Power consumption trend perception method and system based on multi-time scale integrated analysis

Abstract

This invention discloses a method and system for sensing electricity consumption trends based on multi-timescale integrated analysis, relating to the field of power system monitoring technology. It includes: acquiring time-series data of electricity consumption to be monitored; filtering the data using a multi-timescale filter bank with window lengths increasing by a set common ratio to obtain trend subsequences and fluctuation subsequences corresponding to multiple time scales; for each trend subsequence, detecting local turning points where the trend direction reverses, calculating the frequency of occurrence and maximum duration of local turning points within the observation window, and jointly calculating the trend confidence level for that scale; and integrating the trend confidence levels of all scales using a weighting factor negatively correlated with the window length to form a multi-scale trend integrated vector.

Description

Technical Field

[0001] This invention relates to the field of power system monitoring technology, specifically to a method and system for sensing electricity consumption trends based on multi-timescale integrated analysis. Background Technology

[0002] Trend perception of electricity consumption time series is a crucial foundation for power load analysis, anomaly detection, and demand response. Existing technologies typically employ fixed-window moving averages, exponential smoothing, or single-scale filtering methods to extract trend components from electricity consumption series, and determine the trend direction based on the sign of the first derivative of the trend component or a preset threshold. These methods can provide a rough estimate of the trend direction in scenarios with stable electricity load and simple fluctuation characteristics, but they struggle to effectively address the overlapping fluctuations across multiple time scales and sudden short-term disturbances inherent in actual electricity consumption data. Electricity consumption behaviors at different time scales, such as hourly periodic fluctuations, daily load shifts, and longer-term seasonal variations, are simultaneously embedded in the same time series. Single-scale trend extraction methods cannot separate trend subsequences corresponding to different behavioral patterns from the series, resulting in trend perception results that only reflect an approximate trend at a specific level of smoothing, ignoring directional changes occurring at other scales. Isolated fluctuations caused by local spikes, short-term load switching, or acquisition anomalies can easily lead to frequent trend reversals in narrow-window filtering results, while in wide-window filtering results they are over-smoothed, causing the timing of turning points to be delayed and reducing the timeliness of turning signals. Furthermore, most existing trend judgments are based solely on local differential information on the time axis, lacking cross-sectional verification of trend direction consistency across multiple time scales and comprehensive consideration of the evolution characteristics of trend signals at multiple continuous sampling points. A single criterion is prone to false alarms during periods of dense fluctuations and missed alarms due to response delays when true trend reversals occur. How to reasonably quantify the reliability of trend directions at each scale after separating multi-scale trend sub-sequences, and construct a cross-scale and time-axis dual consistency verification mechanism, is an urgent problem to be solved. Summary of the Invention

[0003] This paper provides a method and system for sensing electricity consumption trends that can separate electricity consumption trends from multiple time scales, comprehensively evaluate the reliability of trends at each scale, and use cross-scale gradients and time evolution directions for dual confirmation.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a method for sensing electricity consumption trends based on multi-timescale integrated analysis, comprising: acquiring time series data of electricity consumption to be monitored; filtering the time series data of electricity consumption using a preset multi-timescale filter bank to obtain trend subsequences and fluctuation subsequences corresponding to multiple time scales, wherein the filter window lengths of the multi-timescale filter bank are arranged in ascending order according to a set common ratio; for each trend subsequence, detecting local turning points where the trend direction reverses in the trend subsequence, calculating the frequency of occurrence of the local turning points within a set observation window, and extracting the trend subsequence... Maintaining the maximum duration of a unidirectional trend, the trend confidence level for that time scale is calculated by jointly using the occurrence frequency and the maximum duration. The trend confidence levels for all time scales are integrated using a weighting factor negatively correlated with the scale window length to form a multi-scale trend integration vector. The differential gradients of adjacent elements of the multi-scale trend integration vector on the scale axis are calculated, and combined with the magnitude evolution direction of the integration vector corresponding to multiple consecutive sampling points on the time axis, when the sign of the differential gradient remains consistent within a preset number of consecutive sampling points and the magnitude evolution direction remains consistent within another preset number of consecutive sampling points, an electricity consumption trend sensing signal is output.

[0005] By constructing a multi-scale filter bank with window length increasing proportionally, the original electricity consumption sequence is decomposed into trend components and fluctuation components with different time granularities. This avoids the omission or misjudgment of trends by a single observation scale, thus capturing more comprehensive trend characteristics from transient fluctuations to long-term evolution.

[0006] Preferably, the step of filtering the electricity consumption time series data using a preset multi-time-scale filter bank specifically includes: determining the minimum filter window length and the common ratio of window length increments, generating a set of filter window length sequences that are progressively amplified according to the common ratio; for each filter window length, calculating a central moving average of the electricity consumption time series data using that window length, and using the moving average result as the trend subsequence corresponding to that scale; subtracting the trend subsequence from the electricity consumption time series data to obtain the fluctuation subsequence at that scale. The central moving average can effectively suppress random noise while preserving the trend pattern, allowing the trend components at each scale to more accurately reflect the electricity consumption change pattern over the corresponding time span.

