it is considered to be in the fluctuation time period.

[0176] The short-term data change rate is dynamically adjusted according to the filter window size v The following formula can be used to dynamically adjust the filter window size according to the short-term data change rate: W = W base ×e −k⋅v .

[0177] wherein, k is an adjustment coefficient for controlling the influence of the data change rate on the filter window size. When the short-term data change rate is fast, the filter window size is automatically reduced to better track the rapidly changing signal; conversely, when the change rate is slow, the filter window size is increased to enhance the smoothing effect.

[0178] In the time period of large fluctuation, the filter window size is reduced, for example, the filter window size can be set to half of the base value; in the time period of small fluctuation, the filter window size is expanded, for example, set to twice the base value.

[0179] The short-term fluctuation component is smoothed using the optimized filter window size, for example, the smoothing method includes weighted moving average, etc., to obtain the smoothed time series feature.

[0180] By optimizing the filter window size according to the short-term data change rate or the fluctuation time period, the relationship between smoothing and signal tracking can be better balanced, the smoothing effect can be improved, and the smoothed time series feature can be closer to the change trend of the real signal. In the time period of short-term fluctuation, the filter window size is reduced, which can quickly respond to the change of the signal, retain the key change characteristics, and avoid the loss of important information caused by excessive smoothing.

[0181] In one embodiment, the power supply data is decomposed by using a time series decomposition algorithm to obtain short-term fluctuation components and long-term trend components, including: determining the change range of the substation operation data according to the short-term fluctuation components to obtain a stability distribution; determining the peak-valley difference of the user power consumption data according to the long-term trend components to obtain a peak-valley difference regularity distribution; using a DBSCAN clustering algorithm to perform abnormal fluctuation elimination processing on the stability distribution and the peak-valley difference regularity distribution to obtain a normal fluctuation range; and determining normal short-term fluctuation components and normal long-term trend components according to the normal fluctuation range.

[0182] The change range of the substation operation data is calculated according to the short-term fluctuation components. For example, it can be achieved by calculating the extreme values (maximum and minimum values), standard deviation, etc. of the short-term fluctuation components to obtain a stability distribution. The smaller the standard deviation of the short-term fluctuation components, the more stable the substation operation data; the larger the standard deviation, the greater the fluctuation of the substation operation data.

[0183] The peak-valley difference of the user power consumption data is determined according to the long-term trend components. For example, it can be achieved by calculating the peak value and the valley value in the long-term trend components to obtain a peak-valley difference regularity distribution. For example, the daily or weekly peak value and the valley value of power consumption are analyzed to find the regularity of the changes.

[0184] The stability distribution and the peak-valley difference regularity distribution are analyzed by using a DBSCAN clustering algorithm to identify and eliminate abnormal fluctuation data points.

[0185] The DBSCAN clustering algorithm (Density-Based Spatial Clustering of Applications with Noise) is a density-based spatial clustering algorithm that can identify and eliminate noise points (i.e. abnormal points). In this example, the stability distribution and the peak-valley difference regularity distribution are used as clustering features, and the data points are clustered into two categories: normal fluctuation and abnormal fluctuation. The data points belonging to the noise points in the clustering results are eliminated to obtain the data within the normal fluctuation range.

[0186] In one embodiment, the DBSCAN algorithm is used to process the voltage fluctuation data, and the normal range is concentrated in 10kV±0.2kV, while a sudden mutation to 11kV is marked as an abnormal point. Similarly, if the peak-valley difference suddenly increases to 300kWh on a certain day, which is far more than the normal 150kWh, it will also be separated.

[0187] Preferably, the normal fluctuation range after separation can be set to voltage ±0.2kV, peak-valley difference 120kWh to 180kWh.

[0188] For example, the part of voltage fluctuation exceeding ±0.2 kV is removed, and then the time series is re-synthesized to retain the stable data of 9.8 kV to 10.2 kV, thereby ensuring consistency.

[0189] In one possible implementation, the peak-valley difference data can also be reconstructed after removing abnormal values, to obtain a periodic trend of 120 kWh to 180 kWh per day. This optimization can improve the representativeness of the data to the actual running state, and facilitate accurate analysis.

[0190] Specifically, the fluctuation component reveals the dynamic characteristics of the voltage, the trend component provides the long-term regularity of the power consumption mode, and the abnormal separation ensures the reliability of the data.

