Island micro-grid impact current prediction method and device and electronic equipment

Abstract

The invention provides an impact current prediction method and device for an island micro-grid and electronic equipment, and relates to the technical field of power grids, and the method comprises the steps: inputting the current operation data corresponding to the current operation mode of the island micro-grid into a pre-trained impact current prediction model, obtaining an impact current prediction result corresponding to the current operation mode of the island microgrid output by the impact current prediction model; the impact current prediction model is a neural network model obtained by training based on a training data set and a label corresponding to the training data set. According to the technical scheme, various unknown scenes can be coped with through the impact current prediction model, the determined impact current prediction result of the island microgrid is accurate, an electromagnetic transient simulation model does not need to be reconstructed for each scene, the time cost and the labor cost are saved, and the impact current prediction efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power grids, and in particular, to a method, device, and electronic device for predicting inrush current of an island microgrid. Background Art

[0002] An island microgrid is a power grid operation mode with extremely high risks, showing characteristics such as weak support ability, diverse types of power sources and loads, and harsh operating environments. There are many inrush loads in the island microgrid. Diesel engines, ball mills and other large equipment in the island microgrid can generate inrush currents several times or even more than ten times during normal startup, switching of large equipment, and grid connection switching operations, thereby causing voltage and frequency fluctuations, with the action time ranging from 1 to 10 seconds, which has a great impact on the operation of the microgrid. Even when a fault occurs in the island microgrid, the generated inrush current will impact power electronic devices with weak overcurrent capacity, and even cause problems such as burnout, posing a serious challenge to the safe and reliable operation of the island microgrid. Therefore, predicting the inrush current under various operating scenarios to facilitate the adoption of countermeasures in advance is of great significance for improving the controllability of the island microgrid.

[0003] The existing technology mainly uses electromagnetic transient simulation to analyze the inrush current. If the operating scenario of the island microgrid changes, it is necessary to reconstruct the simulation model. Therefore, the existing methods are difficult to cope with various unknown scenarios, and the simulation analysis method consumes a large amount of time and manpower, seriously affecting the analysis efficiency. Summary of the Invention

[0004] The present invention provides a method, device, and electronic device for predicting inrush current of an island microgrid, to solve the defect in the prior art that if the operating scenario of the island microgrid changes during the inrush current prediction process, it is necessary to reconstruct the simulation model, consuming a large amount of time and manpower, and seriously affecting the analysis efficiency. The technical solution of the present invention can cope with various unknown scenarios through the inrush current prediction model, and the determined inrush current prediction result of the island microgrid is relatively accurate. There is no need to reconstruct the electromagnetic transient simulation model for each scenario, saving time cost and labor cost, and improving the efficiency of inrush current prediction.

[0005] The present invention provides a method for predicting inrush current of an island microgrid, including the following steps.

[0006] Input the current operation data corresponding to the current operation mode of the island microgrid into a pre-trained inrush current prediction model, and obtain the inrush current prediction result corresponding to the current operation mode of the island microgrid output by the inrush current prediction model; the inrush current prediction model is a neural network model trained based on a training data set and labels corresponding to the training data set.

[0007] A method for predicting inrush current of an islanded microgrid provided by the present invention, the inrush current prediction model is obtained by training through the following steps: Obtain the sub-operation data corresponding to each of multiple operation modes of the islanded microgrid; Based on all the sub-operation data, cluster all the operation modes according to multiple operation scenarios, determine multiple clustered operation modes corresponding to each operation scenario respectively, and clustered sub-operation data corresponding to each clustered operation mode respectively; Based on the clustered sub-operation data corresponding to each clustered operation mode respectively, determine the training data set; Obtain the inrush current values corresponding to each of the clustered operation modes respectively, and determine the labels corresponding to the training data set based on the inrush current values corresponding to all the clustered operation modes; Iteratively train the initial neural network model based on the training data set and the labels to obtain the inrush current prediction model.

[0008] A method for predicting inrush current of an islanded microgrid provided by the present invention, the step of clustering all the operation modes according to multiple operation scenarios based on all the sub-operation data to determine multiple clustered operation modes corresponding to each operation scenario respectively includes: Input the sub-operation data corresponding to each operation mode into a pre-constructed self-organizing mapping network to obtain multiple clustered operation modes corresponding to each operation scenario output by the self-organizing mapping network; the self-organizing mapping network is used to cluster all the operation modes according to all the operation scenarios.

[0009] A method for predicting inrush current of an islanded microgrid provided by the present invention, the self-organizing mapping network determines multiple clustered operation modes corresponding to each operation scenario through the following method: Based on the sub-operation data corresponding to each operation mode, perform iterative operations to determine the winning output layer neurons corresponding to each operation mode in each iteration round until the learning rate function corresponding to the current iteration round is greater than a preset learning rate threshold or the current iteration round reaches a preset maximum iteration round. Statistically analyze the winning output layer neurons corresponding to each operation mode in all iteration rounds to obtain the representative output layer neurons corresponding to each operation mode respectively; determine the operation modes with the same representative output layer neurons as multiple clustered operation modes corresponding to the same operation scenario; the winning output layer neuron corresponding to the operation mode is the output layer neuron with the smallest Euclidean distance from the sub-operation data corresponding to the operation mode in the self-organizing mapping network.