[0007] Preferably, the step of jointly calculating the trend confidence level for the time scale using the occurrence frequency and the maximum duration includes: normalizing the occurrence frequency to obtain a reversal density index; dividing the maximum duration by the filter window length of the time scale to obtain a relative persistence index; and harmonic averaging the reversal density index and the relative persistence index, with the result serving as the trend confidence level for the time scale. Using a harmonic average can reasonably balance the mutual constraint between the frequency of trend reversals and the stability of trend persistence, making the confidence level more realistically represent the credibility of the trend at that scale, and avoiding overly sensitive or insensitive single indicators.

[0008] Preferably, the weighting factor negatively correlated with the scale window length can be determined as follows: For each time scale, calculate the inverse of the filter window length for that scale, and divide this inverse by the sum of the inverses of the window lengths for all time scales to obtain a normalization factor. This normalization factor is then used as the weighting factor for that scale. In this way, the narrow-window scale receives a larger integration weight, highlighting the perception of rapidly changing micro-trends, while the wide-window scale participates in the integration with a smaller weight, preserving the stable contribution of macro-trends. This achieves efficient fusion of multi-scale trend information, making the integrated vector both sensitive and robust to trend reversals.

[0009] Preferably, the step of obtaining the differential gradient of adjacent elements of the multi-scale trend integration vector on the scale axis, combined with the magnitude evolution direction of the integration vector, specifically includes: for the multi-scale trend integration vector generated at the current sampling point, calculating the difference between adjacent elements in order of scale from narrow to wide, and dividing the difference by the difference of the logarithmic window length between adjacent scales as the differential gradient; extracting the integration vector of the current sampling point and its preceding multiple consecutive sampling points, calculating the L2 norm of each integration vector as the magnitude, comparing the increasing and decreasing directions of the magnitude of adjacent sampling points, and if the increasing and decreasing directions are both increasing or both decreasing within a consecutive preset number of sampling points, then it is determined that the magnitude evolution direction is consistent. The differential gradient characterizes whether the changing direction of trend intensity is coordinated at multiple scales, and the magnitude evolution direction reflects whether the overall trend energy is continuously increasing or decreasing. The joint decision of the two can reliably filter out accidental shifts caused by random disturbances, improving the certainty and anti-interference ability of trend perception.

[0010] Preferably, detecting local turning points where the trend direction reverses in the trend subsequence includes: performing a first-order difference operation on the trend subsequence, identifying the positions where the signs of the difference results change point by point, marking the data points where the signs change as local turning points, and retaining only the start and end points of intervals with consecutive identical signs as turning boundaries. This processing method accurately marks the trend reversal positions while reducing redundant turning points, thus improving the computational efficiency of subsequent frequency statistics.

[0011] Preferably, the length of the observation window is proportional to the time scale window length corresponding to the trend subsequence, and the observation window is truncated from the current evaluation time towards the historical time. This allows for the use of a longer historical window at long scales to fully accumulate turning information, while a shorter window at short scales maintains keen tracking of recent changes.

[0012] Preferably, the preset number of continuous sampling points used to determine the consistency of the sign of the differential gradient and the other preset number of continuous sampling points used to determine the consistency of the modulus evolution direction satisfy a mapping relationship determined by the time scale configuration scheme, wherein the mapping relationship is a monotonic function of the ratio of the maximum window length to the minimum window length within the multi-time scale filter bank. Through this mapping relationship, the duration of the decision can be dynamically adapted to the selected multi-scale configuration, ensuring the robustness and consistency of the trend-aware signal under different combinations of observation granularities.

[0013] Preferably, the electricity consumption trend sensing signal includes a direction indicator and a reliability level. The direction indicator is determined by the sign of the differential gradient when it remains consistent across consecutive sampling points. The reliability level is determined based on the ratio of the number of sampling points spanned during the duration of the magnitude evolution direction within another preset number to the other preset number. The output signal thus not only indicates the trend direction of electricity consumption changes but also quantifies the reliability of this trend through a level, providing more valuable quantitative data for subsequent applications such as electricity dispatching and anomaly warning.

[0014] This invention also provides a power consumption trend sensing system based on multi-timescale integrated analysis, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the power consumption trend sensing method based on multi-timescale integrated analysis. This system can automatically complete multi-timescale trend decomposition, confidence assessment, weighted integration, and direction determination, and output trend sensing signals with direction and reliability levels in real time, thereby providing accurate and timely trend information for scenarios such as power management and load forecasting.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: By employing a multi-timescale filter bank with window lengths arranged in ascending order of a set common ratio, the electricity consumption time series is filtered to generate trend subsequences corresponding to different scales. This configuration allows the filter bank to cover a wide spectral range from short-term fluctuations to long-term changes with a smaller number of filters while maintaining a moderate density of window length variations, avoiding redundant calculations and scale overlaps caused by mechanical uniform sampling at each scale. In the trend confidence level calculation stage, local turning points where the trend direction reverses are detected for each trend subsequence. The frequency of occurrence of turning points within the observation window is statistically analyzed, and the maximum duration of maintaining a unidirectional trend is extracted. The frequency is normalized to obtain a turning density index, and the maximum duration is divided by the filter window length at that time scale to obtain a relative persistence index. The two are then harmonicly averaged. This measurement method, which combines turning density and relative persistence characteristics, simultaneously suppresses false turnings caused by high-frequency noise and sluggish responses caused by over-smoothing. This ensures that the trend confidence level at each scale can truly reflect the inherent persistence of the trend direction at that scale, avoiding biased judgments caused by relying solely on the number of turning points or the length of trend segments. In the cross-scale integration stage, the trend confidence levels across all time scales are integrated using a weighting factor negatively correlated with the scale window length. This weighting factor is obtained by normalizing the inverse of the filter window length for each scale, assigning higher weights to short-term scales corresponding to narrow windows and lower weights to long-term scales corresponding to wide windows. Through this negatively correlated weighting, the multi-scale trend integration vector can retain information about long-term stable trends while highlighting rapid response capabilities, achieving a trade-off between response speed and anti-interference capabilities among different frequency components. In the triggering logic of the trend sensing signal, the differential gradients of adjacent elements of the multi-scale trend integration vector on the scale axis are calculated and jointly determined by combining them with the evolution direction of the integration vector's magnitude corresponding to multiple consecutive sampling points on the time axis. Only when the sign of the differential gradient remains consistent within a preset number of consecutive sampling points and the evolution direction of the magnitude remains consistent within another preset number of consecutive sampling points is the trend confirmed and the sensing signal output. This dual verification mechanism, which combines gradients across scales with temporal direction, simultaneously locks in trend changes in both spatial and temporal dimensions. It effectively filters out gradient flickering caused by isolated local fluctuations at individual scales, as well as transient magnitude jitter on the time axis. This significantly reduces false alarms induced by transient phenomena at a single scale, which are common in multi-scale analysis, while maintaining the sensitivity to detect true trend reversals. Attached Figure Description

