Micro-grid load prediction optimization algorithm based on deep learning

Abstract

The invention relates to the technical field of power grid load prediction optimization, in particular to a micro-grid load prediction optimization algorithm based on deep learning. Comprising the following steps: collecting real-time operation data of a micro-grid by using a sensor, dividing the micro-grid into different node types based on the collected real-time operation data, and constructing a micro-grid topological structure diagram based on a connection relationship among all nodes; according to an input historical load data set and a micro-grid topological structure diagram, constructing a node load sequence feature and a structure embedding feature for each node, and carrying out operation load prediction on the nodes based on influence weights among the nodes; and according to the prediction result of the node and the prediction output of the adjacent node, carrying out collaborative correction of the prediction value, meanwhile, according to the prediction output of the adjacent node and the error between the actual measurement values, dynamically adjusting the influence weight between the nodes, and carrying out microgrid load prediction based on the adjusted influence weight.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid load prediction optimization, and particularly to a microgrid load prediction optimization algorithm based on deep learning. Background Art

[0002] With the rapid expansion of distributed energy access and the scale of microgrids, the power grid system has put forward higher requirements for the prediction accuracy and dynamic response ability of node-level loads. However, existing load prediction methods mainly rely on traditional statistical models (such as ARIMA), shallow machine learning methods (such as SVR, RF), or sequence neural networks based on RNN / LSTM. Although they can fit the load sequence to a certain extent, there are generally the following key technical bottlenecks: It is unable to effectively integrate the electrical connection relationships between microgrid nodes and ignores the impact of power flow and load conduction paths between nodes on the current node load evolution. It is unable to flexibly adjust the prediction model according to the behavior characteristics of different node types (such as power generation, load, and energy storage nodes) and lacks the ability to model structural input features. In actual engineering deployment, there is a lack of a mechanism to effectively identify and suppress negative impact factors in the graph structure, resulting in the prediction error being easily affected by structural noise. Summary of the Invention

[0003] In view of the above existing problems, the present invention is proposed.

[0004] To solve the above technical problems, the present invention provides the following technical solution: A microgrid load prediction optimization algorithm based on deep learning, including the following steps. Collect real-time operation data of the microgrid using sensors, and based on the collected real-time operation data, divide the microgrid into different types of node types, and construct a microgrid topology structure diagram based on the connection relationships between all nodes. Analyze the influence weights between nodes through a deep learning algorithm, and predict the operating load of nodes using historical loads, specifically: According to the input historical load data set and the microgrid topology structure diagram, construct node load sequence features and structural embedding features for each node, and predict the operating load of the node based on the influence weights between nodes. According to the influence weights between nodes in the microgrid topology structure and the load prediction results for nodes, perform collaborative optimization and prediction of the microgrid global load, specifically: According to the prediction results of nodes, comprehensively correct the prediction values by combining the prediction outputs of adjacent nodes. At the same time, according to the error between the prediction output of adjacent nodes and the actual measurement value, dynamically adjust the influence weights between nodes, and perform microgrid load prediction based on the adjusted influence weights.

[0005] As a preferred solution of the optimization algorithm for microgrid load forecasting based on deep learning according to the present invention, wherein: the specific process of using sensors to collect real-time operation data of the microgrid is as follows: Calculate the active power and reactive power corresponding to the current node. At the same time, determine the power flow direction of the current node, then there is, Active power,

[0006] Reactive power,

[0007] Wherein, 、 respectively represent the voltage data and current data collected at the current node, 、 represent the current phase angle and voltage phase angle corresponding to the current node, represents the active power corresponding to the current node, represents the reactive power corresponding to the current node.

[0008] As a preferred solution of the optimization algorithm for microgrid load forecasting based on deep learning according to the present invention, wherein: the specific process of determining the power flow direction of the current node is as follows: For the current node, select the adjacent nodes of the current node, and calculate the active power flow and reactive power flow from the current node to the adjacent nodes respectively. According to the calculation results, judge the power flow direction of the current node, specifically: Active power from the current node to the adjacent node,

[0009] Reactive power from the current node to the adjacent node,

[0010] Wherein, represents the voltage data of the current node, represents the voltage data of the adjacent node, represents the voltage phase angle of the current node, represents the voltage phase angle of the adjacent node, represents the impedance between the current node and the adjacent node, represents the active power flow from the current node to the adjacent node, represents the reactive power flow from the current node to the adjacent node; According to the calculation results of the active power flow and reactive power flow, judge the power flow direction of the current node, then there is, If the active power flow of the current node satisfies the formula , it indicates that the power of the current node flows from the current node to the adjacent node, and vice versa, it indicates that the power flows from the adjacent node to the current node; If the reactive power flow of the current node satisfies the formula , it indicates that the reactive power of the current node flows from the current node to the adjacent node, and vice versa, it indicates that the reactive power flows from the adjacent node to the current node.

