Method and device for establishing hybrid drive-based industry expansion retrofit digital twin model

Abstract

The application belongs to the technical field of electric power, and provides a method and device for establishing a digital twin model based on hybrid driving for industrial expansion installation, which comprises the steps of multi-element data acquisition, multi-element data preprocessing based on an improved fractional order flexible space method, data-driven model construction based on an improved TCN-BiGRU method, obtaining a hybrid driving model by combining an improved semi-invariant method and the data-driven model, constructing a digital twin dynamic model of a power supply area, and simulating and deducing an industrial expansion installation scheme, and the like. The application utilizes digital twin technology to construct an electric power system model with real-time monitoring and optimization capabilities. The model can accurately simulate the power grid state of an active power supply area, analyze the characteristics of core parameters under different industrial expansion schemes, predict the influence of new energy access on power grid operation, and provide more accurate and flexible scheduling and decision-making basis for power grid management.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electric power, and particularly relates to a method and device for establishing a digital twin model for industrial expansion based on hybrid driving. BACKGROUND

[0002] Traditional power grid load prediction and power dispatching methods often cannot cope with the complexity of new energy access, especially for large-scale new energy installation and industrial expansion demand, lacking effective comprehensive evaluation and optimization tools. In addition, the power system often relies on static models and historical data-based planning methods in actual operation, which cannot dynamically reflect the real-time changes of the power grid state. Therefore, the dynamic and real-time updated industrial expansion model based on digital twin technology can effectively simulate the power system state of the transformer area and predict the impact of new energy access, providing more accurate data support for power grid dispatching and investment decision-making. SUMMARY

[0003] To solve the above technical problems, the application provides a method for establishing a digital twin model for industrial expansion based on hybrid driving to solve the problems in the prior art. The technical scheme adopted by the application is as follows:

[0004] The method for establishing a digital twin model for industrial expansion based on hybrid driving comprises the following steps:

[0005] Step 1: Collecting multi-element data, the multi-element data including static data, dynamic data and environmental data;

[0006] Step 2: Preprocessing the multi-element data based on an improved fractional order flexible space method, including: converting the multi-element data into an m*n matrix, k being the data type, to obtain an original high-dimensional multi-element data feature set m is the data dimension, the high-dimensional multi-element data is mapped to low-dimensional multi-element data, a multi-element local neighborhood graph and a multi-element local objective function are established, and the typical features of the multi-element data are obtained by optimization:

[0007] The multi-element local proximity weight matrix is defined as:

[0008]

[0009] Wherein: is the multi-element proximity weight matrix, represents the length of the neighborhood is r, and t is a time series; is the jth group of data in the kth element data;

[0010] The minimum multi-element local proximity objective function is obtained as:

[0011]

[0012] wherein: X k is an input multi-dimensional data matrix, is a local optimal projection, is a multi-dimensional Laplacian matrix, is a diagonal matrix, argmin() is a mathematical function for finding the parameter value that makes the function take the minimum value, and H is an initial high-dimensional data projection matrix to low-dimensional data;

[0013] The multi-dimensional global proximity weight matrix is defined as:

[0014]

[0015] wherein: is a multi-dimensional global proximity weight matrix, i represents a certain row in the matrix, and j represents a certain column in the matrix;

[0016] The multi-dimensional global proximity objective function is maximized:

[0017]

[0018] wherein: is a global optimal projection, is a multi-dimensional Laplacian matrix, is a diagonal matrix;

[0019] An improved fractional order flexible space learning model is established according to the minimization of the multi-dimensional local proximity and the maximization of the multi-dimensional global proximity function, and the optimal projection matrix H is calculated by the Lagrange multiplier method and the optimal projection matrix H is converted into a low-dimensional feature subset according to the following formula:

[0020]

[0021] wherein: Y k is a processed low-dimensional multi-dimensional data matrix, is an optimal projection matrix;

[0022] Step 3: Construct a data-driven model based on the improved TCN-BiGRU method;

[0023] Step 4: Combine the improved semi-invariant method and the data-driven model of step 3 to obtain a hybrid driving model;

[0024] Step 5: Construct a digital twin district dynamic model and simulate and deduce the industry expansion and installation scheme.

[0025] Further, the static data includes substation topology data, power equipment parameters and equipment use cycle; the dynamic data includes power operation data, equipment state monitoring data, new energy power generation data and load demand; the environmental data includes meteorological data, atmospheric pressure and precipitation data.