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

[0017] Figure 1 This is a flowchart of a method for sensing electricity consumption trends based on multi-timescale integrated analysis; Figure 2 This is a flowchart of the multi-scale trend integration vector construction process; Figure 3 This is a schematic diagram of the trend subsequence of a multi-timescale filter bank; Figure 4 This is a diagram illustrating the direction of electricity consumption trend sensing and the reliability level. Detailed Implementation

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

[0019] See Figure 1 This invention provides a method for sensing electricity consumption trends based on multi-timescale integrated analysis, comprising: S1: acquiring time-series data of electricity consumption to be monitored; S2: filtering the time-series data of electricity consumption using a preset multi-timescale filter bank to obtain trend subsequences and fluctuation subsequences corresponding to multiple time scales, wherein the filter window lengths of the multi-timescale filter bank are arranged in ascending order according to a set common ratio; S3: for each trend subsequence, detecting local turning points where the trend direction reverses, calculating the frequency of occurrence of the local turning points within a set observation window, and extracting the maximum value of the trend subsequence that maintains a unidirectional trend. S4: The trend confidence level of the time scale is calculated by jointly using the occurrence frequency and the maximum duration; S5: The trend confidence levels of all time scales are integrated by applying a weighting factor negatively correlated with the window length of the scale to form a multi-scale trend integration vector; S6: The differential gradient of the adjacent elements of the multi-scale trend integration vector on the scale axis is obtained, and combined with the magnitude evolution direction of the integration vector corresponding to multiple continuous sampling points on the time axis, when the sign of the differential gradient remains consistent within a preset number of continuous sampling points and the magnitude evolution direction remains consistent within another preset number of continuous sampling points, the electricity consumption trend sensing signal is output.

[0020] Example 1: In practice, after acquiring the time-series data of electricity consumption to be monitored, a preset multi-time-scale filter bank is used to filter the electricity consumption time-series data to obtain trend subseries and fluctuation subseries corresponding to multiple time scales. The window lengths of each filter in the multi-time-scale filter bank are arranged in ascending order according to a set common ratio.

[0021] Determine the minimum filter window length and the common ratio of window length increments. The minimum filter window length is denoted as... The minimum filter window length is a positive integer, representing the minimum number of sampling points required for the central moving average. In some embodiments, the minimum filter window length can be set based on the sampling frequency of the electricity consumption time series and the shortest trend period to be captured. For example, for hourly sampled electricity consumption data, the minimum filter window length can be 3 sampling points. The common ratio of the increasing window length is denoted as... The common ratio of the window length increments is a real number greater than 1, used to control the amplification factor of adjacent filter window lengths. The common ratio of the window length increments can be selected based on the required time scale span and filter bank size; for example, the common ratio of the window length increments can be 2. Based on the minimum filter window length and the common ratio of the window length increments, a sequence of filter window lengths is generated, increasing progressively according to the common ratio. The first... The elements are denoted as ,in This is a scale index, with values ​​starting from 1 as positive integers. The largest index is denoted as . The filter window length sequence is generated according to the following relationship:

[0022] in, Minimum filter window length, The common ratio of increasing window lengths, For scale indexing, For the first Each scale corresponds to a filter window length. The largest element in the filter window length sequence does not exceed a predetermined proportion of the total length of the electricity consumption time series to avoid unreliable estimations due to boundary effects. When the generated filter window length exceeds this predetermined proportion, the generation of subsequent elements stops, thereby determining the final filter bank size. .

[0023] For each filter window length in the filter window length sequence With the length of the filtering window A central moving average is calculated for the electricity consumption time series data. The central moving average calculation is implemented as follows: for electricity consumption time series data located at time... The sampling points were selected based on time. Centered on the continuity Each sampling point constitutes a local window, and the local window contains time... oneself and time front and back Each sampling point, within the filter window length Achieving complete symmetry when the value is odd; within the filter window length When the number is even, the local window can be adjusted to make the previously containing window... Points, backward inclusion The data can be processed into an odd-numbered equivalent form by taking the average of adjacent data points. The sum of the electricity consumption values ​​at all sampling points within a local window is then divided by the filter window length. Get the time The moving average at each point in the electricity consumption time series is used. This operation is performed for each point in the time series, and the moving average of all points forms the corresponding scale. The trend subsequence. At the beginning and end of the electricity consumption time series, the local window may exceed the sequence boundary. A truncation method is used to process this, that is, only the available sampling points within the sequence boundary are used to calculate the average value, or a mirror extension method is used to fill in the missing sampling points at the boundary to keep the filter window length constant.

