Power distribution network power reverse transmission risk processing method and device

Abstract

The embodiment of the invention provides a power distribution network power reverse transmission risk processing method and device. The method comprises the steps of obtaining multi-source heterogeneous data, and performing space-time alignment processing through a data alignment algorithm to obtain multi-modal data after space-time alignment; inputting the multi-modal data after time-space alignment into a hybrid model for feature extraction and dynamic attention distribution, and generating an output prediction value and a load demand prediction value of the distributed power supply accessed to the power distribution network in a future time period; based on the output prediction value and the load demand prediction value, determining a power reverse transmission risk assessment index of the distributed power supply to the power distribution network; and if the power reverse transmission risk assessment index is greater than a preset risk threshold, triggering a multi-target adaptive defense decision model based on dynamic game, and generating a power reverse transmission risk coping strategy of the power distribution network. The method is used for actively coping with power reverse transmission and improving the operation safety of a power grid.

Description

Technical Field

[0001] This application relates to the field of power technology, and in particular to a method and apparatus for handling power backflow risk in power distribution networks. Background Technology

[0002] With the high proportion of distributed energy sources such as photovoltaic and wind power being connected to the distribution network, power backfeed has become a significant problem threatening the safe operation of the distribution network in specific scenarios where peak output of distributed power sources overlaps with off-peak loads. Power backfeed can easily lead to a series of safety risks such as voltage exceeding limits, equipment overload, and relay protection malfunctions (such as cascading trips).

[0003] Currently, existing methods for dealing with power backfeed risks are mainly based on the "post-event response" model, which involves passively intervening after power backfeed occurs by adjusting transformer taps and switching reactive power compensation equipment.

[0004] However, the existing response method has a lag in control actions, which reduces the safety of power grid operation. Summary of the Invention

[0005] The distribution network power backfeed risk handling method and apparatus provided in this application are used to proactively respond to power backfeed and improve the safety of power grid operation.

[0006] In a first aspect, embodiments of this application provide a method for handling power backflow risk in a distribution network, including:

[0007] Multi-source heterogeneous data is acquired and spatiotemporally aligned using a data alignment algorithm to obtain spatiotemporally aligned multimodal data; wherein, the multi-source heterogeneous data includes structured data inside the power grid and unstructured data outside the power grid;

[0008] The spatiotemporally aligned multimodal data is input into the hybrid model for feature extraction and dynamic attention allocation to generate output and load demand forecasts for distributed power sources connected to the distribution network in future time periods.

[0009] Based on the predicted output and load demand, the risk assessment index for the power backflow of the distributed power source to the distribution network is determined.

[0010] If the power backfeed risk assessment index is greater than the preset risk threshold, a multi-objective adaptive defense decision model based on dynamic game theory is triggered to generate a power backfeed risk response strategy for the distribution network.

[0011] In one possible implementation, the internal structured data of the power grid includes historical and real-time voltage data, historical and real-time current data, and historical and real-time power data; the external unstructured data of the power grid includes gridded weather forecast data, holiday information, and social event information.

[0012] In one possible implementation, the hybrid model includes a long short-term memory network and a Transformer encoder;

[0013] The step of inputting spatiotemporally aligned multimodal data into a hybrid model for feature extraction and dynamic attention allocation includes:

[0014] Temporal features are extracted from the multimodal data using a long short-term memory network to obtain local temporal features, and the local temporal features are then positionally encoded to obtain an encoded feature sequence.

[0015] The Transformer encoder uses a multi-head attention mechanism to assign attention weights to the feature sequence in order to determine the spatiotemporal features most correlated with power backfeed, and based on the spatiotemporal features, to determine the output prediction value and load demand prediction value.

[0016] In one possible implementation, determining the power backfeed risk assessment index of the distributed power source to the distribution network includes:

[0017] Based on the predicted output and load demand, the node voltage and branch power flow distribution of the distribution network in the future time section are obtained by power flow calculation using the power grid topology parameters.

[0018] Based on the node voltage and branch power flow distribution, the geometric safe distance between the operating state of the distribution network and the boundary of the safe domain in the future time section is determined.

[0019] Obtain the node voltage over-limit risk index and the branch power flow over-limit risk index, and then fuse them to obtain the comprehensive risk index;

[0020] The power backfeed risk assessment index is generated based on the geometric safety distance and / or the comprehensive risk index.

[0021] In one possible implementation, the safety domain boundary includes at least one of line capacity constraints, voltage limit constraints, main transformer capacity constraints, feeder and main transformer capacity matching constraints, distributed power source access constraints, and anti-power backfeeding constraints.

[0022] In one possible implementation, the method further includes:

[0023] Construct a defender, a defender strategy set, an attacker, and an attacker strategy set; wherein, the defender represents the power grid control system of the distribution network, the defender strategy set includes at least one set of coordinated control action combinations, the coordinated control combinations are used to adjust the operating state of the distribution network to eliminate the risk of power backfeed, the attacker represents the uncertainty factor of power backfeed, and the attacker strategy set is used to simulate the most unfavorable uncertainty scenario that leads to power backfeed;

[0024] Based on the system network loss cost, voltage deviation penalty, control equipment action cost, voltage exceeding the upper limit penalty, and branch current exceeding the limit penalty, a defender's revenue function is constructed.

[0025] Based on the degree of voltage exceeding the upper limit, the degree of branch current exceeding the upper limit, the degree of distributed power output exceeding the upper limit, and the degree of power backflow, an attacker's profit function is constructed;

[0026] The constraints are constructed, including power balance constraints, voltage safety constraints, branch capacity constraints, equipment operation constraints, equipment operation limit, power backfeed constraints, and topology connectivity constraints.

[0027] Based on the defender, defender's strategy set, defender's payoff function, attacker, attacker's strategy set, attacker's payoff function, and constraints, a multi-objective adaptive defense decision model for the dynamic game is constructed.

[0028] In one possible implementation, the triggering of a multi-objective adaptive defense decision model based on dynamic game theory to generate a power backflow risk response strategy for the distribution network includes:

[0029] The equilibrium point of the multi-objective adaptive defense decision model is solved by a mixed-integer linear programming method. The equilibrium point is used to characterize the defender's ability to predict the attacker's optimal response to any strategy in the defender's strategy set.

[0030] Based on the equilibrium point, a power backflow risk mitigation strategy for the distribution network is generated.

[0031] In one possible implementation, the method further includes:

[0032] Based on the power backfeed risk response strategy, a set of control instructions is sent to the edge agent. The edge agent is deployed in the substation or key node of the power distribution network to convert the set of control instructions into equipment instructions and send them to the end-side equipment for execution.

[0033] Obtain the actual execution effect of the strategy uploaded by the edge agent;

[0034] Based on the actual implementation effect of the strategy, the model parameters of the hybrid model and / or the model parameters of the multi-target adaptive defense decision model are updated using an online incremental learning algorithm.

