Multi-scale bus load lean prediction method based on source-load cooperation

Abstract

The present application relates to the technical field of data processing, and particularly relates to a multi-scale bus load lean prediction method based on source-load coordination, which comprises the following steps: dividing a time scale into multiple levels such as annual, quarterly, monthly, etc., and counting original load sequences of each bus node and original output sequences of each power source according to each level; recursively propagating from a coarse time scale to a fine time scale, determining a load baseline sequence and a load residual sequence based on a load component sequence of a previous level, superimposing and correcting coupling degrees with mutual information of the load residual sequence and the power source output sequence, and then decoupling an actual load sequence into load component sequences; respectively using the load component sequences of all bus nodes of all levels to perform load prediction, and obtaining a final predicted load. The present application overcomes the defects of the prior art that different time scale coupling mechanisms are not distinguished, and improves the prediction accuracy of the local power grid bus load by multi-scale source-load coordination and layer-by-layer decoupling.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a multi-scale bus load refinement prediction method based on source-load coordination. Background Technology

[0002] Power system load forecasting typically employs machine learning models such as neural networks, using historical load data and power generation output data as input features for training and inference. When considering the source-load synergy, existing methods input both power generation output data and load demand data as unified features into the forecasting model, allowing the model to implicitly learn the coupling relationship between the source and load.

[0003] However, the manifestations of source-load coupling differ fundamentally across different time scales. On a coarse time scale (year / quarter / month), the source-load relationship is reflected as energy balance, while on a fine time scale (hour / minute), it is reflected as fluctuation propagation. However, existing methods mix data from different scales without distinguishing the differences in coupling mechanisms at different scales, resulting in insufficient learning of coupling relationships and affecting prediction results. Especially in local power grids with a high proportion of renewable energy, the high-frequency fluctuations in renewable energy output significantly increase data noise on fine time scales. Existing methods treat all time scales equally, causing fine-scale noise to pollute the prediction model through feature input. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a multi-scale bus load refinement prediction method based on source-load coordination.

[0005] The multi-scale bus load refinement prediction method based on source-load coordination of the present invention adopts the following technical solution: One embodiment of the present invention provides a multi-scale bus load refinement prediction method based on source-load coordination, the method comprising the following steps: The time scale is divided into several levels, including annual, quarterly, and monthly scales. The original load sequence collected at each bus node and the original output sequence collected at each power source output terminal are statistically analyzed according to each level to obtain the actual load sequence of each bus node and the actual output sequence of each power source under each level. Any one of these levels is denoted as the target level. The load component sequence of each bus node in the level above the target level is denoted as L0, and the coupling degree between each bus node and each power source in the level above is denoted as C0. The method for obtaining the load component sequence and coupling degree of each bus node at the target level is as follows: The load baseline sequence of the target level is determined based on the load component sequence L0. The difference between the original load sequence and the load baseline sequence is denoted as the load residual sequence. The difference between the original output sequence and the actual output sequence of each power source under the target level is denoted as the output residual sequence. The mutual information and coupling degree C0 between the load residual sequence and the output residual sequence are superimposed to obtain the coupling degree between each bus node and each power source under the target level. Based on the coupling degree of the target level, the actual load sequence of each bus node under the target level is decoupled to obtain the load component sequence of each bus node under the target level. The final predicted load for each bus node is obtained by using the load component sequence of each bus node at all levels.

[0006] Preferably, the specific steps for determining the target level load baseline sequence based on the load component sequence L0 are as follows: The sampling period for each annual scale is one year, the sampling period for each quarterly scale is one quarter, and the sampling period for each monthly scale is one month; The level above the target level is designated as the reference level; each reference level sampling period contains several target level sampling periods; each element in the load component sequence of any level corresponds to one target level sampling period; Traverse each element in the load component sequence L0, and denote the traversed element as P. The sampling period of the reference level where the element is located is denoteed as the reference period. The sampling periods of several target levels contained in the reference period are denoteed as the target periods. The baseline value x(i) of the i-th target period is P×f(i), where f(i) represents the proportion of the average electricity demand of the i-th target period to the average electricity demand of the reference period. [x(1), x(2), ..., x(N)] are the baseline sequence segments decomposed from element P, and N represents the number of target periods contained in the reference period. For all elements before element P, the baseline sequence segments of all elements are spliced ​​together end to end to form the load baseline sequence of the target level.

[0007] Preferably, the specific method for determining the load residual sequence is as follows: For any element in the load baseline sequence of each bus node under the target level, the target level sampling period corresponding to the original load sequence of each bus node, and the load amount collected within the target level sampling period, is subtracted from that element. For all elements in the load baseline sequence, the load residual sequence is obtained by subtracting them from the original load sequence.

[0008] Preferably, the specific steps for obtaining the load component sequence of each bus node at the target level are as follows: For the actual load sequence of each bus node under the target level, traverse each element in the actual load sequence and record the traversed element as Q. For the target level sampling period where Q is located, the average output of each power source within the target level sampling period is recorded as Q0. Q0 multiplied by C yields the single-source contribution of each bus node under each power source, where C represents the coupling degree between each bus node and each power source at the target level. The sum of the single-source contributions of each bus node under all power sources yields the power drive component value of each bus node at the target level. The difference between Q and the power drive component value is recorded as the load component value of each bus node at the target level. For all elements preceding element Q, the load component values ​​of all elements are concatenated together to form the load component sequence of each bus node under the target level.

[0009] Preferably, the specific steps for obtaining the final predicted load of each bus node by performing load prediction using the load component sequences of each bus node at all levels are as follows: Load forecasting is performed using the load component sequence of each bus node at each level, resulting in the load forecasting results for each bus node at each level. The output prediction result of each power source at each level is predicted using meteorological data. The product of the coupling degree between each bus node and each power source at each level and the output prediction result is used as the power drive prediction result of each bus node at each level. The sum of the power drive prediction result and the load prediction result is used as the final predicted load of each bus node at each level.

[0010] Preferably, the specific steps for obtaining f(i) are as follows: For the reference period in which element P is located, among all sampling periods under the reference level, the sampling periods before the reference period are recorded as historical reference periods; Any historical reference period containing several target level sampling periods is denoted as the historical target period; The average value of all elements corresponding to all historical reference periods within the load component sequence L0 is denoted as the first average value; the average value of the load component values ​​obtained within the i-th historical target period included in all historical reference periods is denoted as the second average value, and the ratio of the second average value to the first average value is denoted as f(i).

[0011] Preferably, when no load component value is obtained within the i-th historical target period, the element corresponding to the i-th historical target period is taken as the load component value in the actual load sequence of each bus node under the target level.

[0012] Preferably, the specific steps for obtaining the coupling degree between each bus node and each power source on an annual scale are as follows: The mutual information between the original load sequence of each bus node and the original output sequence of each power source is denoted as the initial mutual information. On an annual scale, the coupling degree between each bus node and each power source is positively correlated with the initial mutual information and negatively correlated with the topological electrical distance between each bus node and each power source. The topological electrical distance is determined by the admittance matrix of the power grid topology.

[0013] Preferably, the specific steps for obtaining the load component sequence of each bus node at the annual scale are as follows: For the actual output sequence of each power source at the annual scale, each element of the actual output sequence is multiplied by... This yields the first drive sequence for each bus node under each power supply. This indicates the coupling degree between each bus node and each power source on an annual scale; The summation of the first drive sequences of each bus node under all power sources yields the power drive component sequence of each bus node at the annual scale; the difference between the actual load sequence of each bus node at the annual scale and the power drive component sequence of each bus node at the annual scale is denoted as the load component sequence of each bus node at the annual scale.

[0014] Preferably, the specific steps for obtaining the actual load sequence of each bus node at each level are as follows: The original load sequence represents the sequence of active power at each bus node. In the original load sequence, the average active power for each year is calculated, and the sequence of the average active power for all years is recorded as the actual load sequence of each bus node at the annual scale. The average active power for each quarter is calculated, and the sequence of the average active power for all quarters is recorded as the actual load sequence of each bus node at the quarterly scale. The average active power for each month is calculated, and the sequence of the average active power for all months is recorded as the actual load sequence of each bus node at the monthly scale.