[0010] An impact current prediction method for an island microgrid provided by the present invention, which performs iterative operations based on sub-operation data corresponding to each of the operation modes to determine the winning output layer neurons corresponding to each of the operation modes in each iteration round, includes: In each iteration round, for each of the operation modes, based on the sub-operation data corresponding to the operation mode, multiple topological similarity degrees and multiple power flow similarity degrees corresponding to the operation mode, and the weights of each output layer neuron in the self-organizing mapping network corresponding to the current iteration round, determine the Euclidean distance from the operation mode to each output layer neuron, and determine the output layer neuron corresponding to the minimum Euclidean distance as the winning output layer neuron corresponding to the operation mode in the current iteration round; the topological similarity degree and the power flow similarity degree are determined based on the sub-operation data corresponding to all the operation modes; Among them, for each of the operation modes, the weights of each output layer neuron corresponding to the next iteration round are determined based on the sub-operation data corresponding to the operation mode, the weights of the current iteration round corresponding to the winning output layer neuron, the learning rate function and the winning neighborhood radius corresponding to the current iteration round, and the distances from each output layer neuron to the winning output layer neuron.

[0011] An impact current prediction method for an island microgrid provided by the present invention, the impact current prediction model includes a deep learning impact current prediction model and a transfer learning impact current prediction model; The iterative training of the initial neural network model based on the training data set and the labels to obtain the impact current prediction model includes: Iteratively train the initial neural network model based on the training data set and the labels to obtain the deep learning impact current prediction model; the deep learning impact current prediction model is a neural network model based on a convolutional neural network and a bidirectional long short-term memory network and introducing an attention mechanism; Add the new operation data corresponding to the new operation mode to the training data set to obtain the target domain data set, and obtain the target domain labels corresponding to the target domain data set; Based on the target domain data set and the target domain labels, iteratively train the deep learning impact current prediction model by using a preset transfer learning strategy to obtain the transfer learning impact current prediction model.

[0012] An impact current prediction method for an island microgrid provided by the present invention, the preset transfer learning strategy includes a local parameter fine-tuning strategy; Based on the target domain dataset and the target domain labels, iteratively training the deep learning impulse current prediction model using a preset transfer learning strategy to obtain the transfer learning impulse current prediction model, including: Adopting the local parameter fine-tuning strategy, freezing and fixing the weight parameters of a preset network layer in the deep learning impulse current prediction model, and iteratively training the target network layer in the deep learning impulse current prediction model except the preset network layer based on the target domain dataset and the target domain labels, and adjusting the weight parameters of the target network layer to obtain the transfer learning impulse current prediction model.

[0013] According to an impulse current prediction method for an islanded microgrid provided by the present invention, inputting the current operation data corresponding to the current operation mode of the islanded microgrid into a pre-trained impulse current prediction model, including: Determining an operation difference degree index corresponding to the current operation mode based on the current operation data corresponding to the current operation mode of the islanded microgrid and the target domain dataset; When the operation difference degree index is less than or equal to a preset difference degree threshold, inputting the current operation data corresponding to the current operation mode into the deep learning impulse current prediction model; When the operation difference degree index is greater than the preset difference degree threshold, inputting the current operation data corresponding to the current operation mode into the transfer learning impulse current prediction model.

[0014] The present invention also provides an impulse current prediction device for an islanded microgrid, including the following modules: A prediction module, configured to input the current operation data corresponding to the current operation mode of the islanded microgrid into a pre-trained impulse current prediction model to obtain an impulse current prediction result corresponding to the current operation mode of the islanded microgrid output by the impulse current prediction model; the impulse current prediction model is a neural network model trained based on a training dataset and labels corresponding to the training dataset.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, where when the processor executes the computer program, the impulse current prediction method for an islanded microgrid as described in any one of the above is implemented.

[0016] The present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the impulse current prediction method for an islanded microgrid as described in any one of the above is implemented.

[0017] The present invention also provides a computer program product, including a computer program which, when executed by a processor, implements the impact current prediction method for the island microgrid as described in any one of the above.

[0018] The impact current prediction method, device and electronic device for the island microgrid provided by the present invention input the current operation data corresponding to the current operation mode of the island microgrid into a pre-trained impact current prediction model to obtain the impact current prediction result corresponding to the current operation mode of the island microgrid output by the impact current prediction model; the impact current prediction model is a neural network model trained based on a training data set and labels corresponding to the training data set. The technical solution of the present invention can cope with various unknown scenarios through the impact current prediction model, and the determined impact current prediction result of the island microgrid is relatively accurate. There is no need to reconstruct the electromagnetic transient simulation model for each scenario, saving time cost and labor cost, and improving the efficiency of impact current prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0020] Figure 1 It is a schematic flowchart of the impact current prediction method for the island microgrid provided by the present invention.

[0021] Figure 2 It is a schematic structural diagram of the deep learning impact current prediction model provided by the present invention.

[0022] Figure 3 It is a schematic diagram of the preset transfer learning strategy provided by the present invention.

[0023] Figure 4 It is a schematic structural diagram of the impact current prediction device for the island microgrid provided by the present invention.

[0024] Figure 5 It is a schematic structural diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] To make the objectives, technical solutions and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the present invention with reference to the drawings in the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art without creative efforts based on the embodiments in the present invention belong to the scope of protection of the present invention.

[0026] Figure 1 It is a schematic flow chart of the impact current prediction method for the island microgrid provided by the present invention. As Figure 1 shown, the method includes the following: Step 110: Input the current operation data corresponding to the current operation mode of the island microgrid into the pre-trained impact current prediction model to obtain the impact current prediction result corresponding to the current operation mode of the island microgrid output by the impact current prediction model; the impact current prediction model is a neural network model trained based on the training data set and the labels corresponding to the training data set.