[0026] Further, the key information of the processed multi-element data is captured by a time sequence convolutional neural network, and a multi-dimensional convolution is adopted:

[0027]

[0028] Wherein: d is the expansion coefficient, M is the size of the convolution kernel, g(i) is the i-th element in the convolution kernel, is the input data, and s is the pull domain;

[0029] And the residual connection is performed by the following formula:

[0030] h(y)=y+g(y)

[0031] Wherein: y is the data set, h(y) is the residual connection output, and g(y) is the nonlinear transformation;

[0032] The TCN-BIGRU adds two groups of one-way GRU neural network units, and the opposite interfaces are connected to form an improved TCN-BIGRU neural network model with a bidirectional network algorithm;

[0033] The influence of input features on the expansion of the industry is evaluated by the multi-head attention mechanism, and each influence factor is given a corresponding weight, and each matrix is divided into multiple heads to obtain:

[0034]

[0035] A(Y)=Concat(A1,A2,...,A h )W o

[0036] Wherein: A i is the attention output of the i-th head, d k is the key vector dimension, Q i is the i-th value query space, K i is the i-th key space, V i is the i-th value space, W o is the weight vector of the multi-element data, and A(Y) is the output attention of h heads.

[0037] The data-driven model of the expansion of the industry is constructed by the time sequence convolutional neural network, the BiGRU model and the multi-head attention mechanism to predict the power output fluctuation of the substation load, photovoltaic and wind power.

[0038] Further, step 4 comprises: constructing a power system calculation model based on the power topology model;

[0039] The power topology model is an IEEE 34-node system, wherein photovoltaic is accessed at node 13 and node 18, and wind energy is accessed at node 19 and node 22;

[0040] By combining the load normal probability model, the wind-wind and wind-light scene joint probability distribution model, the geometric distance between each pair of scenes a and b in the scene N is calculated, the scene g with the closest geometric distance to the scene a in the scene M is replaced by the scene a to form a new scene set M', and the step is repeated to reduce the number of scenes;

[0041] The node voltage U and the branch power P are expressed as the linear sum of the node injection power variables as follows:

[0042]

[0043] Wherein, J0 and G0 are the partial derivatives of the node injection power imbalance and the branch power with respect to the node voltage amplitude and phase angle, ΔU (s) , ΔP (s) , ΔX (s) are the node voltage variation, branch power variation and node injection power variation after semi-non-dimensional conversion;

[0044] In combination with the Gram-Charlier series, the probability power flow result is obtained, the physical driving model is obtained in combination with the power topology model, the voltage stability and other power grid operation constraints are provided by the physical driving model, the prediction ability of the load and new energy fluctuation is compensated by the data-driven model of step 3, and the model parameters are adjusted to improve the accuracy by comparing the real-time data with the predicted values of the model, and a hybrid driving model is obtained.

[0045] Further, step 5 comprises: injecting the low-dimensional static data and historical data obtained in step 2 into the hybrid driving model to generate an initial twin model, and real-time accessing of sensor data, dynamically updating the substation load, equipment state and new energy access situation, automatically adjusting the key parameters of the twin model to ensure that the model is synchronized with the actual operation state, including the following steps:

[0046] Step 5.1: inputting the new load access demand of the user and the access position and access scale of photovoltaic and energy storage, wherein the new load includes the load access point, load power and load type;

[0047] Step 5.2: simulating the influence of new energy access on the operation of the substation in the twin model, and evaluating the influence of new energy access on power loss and system stability by analyzing voltage fluctuation, current distribution and line load rate change;

[0048] Step 5.3: Analyze the relationship between the newly added load access point and the voltage stability of the transformer area, and the influence of new energy fluctuation on the line load rate and power distribution using the maximum mutual coefficient method;

[0049] Step 5.4: Establish a feedback mechanism to optimize the industry expansion and new energy access scheme, adjust the access point, power distribution and load ratio, and output the optimal access scheme.

[0050] The digital twin model device based on hybrid driving for industry expansion installation includes a data preprocessing module, a hybrid driving module and an industry expansion installation scheme analysis module.