[0015] The beneficial effects of the technical solution of the present invention are: This invention divides the time scale into multiple levels, such as annual, quarterly, and monthly scales, and statistically analyzes the original load sequence collected at each bus node and the original output sequence collected at each power source output terminal according to each level. Then, it uses a recursive approach from coarse to fine time scales to determine the coupling degree between each bus node and each power source at each level. Based on this coupling degree, the actual load sequence of each bus node at each level is decoupled into load component sequences. Finally, load prediction is performed using the load component sequences of each bus node at all levels to obtain the final predicted load of each bus node.

[0016] Compared with the prior art, the present invention has the following further technical effects: (1) By treating the coupling relationship between the source and load in layers at multiple time scales, it avoids to some extent the problem of insufficient learning of coupling relationship and inaccurate prediction results caused by mixing data at different time scales and failing to distinguish the differences in coupling mechanisms at different time scales; (2) By decoupling the source and load relationship layer by layer from the coarse time scale to the fine time scale, the power drive component is stripped first at the coarse time scale with the highest signal-to-noise ratio, and then it is used as the reference benchmark for the next finer scale to process the higher frequency coupling details layer by layer. This avoids to some extent the problem of high frequency fluctuations in new energy output, random fluctuations in the load itself, and measurement noise and other interference components accumulating and contaminating the load component at the coarse time scale layer by layer during the stripping process, thus reducing the stripping error; (3) By determining the load baseline sequence of the target level based on the load component sequence of the previous level, and using the difference between the original load sequence and the load baseline sequence as the target level. The load residual sequence is used to superimpose and correct the coupling degree of the next level by using the mutual information between the load residual sequence and the actual power output sequence of the target level. The coupling degree gradually becomes richer as the time scale is refined, which can reflect the complete coupling relationship between the source and load from energy balance to fluctuation transmission, thereby realizing the dynamic and reliable quantification of the contribution of power output to the bus load; (4) By decoupling the actual load of each level into the load component sequence based on the coupling degree of the target level, the load component sequence mainly reflects the essential law of user side electricity demand and is not excessively affected by the fluctuation of new energy output, providing a reliable data basis for load prediction; (5) By using the load component sequence of each bus node of all levels to predict the load, the prediction model can identify and utilize the change law of bus load under different time scales, capture the energy balance trend at the coarse scale and capture the fluctuation transmission details at the fine scale, thereby effectively improving the prediction accuracy of the bus load of the local power grid with a high proportion of new energy access. Attached Figure Description

[0017] Figure 1 A flowchart of the steps of a multi-scale bus load refinement prediction method based on source-load coordination provided in an embodiment of the present invention; Figure 2 A flowchart illustrating the steps for obtaining the load component sequence for each bus node at each level; Figure 3 A schematic diagram comparing sampling periods at different levels; Figure 4 This is a recursive diagram of the load component sequence at different levels. Detailed Implementation

[0018] The following description, in conjunction with the accompanying drawings, details the specific scheme of the multi-scale bus load refinement prediction method based on source-load coordination provided by this invention.

[0019] Example 1: Please see Figure 1 The diagram illustrates a flowchart of a multi-scale bus load refinement prediction method based on source-load coordination provided by an embodiment of the present invention. The method includes the following steps: Step S1: Divide the time scale into several levels, including: annual scale, quarterly scale, and monthly scale; statistically analyze the original load sequence collected at each bus node and the original output sequence collected at each power source output terminal according to each level to obtain the actual load sequence of each bus node and the actual output sequence of each power source under each level.

[0020] This embodiment targets a local power grid with high penetration of new energy sources, which includes several power sources (photovoltaic power plants, wind power plants, etc.) and several bus nodes (bus nodes connect to loads). Active power time series are collected at each bus node as the raw load series, and active power time series are collected at the output end of each power source as the raw output series. In this embodiment, the above active power is collected synchronously at a sampling frequency of once per second. This embodiment uses the raw load series and raw output series for four full years from 2022 to 2025 as examples for description.

[0021] The time scale is divided into several levels, namely: annual scale, quarterly scale, monthly scale, and weekly scale; other embodiments may further divide it into more levels: daily scale, hourly scale, and minute scale.

[0022] The original load sequence and original power output sequence are statistically analyzed at each level (e.g., annual average, quarterly average, etc.) to obtain the actual load sequence for each bus node and the actual power output sequence for each power source at each level, specifically including: In the original load sequence, the average active power for each year is calculated. The sequence formed by the average active power for all years is denoted as the actual load sequence of each bus node at the annual scale. At this time, each sampling period at the annual scale is one year.

[0023] The average active power for each quarter is calculated, and the sequence of average active power for all quarters is denoted as the actual load sequence of each bus node at the quarterly scale. Each sampling period at the quarterly scale is then considered a quarter.

[0024] The average active power for each month is calculated, and the sequence of average active power for all months is denoted as the actual load sequence of each bus node at the monthly scale. In this case, each sampling period at the monthly scale is one month.

[0025] The average active power is calculated weekly, and the sequence of average active power across all weeks is denoted as the actual load sequence for each bus node on a weekly scale. Each sampling period on the weekly scale is one week.

[0026] The average active power is calculated daily, and the sequence of average active power across all days is denoted as the actual load sequence for each bus node on a daily scale. Each sampling period on the daily scale is one day.

[0027] The average active power is calculated hourly; the sequence of average active power for all hours is denoted as the actual load sequence of each bus node on the hourly scale. Each sampling period on the hourly scale is one hour.

[0028] The average active power is calculated every minute; the sequence of average active power over all minutes is denoted as the actual load sequence of each bus node at the minute scale. Each sampling period at the minute scale is one minute.

[0029] In other embodiments, the average active power can be calculated for each 2-hour period; the sequence of average active power across all periods is recorded as the actual load sequence for each bus node on an hourly scale; in this case, each sampling period on an hourly scale is 2 hours. Alternatively, the average active power can be calculated for each 15-minute period; the sequence of average active power across all periods is recorded as the actual load sequence for each bus node on a minute scale; in this case, each sampling period on a minute scale is 15 minutes.

[0030] The process of statistically analyzing the original power output sequence at each level to obtain the actual power output sequence of each power source at each level is the same as described above, and will not be further described in this embodiment.

[0031] Step S2: After decomposing the actual load into load component sequences at different levels, load prediction is performed using the load component sequences of each bus node at all levels to obtain the final predicted load for each bus node.

[0032] This step utilizes load components at all levels for load forecasting. This process decouples and separates user electricity demand at different scales, enabling the forecasting model to identify and utilize the changing patterns of bus load at different time scales. It captures energy balance trends at a coarse scale and captures the details of fluctuation transmission at a fine scale, thereby improving forecast accuracy.

[0033] Figure 2 As shown, step S2, which involves decomposing the actual load into load components at different levels, specifically includes: Step S201: Any one of the levels is denoted as the target level. The load component sequence of each bus node under the level above the target level is denoted as L0, and the coupling degree between each bus node and each power source under the level above is denoted as C0.

[0034] First, it should be noted that the actual load sequence of each bus node at each level includes the coupling of power output and user electricity demand.

[0035] In this embodiment, the load component sequence at each level represents the load component (i.e., the user's electricity demand) that mainly reflects the user's electricity consumption patterns after deducting the influence of power output, which is included in the actual load sequence. The difference between the actual load sequence at each level and the load component sequence at each level represents the contribution of power output to the bus node load, i.e., the power-driven component sequence. In other words, the actual load sequence at each level can be decoupled and separated into two parts: the load component sequence and the power-driven component sequence.

[0036] Coupling degree describes the degree of influence of power source output on bus nodes, indicating the strength of the coupling or correlation between the two, and is used to characterize the degree of influence of power source output change on load change of each bus node.

