Power Grid Operation Environment Simulation Method Based on Deep Reinforcement Learning

Abstract

The present invention provides a power grid operation environment deduction method for deep reinforcement learning, which relates to the field of data processing technology. The power grid operation environment is classified based on the attributes of equipment in the power grid to obtain power generation side equipment, power storage side equipment and power consumption side equipment; deep learning training is performed according to the historical power data of each device to obtain a power generation sub-neural network, a power storage sub-neural network and a power consumption sub-neural network respectively; the power generation sub-neural network, the power storage sub-neural network and the power consumption sub-neural network are assembled according to the data association relationship between the power generation side equipment, the power storage side equipment and the power consumption side equipment to obtain a total neural deduction network; variable deduction parameters corresponding to the deduction target are obtained and input into the total neural deduction network to obtain a predicted deduction result; the predicted deduction result is compared with the actual result to obtain multi-dimensional training data, and the weights of neurons in the total neural deduction network and the excitation function are trained and updated based on the multi-dimensional training data.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a power grid operation environment deduction method for deep reinforcement learning. Background Art

[0002] At present, with the rapid development of social economy and the increasing demand for electricity in human production and life, the operation and management of power grids have become more and more complicated. The power system includes three main links: power generation, power storage and power consumption. Various types of equipment play a key role in different links. In order to ensure the stability and efficiency of power supply, the power grid needs to operate stably under various environmental conditions and be able to respond to emergencies. Therefore, accurately predicting the operation of the power grid and optimizing the allocation of power resources have become important issues for power management departments.

[0003] Existing deep reinforcement learning methods usually uniformly model the entire power grid, ignoring the different characteristics of the equipment on the power generation side, power storage side, and power consumption side, resulting in low generalization ability and prediction accuracy of the model. In addition, most existing technologies use a single neural network structure, which is insufficient to describe the multi-level and multi-dimensional characteristics of complex power systems. Therefore, in practical applications, the prediction results of existing methods often deviate greatly from the actual situation and cannot meet the needs of high-precision power grid operation prediction.

[0004] Therefore, how to effectively predict and deduce the grid operating environment based on the characteristics of grid equipment and achieve high-precision prediction of grid operation conditions has become a key issue that needs to be solved urgently. Summary of the invention

[0005] The present invention provides a power grid operating environment deduction method for deep reinforcement learning, which can effectively predict and deduce the power grid operating environment according to the characteristics of power grid equipment, and realize high-precision prediction of power grid operation conditions.

[0006] A first aspect of the present invention provides a method for deducing a power grid operation environment for deep reinforcement learning, comprising:

[0007] Based on the properties of the equipment in the power grid, the power grid operating environment is classified into power generation side equipment, power storage side equipment and power consumption side equipment;

[0008] Deep learning training is performed based on the historical power data of the power generation side equipment, the power storage side equipment and the power consumption side equipment to obtain the power generation sub-neural network, the power storage side neural network and the power consumption side neural network respectively;

[0009] According to the data association relationship between the power generation side equipment, the power storage side equipment and the power consumption side equipment, the power generation electronic neural network, the power storage electronic neural network and the power consumption electronic neural network are assembled based on the data association relationship to obtain the overall neural deduction network;

[0010] Obtaining variable deduction parameters corresponding to the deduction target, and inputting the variable deduction parameters into the overall neural deduction network to obtain a predicted deduction result;

[0011] The predicted deduction results are compared with the actual results to obtain multi-dimensional training data, and the weights of neurons in the total neural deduction network and the activation function training and update are performed based on the multi-dimensional training data.

[0012] Optionally, in a possible implementation manner of the first aspect, classifying the power grid operating environment based on attributes of devices in the power grid to obtain power generation side devices, power storage side devices, and power consumption side devices includes:

[0013] Acquire the attributes of the equipment in the power grid, and if it is determined that the equipment has any one of the power generation attribute, the power storage attribute or the power consumption attribute, add the corresponding first label and classify it into the power generation set, the power storage set or the power consumption set;

[0014] If it is determined that the device has any of the power generation attribute, power storage attribute or power consumption attribute, then a corresponding second tag is added and the device is classified into the power generation set, power storage set or power consumption set;

[0015] The power generation group includes power generation side equipment with power generation function, the power storage group includes power storage side equipment with power storage function, and the power consumption group includes power consumption side equipment with power consumption function.

[0016] Optionally, in a possible implementation of the first aspect, performing deep learning training according to historical power data of power generation side equipment, power storage side equipment, and power consumption side equipment to obtain power generation sub-neural network, power storage sub-neural network, and power consumption sub-neural network respectively includes:

[0017] Traversing the IDs of the devices in the power generation set, the power storage set, and the power consumption set in sequence, and acquiring historical power consumption information within a preset time period based on the IDs to obtain historical power data, wherein the historical power consumption information at least includes power information value, environmental information value, and time information value;

[0018] Constructing a corresponding initial neural network based on the types of equipment in the power generation set, the power storage set, and the power consumption set, wherein the initial neural network includes initial neural nodes;

[0019] Based on the objective reference attributes of the power grid operation environment, the initial neural nodes are reviewed and adjusted to obtain the processed power generation neural network, storage neural network and utilization neural network.

[0020] Optionally, in a possible implementation manner of the first aspect, the types of equipment in the power generation set, the power storage set, and the power consumption set construct a corresponding initial neural network, and the initial neural network includes initial neural nodes, including:

[0021] Initialize the initial neural network corresponding to each power generation set, power storage set and power consumption set;

[0022] The initial neural network includes an initial input layer, an initial hidden layer and an initial output layer, traverses all initial input neurons in the initial input layer, and determines the attributes of its input information to obtain corresponding neuron labels;

[0023] The initial hidden neurons and initial output neurons directly or indirectly connected to each initial input neuron are obtained to obtain the corresponding neural sub-paths, and the neuron labels are matched with the neural sub-paths.

[0024] Optionally, in a possible implementation of the first aspect, the review and adjustment processing of the initial neural nodes based on the objective reference attributes of the power grid operation environment to obtain the processed power generation neural network, storage neural network and use neural network includes:

[0025] Obtain objective reference attributes of the region corresponding to the power grid operation environment to obtain objective information, and compare the objective information with the neuron label of the initial input neuron, each objective information having a preset corresponding neuron label;

[0026] If it is determined that the objective information corresponds to the neuron labels one by one, then the power generation neural network, the power storage neural network and the power consumption neural network are obtained based on the initial neural networks corresponding to the power generation set, the power storage set and the power consumption set;

[0027] If it is determined that the objective information does not correspond to the neuron label, the objective information is compared with the neuron label to obtain the different neuron labels. Based on the different neuron labels, the initial neural networks corresponding to the power generation set, the power storage set and the power consumption set are processed to obtain the power generation sub-neural network, the power storage sub-neural network and the power consumption sub-neural network.

[0028] Optionally, in a possible implementation of the first aspect, if it is determined that the objective information does not correspond to the neuron label, the objective information is compared with the neuron label to obtain a different neuron label, and the initial neural networks corresponding to the power generation set, the power storage set, and the power consumption set are processed based on the different neuron labels to obtain the power generation sub-neural network, the power storage sub-neural network, and the power consumption sub-neural network, including:

[0029] Determine the neuron labels that exist in the neuron labels and do not correspond to the objective information as the difference neuron labels;

[0030] Selecting neuron labels and neural sub-paths in the initial neural network based on the different neuron labels, and extracting upper-dimensional neural nodes connected to upper-dimensional connections of each neural sub-node in the neural sub-path;

[0031] The properties of the upper-dimensional neural nodes are extracted, and the corresponding neural sub-paths are adjusted based on the node correction method pre-configured by the different neuron labels and the properties of the upper-dimensional neural nodes to obtain the transmitting electronic neural network, the storing electronic neural network and the using electronic neural network.

[0032] Optionally, in a possible implementation of the first aspect, the extracting the attributes of the upper-dimensional neural nodes, adjusting the corresponding neural sub-paths based on the node correction method pre-configured by the different neuron labels and the attributes of the upper-dimensional neural nodes, and obtaining the transmitting electronic neural network, the storing electronic neural network and the using electronic neural network, include:

[0033] Get the properties of the upper-dimensional neural node corresponding to each neural sub-node. If the upper-dimensional neural node is connected to the initial input neuron through only one path, directly delete the neuron path between the upper-dimensional neural node and the initial input neuron.

