A method for power distribution network dynamic state estimation and risk management collaboration

Abstract

The application discloses a power distribution network dynamic state estimation and risk management collaborative method, comprising the following steps: obtaining power distribution network operation state data, performing power flow calculation to obtain node power data sequence, inputting a prediction model to obtain node power data at a prediction time, and then combining actual measurement data to obtain power distribution network operation state estimation results; subsequently, based on the dynamic state estimation results, typical scenarios are constructed, and corresponding to each scenario, net load prediction data is obtained through a load prediction model, and then a stochastic optimization model is used to obtain operation adjustment strategies of flexible equipment; subsequently, operation decision results are fed back to the dynamic state estimation link, and the state prediction and state estimation process at a subsequent time is adjusted and corrected. The application effectively improves the precision of power distribution network dynamic state estimation and the reliability of risk response optimization, and is suitable for power distribution network real-time operation state sensing and risk management.

Description

Technical Field

[0001] This invention relates to the field of power system state perception and operation optimization technology, and in particular to a method for coordinating dynamic state estimation and risk management of distribution networks. Background Technology

[0002] The power distribution network plays a crucial role as the final link in power transmission, and the development and utilization of distributed renewable energy within it can effectively contribute to achieving carbon peaking and integrated carbon targets. However, with the increasing penetration rate of renewable energy, the randomness of power sources such as wind and solar power increases the uncertainty of distribution network operation, leading to more frequent and severe fluctuations in the network's operating status. Furthermore, the continuous growth in user electricity demand has resulted in insufficient transmission capacity in many distribution networks, causing their operating status to constantly approach or exceed safe limits. Against the backdrop of intensified fluctuations and increased operational risks in distribution networks, the importance and difficulty of accurately sensing the network's operating status and effectively managing operational risks are growing.

[0003] Despite the continuous development of distribution network condition estimation and risk management technologies, existing methods still have significant shortcomings:

[0004] On the one hand, traditional dynamic state estimation methods rely heavily on accurate system state transition formulas. However, the dynamic process of distribution networks is affected by multiple factors such as fluctuations in renewable energy output, random changes in load, and operation control strategies, making it difficult to accurately characterize using formulas. The linear extrapolation method, which is often used, has a large error and can no longer meet the accuracy requirements of complex distribution network scenarios.

[0005] On the other hand, existing risk response and optimization decision-making methods often do not match the dynamic state estimation in terms of time scale, making it difficult to reflect rapid changes in the system state in a timely manner, thus rendering risk response measures ineffective in addressing current risks. Therefore, how to establish an effective collaborative method for dynamic state estimation and risk management of distribution networks is an important problem that needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a method for coordinating dynamic state estimation and risk management of distribution networks based on deep learning and stochastic optimization, which can effectively improve the accuracy of dynamic state estimation and the reliability of risk response strategies.

[0007] The technical solution adopted in this invention is as follows.

[0008] On the one hand, the present invention provides a method for coordinating dynamic state estimation and risk management of distribution networks, comprising:

[0009] S1, acquire the distribution network operation status data of the previous moment, perform power flow calculation based on the distribution network operation status data, and obtain the node power data sequence;

[0010] S2, input the node power data sequence into the pre-trained load prediction model to obtain the node power data sequence for the prediction period including the current prediction time;

[0011] S3, based on the node power data at the current prediction time, perform distribution network operation state estimation to obtain the state prediction quantity at the current prediction time;

[0012] S4. Based on the actual measurement data at the current prediction time, the state prediction is corrected to obtain the state estimation result at the current prediction time.

[0013] S5. Based on the state estimation results, construct multiple typical scenarios of the distribution network operation state at the current predicted time;

[0014] S6. Based on multiple typical scenarios of the state, determine the corresponding node power scenario at the current prediction time, and form a node power data sequence with historical node power data. Through the load prediction model, obtain the node power data sequence for the prediction period corresponding to different typical scenarios and including the next prediction time.

[0015] S7. Based on the node power data at the next prediction time, a stochastic optimization model aimed at minimizing node voltage and branch overload risk is used to calculate the control strategy of distribution network equipment.

[0016] S8, at the next prediction time, the state prediction quantity is corrected based on the actual measurement data at that time and the control strategy of the distribution network equipment to obtain the state estimation result; then proceed to step S5 and repeat steps S5 to S8 to perform state estimation and distribution network equipment control strategy calculation for subsequent time periods.

[0017] Optionally, in the load forecasting model, each sequence transformation module includes:

[0018] Embedding layers are used to project one-dimensional time-series data of the input node power into a high-dimensional feature space;

[0019] The reshaping layer is used to reshape the projected one-dimensional time series data into multiple typical two-dimensional series data;

[0020] A feature extraction layer is used to extract periodic feature data of power data from the multiple typical two-dimensional sequence data using two-dimensional convolutional kernels. The periodic features include intra-period variation features and inter-period variation features.

[0021] The aggregation layer is used to convert the periodic feature data into one-dimensional time series data and perform weighted fusion to obtain the one-dimensional time series data of node power output by the current sequence transformation module.

[0022] In the above scheme, the number of sequence transformation modules is determined by the complexity of the sequence, usually 2 to 3, to avoid excessively large modules that would lead to increased computational complexity and overfitting.

[0023] Optionally, the reshaping layer reshapes the projected one-dimensional time series data into multiple typical two-dimensional sequence data, including:

[0024] Perform a Fast Fourier Transform on the projected one-dimensional time series data, and calculate the average amplitude based on the frequency obtained from the transform. , is represented as: ,in For the original one-dimensional node power data sequence One-dimensional time series data after projection. The length of the original one-dimensional time series data. The original feature dimension can include the active and reactive power of multiple nodes. The projected feature dimension determines the ability of a time series to be represented in a high-dimensional feature space, and depends on the input dimension and system complexity. Indicates the amplitude value. Represents the Fast Fourier Transform. This means that for each frequency, take its... The average magnitude of each feature dimension;

[0025] Select based on the average amplitude of each frequency The typical characteristic frequencies are expressed as: ,in For the first A typical characteristic frequency;

[0026] Based on the period length corresponding to the selected typical characteristic frequencies, the sequence is padded with zeros at the edges corresponding to each frequency to transform the sequence. Remodeling Two-dimensional data , Typical feature frequency dimension The corresponding period length, and we have: .