[0037] It should be noted that, for ease of understanding of coupling degree, this embodiment considers it as an approximation of the following physical quantity: how much of that unit output remains after each power source delivers its unit output to each bus node. This physical quantity cannot be directly fitted because: firstly, multiple power sources operate simultaneously in the power grid, making it difficult to separate the contribution of a single power source; secondly, power flow sensitivity itself varies with the operating point, and this physical quantity is not a constant. Based on this, this embodiment uses the coupling degree index to approximate or numerically evaluate this physical quantity. This embodiment evaluates this physical quantity based on the mutual information between the output of each power source and the contribution of the power source output to the bus node.

[0038] The greater the coupling between any power source and any bus node, the stronger the correlation or coupling between the bus node and the power source, and the more significant the driving effect of the power source output change on the bus node load change.

[0039] Further explanation regarding the load component sequence and coupling degree is that each level corresponds to a load component sequence and coupling degree, and the load component sequence and coupling degree corresponding to each level are recursively derived from the previous level (coarse time scale).

[0040] This embodiment sequentially peels off or decouples the load component sequence from a coarse time scale to a fine time scale because: At a coarse timescale (annual), high-frequency fluctuations in power output are smoothed out during long-period averaging, and the transmission efficiency between source and load tends to stabilize. At this point, the stripping of the power-driven component has the highest confidence. However, when stripping begins at the finest timescale (minute), the data contains multiple interference components, including short-term fluctuations in new energy sources, random fluctuations in the load itself, and measurement noise. These interference components are difficult to separate effectively at the fine timescale, leading to a significant increase in stripping error. More importantly, this error accumulates and propagates upwards along the timescale hierarchy: the stripping deviation at the fine scale will successively contaminate the load components at the daily, monthly, and even annual scales, ultimately rendering the coarse timescale prediction results completely unreliable. Therefore, this embodiment chooses a recursive direction from coarse to fine granular. Essentially, it prioritizes stripping at the coarse timescale with the highest signal-to-noise ratio to obtain a high-confidence load component. This baseline is then used as a reference for the next finer scale, processing higher-frequency coupling details layer by layer downwards, forming a progressive stripping strategy from stable to fluctuating, and from low-risk to high-risk.

[0041] Furthermore, the reason why this embodiment calculates the coupling degree sequentially from the coarse time scale to the fine time scale is as follows: On a coarse timescale (e.g., an annual scale), source-load coupling is mainly manifested as an energy balance relationship (the macroscopic matching between the total annual power output of the power source and the total annual load of the bus). This coupling relationship is most directly and stably constrained by the power grid topology. On a fine timescale, coupling involves multiple layers of demand-side factors such as market response and user behavior. The coupling relationship is complex and time-varying, and is not suitable as a basis for constructing the degree of coupling.

[0042] In subsequent processes, when this embodiment iterates the coupling degree of each level in turn, it superimposes the coupling degree of a level (i.e., C0) on a finer scale (i.e., the target level), so that the source loads in the target level will inevitably be superimposed with the coupling relationship that was not captured by a level.

[0043] Step S202: Determine the load baseline sequence of the target level based on the load component sequence L0. The difference between the original load sequence and the load baseline sequence is recorded as the load residual sequence. The difference between the original output sequence and the actual output sequence of each power source under the target level is recorded as the output residual sequence. The mutual information and coupling degree C0 between the load residual sequence and the output residual sequence are superimposed to obtain the coupling degree between each bus node and each power source under the target level.

[0044] The load baseline sequence represents the benchmark of user electricity demand at the target level after deducting fluctuations in renewable energy output. It only reflects the regularity of user behavior at the target level (such as seasonality) and does not include actual fluctuations at the target level.

[0045] The difference between the load baseline and the load component obtained after subsequent decoupling is that the former is the reference benchmark before decoupling, while the latter is the product after decoupling, which includes the actual power demand of users at this level, but deducts the power output influence.

[0046] The difference between the original load sequence and the load baseline sequence is denoted as the load residual sequence at the target level. Since the load baseline sequence only reflects the user's electricity demand pattern and does not include the influence of power output, while the original load sequence includes power output fluctuations, the load residual sequence obtained by subtracting the two represents the power output situation perceived at the bus node. It is used to describe what load remains at the bus node after deducting the user's basic electricity demand at the target level.

[0047] The difference between the original output sequence and the actual output sequence of each power source under the target level is denoted as the output residual sequence of each power source under the target level, which describes the fluctuation of the original output sequence on the average level of power output of the power sources under the target level.

[0048] For the load residual sequence (i.e., the power output perceived at the bus node) and the output residual sequence of each power source at the target level, calculate the mutual information between them. This mutual information represents the coupling between the bus node and the power source as evaluated only at the target level. The larger the mutual information, the more coupled or consistent the power output perceived at the bus node at the target level is with the actual fluctuation of each power source.

[0049] The coupling degree C0 is fused with the mutual information to obtain the coupling degree at the target level. This fusion operation is equivalent to superimposing the coupling situation at the target level only with the cumulative coupling situation at all previous levels, so that the coupling degree gradually becomes richer with the scale refinement, and can reflect the complete coupling relationship between source and load from energy balance to wave conduction at different time scales.

[0050] Step S203: Decouple the actual load sequence of each bus node under the target level based on the coupling degree of the target level to obtain the load component sequence of each bus node under the target level.

[0051] Based on the coupling degree of the target level, the actual load at the target level is decoupled. That is, according to the relationship between the bus node and the power source, the influence of the power output at the bus node (power source drive component) is subtracted from the actual load to obtain the load component that mainly describes the user's electricity demand. This process achieves the goal of peeling away the influence of the power source layer by layer from coarse to fine, so that the load component at each level can accurately reflect the essential characteristics of the user's electricity consumption behavior at that time scale.

[0052] As a preferred implementation, the load baseline sequence for the target level is determined based on the load component sequence L0, including the following methods: The level above the target level is designated as the reference level, and one sampling period of the reference level contains multiple sampling periods of the target level.

[0053] For example, if the reference level is on an annual scale, then one sampling period for the reference level is one year. If the target level is on a quarterly scale, then one sampling period for the quarterly scale is one quarter. A year contains four quarters, meaning one sampling period for the reference level includes four sampling periods for the target level. Figure 3 As shown.

[0054] For example, if the reference level is on a quarterly scale, then one sampling period for the reference level is one quarter. If the target level is on a monthly scale, then one sampling period for the monthly scale is one month. A quarter contains three months, meaning one sampling period for the reference level includes three sampling periods for the target level. Figure 3 As shown.

[0055] Each element in the load component sequence L0 represents the average electricity demand (i.e., average power) of a user within each sampling period when statistics are performed using the sampling period of the reference level. For example, when the reference level represents an annual scale, each element represents the average electricity demand for the year. As another example, when the reference level represents a quarterly scale, each element represents the average electricity demand for the quarter.

[0056] Traverse each element in the load component L0, and denote the traversed element as P. The sampling period of the reference level where the element is located is denoteed as the reference period (for example, the reference period represents a certain year, and element P represents the average electricity demand in a certain year). The sampling periods of several target levels contained in the reference period are denoteed as target periods. The baseline value x(i) of the i-th target period is P×f(i), where f(i) represents the proportion of the average electricity demand in the i-th target period to the average electricity demand in the reference period. This value is a statistical result, i=1, 2, ..., N, where N represents the number of target periods contained in the reference period.

[0057] For example, when the reference level represents an annual scale (i.e., the reference period represents a certain year), N=4, representing the first, second, third, and fourth quarters of that year, respectively. f(i) represents the proportion of average electricity demand in the i-th quarter to the average electricity demand for the entire year. As another example, when the reference level represents a quarterly scale (i.e., the reference period represents a certain quarter), N=3, representing the first, second, and third months of that quarter, respectively. f(i) represents the proportion of average electricity demand in the i-th month to the average electricity demand for the entire quarter.