[0034] If the upper-dimensional neural node is connected to multiple initial input neurons through multiple paths, then the neuron labels that are differences and the neuron labels that are not differences among the multiple initial input neurons are obtained;

[0035] Comparing the information in the difference neuron labels and the non-difference neuron labels with the pre-configured node correction method, the weight and / or activation function corresponding to the upper dimensional neural node is obtained;

[0036] After adjusting all neural sub-paths, we get the electron-transmitting neural network, the electron-storing neural network and the electron-using neural network.

[0037] Optionally, in a possible implementation manner of the first aspect, the node correction manner includes weights and / or activation functions corresponding to each combination of different neuron labels and non-different neuron labels.

[0038] Optionally, in a possible implementation of the first aspect, according to the data association relationship among the power generation side equipment, the power storage side equipment and the power consumption side equipment, assembling the power generation electronic neural network, the power storage electronic neural network and the power consumption electronic neural network based on the data association relationship to obtain the total neural deduction network includes:

[0039] Obtain the working attributes of the equipment in the power generation side equipment, the power storage side equipment and the power consumption side equipment, and count all the working attributes to obtain the attribute set;

[0040] Combining the work attributes in turn to obtain work attribute groups, determining the association relationship corresponding to each work attribute group, and each work attribute group has a preset association relationship;

[0041] Based on the association relationship, the neural nodes corresponding to the transmitting electronic neural network, the storing electronic neural network and the using electronic neural network are selected and connected to obtain the total neural deduction network, and each association relationship has preset neural nodes that need to be associated and connected.

[0042] Optionally, in a possible implementation of the first aspect, obtaining variable deduction parameters corresponding to the deduction target, and inputting the variable deduction parameters into the overall neural deduction network to obtain the predicted deduction result includes:

[0043] The variable deduction parameters corresponding to the deduction target are obtained and decomposed to obtain multiple decomposition sub-information. The neural nodes are selected according to the information corresponding to each decomposition sub-information and input into the overall neural deduction network to obtain the predicted deduction result.

[0044] Optionally, in a possible implementation of the first aspect, the step of comparing the predicted deduction result with the actual result to obtain multi-dimensional training data, and training and updating the weights of neurons and the activation function in the total neural deduction network based on the multi-dimensional training data includes:

[0045] The prediction results are split according to different dimensions to obtain multiple prediction sub-information, and the prediction sub-information of each dimension is compared with the actual sub-information of the corresponding dimension in the actual result to obtain a comparison sub-information group of multiple dimensions;

[0046] Calculate the comparison difference between the predicted sub-information and the actual sub-information in each comparison sub-information group, and obtain the neuron output by the corresponding predicted sub-information in the total neural deduction network as the target neuron;

[0047] Acquire a chain formed by neurons participating in calculation in an upper dimension directly or indirectly connected to the target neuron as a chain to be trained;

[0048] Based on the comparison difference, the neurons in the training chain are trained and updated to obtain the total neural inference network after training.

[0049] Optionally, in a possible implementation of the first aspect, the updating of neuron training in the training chain based on the comparison difference to obtain a trained total neural deduction network includes:

[0050] Extract neurons other than the target neurons in the chain to be trained as neurons to be trained;

[0051] The number of neurons in the to-be-trained neurons that are in the same sub-neural network as the target neuron is obtained to obtain a first number, and the number of neurons in the to-be-trained neurons that are not in the same sub-neural network as the target neuron is obtained to obtain a second number;

[0052] The number of intervals between the neuron to be trained and the target neuron is obtained, and the weight and / or activation function of the neuron to be trained and updated based on the first number, the second number, the number of intervals, and the comparison difference.

[0053] Optionally, in a possible implementation manner of the first aspect, obtaining the number of intervals between the neuron to be trained and the target neuron, and training and updating the weight and / or activation function of the neuron to be trained based on the first number, the second number, the number of intervals, and the comparison difference, includes:

[0054] Based on the first number and the second number, a first reference adjustment weight of a neuron in the same sub-neural network as the target neuron and a second reference adjustment weight of a neuron not in the same sub-neural network as the target neuron are calculated;

[0055] The first benchmark adjustment weight and the second benchmark adjustment weight are offset-adjusted based on the number of intervals, and the weight and / or activation function training of the neuron to be trained are updated based on the offset-adjusted first benchmark adjustment weight, the second benchmark adjustment weight and the comparison difference.

[0056] The first benchmark adjustment weight and the second benchmark adjustment weight are calculated by the following formula

[0057] ;

[0058] in, is the first benchmark adjustment weight before offset adjustment, is the second benchmark adjustment weight before offset adjustment, is the first calculated weight, is the first quantity, Calculate the weight for the second, is the second quantity, is the first benchmark adjustment weight after offset adjustment, A is the number of intervals, is the first interval coefficient, is the second benchmark adjustment weight after offset adjustment, B is the number of intervals of neurons in different sub-neural networks, is the second interval coefficient;

[0059] The updated weights and / or activation functions are calculated by the following formula,

[0060] ;

[0061] in, is the weight before updating, is the updated weight, is the K value of the activation function before updating, is the K value of the updated activation function.

[0062] According to a second aspect of the present invention, a storage medium is provided, in which a computer program is stored. When the computer program is executed by a processor, it is used to implement the method of the first aspect of the present invention and various possible designs of the first aspect.

[0063] The beneficial effects of the present invention are as follows:

[0064] 1. The present invention can effectively predict and deduce the power grid operating environment according to the characteristics of the power grid equipment, and realize high-precision prediction of the power grid operation status. First, the present invention can classify the power grid operating environment according to the properties of the equipment in the power grid, so that the corresponding neural network can be generated later to accurately deduce the power grid operation status. Secondly, deep learning training can be performed based on the historical power data of different types of equipment to obtain the corresponding power generation neural network, storage neural network and electronic neural network. The prediction accuracy of the network model can be improved through data training, and the power generation neural network, storage neural network and electronic neural network can be assembled according to the data association relationship between different devices to obtain a complex total neural deduction network to improve the accuracy of the power grid prediction results. At the same time, the present invention can also compare and train the predicted results with the actual generated results to increase the iterative training of the neural network model and improve the accuracy of the prediction deduction.

[0065] 2. The present invention can introduce the objective reference attributes of the power grid operating environment, and review and adjust the initial neural nodes, so that the generated electronic neural network, electronic storage neural network and electronic use neural network are more in line with the actual power grid operating environment. Among them, the present invention can determine the corresponding objective information based on the objective reference attributes, and then determine the different neuron labels based on the objective information, so as to determine the upper-dimensional neural nodes, so as to adopt the corresponding node correction method according to the number of paths corresponding to the upper-dimensional neural nodes, and correct the corresponding sub-neural paths, so as to avoid irrelevant factors interfering with the prediction results of the model. This review and adjustment method can significantly improve the prediction accuracy and adaptability of the neural network model, and ensure that the model can have a high prediction accuracy in different power grid operating environments.

[0066] 3. The present invention can perform iterative training processing of multi-dimensional training data on the total neural deduction network to achieve high-precision prediction of the operation of the power grid. Among them, through the data association relationship between the power generation side equipment, the power storage side equipment and the power consumption side equipment, the power generation sub-neural network, the power storage sub-neural network and the power consumption sub-neural network are assembled to form a complete total neural deduction network. Then, the variable deduction parameters of the predicted target date are obtained and input into the total neural deduction network to obtain the predicted deduction results of the power grid. Furthermore, by comparing the predicted results with the actual results, multi-dimensional training data is obtained, and the weights and excitation functions in the total neural deduction network are trained and updated. Through continuous iterative training, the performance of the neural network is optimized, the accuracy and reliability of the prediction are improved, and the problems of poor model generalization ability and insufficient prediction accuracy in the prior art are effectively solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A schematic diagram of a flow chart of a method for deducing a power grid operating environment for deep reinforcement learning provided by the present invention;

[0068] Figure 2 A schematic diagram of a neural sub-pathway provided by the present invention. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0070] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0071] like Figure 1 As shown, the present invention provides a flow chart of a method for deducing a power grid operating environment for deep reinforcement learning, and the method for deducing a power grid operating environment for deep reinforcement learning includes:

[0072] S1, based on the attributes of the equipment in the power grid, the power grid operating environment is classified into power generation side equipment, power storage side equipment and power consumption side equipment.