[0024] In obtaining the corresponding scale After obtaining the trend subsequence, the value of each moment in the original electricity consumption time series data is subtracted from the value of the corresponding moment in the trend subsequence to obtain the corresponding scale. The fluctuation subsequence. The fluctuation subsequence contains high-frequency fluctuation components at this scale, and together with the trend subsequence, constitutes a pair of decomposition results at this scale. For all For each filter window length, the above-mentioned moving average and subtraction operations are performed to obtain the result. All trend subsequences and volatility subsequences corresponding to each time scale.

[0025] See Figure 3 In the graph, the horizontal axis represents time in hours, and the vertical axis represents electricity consumption in kilowatt-hours. The gray solid line in the graph represents the original electricity consumption time series, reflecting the actual fluctuation of electricity load over time. The blue dashed line, orange dotted line, and green dotted line represent the trend subsequences corresponding to different scales in the multi-timescale filter bank with window lengths of 3, 12, and 48, respectively. As described in Example 1, the window length... The values ​​are 3, 12, and 48, all of which are odd numbers, to ensure the symmetry of the center moving average filter.

[0026] As shown in the figure, the trend subsequence with a window length of 3 (blue dashed line) captures the short-term fluctuation trend of electricity consumption well. The curve closely follows the local fluctuations of the original electricity consumption sequence, with relatively obvious fluctuations, reflecting the trend characteristics of electricity consumption at a narrow time scale. The trend subsequence with a window length of 12 (orange dotted line) presents a smoother curve, filtering out some short-term fluctuations and revealing the medium-term trend changes of electricity consumption at a medium time scale. The trend subsequence with a window length of 48 (green dotted line) is further smoothed to reflect the overall trend at a longer time scale. The curve has the least fluctuation, showing the overall change trend of the long-term trend.

[0027] Example 2: In practice, after obtaining trend subsequences and fluctuation subsequences corresponding to multiple time scales, for each trend subsequence, the local turning point where the trend direction reverses is detected, the frequency of occurrence of local turning points within the set observation window is calculated, and the maximum duration of maintaining a unidirectional trend in the trend subsequence is extracted. The trend confidence level of the time scale is calculated by jointly using the frequency of occurrence and the maximum duration.

[0028] When detecting local turning points where the trend direction reverses in a trend subsequence, a first-order differencing operation is performed on the trend subsequence. The positions where the sign of the differencing result changes point by point are identified, and data points where the sign changes are marked as local turning points. Within intervals with consecutive identical signs, only the start and end points of the interval are retained as turning boundaries. In specific implementation, for scale indices of... The trend subsequence is obtained by arranging the trend values ​​corresponding to each time point in the trend subsequence in chronological order, and calculating the difference between the trend values ​​of two adjacent time points from front to back, thus obtaining the first-order difference sequence. Each difference value in the first-order difference sequence can be positive, negative, or zero. When identifying the position of the sign change, the time when the difference value is zero is ignored, and only the jump between positive and negative values ​​is considered. When the trend value of the first-order difference sequence is zero... The sign of the difference value is the same as the first... When the signs of the differences are different and both are not zero, the first difference value will be... The time markers corresponding to the differences are designated as local turning points. Within an interval of consecutive identical symbols, the starting point of the interval is the time when the symbol first appears, and the ending point is the time when the symbol last appears. If other marked local turning points exist within the interval, they are deleted, and only the starting and ending points of the interval are retained as turning boundaries, thus obtaining a simplified set of turning points.

[0029] The length of the observation window is set to be proportional to the time scale window length corresponding to the trend subsequence, and the observation window is truncated from the current evaluation time towards the historical direction. In specific implementation, for scale index... The trend subsequence, with the observation window length set as . Observation window length Filter window length corresponding to the trend subsequence They satisfy a direct proportional relationship, which can be expressed as: ,in It is a direct proportionality constant. The value of is set based on the sampling interval of the electricity consumption time series and the sensitivity requirements for monitoring trend changes. In some embodiments, it is a proportionality constant. The value is set to 10, making the observation window length ten times greater than the filter window length. When truncating the observation window, the right endpoint is set to the current evaluation time, extending towards the historical timeline. Each sampling point, with the left endpoint of the window representing the backward regression from the current evaluation time. The location of each sampling point. The observation window includes the current evaluation time and a total of [number missing] sampling points. The frequency of local turning points is statistically analyzed within an interval consisting of sampling points.

[0030] Within the observation window, the frequency of local turning points is counted, and the frequency is normalized to obtain the turning density index. The turning density index is marked as... Turning density index The value is obtained by dividing the frequency of local turning points by the maximum number of turning points that may appear within the observation window. The frequency of local turning points is the number of simplified local turning points included in the observation window, and the maximum number of turning points that may appear within the observation window is taken as... This corresponds to the extreme case where a turn occurs at every moment within the observation window, except for the first and last points. Turning density index The range of values ​​is a closed interval. The closer the value is to 0, the sparser the trend reversal; the closer the value is to 1, the more frequent the trend reversal.