[0058] [x(1), x(2), ..., x(N)] are the baseline sequence fragments decomposed from element P. For all elements before element P (including element P), the baseline sequence fragments of all elements are spliced ​​together end to end to form the load baseline sequence of the target level.

[0059] In summary, after each element is traversed in load component L0, a load baseline is obtained. As elements are traversed continuously, the load baseline becomes longer and more complete. When the traversal is completed, the complete load baseline of the target level is obtained.

[0060] As a preferred implementation, the difference between the original load sequence and the load baseline sequence is denoted as the load residual sequence at the target level, including the following methods: The above yielded the load baseline sequence for each bus node at the target level. Each element corresponds to a target level sampling period, representing the average electricity demand within that target level sampling period. The load (i.e., active power) collected within the target level sampling period from the original load sequence of each bus node is subtracted from this element.

[0061] For all elements in the load baseline sequence, subtract them from the original load sequence using the method described above to obtain the load residual sequence.

[0062] As a preferred implementation, the specific method for obtaining the mutual information determined between the load residual sequence and the actual output sequence of each power source at the target level is as follows: The difference between the original output sequence of each power source and the actual output sequence of each power source at the target level is denoted as the output residual sequence of each power source at the target level.

[0063] Calculate the mutual information between the load residual sequence and the output residual sequence.

[0064] As an example, the difference between the original output sequence of each power source and the actual output sequence of each power source at the target level is denoted as the output residual sequence of each power source at the target level, and includes the following methods: This method is similar to the process of obtaining the load residual sequence: For each power source's actual output sequence at the target level, any element corresponds to a target level sampling period, representing the average output of the power source within that target level sampling period. This element is subtracted from the output of each power source (i.e., the active power output) collected within that target level sampling period in the original output sequence.

[0065] For all elements in the actual output sequence, subtract them from the original output sequence according to the above method to obtain the output residual sequence.

[0066] As an example, the method for obtaining mutual information is as follows: For any one of the load residual sequence and the output residual sequence, obtain the maximum and minimum values ​​of all elements in the sequence, and divide the closed interval formed by the minimum and maximum values ​​into several sub-intervals. In this embodiment, it is divided into 10 to 20 sub-intervals, with 15 sub-intervals being the preferred embodiment. Except for the last sub-interval, all other sub-intervals are left-closed and right-open.

[0067] Two elements with the same index in the load residual sequence and the output residual sequence are considered as a sample. For the a-th sub-interval of the load residual sequence and the b-th sub-interval of the output residual sequence, if the first element of a sample is in the a-th sub-interval and the second element is in the b-th sub-interval, then it means that the sample falls into the a-th sub-interval and the b-th sub-interval.

[0068] The ratio of the number of samples falling into the a-th and b-th subintervals to the total number of samples is denoted as the joint probability distribution. Obtain the marginal probability distribution , where H represents the number of sub-intervals, and mutual information is calculated based on the joint probability distribution and the marginal probability distribution; the specific calculation formula for mutual information is well known and will not be described in detail in this embodiment.

[0069] Mutual information can capture the nonlinear statistical dependencies between sequences. Compared with the traditional Pearson correlation coefficient, mutual information is more suitable for describing the complex nonlinear coupling relationship between new energy output and load.

[0070] As a preferred implementation, the mutual information is fused with the coupling degree C0 to obtain the coupling degree of the target layer, including the following methods: The coupling degree between each bus node and each power source at the target level is C = C0 + λc, where c represents the mutual information and λ represents an adjustment coefficient greater than 0 and less than 1. The initial value is 0.37 on a quarterly scale, 0.21 on a monthly scale, 0.24 on a weekly scale, 0.17 on a daily scale, 0.13 on an hourly scale, and 0.08 on a minute scale.

[0071] As a preferred implementation, the actual load at the target level is decoupled based on the coupling degree of the target level to obtain the load component sequence of each bus node at the target level. This includes the following methods: For the actual load sequence of each bus node under the target level, traverse each element in the actual load sequence and denote the traversed element as Q, which represents the average active power within a certain sampling period of the target level; denote the average output of each power source within the sampling period as Q0, where Q0 is also the element with the same index as Q in the actual output sequence of each power source under the target level.

[0072] Element Q is processed as follows: Multiplying Q0 by C yields the single-source contribution of each bus node under each power source; the single-source contribution of each bus node under all power sources is then obtained, and these single-source contributions are summed to obtain the power drive component value of each bus node at the target level. The single-source contribution and power drive component values ​​are components determined based on source-load decoupling based on coupling degree, used to characterize the driving effect of power output on the bus node load.

[0073] The difference between Q and the power drive component value is recorded as the load component value of each bus node under the target level.

[0074] For all elements preceding element Q (including element Q), the power drive component values ​​of all elements are concatenated together to form the power drive component sequence of each bus node under the target level; the load component values ​​of all elements are concatenated together to form the load component sequence of each bus node under the target level.

[0075] In summary, after each element is traversed in the actual output sequence, a power drive component sequence and a load component sequence are obtained. As the elements are traversed continuously, these two sequences become longer and more complete. When the traversal is completed, the complete power drive component sequence and load component sequence of each bus node under the target level are obtained.

[0076] As an example, the method for obtaining f(i) above is as follows: First, it should be noted that, as seen from the above processes for obtaining the load baseline sequence and load component sequence, neither the load baseline sequence nor the load component sequence is obtained completely at once. This means that f(i) can be calculated based on the already obtained load component values, then the load baseline can be calculated based on f(i), and then load component values ​​can be obtained again based on the load baseline. This further implies that f(i) is a dynamically updated proportion that is gradually rolled over. The specific method for obtaining f(i) is as follows: For all sampling periods contained in the reference level, for the sampling period where element P is located (i.e., the reference period), the sampling periods before this sampling period are denoted as historical reference periods.

[0077] For example, when the reference level is annual, all sampling periods included under the reference level refer to all years; as another example, when the reference level is quarterly, all sampling periods included under the reference level refer to all quarters of all years; and when the reference level is monthly, all sampling periods included under the reference level refer to all months of all years.

[0078] It is easy to know that any historical reference period also contains several (N) target-level sampling periods, denoted as historical target periods. Each historical target period corresponds to a load component value, which represents the average electricity demand of users after deducting the influence of power output in that historical target period.

[0079] It is also easy to know that any historical reference period corresponds to an element in the load component L0, representing the average electricity demand of users within that historical reference period.

[0080] The average of the elements corresponding to all historical reference periods is denoted as the first average. The average of the load component values ​​of the same (e.g., the i-th) historical target period under all historical reference periods is denoted as the second average, and the ratio of the second average to the first average is denoted as f(i).

[0081] It should be noted that when a power outage occurs (power outages are occasional events), the active power collected in step S1 is considered a missing value. To avoid interference from the power outage on the prediction, this embodiment performs interpolation on the missing value, for example, using the PCHIP algorithm. Other embodiments may also use the Hermite interpolation algorithm or the Lagrange interpolation algorithm. Therefore, in this embodiment, there is no case where the first mean is equal to 0.

[0082] For ease of understanding, this embodiment uses an annual scale for the reference level and a quarterly scale for the target level as an example to briefly describe the process of obtaining f(i): The year in which element P exists (denoted as the current year) is defined. For historical years prior to this year, the average annual electricity demand (i.e., the first mean) is calculated across all historical years. The average electricity demand for the i-th quarter (i.e., the second mean) is also calculated across all historical years. The ratio of the average electricity demand for the i-th quarter to the average annual electricity demand is defined as f(i). The second mean refers to the mean of the load component values. Its purpose is to ensure that the second mean represents the average electricity demand of users after deducting the impact of power output, thereby ensuring that the subsequent load residual can mainly describe: what load remains at the bus node after deducting the basic electricity demand of users (specifically recorded in step S202).

[0083] In summary, the process of obtaining f(i) involves using the ratio of the average electricity demand of each quarter to the average annual electricity demand in historical data. This ratio is then used to allocate the average electricity demand of the current year to each quarter of the current year, thereby obtaining the baseline values ​​of each quarter of the current year, which are the aforementioned [x(1), x(2), ..., x(N)].