[0027] As one implementation method, the step of selecting based on the average amplitude of each frequency... The typical characteristic frequencies are as follows: Each frequency is sorted by its average amplitude, and the top... The frequency is used as the A typical characteristic frequency.

[0028] Optionally, the aggregation layer converts the periodic feature data into one-dimensional time series data and performs weighted fusion, including:

[0029] The output of the feature extraction layer Periodic feature data corresponding to a typical characteristic frequency They were converted into one-dimensional time series data respectively. And truncated to the length of the original one-dimensional time series data;

[0030] Based on the Fourier transform results of the one-dimensional time series data received by the current sequence transformation module's embedded layer, the amplitude corresponding to each frequency is normalized using the Softmax function, and expressed as follows:

[0031] ,in; To address the current issue The Fourier transform is performed on the one-dimensional time series data input to the previous sequence transformation module, resulting in the corresponding... The amplitude of a typical characteristic frequency, for The normalization result;

[0032] Based on the normalization result, the normalized amplitude of each frequency is used as a weighting coefficient. One-dimensional time series data Perform a weighted summation to obtain the aggregated result of the current sequence transformation module, which is represented as:

[0033] ,in For the current number The aggregation layer of the sequence transformation module is for the first... One-dimensional time series data is obtained by transforming the periodic feature data corresponding to the typical feature frequencies. This is the weighted fusion result of the aggregation layer;

[0034] The aggregation result is used as input to the next sequence transformation module, or if the current sequence transformation module is the last sequence transformation module, the aggregation result is output as the prediction result of the load prediction model.

[0035] Optionally, in step S1, power flow calculation is performed based on the distribution network operating status data of the previous moment to obtain the node power data sequence, including:

[0036] Based on the operating state value of the previous moment before the prediction time, the power data of each node is calculated using the network power flow equation. The calculation results are combined with the historical power data of each node with a preset step size to form a node power data sequence containing the current prediction time.

[0037] The formula for calculating the node power data of the time preceding the prediction time is expressed as: ,

[0038] In the formula, for Time Node Power data, and These are the active power and reactive power, respectively. Represents the network power flow equation. For the predicted time The moment before The running status value, for any given time. Operating status The network power flow equation is expressed as:

[0039]

[0040] in, For nodes A set of connected nodes. and They are respectively Time Node The node voltages and their phase angles, and These are the real and imaginary parts of the nodal admittance matrix elements, respectively. and They represent Time Node Net load The active power and reactive power in the process.

[0041] In the above scheme, the operating state value at the previous moment This can represent the initial operating state of the distribution network. The distribution network operating status obtained when initially using the method of this invention to intervene in the distribution network for state estimation and risk management collaborative control.

[0042] Optionally, based on the node power data at the current prediction time, the distribution network operating state is estimated to obtain the state prediction at the current prediction time, including:

[0043] Based on predicted time The node power data is used to calculate the state prediction at the prediction time using the network power flow equation, and is expressed as:

[0044]

[0045] in, The predicted time is calculated from the power flow. State prediction quantity, for Noise in the time-based prediction process for The inverse function; For prediction time The moment before The node power data sequence is used in the load forecasting model. The predicted time The node power data is represented as follows:

[0046] ,

[0047] in, To preset the historical step size, Nodes as input to the load forecasting model exist Net load data at any given time.

[0048] Optionally, in step S4, the state prediction is corrected based on the actual measurement data at the current prediction time to obtain the state estimation result at the current prediction time, including:

[0049] S41, Calculate the predicted state covariance matrix, expressed as:

[0050] ,

[0051] in, To correspond to the current prediction time The predicted state covariance matrix, for The initial state covariance matrix at time 1: If the prediction time is the first prediction, the initial state covariance matrix can be the identity matrix, an empirical value, or determined based on prior knowledge. and For state-based Calculate the state Jacobian matrix and its transpose; For state-based prediction quantities The calculated state Jacobian matrix, where the superscripts -1 and T denote taking the inverse and transpose of the matrix, respectively; for The noise covariance matrix of the prediction process at time step;

[0052] S42, calculate the Kalman gain, expressed as:

[0053] ,

[0054] In the formula, for Kalman gain at time step To measure the noise covariance matrix, To measure the Jacobian matrix, it is a measurement function. Regarding state variables The first-order partial derivative matrix describes how small changes in the state variables affect the measured values. Indicates in Seeking about The partial derivatives, i.e., in Predicting quantities based on state at any time Find the partial derivatives;

[0055] S43, Correct the state prediction based on the actual measurement data to obtain the state estimation result at the current prediction time. and the covariance matrix used to predict the state at the next time step Calculate the state covariance matrix , is represented as:

[0056] ,

[0057]

[0058] In the formula, for The measurement vector at time t, and its relationship with the measurement function The relationship is represented as: ,in for Measurement noise at any given moment;

[0059] For any time Measurement function Represented as: ,

[0060] State prediction Measurement function for variables It can be represented as: ,in and These are node power data predicted based on load forecasting models. , ,through The calculated predicted values ​​of node voltage and phase angle.

[0061] In Kalman filtering, The state covariance matrix Mapped to the measurement space in order to calculate the Kalman gain. This leads to updates in the state estimate and covariance matrix. (Covariance matrix) The uncertainty of the current state estimate is quantified. When constructing typical scenarios of distribution network operation states based on the state estimate results, the covariance matrix... This method guides the sampling process when generating process noise samples, thereby obtaining multiple sets of state scenarios. These can then be reduced using the k-means clustering algorithm to finally construct typical scenarios. The specific process of generating typical scenarios based on state estimation results and state covariance matrix through state sampling and k-means clustering can be found in existing technologies and is not an innovation of this application, so it will not be elaborated upon.