[0073] It is understandable that in actual operation, the power grid requires a variety of equipment to work together to achieve normal operation of the power grid. Many devices constitute the power grid operating environment. Therefore, in order to subsequently predict and deduce the power grid operation based on relevant power data, the equipment in the power grid operating environment can be classified according to the equipment attributes, thereby obtaining the power generation side equipment, the power storage side equipment and the power consumption side equipment, so that the corresponding neural network can be generated subsequently to accurately deduce the power grid operation.

[0074] Among them, the power generation side equipment is the equipment that produces electricity in the power grid, the power storage side equipment is the equipment that stores the produced electricity, and the power consumption side equipment is the equipment that consumes electricity.

[0075] In some embodiments, the specific implementation of step S1 (classifying the power grid operating environment based on the attributes of the devices in the power grid to obtain power generation side devices, power storage side devices and power consumption side devices) includes:

[0076] S11, obtaining the attributes of the equipment in the power grid, if it is determined that the equipment has any one of the power generation attribute, power storage attribute or power consumption attribute, then adding the corresponding first label and classifying it into the power generation set, power storage set or power consumption set.

[0077] It should be noted that the attributes of equipment in the power grid usually include power generation attributes, power storage attributes and power consumption attributes. Each power grid equipment can have one attribute or multiple attributes. Therefore, according to the number of attributes of the equipment, corresponding labels can be added for classification, so that the corresponding sub-neural network can be generated according to the corresponding classification set, so as to realize the subsequent accurate prediction and deduction of the power grid operating environment.

[0078] It is understandable that when it is determined that a device has only one attribute, a first label can be added to the corresponding device and classified into a corresponding set, wherein the first label is a device label with only one attribute, the power generation set is a set of devices with power generation attributes, the power storage set is a set of devices with power storage attributes, and the power consumption set is a set of devices with power consumption attributes.

[0079] For example, when the device is a solar photovoltaic panel, the corresponding device attribute is only one power generation attribute, so the solar photovoltaic panel can be added as the first label and classified into the power generation set. When the device is an electric motor, fan, etc., the corresponding device attribute is only one power consumption attribute, so the first label can be added to the electric motor and fan, and they can be classified into the power consumption set.

[0080] S12: If it is determined that the device has any of the power generation attribute, the power storage attribute or the power consumption attribute, a corresponding second tag is added and the device is classified into the power generation set, the power storage set or the power consumption set.

[0081] It can be understood that when the device is a battery, it can store electricity. When the electrical equipment consumes electricity, the battery can also discharge, so it has storage properties and power generation properties, and has multiple properties. Therefore, a second label can be added to the device with multiple properties, and the corresponding device can be classified into a set with attributes. For example, the battery can be classified into the power generation set, and can also be added to the power storage set.

[0082] The second tag is a device tag with multiple attributes. For example, when the device has both power storage and power generation attributes, a power generation and storage tag may be added to mark the device.

[0083] S13, the power generation set includes power generation side equipment with power generation function, the power storage set includes power storage side equipment with power storage function, and the power consumption set includes power consumption side equipment with power consumption function.

[0084] It can be understood that the devices in different sets have corresponding attribute functions, that is, the power generation equipment in the power generation set has the power generation function, the power storage equipment in the power storage set has the power storage function, and the power consumption equipment in the power consumption set has the power consumption function.

[0085] S2, deep learning training is performed based on the historical power data of the power generation side equipment, the power storage side equipment and the power consumption side equipment to obtain the power generation electronic neural network, the power storage electronic neural network and the power consumption electronic neural network respectively.

[0086] It is understandable that in order to accurately predict the operation of the corresponding power grid and input multi-dimensional data for deduction and prediction, it is necessary to divide various types of equipment into multiple sub-neural networks so that a detailed total neural network can be obtained based on the corresponding related data, so as to accurately predict the power operation conditions in different dates and regions.

[0087] Among them, the historical power data is the power data of the corresponding equipment in the historical time period, the power generation sub-neural network is the sub-neural network corresponding to the power generation equipment, the storage sub-neural network is the sub-neural network corresponding to the equipment with power storage properties, and the power consumption sub-neural network is the sub-neural network corresponding to the equipment with power consumption properties.

[0088] In some embodiments, the specific implementation of step S2 (the deep learning training is performed according to the historical power data of the power generation side equipment, the power storage side equipment and the power consumption side equipment to obtain the power generation sub-neural network, the power storage side neural network and the power consumption side neural network respectively) includes:

[0089] S21, traverse the IDs of the devices in the power generation set, the power storage set and the power consumption set in turn, and obtain the historical power consumption information within a preset time period based on the ID to obtain historical power data, wherein the historical power consumption information at least includes power information value, environmental information value and time information value.

[0090] It can be understood that the device ID means that each device has a unique identifier, which is used to determine the identity of a unique device in a collection. The historical power usage information refers to the relevant information about the power usage of the device within a preset time period (such as data from the past year). The power usage record of each device includes at least power information values ​​(such as current, voltage, etc.), environmental information values ​​(such as temperature, humidity, etc.) and time information values ​​(specific time points). The historical power data is the power data corresponding to the historical power usage information.

[0091] For example, suppose there is a power generation device with an ID of 001. During a certain period of time in the past year, its power information value, environmental information value, and time information value are (50 kWh, 25°C, January 1, 2023). Similarly, we obtain historical data for all devices.

[0092] S22, constructing a corresponding initial neural network based on the types of equipment in the power generation set, the power storage set and the power consumption set, wherein the initial neural network includes initial neural nodes.

[0093] It should be noted that there are certain differences in the corresponding power generation methods in different regions. For example, some regions use wind power generation, some use coal power generation, and some use municipal electricity. Different power generation methods have corresponding input nodes, and thus the initial neural nodes of the corresponding initial neural network will be constructed according to the corresponding types of power generation equipment. Similarly, some regions may not have power storage functions, or the power storage methods in different regions are also different, which can be battery storage, capacitor storage, etc. Therefore, the corresponding equipment in the power storage set has a corresponding power storage initial neural node. The power equipment corresponding to industrial areas and residential areas is different, and the power consumption at the same time may also be different. For example, during the weekend, the corresponding residential area has normal power consumption, but the industrial area may be shut down during the weekend, so the corresponding power consumption will have certain differences. Therefore, corresponding neural nodes are constructed according to different equipment, so that data can be input in combination with the actual corresponding neural nodes in the future, so as to facilitate accurate prediction of the operation of the power grid.

[0094] It can be understood that the device type refers to different devices having different functions and characteristics, such as solar power generation equipment, wind power generation equipment, etc., and the initial neural network refers to the basic neural network model constructed based on the device type, including the initial neural nodes.

[0095] Among them, the initial neural node is the node corresponding to each device in the initial neural network.

[0096] For example, for solar power generation equipment, the corresponding initial neural network structure can be an input layer (including input power information value, environmental information value and time information value), an intermediate layer (several hidden layers) and an output layer (power generation prediction value).

[0097] In some embodiments, the specific implementation of step S22 (the types of equipment in the power generation set, the power storage set, and the power consumption set construct corresponding initial neural networks, and the initial neural network includes initial neural nodes) includes:

[0098] S221, initializing the initial neural network corresponding to each power generation set, power storage set and power consumption set.

[0099] It can be understood that an initial neural network model is created for each device set (power generation, power storage, and power consumption), wherein the initial neural network includes neural nodes of an input layer, a hidden layer, and an output layer.

[0100] S222, the initial neural network includes an initial input layer, an initial hidden layer and an initial output layer, traverses all initial input neurons in the initial input layer, and determines the attributes of its input information to obtain corresponding neuron labels.

[0101] It can be understood that the initial neural network includes an initial input layer, an initial hidden layer and an initial output layer, wherein the initial input layer is the input layer in the initial neural network, the initial hidden layer is the hidden layer in the initial neural network, and the initial output layer is the output layer in the initial neural network.

[0102] It is not difficult to understand that the initial input layer may have multiple initial input neurons, so that all initial input neurons in the initial input layer may be traversed to determine the neuron label corresponding to the input information.

[0103] Among them, the initial input neuron is the neuron corresponding to the initial input layer, the input information is the initial data information input when performing power grid prediction, the neuron label is the label corresponding to each neuron in the initial neural network, and each neuron has a corresponding label. For example, when the initial input neuron is a neuron corresponding to wind energy, the corresponding neuron label can be a wind energy initial input neuron. When the initial input neuron is a neuron corresponding to light energy, the corresponding neuron label can be a light energy initial input neuron. Therefore, when the input information corresponds to the wind energy attribute, that is, wind force level 6, it can be determined that the neuron label corresponding to the corresponding input information is a wind energy initial neuron.