[0191] The DBSCAN clustering algorithm can effectively identify and remove abnormal fluctuation data points, improving the accuracy and reliability of the data. By removing abnormal fluctuation data points and retaining short-term fluctuation components and long-term trend components within the normal fluctuation range, it is helpful to more accurately analyze the characteristics and regularities of the power supply data.

[0192] In one embodiment, after generating the initial prediction distribution according to the feature weight and the preset key feature, the method further comprises: comparing the initial prediction distribution with a preset distribution threshold, if the initial prediction distribution is greater than the preset distribution threshold, adjusting the feature weight using a gradient optimization algorithm to obtain an optimized feature weight; generating an optimized prediction distribution according to the optimized feature weight and the preset key feature; and performing power supply prediction according to the optimized prediction distribution to generate a power supply prediction result.

[0193] If the initial prediction distribution is greater than the preset distribution threshold, it means that the prediction result may be too optimistic or deviate from the normal range, and the feature weight is adjusted using a gradient optimization algorithm to obtain an optimized feature weight.

[0194] In one implementation, if the deviation of the initial prediction distribution from the historical data exceeds the preset distribution threshold, for example, the peak value of the predicted load of a certain substation feeder is set to 500 kW, and the average of the historical data is 480 kW, the deviation is 4%, which exceeds the set threshold of 3%. At this time, the fusion weight is adjusted by the gradient optimization algorithm. In principle, it is based on the deviation direction to gradually iterate, so that the weight is closer to the actual situation.

[0195] For example, the initial weight of voltage stability accounts for 0.4, the peak load accounts for 0.3, and the peak-valley difference accounts for 0.3. After optimization, the weight of the peak load may increase to 0.35, and the weight of the voltage stability may decrease to 0.38.

[0196] Specifically, the load suddenly increases to 510 kW on a certain day, and the voltage decreases to 217 V. The algorithm adjusts the weight according to the historical correlation, and highlights the impact of the load.

[0197] Exemplarily, the adjusted coefficient changes the spatial distribution feature representation. The original feature dominated by voltage fluctuation may turn to be dominated by load change. For example, in certain feeder data, the voltage fluctuation is from 219V to 218V, and the load is increased from 470kW to 520kW. After updating, the feature vector is more inclined to reflect the spatial distribution trend of the load.

[0198] It should be noted that the new dispersion degree is calculated by the feature vector distance. If the vector length of the load change increases by 10%, the dispersion degree will also increase accordingly.

[0199] In an embodiment, the prediction distribution boundary range is calculated by the new dispersion degree. Assuming that the dispersion degree increases to 15%, the boundary range is expanded from the original 470kW to 520kW to 460kW to 530kW. The distribution characteristics show that the voltage slightly decreases when the load increases, for example, the voltage decreases by about 0.5V for every 10kW increase in load. According to this characteristic, the clustering analysis divides the data categories, and the stable category load is around 470kW, and the voltage fluctuation is less than 1V; the high load category load is above 520kW, and the voltage drops to 218V. The difference between the categories is determined by the center distance, such as the stable class center 470kW, the high load class center 525kW, and the difference of 55kW, indicating that the distribution is dispersed.

[0200] Preferably, the fusion weight distribution ratio is adjusted according to the difference degree. If the high load category has a significant impact, the load peak weight can be increased to 0.4, and the peak-valley difference can be reduced to 0.25. The optimized weight coefficient generates an optimized prediction distribution, for example, the load is increased from 460kW to 515kW on a certain day.

[0201] By comparing with the preset distribution threshold, the abnormal initial prediction distribution can be found and adjusted in time, and the accuracy of power supply prediction is improved. Using the gradient optimization algorithm to dynamically adjust the feature weight can better reflect the actual influence of different features on power supply prediction, and improve the performance of the model. By limiting the prediction distribution range, the model can avoid generating extreme prediction values, and the stability and reliability of the model are enhanced.

[0202] In an embodiment, power supply prediction is performed according to the optimized prediction distribution to generate a power supply prediction result, including: calculating user electricity density of different power supply regions according to the optimized prediction distribution; dividing the different power supply regions into grids according to the user electricity density to generate a set of sampling points, wherein each grid represents a sampling point, and the size of the grid is in a negative correlation relationship with the user electricity density; using the inverse distance weighted interpolation algorithm, the predicted load value of an unknown sampling point is calculated by weighting the load value of a known sampling point to obtain the power supply prediction result.