[0062] After obtaining multiple typical scenarios, the node power data at the current prediction time corresponding to different scenarios is obtained through power flow calculation. These are then combined with historical node power data to form node power data sequences, which are input into the load prediction model to obtain node power data sequences for prediction periods that include the next prediction time for different typical scenarios.

[0063] Optionally, in step S7, based on the node power data at the next prediction time, a stochastic optimization model aimed at minimizing node voltage and branch overload risk is used to calculate the distribution network equipment control strategy at the next prediction time, including:

[0064] S71, based on the next prediction time The node power data is obtained through the inverse function of the network power flow equation. Calculate the corresponding state prediction quantities;

[0065] S72, based on node power data and state predictions, uses a stochastic optimization model aimed at minimizing node voltage and branch overload risk to calculate the prediction time. Distribution network equipment control strategies;

[0066] Among them, for the latest current prediction time The objective function of the stochastic optimization model is expressed as:

[0067]

[0068] In the formula, Representing a scene The probability of occurrence Number of future moments considered for risk management This represents the total number of typical scenarios generated based on the state estimation results. and These represent the total number of distribution network nodes and the total number of branch lines, respectively. and These are the high-voltage risk indicators and their weights. and These are the low-voltage risk indicators and their weights. and These are the positive heavy load index for branch roads and its weight, respectively. respectively For the reverse overload index and its index weights;

[0069] Each risk indicator is expressed as follows:

[0070] ,

[0071] ,

[0072] ,

[0073] ;

[0074] in, and These are the upper and lower limits of the node voltage, which can be set to 1.05 times and 0.95 times the reference voltage, respectively. Let be the power flow of the j-th branch at time t. This is the capacity of the branch.

[0075] Regarding the selection of weights for each indicator, the analytic hierarchy process (AHP) and the entropy weight method can be used to calculate subjective and objective weights, respectively, and then combined using coefficients. The final weight is obtained by merging the two types of weights, and is represented as follows:

[0076] ,in Subjective weighting, For objective weighting.

[0077] The risk management and control strategy optimization of this invention can consider various flexible devices for actual distribution networks. For distribution networks that consider energy storage as a flexible device, optionally, the constraints of the stochastic optimization model optimization solution include power flow constraints, upper and lower limits of energy storage charging and discharging power constraints, energy storage state of charge constraints, and predetermined scheduling plan constraints, which are expressed as follows:

[0078] Current constraints: ,

[0079] Upper and lower limits of energy storage charging and discharging power constraints:

[0080] ,

[0081] Energy storage state of charge constraints: ,

[0082] Given scheduling constraints:

[0083] ,

[0084] In the formula, and They are nodes With nodes The active and reactive power between; For nodes A set of connected nodes. For nodes The parent node; and They are nodes With nodes The active and reactive power between; and They are nodes The charging and discharging power of the energy storage device. and They are nodes The injected active and reactive power, and They are nodes With nodes The resistance and reactance between them The reference voltage;

[0085] For the first Energy storage The charging and discharging state at any given moment. and The first Energy storage The power of constant charging and discharging. For energy storage collection;

[0086] For the first Energy storage State of charge at time t, and The first The state-of-charge limit and upper limit of an energy storage system. The time window under consideration and The first The charging and discharging efficiency of the energy storage device;

[0087] As of Time of the first The amount of electricity already consumed by the energy storage The power target corresponding to the established scheduling plan.

[0088] After the above optimization, the optimized results of the energy storage charging and discharging operation strategy in the distribution network can be obtained, including the charging and discharging power of energy storage at each node. These optimization results can be used to guide distribution network scheduling optimization and, in conjunction with actual measurement data, to correct the state estimates obtained from network power flow calculations.

[0089] Optionally, in step S8, at the next prediction time, the predicted state quantity is corrected based on the actual measurement data at that time and the control strategy of the distribution network equipment to obtain the state estimation result, including:

[0090] S81, superimpose the results of the power distribution network equipment control strategy onto the node power data at the predicted time output by the load forecasting model;

[0091] S82, calculates network power flow based on the superimposed node power data to obtain the state prediction quantity;

[0092] S83. By executing the operation of step S4, the state estimation results are obtained, which are used to generate typical scenarios of new distribution network operation states, and then based on these typical scenarios, node power prediction, state estimation and risk management stochastic optimization are performed for subsequent time periods.

[0093] Beneficial effects

[0094] This invention addresses the problem that traditional risk response optimization often fails to match the real-time operational state estimation on a time scale. Based on dynamic state estimation results, it constructs typical scenarios and uses this to obtain node net load sequences through power flow calculations and historical node power data. These sequences are then fed into an established load forecasting model to obtain net load forecast data. Subsequently, stochastic optimization is employed to formulate real-time operational adjustment strategies. The operational decision results are then fed back to the dynamic state estimation stage to adjust and correct the state prediction and estimation processes for subsequent time periods. This effectively improves the accuracy of dynamic state estimation and the reliability of risk response optimization in distribution networks, making it suitable for real-time operational state awareness and risk management in distribution networks.