[0011] As a preferred solution of the microgrid load prediction optimization algorithm based on deep learning according to the present invention, wherein: the division of the microgrid into different types of node types is specifically as follows: If the active power corresponding to the current node satisfies the formula , and the power flow direction of the current node is outwards from the current node, it indicates that the current node type is a power generation node; If the active power corresponding to the current node satisfies the formula , and the reactive power satisfies the formula , and at the same time, the power flow direction of the current node is from the adjacent node into the current node, it indicates that the current node type is a load node; For a node with the node type of an energy storage node, it is judged based on the state of charge corresponding to the node, specifically: Detect whether the current node has a capacitance parameter. If the current node cannot collect the capacitance parameter, it indicates that the current node is a non-energy storage node; If the current node can collect the capacitance parameter, it indicates that the current node type is an energy storage node.

[0012] As a preferred solution of the microgrid load prediction optimization algorithm based on deep learning according to the present invention, wherein: the analysis of the influence weights between nodes by the deep learning algorithm is specifically as follows: For any node in the microgrid topology , extract the node features corresponding to the node , and perform in-depth analysis on the operating state of the node through a convolutional neural network model, and set the analysis result as ; For the initial analysis result of the node , select a node in the microgrid topology that is not the node , and splice the two node features, then there is

[0013] wherein, represents the node features extracted by the node , represents the node The extracted node features represent the concatenated features, which are used for in-depth analysis of the influence weights between nodes. Then, Based on the concatenated features, perform state analysis through a convolutional neural network model and set the analysis result as , and compare the analysis results of the two times and , and based on the comparison result perform in-depth analysis of the influence weights between nodes, specifically: If the comparison result satisfies the formula , it means that the node has a negative influence weight relative to the node ; If the comparison result satisfies the formula , it means that the node has a positive influence weight relative to the node ; If the comparison result satisfies the formula , it means that the node has no influence weight relative to the node .

[0014] As a preferred solution of the microgrid load prediction optimization algorithm based on deep learning according to the present invention, wherein: the operation load of the historical load prediction node is specifically as follows: For each node, based on its historical load sequence, extract a sliding sequence with a time window length of and construct a historical state matrix; Define node type labels, construct an input sample set, and at the same time use a graph convolutional network to extract topological features between nodes and perform time series modeling on the extracted topological features, so as to generate an initial prediction result; According to the generated initial prediction result set, perform error analysis on the prediction result and calculate the prediction error, and re-adjust the graph embedding method based on the prediction error, so as to generate a second-generation prediction value; According to the second-generation prediction value set, by introducing a fitting scoring mechanism, judge the comparison result between the fitting score corresponding to the second-generation prediction value and the fitting score threshold, re-adjust the adjacency structure, and based on the reconstructed adjacency structure, perform load prediction of the node.

[0015] As a preferred solution of the microgrid load prediction optimization algorithm based on deep learning according to the present invention, wherein: the generation of the initial prediction result is specifically as follows: Input the constructed joint feature tensor into the graph convolutional layer to extract node graph embeddings, then

[0016] Among them, represents the normalized adjacency matrix, represents the graph convolution weight, represents the activation function, represents the constructed joint feature tensor, represents the extracted node graph embedding feature; Input the graph embedding feature into the LSTM network to obtain the hidden state output, then there is

[0017] Among them, represents the node corresponding graph embedding feature, represents the node corresponding hidden state output; Based on the hidden state output, generate the initial prediction result, then there is

[0018] Among them, , respectively represent the weight coefficient and bias term of the initial prediction, represents the node corresponding hidden state output, represents the node initial prediction result of the load.

[0019] As a preferred solution of the microgrid load prediction optimization algorithm based on deep learning described in the present invention, where: the load prediction of the node is specifically as follows: Compare the fitting score between the second-generation predicted value and the target load curve, then there is

[0020] Among them, represents the actual load measurement value of the node at time, represents the second-generation predicted value of the node at time, represents the time window length, represents the fitting score of the second-generation predicted value, which is used for reconstructing the adjacency structure. Specifically: Set the fitting score threshold , if the calculated fitting score satisfies the formula , it indicates that the adjacency matrix needs to be readjusted to reconstruct the adjacency structure until the fitting score is higher than the set score threshold. Otherwise, it means that no readjustment of the adjacency structure is required; For the adjusted prediction model, according to the prediction method of the second-generation prediction value, the third-generation prediction value is regenerated, and the fitting score corresponding to the third-generation prediction value is calculated. , the fitting score corresponding to the third-generation prediction value is compared with the fitting score corresponding to the second-generation prediction value until the comparison result satisfies the formula up to this point, it indicates that the load prediction of the node is accurate, and the third-generation prediction value is used as the load prediction result of the node.