[0035] Secondly, embodiments of this application provide a power backflow risk handling device for power distribution networks, comprising:

[0036] The acquisition module is used to acquire multi-source heterogeneous data and perform spatiotemporal alignment processing through a data alignment algorithm to obtain spatiotemporally aligned multimodal data; wherein, the multi-source heterogeneous data includes structured data inside the power grid and unstructured data outside the power grid;

[0037] The prediction module is used to input spatiotemporally aligned multimodal data into a hybrid model for feature extraction and dynamic attention allocation, generating output and load demand predictions for distributed power sources connected to the distribution network in future time periods.

[0038] The determination module is used to determine the power backflow risk assessment index of the distributed power source to the distribution network based on the power output forecast value and the load demand forecast value.

[0039] The generation module is used to trigger a multi-objective adaptive defense decision model based on dynamic game theory to generate a power backflow risk response strategy for the distribution network if the power backflow risk assessment index is greater than a preset risk threshold.

[0040] The distribution network power backfeed risk handling method and apparatus provided in this application use multimodal data as input and a hybrid model to accurately predict power backfeed risk. At the same time, when the power backfeed risk is large, a multi-objective adaptive defense decision model based on dynamic game is triggered, so that power backfeed can be upgraded from passive response to active response, thereby improving the safety of distribution network operation. Attached Figure Description

[0041] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0042] Figure 1 A schematic diagram of the distribution network power backfeed risk handling method provided in this application;

[0043] Figure 2 A schematic diagram of the hybrid model architecture provided in the embodiments of this application;

[0044] Figure 3 This is a schematic diagram of the cloud-edge collaborative execution framework provided in the embodiments of this application;

[0045] Figure 4 A flowchart illustrating the online incremental learning workflow provided in this application embodiment;

[0046] Figure 5 A schematic diagram of the power backflow risk handling device for the power distribution network provided in this application;

[0047] Figure 6 A schematic diagram of the structure of the electronic device provided in this application.

[0048] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0049] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0050] With the significant increase in the penetration rate of distributed generation (DG) in distribution networks, power backflow may occur in certain scenarios where the peak output of DG (such as wind power and solar power) coincides with the off-peak load (e.g., during holidays when there is ample sunshine at midday and industrial and commercial loads are low).

[0051] Power backfeeding can lead to safety risks such as voltage overruns, equipment overloads, and relay protection malfunctions (e.g., cascading trips), threatening the stable operation of the distribution network. Furthermore, the intermittency and volatility of distributed energy resources (e.g., sudden changes in solar intensity and wind speed) further exacerbate the uncertainty of grid operation, thus placing higher demands on the real-time performance, robustness, and coordination of grid dispatching strategies.

[0052] Traditional methods for dealing with power backfeed risk have the following limitations: (1) Lagging perception and insufficient data fusion. Traditional methods mainly rely on structured data from Supervisory Control and Data Acquisition (SCADA) systems, but lack the fusion and utilization of unstructured data from outside the power grid. In addition, when multi-source heterogeneous data have inconsistent spatiotemporal scales (such as minute-level meteorological data and second-level power grid data) and imbalanced modal confidence (such as the difference between the error probability of weather forecasts and the accuracy of power grid measurements), there is a lack of effective processing mechanisms, resulting in a lag in risk perception and difficulty in achieving early warning. (2) Passive decision-making and static optimization. Traditional control strategies are mostly "post-event response" modes, that is, passive intervention is carried out after power backfeed occurs by adjusting transformer taps and switching reactive power compensation equipment. Although some systems adopt static optimization models (such as the minimum network loss strategy based on power flow calculation), they are difficult to cope with the real-time fluctuations and uncertainties of distributed power output and load demand. For example, when extreme weather (such as sudden rain) causes a sudden change in the output of distributed power sources, the static model cannot dynamically adjust the defense strategy, resulting in delayed or overly conservative control actions, which affects the economy and security of the power grid. (3) Difficulty in coordination and conflict of multiple objectives. Traditional methods have a single control objective, usually focusing only on "eliminating the risk of backfeeding power", and fail to dynamically coordinate and optimize multiple objectives such as safety constraints, reducing network losses, improving power quality, and reducing the cost of control actions. In addition, there is a lack of efficient coordination mechanisms among various control devices (such as distributed power sources, energy storage systems, and flexible loads), making it difficult to achieve global optimization. For example, adjusting the transformer tap may improve the voltage level, but may increase the power flow in the branch, leading to overload risks in other nodes, and the existing system lacks the ability to balance conflicts of multiple objectives.

[0053] To address the aforementioned issues, this application proposes a solution that deeply integrates multimodal data for advanced and accurate perception, simulates the impact of uncertainty based on dynamic game theory, and generates an adaptive and collaborative proactive defense strategy. This solution aims to overcome the limitations of traditional passive response modes, achieve a balance between safe operation and economic optimization of the power distribution network, and enhance system resilience.

[0054] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0055] Figure 1 This is a schematic diagram of the distribution network power backfeed risk handling method provided in this application, such as... Figure 1 As shown, the method includes:

[0056] Step 110: Acquire multi-source heterogeneous data and perform spatiotemporal alignment processing using a data alignment algorithm to obtain spatiotemporally aligned multimodal data. The multi-source heterogeneous data includes structured data from within the power grid and unstructured data from outside the power grid.

[0057] Step 120: Input the spatiotemporally aligned multimodal data into the hybrid model for feature extraction and dynamic attention allocation to generate the output forecast and load demand forecast of the distributed power sources connected to the distribution network in future time periods.

[0058] Step 130: Based on the power output forecast and load demand forecast, determine the risk assessment index for power backfeeding to the distribution network by distributed generation.

[0059] Step 140: If the power backfeed risk assessment index is greater than the preset risk threshold, the multi-objective adaptive defense decision model based on dynamic game theory is triggered to generate a power backfeed risk response strategy for the distribution network.

[0060] Regarding step 110 above, structured data from within the power grid can be obtained from the SCADA system; in addition, key information can be extracted and quantified from unstructured data outside the power grid using natural language processing technology.

[0061] Structured data within the power grid characterizes its current and historical conditions, directly reflecting its physical state. Unstructured data outside the power grid represents external environmental data that will affect the grid.

[0062] In addition, for the original multi-source heterogeneous data, data cleaning, noise reduction and normalization can be performed first, and then spatiotemporal alignment can be performed through data alignment algorithms.

[0063] In cases where multi-source heterogeneous data exhibits inconsistencies in spatiotemporal scale and confidence levels, an improved Dynamic Time Warping (DTW) algorithm can be introduced for data alignment during spatiotemporal alignment. This algorithm, based on traditional DTW, combines time-domain and frequency-domain features to calculate a weighted cumulative distance, expressed as follows:

[0064]

[0065] In the above formula, For the aligned multimodal data, , These are the sampling points in the time domain for structured data inside the power grid and unstructured data outside the power grid, respectively. , for , The frequency domain representation is obtained through Fast Fourier Transform (FFT); , These are the weighting coefficients in the time and frequency domains.

[0066] For example, in some embodiments, the structured data inside the power grid includes historical and real-time voltage data, historical and real-time current data, and historical and real-time power data; the unstructured data outside the power grid includes gridded weather forecast data, holiday information, and social event information.

[0067] Historical data can include records of voltage, current, and power at every moment in the past. Real-time data can include the voltage, current, and power readings at the current moment.

