Inertia evaluation method and device for wind and light storage station, medium and program product

Abstract

The invention relates to the technical field of novel power system operation and control, and discloses a wind and light storage station inertia evaluation method and device, a medium and a program product, and the method comprises the steps: processing the data of each cluster through a plurality of evaluation models constructed through deep learning, and carrying out the final evaluation through the integration of a wind and light storage station inertia evaluation aggregation model, the non-linear relation between the output active power and the grid-connected point frequency change rate in the virtual inertia control of the new energy station is fully considered, and the evaluation error caused by simplified processing is remarkably reduced. Meanwhile, by fusing the real-time inertia evaluation result of each sub-cluster and station characteristics, integration from single cluster inertia evaluation to whole station overall inertia evaluation is realized. Furthermore, a plurality of evaluation models and aggregation models are constructed through deep learning, so that the complex nonlinear relationship between the inertia and the operation state of the station can be more accurately captured, and the evaluation deviation is greatly reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of new power system operation and control technology, and particularly relates to a wind-solar-storage station inertia evaluation method, device, medium and program product. BACKGROUND

[0002] Traditional thermal power and hydropower use synchronous motors as generators, and the rotor frequency of the synchronous motor is synchronized with the grid frequency. The synchronous motor can provide real rotational inertia by releasing the kinetic energy of the rotor. New energy power generation (wind power and photovoltaic power) is synchronized with the grid frequency through the control of power electronic converters, and cannot provide the same rotational inertia as the traditional synchronous motor. Therefore, it is necessary to evaluate the inertia of new energy.

[0003] Most of the existing technologies are based on the analysis of the frequency response characteristics of wind turbines and photovoltaic power and the equivalent calculation of the control link to realize the equivalent inertia evaluation of new energy stations. However, the existing technologies use a lot of simplification in the equivalent inertia evaluation of new energy stations, ignore the nonlinear relationship between the output active power and the frequency change rate of the grid connection point in the process of virtual inertia control of new energy stations, and lead to certain errors in the inertia evaluation results, which is particularly obvious in abnormal weather (strong wind, cloudy, blowing sand, etc.). SUMMARY

[0004] Therefore, the present application provides a wind-solar-storage station inertia evaluation method, device, medium and program product to solve the problem that the existing technologies use a lot of simplification in the equivalent inertia evaluation of new energy stations, ignore the nonlinear relationship between the output active power and the frequency change rate of the grid connection point in the process of virtual inertia control of new energy stations, and lead to certain errors in the inertia evaluation results.

[0005] In the first aspect, the present application provides a wind-solar-storage station inertia evaluation method, and the wind-solar-storage station includes a wind turbine group, a photovoltaic turbine group and an energy storage group. The method includes:

[0006] The real-time wind turbine group data set of the wind turbine group, the real-time photovoltaic group data set of the photovoltaic group, the real-time energy storage group data set of the energy storage group, and the real-time station feature set of the wind-solar-storage station are obtained; the real-time wind turbine group data set is input into a target wind turbine group inertia evaluation model for processing to obtain a plurality of first real-time equivalent inertia evaluation values of the wind turbine group; the real-time photovoltaic group data set is input into a target photovoltaic group inertia evaluation model for processing to obtain a plurality of second real-time equivalent inertia evaluation values of the photovoltaic group; the real-time energy storage group data set is input into a target energy storage group inertia evaluation model for processing to obtain a plurality of third real-time equivalent inertia evaluation values of the energy storage group; and the real-time station feature set, the plurality of first real-time equivalent inertia evaluation values, the plurality of second real-time equivalent inertia evaluation values, and the plurality of third real-time equivalent inertia evaluation values are input into a wind-solar-storage station inertia evaluation aggregation model for processing to obtain a target inertia evaluation value of the wind-solar-storage station.

[0007] The wind-solar-storage station inertia evaluation method provided by the application can cover real-time operation characteristics of wind, light, storage, and station levels, and adapt to evaluation requirements under normal and abnormal weather. Further, the target wind turbine group inertia evaluation model, the target photovoltaic group inertia evaluation model, and the target energy storage group inertia evaluation model process the data of each group, respectively, and then the wind-solar-storage station inertia evaluation aggregation model integrates and performs final evaluation, fully considers the nonlinear relationship between output active power and grid point frequency change rate in new energy station virtual inertia control, significantly reduces the evaluation error caused by simplification, and maintains high evaluation accuracy even under abnormal weather such as strong wind, cloudy weather, and blowing sand. At the same time, the wind-solar-storage station inertia evaluation aggregation model fuses the real-time inertia evaluation results of each sub-group and the station features, realizes the integration from single group inertia evaluation to overall station inertia evaluation, and provides comprehensive and accurate station-level inertia data support for power grid dispatching and frequency stability control. Further, the plurality of evaluation models and the aggregation model constructed by deep learning can more accurately capture the complex nonlinear relationship between station inertia and operating state than traditional mathematical methods, avoid the defect that traditional mathematical analysis methods are difficult to represent nonlinear characteristics, and greatly reduce the evaluation deviation.

[0008] In an optional implementation, the method further includes:

[0009] A historical wind turbine group data set of the wind turbine group is obtained; an initial directed graph and an initial node feature matrix are constructed according to the historical wind turbine group data set; and a target wind turbine group inertia evaluation model is constructed by using a graph convolution network for model training based on the historical wind turbine group data set, the initial directed graph, and the initial node feature matrix.

[0010] The inertia evaluation method of the wind and light storage station provided by the application can restore the spatial correlation of the wind turbine group and the equipment characteristics by constructing an initial directed graph and an initial node feature matrix. Further, the model is trained by a graph convolution network, so that the constructed target wind turbine group inertia evaluation model can capture the complex relationships such as the wake effect between wind turbines and the geographical location correlation, provide a suitable model basis for accurate evaluation of the inertia of the wind turbine group, and reduce the evaluation deviation caused by the simplification of the characteristics of the wind turbine group.

[0011] In an optional embodiment, based on the historical wind turbine group data set, the initial directed graph and the initial node feature matrix, a target wind turbine group inertia evaluation model is constructed by training the model using a graph convolution network, comprising:

[0012] The target direction adjacency matrix and the target weight matrix are obtained; the initial directed graph and the initial node feature matrix are input into the input layer of the graph convolution network for processing to obtain a regularized target directed graph and a normalized target node feature matrix; the target directed graph, the target node feature matrix, the target direction adjacency matrix and the target weight matrix are input into the convolution layer of the graph convolution network for processing to obtain an enhanced initial node embedding matrix; the initial node embedding matrix, the target direction adjacency matrix and the target weight matrix are input into the hidden layer of the graph convolution network for processing to obtain a target node embedding matrix; the target node embedding matrix and a preset full connection layer weight matrix are input into the output layer of the graph convolution network for processing to obtain a plurality of first historical equivalent inertia evaluation values in the wind turbine group; and the historical wind turbine group data set is taken as the input, and the plurality of first historical equivalent inertia evaluation values are taken as the output to construct the target wind turbine group inertia evaluation model.

[0013] The inertia evaluation method of the wind and light storage station provided by the application further optimizes the training process of the wind turbine group inertia evaluation model through the targeted processing of each layer of the graph convolution network, ensures that the target wind turbine group inertia evaluation model can accurately learn the spatial dependence and wake influence law between wind turbines, improves the accuracy of the first historical equivalent inertia evaluation value, provides reliable model support for subsequent real-time evaluation of the inertia of the wind turbine group, and avoids the problem that the traditional simplified calculation cannot adapt to the complex correlation characteristics of the wind turbine.

[0014] In an optional embodiment, the method further comprises:

[0015] The historical photovoltaic turbine group data set of the photovoltaic turbine group and the historical energy storage turbine group data set of the energy storage turbine group are obtained; the historical photovoltaic turbine group data set is input into a preset long short-term memory network for training to obtain a target photovoltaic turbine group inertia evaluation model; and the historical energy storage turbine group data set is input into the preset long short-term memory network for training to obtain a target energy storage turbine group inertia evaluation model.

[0016] The wind, light and storage station inertia evaluation method provided by the application trains a preset long short-term memory network and builds a target photovoltaic machine group inertia evaluation model and a target energy storage machine group inertia evaluation model suitable for photovoltaic and energy storage time sequence response characteristics through historical photovoltaic machine group data sets and historical energy storage machine group data sets of the photovoltaic machine group, avoids the error caused by simplifying the photovoltaic and energy storage time sequence characteristics by using unified mathematical derivation in the prior art, improves the accuracy of the photovoltaic and energy storage machine group inertia evaluation, and provides high-quality sub-machine group inertia data for station-level aggregation evaluation.

[0017] In an optional implementation, the historical photovoltaic machine group data set is input into the preset long short-term memory network for training to obtain the target photovoltaic machine group inertia evaluation model, including:

[0018] The historical photovoltaic machine group data set is input into the input layer of the preset long short-term memory network for processing to obtain a normalized photovoltaic machine group time sequence characteristic matrix; the photovoltaic machine group time sequence characteristic matrix is input into the convolution layer of the preset long short-term memory network for processing to obtain a local space-time characteristic matrix; the local space-time characteristic matrix is input into the hidden layer of the preset long short-term memory network for processing to obtain a target feature vector; the target feature vector is input into the full connection layer of the preset long short-term memory network for processing to obtain an inertia feature vector; and the inertia feature vector is input into the output layer of the preset long short-term memory network for processing to obtain a plurality of second historical equivalent inertia evaluation values of the photovoltaic machine group; the historical photovoltaic machine group data set is taken as input, and the plurality of second historical equivalent inertia evaluation values are taken as output to build the target photovoltaic machine group inertia evaluation model.

