Distribution network load characteristic identification method and system, electronic equipment and storage medium

Abstract

The invention belongs to the technical field of power system relay protection, and particularly relates to a distribution network load characteristic identification method and system, electronic equipment and a storage medium. The method comprises the steps that an original power distribution network load curve is processed, a load curve reconstruction model based on cubic spline interpolation is constructed, and time sequence synchronization of multi-source load data is achieved; constructing a power distribution network load identification model based on a deep neural network; an automatic framework search mechanism based on Bayesian optimization is provided, the number of network layers and neuron parameters are dynamically optimized, and an optimal network framework is obtained; based on the monitoring data and the optimal network architecture, decoupling and fine analysis of the equipment-level electrical load are carried out; and carrying out multi-dimensional analysis on the load characteristics of the power distribution network based on decoupling and refined analysis results. According to the method, the engineering applicability of load characteristic identification is remarkably improved, the problem of data distortion under the complex working condition of the distribution network is effectively solved, and the matching precision of the low-frequency load shedding control strategy and the actual working condition of the system is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of power system relay protection, and particularly relates to a distribution network load characteristic identification method, system, electronic device and storage medium, more particularly to a distribution network load characteristic identification method, system, electronic device and storage medium suitable for low-frequency load shedding power shortage calculation. BACKGROUND

[0002] As a key parameter for measuring power quality, the frequency of the power system directly represents the dynamic balance state between active power supply and demand in the system. The stability level of the system frequency essentially depends on the real-time power matching degree between the total power generation and the power load. When the system encounters a sudden active power shortage, if the shortage value exceeds the system rotating reserve capacity, a chain reaction of frequency collapse will be triggered, and at this time, the emergency control must be implemented by the low-frequency load shedding device.

[0003] With the deep development of smart grids, the dynamic characteristics of distribution loads have changed significantly, and the impact on the safety and stability of the power system cannot be ignored. In the process of low-frequency load shedding, if the frequency response characteristics of the distribution load are not considered, the fixed parameters are directly used for power shortage calculation, which will lead to the problems of false cutting or under-cutting of the device, and seriously affect the frequency control accuracy. In addition, at the physical topology level, the low-frequency load shedding monitoring terminal is usually deployed at the head of the distribution feeder, and the monitoring range includes multiple user access points. At the load characteristic level, diversified power equipment is aggregated downstream of the terminal, and various types of equipment have differentiated frequency sensitivity characteristics. These two levels of characteristics bring new challenges to accurately identify the regional load characteristics and accurately estimate the active power shortage: the terminal can only obtain the total load power at the outlet of the feeder, and it is difficult to directly analyze the dynamic response characteristics of the internal load elements.

[0004] Load decomposition provides a new idea to solve this challenge. By decoupling the aggregated load power obtained by the monitoring terminal into device-level loads and combining the frequency response coefficients of various types of power equipment, the load characteristics of the region where the monitoring terminal is located can be effectively identified, and dynamic adaptive parameter support can be provided for the power shortage calculation of low-frequency load shedding. By accurately quantifying the natural recovery capability of the load power at different frequency drop stages, the matching accuracy of the low-frequency load shedding control strategy can be significantly improved. SUMMARY

[0005] In view of the deficiencies in the prior art, the present application provides a distribution network load characteristic identification method, system, electronic device and storage medium. The purpose of the present application is to solve the problem that the aggregated power obtained by the monitoring terminal cannot directly provide support for the development of low-frequency load shedding strategy.

[0006] The technical solution adopted by the present application to achieve the above-mentioned purpose is:

[0007] A distribution network load characteristic identification method, comprising:

[0008] The original distribution network load curve is corrected for outliers and filtered for noise using a data preprocessing method, and a load curve reconstruction model based on cubic spline interpolation is constructed to realize time series synchronization of multi-source load data.

[0009] A deep neural network-based distribution network load identification model is constructed.

[0010] An architecture automatic search mechanism based on Bayesian optimization is proposed, combined with a time series backpropagation algorithm, to dynamically optimize network layers and neuron parameters and obtain the optimal network architecture.

[0011] Based on the monitoring data obtained from the actual low-frequency load shedding system and the optimal network architecture, device-level power load decoupling and refined analysis are performed.

[0012] Based on the decoupling and refined analysis results, device-related parameters are calculated, and a multi-dimensional load static characteristic analysis system is established.

[0013] Based on the decoupling and refined analysis results, load dynamic response parameter calculation based on mathematical modeling is carried out, and multi-dimensional analysis of distribution network load characteristics is completed.

[0014] Further, the original distribution network load curve is corrected for outliers and filtered for noise using a data preprocessing method, and a load curve reconstruction model based on cubic spline interpolation is constructed to realize time series synchronization of multi-source load data, wherein the mean filtering method is selected for data preprocessing, and For the sequence formed by H filter sampling points of power near time t, for time t, the filtered data P t H is:

[0015]

[0016] In the above formula, h is the sampling point sequence number.

