Power distribution network reverse power identification method and system based on photovoltaic-load time sequence coupling

Abstract

The invention provides a power distribution network reverse power identification method and system based on photovoltaic-load time sequence coupling, and the method comprises the steps: obtaining historical operation data and real-time operation data of a target power distribution network, carrying out the alignment processing of photovoltaic output and load power under a unified time grid, and obtaining photovoltaic output time sequence data and load time sequence data; a sliding time window is adopted to construct a time sequence sample, and coupling features representing the coupling relation of photovoltaic-load in the time dimension are extracted to form a coupling feature matrix; training the reverse power identification model by using the reverse power label of the historical sample to obtain a primary model, and inputting the real-time sample into the primary model to obtain a reverse power primary identification result; according to the inverse power initial identification result, optimizing the initial trained inverse power identification model to obtain a trained inverse power identification model; and finally, inputting a real-time sample into the trained model, outputting a reverse power identification result, and generating the occurrence moment, the duration time and the amplitude of a reverse power event.

Description

Technical Field

[0001] This invention relates to the field of distribution network operation monitoring and analysis technology, specifically to a method and system for identifying reverse power in distribution networks based on photovoltaic-load time-series coupling, and more specifically to a method for sensing the operation status and providing risk warnings for distribution networks with a high proportion of distributed photovoltaic access. Background Technology

[0002] With the large-scale integration of renewable energy sources such as distributed photovoltaic (PV) power generation into distribution networks, the distribution network has gradually evolved from a traditional unidirectional power supply mode to a bidirectional power flow operation mode, resulting in significant changes in its operating structure and power characteristics. In particular, distributed PV power generation exhibits significant intermittency, volatility, and randomness. During periods of good sunlight and low local load, PV output can easily exceed load demand, leading to reverse power phenomena in the distribution network. Frequent occurrences of reverse power can not only cause problems such as voltage rise at the grid connection point, protection malfunctions, and equipment overload, but also adversely affect the safe and stable operation of the distribution network. Therefore, accurate and real-time identification of reverse power conditions in the distribution network is of great significance for ensuring its safe and reliable operation.

[0003] During the operation of a distribution network, both photovoltaic (PV) power output and load demand exhibit distinct time-series characteristics. PV power generation is influenced by factors such as irradiance and meteorological conditions, showing strong intraday and seasonal variations. Meanwhile, the load side varies significantly in terms of electricity consumption behavior and temporal distribution due to differences in types such as residential, commercial, and industrial loads. Analyzing PV or load independently is insufficient to accurately reflect their temporal coupling relationship, thus limiting the accuracy and applicability of reverse power identification. Therefore, it is necessary to jointly model PV power output with time-series models of different load types. By comprehensively analyzing the time-series characteristics of both, accurate identification of the occurrence periods and persistence characteristics of reverse power in the distribution network can be achieved.

[0004] Existing reverse power identification methods mostly rely on single measurement information or simple threshold criteria, typically based on instantaneous power direction or electricity metering results. While these methods can reflect the occurrence of reverse power to some extent, they struggle to fully consider the dynamic evolution of photovoltaic output and load demand over time. This is particularly problematic in distribution network scenarios with diverse load types and significant differences in time-series characteristics, easily leading to identification lags or misjudgments. Furthermore, traditional methods often focus on post-hoc judgment, lacking a systematic characterization of the duration and evolution of reverse power, making it difficult to provide accurate basis for subsequent control strategies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for identifying reverse power in distribution networks based on photovoltaic-load time-series coupling.

[0006] A method for identifying reverse power in a distribution network based on photovoltaic-load time-series coupling, according to the present invention, includes: Step S1: Obtain historical and real-time operating data of the target distribution network, and perform time alignment between the historical and real-time operating data of the target distribution network to obtain photovoltaic output time-series data and load time-series data; Step S2: Construct time series samples based on photovoltaic output time series data and load time series data, extract coupling features that characterize the photovoltaic-load time series coupling relationship based on the time series samples, and form a coupling feature matrix; Step S3: Construct an inverse power recognition model, train the inverse power recognition model using the coupling feature matrix of historical samples to obtain the initially trained inverse power recognition model; input the coupling feature matrix of real-time samples into the initially trained inverse power recognition model to obtain the initial inverse power recognition result; Step S4: Calculate the loss function based on the initial inverse power identification result and the verification result, and optimize the initially trained inverse power identification model based on the loss function to obtain the trained inverse power identification model; Step S5: Input the coupling feature matrix of the real-time sample into the trained inverse power recognition model to output the inverse power recognition result, and generate the occurrence time, duration and amplitude of the inverse power event.

[0007] Preferably, step S1 includes: Step S1.1: Obtain historical and real-time operating data from the target identification points of the target distribution network; the target identification points include any one or more of the following: grid connection point, feeder head, transformer substation meter, or low-voltage side of the transformer; the historical and real-time operating data include: raw measured values ​​of active power on the photovoltaic side and / or raw measured values ​​of active power on the load side; Step S1.2: Set a uniform sampling period and construct a uniform time grid so that power data from different sources correspond one-to-one on the same time scale. The time grid formula is:

[0008] in, For the first A unified sampling time; To align the start times of the time period; To standardize the sampling period; The time index is a non-negative integer; Step S1.3: Map the photovoltaic power data and load power data onto the unified time grid using mapping methods including interval average resampling or interpolation alignment; when there are clock asynchronization issues between different devices or fixed communication link delays, further time offset correction is performed to obtain photovoltaic power output time-series data on the unified time grid. With load time series data .