[0104] S223, obtaining the initial hidden neurons and initial output neurons directly or indirectly connected to each initial input neuron to obtain the corresponding neural sub-paths, and making the neuron labels correspond to the neural sub-paths.

[0105] It can be understood that the initial input neuron can be directly or indirectly connected to the initial hidden neuron and the initial output neuron, so that the corresponding connected neural sub-paths can be obtained. Since the initial input neuron is directly or indirectly connected to multiple neurons, the corresponding neural sub-path can be a forked path, and then the label corresponding to the initial input neuron is associated with the neural sub-path for subsequent training and optimization.

[0106] Among them, the neural subpath is the connection path from the input neuron to the hidden layer neuron and then to the output layer neuron, for example, Figure 2 As shown, it is a schematic diagram of a neural sub-path provided by the present invention, in which the initial input neuron, the initial hidden neuron and the initial output neuron are directly or indirectly connected to obtain the corresponding neural sub-path.

[0107] It is not difficult to understand that when the label corresponding to the initial neuron is the wind energy initial neuron, the label of the neural sub-path corresponding to the initial neuron will also be the wind energy initial neuron. The label of each input neuron is matched with its related neural sub-path to facilitate subsequent management according to the label, so as to facilitate the training and optimization of the neural network.

[0108] S23, based on the objective reference attributes of the power grid operation environment, the initial neural nodes are reviewed and adjusted to obtain processed power generation neural networks, storage neural networks, and utilization neural networks.

[0109] It can be understood that objective reference attributes refer to some objective factors that affect the operation of the power grid, such as the power generation method corresponding to the region, ambient temperature, etc. The review and adjustment processing refers to the optimization and adjustment of the initial neural nodes to better adapt to the actual power grid environment.

[0110] For example, when there is no corresponding wind power generation in a certain power grid environment, by adjusting the initial neural network, the paths corresponding to the wind energy neuron labels in the corresponding initial neural network can be updated and deleted to improve the prediction accuracy of the model.

[0111] Through the above-mentioned implementation mode, the present invention can effectively distinguish the different characteristics of the equipment on the power generation side, the power storage side and the power consumption side, conduct targeted deep learning training, and obtain optimized neural network models. These models can more accurately predict the power demand and supply of various types of equipment, thereby improving the efficiency and reliability of power resource allocation.

[0112] In some embodiments, the specific implementation of step S23 (reviewing and adjusting the initial neural nodes based on the objective reference attributes of the power grid operation environment to obtain the processed power generation neural network, storage neural network and use neural network) includes:

[0113] S231, obtaining objective reference attributes of the region corresponding to the power grid operation environment to obtain objective information, and comparing the objective information with the neuron label of the initial input neuron, each objective information has a preset corresponding neuron label.

[0114] It is understandable that objective reference attribute information such as electricity consumption methods and environmental factors corresponding to different regions will have a certain impact on the predicted power grid operation results. For example, area A is vast and sparsely populated, with strong winds, and wind power generation equipment is usually installed, mainly using wind power generation. Area B has concentrated residential areas and uses solar power generation. Therefore, the weather will affect the production of electricity. Therefore, in order to obtain accurate prediction results, it is necessary to compare the objective information of the predicted area with the initial input neural label, so as to determine the corresponding neural network based on the objective information and make accurate predictions.

[0115] Among them, the objective information is the information corresponding to the objective reference attributes, such as the electricity consumption mode is wind energy, solar energy or municipal electricity, etc. Each objective information has a pre-set neuron label. For example, when the objective information is wind speed data, the corresponding wind energy neuron label is pre-set.

[0116] S232, if it is determined that the objective information corresponds one-to-one with the neuron labels, then based on the initial neural networks corresponding to the power generation set, the power storage set and the power consumption set, a power generation sub-neural network, a power storage sub-neural network and a power consumption sub-neural network are obtained.

[0117] It can be understood that when it is determined that the neuron labels corresponding to the objective information of the actual region correspond one-to-one with the corresponding labels in the initial neural network, it can be said that the initial neural network is consistent with the actual region, and the initial neural network model is considered to be suitable for the current power grid operating environment. The power generation neural network, storage neural network and use neural network in the initial neural network are directly used as the final power generation neural network, storage neural network and use neural network.

[0118] S233, if it is determined that the objective information does not correspond to the neuron label, the objective information is compared with the neuron label to obtain a different neuron label, and the initial neural networks corresponding to the power generation set, the power storage set and the power consumption set are processed based on the different neuron labels to obtain a power generation sub-neural network, a power storage sub-neural network and a power consumption sub-neural network.

[0119] It is understandable that when it is determined after comparison that the neuron labels corresponding to the objective information cannot correspond one-to-one with the initial neuron labels, it means that there are neural sub-paths in the corresponding initial neural network that do not correspond to the actual region. In order to improve the accuracy of the power grid prediction results, the difference neuron labels that do not correspond to the neuron labels of the objective information can be determined, so as to facilitate the adjustment of the initial neural network. For example, the sub-paths corresponding to the difference neuron labels can be deleted to obtain an updated initial neural network, and the power generation neural network, storage neural network and power utilization neural network suitable for the current region can be determined.

[0120] Among them, the difference neuron label is the neuron label whose corresponding neuron label in the initial neural network cannot match the objective information. For example, when a wind power tower cannot be built in the corresponding area, the corresponding wind energy neuron label in the initial neural network does not match the objective information. Therefore, the corresponding wind energy neuron label can be used as the difference neuron label.

[0121] Through the above-mentioned implementation mode, the present invention can ensure that the neural network model more accurately reflects the actual situation of the power grid operating environment, thereby improving the prediction accuracy of the model. Specifically, by introducing objective reference attributes of the region, the initial neural network is reviewed and adjusted to ensure that the ultimately generated power generation neural network, storage neural network and power consumption neural network are more adapted to the actual power grid operating environment.

[0122] In some embodiments, the specific implementation of step S233 (if it is determined that the objective information does not correspond to the neuron label, the objective information is compared with the neuron label to obtain a different neuron label, and the initial neural network corresponding to the power generation set, the power storage set and the power consumption set is processed based on the different neuron labels to obtain the power generation sub-neural network, the power storage sub-neural network and the power consumption sub-neural network) includes:

[0123] S2331, determining the neuron labels that exist in the neuron labels and do not correspond to the objective information as the difference neuron labels.

[0124] It is understandable that the different neuron labels are determined so that the neural sub-paths corresponding to the different neuron labels can be quickly determined in the initial neural network later for adjustment.

[0125] S2332, based on the neuron labels of the differences, neuron labels and neural sub-paths in the initial neural network are selected, and upper-dimensional neural nodes connected to upper-dimensional connections of each neural sub-node in the neural sub-path are extracted.

[0126] It can be understood that neuron labels and corresponding neural sub-paths that are consistent with the different neuron labels are selected in the initial neural network, so as to extract the upper-dimensional neural nodes connected to each neural sub-node on the neural sub-path, so as to subsequently determine the path connection method between the upper-dimensional neural nodes and the initial input neurons, and adopt corresponding methods to perform corrections and updates.

[0127] Among them, the neural sub-nodes are the nodes corresponding to each neuron on the neural sub-path, the upper dimension refers to the upper level corresponding to different neuron nodes in the neural sub-path, and the upper dimension neural nodes are the corresponding nodes in the upper dimension, which can be the neuron nodes in the output layer.

[0128] S2333, extract the properties of the upper-dimensional neural nodes, adjust the corresponding neural sub-paths based on the node correction method pre-configured by the different neuron labels and the properties of the upper-dimensional neural nodes, and obtain the transmitting electronic neural network, the storing electronic neural network and the using electronic neural network.

[0129] It can be understood that each node in the initial neural network has corresponding attributes. Therefore, by obtaining the attributes of the upper-dimensional neural nodes and determining the pre-configured node correction method corresponding to the different neuron labels, the corresponding neural sub-paths are adjusted. After the adjustment, a power-generating neural network, a storage neural network and a power-using neural network that are adapted to the corresponding actual environment can be obtained, so as to improve the accuracy of the neural network model and thereby improve the accuracy of the prediction results of the power grid operating environment.

[0130] Among them, the node correction method is a method of correcting and updating the nodes in the neural network, which is preset manually.

[0131] It is not difficult to understand that the attributes of each node in the upper dimension are obtained, so that the path corresponding to the upper dimensional neural node corresponding to the different neuron label is adjusted. For example, when the different neuron label is a wind energy neuron label, the neural sub-path corresponding to the wind energy attribute in the upper dimensional neural node is adjusted according to the corresponding node correction method.