[0068] Among these, gridded weather forecast data divides a region into several finely divided squares (e.g., 1 km x 1 km), providing independent and accurate weather parameter forecasts for each square. Holiday information can be pre-set dates or holidays, with different electricity consumption behaviors occurring on different holidays; for example, electricity load is relatively high during the Spring Festival. Social event information refers to "planned disturbances" occurring at specific times and locations, such as concerts or sporting events, and electricity consumption behaviors may differ at different times.

[0069] Regarding step 120 above, the hybrid model integrates neural networks (adept at handling complex patterns) and physical mechanism models (based on power grid operation patterns). It can automatically extract key clues affecting predictions from massive amounts of aligned multimodal data. For example, the hybrid model discovered that "solar radiation intensity" is the primary feature affecting photovoltaic power generation, while "event start time" is a key feature affecting local loads.

[0070] Dynamic attention allocation refers to the "intelligent focus" of the hybrid model. The hybrid model focuses on different aspects at different times. For example, when predicting photovoltaic output at noon, the hybrid model will pay close attention to "solar radiation" data; when predicting evening load, it will pay more attention to "holidays" and "social events" data.

[0071] The output forecast refers to the amount of electricity that distributed energy sources such as wind and solar power can generate in each future time period. The load demand forecast refers to the amount of electricity that users will need in each future time period.

[0072] Regarding step 130 above, when the local power generation in a certain area far exceeds the local load, the excess power will flow back to the upstream power grid, thus forming power backflow. Power backflow can easily cause a series of safety risks such as voltage exceeding limits, equipment overload, and relay protection malfunctions (such as over-tripping).

[0073] By substituting the predicted power output and load demand values ​​into the power grid model for fast power flow calculation, it is possible to determine which nodes and lines will experience backflow, and the severity of the backflow.

[0074] The risk assessment index can be a quantified hazard level. It is calculated by taking into account factors such as the magnitude of backfeed power, the number of nodes causing voltage overshoot, and the margin from the safety boundary. The higher the risk assessment index, the higher the risk of power backfeed.

[0075] Regarding step 140 above, when the power backfeed risk assessment index of the distribution network exceeds the preset threshold, it is determined that there is a high power backfeed risk, and the subsequent multi-objective adaptive defense decision-making process based on dynamic game will be automatically triggered.

[0076] Dynamic game theory refers to defense decisions that are not one-sided. For example, a dispatcher can instruct distributed power sources to reduce generation (a strategy), but the power source owner may be unwilling (because it affects revenue). By using a multi-objective adaptive defense decision model, the different interests of multiple parties (dispatcher, generator, and user) are taken into account, their interactions and possible reactions are simulated, and the most likely acceptable and effective strategy is found to deal with power backfeed.

[0077] Among them, multiple objectives can refer to the factors that should be considered in generating response strategies, specifically including the primary safety objective (i.e., eliminating the risk of power backflow), the economic objective (to maximize the benefits of distributed power owners and reduce tripping losses), and the reliability objective (to minimize the impact on users' normal power supply).

[0078] In addition, adaptive means that the response strategy is not static. It automatically adjusts and optimizes the strategy based on real-time updated forecast data and power grid status, and has the ability to learn and evolve.

[0079] For example, power backfeed risk response strategies may include: (1) adjusting the grid operation mode and adjusting the reactive power compensation device; (2) issuing instructions to distributed power sources to actively reduce power generation; and (3) incentivizing specific users to increase electricity consumption during peak backfeed periods.

[0080] The distribution network power backfeed risk handling method provided in this application constructs deeply integrated multimodal data and uses a hybrid model for advanced and accurate perception to determine power backfeed risk assessment indicators. Then, based on the idea of ​​dynamic game theory, it simulates the impact of uncertainty to generate adaptive and collaborative proactive defense and response strategies, which can achieve the unity of safe operation and economic optimization of the distribution network.

[0081] Furthermore, in some embodiments, to effectively capture long-term dependencies in time series and dynamically focus on key features, the hybrid model can employ an attention-based LSTM-Transformer hybrid model. Figure 2 This is a schematic diagram of the hybrid model architecture provided in the embodiments of this application, such as... Figure 2 As shown, the hybrid model includes a Long Short-Term Memory (LSTM) network and a Transformer encoder.

[0082] When performing feature extraction and dynamic attention allocation using a hybrid model, the Long Short-Term Memory (LSTM) network is responsible for extracting temporal features from multimodal data, obtaining local temporal features, and encoding the local temporal features at their locations to obtain the encoded feature sequence. The Transformer encoder, on the other hand, is responsible for allocating attention weights to the feature sequence using a multi-head attention mechanism to determine the spatiotemporal features most correlated with power backfeed, and based on the spatiotemporal features, determining the output prediction and load demand prediction values.

[0083] Continue to refer to Figure 2 The specific structure of the hybrid prediction model combining Long Short-Term Memory Network and Transformer encoder is as follows:

[0084] (1) Input sequence: The model receives multivariate time series inputs, such as historical power load, meteorological data and other multidimensional features;

[0085] (2) LSTM layer: Composed of multiple layers of Long Short-Term Memory (LSTM) units, its gating mechanism (input gate, forget gate, output gate) can effectively capture the local short-term patterns and temporal dynamics of the sequence (e.g., the periodic variation of photovoltaic power output from sunrise to sunset in sunny weather, and the different patterns of load on weekdays and weekends), alleviating the gradient vanishing problem. The output is the hidden state sequence for each time step.

[0086] (3) Position encoding: Since the Transformer itself does not have the ability to perceive the sequence order, it is necessary to explicitly inject absolute or relative position information into the feature sequence output by the LSTM through the sine and cosine position encoding functions.

[0087] (4) Transformer Encoder: The core is a multi-head self-attention mechanism. By using multiple attention heads in parallel, the model can jointly focus on and dynamically calculate the importance weights of different input features (light intensity, temperature, wind speed, historical power values, holiday signs, etc.) to the prediction results at different prediction times. For example, when predicting the risk of midday power backflow, the model will automatically assign higher attention weights to features such as "light intensity" and "commercial area load"; while when predicting the risk at night, it may pay more attention to "residential load" and "temperature". Each encoder layer also includes a feedforward neural network, residual connections, and layer normalization to enhance representation ability and training stability.

[0088] (5) Output layer: The feature sequence output by the Transformer encoder is aggregated and then mapped to the final prediction target (such as the power value of multiple future time steps) through a fully connected layer.

[0089] In this embodiment, the temporal features obtained after multimodal fusion are first input into an LSTM layer for preliminary temporal feature extraction, and its output is then fed into a Transformer encoder layer. Through an attention weight matrix, the model can focus on the spatiotemporal features most relevant to power backfeed, and finally output the distributed power output for future time periods through a fully connected layer. and load demand The predicted value.

[0090] This application embodiment utilizes an LSTM-Transformer hybrid model, which enables the model to focus on the spatiotemporal features most relevant to power backfeed, effectively reducing prediction errors for distributed energy processing and load demand, and improving prediction accuracy.