[0009] Preferably, step S2 includes: Step S2.1: Use a sliding time window to process the photovoltaic power output time series data. With load time series data The sample is segmented to form multiple time windows; each sample contains continuous time windows. There are sampling points, and the sliding step size is . This ensures that each sample can cover both the dynamic changes of photovoltaics and loads, and also meet the real-time requirements of online identification; Step S2.2: Construct a net injected power sequence within the sample window to characterize the surplus of photovoltaic output relative to the load;

[0010] in, To unify the time The net injection power is below; For a moment The active power output of photovoltaic power; For a moment The active power of the load; when This indicates that there is a tendency or risk of reverse power transmission to the upper-level power grid at the target identification point; Step S2.3: For each time series sample, extract a set of coupling features that includes at least synchronization / misalignment coupling, coupling strength, inverse power margin, and persistence, and construct a coupling feature matrix based on the set of coupling features; The synchronous / misaligned coupling and coupling strength are characterized by hysteresis correlation features. The correlation coefficient is calculated within a given maximum hysteresis range, and the maximum correlation is taken as the coupling strength index for that sample. The formula is as follows:

[0011] in, For the first The maximum hysteresis correlation coefficient of each sample is used to characterize the coupling strength between photovoltaics and load; The lag is the time shift of the load sequence relative to the photovoltaic sequence; This represents the maximum lag search range; Operators are used to calculate the correlation coefficient; For the sequence by hysteresis Operators for time translation; For the first A sequence vector of photovoltaic power output within a sample window; For the first Load sequence vector within each sample window; Simultaneously, inverse power margin and persistence features are extracted from the net injected power sequence, and these features are concatenated into a sample feature vector in a predetermined order. Stack all sample feature vectors row-wise to form a coupled feature matrix. Each row corresponds to a sample, and each column corresponds to a coupling feature.

[0012] Preferably, in step S3, constructing an inverse power identification model and training the inverse power identification model using the coupling feature matrix of historical samples to obtain an initially trained inverse power identification model includes: Using the coupled feature vectors of historical samples As input to the inverse power identification model, with inverse power labels As a supervisory signal, an inverse power identification model is constructed. The probability of the output sample having inverse power The inverse power recognition model is initially trained by using binary cross-entropy as the loss function and iteratively optimizing it to ensure that the output probability of the model matches the true label. The loss function is:

[0013] in, For training the loss function; These are the parameters for the inverse power identification model; This represents the number of historical samples. For sample index; For the first Inverse power label of each sample, This indicates that reverse power has occurred. This indicates that no reverse power occurred; For the model to the first The predicted probability of each sample having inverse power; It is the natural logarithm function; After training, the inverse power recognition model is obtained. ,in, These are the parameters of the inverse power recognition model when the initial training is completed.

[0014] Preferably, step S3, which involves inputting the coupling feature matrix of the real-time sample into the initially trained inverse power recognition model to obtain the initial inverse power recognition result, includes: Obtain real-time sample coupled feature vectors Couple real-time samples with feature vectors Input the inverse power recognition model trained for the first time Obtain the real-time probability of inverse power occurrence And according to the preset threshold The initial identification result of the output inverse power is determined by the following formula:

[0015] in, This is the initial identification result of inverse power; This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. The inverse power probability output by the initially trained inverse power recognition model; For real-time sample coupled feature vectors; These are the parameters for the initial training of the model; This is the inverse power discrimination threshold.

[0016] Preferably, step S4 includes: The predicted probability output of the initially trained inverse power recognition model for the samples is compared with the preset verification labels to construct a loss function, and the parameters of the initially trained inverse power recognition model are updated using gradient optimization; the parameter update formula is as follows:

[0017] in, The parameters of the inverse power recognition model to be optimized during initial training; The learning rate; For the loss function with respect to the parameters The gradient; The loss function is constructed based on the verification labels; Optimization stops when the loss function decreases and stabilizes with each training epoch, or when the change in loss between two adjacent epochs is less than a preset threshold, thus obtaining a well-trained inverse power recognition model. ,in, The parameters for the optimized inverse power identification model.

[0018] Preferably, the preset verification label includes: The verification results are derived from at least one or more of the following: verification criteria based on the active power direction at the metering point, power direction verification criteria obtained based on power flow calculation or state estimation, and reverse power events recorded by protection devices or reverse power relays; the verification results are converted into verification tags. ,in, This indicates that the verification process has resulted in reverse power. This indicates that no reverse power has occurred during the verification process; and the verification tag is matched with the initial identification output at the corresponding time / time window.

[0019] According to the present invention, a distribution network reverse power identification system based on photovoltaic-load time-series coupling includes: Module M1: Acquires historical and real-time operating data of the target distribution network, and performs time alignment between the historical and real-time operating data of the target distribution network to obtain photovoltaic output time-series data and load time-series data; Module M2: Constructs time-series samples based on photovoltaic output time-series data and load time-series data, extracts coupling features that characterize the photovoltaic-load time-series coupling relationship based on the time-series samples, and forms a coupling feature matrix; Module M3: Constructs an inverse power recognition model, trains the inverse power recognition model using the coupling feature matrix of historical samples to obtain the initially trained inverse power recognition model; inputs the coupling feature matrix of real-time samples into the initially trained inverse power recognition model to obtain the initial inverse power recognition result; Module M4: Calculates the loss function based on the initial inverse power identification result and the verification result, and optimizes the initially trained inverse power identification model based on the loss function to obtain the trained inverse power identification model; Module M5: Inputs the coupling feature matrix of real-time samples into the trained inverse power recognition model, outputs the inverse power recognition result, and generates the occurrence time, duration and amplitude of the inverse power event.