[0095] Meanwhile, the load forecasting model of this invention reshapes the one-dimensional time series of node loads into two-dimensional data through Fast Fourier Transform, and then uses two-dimensional convolutional kernels to simultaneously extract intra- and inter-period time series features, enabling high-precision node net load forecasting. Furthermore, addressing the problem that traditional extended Kalman filter forecasting relies on linear extrapolation and the accuracy of estimation results is limited by the precision of the forecasting step, this invention uses the aforementioned net load forecasting method to replace the forecasting stage of traditional dynamic state estimation, thereby obtaining more accurate and reliable distribution network state estimation results. Attached Figure Description

[0096] Figure 1The diagram shown is a flowchart of the method for coordinating dynamic state estimation and risk management of power distribution networks according to the present invention in an embodiment. Detailed Implementation

[0097] This invention aims to provide a method for the coordinated optimization of dynamic state estimation and risk management of distribution networks based on deep learning and stochastic optimization. The inventive concept specifically includes the following aspects:

[0098] I. Constructing a load forecasting model based on deep learning: Acquire historical and real-time measurement data of the power system, reshape the one-dimensional time series of node net load into two-dimensional data through fast Fourier transform, and use two-dimensional convolution kernels in the sequence transformation module to extract the time series features within and between periods, thereby achieving high-precision node net load forecasting.

[0099] II. Dynamic State Estimation of Power System: The prediction stage of the extended Kalman filter algorithm is optimized. The deep learning load prediction model is used to obtain the node net load prediction data. The system state prediction value is obtained from the node net load prediction value through power flow calculation. The state prediction value is corrected by the real-time measurement data imported by the filtering stage to obtain a more accurate system state estimate.

[0100] III. Power System Risk Management: Based on the dynamic state estimation results, typical scenarios are constructed through state sampling and k-means clustering algorithm. Combined with historical data, these are expanded into typical scenarios with multi-time state sequences. Based on this, the node net load sequence is obtained through power flow calculation, and the net load prediction data is obtained by inputting it into the established load prediction model. Subsequently, stochastic optimization is used to formulate real-time operation adjustment strategies. Then, the operation decision results are fed back to the dynamic state estimation model to adjust and correct the state prediction and estimation process for subsequent time periods.

[0101] The following description, in conjunction with the accompanying drawings and specific embodiments, provides further details.

[0102] Example 1

[0103] This embodiment provides a method for coordinating dynamic state estimation and risk management of distribution networks. The implementation process can be found in the following reference. Figure 1 ,include:

[0104] S1, acquire the distribution network operation status data of the previous moment or the initial moment, perform power flow calculation based on the distribution network operation status data, and obtain the node power data sequence;

[0105] S2, input the node power data sequence into the pre-trained load prediction model to obtain the node power data sequence for the prediction period including the current prediction time;

[0106] S3, based on the node power data at the current prediction time, perform distribution network operation state estimation to obtain the state prediction quantity at the current prediction time;

[0107] S4. Based on the actual measurement data at the current prediction time, the state prediction is corrected to obtain the state estimation result at the current prediction time.

[0108] S5. Based on the state estimation results, construct multiple typical scenarios of the distribution network operation state at the current predicted time;

[0109] S6. Based on multiple typical scenarios of the state, determine the corresponding node power scenario at the current prediction time, and form a node power data sequence with historical node power data. Through the load prediction model, obtain the node power data sequence for the prediction period corresponding to different typical scenarios and including the next prediction time.

[0110] S7. Based on the node power data at the next prediction time, a stochastic optimization model aimed at minimizing node voltage and branch overload risk is used to calculate the control strategy of distribution network equipment.

[0111] S8, at the next prediction time, the state prediction quantity is corrected based on the actual measurement data at that time and the control strategy of the distribution network equipment to obtain the state estimation result; then proceed to step S5 and repeat steps S5 to S8 to perform state estimation and distribution network equipment control strategy calculation for subsequent time periods.

[0112] In this embodiment, the predicted operating state value at the previous moment can be the initial operating state of the distribution network, or the operating state of the distribution network measured or obtained by other means when the method of the present invention is initially used to intervene in the distribution network for state estimation and risk management collaborative control.

[0113] The above method, after step S5, forms the cyclical implementation process of the subsequent power distribution network state dynamic estimation and risk management system. It not only obtains relatively accurate node net load prediction results through the load forecasting model, but also optimizes the real-time operation strategy of flexible equipment based on multiple scenarios based on the prediction results. It corrects the state prediction quantity obtained based on the load forecasting results through the optimization results and measurement data, and further generates multiple state scenarios for subsequent state dynamic estimation and stochastic optimization. It can effectively address the problem that traditional risk response optimization methods do not match the state estimation in real-time operation on the time scale, and effectively improve the accuracy of power distribution network dynamic state estimation and the reliability of risk response optimization.

[0114] Example 2

[0115] Based on Example 1, this example specifically introduces a method for coordinating dynamic state estimation and risk management of distribution networks. Its implementation process includes the following aspects.

[0116] I. Construction of Load Forecasting Model

[0117] 1.1 Logical Architecture of Load Forecasting Model

[0118] The load forecasting model constructed in this embodiment includes multiple stacked sequence transformation modules. The time series of the input node power data will be processed layer by layer through each sequence transformation module. Each sequence transformation module includes:

[0119] Embedding layer, used to embed the original one-dimensional time series of input node power. Projected into a high-dimensional feature space ,in The length of the original one-dimensional time series data. This is the original feature dimension, which can include the net load and active power injection of multiple nodes. The projected feature dimension determines the ability of a time series to be represented in a high-dimensional feature space, and depends on the input dimension and system complexity.

[0120] Reshaping layer, used to project one-dimensional time series data Remodeling A typical two-dimensional sequence data Reshape into i two-dimensional data ;

[0121] The feature extraction layer - Inception module - is used to extract periodic feature data of power data from multiple typical two-dimensional sequence data using two-dimensional convolutional kernels. The periodic features include intra-period variation features and inter-period variation features.

[0122] The aggregation layer is used to transform the two-dimensional periodic feature data extracted from the two-dimensional sequence data by the Inception module back into one-dimensional time series data, and then perform weighted fusion.

[0123] 1.2 Load Forecasting Implementation Process of the Load Forecasting Model

[0124] 1.2.1 In the embedding layer, the original one-dimensional time series of the input node power is... Projected into a high-dimensional feature space ,get .