[0091] Furthermore, when predicting power backfeed risk, the power backfeed risk assessment index is calculated using the minimum geometric safety distance and / or comprehensive risk index across multiple future timeframes. Specifically, the power backfeed risk assessment index can be determined through the following steps:

[0092] Step 11: Based on the output forecast and load demand forecast, use the power grid topology parameters to calculate the node voltage and branch power flow distribution of the distribution network in the future time section.

[0093] Step 12: Based on node voltage and branch power flow distribution, determine the geometric safe distance between the operating state of the distribution network at the future time section and the boundary of the safe domain;

[0094] Step 13: Obtain the node voltage over-limit risk index and the branch power flow over-limit risk index, and fuse them to obtain the comprehensive risk index;

[0095] Step 14: Generate power backfeed risk assessment indicators based on geometric safety distance and / or comprehensive risk index.

[0096] In this embodiment, power backfeed risk prediction is the core link to achieve "advanced perception". Its goal is to transform the cleaned, aligned and feature-fused multimodal data into quantitative and operable assessment indicators for the power backfeed risk of the distribution network in a specific future period (such as the next 1-4 hours).

[0097] For step 11, firstly, power flow calculation and safety boundary comparison are performed, and the predicted values ​​are then compared. and The data, combined with power grid topology parameters, is used to perform fast power flow calculations to obtain the node voltage and branch power flow distribution of the future cross section.

[0098] For step 12, one of the core outputs of power backfeed risk prediction is the calculation of the geometric security distance (GSD) from the current (or predicted) operating point to the boundary of the security domain. The current operating point represents the current operating state of the power grid.

[0099] The geometric safety distance from the operating state of the distribution network to the boundary of the safety domain can be expressed by a basic formula as the distance for a certain safety constraint.

[0100] For example, the security domain boundary includes at least one of the following security constraints: line capacity constraint, voltage limit constraint, main transformer capacity constraint, feeder and main transformer capacity matching constraint, distributed power source access constraint, and anti-power backfeed constraint.

[0101] For (1) line capacity constraint: This reflects the proportion of the current remaining capacity of the line to its maximum capacity. The larger the value of this proportion, the farther the current (or predicted) operating point is from the "safe distance" of full load or overload, and the greater the safety margin.

[0102]

[0103] In the above formula, This indicates the proportion of the line's current remaining capacity to its maximum capacity. For the first The maximum transmission capacity limit of each line; For the flow through the first The first of the lines The load of a feeder or feeder section; Let be the load factor, representing the th The contribution ratio of each load to the total power flow can be set to 1 in a simple case; In order to be with the first The total number of loads related to power flow on each line.

[0104] For (2) voltage limit constraints: calculate the normalized absolute voltage deviation to quantify the degree of current voltage deviation from the limit. The smaller the value, the closer the voltage at the current (or predicted) operating point is to the boundary of the safe operating range.

[0105]

[0106] In the above formula, This indicates the current voltage deviation from the limit; For the first The actual operating voltage of each node; This is the upper / lower voltage limit for this node; The voltage reference value is usually taken as the rated voltage (such as 10 kV or 35 kV) to normalize the absolute voltage deviation, making the GSD dimensionless and easy to compare.

[0107] For (3) main transformer capacity constraint: it reflects the remaining capacity margin when the load of a main transformer needs to be transferred to other normal main transformers after it goes out of operation due to a fault. The larger the value, the stronger the load transfer capability and the safer the system.

[0108]

[0109] In the above formula, Indicates the remaining capacity margin; For the first Rated capacity of the main transformer; For the main change The first power supply The load of the feeder; For all variables from the main variable A collection of power supply feeders.

[0110] For (4) feeder-transformer capacity matching constraint: the feeder's power transfer capacity must match the capacity of its upstream transformer. This is used to evaluate the balance and robustness of the network structure. The higher the matching degree, the more reasonable the network structure.

[0111]

[0112] In the above formula, Indicates the degree of matching; The thermal stability capacity limit for the feeder; The rated capacity of the connected main transformer; This refers to the set of feeders that are connected to this feeder. The set of all feeders connected to the main transformer.

[0113] Regarding (5) distributed generation (DG) access constraints: the potential for DG output increases is quantified, which helps to assess the impact risk on the distribution network when the actual DG output far exceeds the forecast.

[0114]

[0115] In the above formula, There is room for growth in contributing to DG; This is the maximum allowable output of DG; To contribute to DG's predictions.

[0116] For (6) anti-power backfeed constraint: it measures the risk of power backfeed when the total DG output exceeds the sum of the total load and network losses. A negative value indicates that power will be fed back to the upstream power grid, posing a risk of voltage exceeding limits and protection malfunction. The larger the positive value, the stronger the system's ability to resist power backflow.

[0117]

[0118] In the above formula, Indicates the ability to resist power backflow; Total load demand; Contribute to DG's overall efforts; This is due to network loss.

[0119] For step 13, in order to further comprehensively assess the degree of risk, the system will calculate the node voltage over-limit risk index and the branch power flow over-limit risk index.

[0120] Branch line tidal current over-limit risk The calculation can be expressed as:

[0121]

[0122] In the above formula, branch road Power transmission limits; It is a branch trend The probability density function can be obtained from probability current flow or historical error distribution.

[0123] The physical meaning of equation (8) above is the over-limit part. The expected value of the mathematical expression.

[0124] The risk of node voltage exceeding the limit is mainly divided into the risk of voltage exceeding the upper limit. and the risk of voltage exceeding the lower limit The core calculation formula is as follows:

[0125]

[0126]

[0127] In the above formula, For nodes The operating voltage is the real-time or predicted value obtained through power flow calculation; , These are the upper and lower limits for the safe operation of node voltages, which are hard constraints stipulated by the power grid safety operation regulations. Voltages exceeding these limits are considered to be exceeding the limits. , In probabilistic risk assessment, the extreme upper and lower limits that voltage may reach are used to define the range of risk integrals; node voltage The probability density function describes the likelihood of a node voltage occurring within a certain range, and can be obtained through historical data statistics or probabilistic power flow calculations.

[0128] For step 14 above, the final output is the minimum geometric safety distance and / or comprehensive risk index for multiple future time sections, based on which the power backfeed risk assessment index is determined.

[0129] In this embodiment, a risk threshold can be preset. When the power backfeed risk assessment index exceeds the preset risk threshold, it is determined that there is a high power backfeed risk, and the subsequent multi-objective adaptive defense decision-making process based on dynamic game will be automatically triggered.

[0130] The power backfeed risk prediction method provided in this application combines deep learning's precise point prediction with the physical model-based security domain / risk theory to achieve step-by-step quantification of power backfeed risk from "whether it occurs" to "severity," enabling proactive and accurate perception of power backfeed risk and providing precise time windows and decision-making basis for proactive defense.

[0131] Furthermore, in some embodiments, after accurately sensing the risk of power backfeeding, the problem of defending against power backfeeding can be modeled as a Stackelberg game model, and multi-objective adaptive defense decisions can be made based on the Stackelberg game model.

[0132] In the Stackelberg game model, the defender (i.e., the power grid control system) acts as the leader, and the virtual attacker (representing an unfavorable uncertainty scenario that leads to power backflow) acts as the follower. The model construction specifically includes the following steps:

[0133] Step 21: Construct the defender, the defender's policy set, the attacker, and the attacker's policy set.