[0031] When extracting the maximum duration of a unidirectional trend in a trend subsequence, the first-order difference sequence of the trend subsequence is scanned within the observation window to identify the longest consecutive number of sampling points maintaining the same sign. A unidirectional trend refers to the trend subsequence maintaining the same sign for its first-order difference values ​​across multiple consecutive sampling points, either all positive or all negative, indicating a continuously rising or falling trend. The maximum duration of maintaining a unidirectional trend is denoted as... Maximum duration The number of sampling points is used as the unit of measurement to represent the longest span of a unidirectional trend within the observation window.

[0032] When calculating the trend confidence level on a time scale using both occurrence frequency and maximum duration, the maximum duration will be used as the basis for the calculation. Filter window length with time scale The relative persistence index is obtained by dividing by the product, and the relative persistence index is denoted as... Relative persistence index The calculation method is as follows Relative persistence index The relative strength of the ability to characterize the persistence of a unidirectional trend with respect to the filter window length. This will be used to determine the density index. With relative persistence index The harmonic mean is performed, and the result is used as the trend confidence level on the time scale. The trend confidence level is denoted as... Trend confidence level Determined by the following harmonic mean relation:

[0033] in, For scale index The corresponding trend confidence level, This is a turning density index, with values ​​taken from a closed interval. Real numbers within, This is a relative persistence index, with values ​​taken from closed intervals. Real numbers within. Using As a measure of shift towards sparsity, it is compared with the relative persistence index. Perform a harmonic average to achieve a trend confidence level. A larger harmonic mean value indicates a more stable trend and higher confidence level when the turning point density is low and the relative persistence is high. Conversely, a smaller harmonic mean value indicates insufficient trend stability and lower confidence level when the turning point density is high or the relative persistence is low. The harmonic mean is more sensitive to smaller values ​​than the arithmetic mean, preventing misjudgments caused by a single indicator being too high and masking a lower one. Trend Confidence Level The range of values ​​is a closed interval. In practice, the trend confidence level at all time scales is calculated one by one according to the above steps to obtain the trend confidence level corresponding to each scale.

[0034] Example 3: In specific implementation, please refer to Figure 2 After calculating the trend confidence levels for all time scales, the trend confidence levels for all time scales are integrated using a weighting factor that is negatively correlated with the time scale window length to form a multi-scale trend integration vector.

[0035] When determining the weighting factor negatively correlated with the time scale window length, for each time scale, the inverse of the filter window length for that time scale is calculated. This inverse is then divided by the sum of the pre-calculated inverses of the filter window lengths for all time scales to obtain the normalization factor. This normalization factor is used as the weighting factor for the time scale. In specific implementations, the scale index is... The filter window length corresponding to the time scale is denoted as The total number of time scales is denoted as Scale Index The range of values ​​is from arrive Positive integer, filter window length Scale index increases progressively in ascending proportions according to window length. The smaller the corresponding filter window length The narrower, the more accurate the scale index. The larger the corresponding filter window length The wider.

[0036] For the scale index is Calculate the filter window length using the time scale. The reciprocal of the product is obtained. For all After calculating the reciprocal of the filter window length for each time scale, then... The sum of the reciprocals of the lengths of the filter windows is obtained by summing the reciprocals of the lengths of the filter windows, denoted as . , The calculation method is as follows ,in The scale index is represented as The time scale corresponds to the filter window length. To traverse all The summation index for each time scale. The scale index is... The weighting factor corresponding to the time scale is denoted as Weighting factor Determined according to the following relationship:

[0037] in, For scale index The corresponding weighting factor for the time scale. For scale index The corresponding time scale of the filter window length, This represents the total number of time scales in a multi-time-scale filter bank. For scale index The corresponding time scale of the filter window length, Indicates all Sum of the reciprocals of the filter window lengths for each time scale. Weighting factors. The value of satisfies That is, the sum of the weighting factors across all time scales is 1, constituting the normalized weight allocation. Due to the filter window length... The smaller the value, the more the filter window length is reciprocal. The larger the weighting factor, the greater the weighting factor. With filter window length They are negatively correlated; the trend confidence level corresponding to the narrow window length receives a larger weight in the integration process, while the trend confidence level corresponding to the wide window length receives a smaller weight in the integration process.

[0038] In practical implementation, the scale index is... Trend confidence level corresponding to the time scale Weighting factors corresponding to the time scale Multiply by these to obtain the weighted trend confidence level. Perform a multiplication operation at each time scale to obtain The weighted trend confidence level values. The weighted trend confidence level values ​​are arranged in ascending order of scale to construct a multi-scale trend ensemble vector, denoted as [vector name missing]. Multi-scale trend integrated vector For one 3D vector, multi-scale trend ensemble vector The The elements are Multi-scale trend ensemble vector The element arrangement order is strictly consistent with the time scale arrangement order from narrow to wide in the multi-time scale filter bank, so that adjacent elements are continuously progressive on the scale axis.

[0039] Example 4: In practice, after forming a multi-scale trend integration vector, the differential gradient of adjacent elements of the multi-scale trend integration vector on the scale axis is calculated, and the trend perception judgment is made in combination with the magnitude evolution direction of the integration vector corresponding to multiple continuous sampling points on the time axis.