[0027] Specifically, the current operation data may include, for example, the DC voltage, DC current, AC voltage, AC current, active power, and reactive power of each wind turbine in the island microgrid; the DC voltage, DC current, AC voltage, AC current, active power, and reactive power of each photovoltaic power generation device; the voltage, current, state of charge (SOC), and operation state of each energy storage device; the bus voltage and frequency; the voltage, current, and active power of each load; the charging power (electric quantity) and discharging power (electric quantity) of the energy storage converter bidirectional inverter; the operation data of the reactive power compensation device; the current voltage and power at both ends of each inverter; the operation state, internal parameters, and control parameters of the inverter; the impact current data of each node, as well as data such as the number of nodes and the number of transmission lines.

[0028] For the impact current prediction method of the island microgrid provided by the present invention, the current operation data corresponding to the current operation mode of the island microgrid is input into the pre-trained impact current prediction model to obtain the impact current prediction result corresponding to the current operation mode of the island microgrid output by the impact current prediction model; the impact current prediction model is a neural network model trained based on the training data set and the labels corresponding to the training data set. The technical solution of the present invention can handle various unknown scenarios through the impact current prediction model, and the determined impact current prediction result of the island microgrid is relatively accurate. There is no need to reconstruct the electromagnetic transient simulation model for each scenario, saving time and labor costs and improving the efficiency of impact current prediction.

[0029] In one embodiment, the inputting the current operation data corresponding to the current operation mode of the island microgrid into the pre-trained impact current prediction model includes: Determine the operation difference degree index corresponding to the current operation mode based on the current operation data corresponding to the current operation mode of the island microgrid and the target domain data set; When the running difference degree index is less than or equal to the preset difference degree threshold, input the current running data corresponding to the current running mode into the deep learning impulse current prediction model; When the running difference degree index is greater than the preset difference degree threshold, input the current running data corresponding to the current running mode into the transfer learning impulse current prediction model.

[0030] Specifically, the impulse current prediction model includes a deep learning impulse current prediction model and a transfer learning impulse current prediction model. Both models are pre-trained neural network models. After obtaining the current running data corresponding to the current running mode of the island microgrid, the running difference degree index corresponding to the current running mode can be determined based on the current running data corresponding to the current running mode of the island microgrid and the target domain dataset. It should be noted that the current running mode of the island microgrid is not unique, and the current running data corresponding to multiple current running modes of the island microgrid can also be obtained. The above running difference degree index can be determined by the following formula: where, represents the target domain dataset, represents all the current running data, represents the number of samples in the target domain dataset, represents the number of current running data, represents the kernel function, for example, it can be a Gaussian kernel function, represents the th sample in the target domain dataset, represents the th sample in the target domain dataset, represents the th current running data, represents the th current running data.

[0031] Further, when the running difference index is less than or equal to a preset difference threshold, the current running data corresponding to the current running mode can be input into the deep learning inrush current prediction model, and the inrush current prediction result corresponding to the current running mode of the island microgrid output by the deep learning inrush current prediction model can be obtained. When the running difference index is greater than the preset difference threshold, the current running data corresponding to the current running mode can be input into the transfer learning inrush current prediction model, and the inrush current prediction result corresponding to the current running mode of the island microgrid output by the transfer learning inrush current prediction model can be obtained. Wherein, the preset difference threshold can be set as needed, and the embodiments of the present invention do not make specific limitations here.

[0032] In the above embodiment, different inrush current prediction models can be used when the running difference indexes corresponding to the current running modes are different. The differences between the deep learning inrush current prediction model and the transfer learning inrush current prediction model will be discussed in subsequent embodiments. Based on this difference, a more accurate inrush current prediction result can be ensured.

[0033] In one embodiment, the inrush current prediction model is trained through the following steps: Obtain the sub-running data corresponding to each of multiple running modes of the island microgrid; Based on all the sub-running data, cluster all the running modes according to multiple running scenarios, and determine the multiple clustered running modes corresponding to each running scenario, and the clustered sub-running data corresponding to each clustered running mode; Based on the clustered sub-running data corresponding to each clustered running mode, determine the training data set; Obtain the inrush current values corresponding to each of the clustered running modes, and determine the labels corresponding to the training data set based on the inrush current values corresponding to all the clustered running modes; Iteratively train the initial neural network model based on the training data set and the labels to obtain the inrush current prediction model.

[0034] Specifically, the inrush current prediction model can be trained through the following steps: First, the sub-running data corresponding to each of multiple running modes of the island microgrid can be obtained. The number of running modes can be determined by the number of running scenarios and the number of nodes and transmission lines in the island microgrid. For example, the number of running modes can be determined by the following formula: , where represents the number of running scenarios, represents the number of nodes, Indicates the number of transmission lines. Furthermore, based on all the sub-operation data, all the operation modes can be clustered according to multiple operation scenarios to determine multiple clustered operation modes corresponding to each operation scenario, and the clustered sub-operation data corresponding to each clustered operation mode can be further determined.

[0035] Furthermore, based on the clustered sub-operation data corresponding to each clustered operation mode, a training data set can be determined. It is easy to understand that the training data set can include all the clustered sub-operation data. The impulse current value corresponding to each clustered operation mode can also be obtained. The impulse current value can be obtained from the clustered sub-operation data or calculated through simulation, and the label corresponding to the training data set can be determined based on the impulse current values corresponding to all the clustered operation modes. Further, the initial neural network model can be iteratively trained based on the training data set and the label to obtain an impulse current prediction model.

[0036] In the above embodiment, all the operation modes are clustered according to multiple operation scenarios to determine multiple clustered operation modes corresponding to each operation scenario and the clustered sub-operation data corresponding to each clustered operation mode. Furthermore, the training data set and the label corresponding to the training data set are determined, ensuring the training process of the impulse current prediction model and the performance of the impulse current prediction model.