[0019] The wind, light and storage station inertia evaluation method provided by the application optimizes the target photovoltaic machine group inertia evaluation model through fine processing of each layer of the preset long short-term memory network, so that the model can accurately learn the dynamic influence of factors such as irradiance and temperature on photovoltaic inertia response, improves the accuracy of the second historical equivalent inertia evaluation value, further guarantees the reliability of real-time inertia evaluation of the photovoltaic machine group, and solves the problem that the traditional method is difficult to adapt to the time sequence characteristics of photovoltaic output fluctuation.

[0020] In an optional implementation, the method further includes:

[0021] obtaining a plurality of first historical equivalent inertia evaluation values of a wind turbine machine group, a plurality of second historical equivalent inertia evaluation values of a photovoltaic machine group, a plurality of third historical equivalent inertia evaluation values of an energy storage machine group, and a historical station feature set of a wind, light and storage station;

[0022] obtaining a real inertia evaluation value of the wind, light and storage station; and based on the historical station feature set, the plurality of first historical equivalent inertia evaluation values, the plurality of second historical equivalent inertia evaluation values and the plurality of third historical equivalent inertia evaluation values, taking the real inertia evaluation value as a label, and using a preset full connection neural network to build a wind, light and storage station inertia evaluation aggregation model.

[0023] The wind-solar-storage power station inertia evaluation method provided by the application, by using real inertia as a label, integrates historical sub-machine group inertia and power station characteristics with a fully connected neural network, constructs an aggregation model that can learn the dynamic aggregation effect of each sub-machine group inertia and the influence of power station level characteristics, provides an adaptive model architecture for power station level inertia evaluation, and improves the integrity and accuracy of overall inertia evaluation of the power station.

[0024] In an optional implementation, based on the historical power station characteristic set, the plurality of first historical equivalent inertia evaluation values, the plurality of second historical equivalent inertia evaluation values and the plurality of third historical equivalent inertia evaluation values, a wind-solar-storage power station inertia evaluation aggregation model is constructed by using a preset fully connected neural network with the real inertia evaluation value as a label, including:

[0025] The historical power station characteristic set, the plurality of first historical equivalent inertia evaluation values, the plurality of second historical equivalent inertia evaluation values and the plurality of third historical equivalent inertia evaluation values are spliced to obtain a power station aggregation feature matrix; the power station aggregation feature matrix is input into a preset fully connected neural network for processing to obtain a historical inertia evaluation value of the wind-solar-storage power station; the real inertia evaluation value is used as a label, the historical power station characteristic set, the plurality of first historical equivalent inertia evaluation values, the plurality of second historical equivalent inertia evaluation values and the plurality of third historical equivalent inertia evaluation values are used as input, and the historical inertia evaluation value is used as output to construct the wind-solar-storage power station inertia evaluation aggregation model.

[0026] The wind-solar-storage power station inertia evaluation method provided by the application, by feature splicing, obtains an aggregation matrix that fuses sub-machine group inertia and power station characteristics, provides a comprehensive input for the aggregation model, and avoids evaluation deviation caused by scattered features. Further, by preset fully connected neural network processing and outputting the historical inertia evaluation value of the power station, an output basis is provided for the aggregation model, ensuring that the model can accurately learn the relationship between multiple characteristics and power station inertia, further improving the accuracy of the historical inertia evaluation value of the power station, providing reliable support for real-time output of the target inertia evaluation value of the power station, and completely solving the problem that traditional methods cannot integrate multi-dimensional feature evaluation of the overall inertia of the power station.

[0027] In a second aspect, the application provides a wind-solar-storage power station inertia evaluation device, the wind-solar-storage power station including a wind turbine group, a photovoltaic group and an energy storage group; the device includes:

[0028] The first acquisition module is configured to acquire a real-time wind turbine group data set of the wind turbine group, a real-time photovoltaic turbine group data set of the photovoltaic turbine group, a real-time energy storage turbine group data set of the energy storage turbine group, and a real-time station feature set of the wind-solar-storage station; the first processing module is configured to input the real-time wind turbine group data set into a target wind turbine group inertia evaluation model for processing to obtain a plurality of first real-time equivalent inertia evaluation values of the wind turbine group; the second processing module is configured to input the real-time photovoltaic turbine group data set into a target photovoltaic turbine group inertia evaluation model for processing to obtain a plurality of second real-time equivalent inertia evaluation values of the photovoltaic turbine group; the third processing module is configured to input the real-time energy storage turbine group data set into a target energy storage turbine group inertia evaluation model for processing to obtain a plurality of third real-time equivalent inertia evaluation values of the energy storage turbine group; and the fourth processing module is configured to input the real-time station feature set, the plurality of first real-time equivalent inertia evaluation values, the plurality of second real-time equivalent inertia evaluation values, and the plurality of third real-time equivalent inertia evaluation values into a wind-solar-storage station inertia evaluation aggregation model for processing to obtain a target inertia evaluation value of the wind-solar-storage station.

[0029] In a third aspect, the present application provides a computer readable storage medium having stored thereon computer instructions for causing a computer to execute the wind-solar-storage station inertia evaluation method of the first aspect or any of the corresponding embodiments thereof.

[0030] In a fourth aspect, the present application provides a computer program product comprising computer instructions for causing a computer to execute the wind-solar-storage station inertia evaluation method of the first aspect or any of the corresponding embodiments thereof. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the drawings needed in the specific embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0032] Figure 1 is a flowchart of the wind-solar-storage station inertia evaluation method according to an embodiment of the present application;

[0033] Figure 2 is a structural block diagram of the wind-solar-storage station inertia evaluation device according to an embodiment of the present application;

[0034] Figure 3 is a hardware structure schematic diagram of the computer device of an embodiment of the present application. DETAILED DESCRIPTION

[0035] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0036] The embodiment of the present application provides a wind-solar-storage power station inertia evaluation method, through a target wind turbine group inertia evaluation model, a target photovoltaic turbine group inertia evaluation model and a target energy storage turbine group inertia evaluation model, the nonlinear mapping rules of the real-time inertia of the wind, solar and storage turbine groups and the frequency change rate of the power station level are learned respectively, and the real-time power station feature set of the wind-solar-storage power station is integrated, so that the overall inertia evaluation value of the power station can be accurately output, and the errors caused by the simplification processing of the prior art are effectively eliminated, and the evaluation accuracy under abnormal weather is improved.

[0037] According to the embodiment of the present application, a wind-solar-storage power station inertia evaluation method is provided. It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in a different order.

[0038] In the present embodiment, a wind-solar-storage power station inertia evaluation method is provided, which can be used in electronic devices such as computers, mobile phones, tablet computers, etc. The wind-solar-storage power station includes a wind turbine group, a photovoltaic turbine group and an energy storage turbine group.

[0039] Figure 1 The flowchart of the wind-solar-storage power station inertia evaluation method according to the embodiment of the present application is shown in FIG. 1, which includes the following steps: Figure 1

[0040] In step S101, the real-time wind turbine group data set of the wind turbine group, the real-time photovoltaic turbine group data set of the photovoltaic turbine group, the real-time energy storage turbine group data set of the energy storage turbine group and the real-time power station feature set of the wind-solar-storage power station are obtained.

[0041] The real-time wind turbine group data set represents a multi-dimensional data set collected in real time for reflecting the running state of the wind turbine group and the environmental influence factors, which can include environmental and geographical feature data and running and response feature data.

[0042] Further, the environmental and geographical feature data can include real-time numerical weather forecast information (such as real-time wind speed, real-time wind direction) of the area where the wind turbine group is located, and fixed geographical position information (latitude, longitude) of each wind turbine and real-time pitch angle.

[0043] ​Further, the operation and response characteristic data can include real-time active power output value of each wind turbine, real-time frequency deviation of the grid connection point (measured at the wind turbine end), and real-time frequency change rate (measured at the wind turbine end), which can reflect the dynamic response of the wind turbine in the virtual inertia control process.

[0044] Further, the real-time photovoltaic cluster dataset represents a collection of multi-dimensional data collected in real time for reflecting the operation state of the photovoltaic cluster and the illumination influencing factor, which can include real-time irradiance (unit: W / ㎡) of the area where each photovoltaic array is located, real-time ambient temperature (unit: ℃) and other environmental and equipment characteristic data, real-time active power output value of each photovoltaic array, real-time frequency deviation of the grid connection point (measured at the photovoltaic array end), and real-time frequency change rate (measured at the photovoltaic array end) for capturing the inertia response characteristic of the photovoltaic under illumination fluctuation.

[0045] Further, the real-time energy storage cluster dataset represents a collection of multi-dimensional data collected in real time for reflecting the operation state of the energy storage cluster and the charging and discharging capacity, which can include real-time SOC (state of charge, unit: %) and real-time terminal voltage (unit: V) of each electrochemical energy storage bin and other equipment state characteristic data, real-time charging and discharging power (charging is negative and discharging is positive, unit: kW) of each electrochemical energy storage bin, real-time frequency deviation of the grid connection point (measured at the energy storage bin end), and real-time frequency change rate (measured at the energy storage bin end) for reflecting the dynamic process of the energy storage in the virtual inertia control.

[0046] Further, the real-time station characteristic set represents a collection of data collected in real time for reflecting the overall operation state of the wind-solar-storage station and the interaction with the power grid, which can include real-time grid frequency (unit: Hz) and grid connection point voltage amplitude (unit: kV) measured at the grid connection point of the station and external power grid, overall active power output (algebraic sum of active power of all wind turbine clusters, photovoltaic clusters, and energy storage clusters, unit: MW) and overall reactive power output (algebraic sum of reactive power of all clusters, unit: Mvar) of the station, real-time dispatching instructions (such as upper / lower limit of total active power output of the station, reactive power regulation range) issued by the upper-level power grid dispatching institution, and real-time operation limit value (such as output constraint under extreme weather) set internally in the station, and dispatching control instruction data.