[0084] Specifically, for the i-th historical target period included in all historical reference periods, when there is no corresponding load component value (that is, when the load component value has not yet been obtained because the actual load sequence has not yet been traversed), the following method is used for prediction: For the actual load sequence of each bus node under the target level, the element corresponding to any historical target period represents the average electricity demand of users under that historical target period (e.g., under a certain quarter). This element does not deduct the influence of power output, but in this embodiment, it is still used as the load component value corresponding to that historical target period. The purpose is that, since this operation does not deduct the influence of power output, the calculated f(i) has a significant error. However, as mentioned above, f(i) is a dynamically updated proportion that is gradually rolled over. This operation only provides initial or starting conditions for rolling updates. As each element in the actual output sequence is traversed, the load component value is calculated, allowing the influence of power output to be deducted, thereby increasing the accuracy of f(i). However, the accuracy is lower in the early stages of traversal. Since this embodiment involves the original load sequence and original output sequence for 4 years, it means that the f(i) with lower accuracy in the early stages will not participate excessively in the subsequent load forecasting process. When load forecasting is performed later, f(i) has been updated to a relatively accurate value. Therefore, the problem of significant error in f(i) in the early stages of traversal can be ignored.

[0085] Thus, the above provides the recursive relationship between the load component sequence and the coupling degree.

[0086] Furthermore, this embodiment provides the initial conditions for the above recursive relationship, that is, the calculation method for the load component sequence and coupling degree of each bus node at the first level (i.e., the annual scale), as follows: (1) The method for obtaining the coupling degree between each bus node and each power source at the annual scale level is as follows: The annual-scale source-load coupling is mainly manifested as an energy balance relationship (macroscopic matching between the total annual output of the power source and the total annual load of the bus). Therefore, in this embodiment, the mutual information between the original load sequence of each bus node and the original output sequence of each power source is obtained and denoted as the initial mutual information. The coupling degree at the annual-scale level is positively correlated with the initial mutual information.

[0087] Furthermore, this embodiment obtains the complete topology of the power grid, including nodes such as buses, power sources, transmission lines, and transformers (whose impedance, admittance, turns ratio, and other parameters are known), and constructs the node admittance matrix Y. The construction method is derived from the standard theory of power system flow calculation and is a well-known technique; therefore, this embodiment will not elaborate on it in detail.

[0088] For any two nodes m and n in a power grid (e.g., a pair of buses and a power source), the element Y(m,n) of Y represents the injected current response at node m when a unit voltage is applied to node n and the voltages of the other nodes are zero. The modulus |Y(m,n)| of Y(m,n) characterizes the degree of electrical connection between nodes m and n; the larger |Y(m,n)| is, the closer the electrical connection between nodes m and n. Generally, the topological electrical distance d(m,n) between nodes m and n is calculated using the following formula: d(m,n) = |Z(m,m) + Z(n,n) - 2Z(m,n)|, which represents the equivalent Thevenin impedance between nodes m and n. When m = n, d(m,n) = 0, where Z represents the inverse matrix of Y.

[0089] It should be noted that the power flow analysis of the power system in this embodiment was performed using a per-unit method, so the admittance matrix Y is dimensionless. In this embodiment, d(m,n) is also dimensionless, and its value falls between 0.001 and 0.5.

[0090] For the topological electrical distance between each bus node and each power source, the coupling degree at the annual scale level is negatively correlated with the topological electrical distance; that is, the closer the topological electrical distance, the greater the coupling relationship (coupling degree) between the two.

[0091] In the above process, the closer the electrical distance between the topologies and the stronger the dependence between the power output and the load of the bus node (i.e., the greater the mutual information), the greater the coupling degree.

[0092] As an example, the formula for calculating the coupling degree at the annual scale level is: ,in This indicates the degree of coupling at the annual scale level. This represents an exponential function with the natural constant as its base. Indicates mutual information, Indicates topological electrical distance. This represents the attenuation coefficient, which ranges from 2 to 8, with an initial value of 4.1. The fusion weight has a value range of 0 to 1, with an initial value of 0.67.

[0093] In the above formula, the coupling degree is regarded as the superposition of two parts. The first part is determined by the power grid structure and remains unchanged at any level. The second part is determined by the dependence of power output and bus node load. When the above process recursively calculates the coupling degree at a fine time scale, it is to successively superimpose the second part at different levels, so that the coupling degree of the next level (i.e., C0) is superimposed on a finer scale (i.e., the target level). This ensures that the coupling relationship between the source and load in the target level is superimposed on the coupling relationship that was not captured by the previous level. It is equivalent to performing residual correction of the coupling degree for each level based on the dependence between the residual sequences (between the load residual sequence and the power output residual sequence).

[0094] (2) The method for obtaining the load component sequence of each bus node at the annual scale is as follows: For the actual output sequence of each power source at the annual scale, each element of the actual output sequence is multiplied by... This yields the first drive sequence for each bus node under each power supply. Obtain the first drive sequence of each bus node under all power sources. Summing these first drive sequences (summing elements with the same index) yields the power drive component sequence of each bus node on an annual scale.

[0095] The difference between the actual load sequence of each bus node and the power drive component sequence of each bus node at the annual scale is denoted as the load component sequence of each bus node at the annual scale.

[0096] In summary, after providing the starting conditions for the recursion, the actual load sequence of each bus node at each level can be decoupled into a load component sequence and a power drive component sequence through the recursive relationship. The former reflects the power demand of users at each bus node, while the latter represents the contribution of all power outputs to each bus node. Figure 4 The diagram illustrates the recursive process of the load component sequence for each bus node in the first three levels.

[0097] As a preferred implementation, the final predicted load for each bus node is obtained by performing load prediction using the load component sequences of all levels, including the following methods: In this embodiment, a prediction model is assigned to each level, and different prediction models are allowed for different levels. The load component sequence of each bus node under each level is input into the prediction model corresponding to each level to obtain the load prediction result of each bus node under each level.

[0098] In addition, this embodiment takes into account that the power output (i.e., its output active power) is a quantity strongly correlated with the weather. Therefore, meteorological data is obtained through weather forecasts, and the meteorological data is input into the photovoltaic / wind power output model to predict the power output prediction result of each power source at each level. For the coupling degree between each bus node and each power source at each level, the product of the coupling degree and the power output prediction result is used as the power drive prediction result of each bus node at each level.

[0099] The sum of the above power drive prediction results and load prediction results is used as the final predicted load for each bus node at each level.

[0100] For example, in the actual forecasting process, we first select one level from all levels, such as the monthly scale. Then, during the forecasting process described above, we can predict the average load of each bus node within the next month. As another example, if we select the weekly scale, the above process can predict the average load of each bus node within the next week.

[0101] As an example, taking the week-scale approach, the method for predicting the output of each power source at each level includes: By inputting meteorological data for the next week into the photovoltaic / wind power output model, the average output (i.e., the average active power output) of the power source (photovoltaic unit or wind turbine) in the next week can be predicted.

[0102] Example 2: The prediction models assigned to each level in the above embodiments are existing conventional models, which will be described with examples in this embodiment: (1) Annual scale forecast.

[0103] For the load component sequence of each bus node at the annual scale, which represents the annual average electricity demand change of users over the years, the changes are mainly driven by long-term factors such as regional economic development and industrial restructuring. Since the data volume of the load baseline sequence is small (this embodiment only includes the annual average user electricity demand for 4 years), this embodiment uses univariate linear regression for trend extrapolation: the elements in the load baseline sequence are the dependent variables, and the year is the independent variable. A linear trend line (i.e., a linear model) is fitted using the least squares method.

[0104] By inputting the next year into the linear trend line, you can obtain the load forecast results for each bus node on an annual scale, which is the average load (or average electricity demand) for the coming year.