[0037] In one embodiment, the clustering of all the operation modes according to multiple operation scenarios based on all the sub-operation data to determine multiple clustered operation modes corresponding to each operation scenario includes: Inputting the sub-operation data corresponding to each operation mode into a pre-constructed self-organizing map network to obtain multiple clustered operation modes corresponding to each operation scenario output by the self-organizing map network; the self-organizing map network is used to cluster all the operation modes according to all the operation scenarios.

[0038] Specifically, a self-organizing map (SOM) network can be pre-constructed. The SOM network includes n input layer neurons and m×m output layer neurons, and the coordinates of each neuron can be represented by . After obtaining the sub-operation data corresponding to each operation mode, the sub-operation data corresponding to each operation mode can be input into the pre-constructed SOM network, and then multiple clustered operation modes corresponding to each operation scenario output by the SOM network can be obtained.

[0039] In the above embodiment, the use of the SOM network can achieve automatic clustering without manual data annotation. And the SOM network has a high generalization ability and can recognize various previous input samples.

[0040] In one embodiment, the self-organizing mapping network determines multiple post-clustering operation modes corresponding to each of the operation scenarios in the following manner: Based on the sub-operation data corresponding to each of the operation modes, iterative operations are performed to determine the winning output layer neurons corresponding to each of the operation modes in each iteration round until the learning rate function corresponding to the current iteration round is greater than a preset learning rate threshold or the current iteration round reaches a preset maximum number of iteration rounds. The winning output layer neurons corresponding to each of the operation modes in all iteration rounds are statistically analyzed to obtain the representative output layer neurons corresponding to each of the operation modes. The operation modes with the same representative output layer neurons are determined as multiple post-clustering operation modes corresponding to the same operation scenario. The winning output layer neuron corresponding to an operation mode is the output layer neuron in the self-organizing mapping network with the smallest Euclidean distance from the sub-operation data corresponding to the operation mode.

[0041] The self-organizing mapping network can determine multiple post-clustering operation modes corresponding to each operation scenario in the following manner: Specifically, based on the sub-operation data corresponding to each of the operation modes, iterative operations are performed to determine the winning output layer neurons corresponding to each of the operation modes in each iteration round until the learning rate function corresponding to the current iteration round is greater than a preset learning rate threshold or the current iteration round reaches a preset maximum number of iteration rounds. The winning output layer neurons corresponding to each of the operation modes in all iteration rounds can be statistically analyzed to obtain the representative output layer neurons corresponding to each of the operation modes. The preset maximum number of iteration rounds can be set as needed, and the embodiments of the present invention do not make specific limitations here. At the end of the iteration, the winning output layer neurons corresponding to each of the operation modes in each iteration round can be statistically analyzed. For example, in 100 iteration rounds, the winning output layer neurons corresponding to operation mode A include output layer neuron a, output layer neuron b, and output layer neuron c. Among them, output layer neuron a, output layer neuron b, and output layer neuron c win 50 times, 30 times, and 20 times respectively. Then, it can be determined that the output layer neuron a with the most winning times is the representative output layer neuron corresponding to operation mode A.

[0042] Further, after determining the representative output layer neurons corresponding to each operation mode, the operation modes with the same representative output layer neurons can be determined as multiple post-clustering operation modes corresponding to the same operation scenario, that is, multiple post-clustering operation modes corresponding to multiple operation scenarios can be determined.

[0043] Optionally, when the number of operation scenarios corresponding to all the finally determined post-clustering operation modes is less than In this case, the number of output layer neurons can be increased or the preset maximum number of iteration rounds can be increased, and the steps of the above embodiments are repeated until the number of operation scenarios corresponding to all the operation modes after final determination is equal to . This step can effectively mine the typical operation modes and extreme operation modes in the operation process of the island microgrid.

[0044] In the above embodiments, the SOM network can achieve automatic clustering without manual data annotation. And the SOM network has a high generalization ability and can recognize various previous input samples.

[0045] In one embodiment, the iterative operation is performed based on the sub-operation data respectively corresponding to each of the operation modes to determine the winning output layer neurons respectively corresponding to each of the operation modes in each iteration round, including: In each iteration round, for each of the operation modes, based on the sub-operation data corresponding to the operation mode, the multiple topological similarities and multiple power flow similarities corresponding to the operation mode, and the weights of the current iteration round respectively corresponding to each output layer neuron in the self-organizing mapping network, the Euclidean distance from the operation mode to each output layer neuron is determined, and the output layer neuron corresponding to the minimum Euclidean distance is determined as the winning output layer neuron corresponding to the operation mode in the current iteration round; the topological similarity and the power flow similarity are determined based on the sub-operation data corresponding to all the operation modes; Wherein, for each of the operation modes, the weights of the next iteration round respectively corresponding to each output layer neuron are determined based on the sub-operation data corresponding to the operation mode, the weights of the current iteration round corresponding to the winning output layer neuron, the learning rate function corresponding to the current iteration round and the radius of the winning neighborhood, and the distances from each output layer neuron to the winning output layer neuron.