[0203] The user electricity density can be defined as the user electricity consumption per unit area. The power supply region is divided into grids according to the user electricity density to generate a set of sampling points.

[0204] The size of the grid is negatively correlated with the user power density, that is, a smaller grid is divided in a region with high user power density to reflect the power consumption in detail, and a larger grid is divided in a region with low user power density to reduce the calculation amount.

[0205] The basic idea of inverse distance weighted interpolation is that the closer the sampling points, the greater the influence on the unknown point, and the higher the weight.

[0206] Suppose that the user power density of a power supply area A is 200 kW / km2, and the user power density of a region B is 50 kW / km2. The grid is divided according to the user power density. In area A, a 1 km x 1 km grid is divided; in region B, a 2 km x 2 km grid is divided. The center point of each grid is taken as a sampling point. For each unknown sampling point, the load values of the surrounding known sampling points are calculated, and the predicted load value is calculated according to the distance.

[0207] Exemplarily, in the feeder data of a substation in a certain region, the power density of a commercial area is high, and the density of a residential area is low. The interpolation algorithm can adjust the grid accuracy according to the power density, such as refining the grid of the commercial area to 1 km x 1 km and relaxing the grid of the residential area to 2 km x 2 km, to generate a set of sampling points covering the whole area. The preliminary spatial distribution can intuitively reflect the spatial variation trend of the power load.

[0208] For example, there are three known sampling points A, B, and C around the unknown sampling point P, with distances of 1 km, 2 km, and 3 km, and load values of 120 kW, 150 kW, and 180 kW, respectively. The predicted load value of the unknown sampling point P can be obtained by weighted calculation of the load values of the sampling points A, B, and C. The specific weight value can be set according to the actual situation.

[0209] According to the user power density, the grid is divided to reflect the power consumption in different power supply areas in detail, and the accuracy of power supply prediction is improved. In the region with low user power density, a larger grid is divided to reduce the number of sampling points, thereby reducing the calculation amount and improving the calculation efficiency. Through grid division and sampling point generation, the prediction results can be intuitively displayed in the form of a map, which is convenient for decision makers to understand and use. The inverse distance weighted interpolation algorithm can make full use of the information of the known sampling points, reasonably estimate the load value of the unknown sampling points, and improve the reliability of the prediction results.

[0210] In one embodiment, the method further comprises: constructing a spatial gradient matrix according to the load values of the known sampling points and the predicted load values of the unknown sampling points; performing cluster analysis on the spatial gradient matrix to obtain a plurality of classification regions; determining the prediction priority of the plurality of classification regions according to the fluctuation sensitivity of the plurality of classification regions; and sequentially performing power supply prediction according to the prediction priority of the plurality of classification regions to generate power supply prediction results corresponding to the plurality of classification regions.

[0211] The load values of the known sampling points and the predicted load values of the unknown sampling points are collected to construct a spatial gradient matrix, which is used to describe the changes in load values of different regions.

[0212] Table 1

[0213] Sampling points Position (x, y) Load value (kW) A (0,0) 120 B (1,0) 150 C (0,1) 180 D (1,1) 160

[0214] The load gradient values between the sampling points are calculated.

[0215] GAB = -30; GAC = -60; GAD = -40; GBC = -30; GBD = -10; GCD = 20.

[0216] These gradient values are organized into a matrix form to construct a spatial gradient matrix.

[0217] The K-means clustering algorithm is used to perform cluster analysis on the spatial gradient matrix, and the regions are divided into a plurality of classification regions according to the similarity of load changes.

[0218] The prediction priority is determined according to the fluctuation sensitivity of the classification regions. The regions with high fluctuation sensitivity are more sensitive to power supply fluctuations and should be prioritized for prediction. The power supply prediction is sequentially performed on different regions according to the prediction priority to generate power supply prediction results corresponding to each region.

[0219] In the power supply field, the regions with high fluctuation sensitivity are prone to significant fluctuations in power supply load or voltage due to interference, affecting the stability of power supply and the reliability of equipment operation.

[0220] For example, the fluctuation sensitivity of the commercial area is evaluated as 0.9 due to the intensive and frequent fluctuations in electricity consumption, while the fluctuation sensitivity of the surrounding residential area is only 0.4 due to the stable load. The weighted method can give the commercial area a higher weight, for example, 0.7, and the residential area a weight of 0.3, to generate a priority sequence.