[0132] In some embodiments, the specific implementation of step S2333 (the extraction of the properties of the upper-dimensional neural nodes, the adjustment of the corresponding neural sub-paths based on the node correction method pre-configured by the different neuron labels and the properties of the upper-dimensional neural nodes, to obtain the transmitting electronic neural network, the storing electronic neural network and the using electronic neural network) includes:

[0133] S23331, obtain the attributes of the upper-dimensional neural node corresponding to each neural sub-node. If the upper-dimensional neural node is connected to the initial input neuron through only one path, directly delete the neuron path between the upper-dimensional neural node and the initial input neuron.

[0134] It can be understood that since different neuron nodes may have multiple attribute labels in the initial neural network, by obtaining the attributes of the upper-dimensional neural node corresponding to each neural sub-node in turn, the number of connection paths between the upper-dimensional neural node and the initial input neuron can be determined, and then corresponding correction methods can be taken to adjust the paths.

[0135] Among them, when the upper-dimensional neural node is connected to the initial input neuron through only one path, it can be explained that the attributes of the neural nodes on the corresponding path are the same and correspond to different neuron labels, so that the neuron path between the upper-dimensional neural node and the initial input neuron is directly deleted. For example, if the different neuron label is a wind energy neuron label, then the path is different from the prediction of the current regional environment, so that all the neuron paths with wind energy attributes on the sub-path can be deleted to improve the accuracy of the neural network model.

[0136] S23332, if the upper dimensional neural node is connected to multiple initial input neurons through multiple paths, then obtain the difference neuron labels and non-difference neuron labels of the multiple initial input neurons.

[0137] It can be understood that when the upper-dimensional neural node is connected to multiple initial input neurons through multiple paths, it can be said that the corresponding upper-dimensional neural node corresponds to multiple attributes and paths, thereby obtaining the neuron labels of the initial input neurons corresponding to multiple different paths, including different neuron labels and non-different neuron labels.

[0138] It is not difficult to understand that the non-differential neuron labels are neuron labels corresponding to the actual objective attributes.

[0139] For example, when area B has solar power generation but no wind power generation, the labels of the two initial input neurons connected by two paths of the same upper-dimensional neural node with a wind energy neuron label can be obtained, where the neuron label corresponding to the difference of the initial input neuron is the wind energy neuron label, and the other neuron label with no difference is the solar energy neuron label.

[0140] Through the above implementation, the categories of the initial neuron labels corresponding to the different path connections corresponding to the upper-dimensional neural nodes can be determined so as to subsequently adjust the corresponding paths.

[0141] S23333, based on the information in the difference neuron labels and the non-difference neuron labels, compare with the pre-configured node correction method to obtain the weight and / or activation function corresponding to the upper dimensional neural node.

[0142] It can be understood that the information in the difference neuron labels and the non-difference neuron labels is compared with the pre-configured node correction method to determine the weights and / or activation functions corresponding to the upper-dimensional neural nodes, thereby achieving adjustment of the neural sub-paths.

[0143] A1, the weight and / or activation function corresponding to each combination of the different neuron labels and the non-different neuron labels in the node correction method.

[0144] It can be understood that when the corresponding neuron has a corresponding combination of a difference neuron label and a non-difference neuron label, it can be adjusted and corrected according to the correction method corresponding to the node.

[0145] Among them, the weight is a pre-set proportional value. For example, when the corresponding node contains a difference neuron label of wind energy and a light energy label corresponding to the non-difference neuron label, the weight corresponding to wind energy can be 0.5, and the weight corresponding to light energy is 0.5. When there is only light energy but no wind energy, the weight of the corresponding node can be 0.8, which is a weight value set artificially in advance. The excitation function is a data processing function, that is, a processing function during deep learning training.

[0146] S23334, after adjusting all neural sub-paths, the electron-emitting neural network, the electron-storing neural network and the electron-using neural network are obtained.

[0147] It can be understood that after adjusting all neural sub-paths, a power generation neural network, a storage neural network and a power utilization neural network adapted to the current geographical environment are obtained to improve the prediction accuracy of the power grid operating environment.

[0148] S3, according to the data association relationship between the power generation side equipment, the power storage side equipment and the power consumption side equipment, assemble the power generation electronic neural network, the power storage electronic neural network and the power consumption electronic neural network based on the data association relationship to obtain the overall neural inference network.

[0149] It can be understood that devices with different attributes are directly or indirectly related to each other. For example, only the produced electricity can be stored or used, so that the corresponding electricity consumption and storage amount are less than or equal to the power generation. Then, the corresponding sub-neural networks can be assembled according to the data association relationship between the power generation side equipment, the power storage side equipment and the power consumption side equipment to obtain a detailed total power neural network, so that the corresponding parameters can be input into the total neural inference network for accurate prediction.

[0150] Among them, the data association relationship is the association relationship corresponding to the data of different devices, and the total neural deduction network is a complete power deduction neural network, that is, the electronic neural network, the storage electronic neural network and the neural network assembled with the electronic neural network according to the data association relationship, so as to deduce and predict the subsequent operation of the power grid.

[0151] In some embodiments, the specific implementation of step S3 (according to the data association relationship between the power generation side equipment, the power storage side equipment and the power consumption side equipment, assembling the power generation electronic neural network, the power storage electronic neural network and the power consumption electronic neural network based on the data association relationship to obtain the total neural deduction network) includes:

[0152] S31, obtaining the working attributes of the equipment in the power generation side equipment, the power storage side equipment and the power consumption side equipment, and counting all the working attributes to obtain an attribute set.

[0153] It can be understood that the working attributes of each device are extracted from the power generation side, power storage side and power consumption side devices. These attributes can be the power output, efficiency, storage capacity, power demand, etc. of the device, and the working attributes of all devices are counted to form a complete attribute set, so that the corresponding sub-neural networks can be connected and combined by connecting neuron nodes in the attribute set.

[0154] The working attributes are the functional attributes of the equipment, and the attribute set is the set of all working attributes.

[0155] S32, combining the work attributes in sequence to obtain work attribute groups, determining the association relationship corresponding to each work attribute group, and each work attribute group has a preset association relationship.

[0156] It is understandable that the acquired working attributes are combined in pairs to form working attribute groups. For example, power generation and storage capacity, storage capacity and power consumption, etc. For each working attribute group, its preset association relationship is determined. These association relationships can be preset based on physical laws, empirical data or business logic. For example, the relationship between power generation and storage capacity may be determined by a power balance equation.

[0157] It is not difficult to understand that the relationship between power generation and storage capacity is: power generation - power consumption = storage change, and the relationship between storage capacity and power consumption is: storage capacity - power consumption = available power. Therefore, the association relationship can be determined based on the working attribute group, so that the corresponding sub-neural networks can be connected according to the association relationship to obtain the total neural inference network.

[0158] S33, based on the association relationship, the transmitting electronic neural network, the storing electronic neural network and the neural nodes corresponding to the using electronic neural network are selected and connected to obtain the total neural deduction network, and each association relationship has preset neural nodes that need to be associated and connected.

[0159] It can be understood that according to the association relationship in each working attribute group, the corresponding neural nodes are selected from the generating electronic neural network, the storing electronic neural network and the using electronic neural network, and the selected neural nodes are connected according to the preset association relationship to form a total neural deduction network. This process can be understood as building a more complex and complete network that can simulate and predict the operation of the entire power system.

[0160] For example, when the association relationship is the relationship between power generation and power storage or the relationship between power storage and power consumption, the corresponding neural nodes can be selected from the power generation sub-neural network, the power storage sub-neural network and the power consumption sub-neural network, wherein the power generation node is connected with the storage node, and the storage node is connected with the power consumption node, to realize the interconnection of the power generation sub-neural network and the power storage sub-neural network, the power storage sub-neural network and the power consumption sub-neural network to form a complete power flow and storage prediction model.

[0161] Through the above-mentioned implementation mode, the present invention can effectively and organically combine the power generation neural network, the power storage neural network and the power consumption neural network to establish a total neural inference network, which can not only independently predict the power generation, storage and power consumption, but also make a comprehensive prediction of the overall system through the correlation relationship, thereby improving the prediction accuracy and resource allocation efficiency of the power system.

[0162] S4, obtaining the variable deduction parameters corresponding to the deduction target, and inputting the variable deduction parameters into the overall neural deduction network to obtain the predicted deduction results.