[0046] Specifically, in each iteration round for each operation mode , the Euclidean distance from the operation mode to the output layer neuron can be determined based on the sub-operation data corresponding to the operation mode and the weights of the current iteration round corresponding to the output layer neuron in the self-organizing mapping network, where represents the Euclidean distance, where represents the specific data in the sub-operation data, the subscript of represents the number of the specific data, and the superscript represents the transpose, It can be expressed by the following formula: where min represents taking the minimum value, represents the operation mode and the operation mode of the topological similarity, represents the operation mode and the operation mode of the power flow similarity. The operation mode is each operation mode except the operation mode . and can both be determined by the sub-operation data corresponding to the operation mode and the operation mode respectively. The formula is as follows: where, represents the components and the number of components that exist in the operation mode but do not exist in the operation mode ; represents the components and the number of components that exist in the operation mode but do not exist in the operation mode ; represents the components and the number of components that exist in both the operation mode and the operation mode ; represents the node voltage of the th node in the operation mode of the islanded microgrid, represents the active power of the line connecting the th and the th nodes in the operation mode of the islanded microgrid, represents the reactive power of the line connecting the th and the th nodes in the operation mode of the islanded microgrid; and respectively represent the normalized values of the node voltages of the th node in the operation mode and the operation mode ; and respectively represent the operating modes and the operating mode in which the per-unit value of the active power of the line connecting the th th and the and respectively represent the operating mode and the operating mode in which the per-unit value of the reactive power of the line connecting the th and the th nodes, represents the base value of the active power of the line connecting the th and the th nodes, represents the base value of the reactive power of the line connecting the th and the th nodes, represents the per-unit value of the maximum voltage change of node , represents the set of nodes in the islanded microgrid, represents the set of transmission lines in the islanded microgrid.

[0047] Furthermore, after determining the Euclidean distances from the operating mode to each output layer neuron, all the Euclidean distances can be compared, and the output layer neuron corresponding to the smallest Euclidean distance is determined as the winning output layer neuron in the current iteration round for the operating mode . The coordinates of the winning output layer neuron can be expressed as .

[0048] It should be noted that for the operating mode , the weight of the output layer neuron corresponding to the next iteration round is determined based on the sub-operation data corresponding to the operating mode, the weight of the current iteration round corresponding to the winning output layer neuron , the learning rate function corresponding to the current iteration round , and the winning neighborhood radius , as well as the distance from the output layer neuron to the winning output layer neuron . It can be expressed by the following formula: Among them, represents the neighborhood function, represents the preset initial winning neighborhood radius, represents the preset initial learning rate function, represents the preset maximum number of iteration rounds. It can be seen from the above formula that in each iteration round, only the weights of the winning output layer neuron and the output layer neurons belonging to the neighborhood of the winning output layer neuron are updated, while the weights of other output layer neurons remain unchanged.

[0049] In the above embodiment, the determination method of the winning output layer neuron and the specific formula of the iterative process are described, ensuring the clustering of the operation mode, and the setting that the preset learning rate decreases with the iteration rounds ensures the convergence of the iterative process.

[0050] In one embodiment, the impulse current prediction model includes a deep learning impulse current prediction model and a transfer learning impulse current prediction model; Iteratively training the initial neural network model based on the training data set and the labels to obtain the impulse current prediction model includes: Iteratively training the initial neural network model based on the training data set and the labels to obtain the deep learning impulse current prediction model; the deep learning impulse current prediction model is a neural network model based on a convolutional neural network and a bidirectional long short-term memory network and introducing an attention mechanism; Adding new operation data corresponding to the new operation mode to the training data set to obtain a target domain data set, and obtaining target domain labels corresponding to the target domain data set; Iteratively training the deep learning impulse current prediction model based on the target domain data set and the target domain labels by using a preset transfer learning strategy to obtain the transfer learning impulse current prediction model.

[0051] Specifically, the impulse current prediction model includes a deep learning impulse current prediction model and a transfer learning impulse current prediction model. The initial neural network model can be iteratively trained based on the training data set and the labels to obtain the deep learning impulse current prediction model. The deep learning impulse current prediction model is a neural network model based on a convolutional neural network and a bidirectional long short-term memory network and introducing an attention mechanism. Figure 2 is a schematic structural diagram of the deep learning impulse current prediction model provided by the present invention, as Figure 2As shown in the figure, the deep learning impulse current prediction model may include an input layer, a Convolutional Neural Network (CNN) layer, a pooling layer, a Bidirectional Long Short-Term Memory (BiLSTM) layer, an attention layer, a fully connected layer, and an output layer. It is easy to understand that the initial neural network model is also a neural network model based on a convolutional neural network and a bidirectional long short-term memory network, and an attention mechanism is introduced. Further, new operating data corresponding to a new operating mode may be added to the training data set to obtain a target domain data set, and a target domain label corresponding to the target domain data set is obtained. It is easy to understand that the determination process of the target domain label is similar to the determination process of the label corresponding to the training data set. After obtaining the target domain data set and the target domain label, a preset transfer learning strategy is used to iteratively train the deep learning impulse current prediction model to obtain a transfer learning impulse current prediction model.

[0052] In the above embodiment, the deep learning impulse current prediction model is trained with the original training data set, and the transfer learning impulse current prediction model is trained with the updated target domain data set, so that the two models each have different advantages. In the process of model application, a suitable model can be selected according to the operation difference index corresponding to the current operation mode, which has strong engineering practicability and further improves the accuracy of impulse current prediction.

[0053] In one embodiment, the preset transfer learning strategy includes a local parameter fine-tuning strategy; Using the preset transfer learning strategy to iteratively train the deep learning impulse current prediction model based on the target domain data set and the target domain label to obtain the transfer learning impulse current prediction model includes: Using the local parameter fine-tuning strategy, freezing and fixing the weight parameters of a preset network layer in the deep learning impulse current prediction model, and iteratively training the target network layer other than the preset network layer in the deep learning impulse current prediction model based on the target domain data set and the target domain label, and adjusting the weight parameters of the target network layer to obtain the transfer learning impulse current prediction model.