[0105] The process described above, which involves fitting a linear trend line based on the load baseline sequence, constitutes the construction of the prediction model. The process of obtaining the load prediction results is the prediction process of the prediction model.

[0106] In some embodiments, when the amount of data in the load baseline sequence is large, such as more than 4, only the last 4 elements are used as the dependent variable and the years of the most recent 4 years are used as the independent variables for fitting, so as to avoid the problem that changes in electricity demand in too distant years cannot reflect the current electricity consumption status of users.

[0107] (2) Quarterly-scale forecasting.

[0108] After obtaining the average user electricity demand for the next year, the next year is used as the reference period as described in Example 1. At this time, the four quarters within the next year are used as the four target periods included in the reference period. Using the method of obtaining f(i) in Example 1, the proportion of electricity demand in the i-th quarter of the next year to the electricity demand in the reference period is obtained. The product of this proportion and the average electricity demand for the next year is used as the load forecast result of each bus node at the quarterly scale.

[0109] The above process utilizes the seasonal cyclical pattern of electricity load, that is, the proportion of the average electricity demand of the same bus in the same quarter of each year to the average annual electricity demand has statistical stability.

[0110] The process of constructing the prediction model is based on the load component sequence of each bus node at the quarterly scale and obtaining the above proportion using the f(i) acquisition method. The process of obtaining the load prediction result based on this proportion is the prediction process of the prediction model.

[0111] (3) Monthly scale forecasting.

[0112] The load component sequence of each bus node at the monthly scale is input into the STL (Seasonal-Trend decomposition using Loess) algorithm to obtain three parts: trend component, periodic component, and residual component.

[0113] The trend components are extrapolated to the next month using linear regression to obtain the trend prediction results. Specifically, the elements in the trend components are the dependent variables, and the element indices are the independent variables. A linear trend line (i.e., a linear model) is fitted using the least squares method. The index of the last element in the trend component is incremented by 1 and then input into the linear trend line to obtain the trend prediction results for the next month.

[0114] Since the months in the trend component appear cyclically, the month cannot be directly used as the independent variable. The index of the last element plus 1 indicates the index of the next month in the trend component.

[0115] In some embodiments, considering that the trend component contains a relatively large number of elements (e.g., greater than or equal to 24), the linear model can be replaced with a nonlinear model, such as a 5th-order polynomial or a particle filter model.

[0116] Furthermore, the average value of the same period in history is directly reused for the periodic component to obtain the periodic prediction result; specifically, each element in the periodic component corresponds to a month, and the average value of the elements that belong to the same month as the next month is used as the periodic prediction result.

[0117] In this embodiment, the residual result corresponding to the residual component is set to the mean of the residual component; in other embodiments, the residual result can also be set to 0.

[0118] Finally, the trend forecast results, cycle forecast results, and residual results are added together to obtain the load forecast results for each bus node on a monthly scale.

[0119] The process of obtaining trend prediction results, periodic prediction results, and residual results based on the load component sequence is the construction process of the prediction model, and the process of calculating the trend prediction results, periodic prediction results, residual results, and obtaining the load prediction results is the prediction process of the prediction model.

[0120] (4) Daily scale prediction.

[0121] The load forecast results for each bus node at the daily scale include the average daily load (or average daily user electricity demand) over historical dates. In this embodiment, historical dates are classified into weekdays (Monday to Friday, excluding public holidays), weekends (Saturday and Sunday, excluding public holidays), and public holidays (statutory public holidays and their adjusted workdays) according to the following rules.

[0122] For the next day, retrieve all historical dates that belong to the same date category as the next day and record them as the same category day. Each same category day corresponds to an element (i.e., the average daily load) in the load forecast result.

[0123] By summing the elements corresponding to all days of the same category using weighted averages, the load forecast results for each bus node at the daily scale are obtained.

[0124] As an example, the formula for calculating the weight of each element in a weighted summation is: Sort all elements corresponding to the same category of days in chronological order and number them sequentially starting from 1; the ratio of each element's number to the sum of all element numbers is used as the weight of each element.

[0125] In the above process, the classification of dates and the selection of elements of the same date type for the next day constitute the construction process of the prediction model, and the weighted summation process constitutes the prediction process of the prediction model.

[0126] (5) Weekly-scale prediction.

[0127] The load component sequence of each bus node at the weekly scale represents the historical weekly average load (or weekly average user electricity demand). Its variation characteristics lie between monthly and daily variations; therefore, this embodiment combines the monthly forecasting process with the daily forecasting results. Specifically: This embodiment uses the above-mentioned monthly-scale prediction method to predict the load component sequence at the weekly scale, and obtains the first prediction result for the next week. The monthly scale can be directly replaced with the weekly scale. The specific process will not be described in this embodiment.

[0128] In addition, for all days of the coming week, the average of the load forecast results obtained for all days in (4) is calculated to obtain the second forecast result for the coming week. The average of the second forecast result and the first forecast result is used as the load forecast result for each bus node on a weekly scale.

[0129] (6) Hourly / minute scale prediction.

[0130] The hourly / minute scale reflects short-term fluctuations in user electricity consumption behavior at the hourly or even minute level. Its variation is influenced by multiple factors, including date type, time of day, temperature, and user behavior patterns. While exhibiting high fluctuation frequency, it also shows a certain degree of daily periodic repetition. This embodiment employs a K-Nearest Neighbors (KNN)-based prediction method, suitable for real-time prediction of fine-scale intrinsic load. Each hour or minute can be considered as a time period; in other embodiments, 2 hours or 15 minutes may also be considered as a time period. The load component sequence of each bus node at the hour / minute scale represents the average load (or the average user electricity demand) for each time period.

[0131] For a future time period, a multi-dimensional feature vector is constructed to describe the contextual conditions of that time period. This feature vector includes: date type (weekday / weekend / holiday), represented by one-hot encoding; time period number (indicating which time period of the day, for example, 0-23 on an hourly scale, representing the hour of the day); neighboring time period load (elements from the previous and even earlier time periods of the future time period, i.e., the corresponding elements in the load component sequence), used to capture the short-term continuity of the load; and temperature (if any), the forecast temperature value for the future time period, reflecting the impact of temperature on electricity consumption behavior.

[0132] Furthermore, for each historical time period in the load component sequence of each bus node at the hour / minute scale, a feature vector is assigned to each historical time period using the method described above.

[0133] Calculate the Euclidean distance between the feature vector corresponding to each historical time period and the feature vector of the future time period. Note that before calculating the distance, all feature vectors of historical time periods and the feature vectors of the future time periods are normalized (each dimension is subtracted from its mean and then divided by its standard deviation) to eliminate differences in dimensions. Select the K historical time periods with the smallest distance (K is usually 5~15, and K=10 is optimal in this embodiment) as the nearest neighbor samples.

[0134] The elements corresponding to the selected K nearest neighbor samples (that is, the elements corresponding to the load component sequence) are weighted and summed to obtain the predicted value, which is the load prediction result of each bus node on the hour / minute scale.

[0135] As an example, the method for obtaining the weight of each element in a weighted summation is as follows: For each element corresponding to the K nearest neighbor samples, the weight of each element is negatively correlated with the Euclidean distance between the feature vectors corresponding to that element. In other words, the closer the feature vector is to the future time period, the more attention is paid to that element.

[0136] Specifically, for any Euclidean distance q between the feature vectors corresponding to an element, exp(-q) is denoted as the attention coefficient of that element, and the ratio of the attention coefficient of each element to the sum of the attention coefficients of all elements is used as the weight of each element.

[0137] The process of obtaining the elements corresponding to the K nearest neighbor samples based on the load component sequence is the construction process of the prediction model, and the weighted summation process is the prediction process of the prediction model.

[0138] It should be noted that, for any level, such as the annual scale, after obtaining the load forecast result for that level (i.e., the next year), when forecasting any other level below that level, such as forecasting the load forecast result for each month in the next year, the load forecast result for the first month of the next year is first predicted. After the first month, the load forecast result for the second month is then predicted. When predicting the load forecast result for the second month, the active power time series collected at each bus node and the active power time series collected at each power output terminal in the first month need to be spliced ​​together with the original load sequence and the original output sequence described in step S1, and then the steps S1 to S2 of Embodiment 1 are executed again to obtain the load forecast result for the second month.