[0134] Among them, the defender represents the power grid control system of the distribution network, the defender strategy set includes at least one set of coordinated control action combinations, the coordinated control combinations are used to adjust the operating state of the distribution network to eliminate the risk of power backfeeding, the attacker represents the uncertainty factors of power backfeeding, and the attacker strategy set is used to simulate the most unfavorable uncertainty scenario that leads to power backfeeding.

[0135] Step 22: Construct the defender's revenue function based on the system network loss cost, voltage deviation penalty, control equipment action cost, voltage exceeding the upper limit penalty, and branch current exceeding the limit penalty.

[0136] Step 23: Construct the attacker's reward function based on the degree of voltage exceeding the upper limit, the degree of branch current exceeding the upper limit, the degree of distributed power output exceeding the upper limit, and the degree of power backflow.

[0137] Step 24: Construct constraints. These constraints include power balance constraints, voltage safety constraints, branch capacity constraints, equipment operation constraints, equipment operation frequency limits, power backfeed constraints, and topology connectivity constraints.

[0138] Step 25: Based on the defender, the defender's strategy set, the defender's payoff function, the attacker, the attacker's strategy set, the attacker's payoff function, and the constraints, construct a dynamic game-theoretic multi-objective adaptive defense decision model.

[0139] For step 21 above, the defender's (leader's) strategy set It includes all feasible combinations of coordinated control actions, designed to adjust the operating status of the distribution network to eliminate the risk of power backfeeding.

[0140] For example, the defender's strategy set includes the on-load tap changer tap position vector, the capacitor reactor switching state vector, the distributed generation active power output adjustment vector, and the flexible load adjustment vector. Its mathematical representation is a multidimensional mixed-integer decision space:

[0141]

[0142] In the above formula, u is the defense strategy vector, which includes the action combination of all control devices; Let be the tap position vector of the on-load tap-changing transformer, with dimension . Indicates the number of transformers, each component It is an integer, and its value range is determined by the physical limitations of the equipment, representing the tap position; Let the switching state vector of the capacitor reactor be dimensional. This indicates the number of reactive power compensation devices, and each component... Indicates a "disconnected" or "connected" status; For the active power output adjustment vector of distributed power sources, dimension Indicates the number of distributed power sources, each component Indicates the first The active power output setpoint of each distributed power source, among which To its maximum permissible output; For flexible load adjustment vectors, dimension Indicates the number of adjustable loads, each component Indicates the first The active power adjustment of each load, with positive values ​​indicating load reduction and negative values ​​indicating load increase. , These are the lower and upper limits for active power regulation. The constraints include physical constraints such as the number of times equipment can be operated, cooperative operation logic (the correlation between capacitor switching and voltage deviation), and power flow equations.

[0143] The strategy set of virtual attackers (followers) The simulation addresses the most unfavorable uncertainty scenario leading to power backfeed. For example, the attacker's strategy set includes distributed generation output disturbance vectors and load demand disturbance vectors. Mathematically, this is represented by a bounded disturbance parameter space:

[0144]

[0145] In the above formula, Let be the attack strategy vector, representing the perturbation to the predicted value; This is the output disturbance vector of the distributed generation, with the same dimension as the number of distributed generation sources, and each component... Indicates the first The range of the additive disturbance in the output of each distributed power source is determined by the distribution of the prediction error. The upper confidence limit for the historical prediction error of distributed power generation output; This is the load demand disturbance vector, with the same dimension as the number of load nodes, and each component... Indicates the first An additive disturbance to load demand; a positive value indicates an increase in load, and a negative value indicates a decrease in load. The upper confidence limit for historical load forecasting error; disturbance constraint Limit the overall energy of the disturbance to ensure the rationality of the attack scenario. It is the L2 norm. That is, a vector Length, The disturbance budget parameter is calibrated from historical data statistics or extreme weather models.

[0146] In Stackelberg games, the defender, as the leader, chooses the strategy first. The aim is to optimize a multi-objective payoff function; the virtual attacker, acting as a follower, selects the most unfavorable perturbation against the defense strategy. To maximize its damage gains. Strategy set and The design ensures the physical feasibility of the game, in which Embedded power flow constraints in distribution networks Confidence boundaries that reflect uncertainty.

[0147] For step 22, the defender's profit function The aim is to minimize the overall cost of system operation while satisfying security constraints. This includes the defender's profit function. It comprises five parts: system network loss cost; voltage deviation penalty; control equipment operation cost; voltage exceeding the upper limit penalty; and branch current exceeding the upper limit penalty. This is achieved by minimizing... The system achieves economical operation while ensuring safety.

[0148] Defender's Profit Function The specific mathematical expression is:

[0149]

[0150] In the above formula, For the system at time Total active power loss reflects the economic efficiency of system operation; For nodes At any moment The voltage amplitude; This is a reference voltage value, usually the rated voltage; For control equipment At any moment Status (transformer tap position, capacitor switching status, distributed power output, flexible load regulation); For control equipment The motion cost coefficient reflects the mechanical wear and operating costs of equipment movements; For nodes The voltage amplitude; This refers to the upper limit safety threshold for voltage. branch road The current amplitude; branch road Maximum permissible current (thermal stability limit); , , , , These are weighting coefficients used to balance the importance of different objectives; To optimize the time range; The total number of nodes; The number of key voltage monitoring nodes; This refers to the number of critical branches.

[0151] Regarding step 23 above, the attacker's (representing the uncertainty factor causing power backfeed) payoff function aims to maximize the system's insecurity level, simulating the worst-case scenario. The attacker's payoff function comprises four parts: the degree to which voltage exceeds the upper limit; the degree to which branch current exceeds the upper limit; the degree to which distributed power output exceeds the upper limit; and the degree of power backfeed. By maximizing... The virtual attacker simulates the most unfavorable uncertainty scenario, making the defender's strategy robust.

[0152] The mathematical expression for the attacker's profit function is:

[0153]

[0154] In the above formula, For distributed power sources Actual output; For distributed power sources Maximum permissible output; For nodes The amount of power backfeed, when the node injected power is negative, indicates power backfeed; , , , , which is a weighting coefficient, reflecting the attacker's preference for different insecurity indicators; This represents the number of distributed power sources.

[0155] Regarding step 24 above, the power balance constraints are as follows:

[0156]

[0157]

[0158] In the above formula, , For nodes The injected active power and reactive power; For nodes , The voltage phase angle difference between them; , Let be the real and imaginary parts of the nodal admittance matrix.

[0159] Voltage safety constraints are as follows:

[0160]

[0161] Branch capacity constraints are as follows:

[0162]

[0163] In the above formula, branch road Apparent power; branch road The maximum permissible transmission power.

[0164] The equipment operating constraints are as follows:

[0165]

[0166]

[0167]

[0168]

[0169]

[0170] In the above formula, , These are the lower and upper limits of the transformer tap position; , For the active and reactive power outputs of distributed power sources; , For the lower and upper limits of active power output of distributed generation; , For the lower and upper limits of reactive power output of distributed generation; , These are the lower and upper limits for flexible load adjustment.