[0054] Specifically, the preset transfer learning strategy includes a local parameter fine-tuning strategy, and the local parameter adjustment strategy may be as follows: Adopt a local parameter fine-tuning strategy, freeze and fix the weight parameters of the preset network layer in the deep learning impulse current prediction model, and iteratively train the target network layer in the deep learning impulse current prediction model except the preset network layer based on the target domain dataset and target domain labels, adjust the weight parameters of the target network layer, and obtain a transfer learning impulse current prediction model. Figure 3 It is a schematic diagram of the preset transfer learning strategy provided by the present invention. As Figure 3 shown, the weight parameters of the Attention layer in the deep learning impulse current prediction model can be frozen and fixed, and the weight parameters of other network layers can be adjusted to obtain a transfer learning impulse current prediction model.

[0055] Optionally, the preset transfer learning strategy may further include a global parameter fine-tuning strategy and a global parameter freezing strategy.

[0056] As Figure 3 shown, the global parameter fine-tuning strategy is to directly perform iterative training on the deep learning impulse current prediction model based on the target domain dataset and target domain labels to obtain a transfer learning impulse current prediction model.

[0057] The global parameter freezing strategy can freeze and fix the weight parameters of all network layers in the deep learning impulse current prediction model, add a new fully connected layer, and perform iterative training on the new fully connected layer in the deep learning impulse current prediction model based on the target domain dataset and target domain labels, adjust the weight parameters of the new fully connected layer, and obtain a transfer learning impulse current prediction model. As Figure 3 shown, the weight parameters of all network layers in the deep learning impulse current prediction model can be frozen and fixed, and fully connected layer 1 and fully connected layer 2 are added, and then the weight parameters of fully connected layer 1 and fully connected layer 2 are fine-tuned.

[0058] In the above embodiment, by adopting the local parameter fine-tuning strategy, the weight parameters of the preset network layer in the deep learning impulse current prediction model are frozen and fixed, and the target network layer in the deep learning impulse current prediction model except the preset network layer is iteratively trained based on the target domain dataset and target domain labels. It not only utilizes the results of training the deep learning impulse current prediction model in the previous steps, but also adjusts the target network layer based on the target domain dataset and target domain labels, so that the prediction performance of the transfer learning impulse current prediction model is further optimized.

[0059] Next, the impulse current prediction device for an island microgrid provided by the present invention will be described. The impulse current prediction device for an island microgrid described below can be mutually corresponding and referred to with the impulse current prediction method for an island microgrid described above.

[0060] Figure 4is a schematic structural diagram of the impact current prediction device for an islanded microgrid provided by the present invention. As Figure 4 shown, the impact current prediction device 400 for the islanded microgrid includes the following modules: A prediction module 410, configured to input the current operation data corresponding to the current operation mode of the islanded microgrid into a pre-trained impact current prediction model, and obtain the impact current prediction result corresponding to the current operation mode of the islanded microgrid output by the impact current prediction model; the impact current prediction model is a neural network model trained based on a training data set and labels corresponding to the training data set.

[0061] In one embodiment, the prediction module 410 is specifically configured to: Determine an operation difference degree index corresponding to the current operation mode based on the current operation data corresponding to the current operation mode of the islanded microgrid and the target domain data set; When the operation difference degree index is less than or equal to a preset difference degree threshold, input the current operation data corresponding to the current operation mode into the deep learning impact current prediction model; When the operation difference degree index is greater than the preset difference degree threshold, input the current operation data corresponding to the current operation mode into the transfer learning impact current prediction model.

[0062] In one embodiment, the impact current prediction device for the islanded microgrid further includes a model training module, and the model training module is specifically configured to: Obtain sub-operation data corresponding to multiple operation modes of the islanded microgrid respectively; Based on all the sub-operation data, cluster all the operation modes according to multiple operation scenarios, determine multiple clustered operation modes respectively corresponding to each operation scenario, and clustered sub-operation data respectively corresponding to each clustered operation mode; Based on the clustered sub-operation data respectively corresponding to each clustered operation mode, determine the training data set; Obtain the impact current values respectively corresponding to each clustered operation mode, and determine the labels corresponding to the training data set based on the impact current values corresponding to all the clustered operation modes; Iteratively train an initial neural network model based on the training data set and the labels to obtain the impact current prediction model.

[0063] In one embodiment, the model training module is further specifically configured to: Input the sub-operation data corresponding to each of the above operation modes into a pre-constructed self-organizing mapping network to obtain multiple clustered operation modes corresponding to each of the above operation scenarios output by the self-organizing mapping network; the self-organizing mapping network is used to cluster all the operation modes according to all the operation scenarios.

[0064] In one embodiment, the self-organizing mapping network determines multiple clustered operation modes corresponding to each of the operation scenarios in the following manner: Perform iterative operations based on the sub-operation data corresponding to each of the operation modes to determine the winning output layer neurons corresponding to each of the operation modes in each iteration round until the learning rate function corresponding to the current iteration round is greater than a preset learning rate threshold or the current iteration round reaches a preset maximum iteration round. Statistically analyze the winning output layer neurons corresponding to each of the operation modes in all iteration rounds to obtain the representative output layer neurons corresponding to each of the operation modes; determine the operation modes with the same representative output layer neurons as the multiple clustered operation modes corresponding to the same operation scenario; the winning output layer neuron corresponding to the operation mode is the output layer neuron in the self-organizing mapping network with the smallest Euclidean distance from the sub-operation data corresponding to the operation mode.