[0020] Preferably, the module M1 includes: Module M1.1: Acquires historical and real-time operating data from the target identification points of the target distribution network; the target identification points include any one or more of the following: grid connection point, feeder head, transformer substation meter, or low-voltage side of the transformer; the historical and real-time operating data include: raw measured values ​​of active power on the photovoltaic side and / or raw measured values ​​of active power on the load side; Module M1.2: Sets a uniform sampling period and constructs a uniform time grid, so that power data from different sources correspond one-to-one on the same time scale. The time grid formula is:

[0021] in, For the first A unified sampling time; To align the start times of the time period; To standardize the sampling period; The time index is a non-negative integer; Module M1.3: Maps photovoltaic power data and load power data onto the unified time grid using mapping methods including interval average resampling or interpolation alignment; when there are clock asynchronization issues between different devices or fixed communication link delays, further time offset correction is performed to obtain photovoltaic power output time-series data on the unified time grid. With load time series data ; The module M2 includes: Module M2.1: Employs a sliding time window for photovoltaic power output timing data. With load time series data The sample is segmented to form multiple time windows; each sample contains continuous time windows. There are sampling points, and the sliding step size is . This ensures that each sample can cover both the dynamic changes of photovoltaics and loads, and also meet the real-time requirements of online identification; Module M2.2: Constructs a net injected power sequence within the sample window to characterize the surplus of photovoltaic output relative to the load;

[0022] in, To unify the time The net injection power is below; For a moment The active power output of photovoltaic power; For a moment The active power of the load; when This indicates that there is a tendency or risk of reverse power transmission to the upper-level power grid at the target identification point; Module M2.3: For each time series sample, extract a set of coupling features that includes at least synchronization / misalignment coupling, coupling strength, inverse power margin, and persistence, and construct a coupling feature matrix based on the set of coupling features; The synchronous / misaligned coupling and coupling strength are characterized by hysteresis correlation features. The correlation coefficient is calculated within a given maximum hysteresis range, and the maximum correlation is taken as the coupling strength index for that sample. The formula is as follows:

[0023] in, For the first The maximum hysteresis correlation coefficient of each sample is used to characterize the coupling strength between photovoltaics and load; The lag is the time shift of the load sequence relative to the photovoltaic sequence; This represents the maximum lag search range; Operators are used to calculate the correlation coefficient; For the sequence by hysteresis Operators for time translation; For the first A sequence vector of photovoltaic power output within a sample window; For the first Load sequence vector within each sample window; Simultaneously, inverse power margin and persistence features are extracted from the net injected power sequence, and these features are concatenated into a sample feature vector in a predetermined order. Stack all sample feature vectors row-wise to form a coupled feature matrix. Each row corresponds to a sample, and each column corresponds to a coupling feature.

[0024] Preferably, the inverse power recognition model is constructed in module M3, and the inverse power recognition model is trained using the coupling feature matrix of historical samples to obtain the initially trained inverse power recognition model, including: Using the coupled feature vectors of historical samples As input to the inverse power identification model, with inverse power labels As a supervisory signal, an inverse power identification model is constructed. The probability of the output sample having inverse power The inverse power recognition model is initially trained by using binary cross-entropy as the loss function and iteratively optimizing it to ensure that the output probability of the model matches the true label. The loss function is:

[0025] in, For training the loss function; These are the parameters for the inverse power identification model; This represents the number of historical samples. For sample index; For the first Inverse power label of each sample, This indicates that reverse power has occurred. This indicates that no reverse power occurred; For the model to the first The predicted probability of each sample having inverse power; It is the natural logarithm function; After training, the inverse power recognition model is obtained. ,in, These are the parameters of the inverse power recognition model when the initial training is completed; In module M3, the coupling feature matrix of the real-time samples is input into the initially trained inverse power recognition model to obtain the initial inverse power recognition result, including: Obtain real-time sample coupled feature vectors Couple real-time samples with feature vectors Input the inverse power recognition model trained for the first time Obtain the real-time probability of inverse power occurrence And according to the preset threshold The initial identification result of the output inverse power is determined by the following formula:

[0026] in, This is the initial identification result of inverse power; This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. The inverse power probability output by the initially trained inverse power recognition model; For real-time sample coupled feature vectors; These are the parameters for the initial training of the model; The inverse power discrimination threshold; The module M4 includes: The predicted probability output of the initially trained inverse power recognition model for the samples is compared with the preset verification labels to construct a loss function, and the parameters of the initially trained inverse power recognition model are updated using gradient optimization; the parameter update formula is as follows:

[0027] in, The parameters of the inverse power recognition model to be optimized during initial training; The learning rate; For the loss function with respect to the parameters The gradient; The loss function is constructed based on the verification labels; Optimization stops when the loss function decreases and stabilizes with each training epoch, or when the change in loss between two adjacent epochs is less than a preset threshold, thus obtaining a well-trained inverse power recognition model. ,in, The parameters for the optimized inverse power identification model.