[0171] The number of times the equipment can be operated is limited as follows:

[0172]

[0173]

[0174] In the above formula, This represents the maximum permissible number of operations for the transformer tap changer within the optimization period. This represents the maximum number of times the capacitor can be switched on and off within the optimization cycle.

[0175] The power backfeed constraint is as follows:

[0176]

[0177] In the above formula, For nodes The injection power, when The time indicates power reversal.

[0178] The topological connectivity constraints are as follows:

[0179]

[0180]

[0181] In the above formula, branch road The connectivity state is 0 for disconnection and 1 for closure; For the set of all branches; For any loop in the distribution network, Indicates a loop The length.

[0182] Regarding step 25 above, in some embodiments, a mixed-integer linear programming method can be used to solve the equilibrium point of the multi-objective adaptive defense decision model; and based on the equilibrium point, a power backflow risk response strategy for the distribution network can be generated.

[0183] The equilibrium point is used to characterize the defender's ability to anticipate the attacker's optimal response to any strategy in the defender's strategy set.

[0184] In this embodiment, the Mixed Integer Linear (MILP) method is used to solve for the equilibrium point of the Stackelberg game. Under this equilibrium, the defender can anticipate the optimal response of the virtual attacker to any of its strategies, thus formulating a robust optimal strategy that ensures the safe and stable operation of the distribution network even under the most unfavorable uncertainty scenarios. The specific solution process is as follows:

[0185] (1) Transformation of the two-layer optimization problem: The Stackelberg game is transformed into a two-layer optimization problem, with the upper layer being the optimization problem of the defender and the lower layer being the optimization problem of the virtual attacker;

[0186] (2) Introduction of Karush-Kuhn-Tucker (KKT) conditions: The optimality conditions (KKT conditions) of the lower-level optimization problem are introduced into the upper-level optimization problem, transforming the two-level problem into a single-level mixed integer nonlinear programming problem;

[0187] (3) Linearization: The nonlinear constraints are linearized by methods such as the Big M method and McCormick envelope to obtain the mixed integer linear programming problem;

[0188] (4) Solving with commercial solvers: Using commercial solvers such as CPLEX and Gurobi to solve MILP problems and obtain the equilibrium strategy of the game.

[0189] This application embodiment constructs a Stackelberg game model as a multi-objective adaptive defense decision model based on dynamic game theory, realizing the application of Stackelberg game theory to power backfeed risk defense. This upgrades the decision-making process from passive response to active prediction of virtual attacker behavior and formulation of optimal response strategies, significantly improving the scientific nature and robustness of the decision.

[0190] Furthermore, after generating response strategies using the Stackelberg game model, these strategies can be implemented through cloud-edge collaborative execution and online incremental learning, forming a crucial self-optimizing closed loop. The core of this approach lies in leveraging a cloud-edge collaborative architecture to balance global optimization with local real-time response, and using an online incremental learning mechanism to enable the system to adapt to dynamic changes in the power grid.

[0191] For example, Figure 3 This is a schematic diagram of the cloud-edge collaborative execution framework provided in the embodiments of this application, as shown below. Figure 3 As shown, it includes cloud-based intelligent agents, edge intelligent agents, and end-device devices. By adopting a three-layer collaborative architecture of cloud-edge-device, reliable, low-latency execution of control commands and status feedback are achieved.

[0192] Cloud-based intelligent agent: As the global decision-making center, it is responsible for receiving the optimal control instruction set generated by the game decision-making module. It then distributes this information to the designated edge controller. Simultaneously, it receives local status feedback uploaded from the edge side, performs global performance evaluation, and logs the data.

[0193] Edge intelligent agents: Deployed in substations or critical nodes. They receive instructions from the cloud, perform local power flow calculations, voltage limit judgments, and other safety checks, and translate these into specific equipment instructions, such as adjusting transformer taps and switching capacitors and reactors. Simultaneously, they collect local real-time data and upload it to the cloud.

[0194] End-side equipment: actuators, such as adjustable transformers, smart switches, distributed energy controllers, smart loads, etc., strictly execute the instructions issued by the edge controller.

[0195] In addition, to ensure instruction security, a dual-channel verification mechanism can be adopted before instruction execution: (1) Physical layer verification: Inject the instruction to be executed into the digital twin system for rapid simulation (such as based on PSCAD or EMTDP) to predict whether the critical node voltage and branch current in the next few seconds to minutes meet the above formula. (2) Logic layer verification: Formal verification tools are used to check whether the instruction sequence violates preset security rules (such as topology connectivity protection and minimum power supply duration for critical loads). Only when both layers of verification pass, the instruction is marked as "safe and executable" and sent to the end-side device.

[0196] In some embodiments, under the cloud-edge collaborative execution framework, during online incremental learning, the cloud agent will issue a set of control instructions to the edge agent based on the power backfeed risk response strategy; obtain the actual execution effect of the strategy uploaded by the edge agent; and update the model parameters of the hybrid model and / or the model parameters of the multi-target adaptive defense decision model using an online incremental learning algorithm based on the actual execution effect of the strategy.

[0197] Among them, edge agents are deployed in substations or key nodes of the power distribution network to convert control command sets into equipment commands and send them to end-side devices for execution.

[0198] Online incremental learning mechanisms can continuously monitor the actual effects of strategy execution and update the parameters of the LSTM-Transformer model and the Stackelberg game decision model for power backfeed risk prediction to adapt to changes in power grid structure, shifts in operating modes, and evolution of the external environment. The core idea of ​​incremental learning is to gradually update model parameters using newly arriving data while maintaining existing knowledge, thus avoiding "catastrophic forgetting," where new knowledge overwrites old knowledge, leading to model performance degradation.

[0199] Taking the updating of the risk prediction LSTM-Transformer hybrid model through online incremental learning as an example, the goal is to enable the risk prediction LSTM-Transformer hybrid model to adapt to the dynamic changes in distributed power output and load demand, avoiding performance degradation caused by data distribution offset. Specifically, an EWC combined approach with online mini-batch gradient descent can be used. EWC calculates the importance weights of the parameters to prevent "catastrophic forgetting" of old knowledge during the update process. The total loss function... Includes prediction error loss and EWC regularization term:

[0200]

[0201] In the above formula, To predict error loss (optional mean squared error), it is calculated based on new data batches; The regularization strength coefficient balances the retention of both old and new knowledge; The diagonal elements of the Fisher information matrix, quantization parameters The importance of historical data; The optimal values ​​of the parameters obtained by training on historical data; This is the trainable parameter matrix for the model.

[0202] The specific online update process is as follows: (1) Data buffer: Maintain a sliding window buffer to store recent meteorological, load series and other multimodal data; (2) Trigger update: When the prediction error continues to exceed the threshold or a change in data distribution is detected, start incremental learning; (3) Parameter update: Use an adaptive learning rate optimizer to perform gradient descent.

[0203]

[0204] In the above formula, The updated parameters; These are the parameters before the update; This is the learning rate.

[0205] (4) Model Validation: Evaluate the performance of the updated model on the validation set, and deploy new parameters only if the accuracy is better than the old model. The adaptive learning mechanism is as follows: ① The learning rate is dynamically adjusted using a time decay strategy to prevent overfitting:

[0206]

[0207] In the above formula, The initial learning rate; This is the attenuation factor.