[0051] The data preprocessing module communicates with the power grid system to obtain multi-element data, reduces the dimension of high-dimensional data by improving the fractional order flexible space method, and finally transmits the hybrid driving module;

[0052] The hybrid driving module uses the improved TCN-BIGRU method to perform deep learning on the low-dimensional data in step 1 to form a data-driven module, combines the IEEE 34-node system and the improved semi-nonlinear load flow calculation method to obtain a physical driving model, modifies the data-driven model to the physical driving model, and transmits the results to the industry expansion installation analysis module;

[0053] The industry expansion installation scheme analysis module communicates with the power grid, takes the user installation data as input, analyzes the node voltage change, branch power change and node injection power change, finds the scheme with the smallest influence on the stability of the power grid, and obtains the optimal installation scheme.

[0054] The present application has the following advantages:

[0055] Based on the digital twin technology, the present application can obtain the running state of the transformer area power grid in real time, monitor the real-time data of the transformer, feeder, load and new energy generation. Secondly, it can be dynamically updated in the background of new energy access, and reflect the changes of the power grid in real time, help the power grid managers to take timely measures to avoid equipment overload or power waste. Finally, the characteristics of wind-solar-storage cooperation are dynamically optimized, the capacity of the power grid to accept new energy is improved, and the stability of the power grid operation is affected by the large fluctuation of new energy generation. BRIEF DESCRIPTION OF DRAWINGS

[0056] Fig. 1 The method flowchart of the present application is shown in the figure;

[0057] Fig. 2 The TCN-BiGRU model structure diagram is shown in the figure;

[0058] Fig. 3 The IEEE 34-node system diagram is shown in the figure. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present application will be clearly and completely described below Figs. 1-3 It should be apparent that the described embodiments are only a part of the embodiments of the present application, and not all the embodiments, and if not specifically indicated, the technical means used in the embodiments are conventional means familiar to those skilled in the art.

[0060] The present application utilizes digital twin technology to construct a power system model with real-time monitoring and optimization capabilities. This model can accurately simulate the grid state of active substations, analyze the characteristics of core parameters under different industry expansion schemes, predict the impact of new energy access on grid operation, and provide more accurate and flexible scheduling and decision-making basis for grid management. First, multiple types of data such as substations, equipment, and environment are obtained, and the multiple data are preprocessed according to the improved fractional order flexible space method. Second, a data-driven model is constructed based on the improved TCN-BiGRU method. Third, a hybrid driving model is obtained by combining the improved semi-invariant method and the data-driven model. Finally, static data and historical data are injected into the hybrid driving model to generate an initial twin model, and the industry expansion installation scheme is simulated and deduced.

[0061] The industry expansion digital twin model establishment method based on hybrid driving includes the following steps:

[0062] Step 1: Collect multiple data, including static data, dynamic data, and environmental data.

[0063] As Fig. 2 The collection of multiple data provides rich information support for the construction of the digital twin substation model, including static data, dynamic data, and environmental data. The static data includes substation topology structure data, power equipment parameters, and equipment service life data. The dynamic data is collected by sensors, SCADA systems, and other devices under the running state of the substation, including power operation data, equipment state monitoring data, new energy generation data, and load demand data. The environmental data is the external factor data under the condition of substation equipment and new energy access, including meteorological data, atmospheric pressure, and precipitation data.

[0064] Step 2: Preprocess the multiple data based on the improved fractional order flexible space method.

[0065] The multiple data obtained in step 1 are converted into an m*n matrix, where k is the data type, and the original high-dimensional multiple data feature set is obtained m is the data dimension. By mapping high-dimensional multiple data to low-dimensional multiple data and preserving the similarity between data, the typical features of multiple data are obtained by establishing a multi-element local neighborhood graph and a multi-element local objective function.

[0066] The multivariate local proximity weight matrix is defined as:

[0067]

[0068] wherein: is a multivariate proximity weight matrix, represents a neighborhood of length r, and t is a time series. is the jth group of data in the kth data.

[0069] The minimized multivariate local proximity objective function is obtained as:

[0070]

[0071]

[0072] wherein: is a locally optimal projection, is a multivariate Laplacian matrix, is a diagonal matrix of argmin() is a mathematical function for finding the value of an argument that minimizes a function, and H is an initial high-dimensional data projection matrix to low-dimensional data.

[0073] The multivariate global proximity weight matrix is defined as:

[0074]

[0075] wherein: is a multivariate global proximity weight matrix; i represents a row in the matrix, and j represents a column in the matrix.

[0076] The maximized multivariate global proximity objective function is:

[0077]

[0078] wherein: is a globally optimal projection, is a multivariate Laplacian matrix, is a diagonal matrix of .