[0017] Further, the deep neural network-based distribution network load identification model is constructed with a long short-term memory neural network as the core module for time series feature extraction, and an end-to-end load characteristic identification framework is constructed. LSTM introduces a gating mechanism and a cell state memory unit. The input gate determines the update amplitude of the current time feature. The forgetting gate calculates the retention coefficient of the historical feature. The output gate determines the information content output at the current time. The cell state realizes information transmission across time steps through a loop connection. The LSTM unit calculation formula is:

[0018] f t = σf (W fx x t +W fh h t-1 +b f )

[0019] i t =σ i (W ix x t +W ih h t-1 +b i )

[0020] o t =σ i (W ox x t +W oh h t-1 +b o )

[0021]

[0022]

[0023] h t =φ(c t )·o t

[0024] In the above formula: f t is the forget gate, i t is the input gate, o t is the output gate, W fx , W fh , b f respectively represent the input weight matrix, the historical state weight matrix, and the bias of the forget gate, W ix , W ih , b i respectively represent the input weight matrix, the historical state weight matrix, and the bias of the input gate, W ox , W oh , b o respectively represent the input weight matrix, the historical state weight matrix, and the bias of the output gate, W cx , W ch , b c respectively represent the input weight matrix, the historical state weight matrix, and the bias of the cell state, x t is the input at t time, h t is the output at t time, b f , b i , b o , b c respectively represent the corresponding bias matrix, represents the candidate state, ct It represents a memory state.

[0025] Furthermore, the proposed automatic architecture search mechanism is based on Bayesian optimization and combined with the temporal backpropagation algorithm to dynamically optimize the number of network layers and neuron parameters to obtain the optimal network architecture. This includes a probabilistic surrogate model and a sampling function. The probabilistic surrogate model estimates the true function distribution based on finite sampled data. The model is a Gaussian process, and for a sample of length m, X = (x1, x2, ..., x...)... m The Gaussian probability distribution is expressed as G(μ,k), where μ is the mean function and k is the covariance function. Based on the probabilistic surrogate model, the posterior probability distribution of the true function is calculated, and the mean and variance of the candidate points are estimated. Based on this, the position of the candidate points for the next iteration is found. The upper confidence limit is used as the acquisition function, specifically:

[0026] g UCB =μ+βσ

[0027] In the above formula: β represents the balance factor, μ and σ represent the mean and variance respectively, and g UCB This represents the data acquisition function.

[0028] Furthermore, based on the monitoring data and optimal network architecture obtained from the actual low-frequency load shedding system on-site, the decoupling and refined analysis of equipment-level power load are achieved. Specifically, the low-frequency load shedding system, through intelligent monitoring terminals deployed at key nodes of the power grid, senses the operating status of lines participating in low-frequency load shedding in real time, records the action time sequence, execution strategy mode, and priority logic relationship digital information during load shedding, as well as load response curves and core analog parameters of the spatiotemporal distribution of load shedding, forming a multi-dimensional data set with strict temporal correlation. The monitored low-frequency load shedding data, accumulated over a long period, forms a large database of distribution network load characteristics covering different seasons and operating conditions. Through a distribution network load identification model based on deep neural networks combined with an automatic architecture search mechanism, a mapping relationship is established between the monitored load at the key points and the operating status of equipment groups, achieving layer-by-layer decoupling from the upper-level monitoring terminal to the lower-level power consumption unit and precise separation of power components from various types of power equipment.

[0029] Further, based on the decoupling and refined analysis results, the device-related parameters are calculated, and a multi-dimensional static load characteristic analysis system is established; wherein, the multi-dimensional static load characteristic analysis system is established by multiple interrelated and each with a focus analysis dimension, to construct a "power consumption mode-energy characteristic-behavior rule" load characteristic cognitive framework, including a device load proportion matrix, an electric energy consumption analysis, and a running time probability distribution parameter, wherein: the device load proportion matrix records the power consumption mode of the lower-level user unit accessed by the low-frequency load shedding monitoring terminal, representing the power contribution of different devices in each grid; the electric energy consumption analysis provides support for user type judgment in order of magnitude; the running time probability distribution focuses on the deep mining of user behavior rules, analyzes the habit of preferring to use a certain device in a certain time period, and describes the time preference of user power consumption in different scenarios.