[0208] ② By constraining the changes in important parameters through EWC regularization terms, we can ensure that the model does not forget historical patterns (such as holiday load characteristics) when adapting to new data (such as seasonal load changes).

[0209] Alternatively, the payoff function weights and constraint parameters of the Stackelberg game decision model can be dynamically adjusted through online incremental learning, enabling the defense strategy to respond to changes in the power grid's operating state. (Defender payoff function) Weight adaptation:

[0210]

[0211] In the above formula, This is the weight vector of the payoff function; As a comprehensive performance indicator, it can be estimated through historical strategy execution records; The weight learning rate is estimated using the policy gradient method.

[0212] Among them, the attacker's profit function The weight vector can be dynamically updated based on statistical results from historical attack scenarios. This allows for a more accurate simulation of worst-case scenarios. Additionally, the constraint boundary parameters can be adjusted based on the frequency of voltage / current exceeding limits.

[0213]

[0214] In the above formula, For constraint parameters, such as , , Equal boundary; To constrain the degree of violation; To constrain the learning rate parameter.

[0215] Figure 4 The online incremental learning workflow diagram provided for the embodiments of this application is as follows: Figure 4 As shown, its workflow is as follows:

[0216] 1. Data Acquisition: Real-time collection of power grid status, control strategy execution effectiveness, and external environment data;

[0217] 2. Model Update Trigger: When the prediction error satisfies the following formula 34, the prediction model is triggered for update. The specific formula is as follows:

[0218]

[0219] In the above formula, For the data prediction value vector, For the data's true value vector, This is the error threshold.

[0220] 3. Incremental Learning Engine: Samples historical and latest data from the buffer, calculates gradients and regularization terms, and updates the prediction model parameters. With game model parameters , , .

[0221] 4. Incremental Learning Engine: Samples historical and latest data from the buffer, calculates gradients and regularization terms, and updates the prediction model parameters. With game model parameters , , .

[0222] This application's embodiments utilize a cloud-edge collaborative architecture to achieve collaborative optimization of multiple control devices and secure and reliable execution of strategies. Topology reconfiguration using existing distribution network line switches is one of the primary defense methods, eliminating the need for significant new hardware investment and achieving cost-effective defense. This makes the entire defense system both collaborative and economical. Furthermore, combined with an incremental learning mechanism, it can continuously learn and evolve from actual operation, possessing adaptive optimization capabilities and significantly improving the resilience of the distribution network in dealing with a high proportion of distributed energy access.

[0223] Figure 5 The schematic diagram of the power backflow risk handling device for the distribution network provided in this application is as follows: Figure 5As shown, the power backflow risk handling device 50 for power distribution networks provided in this embodiment includes:

[0224] The acquisition module 510 is used to acquire multi-source heterogeneous data and perform spatiotemporal alignment processing using a data alignment algorithm to obtain spatiotemporally aligned multimodal data. The multi-source heterogeneous data includes structured data from within the power grid and unstructured data from outside the power grid.

[0225] The prediction module 520 is used to input spatiotemporally aligned multimodal data into a hybrid model for feature extraction and dynamic attention allocation, and to generate output and load demand predictions for distributed power sources connected to the distribution network in future time periods.

[0226] Module 530 is used to determine the risk assessment indicators for power backflow from distributed generation sources to the distribution network based on output forecasts and load demand forecasts.

[0227] The generation module 540 is used to trigger a multi-objective adaptive defense decision model based on dynamic game theory to generate a power backflow risk response strategy for the distribution network if the power backflow risk assessment index is greater than the preset risk threshold.

[0228] In one possible implementation, the internal structured data of the power grid includes historical and real-time voltage data, historical and real-time current data, and historical and real-time power data; the external unstructured data of the power grid includes gridded weather forecast data, holiday information, and social event information.

[0229] In one possible implementation, the hybrid model includes a Long Short-Term Memory network and a Transformer encoder; correspondingly, the prediction module can be used for:

[0230] Temporal features are extracted from multimodal data using a long short-term memory network to obtain local temporal features, and these local temporal features are then encoded at their locations to obtain an encoded feature sequence.

[0231] The Transformer encoder uses a multi-head attention mechanism to assign attention weights to the feature sequence in order to determine the spatiotemporal features most correlated with power backfeed, and based on the spatiotemporal features, to determine the output forecast and load demand forecast.

[0232] In one possible implementation, a module is defined, which can specifically be used for:

[0233] Based on the predicted output and load demand, the node voltage and branch power flow distribution of the distribution network in the future time section are obtained by power flow calculation using the power grid topology parameters.

[0234] Based on node voltage and branch power flow distribution, the geometric safe distance between the operating state of the distribution network at future time sections and the boundary of the safe domain is determined.

[0235] Obtain the node voltage over-limit risk index and the branch power flow over-limit risk index, and then fuse them to obtain the comprehensive risk index;

[0236] Based on geometric safety distance and / or comprehensive risk index, generate power backfeed risk assessment indicators.

[0237] In one possible implementation, the safety domain boundary includes at least one of the following: line capacity constraint, voltage limit constraint, main transformer capacity constraint, feeder and main transformer capacity matching constraint, distributed power source access constraint, and anti-power backfeed constraint.

[0238] In one possible implementation, a decision model building module is also included, for:

[0239] Construct a defender, a defender strategy set, an attacker, and an attacker strategy set; wherein, the defender represents the power grid control system of the distribution network, the defender strategy set includes at least one set of coordinated control action combinations, the coordinated control combinations are used to adjust the operating state of the distribution network to eliminate the risk of power backfeed, the attacker represents the uncertainty factors of power backfeed, and the attacker strategy set is used to simulate the most unfavorable uncertainty scenario that leads to power backfeed.

[0240] Based on the system network loss cost, voltage deviation penalty, control equipment action cost, voltage exceeding the upper limit penalty, and branch current exceeding the limit penalty, a defender's revenue function is constructed.

[0241] Based on the degree of voltage exceeding the upper limit, the degree of branch current exceeding the upper limit, the degree of distributed power output exceeding the upper limit, and the degree of power backflow, an attacker's profit function is constructed;

[0242] The constraints are constructed, including power balance constraints, voltage safety constraints, branch capacity constraints, equipment operation constraints, equipment operation limit, power backfeed constraints, and topology connectivity constraints.

[0243] Based on the defender, the defender's policy set, the defender's payoff function, the attacker, the attacker's policy set, the attacker's payoff function, and constraints, a multi-objective adaptive defense decision model for dynamic game theory is constructed.

[0244] In one possible implementation, the generation module is specifically used to: solve for the equilibrium point of the multi-objective adaptive defense decision model using a mixed-integer linear programming method; and generate a power backflow risk response strategy for the distribution network based on the equilibrium point.

[0245] The equilibrium point is used to characterize the defender's ability to anticipate the attacker's optimal response to any strategy in the defender's strategy set.

[0246] In one possible implementation, the defender's strategy set includes the on-load tap changer tap position vector, the capacitor reactor switching state vector, the distributed generation active power output adjustment vector, and the flexible load adjustment vector; the attacker's strategy set includes the distributed generation output disturbance vector and the load demand disturbance vector.