[0047] Specifically, the original wind power dataset of the wind turbine cluster, the real-time photovoltaic cluster dataset of the photovoltaic cluster, the original energy storage dataset of the energy storage cluster, and the original station characteristic set of the wind-solar-storage station can be collected.

[0048] Further, the collected raw data can also be screened to remove invalid or abnormal data (such as active power drop to 0 caused by fan failure, irradiance exceeding the physically reasonable range (0-1200W / ㎡) caused by sensor error, SOC less than 0 or greater than 100% of energy storage, etc.), and a small amount of missing data is filled by adjacent time data interpolation method to ensure data set integrity.

[0049] Further, the corresponding real-time wind turbine group data set, real-time photovoltaic turbine group data set, real-time energy storage turbine group data set and real-time station feature set of the wind-solar-storage station are obtained by preprocessing.

[0050] In step S102, the real-time wind turbine group data set is input into the target wind turbine group inertia evaluation model for processing to obtain a plurality of first real-time equivalent inertia evaluation values of the wind turbine group.

[0051] The target wind turbine group inertia evaluation model represents an inertia evaluation model specially adapted to the characteristics of the wind turbine group based on a graph convolutional neural network. In this embodiment, the graph convolutional neural network GCN is a direction-enhanced graph convolution DE-GCN, which combines the complex spatial topological structure (such as the geographical position of the wind turbine and the wake effect) and the aerodynamic coupling characteristics of the wind turbine group. By learning the correlation rules between wind turbines and the inertia response characteristics of the wind turbines in historical wind data (including environmental, operating and response data), the real-time equivalent inertia of the wind turbine group can be accurately evaluated, and the equivalent inertia evaluation value of each wind turbine can be output, rather than simply equivalent the wind turbine group to a single wind turbine for evaluation.

[0052] Specifically, the real-time wind turbine group data set reflecting the real-time operating state (such as real-time wind speed and active power) and environmental characteristics (such as real-time wind direction) of the wind turbine group is input into the target wind turbine group inertia evaluation model which has been trained. The model can dynamically respond to the real-time operating changes (such as sudden changes in wind speed and wind direction adjustment) of the wind turbine group.

[0053] Further, the target wind turbine group inertia evaluation model can automatically extract the spatial correlation characteristics (such as the wake effect between different wind turbines) and real-time inertia response characteristics (such as the correlation between active power and frequency change rate) of the wind turbine group in real-time data through the built-in DE-GCN architecture. After convolution, feature aggregation and other calculations, the real-time equivalent inertia evaluation value corresponding to each wind turbine in the wind turbine group is output, i.e. a plurality of first real-time equivalent inertia evaluation values.

[0054] Further, compared with the simplified processing method of the prior art which equates the wind turbine group to a single wind turbine and ignores the spatial correlation and aerodynamic coupling between wind turbines, the GCN model adapted to the characteristics of the wind turbine in this embodiment can accurately capture the influence of the wake effect and geographical position correlation between wind turbines on inertia, significantly reducing the evaluation error caused by the simplified processing.

[0055] Step S103, input the real-time photovoltaic plant data set into the target photovoltaic plant inertia evaluation model for processing, to obtain a plurality of second real-time equivalent inertia evaluation values of the photovoltaic plant.

[0056] The target photovoltaic plant inertia evaluation model is a special inertia evaluation model based on a long short-term memory network (LSTM) and integrated with a one-dimensional convolution layer. The model can adapt to the characteristics of the photovoltaic plant, such as the centralized arrangement, time-dependent inertia response, and output fluctuation correlation between arrays. Through learning the correlation between photovoltaic output and inertia response in historical photovoltaic data (irradiance, temperature, active power, etc.), the model can accurately process the inertia dynamic changes of the photovoltaic plant under different light and temperature conditions, and finally output the equivalent inertia evaluation values of each array or the whole photovoltaic plant, avoiding the problem that the traditional simplified model cannot adapt to the time-series response characteristics of photovoltaic.

[0057] Specifically, the real-time photovoltaic plant data set reflecting the real-time operation state (such as the active power of each array) and environmental characteristics (such as real-time irradiance and temperature) of the photovoltaic plant is input into the target photovoltaic plant inertia evaluation model which has been trained.

[0058] Further, the target photovoltaic plant inertia evaluation model can automatically extract the local spatio-temporal correlation features (such as the output fluctuation coupling between adjacent arrays) and the time-series features (such as the inertia dynamic adjustment under sudden irradiance changes) of the inertia response through the built-in "one-dimensional convolution layer + LSTM" architecture, and output the real-time equivalent inertia evaluation values of each array in the photovoltaic plant, i.e., a plurality of second real-time equivalent inertia evaluation values.

[0059] Step S104, input the real-time energy storage plant data set into the target energy storage plant inertia evaluation model for processing, to obtain a plurality of third real-time equivalent inertia evaluation values of the energy storage plant.

[0060] The target energy storage plant inertia evaluation model is a model based on a long short-term memory network (LSTM) and suitable for the energy storage plant with fast charge-discharge response speed and strong correlation between inertia contribution and SOC (state of charge). Further, the model can learn the mapping relationship between energy storage dynamic response and inertia output in historical energy storage data (charge-discharge power, SOC, terminal voltage, etc.), capture the fast inertia adjustment process of the energy storage plant under grid frequency fluctuation, accurately reflect the inertia contribution of the energy storage under different SOC states and charge-discharge modes, and finally output the equivalent inertia evaluation values of each energy storage bin or the whole energy storage plant, solving the problem that the traditional model cannot adapt to the fast dynamic characteristics of the energy storage.

[0061] Specifically, the real-time energy storage machine group data set reflecting the real-time operation state of the energy storage machine group (such as the charging and discharging power and SOC of each energy storage warehouse) is input into the trained target energy storage machine group inertia evaluation model.

[0062] Further, the target energy storage machine group inertia evaluation model can automatically capture the rapid charging and discharging response process of the energy storage under the frequency fluctuation of the power grid through the fast time sequence learning ability of the LSTM architecture, thereby accurately calculating the inertia contribution of each energy storage warehouse under different SOC and charging and discharging power, and finally outputting the real-time equivalent inertia evaluation value corresponding to each energy storage warehouse in the energy storage machine group, i.e., multiple third real-time equivalent inertia evaluation values.

[0063] Step S105, input the real-time station feature set, the multiple first real-time equivalent inertia evaluation values, the multiple second real-time equivalent inertia evaluation values and the multiple third real-time equivalent inertia evaluation values into the wind-solar-storage station inertia evaluation aggregation model for processing to obtain the target inertia evaluation value of the wind-solar-storage station.

[0064] The wind-solar-storage station inertia evaluation aggregation model represents a station-level inertia integration evaluation model constructed based on a fully connected neural network (MLP), which can fuse the real-time inertia evaluation results (first, second and third real-time equivalent inertia evaluation values) of the wind power machine group, the photovoltaic machine group and the energy storage machine group and the station-level operation features (such as the grid connection point frequency, total output and dispatching instruction), and learn the dynamic aggregation effect of the inertia of each sub-machine group at the grid connection point and the influence of the station-level features on the overall inertia, and finally output the equivalent inertia evaluation value of the overall wind-solar-storage station, realizing the integration from “sub-machine group evaluation” to “overall station evaluation” and filling the gap of the traditional technology which cannot evaluate the overall station inertia.

[0065] Specifically, the real-time station feature set reflecting the global operation state of the station is integrated with the multiple first real-time equivalent inertia evaluation values of the wind power machine group, the multiple second real-time equivalent inertia evaluation values of the photovoltaic machine group and the multiple third real-time equivalent inertia evaluation values of the energy storage machine group, and is input into the trained wind-solar-storage station inertia evaluation aggregation model.

[0066] Further, the wind-solar-storage station inertia evaluation aggregation model can learn the dynamic superposition effect of the inertia of each sub-machine group at the grid connection point (such as the coordinated provision of inertia by wind power and energy storage) and the influence of the station-level features (such as the dispatching instruction and the grid frequency) on the overall inertia through the nonlinear feature fusion ability of the built-in fully connected neural network, and output the equivalent inertia evaluation value of the overall wind-solar-storage station, i.e., the target inertia evaluation value, after calculation.

[0067] The inertia evaluation method of the wind-solar-storage station provided by the embodiment can cover the full-dimensional real-time operation characteristics of wind, light, storage and station level, and adapt to the evaluation needs under normal and abnormal weather, by obtaining the real-time wind turbine group data set of the wind turbine group, the real-time photovoltaic turbine group data set of the photovoltaic turbine group, the real-time energy storage turbine group data set of the energy storage turbine group and the real-time station feature set of the wind-solar-storage station. Further, the data of each turbine group is processed by the target wind turbine group inertia evaluation model, the target photovoltaic turbine group inertia evaluation model and the target energy storage turbine group inertia evaluation model respectively, and then the final evaluation is carried out through the wind-solar-storage station inertia evaluation aggregation model integration, which fully considers the nonlinear relationship between the output active power and the frequency change rate of the grid connection point in the virtual inertia control of the new energy station, significantly reduces the evaluation error caused by the simplification processing, and especially in the abnormal weather such as strong wind, cloudy and blowing sand, still can maintain high evaluation accuracy. At the same time, by fusing the real-time inertia evaluation results of each sub-turbine group and the station features through the wind-solar-storage station inertia evaluation aggregation model, the integration from single turbine group inertia evaluation to overall station inertia evaluation is realized, which provides comprehensive and accurate station level inertia data support for power grid dispatching, frequency stability control and the like. Further, compared with the traditional mathematical method, the multiple evaluation models and aggregation models constructed by deep learning can more accurately capture the complex nonlinear relationship between the station inertia and the running state, avoid the defects that the traditional mathematical analysis method is difficult to represent the nonlinear characteristics, and greatly reduce the evaluation deviation.