[0125] 1.2.2 In the reshaping layer, a Fast Fourier Transform is performed on the projected one-dimensional time series data, and the average amplitude is calculated based on the frequency obtained from the transform. , is represented as: ,in For the original one-dimensional node power data sequence One-dimensional time series data after projection. This represents the average of the magnitudes of multiple feature dimensions at each frequency after the sequence has been decomposed by FFT. Indicates the amplitude value. Represents the Fast Fourier Transform. The length of the original one-dimensional time series data. This is the original feature dimension, which can include the net load and reactive power of multiple nodes. The projected feature dimension determines the ability of a time series to be represented in a high-dimensional feature space, and depends on the input dimension and system complexity.

[0126] Select based on the average amplitude of each frequency The typical characteristic frequencies are expressed as: ,in For the first A typical characteristic frequency;

[0127] In some possible implementations, when selecting typical characteristic frequencies, the frequencies can be sorted by their average amplitude, and the top frequencies can be ranked accordingly. The frequency is used as the A typical characteristic frequency;

[0128] Based on the period length corresponding to the selected typical characteristic frequencies, the sequence is padded with zeros at the edges corresponding to each frequency to transform the sequence. Remodeling Two-dimensional data , Typical feature frequency dimension The corresponding period length, and we have: .

[0129] 1.2.3 Simultaneously extract the aforementioned data using the two-dimensional convolutional kernel of the Inception module. The periodic and periodic characteristics of two-dimensional data are used to obtain a two-dimensional output, which is then input into the aggregation layer.

[0130] 1.2.4 In the aggregation layer, the output of the feature extraction layer is... Periodic feature data corresponding to a typical characteristic frequency They were converted into one-dimensional time series data respectively. And truncated to the length of the original one-dimensional time series data;

[0131] Based on the Fourier transform results of the one-dimensional time series data received by the current sequence transformation module's embedded layer, the amplitude corresponding to each frequency is normalized using the Softmax function, and expressed as follows:

[0132] ,in; To address the current issue The Fourier transform is performed on the one-dimensional time series data input to the previous sequence transformation module, resulting in the corresponding... The amplitude of a typical characteristic frequency, for The normalization result;

[0133] Based on the normalization result, the normalized amplitude of each frequency is used as a weighting coefficient. One-dimensional time series data Perform a weighted summation to obtain the aggregated result of the current sequence transformation module, which is represented as:

[0134] ,in For the current number The aggregation layer of the sequence transformation module is for the first... One-dimensional time series data is obtained by transforming the periodic feature data corresponding to the typical feature frequencies. This is the weighted fusion result of the aggregation layer;

[0135] The aggregation result is used as input to the next sequence transformation module, or if the current sequence transformation module is the last sequence transformation module, the aggregation result is output as the prediction result of the load prediction model.

[0136] like The total number of sequence transformation modules in the load forecasting model; the two-dimensional output of the Inception module in the aggregation layer of the last sequence transformation module. First, it is converted into a one-dimensional time series. And truncated to the initial sequence length T. Then, using the Softmax function, based on the first... Input of each sequence transformation module The Fourier transform is used to calculate the amplitude, which is then normalized, and then the normalized amplitude is used as the basis for the normalization. right Perform a weighted summation to obtain the first result. The output of each sequence transformation module is a one-dimensional time series. As the output of the load forecasting model, it includes node power data at the current forecast time.

[0137] II. Dynamic Estimation of Distribution Network Status Based on Load Forecasting Results

[0138] In summary, this embodiment utilizes the multi-cycle learning capability of a deep learning-based load forecasting model to improve and extend the Kalman filter state estimation algorithm.

[0139] 2.1 Definition of State Vector and Measurement Function

[0140] In this embodiment, the distribution network operation state vector is defined as follows: , express Time Node Voltage amplitude at that point express Time Node The phase angle at the point. Measurement vector at time The relationship between the state vector and the state vector is expressed by the following formula:

[0141]

[0142] in For measurement functions, For measuring noise.

[0143] Measurement function The relationship between node injected power and node voltage phase angle is expressed as:

[0144]

[0145] The network power flow equation is expressed as:

[0146]

[0147] in, and They represent Time of the first Net load of each node Active power and reactive power in and Let be the real and imaginary parts of the nodal admittance matrix elements. For nodes A set of connected nodes.

[0148] 2.2 Dynamic estimation process of distribution network operation status based on load forecasting method

[0149] 2.2.1 Based on initial state or the system state at the previous moment Using network power flow equations The net load input sequence for each node is calculated and expressed as follows: ,in, for Time Node Net load, and These are the active power and reactive power, respectively. For the predicted time The moment before The running status value;

[0150] 2.2.2 Combining the current time and preset... Historical node net load data at each step size is used to obtain the time series of node net load data. The net load forecasting result is then generated using a load forecasting model, and is represented as follows:

[0151] ,

[0152] in, Nodes as input to the load forecasting model exist Net load data at any given time.

[0153] 2.2.3 Based on the node net load forecast, the state prediction is obtained through power flow calculation, and is expressed as:

[0154] ,

[0155] in, The predicted time is calculated from the power flow. State prediction quantity, for Noise in the time-based prediction process for The inverse function;

[0156] 2.2.4 Based on the actual measurement data at the current prediction time, the state prediction is corrected to obtain the state estimation result at the current prediction time, specifically including:

[0157] S41, Calculate the predicted state covariance matrix, expressed as:

[0158] ,

[0159] in, To correspond to the current prediction time The predicted state covariance matrix, for The state covariance matrix at time t; and For state-based Calculate the state Jacobian matrix and its transpose; For state-based prediction quantities The calculated state Jacobian matrix, where the superscripts -1 and T denote taking the inverse and transpose of the matrix, respectively; for The noise covariance matrix of the prediction process at time step;

[0160] S42, calculate the Kalman gain, expressed as:

[0161] ,

[0162] In the formula, for Kalman gain at time step To measure the noise covariance matrix, To measure the Jacobian matrix, it is a measurement function. Regarding state variables The first-order partial derivative matrix describes how small changes in the state variables affect the measured values. Indicates in Seeking about The partial derivatives, i.e., in Predicting quantities based on state at any time Find the partial derivatives;

[0163] S43, Correct the state prediction based on the actual measurement data to obtain the state estimation result at the current prediction time. and the covariance matrix used to predict the state at the next time step Calculate the state covariance matrix , is represented as:

[0164] ,

[0165]

[0166] In the formula, for The measurement vector at time t, and its relationship with the measurement function The relationship is represented as: ,in for Measurement noise at any given moment.