[0030] Further, based on the decoupling and refined analysis results, the device-related parameters are calculated, and a multi-dimensional static load characteristic analysis system is established; wherein, the multi-dimensional static load characteristic analysis system is established by multiple interrelated and each with a focus analysis dimension, to construct a "power consumption mode-energy characteristic-behavior rule" load characteristic cognitive framework, including a device load proportion matrix, an electric energy consumption analysis, and a running time probability distribution parameter, wherein: the device load proportion matrix records the power consumption mode of the lower-level user unit accessed by the low-frequency load shedding monitoring terminal, representing the power contribution of different devices in each grid; the electric energy consumption analysis provides support for user type judgment in order of magnitude; the running time probability distribution focuses on the deep mining of user behavior rules, analyzes the habit of preferring to use a certain device in a certain time period, and describes the time preference of user power consumption in different scenarios.

[0031]

[0032] In the above formula, P t is the load sequence obtained after cubic spline interpolation, α0, α1, α2, α3 represent the proportion of each type of load in the total load respectively, f req is the system frequency, f reqN is the rated frequency, P tn is the load size under the rated frequency; each proportion satisfies:

[0033] α0+α1+α2+α3+···=1

[0034] In the above formula, α0 represents the proportion of frequency-insensitive load, the load presents power-frequency decoupling characteristics, the active power consumption is irrelevant to the system frequency change, and typical representative loads include incandescent lamp lighting devices and resistance heating furnaces; α1 represents linear frequency-dependent load, the active power of such load is positively correlated with the system frequency in the first order, and it covers rotating machinery devices; α2 represents quadratic frequency-dependent load, and the typical representative is eddy current loss; α3 and above represent high-order frequency-dependent load, and the typical characteristic device is the high-pressure feed water pump set matched with the thermal power generator set;

[0035] When the system frequency changes due to fault factors, the load power also changes, and the change characteristics are different for different proportions of loads, the input power and the load power reach a dynamic balance at the new frequency, and the active power shortage of low-frequency load shedding should be based on the new balance point; based on the obtained multi-dimensional static load characteristics, the proportion of each type of load in the terminal monitoring load is evaluated in detail, and the α0, α1, α2, α3 parameters of the grid area are obtained;

[0036] The load power-frequency dynamic response characteristic is characterized by a frequency response coefficient K, specifically:

[0037]

[0038] In the above formula, The above formula represents a differential calculation.

[0039] A distribution network load characteristic identification system comprises:

[0040] A data preprocessing and load reconstruction module is configured to correct outliers and filter noise of original distribution network load curves by using a data preprocessing method, and to construct a load curve reconstruction model based on cubic spline interpolation, thereby realizing time series synchronization of multi-source load data.

[0041] A deep neural network module is configured to construct a distribution network load identification model based on a deep neural network.

[0042] An architecture search module based on Bayesian optimization is configured to propose an architecture automatic search mechanism based on Bayesian optimization, dynamically optimize network layers and neuron parameters in combination with a time series back propagation algorithm, and obtain an optimal network architecture.

[0043] A device-level power load decoupling and fine analysis module is configured to realize decoupling and fine analysis of device-level power loads based on monitoring data obtained from an actual low-frequency load shedding system and the optimal network architecture.

[0044] A multi-dimensional load static analysis module is configured to calculate device-related parameters based on decoupling and fine analysis results, and establish a multi-dimensional load static characteristic analysis system.

[0045] A multi-dimensional analysis module of distribution network load characteristics is configured to carry out load dynamic response parameter calculation based on mathematical modeling based on decoupling and fine analysis results, and complete multi-dimensional analysis of distribution network load characteristics.

[0046] An electronic device comprises a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, and the processor implements the steps of any one of the distribution network load characteristic identification methods.

[0047] A computer storage medium has a computer program stored thereon, and the computer program is executed by a processor to implement the steps of any one of the distribution network load characteristic identification methods.

[0048] The present application has the following advantages and benefits:

[0049] The application constructs a full-process analysis system of "data cleaning-load identification-characteristic evaluation", which significantly improves the engineering applicability of load characteristic identification. Based on the mixed processing of mean filtering and cubic spline interpolation, the data distortion problem under complex working conditions of distribution network is effectively overcome, and a high-credibility data benchmark is established for subsequent analysis; the collaborative application of deep neural network and Bayesian optimization realizes the collaborative optimization of time series feature extraction and model architecture, so that the load identification model has the dynamic adjustment ability of self-adapting to complex working conditions, realizes the layer-by-layer decoupling of gateway load and the precise stripping of equipment-level load; the proposed multi-dimensional characteristic analysis framework realizes the panoramic portrayal of load static and dynamic characteristics, provides accurate load dynamic response parameters for low-frequency load shedding strategy making, and effectively improves the matching accuracy of low-frequency load shedding control strategy and actual working conditions of the system. Compared with the prior art, the application maintains the robustness of the algorithm while considering the feasibility of engineering implementation, and provides important technical support for frequency emergency control in the background of new power system. BRIEF DESCRIPTION OF DRAWINGS

[0050] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:

[0051] Figure 1 is a flow chart of a distribution network load characteristic identification method of the application;

[0052] Figure 2 is a distribution network load identification model architecture based on deep neural network in the application;

[0053] Figure 3 is the electric energy proportion of each device in the application. DETAILED DESCRIPTION

[0054] In order to more clearly understand the above-mentioned purposes, features and advantages of the application, the application will be further described in detail below in conjunction with the drawings and specific embodiments. It should be noted that the embodiments of the application and the features in the embodiments can be combined with each other without conflict.