[0068] In some optional embodiments, the target wind turbine group inertia evaluation model in the above step S102 is obtained through the following steps:

[0069] Step a1, obtaining a historical wind turbine group data set of the wind turbine group. The specific process can refer to the description of the real-time wind turbine group data set in the above step S101, which will not be repeated here.

[0070] Step a2, constructing an initial directed graph and an initial node feature matrix according to the historical wind turbine group data set.

[0071] Specifically, the information reflecting the spatial correlation (wake effect, geographical position) and equipment characteristics of the wind turbine group in the historical wind data can be used to abstract the wind turbine group into a directed graph structure with quantitative correlation, and each wind turbine node is given multi-dimensional features, thereby forming an input format that can be directly processed by a subsequent graph convolution network (GCN), avoiding the evaluation deviation caused by the traditional model ignoring the complex correlation between wind turbines.

[0072] Firstly, the following data can be selected from the historical wind turbine group data set:

[0073] (1) Wind turbine geographical position and meteorological data, which can include the historical latitude longitude (λ i), as well as the historical dominant wind direction (the high-frequency wind direction range counted by month, such as the dominant wind direction if the northerly wind accounts for the highest proportion in the current month), and the historical record of the orientation of the wind turbine blades.

[0074] (2) Data on wind turbine operation and wake correlation, which may include the historical active power of each wind turbine and the historical wake influence range (obtained from the wind turbine manufacturer's technical manual, such as the wake influence radius).

[0075] Secondly, the mathematical expression of the initial directed graph is defined as G = (V, E, A). Where V represents the set of N graph nodes of the wind turbine group, each node representing a wind turbine; E represents the set of graph edges of the wind turbine group; and A represents the adjacency matrix of the wind turbine group, which is the quantification of the spatial dependencies between the N graph nodes.

[0076] Each wind turbine in the wind turbine cluster can be viewed as a node in a directed graph G. If the wind turbine cluster has N turbines, then the node set V = {v1, v2, ..., v...} N}, where each node v i This corresponds to a single wind turbine and is used to identify the wind turbine's role in the directed graph.

[0077] Furthermore, taking the prevailing wind direction of the month as a reference direction and considering the wake effect of the wind turbine, the positive direction of the directed side is defined as the direction from the upwind direction to the downwind direction.

[0078] Furthermore, for any two wind turbines i and j, the distance d between them can be calculated using the following relationship (1). ij :

[0079]

[0080] In the formula: Indicates the dimension of wind turbine i; λ represents the dimension of wind turbine j; j Indicates the longitude of fan j; λ i R represents the longitude of wind turbine i; R represents the radius of the Earth.

[0081] Furthermore, if and only if wind turbine i is upwind of wind turbine j, and the distance d between the two wind turbines is... ij When the influence range of the wake effect is satisfied, construct a directed edge e. ij ∈E, the direction is from upwind to downwind, that is, from wind turbine i to wind turbine j; if the wake effect condition is not met, then this directed edge does not exist. In this way, the graph edge set E of the wind turbine group is constructed.

[0082] Furthermore, the definition of each element in the adjacency matrix A is shown in the following relation (2):

[0083]

[0084] wherein β represents a wake weight distribution coefficient, which is a learnable parameter; σ ω represents a wake attenuation coefficient, which needs to be provided by the fan manufacturer.

[0085] Further, the feature information of each fan can be taken as a node feature vector, which can include wind speed, wind direction, real-time active power, and local inertia response coefficient (initial value is 0, which is updated by the network subsequently).

[0086] Further, if there are N fans, the feature vector of each fan is X wind ={x1,x2,…,x N}, wherein x i is the feature vector of the i-th fan.

[0087] Further, the feature vectors of the N fans are arranged in rows to form a corresponding initial node feature matrix, ensuring that each element in the matrix corresponds to the historical features of the corresponding fan.

[0088] Step a3, based on the historical wind turbine cluster dataset, the initial directed graph and the initial node feature matrix, a graph convolutional network is used for model training and construction of the target wind turbine cluster inertia evaluation model.

[0089] wherein the graph convolutional network (GCN) represents a deep learning model for processing graph structure data, which can utilize the topological structure of the graph (such as the connection relationship between nodes) and the node features, and update the feature representation of the current node by convolution operation to aggregate the information of adjacent nodes. In the wind turbine cluster and other scenarios of the present embodiment, the spatial dependence relationship between wind turbines due to wake effect, electrical distance, etc. can be effectively captured, which is helpful for the evaluation of the inertia and other characteristics of the wind turbine cluster.

[0090] Specifically, the above step a3 includes:

[0091] Step a31, obtaining a target directional adjacency matrix and a target weight matrix.

[0092] Specifically, referring to the historical wind direction data of the wind turbine cluster, the 360° wind direction can be divided into K partitions, such as 0°-120°, 120°-240°, and 240°-360°. Each partition corresponds to a directional adjacency matrix B k (k = 1, 2, …, k).

[0093] Further, for each pair of fans i and j, it is determined whether the azimuth angle falls within the k-th wind direction partition, and if it falls within the partition and there is a wake effect, B k,ij = 1; otherwise B k,ij= 0, thereby obtaining K target directional adjacency matrices of N x N.

[0094] Further, the target weight matrix can include a basic weight matrix W (l) and a directional weight matrix wherein, l represents the number of network layers, and is initially W (0) ,

[0095] Further, the basic weight matrix W (l) has a dimension of D x F (D is the node feature dimension, and F is the convolutional layer output feature dimension, which is set by a hyperparameter); the directional weight matrix also has a dimension of D x F, and both are randomly generated initial values by the Xavier initialization method.

[0096] Step a32, inputting the initial directed graph and the initial node feature matrix into the input layer of the graph convolutional network to obtain a regularized target directed graph and a normalized target node feature matrix.

[0097] Specifically, the adjacency matrix of the initial directed graph is extracted, and the adjacency matrix is regularized to obtain a regularized adjacency matrix, and then the node set V and the directed edge set E of the initial directed graph are combined to form a regularized target directed graph.

[0098] Further, each type of feature (wind speed, wind direction, real-time active power, and local inertia response coefficient) in the initial node feature matrix is respectively processed by Min-Max normalization, and the feature values are mapped to the interval [0, 1], thereby obtaining a normalized target node feature matrix.

[0099] Step a33, inputting the target directed graph, the target node feature matrix, the target directional adjacency matrix, and the target weight matrix into the convolutional layer of the graph convolutional network to obtain an enhanced initial node embedding matrix.

[0100] Specifically, the input parameters of the convolutional layer can include the regularized adjacency matrix of the target directed graph, the target node feature matrix, K target directional adjacency matrices, the basic weight matrix, and the directional weight matrix.

[0101] Further, the directional enhancement type graph convolutional calculation can be performed according to the following relationship (3):

[0102]

[0103] In the formula, H (l) represents the initial node embedding matrix of the lth layer. denotes the regularized adjacency matrix in the target directed graph; and denotes a nonlinear activation function, and in the embodiment, Exponential Linear Unit (ELU) is selected as the activation function considering that inertia evaluation is sensitive to noise.

[0104] wherein the regularized adjacency matrix is The aggregated adjacency node features are multiplied by the basic weight matrix W (l) to realize linear mapping of the node features and fusion of the spatial correlation.

[0105] Further, for each wind direction partition k, the directional adjacency matrix B k is extracted to extract the correlation features between the wind direction partitions, and multiplied by the directional weight matrix W to realize enhanced fusion of the different wind direction features.

[0106] Further, after the two parts of results are added, an Exponential Linear Unit (ELU) activation function is substituted to obtain an enhanced initial node embedding matrix H (l+1) .

[0107] Further, the initial node embedding matrix H (0) output by the convolutional layer can be obtained through the above process.

[0108] In step a34, the initial node embedding matrix, the target directional adjacency matrix, and the target weight matrix are input into the hidden layer of the graph convolutional network to obtain a target node embedding matrix.

[0109] Specifically, the hidden layer of the graph convolutional network is set to three layers of directional enhancement type graph convolution (DE-GCN).

[0110] Further, for each layer, the above relationship (3) can be calculated.

[0111] First, the spatial correlation features and the wind direction correlation features of the node embedding of the previous layer are aggregated by using the regularized adjacency matrix A and the directional adjacency matrix B k , and then the features are linearly mapped by the weight matrix W (l) .

[0112] Second, after the calculation of each layer is completed, a Dropout layer (with a probability of 0.2-0.5) is added to randomly close the feature connection of part of the nodes to perform overfitting suppression.

[0113] Then, an ELU activation function is used to obtain the node embedding matrix H (l+1) output by each layer.

[0114] ​Finally, after processing through 3 hidden layers, the node embedding matrix H output by the 4th layer is obtained. (4) That is, the target node embedding matrix.

[0115] Step a35: Input the target node embedding matrix and the preset fully connected layer weight matrix into the output layer of the graph convolutional network for processing to obtain multiple first historical equivalent inertia evaluation values ​​in the wind turbine group.