[0079] The improved fractional flexible space learning model is established according to the minimized multivariate local proximity and the maximized multivariate global proximity function, and the optimal projection matrix H is obtained by Lagrange multiplier method. The optimal projection matrix H is converted into a low-dimensional feature subset according to the following formula.

[0080]

[0081] wherein: Y k is the processed low-dimensional multivariate data matrix, is the optimal projection matrix, X k is the input multivariate data matrix.

[0082] Step 3: Construct a data-driven model based on the improved TCN-BiGRU method;

[0083] Through the time sequence convolutional neural network, the key information of the processed multivariate data is captured, and multi-dimensional convolution is adopted:

[0084]

[0085] wherein: d is the dilation coefficient, M is the size of the convolution kernel, g(i) is the i-th element in the convolution kernel, is the input data, and s is the pull domain.

[0086] And the residual connection is made by the following formula:

[0087] h(y) = y + g(y)

[0088] wherein: y is the data set, h(y) is the residual connection output, and g(y) is the nonlinear transformation.

[0089] Through the TCN-BIGRU, two groups of one-way GRU neural network units are added, and the improved TCN-BIGRU neural network model with bidirectional network algorithm is composed by connecting the opposite groups, which expands the attention field and analyzes the state of the previous and subsequent time;

[0090] The influence of the input features on the expansion equipment is evaluated by the multi-head attention mechanism, and each influence factor is given a corresponding weight, and each matrix is divided into multiple heads to obtain:

[0091]

[0092] A(Y) = Concat(A1, A2,..., A h )W o

[0093] wherein: A i is the attention output of the i-th head, d k is the key vector dimension, Q i is the i-th value query space, K i is the i-th key space, V i is the i-th value space, W o is the weight vector of the multivariate data, and A(Y) is the output attention of h heads.

[0094] The data-driven model of the industry expansion is composed of a time convolution neural network, a BiGRU model and a multi-head attention mechanism to predict the power output fluctuation of the transformer area load, photovoltaic and wind power. Fig. 2 The TCN-BiGRU model structure diagram is shown.

[0095] Step 4: A hybrid driving model is obtained by combining the improved semi-invariant method and the data-driven model of step 3.

[0096] The power system calculation model is constructed based on the power topology model. The power topology model is the IEEE 34-node system shown in FIG. 3. The photovoltaic is connected to node 13 and node 18, and the wind power is connected to node 19 and node 22. Based on the power topology model, the improved semi-invariant method is used to calculate the node voltage, current distribution and power flow.

[0097] By combining the load normal probability model, the wind-wind and wind-light scene joint probability distribution model, the geometric distance between each pair of scenes a and b in scene N is calculated. The scene g with the closest geometric distance to scene a in scene M is replaced by scene a to form a new scene set M'. The step is repeated to reduce the number of scenes.

[0098] The node voltage U and branch power P are expressed as the linear sum of the node injection power variables as follows:

[0099]

[0100] Where J0 and G0 are the partial derivatives of the node injection power imbalance and the branch power with respect to the node voltage amplitude and phase angle, respectively, and ΔU (s) , ΔP (s) , ΔX (s) are the node voltage change, branch power change and node injection power change after semi-invariant conversion.

[0101] The Gram-Charlier series is combined to obtain the probability power flow result. The physical driving model is obtained by combining the power topology model, which provides the voltage stability and other power grid operation constraints. The data-driven model of step 3 compensates for the prediction ability of load and new energy fluctuations, and compares the real-time data with the predicted values of the model to adjust the model parameters to improve the accuracy, and obtains the hybrid driving model.

[0102] Step 5: Construct a digital twin transformer area dynamic model and simulate and deduce the industry expansion scheme.

[0103] The low-dimensional static data and historical data obtained in step 2 are injected into the hybrid driving model to generate an initial twin model, and real-time sensor data is accessed to dynamically update the load of the transformer area, the state of the equipment and the access of new energy, and automatically adjust the key parameters (such as line impedance, load distribution) of the twin model, so as to ensure that the model is synchronized with the actual operation state.

[0104] Step 5.1: input the user's new load access demand and the access location and access scale of photovoltaic and energy storage, wherein the new load includes a load access point, a load power and a load type (residential, industrial and commercial).

[0105] Step 5.2: simulate the influence of new energy access on the operation of the transformer area in the twin model, and evaluate the influence of new energy access on power loss and system stability by analyzing voltage fluctuation, current distribution and line load rate change.