[0065] In one embodiment, the performing iterative operations based on the sub-operation data corresponding to each of the operation modes to determine the winning output layer neurons corresponding to each of the operation modes in each iteration round includes: In each iteration round, for each of the operation modes, based on the sub-operation data corresponding to the operation mode, multiple topological similarity degrees and multiple power flow similarity degrees corresponding to the operation mode, and the weights of each output layer neuron in the self-organizing mapping network corresponding to the current iteration round, determine the Euclidean distance from the operation mode to each output layer neuron, and determine the output layer neuron corresponding to the minimum Euclidean distance as the winning output layer neuron corresponding to the operation mode in the current iteration round; the topological similarity degree and the power flow similarity degree are determined based on the sub-operation data corresponding to all the operation modes; wherein, for each of the operation modes, the weights of each output layer neuron corresponding to the next iteration round are determined based on the sub-operation data corresponding to the operation mode, the weights of the current iteration round corresponding to the winning output layer neuron, the learning rate function corresponding to the current iteration round and the radius of the winning neighborhood, and the distances from each output layer neuron to the winning output layer neuron.

[0066] In one embodiment, the inrush current prediction model includes a deep learning inrush current prediction model and a transfer learning inrush current prediction model; the model training module is specifically further configured to: Iteratively train the initial neural network model based on the training data set and the labels to obtain the deep learning impulse current prediction model; the deep learning impulse current prediction model is a neural network model based on a convolutional neural network and a bidirectional long short-term memory network, and introducing an attention mechanism; Add the new operation data corresponding to the new operation mode to the training data set to obtain a target domain data set, and obtain the target domain labels corresponding to the target domain data set; Based on the target domain data set and the target domain labels, iteratively train the deep learning impulse current prediction model using a preset transfer learning strategy to obtain the transfer learning impulse current prediction model.

[0067] In one embodiment, the preset transfer learning strategy includes a local parameter fine-tuning strategy; specifically, the model training module is further configured to: Adopt the local parameter fine-tuning strategy to freeze and fix the weight parameters of the preset network layer in the deep learning impulse current prediction model, and iteratively train the target network layer in the deep learning impulse current prediction model except the preset network layer based on the target domain data set and the target domain labels, and adjust the weight parameters of the target network layer to obtain the transfer learning impulse current prediction model.

[0068] The impulse current prediction device for an island microgrid provided by the present invention inputs the current operation data corresponding to the current operation mode of the island microgrid into a pre-trained impulse current prediction model to obtain an impulse current prediction result corresponding to the current operation mode of the island microgrid output by the impulse current prediction model; the impulse current prediction model is a neural network model trained based on a training data set and the labels corresponding to the training data set. The technical solution of the present invention can cope with various unknown scenarios through the impulse current prediction model, and the determined impulse current prediction result of the island microgrid is relatively accurate, without having to reconstruct the electromagnetic transient simulation model for each scenario, saving time and labor costs and improving the efficiency of impulse current prediction.

[0069] Figure 5 Illustrates a schematic physical structure diagram of an electronic device, as Figure 5 shown. The electronic device may include: a processor 510, a communication interface 520, a memory 530, and a communication bus 540. Among them, the processor 510, the communication interface 520, and the memory 530 communicate with each other through the communication bus 540. The processor 510 can call the logical instructions in the memory 530 to execute the impulse current prediction method for the island microgrid, and the method includes: Input the current operation data corresponding to the current operation mode of the island microgrid into a pre-trained inrush current prediction model to obtain the inrush current prediction result corresponding to the current operation mode of the island microgrid output by the inrush current prediction model; the inrush current prediction model is a neural network model trained based on a training data set and the labels corresponding to the training data set.

[0070] In addition, when the logical instructions in the above-mentioned memory 530 are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs that can store program codes.

[0071] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the inrush current prediction method of the island microgrid provided by the above-mentioned various methods. The method includes: Input the current operation data corresponding to the current operation mode of the island microgrid into a pre-trained inrush current prediction model to obtain the inrush current prediction result corresponding to the current operation mode of the island microgrid output by the inrush current prediction model; the inrush current prediction model is a neural network model trained based on a training data set and the labels corresponding to the training data set.

[0072] On another aspect, the present invention also provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it is implemented to execute the inrush current prediction method of the island microgrid provided by the above-mentioned various methods. The method includes: Input the current operation data corresponding to the current operation mode of the island microgrid into a pre-trained inrush current prediction model to obtain the inrush current prediction result corresponding to the current operation mode of the island microgrid output by the inrush current prediction model; the inrush current prediction model is a neural network model trained based on a training data set and the labels corresponding to the training data set.

[0073] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without creative work.

[0074] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the essence of the above technical solution or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features. These modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting inrush current of an isolated island microgrid, characterized in that: include: The current operating data corresponding to the current operating mode of the isolated island microgrid is input into a pre-trained impulse current prediction model to obtain an impulse current prediction result corresponding to the current operating mode of the isolated island microgrid output by the impulse current prediction model; the impulse current prediction model is a neural network model trained based on a training data set and labels corresponding to the training data set.

2. The method for predicting the inrush current of an isolated island microgrid according to claim 1, characterized in that: The impulse current prediction model is obtained by training through the following steps: Obtain sub-operation data corresponding to multiple operation modes of the isolated island microgrid; Based on all the sub-operation data, cluster all the operation modes according to multiple operation scenarios, determine multiple clustered operation modes corresponding to each of the operation scenarios, and clustered sub-operation data corresponding to each of the clustered operation modes; Determining the training data set based on the clustering sub-operation data corresponding to each of the post-clustering operation modes; Obtaining the inrush current values ​​corresponding to the respective post-clustering operation modes, and determining the label corresponding to the training data set based on the inrush current values ​​corresponding to all the post-clustering operation modes; The initial neural network model is iteratively trained based on the training data set and the labels to obtain the inrush current prediction model.