[0116] Specifically, the output layer employs a two-layer fully connected network, and its input parameters may include the target node embedding matrix and a preset fully connected layer weight matrix (W of the first layer). fc1 : F×M, where M is the intermediate feature dimension; second layer W fc2 : M×1) and the corresponding bias term b fc1 b fc2 The initial values ​​of the weights and biases are generated through Xavier initialization.

[0117] First, in the first fully connected layer, according to F... mid =σ ReLU (H (4) W fc1 +b fc1 Calculate and obtain the N×M intermediate feature matrix F. mid Among them, σ ReLU (x) = max(0,x) represents the ReLU activation function.

[0118] Secondly, in the second fully connected layer, according to F out =F mid W fc2 +b fc2 Calculate and obtain an N×1 vector, where each element corresponds to the calculated equivalent inertia value of a wind turbine.

[0119] Then, the loss function shown in the following relation (4) is introduced:

[0120]

[0121] In the formula: H represents the calculated equivalent inertia of the i-th wind turbine obtained from the hidden layer. i The actual equivalent inertia of the i-th wind turbine can be obtained from the wind turbine manufacturer; H ref λ represents the theoretical inertia reference value of the wind turbine group, which can be obtained from the wind turbine manufacturer; λ represents the balance hyperparameter.

[0122] Finally, the loss value L is calculated using the backpropagation algorithm, and the weight matrices of the output layer and the preceding layers are updated until the loss value converges. Finally, N converged wind turbine inertia evaluation values ​​are output, which are multiple first historical equivalent inertia evaluation values.

[0123] Step a36, taking the historical wind turbine cluster dataset as input and a plurality of first historical equivalent inertia evaluation values as output, constructing a target wind turbine cluster inertia evaluation model.

[0124] Specifically, taking the historical wind turbine cluster dataset as model input data and a plurality of first historical equivalent inertia evaluation values as model output data, the model is trained until a trained target wind turbine cluster inertia evaluation model is obtained.

[0125] In some optional embodiments, the target photovoltaic cluster inertia evaluation model in step S103 and the target energy storage cluster inertia evaluation model in step S104 are obtained by the following steps:

[0126] Step b1, obtaining a historical photovoltaic cluster dataset of a photovoltaic cluster and a historical energy storage cluster dataset of an energy storage cluster. The specific process can refer to the description of the real-time photovoltaic cluster dataset and the real-time energy storage cluster dataset in step S101 above, which will not be described here.

[0127] Step b2, inputting the historical photovoltaic cluster dataset into a preset long short-term memory network for training to obtain a target photovoltaic cluster inertia evaluation model.

[0128] Wherein, the preset long short-term memory network (Long Short-Term Memory, LSTM) represents an improved recurrent neural network (RNN), mainly used for processing time series data, and can control the flow of information through the gating mechanism (input gate, forget gate, output gate), thereby effectively solving the gradient disappearance or gradient explosion problem of traditional RNN in long sequence training process, and is good at capturing long-term dependencies in data. In the photovoltaic and energy storage cluster scenario in this embodiment, since the inertia response sequence has time dependence, LSTM can learn the time sequence characteristics of photovoltaic and energy storage cluster inertia related data, and realize the evaluation of its inertia.

[0129] Specifically, the above step b2 includes:

[0130] Step b21, inputting the historical photovoltaic cluster dataset into the input layer of the preset long short-term memory network for processing to obtain a normalized photovoltaic cluster time sequence feature matrix.

[0131] Specifically, the historical photovoltaic cluster dataset, the frequency deviation of the grid connection point and the frequency change rate can be input into the input layer of the preset long short-term memory network and normalized. Wherein, the Min-Max normalization method can be used.

[0132] Further, all normalized data are arranged according to time sequence and feature dimension to form a normalized photovoltaic cluster time sequence feature matrix.

[0133] Further, by normalization processing, the dimensional and numerical range differences between different photovoltaic data (such as active power, irradiance, temperature, etc.) can be eliminated, the data distribution is more uniform, the learning efficiency and stability of the LSTM network for data are improved, and the model training deviation caused by large data magnitude gap is avoided.

[0134] Step b22, inputting the photovoltaic plant time sequence feature matrix into the convolution layer of the preset long short-term memory network to obtain a local space-time feature matrix.

[0135] Specifically, a one-dimensional convolution layer is added in the LSTM, and appropriate convolution kernel size, number and other parameters are set.

[0136] Further, the convolution kernel slides on the photovoltaic plant time sequence feature matrix, and extracts local features through convolution operation. Then, the convolution result is processed and the non-linear expression ability of the features is enhanced by using an activation function.

[0137] Further, after the convolution and activation operation on the entire photovoltaic plant time sequence feature matrix, a series of new feature maps are obtained and combined to form the final local space-time feature matrix.

[0138] Further, by using the one-dimensional convolution layer to extract features from the input photovoltaic plant time sequence feature matrix, the local space-time coupling characteristics inside the photovoltaic plant can be captured, and more valuable input is provided for the subsequent LSTM hidden layer to further learn deep features.

[0139] Step b23, inputting the local space-time feature matrix into the hidden layer of the preset long short-term memory network to obtain a target feature vector.

[0140] By using the hidden layer with a double-layer bidirectional LSTM structure to process the local space-time feature matrix, the inertia response characteristics of the photovoltaic plant before and after the frequency disturbance can be captured, and the feature sequence containing multiple time steps can be converted into a target feature vector representing the overall characteristics.

[0141] Specifically, the hidden layer adopts a double-layer bidirectional LSTM structure, the first layer LSTM extracts features in the forward and reverse directions according to the time steps of the local space-time feature matrix, and captures features in different time directions; the second layer LSTM further deepens the learning of the time sequence features based on the first layer.

[0142] Further, the output of the last time step integrates the information of all previous time steps, so the output of the last time step is taken as the output of the hidden layer and a target feature vector is obtained.

[0143] Further, by mining the time dependence and deep temporal characteristics of the inertia response of the photovoltaic plant group, the characteristics of multiple time steps in the local space-time feature matrix are integrated into a more representative target feature vector, facilitating the mapping of the inertia evaluation space by the subsequent fully connected layer.

[0144] Step b24, input the target feature vector into the fully connected layer of the preset long short-term memory network to obtain an inertia feature vector.

[0145] Specifically, the fully connected layer selects a rectified linear unit (ReLU) as an activation function.

[0146] Further, after inputting the target feature vector into the fully connected layer, the inertia feature vector can be finally obtained through weight matrix multiplication and ReLU activation function processing, and then through the dropout operation to prevent overfitting.

[0147] Further, by converting the deep features extracted by the hidden layer into inertia evaluation related feature representations, the model generalization ability is improved by providing a direct basis for the output layer to calculate the inertia evaluation value and adopting dropout (randomly discarding part of the neuron connections) to prevent overfitting.

[0148] Step b25, input the inertia feature vector into the output layer of the preset long short-term memory network to obtain a plurality of second historical equivalent inertia evaluation values of the photovoltaic plant group.

[0149] Specifically, the loss function of the output layer is defined as shown in the following relation (5):

[0150] L = |H out -H cal |(5)

[0151] In the formula: H out represents the inertia feature vector output by the fully connected layer; H cal represents the inertia calculation result, which can be independently calculated based on the conventional inertia evaluation method using the input data, and is used as the true label for subsequent model training.

[0152] Further, after inputting the inertia feature vector into the output layer, H out is calculated, then the loss value is calculated according to the loss function, the parameters of each layer of the network are updated through the back propagation algorithm, the model is continuously optimized, and finally a plurality of second historical equivalent inertia evaluation values of the photovoltaic plant group are obtained.

[0153] Step b26, input the historical photovoltaic plant data set as input and output a plurality of second historical equivalent inertia evaluation values to construct a target photovoltaic plant inertia evaluation model.

[0154] Specifically, the historical photovoltaic plant data set is taken as the model input data, and a plurality of second historical equivalent inertia evaluation values are taken as the model output data for model training until a trained target photovoltaic plant inertia evaluation model is obtained.

[0155] In step b3, the historical energy storage plant data set is input into a preset long short-term memory network for training to obtain a target energy storage plant inertia evaluation model. The specific process can refer to the process of constructing the target photovoltaic plant inertia evaluation model in step b2 described above, and will not be described here.

[0156] In some optional embodiments, the inertia evaluation aggregation model of the wind-solar-storage station in step S105 is obtained by the following steps:

[0157] In step c1, a plurality of first historical equivalent inertia evaluation values of the wind power plant, a plurality of second historical equivalent inertia evaluation values of the photovoltaic plant, a plurality of third historical equivalent inertia evaluation values of the energy storage plant, and a historical station feature set of the wind-solar-storage station are obtained.

[0158] The specific process can refer to the processes of steps S102 to S104 described above, and will not be described here.

[0159] In step c2, a real inertia evaluation value of the wind-solar-storage station is obtained.

[0160] Specifically, according to the description in step b25, the real inertia evaluation value H of the wind-solar-storage station can be calculated independently based on the conventional inertia evaluation method using the input data. cal .

[0161] In step c3, based on the historical station feature set, the plurality of first historical equivalent inertia evaluation values, the plurality of second historical equivalent inertia evaluation values, and the plurality of third historical equivalent inertia evaluation values, a real inertia evaluation value is taken as a label, and a preset multilayer perceptron (MLP) neural network is used to construct an inertia evaluation aggregation model of the wind-solar-storage station.

[0162] The preset multilayer perceptron (MLP) neural network represents a neural network structure composed of an input layer, a plurality of hidden layers, and an output layer, and the neurons between the layers are fully connected, which is used to learn the dynamic aggregation effect of the inertia contribution of the wind power plant, the photovoltaic plant, and the energy storage plant at the point of common coupling, and the influence of the station-level input features (such as the grid frequency and voltage at the point of common coupling) on the aggregation effect, so as to fuse the inertia evaluation results of each plant with the characteristics of the station itself, and output the overall inertia evaluation value of the wind-solar-storage station.