[0106] Step 5.3: analyze the relationship between the new load access point and the voltage stability of the transformer area, and the influence of new energy fluctuation on line load rate and power distribution by using the maximum mutual trust coefficient method.

[0107] Step 5.4: establish a feedback mechanism to output the optimal access scheme by optimizing the industry expansion and new energy access scheme, adjusting the access point, power distribution and load proportion.

[0108] The industry expansion installation digital twin model device based on hybrid driving includes a data preprocessing module, a hybrid driving module and an industry expansion installation scheme analysis module.

[0109] The data preprocessing module communicates with the power grid system to obtain multi-element data, reduces the dimension of high-dimensional data by improving the fractional order flexible space method, and finally transmits the hybrid driving module;

[0110] The hybrid driving module uses the improved TCN-BIGRU method to perform deep learning on the low-dimensional data in step 1 to form a data-driven module, combines the IEEE 34-node system and the improved semi-nonlinear power flow calculation method to obtain a physical driving model, modifies the data-driven model to the physical driving model, and transmits the result to the industry expansion installation analysis module;

[0111] The industry expansion installation scheme analysis module communicates with the power grid, takes the user installation data as input, analyzes the voltage change of the user input node, the branch power change and the node injected power change, finds the scheme with the smallest influence on the stability of the power grid, and obtains the optimal installation scheme.

[0112] The above described embodiments are only to illustrate the preferred modes of the present application, and are not intended to limit the scope of the present application. Any modification, variation, modification, and replacement of the technical solutions of the present application made by those skilled in the art without departing from the design spirit of the present application shall fall within the protection scope determined by the claims of the present application.

Claims

1. A method for establishing a hybrid drive-based industry expansion retrofit digital twin model, characterized in that, Includes the following steps: Step 1: Collect multivariate data, which includes: static data, dynamic data, and environmental data; Step 2: preprocessing of multivariate data based on improved fractional flexible space method, including: converting multivariate data into an m*n matrix, k being the data category, to obtain an original high-dimensional multivariate data feature set m being the data dimension, mapping the high-dimensional multivariate data to low-dimensional multivariate data, establishing a multivariate local neighborhood graph and a multivariate local objective function, and optimizing to obtain typical features of the multivariate data: Define the multivariate local proximity weight matrix as follows: wherein: is a plurality of proximity weight matrices, represents is a neighborhood of length r, t is a time series; is the jth group of data in the kth metadata; The objective function for minimizing the multivariate local proximity is obtained as follows: wherein: X k is an input multi-dimensional data matrix, is a locally optimal projection, is a multi-dimensional Laplacian matrix, is is a diagonal matrix, argmin() is a mathematical function for finding the argument value that makes a function take a minimum value, and H is an initial high-dimensional data projection matrix to low-dimensional data. Define the multivariate global proximity weight matrix as follows: wherein: is a multivariate global proximity weight matrix, i represents a row in the matrix, and j represents a column in the matrix; Maximize the multivariate global proximity objective function: in: For the best global projection, It is a multivariate Laplace matrix. for a diagonal matrix; The improved fractional flexible space learning model is established according to the minimization of the multivariate local proximity and the maximization of the multivariate global proximity function, and the optimal projection matrix is calculated through the Lagrange multiplier method and the optimal projection matrix is converted into a low-dimensional feature subset according to the following formula: wherein: Y k is the processed low-dimensional multivariate data matrix, is the optimal projection matrix; Step 3: Construct a data-driven model based on the improved TCN-BiGRU method; Step 4: Combine the improved semi-invariant method and the data-driven model from Step 3 to obtain the hybrid-driven model; Step 5: Construct a dynamic model of the digital twin transformer substation and simulate and extrapolate the business expansion application scheme.

2. The method of claim 1, wherein the method further comprises: In step 1, static data includes transformer area topology data, power equipment parameters, and equipment usage cycles; dynamic data includes power operation data, equipment status monitoring data, new energy power generation data, and load demand; environmental data includes meteorological data, atmospheric pressure, and precipitation data.