[0163] It is understandable that since the environment corresponding to different dates is different, and the corresponding data will also have a certain impact on the operation of the power grid, for example, when the temperature is higher or lower, people may need to use electrical equipment to cool down or heat up in order to carry out normal activities, then the corresponding power generation, power consumption and other data will fluctuate accordingly. Therefore, the variable deduction parameters of the corresponding date can be input into the total neural deduction network to obtain the predicted deduction results, and when the more parameters and the higher the accuracy, the accuracy of the corresponding deduced results will also increase accordingly.

[0164] Among them, the deduction target is the date on which the power grid operation needs to be deduced and predicted, the variable deduction parameters are the parameters of various factors that affect the power grid operation corresponding to the deduction target, such as weather data, temperature, humidity, and whether it is a holiday, etc. The prediction deduction result is the power grid data result predicted by the total neural deduction network.

[0165] In some embodiments, the specific implementation of step S4 (according to the data association relationship between the power generation side equipment, the power storage side equipment and the power consumption side equipment, assembling the power generation electronic neural network, the power storage electronic neural network and the power consumption electronic neural network based on the data association relationship to obtain the total neural deduction network) includes:

[0166] S41, obtain the variable deduction parameters corresponding to the deduction target and decompose them to obtain multiple decomposition sub-information, select neural nodes according to the information corresponding to each decomposition sub-information and input them into the overall neural deduction network to obtain the predicted deduction result.

[0167] It can be understood that the relevant variable deduction parameters corresponding to the predicted date for which the power grid operation situation needs to be deduced are decomposed to obtain multiple decomposition sub-information. For example, when the power grid operation environment on the 15th needs to be predicted, the temperature, humidity, wind speed and other data corresponding to the weather on the 15th can be decomposed to obtain multiple decomposition sub-information, such as temperature 38°, humidity 50%, wind force 2, so as to select the neural node input temperature value 38° corresponding to the temperature, the neural node input 50% corresponding to the humidity, and the neural node input 2 corresponding to the wind force, so as to obtain the prediction and deduction results through calculation and transmission of each node in the total neural network, for example, the wind power generation, storage capacity and power consumption corresponding to the 15th are predicted.

[0168] The decomposed sub-information is different types of data information after decomposing the variable deduction parameters.

[0169] S5, compare the predicted deduction results with the actual results to obtain multi-dimensional training data, and based on the multi-dimensional training data, train and update the weights of neurons and the activation function in the total neural deduction network.

[0170] It is understandable that the predicted deduction results may deviate from the actual results to a certain extent. For example, if the electricity consumption in the predicted deduction results is less than the actual electricity consumption, the corresponding electricity consumption comparison difference can be obtained, so that the difference between the predicted result and the actual result can be used as training data. Therefore, a variety of data differences are obtained as multi-dimensional training data, so that the neural network can update the weights and activation functions of the neurons in the total neural deduction network according to the corresponding multi-dimensional training data. For example, the weight values ​​related to electricity consumption can be increased.

[0171] Among them, multi-dimensional training data refers to multiple types of data that compare the predicted results with the actual results, such as wind power generation, solar power generation, wind energy storage and electricity consumption.

[0172] It is not difficult to understand that by constantly comparing the prediction results with the actual data and iteratively training, the network can be continuously optimized, the accuracy and reliability of the prediction can be improved, and efficient management and optimization of the power system can be achieved.

[0173] In some embodiments, the specific implementation of step S5 (comparing the predicted deduction results with the actual results to obtain multi-dimensional training data, and training and updating the weights of neurons and the activation function in the total neural deduction network based on the multi-dimensional training data) includes:

[0174] S51, split the prediction and deduction results according to different dimensions to obtain multiple prediction sub-information, compare the prediction sub-information of each dimension with the actual sub-information of the corresponding dimension in the actual result, and obtain a comparison sub-information group of multiple dimensions.

[0175] It is understandable that the prediction results output by the total neural inference network are split according to predetermined dimensions. For example, they are split according to dimensions such as power generation, storage, and power consumption to obtain corresponding prediction sub-information, and the prediction sub-information of each dimension is compared with the actual sub-information in the actual occurrence result of the corresponding dimension to obtain comparison sub-information groups of multiple dimensions.

[0176] Among them, the predicted sub-information is the data information of different dimensions in the prediction and deduction results, for example, the wind power consumption is 500 degrees, the wind power storage capacity is 200 degrees, etc. The actual sub-information is the data information corresponding to each dimension in the actual result. For example, the actual wind power consumption on the 15th is 600 degrees. The predicted sub-information and actual sub-information of the same dimension can be compared to obtain a comparison sub-information group. For example, the predicted sub-information and actual sub-information of the power consumption dimension are compared and matched to obtain a comparison sub-information group of the power consumption dimension.

[0177] S52, calculating the comparison difference between the predicted sub-information and the actual sub-information in each comparison sub-information group, and obtaining the neuron output by the corresponding predicted sub-information in the total neural inference network as the target neuron.

[0178] It can be understood that the comparison difference is the data difference between the predicted sub-information and the actual sub-information in the same comparison sub-information group, and the target neuron is the neuron corresponding to the output of the corresponding predicted sub-information in the total neural inference network.

[0179] S53, obtaining a chain formed by neurons participating in calculation in an upper dimension directly or indirectly connected to the target neuron as a chain to be trained.

[0180] It can be understood that the chain to be trained is a path chain formed by the upper-level neurons participating in the calculation that are directly or indirectly connected to the target neuron. For example, when the target neuron is an electricity consumption output neuron, the corresponding chain to be trained is a neuron chain formed by connecting the neurons participating in the calculation of electricity consumption.

[0181] S54, based on the comparison difference, the neurons in the training chain are trained and updated to obtain the total neural inference network after training.

[0182] It can be understood that the comparison difference is used to adjust the weights and activation functions of the neurons in the chain to be trained through the backpropagation algorithm until the error is reduced to an acceptable range.

[0183] In some embodiments, the specific implementation of step S54 (the updating of neuron training in the training chain based on the comparison difference to obtain the trained total neural deduction network) includes:

[0184] S541, extracting neurons other than the target neuron in the chain to be trained as neurons to be trained.

[0185] It is understandable that in order to perform information adjustment training on neurons in a neural network, neurons other than target neurons in the chain to be trained can be used as neurons to be trained, so that information adjustment can be performed on the determined neurons to be trained later.

[0186] S542, obtaining the number of neurons in the to-be-trained neurons that are in the same sub-neural network as the target neuron to obtain a first number, and obtaining the number of neurons in the to-be-trained neurons that are not in the same sub-neural network as the target neuron to obtain a second number.

[0187] It can be understood that since the degree of influence on the output result is different whether the neurons to be trained and the target neurons are in the same sub-neural network or not, the adjustment proportion weight of the corresponding neural network is also different. Therefore, different degrees of adjustment can be made according to whether the neurons to be trained and the target neurons are in the same sub-neural network.

[0188] The first number is the number of neurons in the same sub-neural network as the target neuron, and the second number is the number of neurons in a different sub-neural network than the target neuron.

[0189] S543, obtaining the number of intervals between the neuron to be trained and the target neuron, and training and updating the weight and / or activation function of the neuron to be trained based on the first number, the second number, the number of intervals, and the comparison difference.

[0190] It can be understood that the number of intervals reflects the distance between the neuron to be trained and the target neuron. By considering this distance, the weights of the neurons can be adjusted more accurately. Neurons that are closer may have a greater impact on the target neuron, so a larger weight adjustment is required, while neurons that are farther away may have a smaller impact on the target neuron, so the weight adjustment can be reduced accordingly.

[0191] The number of intervals is the number of neurons between the neuron to be trained and the target neuron.

[0192] Therefore, the weights and / or activation functions of the neurons to be trained can be trained and updated according to the first quantity, the second quantity, the interval quantity, and the comparison difference, so as to improve the accuracy of the adjustment data, thereby improving the training effect and prediction performance of the entire neural network.

[0193] In some embodiments, the specific implementation of step S543 (obtaining the number of intervals between the neuron to be trained and the target neuron, and training and updating the weight and / or activation function of the neuron to be trained based on the first number, the second number, the number of intervals, and the comparison difference) includes:

[0194] S5431, based on the first number and the second number, obtain the first benchmark adjustment weight of the neurons in the same sub-neural network as the target neuron, and the second benchmark adjustment weight of the neurons not in the same sub-neural network as the target neuron.