[0163] Specifically, the above step c3 includes:

[0164] Step c31, splicing the historical station feature set, the plurality of first historical equivalent inertia evaluation values, the plurality of second historical equivalent inertia evaluation values and the plurality of third historical equivalent inertia evaluation values to obtain a station aggregated feature matrix.

[0165] Step c32, inputting the station aggregated feature matrix into a preset fully connected neural network to obtain a historical inertia evaluation value of the wind-solar-storage station.

[0166] Step c33, constructing an inertia evaluation aggregated model of the wind-solar-storage station by taking the real inertia evaluation value as a label, taking the historical station feature set, the plurality of first historical equivalent inertia evaluation values, the plurality of second historical equivalent inertia evaluation values and the plurality of third historical equivalent inertia evaluation values as inputs and taking the historical inertia evaluation value as an output.

[0167] Specifically, the historical station feature set, the plurality of first historical equivalent inertia evaluation values, the plurality of second historical equivalent inertia evaluation values and the plurality of third historical equivalent inertia evaluation values can be spliced in the feature dimension according to the correspondence between the feature dimension and the time sequence, and a station aggregated feature matrix with higher dimension and containing station multi-aspect features and inertia evaluation information of each machine group can be obtained.

[0168] Further, the preset fully connected neural network (MLP) contains several hidden layers, each layer is composed of a plurality of neurons, and full connection is realized between the neurons.

[0169] Further, the station aggregated feature matrix is input into an input layer of the fully connected neural network, and then the linear transformation and activation function of each hidden layer are used to learn the dynamic aggregated effect of inertia contribution of the wind power, photovoltaic and energy storage machine groups at the grid connection point and the influence of the station level input features thereon and gradually extract features, so as to obtain the historical inertia evaluation value of the wind-solar-storage station at the output layer.

[0170] Further, by means of supervised learning, the parameters of the fully connected neural network are adjusted by taking the real station inertia value as a reference, so that the model can accurately learn the aggregation rule of the station level inertia, thereby constructing an aggregated model that can be used for inertia evaluation of the wind-solar-storage station.

[0171] Specifically, the real inertia evaluation value is taken as a label for model training, the historical station feature set, the plurality of first historical equivalent inertia evaluation values, the plurality of second historical equivalent inertia evaluation values and the plurality of third historical equivalent inertia evaluation values are taken as inputs of the model, and the historical inertia evaluation value is taken as an output of the model.

[0172] Further, define the loss function, and update the parameters of the fully connected neural network through the back propagation algorithm, fix the parameters of the wind power, photovoltaic and energy storage machine group model, only train the parameters of the newly added aggregation model (fully connected neural network), until the prediction error of the model meets the requirements, thereby constructing the wind-solar-storage station inertia evaluation aggregation model.

[0173] In an example, a deep learning-based wind-solar-storage station inertia evaluation method is provided, which evaluates the equivalent inertia of the wind-solar-storage station by using a deep learning method, and considers the differences in frequency control and response characteristics of the wind power, photovoltaic and energy storage machine groups in the wind-solar-storage station, and respectively adopts a targeted equivalent inertia evaluation mechanism, specifically including:

[0174] (I) Wind turbine group inertia evaluation method based on graph convolutional neural network.

[0175] The wind turbine group has complex spatial topological structure and aerodynamic coupling characteristics, and the equivalent of the wind turbine group as a wind turbine for evaluation cannot accurately describe the associated influence between wind turbines. Therefore, the graph convolutional neural network (GCN) is applied to the inertia modeling of the wind turbine group, and the specific modeling method is as follows:

[0176] 1. Data collection and preprocessing.

[0177] The numerical weather forecast information of the actual wind turbine group is obtained, including wind speed, wind direction and other information; the geographic location information is obtained, including the geographic location of the wind turbine, the orientation of the wind turbine blade and other information; considering that it is difficult to obtain a large amount of frequency response data of the actual wind turbine group, a wind turbine group simulation model is established according to the actual data obtained, and the frequency response data (including real-time active power, frequency deviation and frequency change rate) of each wind turbine is obtained by setting various frequency response conditions. Further, the simulation data is cleaned to remove outliers.

[0178] 2. Construction of directed graph.

[0179] The output active power of wind turbines at different positions is affected by wind speed and wind direction, and considering the influence of factors such as wake effect and wind turbine blade orientation, the topological graph of the wind turbine group can be regarded as a directed graph. Each wind turbine is regarded as a node of the directed graph, and the electrical distance between wind turbines is regarded as the edge weight, which is defined as follows:

[0180] (1) A directed graph G=(V,E,A) is constructed, wherein: V represents a set of N graph nodes of the wind turbine group, each node represents a wind turbine; E represents a graph edge set of the wind turbine group; A represents an adjacency matrix of the wind turbine group, which is a quantification of the spatial dependence between N graph nodes.

[0181] (2) The directed edge E represents the distance d ijThe direction of the wind turbine is defined as the upwind direction, and the downwind direction is defined as the direction of the wind turbine, considering the wake effect of the wind turbine, and the reference direction is the dominant wind direction of the next month.

[0182] (3) The definition of each element in the adjacency matrix A is shown in the above relation (2).

[0183] (4) The characteristic information of each wind turbine is taken as the node feature vector, including wind speed, wind direction, real-time active power, and local inertia response coefficient (initial value is 0, updated by network learning), denoted as X wind ={x1,x2,…,x N}, where x i is the feature vector of the i-th wind turbine.

[0184] 3. Graph neural network model construction.

[0185] The architecture of the graph neural network used in this example is selected as a graph convolutional network (GCN), which is trained on wind turbine groups and photovoltaic turbine groups respectively. The feature representation of the current node is updated by aggregating the information of adjacent nodes, and the GCN includes an input layer, a hidden layer and an output layer, and each layer is designed as follows:

[0186] (1) Input layer: the constructed graph structure (adjacency matrix A and node feature matrix X) is input into the graph convolutional neural network of the wind turbine group, the dimension of the input layer matches the dimension of the node feature vector, and the normalization of X and the regularization of A are performed.

[0187] (2) Convolution layer: in the convolution layer, a direction-enhanced graph convolution (DE-GCN) is used, and the update calculation formula can be represented by the above relation (3).

[0188] (3) Hidden layer: based on the deep experimental results in the related literature, 3 layers of DE-GCN are selected, and Dropout is added between layers to prevent overfitting.

[0189] (4) Output layer: two layers of fully connected network are used, and the law of conservation of momentum is added in the loss function, and the formula is shown in the above relation (4).

[0190] In the hidden layer (3 layers of DE-GCN), the equivalent inertia calculation value is obtained by iterative direction-enhanced graph convolution: the information of the upwind neighbor is aggregated in each layer, the node feature vector is updated, and the local inertia response coefficient is gradually optimized (starting from 0). After 3 layers of DE-GCN, the local inertia response coefficient in H(3) is the equivalent inertia calculation value. The two layers of fully connected network in the output layer take H(3) as input to generate the final output, but the hidden layer has provided its essential representation. This process effectively combines the graph structure (wind direction, wake effect) and node features, enabling the model to accurately assess the inertia of the wind turbine group.

[0191] (ii) LSTM-based photovoltaic / energy storage fleet inertia evaluation method.

[0192] Considering that the arrangement of photovoltaic and energy storage is relatively concentrated, the inertia response sequence has time dependence, therefore both photovoltaic and energy storage fleet learn photovoltaic / energy storage fleet inertia related data using LSTM, the specific steps are as follows:

[0193] 1. Input layer design.

[0194] The input data of photovoltaic includes the active power, irradiance, and temperature of each photovoltaic array; the input data of energy storage includes the charge and discharge power, SOC, and terminal voltage of each electrochemical energy storage bin; the system input data includes the frequency deviation and frequency change rate of the grid-connected point, and the input data is normalized.

[0195] 2. One-dimensional convolution layer.

[0196] Considering the output fluctuation correlation between photovoltaic arrays and the response delay of energy storage, a one-dimensional convolution layer is added to LSTM to extract the local space-time features of photovoltaic and energy storage, and automatically capture the coupling characteristics inside the fleet. Taking photovoltaic as an example, the local space-time feature is that some photovoltaic panels are in the shadow area due to cloud cover, and the spatial characteristics of energy storage are weak, mainly considering the response delay of energy storage, which belongs to the time characteristics.

[0197] 3. LSTM hidden layer.

[0198] The hidden layer adopts a double-layer bidirectional LSTM structure, and takes the last time step as the output of the hidden layer, so as to capture the inertia response characteristics before and after the frequency disturbance. The one-dimensional convolution layer processes the entire input time sequence, and the output is also a complete time sequence (feature sequence) containing multiple time step feature vectors.

[0199] 4. Fully connected layer.

[0200] The fully connected layer selects the rectified linear unit (ReLU) as the activation function, and uses dropout to prevent overfitting, so as to fuse the high-order features output by the LSTM hidden layer (the space-time features of photovoltaic and energy storage at the last time step of the hidden layer output), and map to the inertia evaluation space.

[0201] 5. Output layer.

[0202] Among them, the definition of the loss function is the above relation (5).

[0203] 6. Model training

[0204] The data collected from the new energy station is divided into training set, validation set and test set, and is divided according to the proportion of 6:3:1. According to the characteristics of the selected data, appropriate learning rate, iteration number and other hyperparameters are set, and model training is started.

[0205] (Three) Constructing a station-level inertia aggregation model.