[0028] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention reduces timing misalignment caused by different sampling periods, clock offsets and communication delays by aligning multi-source data of photovoltaic output and load power under a unified time grid, making the photovoltaic-load relationship comparable and modelable on the same time axis; 2. This invention addresses the reverse power formation mechanism in distribution networks by constructing time-series samples using a sliding time window. It extracts coupling features from these samples that characterize the synchronization / misalignment, coupling strength, net injection margin, and persistence of photovoltaics and loads in the time dimension, forming a coupling feature matrix. This provides a physically meaningful and noise-robust input representation for reverse power identification. 3. This invention uses historical sample inverse power labels for supervised training to obtain the initial model, and introduces verification results based on metered power direction, power flow / state estimation or protection device records to constrain the model and optimize parameters, so that the model can maintain stable recognition performance under actual operating conditions. 4. This invention not only outputs the identification result of whether reverse power has occurred, but also generates event-level information, including the occurrence time, duration and amplitude of the reverse power event, providing direct basis for distribution network operation alarms, dispatching decisions, distributed photovoltaic absorption analysis and reverse power flow management, thereby improving the security, stability and observability of distribution networks with a high proportion of distributed photovoltaic access. 5. This invention can fully utilize the time-series coupling information between photovoltaics and loads, improve the accuracy and robustness of reverse power identification, realize online identification and quantitative characterization of reverse power events, and is suitable for distribution network operation monitoring and alarm scenarios with a high proportion of distributed photovoltaic access. 6. This invention can fully utilize the correlation and misalignment characteristics between photovoltaic power output and load demand in the time dimension to achieve accurate identification and event-level quantitative characterization of reverse power events in the distribution network, reduce the risk of false alarms and missed alarms caused by metering noise, communication delays, photovoltaic volatility and load uncertainty, thereby improving the safety and stability of distribution networks with a high proportion of distributed photovoltaic access. Attached Figure Description

[0029] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart of a distribution network reverse power identification method based on photovoltaic-load time-series coupling.

[0030] Figure 2 This is a structural diagram of the inverse power identification model.

[0031] Figure 3 This is a schematic diagram of a distribution network reverse power identification system based on photovoltaic-load time-series coupling. Detailed Implementation

[0032] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0033] Example 1 According to the present invention, a method for identifying reverse power in a distribution network based on photovoltaic-load time-series coupling is provided, such as... Figures 1-2 As shown, it includes the following steps: Step 1: Obtain historical and real-time operating data of the target distribution network and align them by time to obtain photovoltaic output time-series data and load time-series data; Step 2: Construct time-series samples based on photovoltaic output time-series data and load time-series data, and extract coupling features that characterize the photovoltaic-load time-series coupling relationship to form a coupling feature matrix; Step 3: Train the inverse power recognition model using the inverse power labels of historical samples to obtain the initially trained inverse power recognition model; input the real-time samples into the initially trained inverse power recognition model to obtain the initial inverse power recognition result. Step 4: Calculate the loss function based on the initial inverse power identification results and verification results (based on metered power direction, power flow / state estimation or protection device records), optimize the model parameters, and obtain the trained inverse power identification model; Step 5: Input the real-time samples into the trained inverse power recognition model, output the inverse power recognition results, and generate the occurrence time, duration, and amplitude of the inverse power event; The reverse power identification model is constructed based on the coupling feature matrix extracted from the time series samples built from photovoltaic output time series data and load time series data. It is used to identify and judge whether reverse power occurs in the target distribution network within the time period corresponding to the real-time sample, and output the occurrence time, duration and amplitude of the reverse power event.

[0034] Specifically, step 1 includes: First, historical and real-time operating data are obtained from the target identification points of the target distribution network. These target identification points can be locations such as grid connection points, feeder heads, transformer substation meters, or the low-voltage side of the transformer. The historical and real-time operating data at least include the raw measured values ​​of active power on the photovoltaic side. and the original measured value of active power on the load side ,in, The timestamp corresponding to the data point; The raw data from the photovoltaic side is represented as a set of discrete points with timestamps;

[0035] The raw load-side data is represented as follows:

[0036] in, This is the original data set from the photovoltaic side. This is the original data set from the load side; Let i be the timestamp of the i-th photovoltaic data point. The timestamp of the j-th load data point; and These represent the number of raw data points on the photovoltaic side and the load side, respectively; the subscript "raw" indicates unaligned and uncleaned raw measurements.

[0037] Set a uniform sampling period Δt and construct a uniform time grid, taking the aligned start and end times as follows: and Construct a standard sampling time sequence

[0038]

[0039] in, To standardize the k-th sampling time on the time grid, To standardize the sampling period, k is the time index, and K is the last index of the time grid. Indicates floor operation; unified time grid This is used to map power sequences from different data sources to the same time axis, thereby achieving time alignment.

[0040] To eliminate potential fixed clock skew or communication delays from different data sources, a sequence for skew estimation is constructed on a unified time grid, and the cross-correlation coefficient is calculated, defined on the lag. The correlation coefficients are as follows:

[0041] And in Within the search range, select the offset that maximizes the absolute value of the correlation coefficient.

[0042] Where ρ(δ) is the correlation coefficient between the photovoltaic sequence and the load sequence with a lag of δ, and δ is the candidate time offset. The maximum search offset range is given by δ*, where δ* is the estimated optimal offset. and These are the mean values ​​of the photovoltaic power series and the load power series used to calculate the correlation coefficient, respectively; argmax(·) represents the independent variable that maximizes the objective function; subsequently, the timestamp of one of the data sources is corrected, for example, the timestamp of the photovoltaic side is corrected to...