3. The method for predicting the inrush current of an isolated island microgrid according to claim 2, characterized in that: The clustering of all the operation modes according to a plurality of operation scenarios based on all the sub-operation data, and determining a plurality of clustered operation modes corresponding to each of the operation scenarios, includes: The sub-operation data corresponding to each of the operation modes are input into a pre-constructed self-organizing map network to obtain a plurality of clustered operation modes corresponding to each of the operation scenarios output by the self-organizing map network; the self-organizing map network is used to cluster all of the operation modes according to all of the operation scenarios.

4. The method for predicting the inrush current of an isolated island microgrid according to claim 3, characterized in that: The self-organizing map network determines a plurality of post-clustering operation modes corresponding to each of the operation scenarios in the following manner: Iterative calculations are performed based on the sub-operation data corresponding to each of the operation modes, and winning output layer neurons corresponding to each of the operation modes in each iteration round are determined, until the learning rate function corresponding to the current iteration round is greater than a preset learning rate threshold or the current iteration round reaches a preset maximum iteration round, and the winning output layer neurons corresponding to each of the operation modes in all iteration rounds are counted to obtain representative output layer neurons corresponding to each of the operation modes; The same operating modes representing output layer neurons are determined as multiple clustered operating modes corresponding to the same operating scenario; the winning output layer neuron corresponding to the operating mode is the output layer neuron with the smallest Euclidean distance to the sub-operation data corresponding to the operating mode in the self-organizing mapping network.

5. The method for predicting the inrush current of an isolated island microgrid according to claim 4, characterized in that: The iterative operation is performed based on the sub-operation data corresponding to each of the operation modes to determine the winning output layer neurons corresponding to each of the operation modes in each iteration round, including: In each iteration round, for each of the operation modes, based on the sub-operation data corresponding to the operation mode, multiple topological similarities and multiple power flow similarities corresponding to the operation mode, and the current iteration round weights corresponding to each output layer neuron in the self-organizing map network, determine the Euclidean distance from the operation mode to each output layer neuron, and determine the output layer neuron corresponding to the minimum Euclidean distance as the winning output layer neuron corresponding to the operation mode in the current iteration round; the topological similarity and the power flow similarity are determined based on the sub-operation data corresponding to all the operation modes; Among them, for each of the operating modes, the next iteration round weight corresponding to each of the output layer neurons is determined based on the sub-operation data corresponding to the operating mode, the current iteration round weight corresponding to the winning output layer neuron, the learning rate function and winning neighborhood radius corresponding to the current iteration round, and the distance from each of the output layer neurons to the winning output layer neuron.

6. The method for predicting the inrush current of an isolated island microgrid according to any one of claims 2 to 5, characterized in that: The impulse current prediction model includes a deep learning impulse current prediction model and a transfer learning impulse current prediction model; The iterative training of the initial neural network model based on the training data set and the label to obtain the inrush current prediction model includes: Iteratively training the initial neural network model based on the training data set and the labels to obtain the deep learning impulse current prediction model; the deep learning impulse current prediction model is a neural network model based on a convolutional neural network and a bidirectional long short-term memory network, and introducing an attention mechanism; Adding new operation data corresponding to the new operation mode to the training data set to obtain a target domain data set, and obtaining a target domain label corresponding to the target domain data set; Based on the target domain data set and the target domain label, the deep learning impulse current prediction model is iteratively trained using a preset transfer learning strategy to obtain the transfer learning impulse current prediction model.

7. The method for predicting the inrush current of an isolated island microgrid according to claim 6, characterized in that: The preset transfer learning strategy includes a local parameter fine-tuning strategy; The method of iteratively training the deep learning impulse current prediction model based on the target domain data set and the target domain label using a preset transfer learning strategy to obtain the transfer learning impulse current prediction model includes: The local parameter fine-tuning strategy is adopted to freeze the weight parameters of the preset network layer in the deep learning impulse current prediction model, and the target network layer other than the preset network layer in the deep learning impulse current prediction model is iteratively trained based on the target domain data set and the target domain label, and the weight parameters of the target network layer are adjusted to obtain the transfer learning impulse current prediction model.

8. The method for predicting the inrush current of an isolated island microgrid according to claim 6, characterized in that: The step of inputting current operation data corresponding to the current operation mode of the isolated island microgrid into the pre-trained inrush current prediction model includes: Determine an operation difference index corresponding to the current operation mode based on current operation data corresponding to the current operation mode of the isolated microgrid and the target domain data set; When the operation difference index is less than or equal to a preset difference threshold, inputting current operation data corresponding to the current operation mode into the deep learning inrush current prediction model; When the operation difference index is greater than the preset difference threshold, the current operation data corresponding to the current operation mode is input into the transfer learning inrush current prediction model.

9. An inrush current prediction device for an isolated island microgrid, characterized in that: include: A prediction module is used to input current operating data corresponding to the current operating mode of the isolated island microgrid into a pre-trained impulse current prediction model to obtain an impulse current prediction result corresponding to the current operating mode of the isolated island microgrid output by the impulse current prediction model; the impulse current prediction model is a neural network model trained based on a training data set and labels corresponding to the training data set.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the method for predicting the impact current of the isolated island microgrid according to any one of claims 1 to 8 is implemented.

Images