[0167] In Kalman filtering, The state covariance matrix Mapped to the measurement space in order to calculate the Kalman gain. This leads to updates in the state estimate and covariance matrix. (Covariance matrix) The uncertainty of the current state estimate is quantified. When constructing typical scenarios of distribution network operation states based on the state estimate results, the covariance matrix... This is used to guide the sampling process when generating process noise samples, thereby sampling multiple sets of state scenarios, which can then be reduced using the k-means clustering algorithm and finally constructed as typical scenarios.

[0168] After obtaining multiple typical scenarios, the node power data at the current prediction time corresponding to different scenarios is obtained through power flow calculation. These are then combined with historical node power data to form node power data sequences, which are input into the load forecasting model to obtain node power data sequences for the prediction period corresponding to different typical scenarios, including the next prediction time. This prediction result is used for both distribution network operation status estimation and stochastic optimization for risk management, resulting in operation adjustment strategies for flexible devices such as energy storage in the distribution network.

[0169] When estimating the distribution network operation status at the next time step, the state prediction obtained from the load forecasting model is corrected based on actual measurement data and distribution network equipment control strategies to obtain the state estimation result, which specifically includes:

[0170] S81, superimpose the results of the power distribution network equipment control strategy onto the node power data at the predicted time output by the load forecasting model;

[0171] S82, calculates network power flow based on the superimposed node power data to obtain the state prediction quantity;

[0172] S83. By executing the operation of step S4, the state estimation results are obtained, which are used to generate typical scenarios of new distribution network operation states, and then based on these typical scenarios, node power prediction, state estimation and risk management stochastic optimization are performed for subsequent time periods.

[0173] III. System Risk Management Based on Dynamic Estimation of Distribution Network Operation Status

[0174] The risk management method in this embodiment samples multiple scenarios of state variables and reduces and constructs typical scenarios based on the k-means clustering algorithm. Then, it transforms them into multiple net load scenarios through power flow calculation and extends them to multiple time periods by combining historical data. It imports the load prediction model to predict the future net load. Based on the predicted net load scenarios, it adopts stochastic optimization to formulate and implement adjustment strategies. Finally, it feeds back the charging and discharging power of flexible equipment obtained from the operation decision to the dynamic state estimation model so as to adjust and correct the state prediction and estimation process for subsequent time periods.

[0175] 3.1 Definition of Power System Risk Management Indicators

[0176] To comprehensively address various unsafe operating conditions, this embodiment establishes calculation methods for node voltage risk and branch overload risk under different scenarios such as heavy load and excess renewable energy output. To reflect the significantly increased impact of operating conditions deviating further from the safe range, a quadratic indicator is used to quantitatively assess the risk. For different types of risks, the overall operating risk of the distribution network is obtained by summing subjective and objective weights.

[0177] 3.1.1 High Voltage Risk Indicators Low voltage risk indicators They are defined as follows:

[0178]

[0179]

[0180] in, and These are the upper and lower limits of the node voltage, which can be set to 1.05 times and 0.95 times the reference voltage, respectively.

[0181] 3.1.2 Branch Road Forward Heavy Load Index With reverse overload index They are defined as follows:

[0182] ,

[0183]

[0184] Let be the power flow of the j-th branch at time t. For the branch capacity, a load level exceeding 70% of the capacity is considered a heavy load.

[0185] 3.1.3 Overall operational risks of the power distribution system Through the weighted summation of various risk indicators, among which The risk weights are expressed as follows:

[0186]

[0187] In determining the weights, the analytic hierarchy process (AHP) and the entropy weight method were used to calculate the subjective and objective weights, respectively, and then the weights were determined by combining the coefficients. The final weight is obtained by merging the two types of weights, and is represented as follows: ,in Subjective weighting, For objective weighting.

[0188] 3.2 Establishing a stochastic optimization model for system risk management

[0189] In this embodiment, the optimization objective of system risk management is to minimize node voltage and branch overload risk. The objective function of the stochastic optimization model is expressed as:

[0190]

[0191] In the formula, Representing a scene The probability of occurrence Number of future moments considered for risk management This represents the total number of typical scenarios generated based on the state estimation results. and These represent the total number of distribution network nodes and the total number of branch lines, respectively. and These are the high-voltage risk indicators and their weights. and These are the low-voltage risk indicators and their weights. and These are the positive heavy load index for branch roads and its weight, respectively. respectively This refers to the reverse overload metric and its weights.

[0192] 3.3 Establish constraints for optimizing and adjusting risk management and operational strategies

[0193] The risk management optimization method can consider various flexible devices. In this embodiment, energy storage is selected as a representative flexible device. Risk management needs to meet the following constraints: node power balance constraints, energy storage charging and discharging power constraints, energy storage state of charge safety range constraints, energy storage state of charge and its operating power relationship constraints, and the relationship between remaining power and scheduling plan. The specific forms of the relevant constraints are as follows:

[0194] 3.3.1 Risk management must satisfy current flow constraints, expressed as:

[0195] ,

[0196] in and They are nodes With nodes The active and reactive power between; For nodes A set of connected nodes. For nodes The parent node; and They are nodes With nodes The active and reactive power between; and They are nodes The charging and discharging power of the energy storage device. and They are nodes The injected active and reactive power, and They are nodes With nodes The resistance and reactance between them This is the reference voltage.