[0247] In one possible implementation, an update module is also included, for:

[0248] Based on the power backfeed risk response strategy, control command sets are sent to edge agents. The edge agents are deployed in substations or key nodes of the distribution network to convert the control command sets into equipment commands and send them to the end-side equipment for execution.

[0249] Obtain the actual execution effect of the strategy uploaded by the edge agent;

[0250] Based on the actual implementation effect of the strategy, an online incremental learning algorithm is used to update the model parameters of the hybrid model and / or the model parameters of the multi-target adaptive defense decision model.

[0251] The power backflow risk handling device for distribution networks provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0252] Figure 6 A schematic diagram of the structure of the electronic device provided in this application. Figure 6 As shown, the electronic device 60 provided in this embodiment includes at least one processor 601 and a memory 602. Optionally, the electronic device 60 further includes a communication component 603. The processor 601, memory 602, and communication component 603 are connected via a bus.

[0253] In a specific implementation, at least one processor 601 executes computer execution instructions stored in memory 602, causing at least one processor 601 to perform the above-described method.

[0254] The specific implementation process of processor 601 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0255] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0256] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0257] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0258] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0259] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0260] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0261] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for handling the risk of power backflow in a distribution network, characterized in that, include: Multi-source heterogeneous data is acquired and spatiotemporally aligned using a data alignment algorithm to obtain spatiotemporally aligned multimodal data; wherein, the multi-source heterogeneous data includes structured data inside the power grid and unstructured data outside the power grid; The spatiotemporally aligned multimodal data is input into the hybrid model for feature extraction and dynamic attention allocation to generate output and load demand forecasts for distributed power sources connected to the distribution network in future time periods. Based on the predicted output and load demand, the risk assessment index for the power backflow of the distributed power source to the distribution network is determined. If the power backfeed risk assessment index is greater than the preset risk threshold, a multi-objective adaptive defense decision model based on dynamic game theory is triggered to generate a power backfeed risk response strategy for the distribution network.

2. The method according to claim 1, characterized in that, The internal structured data of the power grid includes historical and real-time voltage data, historical and real-time current data, and historical and real-time power data; the external unstructured data of the power grid includes gridded weather forecast data, holiday information, and social event information.

3. The method according to claim 1, characterized in that, The hybrid model includes a long short-term memory network and a Transformer encoder; The step of inputting spatiotemporally aligned multimodal data into a hybrid model for feature extraction and dynamic attention allocation includes: Temporal features are extracted from the multimodal data using a long short-term memory network to obtain local temporal features, and the local temporal features are then positionally encoded to obtain an encoded feature sequence. The Transformer encoder uses a multi-head attention mechanism to assign attention weights to the feature sequence in order to determine the spatiotemporal features most correlated with power backfeed, and based on the spatiotemporal features, to determine the output prediction value and the load demand prediction value.

4. The method according to claim 3, characterized in that, The determination of the power backfeed risk assessment indicators of the distributed power source to the distribution network includes: Based on the predicted output and load demand, the node voltage and branch power flow distribution of the distribution network in the future time section are obtained by power flow calculation using the power grid topology parameters. Based on the node voltage and branch power flow distribution, the geometric safe distance between the operating state of the distribution network and the boundary of the safe domain in the future time section is determined. Obtain the node voltage over-limit risk index and the branch power flow over-limit risk index, and then fuse them to obtain the comprehensive risk index; The power backfeed risk assessment index is generated based on the geometric safety distance and / or the comprehensive risk index.

5. The method according to claim 4, characterized in that, The safety domain boundary includes at least one of the following: line capacity constraint, voltage limit constraint, main transformer capacity constraint, feeder and main transformer capacity matching constraint, distributed power source access constraint, and anti-power backfeed constraint.

6. The method according to claim 1, characterized in that, The method further includes: Construct a defender, a defender strategy set, an attacker, and an attacker strategy set; wherein, the defender represents the power grid control system of the distribution network, the defender strategy set includes at least one set of coordinated control action combinations, the coordinated control combinations are used to adjust the operating state of the distribution network to eliminate the risk of power backfeed, the attacker represents the uncertainty factor of power backfeed, and the attacker strategy set is used to simulate the most unfavorable uncertainty scenario that leads to power backfeed; Based on the system network loss cost, voltage deviation penalty, control equipment action cost, voltage exceeding the upper limit penalty, and branch current exceeding the limit penalty, a defender's revenue function is constructed. Based on the degree of voltage exceeding the upper limit, the degree of branch current exceeding the upper limit, the degree of distributed power output exceeding the upper limit, and the degree of power backflow, an attacker's profit function is constructed; The constraints are constructed, including power balance constraints, voltage safety constraints, branch capacity constraints, equipment operation constraints, equipment operation limit, power backfeed constraints, and topology connectivity constraints. Based on the defender, defender's strategy set, defender's payoff function, attacker, attacker's strategy set, attacker's payoff function, and constraints, a multi-objective adaptive defense decision model for the dynamic game is constructed.

7. The method according to claim 6, characterized in that, The triggering of the multi-objective adaptive defense decision model based on dynamic game theory generates a power backflow risk response strategy for the distribution network, including: The equilibrium point of the multi-objective adaptive defense decision model is solved by a mixed-integer linear programming method. The equilibrium point is used to characterize the defender's ability to predict the attacker's optimal response to any strategy in the defender's strategy set. Based on the equilibrium point, a power backflow risk mitigation strategy for the distribution network is generated.

8. The method according to claim 6, characterized in that, The defender's strategy set includes the on-load tap changer tap position vector, the capacitor and reactor switching state vector, the distributed generation active power output adjustment vector, and the flexible load adjustment vector; the attacker's strategy set includes the distributed generation output disturbance vector and the load demand disturbance vector.

9. The method according to claim 1, characterized in that, The method further includes: Based on the power backfeed risk response strategy, a set of control instructions is sent to the edge agent. The edge agent is deployed in the substation or key node of the power distribution network to convert the set of control instructions into equipment instructions and send them to the end-side equipment for execution. Obtain the actual execution effect of the strategy uploaded by the edge agent; Based on the actual implementation effect of the strategy, the model parameters of the hybrid model and / or the model parameters of the multi-target adaptive defense decision model are updated using an online incremental learning algorithm.

10. A power backfeed risk handling device for power distribution networks, characterized in that, include: The acquisition module is used to acquire multi-source heterogeneous data and perform spatiotemporal alignment processing through a data alignment algorithm to obtain spatiotemporally aligned multimodal data; wherein, the multi-source heterogeneous data includes structured data inside the power grid and unstructured data outside the power grid; The prediction module is used to input spatiotemporally aligned multimodal data into a hybrid model for feature extraction and dynamic attention allocation, generating output and load demand predictions for distributed power sources connected to the distribution network in future time periods. The determination module is used to determine the power backflow risk assessment index of the distributed power source to the distribution network based on the power output forecast value and the load demand forecast value. The generation module is used to trigger a multi-objective adaptive defense decision model based on dynamic game theory to generate a power backflow risk response strategy for the distribution network if the power backflow risk assessment index is greater than a preset risk threshold.

Images