[0221] In one embodiment, the method further comprises: comparing the power supply prediction results corresponding to the plurality of classification regions with the actual power supply values corresponding to the plurality of classification regions to determine the update time range corresponding to the plurality of classification regions; performing smoothing processing on the smoothed time series features within the update time range to obtain updated smoothed time series; and inputting the updated smoothed time series into a convolutional neural network for feature extraction processing to obtain an updated multi-dimensional feature vector.

[0222] By comparing the prediction results with the actual power supply values, determining the update time range can identify areas and time periods with large prediction errors, and optimize the model accordingly.

[0223] Calculate the error (such as mean square error, absolute percentage error) between the predicted value and the actual value of each area, and set an error threshold (such as error > 10% is considered significant deviation). For areas that exceed the threshold, extract the time period with significant error (such as the last n hours or n days) as the update time range.

[0224] Adopt adaptive filtering or weighted moving average, dynamically adjust parameters according to data fluctuations, that is, perform secondary smoothing operation, smooth the data in the update time range (such as 3 hours of load data of area B), output the updated smoothed time series (such as the load curve after removing abnormal peaks).

[0225] For example, the original load data of area B has minute-level jumps (such as 500kW→550kW→500kW), and after secondary smoothing, it becomes a gentle rise (500kW→510kW→520kW). Input the updated smoothed time series into the convolutional neural network for feature extraction processing, and get the updated multi-dimensional feature vector.

[0226] Through error comparison, accurately locate the areas and time periods that need to be optimized, avoid resource waste caused by global update. Through secondary smoothing to reduce noise, make the model learn the more real load change rule.

[0227] Specifically, when extracting the spatio-temporal correlation matrix from the updated training set (updated smoothed time series) through the convolutional neural network, it can be understood as using the convolutional layer of the network to capture the correlation characteristics in the time and space dimensions.

[0228] For example, in the power supply scenario of a certain city industrial park, the training set may contain multi-day load data, such as the daily 24-hour electricity consumption records of a certain substation. The convolutional neural network scans these data through a sliding window, extracts the correlation between load fluctuations over time and spatial distribution, forms a matrix, and represents the dependence between different time points and areas. When combining the sampling point representation to obtain the preliminary spatial distribution feature, it can start from the load data of key nodes.

[0229] Specifically, five main substations in the industrial park may be selected as sampling points, each recording real-time load, such as 800kW, 920kW, 650kW, etc. at a certain time on a certain day. The convolutional neural network maps these sampling point data to the spatial distribution, generating a preliminary load distribution feature map reflecting the spatial differences in power supply capacity within the park.

[0230] As an example but not limitation, when processing the spatio-temporal correlation matrix by the weighted aggregation method, weights can be assigned according to the importance of the sampling points.

[0231] In an implementation, the substations near the industrial core area are given higher weights, such as 0.4, due to high load demand, while the edge area weight is 0.2.

[0232] After weighted aggregation, the convergence speed distribution range reflected by the matrix may show that the core area changes fast and the edge area changes slowly. If the convergence speed of the core area is fast, it indicates that the load fluctuation is severe, and the number of sampling points can be increased to 7 to cover a wider area. The updated spatial distribution representation may show that the load transition from the core area 920 kW to the edge is clearer, which helps to identify weak power supply points. If the convergence speed exceeds the preset threshold, such as a change of more than 50 kW per minute, the spatial distribution needs to be adjusted again.

[0233] Preferably, more accurate distribution prediction features can be obtained by increasing the time resolution from hourly sampling to half-hourly sampling.

[0234] This method realizes efficient self-updating of the prediction system through the closed loop of "local optimization → feature enhancement → model iteration", and is suitable for dynamic changing power demand scenarios.

[0235] The above mainly introduces the method of the embodiments of the application in combination with the drawings. It should be understood that although each step in the flowchart involved in each embodiment as described above is displayed in sequence, these steps are not necessarily executed in sequence as shown in the figure. Unless explicitly stated herein, the execution of these steps has no strict sequence limitation, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in each embodiment as described above can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be executed in rotation or alternation with other steps or steps or stages in other steps. The following introduces a system of the embodiments of the application in combination with the drawings. In order to be brief, in the following introduction of the system, appropriate omissions will be made, and the related content can be referred to the related description in the above method, which will not be introduced again.

[0236] Figure 3 is a structural schematic diagram of an intelligent prediction system for power supply capacity of a distribution network provided by the embodiments of the application.