[0055] In the following description, many specific details are set forth in order to provide a thorough understanding of the application, but the application can also be implemented in other ways different from those described herein, therefore, the scope of protection of the application is not limited by the specific embodiments disclosed below.

[0056] The technical solutions of some embodiments of the application will be described below with reference to the drawings. Figures 1-3

[0057] Embodiment 1

[0058] The application provides an embodiment, which is a distribution network load characteristic identification method. As shown in Figure 1 ​Fig. 1 is a flow chart of a power distribution network load characteristic identification method according to the present application. Figure 1 Fig. 1 is a flow chart of a power distribution network load characteristic identification method according to the present application.

[0059] The method comprises the following steps:

[0060] Step 1. The original power distribution network load curve is subjected to abnormal value correction and noise filtering processing by using a data preprocessing method, and a load curve reconstruction model based on cubic spline interpolation is constructed to realize time sequence synchronization of multi-source load data. The load curve after data preprocessing by selecting mean filtering method retains the original fluctuation characteristics while effectively eliminating random peak interference. The multi-source data after cubic spline interpolation reconstruction is accurately aligned in time domain resolution, laying a reliable data foundation for subsequent fine decomposition of device-level load.

[0061] Step 2. A power distribution network load identification model based on deep neural network is constructed as shown in Fig. 2. Figure 2 Figure 2 Fig. 2 is a deep neural network-based power distribution network load identification model architecture according to the present application.

[0062] Step 3. An architecture automatic search mechanism based on Bayesian optimization is proposed, combined with time sequence back propagation algorithm, to dynamically optimize network layer and neuron parameters and obtain optimal network architecture. The results show that the proposed architecture search mechanism exhibits good convergence characteristics within a limited time range after multiple rounds of iterative optimization, and the load decomposition accuracy presents a continuous improvement trend.

[0063] Step 4. Based on the monitoring data obtained from the actual low-frequency load shedding system and the optimal network architecture in step 3, device-level electricity load is decoupled and finely analyzed.

[0064] Step 5. Based on the decoupling and fine analysis results in step 4, device load proportion matrix, power consumption analysis, running time probability distribution and other parameters are calculated to establish a multi-dimensional load static characteristic analysis system. The calculation results are shown in Fig. 5. Figure 3 Figure 3 Fig. 5 is a multi-dimensional load static characteristic analysis system according to the present application.

[0065] Step 6. Based on the decoupling and fine analysis results in step 4, load dynamic response parameter calculation based on mathematical modeling is carried out to finally complete the multi-dimensional analysis of power distribution network load characteristics.

[0066] ​​In step 1, the data preprocessing method is mainly aimed at the data distortion problems caused by the high-frequency noise and abnormal sampling under the complex operation environment of the power distribution network and the coupling of multiple factors. Nonlinear factors such as voltage transient fluctuation, random interference of load, and time-varying loss of line cause the load curve to present non-stationary characteristics. If such non-stationary distorted data is not preprocessed, the load characteristics of the equipment may be covered by high-frequency noise, and the time sequence correlation of the front and rear data will be destroyed. Abnormal points in the sampling process will distort the statistical characteristic distribution of the load curve, destroy the smoothness and continuity of the load curve, reduce the physical interpretability of the load curve, and even cause the analysis result to deviate from the real load variation law, resulting in distortion of the identification of the load characteristics of the power distribution network.

[0067] In step 1, the mean filtering method is selected for data preprocessing. This method can effectively suppress Gaussian white noise and other random interference through low-complexity neighborhood mean operation without prior assumptions, and is especially suitable for filtering transient spikes caused by abnormal sampling. The specific principle is that a sequence of H sampling points is formed around t time. For t time, the filtered data P filter t H is:

[0068]

[0069] In the above formula, h is the sampling point sequence number. In addition to high-frequency noise and abnormal samples, the problem of multi-source heterogeneous sampling also seriously restricts the effective identification of load characteristics. The monitoring terminal is difficult to achieve complete synchronization and alignment of multi-channel data due to the difference in hardware clock accuracy. In addition, the statistical period of different terminals is different, and the load curve fluctuation exists in the scale mismatch in the time domain resolution. Therefore, in step 1, a cubic spline interpolation method is used to construct a load reconstruction model. This method can overcome the curvature mutation caused by traditional linear interpolation, ensure numerical stability, and realize accurate conversion of non-uniform sampling data to equal-interval high-resolution load curves, establishing a unified time sequence reference for subsequent fine analysis of device-level electricity load. The load sequence obtained after cubic spline interpolation is P t .