[0040] When calculating the differential gradient of adjacent elements in the multi-scale trend ensemble vector along the scale axis, for the multi-scale trend ensemble vector generated at the current sampling point, the differences between adjacent elements are calculated in ascending order of scale. The difference is then divided by the difference in the logarithmic window length between adjacent scales to obtain the differential gradient. The multi-scale trend ensemble vector is denoted as... Multi-scale trend integrated vector The number of elements contained is equal to the total number of time scales in the multi-timescale filter bank. Multi-scale trend ensemble vector The The elements are denoted as ,in For scale index, the value is from arrive positive integers, Corresponding to the narrowest filter window length, The corresponding widest filter window length. The filter window length corresponding to each time scale is denoted as . Filter window length With scale index It increases with the increase of . The gradient of the difference between adjacent elements is denoted as . Gradient difference between adjacent elements The first in the multi-scale trend integration vector element With the element The difference, divided by the first The filter window length corresponding to each time scale With the The filter window length corresponding to each time scale The difference is calculated from the natural logarithm of the gradient. The calculation relationship is as follows:

[0041] in, For scale index The corresponding gradient of the difference between adjacent elements, The range of values ​​is from arrive Positive integers; Multi-scale trend integration vector The Middle One element; Multi-scale trend integration vector The Middle One element; For scale index The length of the filter window corresponding to the time scale; For scale index The length of the filter window corresponding to the time scale; Represents the natural logarithm operation. Gradient of the difference between adjacent elements. The numerical value reflects the rate of change of the trend confidence level across different time scales, and the gradient of the difference between adjacent elements. The positive or negative sign indicates the direction of increase or decrease in the trend confidence level as the scale is relaxed. Based on the above relationship, for... from arrive Calculate one by one, and get The gradient values ​​of the differences between adjacent elements.

[0042] In practice, a multi-scale trend ensemble vector is extracted from the current sampling point and its preceding multiple consecutive sampling points to determine the direction of modulus evolution. Let the time stamp of the current sampling point be... The number of preset sampling points used to determine the consistency of the modulus evolution direction is denoted as . , A positive integer greater than or equal to 2. Extract timestamps from... arrive Continuous The multi-scale trend ensemble vector corresponding to each sampling point. For time-labeled... The multi-scale trend ensemble vector corresponding to the sampling points is denoted as . Calculate multi-scale trend ensemble vector The L2 norm is used as the modulus, and the modulus is denoted as . , module length The calculation method is to take the square root of the sum of squares of all elements in the multi-scale trend integration vector. The modulus length is calculated for each sampling point to obtain the modulus length sequence. .

[0043] When comparing the increasing or decreasing direction of the modulus of adjacent sampling points, the magnitude relationship between two adjacent modulus values ​​in the modulus sequence is compared sequentially from front to back. For time stamps... and ,like Then determine the time stamp. The direction of increase or decrease in the modulus at that point is increasing; if If the direction of increase or decrease in the modulus is determined to be decreasing; if If so, then the direction of increase or decrease in the modulus is determined to be constant. In continuous... Within an interval consisting of sampling points, if and only if all The modulus of all adjacent comparison pairs increases or decreases in the direction of increment, or all... When the modulus of adjacent comparison pairs all decrease in direction, the modulus evolution direction is considered to be consistent. Consistent modulus evolution direction indicates that the overall trend strength of the multi-scale trend integration vector shows a continuous strengthening or weakening trend over time. In some implementations, a preset number of sampling points is used. The value of is 3. When the sampling period of the electricity consumption time series is 1 hour, the three consecutive sampling points correspond to a time span of 3 hours, which is consistent with the typical duration of short-term electricity load fluctuations. It can filter out single-point jump interference while retaining effective trend evolution information.

[0044] Example 5: In practice, the differential gradient of adjacent elements of the multi-scale trend integration vector on the scale axis is obtained, and combined with the magnitude evolution direction of the integration vector corresponding to multiple consecutive sampling points on the time axis, the power consumption trend sensing signal is output when the sign of the differential gradient remains consistent within a preset number of consecutive sampling points and the magnitude evolution direction remains consistent within another preset number of consecutive sampling points.

[0045] A preset number of continuous sampling points used to determine the consistency of the sign of the differential gradient, and another preset number of continuous sampling points used to determine the consistency of the modulus evolution direction, the values ​​of which satisfy a mapping relationship determined by the time scale configuration scheme, wherein the mapping relationship is a monotonic function of the ratio of the maximum filter window length to the minimum filter window length in the multi-time scale filter bank.

[0046] In practical implementation, the maximum filter window length within a multi-timescale filter bank is denoted as... The minimum filter window length within a multi-timescale filter bank is denoted as... Maximum filter window length The maximum value among all filter window lengths in a multi-timescale filter bank. In the case of a certain time scale Equal to the scale index is Corresponding filter window length Minimum filter window length This is the minimum value among all filter window lengths in a multi-timescale filter bank. Equal to the filter window length corresponding to scale index 1 The scale span ratio is defined as the ratio of the maximum filter window length to the minimum filter window length. Scale span ratio The calculation method is as follows Scale span ratio The value of is a real number greater than 1, and the scale span ratio is... The size reflects the breadth of the time range covered by the multi-timescale filter bank.