[0237] As Figure 3 shown, the system 1000 includes the following units.

[0238] The acquisition unit 1001 is configured to acquire power supply data, the power supply data including at least one of substation operation data, feeder load data, and user power consumption data.

[0239] The decomposition unit 1002 is configured to perform decomposition processing on the power supply data by using a time series decomposition algorithm to obtain a short-term fluctuation component and a long-term trend component, the short-term fluctuation component being used to represent short-term changes in the power supply data, and the long-term trend component being used to represent long-term changes in the power supply data.

[0240] The smoothing unit 1003 is configured to perform smoothing processing on the short-term fluctuation component by using an adaptive filter to obtain a smoothed time series feature, wherein the adaptive filter adjusts corresponding filter parameters according to the long-term trend component.

[0241] The convolution unit 1004 is configured to perform feature extraction processing on the smoothed time series feature by using a convolutional neural network to obtain a multi-dimensional feature vector, the multi-dimensional feature vector being used to represent an association relationship between the substation operation data, the feeder load data, and the user power consumption data.

[0242] The fusion unit 1005 is configured to extract a preset key feature from the multi-dimensional feature vector, and perform data fusion processing on the preset key feature by using a weighted aggregation algorithm to determine a feature weight corresponding to the preset key feature.

[0243] The calculation unit 1006 is configured to generate an initial prediction distribution according to the feature weight and the preset key feature.

[0244] The prediction unit 1007 is configured to perform power supply prediction according to the initial prediction distribution to generate a power supply prediction result.

[0245] It should be noted that the information interaction and execution process between the above units are based on the same concept as the method embodiments of the present application, and the specific functions and technical effects thereof can be referred to the method embodiments part, which will not be repeated here.

[0246] Figure 4 is a structural schematic diagram of a computer device provided by the embodiment of the present application. As shown in Figure 4 the computer device 3000 of the embodiment includes at least one processor 3100 (only one processor is shown), a memory 3200, and a computer program 3210 stored in the memory 3200 and executable on the at least one processor 3100, and the processor 3100 executes the computer program 3210, so that the computer device implements the steps in the above embodiments. Figure 4

[0247] ​The processor 3100 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gates or transistor logic, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can be any conventional processor.

[0248] The memory 3200 can be an internal storage unit of the computer device 3000 in some embodiments, for example, a hard disk or a memory of the computer device 3000. The memory 3200 can also be an external storage device of the computer device 3000 in other embodiments, for example, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. Further, the memory 3200 can include both the internal storage unit and the external storage device of the computer device 3000. The memory 3200 is used to store an operating system, application programs, boot loader data, and other programs, for example, program codes of computer programs, etc. The memory 3200 can also be used to temporarily store data that has been output or will be output.

[0249] It can be clearly understood by those skilled in the art that, for the convenience and brevity of description, only the division of the above functional units is taken as an example, and in actual application, the above functions can be completed by different functional units or modules according to needs, that is, the internal structure of the apparatus is divided into different functional units or modules to complete all or part of the functions described above. Each functional unit in the embodiment can be integrated in one processing unit, or each unit can exist physically independently, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit. In addition, the specific names of the functional units are only for the convenience of mutual distinction, and do not limit the protection scope of the present application. The specific working process of the unit in the system can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0250] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program. The computer program is executed by a computer device, so that the computer device implements steps in each method embodiment.

[0251] The embodiment of the present application provides a computer program product. When the computer program product is run on a computer device, the computer device can implement the above-mentioned methods.

[0252] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the present application implements all or part of the processes in the above-mentioned embodiment methods, which can be completed by instructing related hardware through a computer program. The computer program can be stored in a computer readable storage medium. When the computer program is executed by a processor, the computer device can implement the steps of each method embodiment. The computer program includes computer program code, which can be in the form of source code, object code, executable files or some intermediate forms. The computer readable medium at least includes any entity or device capable of carrying the computer program code to the photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunications signal and a software distribution medium. For example, a U disk, a mobile hard disk, a magnetic disk or an optical disk, etc. In some jurisdictions, according to legislation and patent practice, the computer readable medium cannot be an electrical carrier signal and a telecommunications signal.

[0253] It should be understood that the size of the serial number of each step in the above-mentioned embodiment does not mean the execution order. The execution order of each process should be determined according to its function and inherent logic, and should not constitute any limitation on the implementation process of the embodiment of the present application. In the description, specific details such as specific system structures, technologies are proposed for illustration but not for limitation, so as to thoroughly understand the embodiments of the present application. However, it should be clear for those skilled in the art that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed description of well-known systems, devices, circuits and methods is omitted to avoid unnecessary details that hinder the description of the present application.