[0070] ​In step 2, the distribution network load identification model based on deep neural networks uses a long short-term memory (LSTM) neural network as the core module for time-series feature extraction to construct an end-to-end load characteristic identification framework. LSTM is an improved structure proposed to address the gradient vanishing defect of traditional recurrent neural networks (RNNs) in long-term time-series modeling. During error backpropagation, the gradient information of historical moments decays exponentially with increasing time steps in traditional RNNs, making it difficult to model long-cycle load fluctuations. LSTM solves this problem by introducing a gating mechanism and cell state memory units: the input gate determines the update magnitude of the current feature; the forget gate calculates the retention coefficient of historical features; and the output gate determines the information content output at the current moment. The cell state, through recurrent connections, enables information transfer across time steps, significantly reducing the impact of gradient vanishing. The calculation formulas for each unit of LSTM are as follows:

[0071] f t =σ f (W fx x t +W fh h t-1 +b f )

[0072] i t =σ i (W ix x t +W ih h t-1 +b i )

[0073] o t =σ i (W ox x t +W oh h t-1 +b o )

[0074]

[0075]

[0076] h t =φ(c t )·o t

[0077] In the above formula: f t For the Gate of Oblivion, i t For the input gate, o t For output gate, W fx Wfh , b f represent input weight matrix, history state weight matrix, bias of forget gate respectively, W ix , W ih , b i represent input weight matrix, history state weight matrix, bias of input gate respectively, W ox , W oh , b o represent input weight matrix, history state weight matrix, bias of output gate respectively, W cx , W ch , b c represent input weight matrix, history state weight matrix, bias of cell state respectively, x t is input at t moment, h t is output at t moment, b f , b i , b o , b c represent corresponding bias matrix respectively, represent candidate state, c t represent memory state.

[0078] In step 3, the architecture automatic search mechanism based on Bayesian optimization takes the number of layers and nodes of the long short-term memory neural network as the search object, and performs global optimization based on Bayesian theorem and probability model. Compared with traditional optimization methods, the main advantages of the architecture automatic search mechanism based on Bayesian optimization include:

[0079] 1) Reduce the number of evaluations by active sampling, quickly converge in a limited number of iterations, greatly improve the search efficiency, and effectively avoid local optimal trap.

[0080] 2) Low dependence on target function form, can handle high-dimensional black box problem without gradient information, and has wide applicability in complex engineering optimization problems. In the iteration process, the mechanism constantly updates the posterior distribution through Bayesian inference, and gradually realizes accurate estimation of the real function. Then, based on the posterior distribution updated each time, a collection function is constructed to balance the exploration and development strategy, and the next optimal evaluation point is determined to maximize its own value.

[0081] The mechanism includes two core modules: a probability agent model and a collection function. The probability agent model estimates the distribution of the real function based on limited sampling data, and the most commonly used model is Gaussian Process (GP). For a sample X = (x1, x2,..., x m), the expression form of Gaussian probability distribution is G(μ, k), wherein μ is a mean function, and k is a covariance function. Based on the probabilistic surrogate model, the posterior probability distribution of the real function can be calculated, and the mean and variance of the candidate point can be estimated, and on this basis, the position of the next iteration candidate point is found. In step 2, the upper limit of confidence is taken as the acquisition function, specifically:

[0082] g UCB = μ + βσ

[0083] In the above formula: β represents a balance factor, μ and σ represent mean and variance respectively, and g UCB represents the acquisition function.

[0084] Preferably, in step 4, the low-frequency load shedding monitoring system realizes real-time sensing of the operation state of the line participating in low-frequency load shedding through the intelligent monitoring terminal deployed at the key nodes of the power grid, and records the digital quantity information such as the action time sequence, the execution strategy mode and the priority logic relationship in the process of load shedding, as well as the core analog quantity parameters such as the load response curve and the time and space distribution of load shedding amount, forming a multi-dimensional data set with strict time sequence correlation. After a long time of accumulation, the monitored low-frequency load shedding load data can form a large database of power distribution network load characteristics covering different seasons and operating conditions. On this basis, by using the power distribution network load identification model based on deep neural network in step 2, combined with the automatic search mechanism in step 3, the mapping relationship between the key load monitored by the low-frequency load shedding and the operation state of the device group is established, realizing the layer-by-layer decoupling of the upper monitoring terminal to the lower power unit, and the accurate separation of power components of various power devices.