[0195] It can be understood that the first benchmark adjustment weight refers to the benchmark weight value for adjusting neurons in the same sub-neural network as the target neuron, and the second benchmark adjustment weight refers to the benchmark weight value for adjusting neurons that are not in the same sub-neural network as the target neuron, which is calculated by the first quantity and the second quantity.

[0196] S5432, offset-adjust the first benchmark adjustment weight and the second benchmark adjustment weight based on the number of intervals, and update the weight and / or activation function training of the neuron to be trained based on the offset-adjusted first benchmark adjustment weight, the second benchmark adjustment weight, and the comparison difference.

[0197] It is understandable that since different numbers of intervals will also have a certain impact on the degree of adjustment offset, the first benchmark adjustment weight and the second benchmark adjustment weight can be offset adjusted according to the determined number of intervals, thereby obtaining the adjusted first benchmark adjustment weight, the second benchmark adjustment weight, and the comparison difference to update the weights of the neurons to be trained and / or the excitation function training.

[0198] S5433, calculate the first benchmark adjustment weight and the second benchmark adjustment weight by the following formula

[0199]

[0200] in, is the first benchmark adjustment weight before offset adjustment, is the second benchmark adjustment weight before offset adjustment, is the first calculation weight, is the first quantity, Calculate the weight for the second, is the second quantity, is the first benchmark adjustment weight after offset adjustment, A is the number of intervals, is the first interval coefficient, is the second benchmark adjustment weight after offset adjustment, B is the number of intervals of neurons in different sub-neural networks, is the second interval coefficient.

[0201] It is understandable that the first number , the second quantity Have corresponding first calculation weights respectively and the second calculates the weights , where the first calculation weight and the second calculates the weights is pre-configured so that the first quantity and first calculate the weights The product of the first quantity and first calculate the weights The product of and the second calculates the weights The sum of the products of , thereby calculating the first benchmark adjustment weight before the offset adjustment , when the first number is larger, the first number and first calculate the weights The product will also be larger, and thus, the first benchmark adjustment weight before the offset adjustment It will also increase with the increase of the numerator, which means that when the number of neurons to be trained in the same sub-neural network as the target neuron increases, the corresponding adjusted baseline adjustment weight will also increase.

[0202] Similarly, the second quantity and the second calculates the weights The product of the first quantity and first calculate the weights The product of and the second calculates the weights The sum of the products of , thereby calculating the second benchmark adjustment weight before the offset adjustment , when the second number is larger, the second number and the second calculates the weights The product will also be larger, and thus, the second benchmark adjustment weight before the offset adjustment It will also increase with the increase of the numerator, which means that when the number of neurons to be trained that are not in the same sub-neural network as the target neuron increases, the corresponding adjusted baseline adjustment weight will also increase.

[0203] It is not difficult to understand that when the number of intervals A is larger, the first benchmark adjustment weight after the corresponding offset adjustment is will become smaller, it can be said that the weight of the corresponding neuron is lower, that is, the more intervals there are, the farther the corresponding neurons may be, the less influence they have on the target neuron, so the weight adjustment can be reduced accordingly. Similarly, when the number of intervals is smaller, the first benchmark adjustment weight after the corresponding offset adjustment is will become larger, which means that the weight of the corresponding neuron is higher, that is, the smaller the number of intervals, the closer the corresponding neurons may be, the greater the impact on the target neuron, so the weight adjustment can be increased accordingly, where the first interval coefficient The coefficient value for weight adjustment according to the number of intervals is preset artificially.

[0204] Similarly, when the number of intervals B between neurons in different sub-neural networks is larger, the corresponding second benchmark adjustment weight after offset adjustment is will become smaller, it can be said that the weight of the corresponding neuron is lower, that is, the more intervals there are, the farther the corresponding neurons may be, the less influence they have on the target neuron, so the weight adjustment can be reduced accordingly. Similarly, when the number of intervals is smaller, the second benchmark adjustment weight after the corresponding offset adjustment is will become larger, which means that the weight of the corresponding neuron is higher, that is, the smaller the number of intervals, the closer the corresponding neurons may be, the greater the impact on the target neuron, so the weight adjustment can be increased accordingly, where the second interval coefficient The coefficient value for weight adjustment according to the number of intervals is preset artificially.

[0205] It is worth mentioning that the first benchmark adjustment weight after offset adjustment and the second benchmark adjustment weight after offset adjustment will be calculated based on the first quantity, the second quantity and the interval quantity, so as to subsequently adjust the weights and / or activation functions of the corresponding sub-neural networks.

[0206] S5434, calculating the updated weight and / or activation function by the following formula:

[0207]

[0208] in, is the weight before updating, is the updated weight, is the K value of the activation function before updating, is the K value of the updated activation function.

[0209] It is understandable that when the offset-adjusted first reference adjustment weight The larger the value, the larger the corresponding weight change will be. On the contrary, when the first benchmark after the offset adjustment is adjusted, the weight The smaller the value, the smaller the corresponding weight change will be. Similarly, when the second benchmark after the offset adjustment adjusts the weight The larger the value, the larger the change range of the corresponding activation function will be. On the contrary, when the second benchmark adjustment weight after the offset adjustment is The smaller the value, the smaller the change amplitude of the corresponding activation function will be, so that the performance of the neural network can be continuously optimized through continuous iterative training to improve the accuracy and reliability of the prediction.

[0210] The present invention also provides a storage medium, in which a computer program is stored. When the computer program is executed by a processor, it is used to implement the methods provided by the various embodiments described above.

[0211] Among them, the storage medium can be a computer storage medium or a communication medium. The communication medium includes any medium that facilitates the transmission of a computer program from one place to another. The computer storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer. For example, the storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an application-specific integrated circuit (Application Specific Integrated Circuits, referred to as: ASIC). In addition, the ASIC can be located in a user device. Of course, the processor and the storage medium can also exist in a communication device as discrete components. The storage medium can be a read-only memory (ROM), a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, and an optical data storage device, etc.

[0212] The present invention also provides a program product, which includes an execution instruction, which is stored in a storage medium. At least one processor of a device can read the execution instruction from the storage medium, and at least one processor executes the execution instruction so that the device implements the methods provided in the above various embodiments.

[0213] In the above-mentioned terminal or server embodiments, it should be understood that the processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the present invention may be directly implemented as being executed by a hardware processor, or may be executed by a combination of hardware and software modules in the processor.

[0214] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A power grid operation environment deduction method for deep reinforcement learning, characterized in that: S1, classify the power grid operating environment based on the attributes of the equipment in the power grid to obtain power generation side equipment, power storage side equipment and power consumption side equipment; S2, performing deep learning training based on historical power data of power generation side equipment, power storage side equipment and power consumption side equipment to obtain power generation sub-neural network, power storage sub-neural network and power consumption sub-neural network respectively; S3, obtaining data associations among power generation equipment, power storage equipment and power consumption equipment, and assembling the power generation neural network, power storage neural network and power consumption neural network based on the data associations to obtain a total neural deduction network; S4, obtaining variable deduction parameters corresponding to the deduction target, and inputting the variable deduction parameters into the overall neural deduction network to obtain a prediction deduction result; S5, comparing the predicted deduction results with the actual results to obtain multi-dimensional training data, and training and updating the weights and activation functions of the neurons in the total neural deduction network based on the multi-dimensional training data; The method of classifying the power grid operating environment based on the attributes of the devices in the power grid to obtain power generation side devices, power storage side devices and power consumption side devices includes: Acquire the attributes of the equipment in the power grid, and if it is determined that the equipment has any one of the power generation attribute, the power storage attribute or the power consumption attribute, add the corresponding first label and classify it into the power generation set, the power storage set or the power consumption set; If it is determined that the device has any of the power generation attribute, power storage attribute or power consumption attribute, then a corresponding second tag is added and the device is classified into the power generation set, power storage set or power consumption set; The power generation set includes power generation side equipment with power generation function, the power storage set includes power storage side equipment with power storage function, and the power consumption set includes power consumption side equipment with power consumption function; The deep learning training is performed based on the historical power data of the power generation side equipment, the power storage side equipment and the power consumption side equipment to obtain the power generation sub-neural network, the power storage side neural network and the power consumption side neural network respectively, including: Traversing the IDs of the devices in the power generation set, the power storage set, and the power consumption set in sequence, and acquiring historical power consumption information within a preset time period based on the IDs to obtain historical power data, wherein the historical power consumption information at least includes power information value, environmental information value, and time information value; Constructing a corresponding initial neural network based on the types of equipment in the power generation set, the power storage set, and the power consumption set, wherein the initial neural network includes initial neural nodes; Based on the objective reference attributes of the power grid operation environment, the initial neural nodes are reviewed and adjusted to obtain the processed power generation neural network, storage neural network and power consumption neural network, wherein the objective reference attributes are the attributes of objective factors that affect the operation of the power grid; The types of equipment in the power generation set, power storage set and power consumption set construct corresponding initial neural networks, and the initial neural network includes initial neural nodes, including: Initialize the initial neural network corresponding to each power generation set, power storage set and power consumption set; The initial neural network includes an initial input layer, an initial hidden layer and an initial output layer, traverses all initial input neurons in the initial input layer, and determines the attributes of its input information to obtain corresponding neuron labels; Obtaining the initial hidden neurons and initial output neurons directly or indirectly connected to each initial input neuron to obtain the corresponding neural sub-paths, and making the neuron labels correspond to the neural sub-paths; The initial neural nodes are reviewed and adjusted based on the objective reference attributes of the power grid operation environment to obtain the processed power generation neural network, storage neural network and use neural network, including: Obtain objective reference attributes of the region corresponding to the power grid operation environment to obtain objective information, and compare the objective information with the neuron label of the initial input neuron, each objective information having a preset corresponding neuron label; If it is determined that the objective information corresponds to the neuron labels one by one, then the power generation neural network, the power storage neural network and the power consumption neural network are obtained based on the initial neural networks corresponding to the power generation set, the power storage set and the power consumption set; If it is determined that the objective information does not correspond to the neuron label, the objective information is compared with the neuron label to obtain the different neuron labels. Based on the different neuron labels, the initial neural networks corresponding to the power generation set, the power storage set and the power consumption set are processed to obtain the power generation sub-neural network, the power storage sub-neural network and the power consumption sub-neural network.