[0254] It should be understood that when used in the specification and the appended claims of the present application, the term "comprising" indicates the existence of the described features, integers, steps, operations, elements and / or components, but does not exclude one or more other features, integers, steps, operations, elements, components and / or sets thereof.

[0255] It should also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items, and that the term "at least one of" followed by a list of two or more items means any single one of the items in the list, and that the term "one or more of" followed by a list of two or more items means any single one or plurality of the items in the list.

[0256] In addition, the terms "comprise", "comprising", "include", "including", "have", "has", "contain", "containing", "provide", "providing", "involving", "carry", "carrying", "utilize", "utilizing", "characterized by" and the like, when used in the description and the following claims, are each intended to mean that the stated item, material, etc. is included in or comprises a part of the claimed subject matter, but does not exclude the presence of additional items, materials, etc. in order to convey the idea that the claimed subject matter can include, but does not necessarily exclude, the presence of additional items, materials, etc.

[0257] In the above embodiments, the description of each embodiment is focused on, and the parts not described or recorded in a certain embodiment can be referred to the relevant description of other embodiments.

[0258] Those skilled in the art can understand that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software mode depends on the specific application and design constraints of the technical solutions. The skilled person can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0259] In the embodiments provided in the present application, it should be understood that the disclosed devices, computer devices and methods can be implemented in other ways. For example, the above-described device and computer device embodiments are only schematic. The division of units is only a logical function division. In actual implementation, additional division modes can be used. For example, multiple units or components can be combined or integrated into another system, or some features can be omitted or not implemented. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or other forms.

[0260] The above-described embodiments are only used to illustrate the technical solutions of the present application, but not limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A method for intelligent prediction of power supply capability of a power distribution network, characterized in that, The method comprises: obtaining power supply data, the power supply data comprising at least one of substation operation data, feeder load data, and user power consumption data; decomposing the power supply data using a time series decomposition algorithm to obtain a short-term fluctuation component and a long-term trend component, the short-term fluctuation component being used to represent short-term changes in the power supply data, and the long-term trend component being used to represent long-term changes in the power supply data; smoothing the short-term fluctuation component using an adaptive filter to obtain smoothed time series features, wherein the adaptive filter adjusts corresponding filter parameters according to the long-term trend component; extracting features from the smoothed time series features using a convolutional neural network to obtain a multi-dimensional feature vector, the multi-dimensional feature vector being used to represent the correlation between substation operation data, feeder load data, and user power consumption data; extracting preset key features from the multi-dimensional feature vector, and performing data fusion on the preset key features using a weighted aggregation algorithm to determine feature weights corresponding to the preset key features; generating an initial prediction distribution according to the feature weights and the preset key features; performing power supply prediction according to the initial prediction distribution to generate a power supply prediction result.

2. The method of claim 1, wherein, dividing the smoothed time series features into a first time period and a second time period according to the data change rate of the smoothed time series features, the first time period being smaller than the second time period; the convolutional neural network comprises at least two convolutional layers, and the extracting features from the smoothed time series features using the convolutional neural network to obtain a multi-dimensional feature vector comprises: extracting features from the smoothed time series features corresponding to the first time period using a first convolutional layer to obtain a first feature vector; extracting features from the smoothed time series features corresponding to the second time period using a second convolutional layer to obtain a second feature vector, the size of the convolutional kernel of the first convolutional layer being smaller than the size of the convolutional kernel of the second convolutional layer; obtaining the multi-dimensional feature vector according to the first feature vector and the second feature vector.

3. The method of claim 1, wherein, The smoothing of the short-term fluctuation component using the adaptive filter to obtain the smoothed time series features comprises: determining a long-term trend direction and a data change rate according to the long-term trend component; determining a filter window size and a filter weight coefficient of the adaptive filter according to the long-term trend direction and the data change rate; smoothing the short-term fluctuation component according to the filter window size and the filter weight coefficient to obtain the smoothed time series features.