[0085] Preferably, in step 5, the multi-dimensional load static characteristic analysis system is established, which constructs a load characteristic cognition framework of "power mode-energy feature-behavior rule" through multiple interrelated and each focused analysis dimensions, provides multi-perspective data support for safe and stable operation of the power distribution network, including device load proportion matrix, power consumption analysis, operating time probability distribution parameters, etc. Among them: the device load proportion matrix records the power mode of the lower user unit accessed by the low-frequency load shedding monitoring terminal, representing the power contribution of different devices in each grid; the power consumption analysis can provide support for user type judgment in order of magnitude; the operating time probability distribution focuses on the deep mining of user behavior rule, analyzes the living habit of the user who tends to use a certain device in a certain time period, and describes the time preference of the user in different scenarios.

[0086] Preferably, in step 6, based on the decoupling and fine analysis results, a load dynamic response parameter calculation framework based on mathematical modeling is carried out, which can characterize the load in power-frequency characteristics. Specifically:

[0087]

[0088] In the above formula, P t is the load sequence obtained after cubic spline interpolation, and a0, a1, a2, a3 represent the proportions of various types of loads in the total load, respectively, f req is the system frequency, f reqN is the rated frequency, and P tn is the load size at the rated frequency. Each proportion satisfies:

[0089] a0+a1+a2+a3+…=1

[0090] In the above formula, a0 represents the proportion of frequency-insensitive loads, which present power-frequency decoupling characteristics, and the active power consumption of which is irrelevant to the system frequency. Typical representative loads include incandescent lamp lighting devices, resistance heating furnaces, etc.; a1 represents linear frequency-dependent loads, the active power of which is positively correlated with the system frequency in the first order, mainly covering rotating mechanical devices; a2 represents quadratic frequency-dependent loads, typical representative being eddy current loss; and a3 and above represent high-order frequency-dependent loads, typical characteristic devices being high-pressure feed water pump sets matched with thermal power generators.

[0091] When the system frequency changes due to faults or other factors, the load power will also change, and the change characteristics are different for various types of loads. The input power and the load power reach a dynamic balance at the new frequency, and the active power shortage of low-frequency load shedding should be based on the new balance point. Based on the multi-dimensional load static characteristics obtained in step 5, the device usage of the lower-level users connected to the low-frequency load shedding monitoring terminal in different seasons, months and time points can be clearly depicted. By statistically analyzing the load distribution of various types of loads containing devices under the terminal grid, the proportions of various types of loads in the terminal monitoring load can be evaluated in detail, and finally the parameters such as a0, a1, a2 and a3 of the grid area can be obtained.

[0092] The dynamic response characteristics of the load power to the frequency are characterized by the frequency response coefficient K, which is specifically:

[0093]

[0094] In the above formula, characterizes the differential calculation.

[0095] Embodiment 2

[0096] The application further provides an embodiment, which is a distribution network load characteristic identification system, more specifically, a distribution network load characteristic identification system suitable for low-frequency load shedding power shortage calculation. The system is used to realize the distribution network load characteristic identification method of embodiment 1.

[0097] A distribution network load characteristic identification system, specifically comprising:

[0098] Data preprocessing and load reconstruction module: used for median filtering of load curve, realization of abnormal point correction and noise processing, and load curve reconstruction based on cubic spline interpolation, realization of time synchronization of multi-source data.

[0099] Deep neural network module: used for establishing long short-term memory neural network, extracting load curve time sequence characteristics, and constructing end-to-end load characteristic identification framework;

[0100] Architecture search module based on Bayesian optimization: used for searching optimal number of layers and nodes of the established long short-term neural network;

[0101] Device-level power consumption load decoupling and fine analysis module: used for decomposing the load curve monitored by the low-frequency load shedding terminal into device-level load;

[0102] Multi-dimensional load static analysis module: used for analyzing static load characteristics of the distribution network;

[0103] Multi-dimensional analysis module of load characteristics of the distribution network: used for analyzing dynamic load characteristics of the distribution network.

[0104] Embodiment 3

[0105] Based on the same inventive concept, the embodiments of the present application also provide a computer electronic device, including a storage medium, a processor, and a computer program stored in the storage medium and executable on the processor. The processor executes the computer program to realize the steps of the power distribution network load characteristic identification method according to any one of the embodiments 1 or 2.

[0106] Embodiment 4

[0107] Based on the same inventive concept, the embodiments of the present application also provide a computer storage medium, which has a computer program stored thereon. The computer program is executed by a processor to realize the steps of the power distribution network load characteristic identification method according to any one of the embodiments 1 or 2.

[0108] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented 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.

[0109] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps in the flowchart block or blocks.

[0110] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps in the flowchart block or blocks.

[0111] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flow or blocks Figure 1 means for functionally implementing the steps in the flowchart block or blocks.