[0139] Note that when repeating steps S1-S2 of Embodiment 1 to obtain the load forecast results for the second month, since it is less than a year and less than a quarter, the average active power of the current year and the current quarter cannot be calculated when obtaining the actual load sequence in S1. This results in insufficient length of the actual load sequence at the annual and quarterly scales (lacking the average active power of the current year and the current quarter), which in turn leads to the missing last element of the load component sequence at the annual and quarterly scales obtained in S2. To address this issue, this embodiment uses the load forecast results at the annual and quarterly scales as the last element of the load component sequence at the annual and quarterly scales.

[0140] In some embodiments, if the load forecast for the first month of the coming year is predicted before the first month has ended (e.g., only the middle of the month has passed), then the load forecast for the second month is not predicted, and a prompt is made that only other levels at the monthly scale can be selected for forecasting (e.g., the load forecast for the current week or day).

[0141] In other embodiments, the above-described prediction model can be used to predict the load forecast results for multiple future months (e.g., three months) at once. However, since predicting multiple months at once can result in lower accuracy, this embodiment is implemented as an optional embodiment.

[0142] In addition, it should be noted in this embodiment that the photovoltaic / wind power output model is also a well-known and conventional model. This embodiment will provide an example as follows: In one embodiment, the photovoltaic power output model is based on a physical model of solar radiation conversion (single-diode / dual-diode equivalent circuit model). Based on meteorological data (solar irradiance, ambient temperature) and component parameters for each future time period, the output power (i.e., the predicted output of photovoltaic power for each future time period) is directly calculated using the IV characteristic equation of the photovoltaic cells. The core parameters of this model (series resistance, parallel resistance, ideality factor, photocurrent, and reverse saturation current) are obtained from manufacturer datasheets or through field calibration. The model accuracy depends on the accuracy of the irradiance forecast; under clear sky conditions, the error can be controlled within 3%.

[0143] The output forecasts for each day / week, or even each month / quarter / year, are equal to the average of the output forecasts for all periods included.

[0144] If it is impossible to predict meteorological data for each period of the month / quarter / year, the average solar irradiance and ambient temperature for the month / quarter / year can be predicted directly. Weather prediction is a well-known technique and will not be described in detail in this embodiment.

[0145] In other embodiments, the irradiance, cloud cover, and temperature forecast fields output by meteorological models (such as WRF-Solar and ECMWF) are used as input. Combined with the geographical location, tilt angle, and azimuth angle of the photovoltaic power station, the effective irradiance received by the modules is calculated using a slope irradiance conversion model (such as the Perez model or Hay-Davies model). Then, the power output prediction is obtained through a physical model of solar radiation conversion. This is currently the mainstream technical route for day-ahead / intra-day photovoltaic forecasting.

[0146] In some embodiments, the wind speed-power characteristic curve of the wind turbine (provided by the turbine manufacturer) is used to convert the wind speed forecast value at the hub height into the power output forecast value (i.e., the power output forecast result of the wind power source). This model has a simple structure and strong physical interpretability, and is the basic method for wind power forecasting.

[0147] In other embodiments, wind field forecasts (wind speed, wind direction, air pressure, and temperature) output from mesoscale models such as WRF are used as input. Combined with the turbine layout and terrain elevation data of the wind farm, the data is downscaled using micro-meteorological models (such as WASP and WindSim) to obtain the hub wind speed of each turbine. The wind speed values ​​are then converted into predicted power output values. This process is suitable for large-scale wind farms with complex terrain.

[0148] The photovoltaic / wind power output models exemplified above are all mainstream prediction models, and will not be described in detail in this embodiment.

[0149] Example 3: In Example 1, the process of predicting the load at each bus node under each level depends on the following parameter: attenuation coefficient. fusion weight The corresponding adjustment coefficient λ for each level. These parameters are independent of the power grid structure, the specific operating state of the power grid, and the demand-side response.

[0150] In this embodiment, the above parameters are learned online.

[0151] Before proceeding with online learning, it should be noted that the initial values ​​of the parameters mentioned above in Example 1 were obtained through optimization. This example further provides the specific optimization method: The above embodiment is described based on the original load sequence and original power output sequence of any local power grid over the previous four years.

[0152] In this embodiment, an original load sequence and original output sequence with a duration of 2 years are randomly selected from the 4 years as an optimized sample, and the original load sequence and original output sequence after the optimized sample are used as labels.

[0153] For several local power grids with high penetration rates of new energy sources, multiple optimized samples and labels are obtained. These optimized samples and labels form a dataset, and the aforementioned parameters are used as parameters to be optimized. It should be noted that since the above parameters are independent of the power grid structure, the specific operating state of the power grid, and demand-side response, samples from different local power grids can be placed in the same dataset.

[0154] Optimize (or fit) the parameters to be optimized using the above dataset; specifically: Each optimized sample is input into Example 1. Example 1 outputs the final predicted load for each bus node at each level, which represents the predicted average user electricity demand for each bus node at each level (e.g., within one year after the optimized sample / in the first quarter after the optimized sample / in the first month after the optimized sample). This example uses the original load sequence in the tags to statistically calculate the actual load of each bus node at each level (e.g., within the first year of the tag / in the first quarter of the tag / in the first month of the tag), and uses this as the true value of the final predicted load. The average of the squared differences between the actual load and the final predicted load for all bus nodes at all levels is used as the cost function for the parameters to be optimized.

[0155] Specifically, if the time period corresponding to a sample or label does not span a full year, quarter, or month, then Example 1 does not need to output the predicted load at the corresponding scale. When obtaining the cost function, the predicted load at these scales does not need to be squared with the true value, nor does it need to be included in the calculation of the mean. The purpose is to ensure the perceptual integrity of the time scale at each level.

[0156] Given the parameters to be optimized and their cost function, an optimization algorithm is used to iteratively update the parameters to be optimized, thereby minimizing the cost function.

[0157] As an example, the differential evolution algorithm is used to iteratively update the parameters to be optimized. The population size is set to 50, the mutation factor to 0.5, the crossover probability to 0.9, the maximum number of iterations to 200, and the population initialization method is random initialization. Other embodiments may use different values ​​for these parameters. Since the differential evolution algorithm is a well-known and conventional technique, this embodiment does not specifically limit its parameter settings (these parameter settings do not affect the further technical effects of this embodiment), nor will it describe the specific implementation process of the algorithm. Simulated annealing may also be used in some embodiments; this embodiment does not specifically limit the optimization algorithm used.

[0158] The method for learning the above parameters online is as follows: Every year, for the raw load and output sequences collected in the most recent year, and the raw load and output sequences collected in the previous two years, a dataset is constructed according to the above method. The parameters are then optimized again using the current values ​​of the above parameters (i.e. the results of the last optimization) as the initial values, so as to realize the online updating of the parameters.

[0159] In other embodiments, online updates may be performed monthly or quarterly, which will not be described in this embodiment. The aforementioned online updates allow the forecasting process to adapt to dynamic changes on both the grid and demand sides.