[0197] 3.3.2 Risk management needs to meet the upper and lower limits of energy storage charging and discharging power constraints:

[0198]

[0199] in, For the first Energy storage The charging and discharging state at any given moment. and The first Energy storage The power of constant charging and discharging. It is an energy storage collection.

[0200] 3.3.3 Risk management needs to meet the energy storage state of charge constraints:

[0201]

[0202] in, For the first Energy storage State of charge at time t, and The first The state-of-charge limit and upper limit of an energy storage system. The time window under consideration and The first The charging and discharging efficiency of an energy storage device.

[0203] 3.3.4 The total charge and discharge amount within the time window must conform to the established scheduling plan. The corresponding constraints are expressed as follows:

[0204]

[0205] in, As of Time of the first The amount of electricity already consumed by the energy storage The power target corresponding to the established scheduling plan.

[0206] By solving the objective function, the optimization results of the energy storage charging and discharging operation strategy in the distribution network can be obtained, including the charging and discharging power of energy storage at each node. These optimization results can be used to guide distribution network scheduling optimization and, in conjunction with actual measurement data, to correct the state estimates obtained from network power flow calculations.

[0207] Example 3

[0208] This embodiment describes a computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for coordinating dynamic state estimation and risk management of a power distribution network as described in Embodiment 1 or 2.

[0209] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0210] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0211] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0212] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0213] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for coordinating dynamic state estimation and risk management of distribution networks, characterized in that, include: S1, acquire the distribution network operation status data of the previous moment, perform power flow calculation based on the distribution network operation status data, and obtain the node power data sequence; S2, input the node power data sequence into the pre-trained load prediction model to obtain the node power data sequence for the prediction period including the current prediction time; S3, based on the node power data at the current prediction time, perform distribution network operation state estimation to obtain the state prediction quantity at the current prediction time; S4. Based on the actual measurement data at the current prediction time, the state prediction is corrected to obtain the state estimation result at the current prediction time. S5. Based on the state estimation results, construct multiple typical scenarios of the distribution network operation state at the current predicted time; S6. Based on multiple typical scenarios of the state, determine the corresponding node power scenario at the current prediction time, and form a node power data sequence with historical node power data. Through the load prediction model, obtain the node power data sequence for the prediction period corresponding to different typical scenarios and including the next prediction time. S7. Based on the node power data at the next prediction time, a stochastic optimization model aimed at minimizing node voltage and branch overload risk is used to calculate the control strategy of distribution network equipment. S8, at the next prediction time, the state prediction quantity is corrected based on the actual measurement data at that time and the control strategy of the power distribution network equipment to obtain the state estimation result; Proceed to step S5 and repeat steps S5 to S8 to perform state estimation and distribution network equipment control strategy calculation for subsequent time periods.

2. The method according to claim 1, characterized in that, In the load forecasting model, each sequence transformation module includes: Embedding layers are used to project one-dimensional time-series data of the input node power into a high-dimensional feature space; The reshaping layer is used to reshape the projected one-dimensional time series data into multiple typical two-dimensional series data; A feature extraction layer is used to extract periodic feature data of power data from the multiple typical two-dimensional sequence data using two-dimensional convolutional kernels. The periodic features include intra-period variation features and inter-period variation features. The aggregation layer is used to convert the periodic feature data into one-dimensional time series data and perform weighted fusion to obtain the one-dimensional time series data of node power output by the current sequence transformation module.

3. The method according to claim 2, characterized in that, The reshaping layer reshapes the projected one-dimensional time series data into multiple typical two-dimensional sequence data, including: Perform a Fast Fourier Transform on the projected one-dimensional time series data, and calculate the average amplitude based on the frequency obtained from the transform. , is represented as: ,in For the original one-dimensional node power data sequence One-dimensional time series data after projection. The length of the original one-dimensional time series data. This is the original feature dimension, which can include the net load and reactive power of multiple nodes. The projected feature dimension determines the representation capability of the time series in a high-dimensional feature space, and depends on the input dimension and system complexity. Indicates the amplitude value. Represents the Fast Fourier Transform. This means that for each frequency, take its... The average magnitude of each feature dimension; Select based on the amplitude of each frequency and its average value The typical characteristic frequencies are expressed as: ,in For the first A typical characteristic frequency; Based on the period length corresponding to the selected typical characteristic frequencies, the sequence is padded with zeros at the edges corresponding to each frequency to transform the sequence. Remodeling Two-dimensional data , Typical feature frequency dimension The corresponding period length, and we have: .

4. The method according to claim 3, characterized in that, The aggregation layer converts the periodic feature data into one-dimensional time series data and performs weighted fusion, including: The output of the feature extraction layer Periodic feature data corresponding to a typical characteristic frequency They were converted into one-dimensional time series data respectively. And truncated to the length of the original one-dimensional time series data; Based on the Fourier transform results of the one-dimensional time series data received by the current sequence transformation module's embedded layer, the amplitude corresponding to each frequency is normalized using the Softmax function, and expressed as follows: ,in; To address the current issue The Fourier transform is performed on the one-dimensional time series data input to the previous sequence transformation module, resulting in the corresponding... The amplitude of a typical characteristic frequency, for The normalization result; Based on the normalization result, the normalized amplitude of each frequency is used as a weighting coefficient. One-dimensional time series data Perform a weighted summation to obtain the aggregated result of the current sequence transformation module, which is represented as: ,in For the current number The aggregation layer of the sequence transformation module is for the first... One-dimensional time series data is obtained by transforming the periodic feature data corresponding to the typical feature frequencies. This is the weighted fusion result of the aggregation layer; The aggregation result is used as input to the next sequence transformation module, or if the current sequence transformation module is the last sequence transformation module, the aggregation result is output as the prediction result of the load prediction model.