3. The method of claim 1, wherein the method further comprises: Step 3 includes: extracting key information from the processed multivariate data using a temporal convolutional neural network, employing multidimensional convolution: where: d is the dilation coefficient, M is the size of the convolution kernel, g(i) is the i-th element in the convolution kernel, is the input data, s is the pullback field; And the residuals are connected using the following formula: h(y) = y + g(y) Where: y is the dataset, h(y) is the residual connection output, and g(y) is the nonlinear transformation; An improved TCN-BIGRU neural network model with a bidirectional network algorithm is formed by adding two unidirectional GRU neural network units to TCN-BIGRU and connecting them in opposite directions. The influence of input features on business expansion applications is evaluated using a multi-head attention mechanism, and each influencing factor is assigned a corresponding weight. Each matrix is ​​divided into multiple heads, resulting in: A(Y) = Concat(A1, A2,..., A h )W o where: A i is the attention output for the i-th head, d k is the key vector dimension, Q i is the i-th value query space, K i is the i-th key space, V i is the i-th value space, W o is the weight vector for the multi-modal data, A(Y) is the output attention of h heads added together; A data-driven model for business expansion and installation applications is constructed using a temporal convolutional neural network, a BiGRU model, and a multi-head attention mechanism to predict fluctuations in transformer load, photovoltaic, and wind power output.

4. The method of claim 1, wherein the method further comprises: Step 4 includes: constructing a power system computational model based on the power topology model; The power topology model is an IEEE 34-node system, in which photovoltaic power is connected at nodes 13 and 18, and wind power is connected at nodes 19 and 22. By combining the load normal probability model, the wind-wind and wind-light joint probability distribution models, the geometric distance between each pair of scenes a and b in scene N is calculated. Then, scene g, which is the closest to scene a in scene M, is replaced with scene a to form a new scene set M'. This step is repeated to reduce the number of scenes. The linear sum of the node voltage U and branch power P representing the node injected power variable is expressed as follows: Wherein: J0, G0 are the partial derivatives of the node injection power imbalance and the branch power to the node voltage amplitude and phase angle, ΔU (s) , ΔP (s) , Δx (s) are the node voltage change, branch power change and node injection power change after half-units conversion; By combining Gram-Charlier series, probabilistic power flow results are obtained. By combining power topology model, physical driving model is obtained. The physical driving model provides voltage stability grid operation constraints. The data driving model from step 3 compensates for the predictive ability of load and renewable energy fluctuations. Furthermore, by comparing real-time data with the model's predicted values, the model parameters are adjusted to improve accuracy, resulting in a hybrid driving model.

5. The hybrid drive-based industry expansion retrofit digital twin model establishing method according to claim 1, characterized in that, Step 5 includes: injecting the low-dimensional static data and historical data obtained in step 2 into the hybrid driving model to generate an initial twin model, and accessing sensor data in real time to dynamically update the load of the transformer area, the state of the equipment and the access of new energy, automatically adjust the key parameters of the twin model, and ensure that the model is synchronized with the actual running state, including the following steps: Step 5.1: input the user's new load access demand and the access location and access of photovoltaic and energy storage, and the access of large-scale, wherein the new load includes the load access point, the load power and the load type; Step 5.2: simulate the influence of new energy access on the operation of the transformer area in the twin model, and evaluate the influence of new energy access on power loss and system stability by analyzing voltage fluctuation, current distribution and line load rate change; Step 5.3: using the maximum mutual trust coefficient method, analyze the relationship between the new load access point and the voltage stability of the transformer area, and the influence of new energy fluctuation on line load rate and power distribution; Step 5.4: establish a feedback mechanism to optimize the industry expansion and new energy access scheme, adjust the access point, power distribution and load proportion, and output the optimal access scheme.

6. The device based on hybrid drive industry expansion installation digital twin model, characterized in that, The industry expansion installation digital twin model based on the hybrid driving method according to any one of claims 1-5, comprising: a data preprocessing module, a hybrid driving module, and an industry expansion installation scheme analysis module; wherein: The data preprocessing module communicates with the power grid system to obtain multi-element data, reduces the dimension of high-dimensional data by improving the fractional order flexible space method, and finally transmits the hybrid driving module; The hybrid driving module uses the improved TCN-BIGRU method to perform deep learning on the low-dimensional data in step 1 to form a data-driven module, combines the IEEE 34-node system and the improved semi-nonlinear power flow calculation method to obtain a physical driving model, modifies the data-driven model to the physical driving model, and transmits the results to the industry expansion installation analysis module; The industry expansion installation scheme analysis module communicates with the power grid, takes the user installation data as input, analyzes the node voltage change, branch power change and node injected power change, finds the scheme with the smallest influence on the stability of the power grid, and obtains the optimal installation scheme.

Images