[0043] in, This is the corrected timestamp of the i-th photovoltaic data point. Next, the power data from both the photovoltaic and load sides are uniformly mapped to the constructed unified time grid. The above method achieves synchronous alignment of two types of data under the same sampling period and timestamp. During the alignment process, the interval averaging resampling or interpolation alignment method can be selected according to the data sampling characteristics: when the original data sampling intervals are unequal or there is jitter, the interval averaging method is preferred to calculate the equivalent power in each sampling interval; when the original data sampling is relatively stable and the timestamp is close to the grid point, the interpolation method can be used to obtain the power value at the grid time, thus obtaining the aligned photovoltaic power sequence and load power sequence. For ease of representation, the aligned discrete sequence is uniformly denoted as:

[0044] in, For the first time grid on a unified time grid A standard sampling time, and At time respectively The photovoltaic aligned power value and the load aligned power value, This represents a discrete-time index.

[0045] To ensure the continuity and reliability of the aligned time-series data, data quality processing is performed on the aligned sequence: when a small number of missing points occur, interpolation or forward padding is used to fill them in; when long-term missing points occur, the historical average of the same hour, the daily curve of the same type, or samples from the corresponding time period are directly removed to avoid misleading subsequent coupling feature extraction. For outliers caused by communication jitter or metering spikes, robust statistics or threshold rules are used for identification, and the outliers are replaced with neighborhood statistics or subjected to amplitude limiting; if necessary, the sequence is lightly smoothed and filtered to suppress high-frequency noise while preserving the rapid fluctuation characteristics of photovoltaic data as much as possible.

[0046] Output on a unified time grid The synchronized photovoltaic output time-series data and load time-series data are used as input data for subsequent construction of photovoltaic-load time-series coupling characteristics and training / online identification of inverse power identification models.

[0047] Specifically, step 2 includes: A unified time grid is obtained after time alignment is completed. Photovoltaic power output time series data With load time series data Time series samples are constructed using a sliding time window method; let the length of the sliding time window be... The sliding step size is Then the first The starting index corresponding to each sample is The photovoltaic and load sequences within this sample are represented as follows:

[0048] in, For the first A photovoltaic power sequence vector of samples, For the first A load power sequence vector of each sample, For the sample window length, The sliding step size of the sample window. For the first The start time index of each sample.

[0049] To characterize the coupling relationship between photovoltaic (PV) and load over time, coupling features reflecting synchronization / dislocation, strong / weak correlation, and backfeed margin are extracted for each sample. Preferably, these features include hysteresis correlation features and net injection margin features. The hysteresis correlation feature reflects the degree of correlation between PV changes and load changes under different time misalignments, and is taken within a given maximum hysteresis range. The maximum correlation coefficient and its corresponding lag within the range are used as characters for coupling strength and coupling phase, and are defined as follows:

[0050] in, For the first Each sample is in lag. The correlation coefficient below, Operators for calculating correlation coefficients, For sequence by Operators for time shifting, To maximize the search range with lag, The maximum correlation coefficient is used to measure the coupling strength between photovoltaics and load within the sample. The corresponding optimal lag is used to characterize the temporal misalignment relationship between the two.

[0051] To reflect the direct driving relationship of reverse power formation, a net injection margin feature is introduced to characterize the degree to which photovoltaic output exceeds the load. The net injection sequence within the sample is defined as follows: The maximum net injection value within the sample window and the percentage of positive values ​​are used as coupling margin features, defined as follows:

[0052] in, For the first The maximum net injection value within each sample window is used to characterize the peak margin of the potential inverse power. The percentage of time during which net injection is positive is used to characterize the persistence of reverse power risk over time. An indicator function that returns 1 if the condition within the parentheses is true, and 0 otherwise.

[0053] By concatenating the aforementioned coupling features with other optional statistical features, the first... Coupled feature vectors of each sample And stack all sample features row by row to form a coupled feature matrix. Each row corresponds to a time series sample, and each column corresponds to a coupling feature, thus providing input for the training and online recognition of the subsequent inverse power recognition model.

[0054] Specifically, step 3 includes: In this invention, the historical sample coupling feature vector is obtained based on coupling feature extraction. and its corresponding inverse power tag Construct a training dataset, in which , Given the number of historical samples; establish an inverse power identification model. Its input is the coupled feature vector, and its output is the predicted probability of inverse power. The model parameters are optimized iteratively using binary cross-entropy as the training objective function. To minimize the loss function, we obtain the inverse power recognition model trained initially. Its loss function is defined as

[0055] in, For training the loss function; These are the parameters for the inverse power identification model; This represents the number of historical samples. For sample index; For the first Inverse power label of each sample, This indicates that reverse power has occurred. This indicates that no reverse power occurred; For the model to the first The predicted probability of each sample having inverse power; It is the natural logarithm function.

[0056] Couple the feature vectors of real-time samples Input the inverse power recognition model trained for the first time The real-time probability of inverse power occurrence is obtained, and a preset discrimination threshold is used. Initial identification result of output inverse power Defined as

[0057] in, For real-time sample coupled feature vectors, The inverse power discrimination threshold, This indicates that reverse power has occurred. This indicates that no reverse power has occurred.