4. The method of claim 3, wherein, The smoothing of the short-term fluctuation component according to the filter window size and the filter weight coefficient to obtain the smoothed time series features comprises: calculating a short-term data change rate according to the short-term fluctuation component, and optimizing the filter window size of the adaptive filter according to the short-term data change rate to obtain an optimized filter window size; or determining a fluctuation time period according to the short-term fluctuation component, optimizing a filter window size of the adaptive filter according to the fluctuation time period, and obtaining an optimized filter window size; performing smoothing processing on the short-term fluctuation component according to the optimized filter window size, and obtaining a smoothed time series feature.

5. The method of claim 1, wherein, The decomposing the power supply data by using the time series decomposition algorithm to obtain the short-term fluctuation component and the long-term trend component includes: determining a change range of the substation operation data according to the short-term fluctuation component, and obtaining a stability distribution; determining a peak-valley difference of the user power consumption data according to the long-term trend component, and obtaining a peak-valley difference regularity distribution; performing abnormal fluctuation elimination processing on the stability distribution and the peak-valley difference regularity distribution by using a DBSCAN clustering algorithm, and obtaining a normal fluctuation range; determining a normal short-term fluctuation component and a normal long-term trend component according to the normal fluctuation range.

6. The method according to any one of claims 1-5, characterized in that, After the initial prediction distribution is generated according to the feature weight and the preset key feature, the method further includes: comparing the initial prediction distribution with a preset distribution threshold value, and if the initial prediction distribution is greater than the preset distribution threshold value, adjusting the feature weight by using a gradient optimization algorithm to obtain an optimized feature weight; generating an optimized prediction distribution according to the optimized feature weight and the preset key feature; performing power supply prediction according to the optimized prediction distribution to generate a power supply prediction result.

7. The method of claim 6, wherein, The power supply prediction according to the optimized prediction distribution to generate a power supply prediction result includes: calculating user power consumption densities of different power supply regions according to the optimized prediction distribution; dividing the different power supply regions into grids according to the user power consumption densities to generate a set of sampling points, wherein each grid represents a sampling point, and the size of the grid is in a negative correlation relationship with the user power consumption density; using an inverse distance weighted interpolation algorithm to calculate predicted load values of unknown sampling points by weighting load values of known sampling points, and obtaining the power supply prediction result.

8. The method of claim 7, wherein, The method further includes: constructing a spatial gradient matrix according to the load values of the known sampling points and the predicted load values of the unknown sampling points; performing clustering analysis on the spatial gradient matrix to obtain a plurality of classification regions; determining prediction priorities of the plurality of classification regions according to fluctuation sensitivities of the plurality of classification regions; performing power supply prediction on the plurality of classification regions in sequence according to the prediction priorities of the plurality of classification regions to generate power supply prediction results corresponding to the plurality of classification regions.

9. The method of claim 8, wherein, The method further includes: comparing the power supply prediction results corresponding to the plurality of classification regions with actual power supply values corresponding to the plurality of classification regions to determine update time ranges corresponding to the plurality of classification regions; performing smoothing processing on the smoothed time series features in the update time ranges to obtain updated smoothed time series; inputting the updated smoothed time series into the convolutional neural network to perform feature extraction processing to obtain an updated multi-dimensional feature vector.

10. A system for intelligent forecasting of power supply capability of a power distribution network, characterized in that, The method includes: an acquisition unit configured to acquire power supply data, the power supply data including at least one of substation operation data, feeder load data, and user power consumption data; The decomposition unit is configured to perform decomposition processing on the power supply data by using a time series decomposition algorithm to obtain a short-term fluctuation component and a long-term trend component, the short-term fluctuation component being used to represent short-term changes in the power supply data, and the long-term trend component being used to represent long-term changes in the power supply data. The smoothing unit is configured to perform smoothing processing on the short-term fluctuation component by using an adaptive filter to obtain a smoothed time series feature, wherein the adaptive filter adjusts corresponding filter parameters according to the long-term trend component. The convolution unit is configured to perform feature extraction processing on the smoothed time series feature by using a convolutional neural network to obtain a multi-dimensional feature vector, the multi-dimensional feature vector being used to represent an association relationship among substation operation data, feeder load data, and user power consumption data. The fusion unit is configured to extract a preset key feature from the multi-dimensional feature vector, perform data fusion processing on the preset key feature by using a weighted aggregation algorithm, and determine a feature weight corresponding to the preset key feature. The calculation unit is configured to generate an initial prediction distribution according to the feature weight and the preset key feature. The prediction unit is configured to perform power supply prediction according to the initial prediction distribution to generate a power supply prediction result.

Images