[0112] Finally, it should be noted that the above-mentioned embodiments are merely used to illustrate the technical solutions of the present application, rather than limiting the same. Although the present application is described in detail with reference to the above embodiments, it should be understood by those skilled in the art that the specific embodiments of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application, and any modification or replacement without departing from the spirit and scope of the present application should be included in the protection scope of the claims of the present application.

Claims

1. A method for identifying the characteristics of distribution network loads, characterized by: include: Data preprocessing methods are used to correct outliers and filter noise in the original distribution network load curves, and a load curve reconstruction model based on cubic spline interpolation is constructed to achieve time-series synchronization of multi-source load data. Construct a distribution network load identification model based on deep neural networks; We propose an automatic architecture search mechanism based on Bayesian optimization, which combines the temporal backpropagation algorithm to dynamically optimize the number of network layers and neuron parameters to obtain the optimal network architecture. Based on the monitoring data obtained from the actual low-frequency load shedding system on site and the optimal network architecture, we conduct decoupling and refined analysis of equipment-level power load. Based on the decoupling and refined analysis results, relevant equipment parameters are calculated, and a multi-dimensional load static characteristic analysis system is established. Based on the decoupling and refined analysis results, load dynamic response parameters are calculated based on mathematical modeling, and a multi-dimensional analysis of the load characteristics of the distribution network is completed.

2. The method for identifying distribution network load characteristics according to claim 1, characterized in that: The method utilizes data preprocessing to correct outliers and filter noise in the original distribution network load curve, and constructs a load curve reconstruction model based on cubic spline interpolation to achieve time-series synchronization of multi-source load data. Specifically, the mean filtering method is selected for data preprocessing, and P is set as follows: t (1) ,P t (2) ,P t (3) ,...,P t (Hfilter) H around time t filter The sequence formed by the power of each sampling point, for time t, the filtered data P t H for: In the above formula: h is the sampling point sequence number.

3. The method, system, electronic device, and storage medium for identifying distribution network load characteristics according to claim 1, characterized in that: The proposed deep neural network-based distribution network load identification model uses a long short-term memory (LSTM) neural network as the core module for time-series feature extraction, constructing an end-to-end load characteristic identification framework. The LSTM incorporates a gating mechanism and cell state memory units. The input gate determines the update magnitude of the current feature; the forget gate calculates the retention coefficient of historical features; the output gate determines the information content output at the current moment; and the cell state achieves information transfer across time steps through recurrent connections. The calculation formulas for each LSTM unit are as follows: f t =σ f (W fx x t +W fh h t-1 +b f ) i t =σ i (W ix x t +W ih h t-1 +b i ) o t =σ i (W ox x t +W oh h t-1 +b o ) h t =φ(c t )·o t In the above formula: f t For the Gate of Oblivion, i t For the input gate, o t For output gate, W fx W fh b f W represents the input weight matrix, historical state weight matrix, and bias of the forget gate, respectively. ix W ih b i W represents the input weight matrix, historical state weight matrix, and bias of the input gate, respectively. ox W oh b o W represents the input weight matrix, historical state weight matrix, and bias of the output gate, respectively. cx W ch b c The input weight matrix, historical state weight matrix, and bias, respectively, represent the cell state. t Input at time t, h t For the output at time t, b f b i b o b c These represent the corresponding bias matrices. Representing the candidate state, c t It represents a memory state.

4. The method for identifying distribution network load characteristics according to claim 1, characterized in that: The proposed automatic architecture search mechanism is based on Bayesian optimization and combines it with the temporal backpropagation algorithm to dynamically optimize the number of network layers and neuron parameters to obtain the optimal network architecture. This mechanism includes a probabilistic surrogate model and a sampling function. The probabilistic surrogate model estimates the true function distribution based on finite sampled data. The model is a Gaussian process, and for a sample of length m, X = (x1, x2, ..., x...)... m The Gaussian probability distribution is expressed as G(μ,k), where μ is the mean function and k is the covariance function. Based on the probabilistic surrogate model, the posterior probability distribution of the true function is calculated, and the mean and variance of the candidate points are estimated. Based on this, the position of the candidate points for the next iteration is found. The upper confidence limit is used as the acquisition function, specifically: g UCB =μ+βσ In the above formula: β represents the balance factor, μ and σ represent the mean and variance respectively, and g UCB This represents the data acquisition function.