[0058] Specifically, step 4 includes: First, the initial inverse power identification results obtained from the initial training of the inverse power identification model are correlated and aligned with the verification results. The verification results can originate from the active power direction criterion at the metering point, the power direction calculated from power flow / state estimation, or the reverse power alarm recorded by the protection device, etc. Taking a sample as an example, let the probability of the inverse power output by the inverse power identification model for that sample be . The sample labels given in the verification results are ,in This indicates that the verification determines that the sample has inverse power. This indicates that the verification determines that the sample did not experience inverse power; based on the above two factors, the model loss function is constructed, preferably using the binary classification cross-entropy form:

[0059] in, The number of samples participating in the validation and optimization. For inverse power identification model parameters, For the model to the first The predicted probability of each sample having inverse power. The label corresponding to the verification result.

[0060] Secondly, a gradient-based optimization method is used to update the model parameters, thereby improving the loss function. Gradually decrease the learning rate to improve the consistency between the model output and the validation results; for example, use a learning rate of... The parameter update strategy iteratively optimizes the parameters:

[0061] in, For learning rate, The gradient of the loss function with respect to the model parameters is used. Optimization stops when the change in the loss function between adjacent iterations is less than a preset threshold or when a preset number of training iterations is reached, resulting in a well-trained inverse power identification model. This trained model can output more accurate inverse power identification results under new real-time sample inputs, thereby improving the reliability and robustness of identifying inverse power events in the distribution network.

[0062] Specifically, step 5 includes: In this invention, a preset sampling period is used. Continuously acquire and align real-time photovoltaic power output data with real-time load data, construct real-time samples using a sliding time window, and extract the corresponding coupled feature vectors. ,in The index of the real-time sample on a unified time grid is used; the coupled feature vector is input into the trained inverse power recognition model to obtain the inverse power discriminant output (probability or discriminant value) corresponding to the current time window, and the inverse power recognition result of the current window is output according to the preset threshold. ,in This indicates that reverse power has occurred. This indicates that no inverse power has occurred; the identification results of each time window are stored in chronological order to form a continuous inverse power determination sequence.

[0063] To suppress false alarms and jitter caused by short-term fluctuations, a continuous criterion is used to aggregate inverse power events: when continuous A time window satisfies The inverse power event is determined at a specific time, and the time of the event occurrence is recorded as the corresponding start time. When subsequent consecutive A time window satisfies The inverse power event is determined to have ended, and the time of event termination is recorded as the corresponding end time. ;in, and These are preset continuous judgment parameters used to balance detection sensitivity and stability.

[0064] Secondly, during the event duration, the reverse power amplitude is calculated based on real-time power data. The reverse power amplitude can be the absolute value of the reverse active power at the grid connection point / feeder head end as the event amplitude, or the positive value of the net injected power can be used as the equivalent reverse power amplitude. In one implementation, the net injected power is used as the event amplitude. Based on this, the maximum reverse power within the event interval is used as the event amplitude. Simultaneously, the average reverse power or integral energy can be statistically analyzed as an extended indicator; among which, the event amplitude... Used to characterize the intensity of inverse power events.

[0065] Finally, based on the start time of the event and the end time Calculate event duration It outputs inverse power event information, which includes at least the event occurrence time. Event End Time Duration of the event and event amplitude For multiple events that may occur on the same feeder or in the same distribution area, an event list is generated in chronological order to enable online identification, recording, and alarm output of reverse power events.

[0066] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0067] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for identifying reverse power in a distribution network based on photovoltaic-load time-series coupling, characterized in that, include: Step S1: Obtain historical and real-time operating data of the target distribution network, and perform time alignment between the historical and real-time operating data of the target distribution network to obtain photovoltaic output time-series data and load time-series data; Step S2: Construct time series samples based on photovoltaic output time series data and load time series data, extract coupling features that characterize the photovoltaic-load time series coupling relationship based on the time series samples, and form a coupling feature matrix; Step S3: Construct an inverse power recognition model, train the inverse power recognition model using the coupling feature matrix of historical samples to obtain the initially trained inverse power recognition model; input the coupling feature matrix of real-time samples into the initially trained inverse power recognition model to obtain the initial inverse power recognition result; Step S4: Calculate the loss function based on the initial inverse power identification result and the verification result, and optimize the initially trained inverse power identification model based on the loss function to obtain the trained inverse power identification model; Step S5: Input the coupling feature matrix of the real-time sample into the trained inverse power recognition model to output the inverse power recognition result, and generate the occurrence time, duration and amplitude of the inverse power event.

2. The method for identifying reverse power in a distribution network based on photovoltaic-load time-series coupling according to claim 1, characterized in that, Step S1 includes: Step S1.1: Obtain historical and real-time operating data from the target identification points of the target distribution network; the target identification points include any one or more of the following: grid connection point, feeder head, transformer substation meter, or low-voltage side of the transformer; the historical and real-time operating data include: raw measured values ​​of active power on the photovoltaic side and / or raw measured values ​​of active power on the load side; Step S1.2: Set a uniform sampling period and construct a uniform time grid so that power data from different sources correspond one-to-one on the same time scale. The time grid formula is: in, For the first A unified sampling time; To align the start times of the time period; To standardize the sampling period; The time index is a non-negative integer; Step S1.3: Map the photovoltaic power data and load power data onto the unified time grid using mapping methods including interval average resampling or interpolation alignment; when there are clock asynchronization issues between different devices or fixed communication link delays, further time offset correction is performed to obtain photovoltaic power output time-series data on the unified time grid. With load time series data .