[0160] In summary, considering all embodiments of the present invention, since power grid load is affected by factors such as regional economic development, industrial restructuring, and user behavior, the present invention and existing prediction schemes (such as prediction schemes based on neural network models or machine learning models) cannot accurately predict bus load. However, compared with the prior art, the present embodiment has at least the following further technical effects, making the prediction results of the present embodiment more reliable than those of the prior art: By dividing the data into multiple scales and performing layer-by-layer statistics, and by decoupling the data layer by layer from coarse time scale to fine time scale, the source-load coupling relationship is differentiated at different time scales. This overcomes the technical shortcomings of existing methods that mix data of different scales and do not learn the coupling relationship sufficiently, and improves the accuracy of bus load prediction to a certain extent. By prioritizing the stripping of the power drive component on the coarse time scale with the highest signal-to-noise ratio and then progressively pushing it downwards, the technical problem of interference components such as high-frequency fluctuations of new energy sources, random load fluctuations, and measurement noise on the fine time scale accumulating upwards and contaminating the intrinsic load components on the coarse time scale is avoided. This effectively reduces the stripping error and ensures the accuracy of the load components at each level. By using the mutual information between the load residual sequence and the power output sequence to progressively correct the coupling degree, the coupling degree can fully reflect the coupling relationship between the source and load from energy balance to fluctuation transmission. This enables dynamic and reliable quantification of the contribution of power output to the bus load, providing a reliable basis for load decoupling and load forecasting. By decoupling the actual load at each level into load component sequences that mainly reflect the electricity demand patterns of users, high-quality intrinsic load components that are not excessively affected by fluctuations in renewable energy output are obtained, providing a high-quality data foundation for the prediction model. By utilizing the load component sequences of each bus node at all levels for load prediction, the prediction model can identify and utilize the changing patterns of bus load at different time scales. It captures energy balance trends at a coarse scale and captures the details of fluctuation transmission at a fine scale, thereby effectively improving the prediction accuracy of bus load.

[0161] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-scale bus load refinement prediction method based on source-load coordination, characterized in that, The method includes the following steps: The time scale is divided into several levels, including annual, quarterly, and monthly scales. The original load sequence collected at each bus node and the original output sequence collected at each power source output terminal are statistically analyzed according to each level to obtain the actual load sequence of each bus node and the actual output sequence of each power source under each level. Any one of these levels is denoted as the target level. The load component sequence of each bus node in the level above the target level is denoted as L0, and the coupling degree between each bus node and each power source in the level above is denoted as C0. The method for obtaining the load component sequence and coupling degree of each bus node at the target level is as follows: The load baseline sequence of the target level is determined based on the load component sequence L0. The difference between the original load sequence and the load baseline sequence is denoted as the load residual sequence. The difference between the original output sequence and the actual output sequence of each power source under the target level is denoted as the output residual sequence. The mutual information and coupling degree C0 between the load residual sequence and the output residual sequence are superimposed to obtain the coupling degree between each bus node and each power source under the target level. Based on the coupling degree of the target level, the actual load sequence of each bus node under the target level is decoupled to obtain the load component sequence of each bus node under the target level. The final predicted load for each bus node is obtained by using the load component sequence of each bus node at all levels.

2. The multi-scale bus load refinement prediction method based on source-load coordination according to claim 1, characterized in that, The specific steps for determining the target level load baseline sequence based on the load component sequence L0 are as follows: The sampling period for each annual scale is one year, the sampling period for each quarterly scale is one quarter, and the sampling period for each monthly scale is one month; The level above the target level is designated as the reference level; Each reference level sampling period contains several target level sampling periods; Each element in the load component sequence at any level corresponds to a target level sampling period; Traverse each element in the load component sequence L0, and denote the traversed element as P. The sampling period of the reference level where the element is located is denoteed as the reference period. The sampling periods of several target levels contained in the reference period are denoteed as the target periods. The baseline value x(i) of the i-th target period is P×f(i), where f(i) represents the proportion of the average electricity demand of the i-th target period to the average electricity demand of the reference period. [x(1), x(2), ..., x(N)] are the baseline sequence segments decomposed from element P, and N represents the number of target periods contained in the reference period. For all elements before element P, the baseline sequence segments of all elements are spliced ​​together end to end to form the load baseline sequence of the target level.

3. The multi-scale bus load refinement prediction method based on source-load coordination according to claim 2, characterized in that, The specific method for determining the load residual sequence is as follows: For any element in the load baseline sequence of each bus node under the target level, the target level sampling period corresponding to the original load sequence of each bus node, and the load amount collected within the target level sampling period, is subtracted from that element. For all elements in the load baseline sequence, the load residual sequence is obtained by subtracting them from the original load sequence.

4. The multi-scale bus load refinement prediction method based on source-load coordination according to claim 2, characterized in that, The specific steps for obtaining the load component sequence of each bus node at the target level are as follows: For the actual load sequence of each bus node under the target level, traverse each element in the actual load sequence and record the traversed element as Q. For the target level sampling period where Q is located, the average output of each power source within the target level sampling period is recorded as Q0. Q0 multiplied by C yields the single-source contribution of each bus node under each power source, where C represents the coupling degree between each bus node and each power source at the target level. The sum of the single-source contributions of each bus node under all power sources yields the power drive component value of each bus node at the target level. The difference between Q and the power drive component value is recorded as the load component value of each bus node at the target level. For all elements preceding element Q, the load component values ​​of all elements are concatenated together to form the load component sequence of each bus node under the target level.

5. The multi-scale bus load refinement prediction method based on source-load coordination according to claim 1, characterized in that, The specific steps for obtaining the final predicted load for each bus node by using the load component sequences of each bus node at all levels are as follows: Load forecasting is performed using the load component sequence of each bus node at each level, resulting in the load forecasting results for each bus node at each level. The output prediction result of each power source at each level is predicted using meteorological data. The product of the coupling degree between each bus node and each power source at each level and the output prediction result is used as the power drive prediction result of each bus node at each level. The sum of the power drive prediction result and the load prediction result is used as the final predicted load of each bus node at each level.

6. The multi-scale bus load refinement prediction method based on source-load coordination according to claim 4, characterized in that, The specific steps to obtain f(i) are as follows: For the reference period in which element P is located, among all sampling periods under the reference level, the sampling periods before the reference period are recorded as historical reference periods; Any historical reference period containing several target level sampling periods is denoted as the historical target period; The average value of the elements corresponding to all historical reference periods within the load component sequence L0 is denoted as the first average value; the average value of the load component values ​​obtained within the i-th historical target period included in all historical reference periods is denoted as the second average value, and the ratio of the second average value to the first average value is denoted as f(i).

7. The multi-scale bus load refinement prediction method based on source-load coordination according to claim 6, characterized in that, When no load component value is obtained within the i-th historical target period, the element corresponding to the i-th historical target period is taken as the load component value in the actual load sequence of each bus node under the target level.

8. The multi-scale bus load refinement prediction method based on source-load coordination according to claim 1, characterized in that, The specific steps for obtaining the coupling degree between each bus node and each power source at the annual scale are as follows: The mutual information between the original load sequence of each bus node and the original output sequence of each power source is denoted as the initial mutual information. On an annual scale, the coupling degree between each bus node and each power source is positively correlated with the initial mutual information and negatively correlated with the topological electrical distance between each bus node and each power source. The topological electrical distance is determined by the admittance matrix of the power grid topology.

9. The multi-scale bus load refinement prediction method based on source-load coordination according to claim 8, characterized in that, The specific steps for obtaining the load component sequence of each bus node at the annual scale are as follows: For the actual output sequence of each power source at the annual scale, each element of the actual output sequence is multiplied by... This yields the first drive sequence for each bus node under each power supply. This indicates the coupling degree between each bus node and each power source on an annual scale. The summation of the first drive sequence of each bus node under all power sources yields the power drive component sequence of each bus node on an annual scale. The difference between the actual load sequence of each bus node and the power drive component sequence of each bus node at the annual scale is denoted as the load component sequence of each bus node at the annual scale.

10. The multi-scale bus load refinement prediction method based on source-load coordination according to claim 1, characterized in that, The specific steps for obtaining the actual load sequence of each bus node at each level are as follows: The original load sequence represents the sequence of active power at each bus node. In the original load sequence, the average active power for each year is calculated, and the sequence of the average active power for all years is recorded as the actual load sequence of each bus node at the annual scale. The average active power for each quarter is calculated, and the sequence of the average active power for all quarters is recorded as the actual load sequence of each bus node at the quarterly scale. The average active power for each month is calculated, and the sequence of the average active power for all months is recorded as the actual load sequence of each bus node at the monthly scale.

Images