[0047] The preset number of consecutive sampling points used to determine the consistency of the sign of the differential gradient is denoted as . , It is a positive integer greater than or equal to 2. Another preset number of consecutive sampling points used to determine the consistency of the modulus evolution direction is denoted as . , A positive integer greater than or equal to 2. Preset number of consecutive sampling points. With another preset number of continuous sampling points The two sides satisfy a mapping relationship determined by the time scale configuration scheme, and the mapping relationship is the scale span ratio. The monotonic function. The specific form of a monotonic function is, and Each varies with the scale span ratio It increases monotonically with increasing magnitude. The following relationship holds:

[0048] in, This is the preset number of consecutive sampling points used to determine the consistency of the sign of the differential gradient. The scale span ratio is the ratio of the maximum filter window length to the minimum filter window length. Represents the natural logarithm operation. A positive proportionality coefficient Based on the offset, This indicates a rounding up operation. Scale factor. and base offset All are non-negative constants and are configured during system initialization. In one configuration method, the scaling factor... The value is 1.0, which is the base offset. The value is 2, when the scale span ratio is... It was 8 o'clock. Another preset number of consecutive sampling points The scale span ratio is obtained through the same form of relationship. Sure, Configure using independent coefficient parameters.

[0049] When determining the consistency of the sign of the differential gradient, for continuous... For each of the sampling points, calculate the difference gradient between adjacent elements on the scale axis and extract the sign if and only if the samples are continuous. The differential gradients of all adjacent elements within a sampling point are either positive or consecutive. When the signs of the differential gradients of all adjacent elements within a given sampling point are negative, it is determined that the signs of the differential gradients remain consistent within a preset number of consecutive sampling points. When determining the consistency of the modulus evolution direction, another preset number of consecutive sampling points is used. Extract the current sampling point and its preceding continuous segments. Each sampling point is compared with the direction of increase or decrease in the magnitude of adjacent sampling points; if and only if continuous If the modulus of all adjacent comparison pairs within a sampling point increases or decreases in the direction of increase or decrease, or if the modulus of all adjacent comparison pairs decreases in the direction of increase or decrease, it is determined that the modulus evolution direction remains consistent within another preset number of consecutive sampling points.

[0050] When both the consistency condition for the sign of the differential gradient and the consistency condition for the direction of modulus evolution are satisfied, a power consumption trend sensing signal is output. The power consumption trend sensing signal includes a direction indicator and a reliability level. The direction indicator is determined by the sign of the differential gradient in a continuous sequence. The sign of the sample is determined by maintaining consistency across all sampling points; a positive sign indicates an upward trend, and a negative sign indicates a downward trend. The reliability level is based on another preset number of consecutive sampling points. The number of sampling points traversed during the duration of the internal model's long evolution direction and the number of another preset consecutive sampling points. The ratio is determined. The number of sampling points traversed during the duration of the modulus evolution direction is the count of consecutive sampling points from the first sampling point in the modulus evolution direction that aligns with the current direction to the current sampling point, denoted as . Continuous sampling point counting The value range is from 1 to A positive integer. The reliability level is denoted as... Reliability level Counting from consecutive sampling points Divide by another preset number of consecutive sampling points To obtain, that is Reliability level The value of is a closed interval. Real numbers within, reliability level The closer the result is to 1, the higher the reliability level of the trend perception result. A value of 1 indicates that the modulus evolution direction is in all directions. The data points are completely consistent. The electricity consumption trend sensing signal outputs both the direction indicator and the reliability level, which are used by the electricity management department for load scheduling and trend early warning.

[0051] See Figure 4 In the figure, the horizontal axis represents time in hours, and the vertical axis represents the reliability level H of electricity consumption trend sensing, with a value range of [0,1]. The curve in the figure shows the changing trend of the reliability level H. The red upper triangle symbol indicates an upward trend, and the green lower triangle symbol indicates a downward trend. The reliability level H is determined based on the duration of the evolution direction of the multi-scale trend integration vector magnitude in Example 5 and the preset number of continuous sampling points. The ratio is calculated to reflect the reliability of the trend perception results.

[0052] The graph shows that between approximately hour 80 and hour 160, the reliability level H rapidly increased and stabilized in a high range close to 1, accompanied by numerous red upward trend indicators, indicating that the electricity consumption trend sensing system output a continuous and highly reliable upward trend signal during this period. Subsequently, the reliability level rapidly decreased to near 0, indicating the end of the trend signal. Similarly, between approximately hour 300 and hour 380, the reliability level also rapidly increased to near 1, and green downward trend indicators appeared, indicating that the electricity consumption trend system detected a continuous and highly reliable downward trend signal; after this period, the reliability level again rapidly dropped back to near 0.

[0053] The rapid rise and fall of reliability levels and the concentrated appearance of directional indicators in the figure are consistent with the positive and negative signs of the differential gradient between adjacent elements of the multi-scale trend integrated vector in Example 5. Maintaining consistency across all sampling points and ensuring the continuous evolution direction of the modulus. The trend-sensing signal characteristics are output under consistent conditions across all sampling points. The figure shows two distinct trend signal phases representing the continuous rise and fall of applied electricity, respectively, with a reliability level H value close to 1, indicating high stability of the sensed trend and a high confidence level in the trend change. The reliability level for the remaining time period in the figure is close to zero, indicating that the trend-sensing system did not detect a significant and stable trend signal.