3. The method for identifying reverse power in a distribution network based on photovoltaic-load time-series coupling according to claim 1, characterized in that, Step S2 includes: Step S2.1: Use a sliding time window to process the photovoltaic power output time series data. With load time series data The sample is segmented to form multiple time windows; each sample contains continuous time windows. There are sampling points, and the sliding step size is . This ensures that each sample can cover both the dynamic changes of photovoltaics and loads, and also meet the real-time requirements of online identification; Step S2.2: Construct a net injected power sequence within the sample window to characterize the surplus of photovoltaic output relative to the load; in, To unify the time The net injection power is below; For a moment The active power output of photovoltaic power; For a moment The active power of the load; when This indicates that there is a tendency or risk of reverse power transmission to the upper-level power grid at the target identification point; Step S2.3: For each time series sample, extract a set of coupling features that includes at least synchronization / misalignment coupling, coupling strength, inverse power margin, and persistence, and construct a coupling feature matrix based on the set of coupling features; The synchronous / misaligned coupling and coupling strength are characterized by hysteresis correlation features. The correlation coefficient is calculated within a given maximum hysteresis range, and the maximum correlation is taken as the coupling strength index for that sample. The formula is as follows: in, For the first The maximum hysteresis correlation coefficient of each sample is used to characterize the coupling strength between photovoltaics and load; The lag is the time shift of the load sequence relative to the photovoltaic sequence; This represents the maximum lag search range; Operators are used to calculate the correlation coefficient; For the sequence by hysteresis Operators for time translation; For the first A sequence vector of photovoltaic power output within a sample window; For the first Load sequence vector within each sample window; Simultaneously, inverse power margin and persistence features are extracted from the net injected power sequence, and these features are concatenated into a sample feature vector in a predetermined order. Stack all sample feature vectors row-wise to form a coupled feature matrix. Each row corresponds to a sample, and each column corresponds to a coupling feature.

4. The method for identifying reverse power in a distribution network based on photovoltaic-load time-series coupling according to claim 1, characterized in that, In step S3, an inverse power identification model is constructed. This model is trained using the coupling feature matrix of historical samples to obtain the initially trained inverse power identification model, including: Using the coupled feature vectors of historical samples As input to the inverse power identification model, with inverse power labels As a supervisory signal, an inverse power identification model is constructed. The probability of the output sample having inverse power The inverse power recognition model is initially trained by using binary cross-entropy as the loss function and iteratively optimizing it to ensure that the output probability of the model matches the true label. The loss function is: in, For training the loss function; These are the parameters for the inverse power identification model; This represents the number of historical samples. For sample index; For the first Inverse power label of each sample, This indicates that reverse power has occurred. This indicates that no reverse power occurred; For the model to the first The predicted probability of each sample having inverse power; It is the natural logarithm function; After training, the inverse power recognition model is obtained. ,in, These are the parameters of the inverse power recognition model when the initial training is completed.

5. The method for identifying reverse power in a distribution network based on photovoltaic-load time-series coupling according to claim 1, characterized in that, In step S3, the coupling feature matrix of the real-time sample is input into the initially trained inverse power recognition model to obtain the initial inverse power recognition result, including: Obtain real-time sample coupled feature vectors Couple real-time samples with feature vectors Input the inverse power recognition model trained for the first time Obtain the real-time probability of inverse power occurrence And according to the preset threshold The initial identification result of the output inverse power is determined by the following formula: in, This is the initial identification result of inverse power; This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. The inverse power probability output by the initially trained inverse power recognition model; For real-time sample coupled feature vectors; These are the parameters for the initial training of the model; This is the inverse power discrimination threshold.

6. The method for identifying reverse power in a distribution network based on photovoltaic-load time-series coupling according to claim 1, characterized in that, Step S4 includes: The predicted probability output of the initially trained inverse power recognition model for the samples is compared with the preset verification labels to construct a loss function, and the parameters of the initially trained inverse power recognition model are updated using gradient optimization; the parameter update formula is as follows: in, The parameters of the inverse power recognition model to be optimized during initial training; The learning rate; For the loss function with respect to the parameters The gradient; The loss function is constructed based on the verification labels; Optimization stops when the loss function decreases and stabilizes with each training epoch, or when the change in loss between two adjacent epochs is less than a preset threshold, thus obtaining a well-trained inverse power recognition model. ,in, The parameters for the optimized inverse power identification model.

7. The method for identifying reverse power in a distribution network based on photovoltaic-load time-series coupling according to claim 6, characterized in that, The preset verification tags include: The verification results are derived from at least one or more of the following: verification criteria based on the active power direction at the metering point, power direction verification criteria obtained based on power flow calculation or state estimation, and reverse power events recorded by protection devices or reverse power relays; the verification results are converted into verification tags. ,in, This indicates that the verification process has resulted in reverse power. This indicates that no reverse power has occurred during the verification process; and the verification tag is matched with the initial identification output at the corresponding time / time window.

8. A distribution network reverse power identification system based on photovoltaic-load time-series coupling, characterized in that, include: Module M1: Acquires historical and real-time operating data of the target distribution network, and performs time alignment between the historical and real-time operating data of the target distribution network to obtain photovoltaic output time-series data and load time-series data; Module M2: Constructs time-series samples based on photovoltaic output time-series data and load time-series data, extracts coupling features that characterize the photovoltaic-load time-series coupling relationship based on the time-series samples, and forms a coupling feature matrix; Module M3: Constructs an inverse power recognition model, trains the inverse power recognition model using the coupling feature matrix of historical samples to obtain the initially trained inverse power recognition model; inputs the coupling feature matrix of real-time samples into the initially trained inverse power recognition model to obtain the initial inverse power recognition result; Module M4: Calculates the loss function based on the initial inverse power identification result and the verification result, and optimizes the initially trained inverse power identification model based on the loss function to obtain the trained inverse power identification model; Module M5: Inputs the coupling feature matrix of real-time samples into the trained inverse power recognition model, outputs the inverse power recognition result, and generates the occurrence time, duration and amplitude of the inverse power event.