5. The method for identifying distribution network load characteristics according to claim 1, characterized in that: Based on the monitoring data and optimal network architecture obtained from the actual low-frequency load shedding system on site, the system achieves decoupling and refined analysis of equipment-level power load. The low-frequency load shedding system, through intelligent monitoring terminals deployed at key nodes of the power grid, senses the operating status of lines participating in low-frequency load shedding in real time, records the action time sequence, execution strategy mode, and priority logic relationship digital information during load shedding, as well as load response curves and core analog parameters of the spatiotemporal distribution of load shedding, forming a multi-dimensional data set with strict temporal correlation. The monitored low-frequency load shedding data, accumulated over a long period, forms a large database of distribution network load characteristics covering different seasons and operating conditions. Through a distribution network load identification model based on deep neural networks combined with an automatic architecture search mechanism, a mapping relationship is established between the monitored load at the low-frequency load shedding point and the operating status of equipment groups, achieving layer-by-layer decoupling from the upper-level monitoring terminal to the lower-level power consumption unit and precise separation of power components from various types of power equipment.

6. The method for identifying distribution network load characteristics according to claim 1, characterized in that: Based on the decoupling and refined analysis results, calculate the device-related parameters and establish a multi-dimensional load static characteristic analysis system. Among them, establishing a multi-dimensional load static characteristic analysis system is to construct a load characteristic cognitive framework of "power consumption mode - energy characteristics - behavior rules" through multiple interrelated and focused analysis dimensions, including the device load proportion matrix, power consumption analysis, and operation time probability distribution parameters. Specifically: The device load proportion matrix records the power consumption modes of the subordinate user units connected to the low-frequency load shedding monitoring terminal, representing the power contribution degrees of different devices in each grid; the power consumption analysis provides support for user type judgment in terms of magnitude; the operation time probability distribution focuses on the in-depth exploration of user behavior rules, analyzes the living habits of being more inclined to use a certain device in a certain time period, and depicts the time preference of users' power consumption in different scenarios.

7. The method for identifying distribution network load characteristics according to claim 1, characterized in that: Based on the decoupling and refined analysis results, carry out the calculation of load dynamic response parameters based on mathematical modeling, and complete the multi-dimensional analysis of the load characteristics of the distribution network. Among them: In the above formula, P t The load sequence is obtained after cubic spline interpolation, where α0, α1, α2, and α3 represent the proportion of each type of load to the total load, and f req f is the system frequency. reqN For the rated frequency, P tn The load size is at the rated frequency; the proportions satisfy: α0+α1+α2+α3+···＝1 In the above formula, α0 represents the proportion of frequency-insensitive loads. The load shows power-frequency decoupling characteristics, and the active power consumption is independent of the system frequency change. Typical representative loads include incandescent lighting devices and resistance heating furnaces; α1 represents linearly frequency-related loads. The active power of such loads is linearly positively correlated with the system frequency in the first order, covering rotating machinery equipment; α2 represents quadratically frequency-related loads, and the typical representative is eddy current loss; α3 and above represent high-order frequency-related loads, and the typical characteristic device is the high-pressure feed water pump group supporting the thermal power generation unit. When the system frequency changes due to fault factors, the load power also changes accordingly. The change characteristics vary depending on the proportion of various loads. The input power and the load power reach dynamic balance at the new frequency, and the low-frequency load shedding active power deficit should be based on the new balance point. Based on the obtained multi-dimensional load static characteristics, conduct a detailed assessment of the proportion of various loads in the terminal monitoring load, and obtain the α0, α1, α2, and α3 parameters of this part of the grid area. The dynamic response characteristic of the load power to the frequency is characterized by the frequency response coefficient K, specifically: In the above formula, Characterize differential calculation.

8. A distribution network load characteristic identification system, characterized in that: Including: Data preprocessing and load reconstruction module: Used to perform outlier correction and noise filtering on the original distribution network load curve using data preprocessing methods, and construct a load curve reconstruction model based on cubic spline interpolation to achieve the time series synchronization of multi-source load data. Deep neural network module: Used to construct a distribution network load identification model based on a deep neural network. Architecture search module based on Bayesian optimization: Used to propose an automatic architecture search mechanism based on Bayesian optimization, combine with the time series backpropagation algorithm, dynamically optimize the network layer number and neuron parameters, and obtain the optimal network architecture. Device-level power consumption load decoupling and refined analysis module: Used to achieve the decoupling and refined analysis of device-level power consumption loads based on the monitoring data obtained from the actual low-frequency load shedding system on site and the optimal network architecture. Multi-dimensional load static analysis module: Used to calculate device-related parameters based on the decoupling and refined analysis results and establish a multi-dimensional load static characteristic analysis system. Multi-dimensional analysis module for distribution network load characteristics: Based on the decoupling and refined analysis results, it performs calculations of load dynamic response parameters based on mathematical modeling, and completes multi-dimensional analysis of distribution network load characteristics.

9. An electronic device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the distribution network load characteristic identification method according to any one of claims 1-7.

10. A computer storage medium, characterized in that: The computer storage medium contains a computer program, which, when executed by a processor, implements the steps of a distribution network load characteristic identification method as described in any one of claims 1-7.

Images