2. The power grid operation environment deduction method for deep reinforcement learning according to claim 1 is characterized in that: If it is determined that the objective information does not correspond to the neuron label, the objective information is compared with the neuron label to obtain a different neuron label, and the initial neural network corresponding to the power generation set, the power storage set and the power consumption set is processed based on the different neuron labels to obtain a power generation sub-neural network, a power storage sub-neural network and a power consumption sub-neural network, including: Determine the neuron labels that exist in the neuron labels and do not correspond to the objective information as the difference neuron labels; Selecting neuron labels and neural sub-paths in the initial neural network based on the different neuron labels, and extracting upper-dimensional neural nodes connected to upper-dimensional connections of each neural sub-node in the neural sub-path; The properties of the upper-dimensional neural nodes are extracted, and the corresponding neural sub-paths are adjusted based on the node correction method pre-configured by the different neuron labels and the properties of the upper-dimensional neural nodes to obtain the transmitting electronic neural network, the storing electronic neural network and the using electronic neural network.

3. The power grid operation environment deduction method for deep reinforcement learning according to claim 2 is characterized in that: The extracting the properties of the upper dimensional neural nodes, adjusting the corresponding neural sub-paths based on the node correction method pre-configured by the different neuron labels and the properties of the upper dimensional neural nodes, and obtaining the transmitting electronic neural network, the storing electronic neural network and the using electronic neural network, including: Get the properties of the upper-dimensional neural node corresponding to each neural sub-node. If the upper-dimensional neural node is connected to the initial input neuron through only one path, directly delete the neuron path between the upper-dimensional neural node and the initial input neuron. If the upper-dimensional neural node is connected to multiple initial input neurons through multiple paths, then the difference neuron labels and non-difference neuron labels in the multiple initial input neurons are obtained; Comparing the information in the difference neuron labels and the non-difference neuron labels with the pre-configured node correction method, the weight and / or activation function corresponding to the upper dimensional neural node is obtained; After adjusting all neural sub-paths, we get the electron-emitting neural network, the electron-storing neural network and the electron-using neural network.

4. The power grid operation environment deduction method for deep reinforcement learning according to claim 3 is characterized in that: The node correction method has weights and / or activation functions corresponding to each combination of different neuron labels and non-different neuron labels.

5. The power grid operation environment deduction method for deep reinforcement learning according to claim 1, characterized in that: According to the data association relationship between the power generation side equipment, the power storage side equipment and the power consumption side equipment, the power generation electronic neural network, the power storage electronic neural network and the power consumption electronic neural network are assembled based on the data association relationship to obtain the total neural deduction network, including: Obtain the working attributes of the equipment in the power generation side equipment, the power storage side equipment and the power consumption side equipment, and count all the working attributes to obtain the attribute set; Combining the work attributes in turn to obtain work attribute groups, determining the association relationship corresponding to each work attribute group, and each work attribute group has a preset association relationship; Based on the association relationship, the neural nodes corresponding to the transmitting electronic neural network, the storing electronic neural network and the using electronic neural network are selected and connected to obtain the total neural deduction network, and each association relationship has preset neural nodes that need to be associated and connected.

6. The power grid operation environment deduction method for deep reinforcement learning according to claim 5 is characterized in that: The step of obtaining the variable deduction parameters corresponding to the deduction target and inputting the variable deduction parameters into the overall neural deduction network to obtain the prediction deduction result includes: The variable deduction parameters corresponding to the deduction target are obtained and decomposed to obtain multiple decomposition sub-information. The neural nodes are selected according to the information corresponding to each decomposition sub-information and input into the overall neural deduction network to obtain the predicted deduction result.

7. The power grid operation environment deduction method for deep reinforcement learning according to claim 5, characterized in that: The method of comparing the predicted deduction results with the actual results to obtain multi-dimensional training data, and training and updating the weights and activation functions of neurons in the total neural deduction network based on the multi-dimensional training data, includes: The prediction results are split according to different dimensions to obtain multiple prediction sub-information, and the prediction sub-information of each dimension is compared with the actual sub-information of the corresponding dimension in the actual result to obtain a comparison sub-information group of multiple dimensions; Calculate the comparison difference between the predicted sub-information and the actual sub-information in each comparison sub-information group, and obtain the neuron output by the corresponding predicted sub-information in the total neural deduction network as the target neuron; Acquire a chain formed by neurons participating in calculation in an upper dimension directly or indirectly connected to the target neuron as a chain to be trained; Based on the comparison difference, the neurons in the training chain are trained and updated to obtain the trained total neural inference network.

8. The power grid operation environment deduction method for deep reinforcement learning according to claim 7 is characterized in that: The neurons in the training chain are trained and updated based on the comparison difference to obtain the trained total neural deduction network, including: Extract neurons other than the target neurons in the chain to be trained as neurons to be trained; The number of neurons in the to-be-trained neurons that are in the same sub-neural network as the target neuron is obtained to obtain a first number, and the number of neurons in the to-be-trained neurons that are not in the same sub-neural network as the target neuron is obtained to obtain a second number; The number of intervals between the neuron to be trained and the target neuron is obtained, and the weight and / or activation function of the neuron to be trained and updated based on the first number, the second number, the number of intervals, and the comparison difference.

9. The power grid operation environment deduction method for deep reinforcement learning according to claim 8, characterized in that: The step of obtaining the number of intervals between the neuron to be trained and the target neuron, and training and updating the weight and / or activation function of the neuron to be trained based on the first number, the second number, the number of intervals, and the comparison difference includes: Based on the first number and the second number, a first reference adjustment weight of a neuron in the same sub-neural network as the target neuron and a second reference adjustment weight of a neuron not in the same sub-neural network as the target neuron are calculated; Performing an offset adjustment on the first benchmark adjustment weight and the second benchmark adjustment weight based on the number of intervals, and training and updating the weight and / or activation function of the neuron to be trained based on the offset-adjusted first benchmark adjustment weight, the second benchmark adjustment weight, and the comparison difference; The first benchmark adjustment weight and the second benchmark adjustment weight are calculated by the following formula ， in, is the first benchmark adjustment weight before offset adjustment, is the second benchmark adjustment weight before offset adjustment, is the first calculation weight, is the first quantity, Calculate the weight for the second, is the second quantity, is the offset-adjusted first benchmark adjustment weight, is the number of intervals, is the first interval coefficient, Adjust the weights for the offset-adjusted second benchmark, is the number of intervals between neurons in different sub-neural networks, is the second interval coefficient; The updated weights and / or activation functions are calculated by the following formula, ， in is the weight before updating, is the updated weight, is the K value of the activation function before updating, is the K value of the updated activation function.

Images