[0206] 1. Constructing station-level input features.

[0207] The real-time evaluation results of each machine group obtained in (one) and (two) are taken as the machine group input features, that is, the equivalent inertia evaluation values Hwind(t), HPV(t) and Hs(t) output by the wind power, photovoltaic and energy storage models. The station-level input features include the grid frequency measured at the grid connection point, the grid connection point voltage, the total active / reactive power output of the station and the station-level dispatching instructions / limits.

[0208] 2. Aggregation model architecture.

[0209] The inertia evaluation outputs H wind (t), H PV (t) and H s (t) of the wind power, photovoltaic and energy storage machine group models are taken as intermediate features, which are spliced with the station-level input features and then input into an additional fully connected neural network (MLP). The target of the MLP is to learn the dynamic aggregation effect of the inertia contributions of the wind power, photovoltaic and energy storage machine groups at the grid connection point and the influence of the station-level input features thereon, and the output is H total (t).

[0210] 3. Model training.

[0211] The real equivalent inertia H true (t) of the station level is taken as the supervision signal, which is fitted / calculated by simulating the response of the entire wind-solar-storage station under different working conditions and disturbances. The parameters of the wind power, photovoltaic and energy storage machine group models are fixed, and the parameters of the newly added aggregation model (MLP / LSTM / GCN) are trained using the station-level data.

[0212] The wind-solar-storage station inertia evaluation method based on deep learning provided in the present example considers the characteristics of the installation location, output characteristics and the like of the wind power, photovoltaic and energy storage machine groups, specifically selects a deep learning method to establish a model for training, and constructs a station-level aggregation model, so that the evaluation result of the wind-solar-storage station is more accurate.

[0213] An inertia evaluation device for a wind-solar-storage station is also provided in the embodiment. The device is used to implement the above-mentioned embodiments and preferred embodiments, and details thereof have been described above. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, implementation in hardware or a combination of software and hardware is also possible and contemplated.

[0214] The embodiment provides an inertia evaluation device for a wind-solar-storage station. The wind-solar-storage station includes a wind turbine group, a photovoltaic group, and an energy storage group. As shown in the figure, the device includes: Figure 2

[0215] A first obtaining module 201 is configured to obtain a real-time wind turbine group data set of the wind turbine group, a real-time photovoltaic group data set of the photovoltaic group, a real-time energy storage group data set of the energy storage group, and a real-time station feature set of the wind-solar-storage station.

[0216] A first processing module 202 is configured to input the real-time wind turbine group data set into a target wind turbine group inertia evaluation model for processing to obtain a plurality of first real-time equivalent inertia evaluation values of the wind turbine group.

[0217] A second processing module 203 is configured to input the real-time photovoltaic group data set into a target photovoltaic group inertia evaluation model for processing to obtain a plurality of second real-time equivalent inertia evaluation values of the photovoltaic group.

[0218] A third processing module 204 is configured to input the real-time energy storage group data set into a target energy storage group inertia evaluation model for processing to obtain a plurality of third real-time equivalent inertia evaluation values of the energy storage group.

[0219] A fourth processing module 205 is configured to input the real-time station feature set, the plurality of first real-time equivalent inertia evaluation values, the plurality of second real-time equivalent inertia evaluation values, and the plurality of third real-time equivalent inertia evaluation values into a wind-solar-storage station inertia evaluation aggregation model for processing to obtain a target inertia evaluation value of the wind-solar-storage station.

[0220] In some optional embodiments, the device includes:

[0221] A second obtaining module is configured to obtain a historical wind turbine group data set of the wind turbine group.

[0222] A first constructing module is configured to construct an initial directed graph and an initial node feature matrix according to the historical wind turbine group data set.

[0223] A second constructing module is configured to perform model training using a graph convolution network and construct a target wind turbine group inertia evaluation model based on the historical wind turbine group data set, the initial directed graph, and the initial node feature matrix.

[0224] ​In some optional embodiments, the second construction module comprises:

[0225] The acquisition sub-module is configured to acquire the target direction adjacency matrix and the target weight matrix.

[0226] The first processing sub-module is configured to input the initial directed graph and the initial node feature matrix into an input layer of the graph convolution network for processing to obtain a regularized target directed graph and a normalized target node feature matrix.

[0227] The second processing sub-module is configured to input the target directed graph, the target node feature matrix, the target direction adjacency matrix and the target weight matrix into a convolution layer of the graph convolution network for processing to obtain an enhanced initial node embedding matrix.

[0228] The third processing sub-module is configured to input the initial node embedding matrix, the target direction adjacency matrix and the target weight matrix into a hidden layer of the graph convolution network for processing to obtain a target node embedding matrix.

[0229] The fourth processing sub-module is configured to input the target node embedding matrix and a preset fully connected layer weight matrix into an output layer of the graph convolution network for processing to obtain a plurality of first historical equivalent inertia evaluation values in the wind turbine group.

[0230] The first construction sub-module is configured to input the historical wind turbine group dataset as input and output the plurality of first historical equivalent inertia evaluation values as output to construct a target wind turbine group inertia evaluation model.

[0231] In some optional embodiments, the method comprises:

[0232] The third acquisition module is configured to acquire a historical photovoltaic turbine group dataset of a photovoltaic turbine group and a historical energy storage turbine group dataset of an energy storage turbine group.

[0233] The first training module is configured to input the historical photovoltaic turbine group dataset into a preset long short-term memory network for training to obtain a target photovoltaic turbine group inertia evaluation model.

[0234] The second training module is configured to input the historical energy storage turbine group dataset into the preset long short-term memory network for training to obtain a target energy storage turbine group inertia evaluation model.

[0235] In some optional embodiments, the first training module comprises:

[0236] The fifth processing sub-module is configured to input the historical photovoltaic turbine group dataset into an input layer of the preset long short-term memory network for processing to obtain a normalized photovoltaic turbine group time series feature matrix.

[0237] The sixth processing sub-module is configured to input the photovoltaic turbine group time series feature matrix into a convolution layer of the preset long short-term memory network for processing to obtain a local spatiotemporal feature matrix.

[0238] a seventh processing submodule, configured to input the local spatiotemporal feature matrix into a hidden layer of a preset long short-term memory network for processing to obtain a target feature vector.

[0239] an eighth processing submodule, configured to input the target feature vector into a fully connected layer of the preset long short-term memory network for processing to obtain an inertia feature vector.

[0240] a ninth processing submodule, configured to input the inertia feature vector into an output layer of the preset long short-term memory network for processing to obtain a plurality of second historical equivalent inertia evaluation values of the photovoltaic plant.

[0241] a second construction submodule, configured to take the historical photovoltaic plant dataset as input and take the plurality of second historical equivalent inertia evaluation values as output to construct a target photovoltaic plant inertia evaluation model.

[0242] In some optional embodiments, the apparatus comprises:

[0243] a fourth acquisition module, configured to acquire a plurality of first historical equivalent inertia evaluation values of the wind power plant, a plurality of second historical equivalent inertia evaluation values of the photovoltaic plant, a plurality of third historical equivalent inertia evaluation values of the energy storage plant, and a historical station feature set of the wind-solar-storage station.

[0244] a fifth acquisition module, configured to acquire a real inertia evaluation value of the wind-solar-storage station.

[0245] a third construction module, configured to take the historical station feature set, the plurality of first historical equivalent inertia evaluation values, the plurality of second historical equivalent inertia evaluation values, and the plurality of third historical equivalent inertia evaluation values as input, take the real inertia evaluation value as a label, and construct a wind-solar-storage station inertia evaluation aggregation model by using a preset fully connected neural network.

[0246] In some optional embodiments, the third construction module comprises:

[0247] a splicing submodule, configured to splice the historical station feature set, the plurality of first historical equivalent inertia evaluation values, the plurality of second historical equivalent inertia evaluation values, and the plurality of third historical equivalent inertia evaluation values to obtain a station aggregation feature matrix.

[0248] a tenth processing submodule, configured to input the station aggregation feature matrix into the preset fully connected neural network for processing to obtain a historical inertia evaluation value of the wind-solar-storage station.

[0249] a third construction submodule, configured to take the real inertia evaluation value as a label, take the historical station feature set, the plurality of first historical equivalent inertia evaluation values, the plurality of second historical equivalent inertia evaluation values, and the plurality of third historical equivalent inertia evaluation values as input, and take the historical inertia evaluation value as output to construct the wind-solar-storage station inertia evaluation aggregation model.

[0250] Further functional descriptions of the above modules are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0251] In this embodiment, the inertia assessment device for wind and solar storage stations is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0252] This invention also provides a computer device having the above-described features. Figure 2 The inertia assessment device for wind and solar storage stations is shown.

[0253] Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 3 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 3 Take a processor 10 as an example.

[0254] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0255] The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the method shown in the above embodiments.

[0256] The memory 20 can include a program storage area and a data storage area. The program storage area can store an operating system, application programs required for at least one function, etc. The data storage area can store data created by the computer device, etc. In addition, the memory 20 can include a high-speed random access memory, and can also include a non-transitory memory such as at least one disk storage device, a flash memory device, or other non-transitory solid state memory device. In some alternative embodiments, the memory 20 can optionally include memory that is remotely located with respect to the processor 10, and which can be connected to the computer device through a network. Examples of such networks include, but are not limited to, the Internet, an enterprise intranet, a local area network, a mobile communications network, and combinations thereof.

[0257] The memory 20 can include a volatile memory, such as a random access memory, and / or can include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid state memory device. The memory 20 can also include an array of multi-state flash memory cells, which can be programmed to store one or more bits per cell. For example, multi-state flash memory cells can store two or more bits per cell. In a particular embodiment, the memory 20 can include a three-state flash memory cell, which can be programmed to store one or two bits per cell. In some embodiments, the memory 20 can include a combination of storage devices, such as one or more flash memory devices combined with one or more dynamic random access memory (DRAM) devices.