[0054] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for sensing electricity consumption trends based on multi-timescale integrated analysis, characterized in that, include: S1: Obtain the time series data of the electricity consumption to be monitored; S2: The electricity consumption time series data is filtered by a preset multi-time-scale filter bank to obtain trend subsequences and fluctuation subsequences corresponding to multiple time scales, wherein the window lengths of each filter in the multi-time-scale filter bank are arranged in ascending order according to a set common ratio. S3: For each trend subsequence, detect the local turning point where the trend direction reverses in the trend subsequence, calculate the frequency of occurrence of the local turning point within the set observation window, and extract the maximum duration of maintaining a unidirectional trend in the trend subsequence. Calculate the trend confidence level of the time scale by jointly using the frequency of occurrence and the maximum duration. S4: The trend confidence levels across all time scales are integrated using weighted factors that are negatively correlated with the window length of that scale, forming a multi-scale trend integration vector. S5: Calculate the differential gradient of the adjacent elements of the multi-scale trend integration vector on the scale axis, and combine it with the magnitude evolution direction of the integration vector corresponding to multiple consecutive sampling points on the time axis. When the sign of the differential gradient remains consistent within a preset number of consecutive sampling points and the magnitude evolution direction remains consistent within another preset number of consecutive sampling points, output the electricity consumption trend sensing signal. Step S3, which involves jointly calculating the trend confidence level for the time scale using the frequency of occurrence and the maximum duration, includes: The frequency of occurrence is normalized to obtain the turning density index. The maximum duration is divided by the filter window length of the time scale to obtain the relative duration index. The turning density index and the relative duration index are harmonic averaged, and the result is used as the trend confidence level of the time scale.

2. The electricity consumption trend sensing method based on multi-timescale integrated analysis according to claim 1, characterized in that, Step S2, which involves filtering the electricity consumption time series data using a preset multi-time-scale filter bank, specifically includes: Determine the minimum filter window length and the common ratio of window length increments, and generate a set of filter window length sequences that are progressively increased according to the stated common ratio; For each of the filter window lengths, a central moving average is calculated on the electricity consumption time series data with that window length, and the moving average result is used as the trend subsequence corresponding to that scale. Subtracting the trend subsequence from the electricity consumption time series data yields the fluctuation subsequence at this scale.

3. The electricity consumption trend sensing method based on multi-timescale integrated analysis according to claim 1, characterized in that, The weighting factor negatively correlated with the scale window length mentioned in step S4 is determined in the following way: For each time scale, calculate the reciprocal of the filter window length for that scale, and divide this reciprocal by the sum of the reciprocals of the window lengths for all time scales obtained in advance to obtain the normalization factor. Use this normalization factor as the weighting factor for that scale.

4. The electricity consumption trend sensing method based on multi-timescale integrated analysis according to claim 1, characterized in that, Step S5, which involves calculating the differential gradient between adjacent elements of the multi-scale trend ensemble vector on the scale axis and combining this with the magnitude evolution direction of the ensemble vector corresponding to multiple consecutive sampling points on the time axis, specifically includes: For the multi-scale trend ensemble vector generated at the current sampling point, the difference between adjacent elements is calculated in order of scale from narrow to wide, and the difference is divided by the difference in logarithmic window length between adjacent scales to obtain the difference gradient. Extract the integrated vector of the current sampling point and its preceding multiple consecutive sampling points, calculate the L2 norm of each integrated vector as the modulus, compare the increasing and decreasing directions of the modulus of adjacent sampling points, and if the increasing and decreasing directions are both increasing or both decreasing within a consecutive preset number of sampling points, it is determined that the modulus evolution direction remains consistent.

5. The electricity consumption trend sensing method based on multi-timescale integrated analysis according to claim 1, characterized in that, Step S3, which involves detecting local reversal points where the trend direction reverses in the trend subsequence, specifically includes: Perform a first-order difference operation on the trend subsequence, identify the position of the positive or negative sign change of the difference result point by point, mark the data point where the sign jump occurs as the local turning point, and retain only the start and end points of the interval as the turning boundary within the interval with consecutive identical signs.

6. The electricity consumption trend sensing method based on multi-timescale integrated analysis according to claim 1, characterized in that, In step S3, the length of the observation window is proportional to the time scale window length corresponding to the trend subsequence, and the observation window is truncated from the current evaluation time towards the historical direction.

7. The electricity consumption trend sensing method based on multi-timescale integrated analysis according to claim 1, characterized in that, In step S5, a preset number of continuous sampling points used to determine the consistency of the positive and negative signs of the differential gradient, and another preset number of continuous sampling points used to determine the consistency of the modulus evolution direction, satisfy a mapping relationship determined by the time scale configuration scheme, wherein the mapping relationship is a monotonic function of the ratio of the maximum window length to the minimum window length within the multi-time scale filter bank.

8. The electricity consumption trend sensing method based on multi-timescale integrated analysis according to claim 1, characterized in that, The electricity consumption trend sensing signal in step S5 includes a direction indicator and a reliability level. The direction indicator is determined by the sign of the differential gradient when the sign remains consistent within consecutive sampling points. The reliability level is determined based on the ratio of the number of sampling points traversed during the duration of the evolution direction of the inner modulus of another preset number to the other preset number.

9. A power consumption trend sensing system based on multi-timescale integrated analysis, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the electricity consumption trend sensing method based on multi-timescale integrated analysis as described in any one of claims 1 to 8.

Images