9. The distribution network reverse power identification system based on photovoltaic-load time-series coupling according to claim 8, characterized in that, The module M1 includes: Module M1.1: Acquires historical and real-time operating data from the target identification points of the target distribution network; the target identification points include any one or more of the following: grid connection point, feeder head, transformer substation meter, or low-voltage side of the transformer; the historical and real-time operating data include: raw measured values ​​of active power on the photovoltaic side and / or raw measured values ​​of active power on the load side; Module M1.2: Sets a uniform sampling period and constructs a uniform time grid, so that power data from different sources correspond one-to-one on the same time scale. The time grid formula is: in, For the first A unified sampling time; To align the start times of the time period; To standardize the sampling period; The time index is a non-negative integer; Module M1.3: Maps photovoltaic power data and load power data onto the unified time grid using mapping methods including interval average resampling or interpolation alignment; when there are clock asynchronization issues between different devices or fixed communication link delays, further time offset correction is performed to obtain photovoltaic power output time-series data on the unified time grid. With load time series data ; The module M2 includes: Module M2.1: Employs a sliding time window for photovoltaic power output timing data. With load time series data The sample is segmented to form multiple time windows; each sample contains continuous time windows. There are sampling points, and the sliding step size is . This ensures that each sample can cover both the dynamic changes of photovoltaics and loads, and also meet the real-time requirements of online identification; Module M2.2: Constructs a net injected power sequence within the sample window to characterize the surplus of photovoltaic output relative to the load; in, To unify the time The net injection power is below; For a moment The active power output of photovoltaic power; For a moment The active power of the load; when This indicates that there is a tendency or risk of reverse power transmission to the upper-level power grid at the target identification point; Module M2.3: For each time series sample, extract a set of coupling features that includes at least synchronization / misalignment coupling, coupling strength, inverse power margin, and persistence, and construct a coupling feature matrix based on the set of coupling features; The synchronous / misaligned coupling and coupling strength are characterized by hysteresis correlation features. The correlation coefficient is calculated within a given maximum hysteresis range, and the maximum correlation is taken as the coupling strength index for that sample. The formula is as follows: in, For the first The maximum hysteresis correlation coefficient of each sample is used to characterize the coupling strength between photovoltaics and load; The lag is the time shift of the load sequence relative to the photovoltaic sequence; This represents the maximum lag search range; Operators are used to calculate the correlation coefficient; For the sequence by hysteresis Operators for time translation; For the first A sequence vector of photovoltaic power output within a sample window; For the first Load sequence vector within each sample window; Simultaneously, inverse power margin and persistence features are extracted from the net injected power sequence, and these features are concatenated into a sample feature vector in a predetermined order. Stack all sample feature vectors row-wise to form a coupled feature matrix. Each row corresponds to a sample, and each column corresponds to a coupling feature.

10. The distribution network reverse power identification system based on photovoltaic-load time-series coupling according to claim 8, characterized in that, The module M3 constructs an inverse power identification model, which is trained using the coupling feature matrix of historical samples to obtain an initially trained inverse power identification model, including: Using the coupled feature vectors of historical samples As input to the inverse power identification model, with inverse power labels As a supervisory signal, an inverse power identification model is constructed. The probability of the output sample having inverse power The inverse power recognition model is initially trained by using binary cross-entropy as the loss function and iteratively optimizing it to ensure that the output probability of the model matches the true label. The loss function is: in, For training the loss function; These are the parameters for the inverse power identification model; This represents the number of historical samples. For sample index; For the first Inverse power label of each sample, This indicates that reverse power has occurred. This indicates that no reverse power occurred; For the model to the first The predicted probability of each sample having inverse power; It is the natural logarithm function; After training, the inverse power recognition model is obtained. ,in, These are the parameters of the inverse power recognition model when the initial training is completed; In module M3, the coupling feature matrix of the real-time samples is input into the initially trained inverse power recognition model to obtain the initial inverse power recognition result, including: Obtain real-time sample coupled feature vectors Couple real-time samples with feature vectors Input the inverse power recognition model trained for the first time Obtain the real-time probability of inverse power occurrence And according to the preset threshold The initial identification result of the output inverse power is determined by the following formula: in, This is the initial identification result of inverse power; This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. The inverse power probability output by the initially trained inverse power recognition model; For real-time sample coupled feature vectors; These are the parameters for the initial training of the model; The inverse power discrimination threshold; The module M4 includes: The predicted probability output of the initially trained inverse power recognition model for the samples is compared with the preset verification labels to construct a loss function, and the parameters of the initially trained inverse power recognition model are updated using gradient optimization; the parameter update formula is as follows: in, The parameters of the inverse power recognition model to be optimized during initial training; The learning rate; For the loss function with respect to the parameters The gradient; The loss function is constructed based on the verification labels; Optimization stops when the loss function decreases and stabilizes with each training epoch, or when the change in loss between two adjacent epochs is less than a preset threshold, thus obtaining a well-trained inverse power recognition model. ,in, The parameters for the optimized inverse power identification model.

Images