[0258] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0259] The embodiments of the present application also provide a computer readable storage medium, and the method according to the embodiments of the present application can be implemented in hardware, firmware, or recorded in a storage medium, or be implemented as computer code to be originally stored in a remote storage medium or a non-transitory machine readable storage medium downloaded through a network and stored in a local storage medium, so that the method described herein can be processed by such software on a storage medium using a general purpose computer, a special purpose processor, or programmable or special purpose hardware. The storage medium can be a disk, an optical disk, a read-only memory, a random access memory, a flash memory, a hard disk, or a solid state disk, etc. Further, the storage medium can also include a combination of the above-mentioned kinds of storage devices. It can be understood that the computer, the processor, the microprocessor controller, or the programmable hardware includes a storage component that can store or receive software or computer code, which, when accessed and executed by the computer, the processor, or the hardware, implements the method shown in the above embodiments.

[0260] Part of the present application can be applied as a computer program product, for example, computer program instructions, when executed by a computer, through the operation of the computer, the method and / or technical solutions according to the present application can be called or provided. Those skilled in the art should understand that the form of computer program instructions in computer readable medium includes but is not limited to source file, executable file, installation package file and the like, and accordingly, the way of computer program instructions executed by computer includes but is not limited to: the computer directly executes the instructions, or the computer compiles the instructions and then executes the corresponding compiled program, or the computer reads and executes the instructions, or the computer reads and installs the instructions and then executes the corresponding installed program. Here, the computer readable medium can be any available computer readable storage medium or communication medium accessible to the computer.

[0261] Although the embodiments of the present application are described in conjunction with the drawings, various modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the present application, and such modifications and changes fall within the scope defined by the appended claims.

Claims

1. A method for evaluating the inertia of a wind-solar-storage storage station, characterized in that, The wind-solar-storage station includes a wind turbine cluster, a photovoltaic cluster, and an energy storage cluster; the method includes: Obtain the real-time wind turbine cluster dataset, the real-time photovoltaic cluster dataset, the real-time energy storage cluster dataset, and the real-time site feature set of the wind-solar-storage power station. The real-time wind turbine cluster dataset is input into the target wind turbine cluster inertia evaluation model for processing to obtain multiple first real-time equivalent inertia evaluation values ​​of the wind turbine cluster. The real-time photovoltaic cluster dataset is input into the target photovoltaic cluster inertia assessment model for processing to obtain multiple second real-time equivalent inertia assessment values ​​for the photovoltaic cluster. The real-time energy storage cluster dataset is input into the target energy storage cluster inertia evaluation model for processing to obtain multiple third real-time equivalent inertia evaluation values ​​of the energy storage cluster. The real-time site feature set, the multiple first real-time equivalent inertia evaluation values, the multiple second real-time equivalent inertia evaluation values, and the multiple third real-time equivalent inertia evaluation values ​​are input into the wind-solar-storage site inertia evaluation aggregation model for processing to obtain the target inertia evaluation value of the wind-solar-storage site.

2. The method according to claim 1, characterized in that, The method further includes: Obtain the historical wind turbine cluster dataset of the aforementioned wind turbine cluster; An initial directed graph and an initial node feature matrix are constructed based on the historical wind turbine cluster dataset. Based on the historical wind turbine cluster dataset, the initial directed graph, and the initial node feature matrix, a graph convolutional network is used to train the model and construct the inertia evaluation model for the target wind turbine cluster.

3. The method according to claim 2, characterized in that, Based on the historical wind turbine cluster dataset, the initial directed graph, and the initial node feature matrix, a graph convolutional network is used to train the model and construct the inertia evaluation model for the target wind turbine cluster, including: Obtain the target direction adjacency matrix and the target weight matrix; The initial directed graph and the initial node feature matrix are input into the input layer of the graph convolutional network for processing to obtain a regularized target directed graph and a normalized target node feature matrix. The target directed graph, the target node feature matrix, the target orientation adjacency matrix, and the target weight matrix are input into the convolutional layer of the graph convolutional network to obtain the enhanced initial node embedding matrix. The initial node embedding matrix, the target orientation adjacency matrix, and the target weight matrix are input into the hidden layer of the graph convolutional network to obtain the target node embedding matrix; The target node embedding matrix and the preset fully connected layer weight matrix are input into the output layer of the graph convolutional network for processing to obtain multiple first historical equivalent inertia evaluation values ​​in the wind turbine group. Using the historical wind turbine cluster dataset as input and the multiple first historical equivalent inertia assessment values ​​as output, the inertia assessment model of the target wind turbine cluster is constructed.

4. The method according to claim 1, characterized in that, The method further includes: Obtain the historical photovoltaic cluster dataset and the historical energy storage cluster dataset of the photovoltaic cluster; The historical photovoltaic cluster dataset is input into a preset long short-term memory network for training to obtain the target photovoltaic cluster inertia evaluation model; The historical energy storage cluster dataset is input into the preset long short-term memory network for training to obtain the inertia evaluation model of the target energy storage cluster.

5. The method according to claim 4, characterized in that, The historical photovoltaic (PV) cluster dataset is input into a preset long short-term memory (LSTM) network for training to obtain the target PV cluster inertia assessment model, including: The historical photovoltaic cluster dataset is input into the input layer of the preset long short-term memory network for processing to obtain the normalized photovoltaic cluster time-series feature matrix. The photovoltaic cluster time-series feature matrix is ​​input into the convolutional layer of the preset long short-term memory network for processing to obtain the local spatiotemporal feature matrix; The local spatiotemporal feature matrix is ​​input into the hidden layer of the preset long short-term memory network for processing to obtain the target feature vector; The target feature vector is input into the fully connected layer of the preset long short-term memory network for processing to obtain the inertia feature vector; The inertia feature vector is input into the output layer of the preset long short-term memory network for processing to obtain multiple second historical equivalent inertia evaluation values ​​of the photovoltaic cluster. Using the historical photovoltaic cluster dataset as input and the multiple second historical equivalent inertia assessment values ​​as output, the inertia assessment model of the target photovoltaic cluster is constructed.

6. The method according to claim 1, characterized in that, The method further includes: Obtain multiple first historical equivalent inertia assessment values ​​of the wind turbine group, multiple second historical equivalent inertia assessment values ​​of the photovoltaic group, multiple third historical equivalent inertia assessment values ​​of the energy storage group, and the historical site feature set of the wind-solar-storage station; Obtain the actual inertia assessment value of the wind and solar storage station; Based on the historical site feature set, the multiple first historical equivalent inertia assessment values, the multiple second historical equivalent inertia assessment values, and the multiple third historical equivalent inertia assessment values, and using the actual inertia assessment value as a label, the wind, solar and energy storage site inertia assessment aggregation model is constructed using a preset fully connected neural network.

7. The method according to claim 6, characterized in that, Based on the historical site feature set, the multiple first historical equivalent inertia assessment values, the multiple second historical equivalent inertia assessment values, and the multiple third historical equivalent inertia assessment values, and using the actual inertia assessment value as the label, a pre-defined fully connected neural network is used to construct the wind-solar-storage site inertia assessment aggregation model, including: The historical site feature set, the multiple first historical equivalent inertia evaluation values, the multiple second historical equivalent inertia evaluation values, and the multiple third historical equivalent inertia evaluation values ​​are concatenated to obtain the site aggregate feature matrix; The aggregated feature matrix of the wind and solar storage station is input into a preset fully connected neural network for processing to obtain the historical inertia evaluation value of the wind and solar storage station. Using the actual inertia assessment value as the label, and taking the historical site feature set, the multiple first historical equivalent inertia assessment values, the multiple second historical equivalent inertia assessment values, and the multiple third historical equivalent inertia assessment values ​​as inputs, and the historical inertia assessment value as the output, a wind, solar, and energy storage site inertia assessment aggregation model is constructed.

8. A device for assessing the inertia of a wind-solar-storage storage station, characterized in that, The wind-solar-storage station includes a group of wind turbines, a group of photovoltaic generators, and a group of energy storage units; the device includes: The first acquisition module is used to acquire the real-time wind turbine cluster dataset, the real-time photovoltaic cluster dataset, the real-time energy storage cluster dataset, and the real-time site feature set of the wind-solar-storage station. The first processing module is used to input the real-time wind turbine cluster dataset into the target wind turbine cluster inertia evaluation model for processing, and obtain multiple first real-time equivalent inertia evaluation values ​​of the wind turbine cluster. The second processing module is used to input the real-time photovoltaic cluster dataset into the target photovoltaic cluster inertia evaluation model for processing, and obtain multiple second real-time equivalent inertia evaluation values ​​of the photovoltaic cluster. The third processing module is used to input the real-time energy storage cluster dataset into the target energy storage cluster inertia evaluation model for processing, and obtain multiple third real-time equivalent inertia evaluation values ​​of the energy storage cluster. The fourth processing module is used to input the real-time site feature set, the multiple first real-time equivalent inertia evaluation values, the multiple second real-time equivalent inertia evaluation values, and the multiple third real-time equivalent inertia evaluation values ​​into the wind-solar-storage site inertia evaluation aggregation model for processing, so as to obtain the target inertia evaluation value of the wind-solar-storage site.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the wind and solar storage station inertia assessment method according to any one of claims 1 to 7.

10. A computer program product, characterized in that, Includes computer instructions for causing a computer to execute the wind and solar storage station inertia assessment method according to any one of claims 1 to 7.

Images