5. The method according to claim 1, characterized in that, In step S1, power flow calculation is performed based on the distribution network operating status data from the previous moment to obtain the node power data sequence, including: Based on the operating state value of the previous moment before the prediction time, the power data of each node is calculated using the network power flow equation. The calculation results are combined with the historical power data of each node with a preset step size to form a node power data sequence containing the current prediction time. The formula for calculating the node power data of the time preceding the prediction time is expressed as: , In the formula, for Time Node Power data, and These are the active power and reactive power, respectively. Represents the network power flow equation. For the predicted time The moment before The running status value; for any given time Operating status The network power flow equation is expressed as: in, For nodes A set of connected nodes. and They are respectively Time Node The node voltages and their phase angles, and These are the real and imaginary parts of the nodal admittance matrix elements, respectively. and They represent Time Node Net load The active power and reactive power in the process.

6. The method according to claim 5, characterized in that, Based on the node power data at the current prediction time, the distribution network operating state is estimated to obtain the state prediction at the current prediction time, including: Based on predicted time The node power data is used to calculate the state prediction at the prediction time using the network power flow equation, and is expressed as: in, The predicted time is calculated from the power flow. State prediction quantity, for Noise in the time-based prediction process for The inverse function; For prediction time The moment before The node power data sequence is used in the load forecasting model. The predicted time The node power data is represented as follows: ， in, To preset the historical step size, Nodes as input to the load forecasting model exist Power data at any given time.

7. The method according to claim 6, characterized in that, In step S4, based on the actual measurement data at the current prediction time, the state prediction is corrected to obtain the state estimation result at the current prediction time, including: S41, Calculate the predicted state covariance matrix, expressed as: ， in, To correspond to the current prediction time The predicted state covariance matrix, for The state covariance matrix at time t; and For state-based Calculate the state Jacobian matrix and its transpose; For state-based prediction quantities The calculated state Jacobian matrix, where the superscripts -1 and T denote taking the inverse and transpose of the matrix, respectively; for The noise covariance matrix of the prediction process at time step; S42, calculate the Kalman gain, expressed as: ， In the formula, for Kalman gain at time step To measure the noise covariance matrix, To measure the Jacobian matrix, is the measurement function. Regarding state variables The first-order partial derivative matrix, Indicates in Seeking about The partial derivatives, i.e., in Predicting quantities based on state at any time Find the partial derivatives; S43, Correct the state prediction based on the actual measurement data to obtain the state estimation result at the current prediction time. and the covariance matrix used to predict the state at the next time step Calculate the state covariance matrix , is represented as: ， In the formula, for The measurement vector at time t, and its relationship with the measurement function The relationship is represented as: ,in for Measurement noise at any given moment; For any time Measurement function Represented as: , State prediction Measurement function for variables It can be represented as: ,in and These are node power data predicted based on load forecasting models. , ,through The calculated predicted values ​​of node voltage and phase angle.

8. The method according to claim 7, characterized in that, In step S7, based on the node power data at the next prediction time, a stochastic optimization model aimed at minimizing node voltage and branch overload risk is used to calculate the distribution network equipment control strategy at the next prediction time, including: S71, based on the next prediction time The node power data is obtained through the inverse function of the network power flow equation. Calculate the corresponding state prediction quantities; S72, based on node power data and state predictions, uses a stochastic optimization model aimed at minimizing node voltage and branch overload risk to calculate the prediction time. Distribution network equipment control strategies; Among them, for the latest current prediction time The objective function of the stochastic optimization model is expressed as: In the formula, Representing a scene The probability of occurrence Number of future moments considered for risk management This represents the total number of typical scenarios generated based on the state estimation results. and These represent the total number of distribution network nodes and the total number of branch lines, respectively. and These are the high-voltage risk indicators and their weights. and These are the low-voltage risk indicators and their weights. and These are the positive heavy load index for branch roads and its weight, respectively. respectively For the reverse overload index and its index weights; Each risk indicator is expressed as follows: ， ， ， ； in, and These are the upper and lower limits of the node voltage, respectively. Let be the power flow of the j-th branch at time t. This is the capacity of the branch.

9. The method according to claim 8, characterized in that, The constraints for solving the stochastic optimization model include power flow constraints, upper and lower limits of energy storage charging and discharging power constraints, energy storage state of charge constraints, and predetermined scheduling plan constraints, which are expressed as follows: Current constraints: , Upper and lower limits of energy storage charging and discharging power constraints: ， Energy storage state of charge constraints: ， Given scheduling constraints: ， In the formula, and They are nodes With nodes The active and reactive power between; For nodes A set of connected nodes. For nodes The parent node; and They are nodes With nodes The active and reactive power between; and They are nodes The charging and discharging power of the energy storage device. and They are nodes The injected active and reactive power, and They are nodes With nodes The resistance and reactance between them The reference voltage; For the first Energy storage The charging and discharging state at any given moment. and The first Energy storage The power of constant charging and discharging. For energy storage collection; For the first Energy storage State of charge at time t, and The first The state-of-charge limit and upper limit of an energy storage system. The time window under consideration and The first The charging and discharging efficiency of the energy storage device; As of Time of the first The amount of electricity already consumed by the energy storage The power target corresponding to the established scheduling plan.

10. The method according to any one of claims 1-9, characterized in that, In step S8, at the next prediction time, the predicted state value is corrected based on the actual measurement data at that time and the control strategy of the distribution network equipment to obtain the state estimation result, including: S81, superimpose the results of the power distribution network equipment control strategy onto the node power data at the predicted time output by the load forecasting model; S82, calculates network power flow based on the superimposed node power data to obtain the state prediction quantity; S83, by executing the operation of step S4, the state estimation result is obtained, which is used to generate a typical scenario of new distribution network operation state.

Images