[0021] As a preferred solution of the microgrid load prediction optimization algorithm based on deep learning described in the present invention, wherein: the collaborative correction of the prediction value is specifically as follows: For the prediction result of node , according to its set of adjacent nodes , by comprehensively considering the prediction outputs of adjacent nodes, the collaborative correction of the prediction value is carried out. Then,

[0022] wherein, represents the prediction value of node after comprehensively considering adjacent nodes, represents the third-generation prediction value of node , represents the third-generation prediction value of node , represents the influence weight between node and node , represents the set of adjacent nodes of node , which is a set of nodes in the topological structure that can form an edge relationship with node , represents the collaborative correction coefficient, specifically: If the prediction result of node is compared with the prediction results of the set of adjacent nodes and satisfies the formula , it indicates that the node prediction is disturbed by adjacent nodes, and the correction coefficient is increased to achieve the correction of adjacent nodes.

[0023] As a preferred solution of the microgrid load prediction optimization algorithm based on deep learning described in the present invention, wherein: the microgrid load prediction based on the adjusted influence weight is specifically as follows: For the prediction value of node The error between the actual measurement values is used to calculate the weight adjustment factor between each pair of nodes. Then,

[0024] where, represents the adjustment coefficient, represents the node the predicted value after integrating adjacent nodes, represents the node the third-generation predicted value of, node the actual measurement result of the load, represents the node and the node the weight adjustment factor between them, which is used to realize the dynamic adjustment of the influence weight between nodes. Then,

[0025] where, represents the influence weight between the node and the node ; represents the node and the node the weight adjustment factor between them, represents the node and the node the adjusted influence weight between them; Repeat the collaborative correction of the predicted value by integrating adjacent nodes and the dynamic adjustment process of the influence weight until the calculated weight adjustment factor is less than the set weight adjustment factor threshold. At this time, the set of predicted results generated in the current iteration:

[0026] where, represents the optimized result of the microgrid load prediction, represents the optimized result of the load prediction of the first node, represents the th node's optimized result of the load prediction.

[0027] Advantages of the present invention: By constructing a topological time series joint modeling framework that integrates a graph convolutional network and a GRU structure, the present invention realizes the joint modeling of the topological structure information between nodes and the time series load data in the microgrid, effectively captures the complex correlation relationships between nodes, and significantly improves the adaptability and prediction accuracy of the model to sudden load changes; By introducing the "structure - time" joint embedding tensor construction mechanism and combining node types (such as energy storage nodes, load nodes, and power generation nodes) and their power flow direction characteristics, the adaptive generation of prediction samples and classification modeling for different node functional characteristics are realized, enhancing the generalization ability of the model; Through the dynamic adjustment of adjacent weights and structure optimization mechanism guided by errors, the automatic learning of load conduction relationships between nodes and the self - optimization of the network structure are realized, significantly improving the model consistency score and generalization stability in the scenario of multi - node and multi - load collaborative prediction. Description of the Drawings

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts. Among them: Figure 1 It is a schematic structural diagram of the overall method steps of the micro - grid load prediction optimization algorithm based on deep learning of the present invention. Detailed Embodiments

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the protection scope of the present invention.

[0030] Embodiment 1 Referring to Figure 1 , this is the first embodiment of the present invention, providing a micro - grid load prediction optimization algorithm based on deep learning, including the following steps. S1: Collect the operation data of the micro - grid and construct a micro - grid topological structure diagram according to the collected operation data.

[0031] Specifically, the collection of the microgrid operation data is the real-time operation data of the microgrid collected by deploying a sensor network and using the sensors in the sensor network. The sensor network divides the microgrid into multiple nodes, and sensors are deployed at each node. The microgrid topology structure is composed of the node types in the microgrid and the connection edges between different nodes. The node types are determined by deploying sensors at the operation nodes of the microgrid and judging the current node types according to the real-time operation data of the microgrid collected by the sensors, including power generation nodes, energy storage nodes, and load nodes. The connection edges are the connection relationships between different nodes, and the structure formed by the connection edges of all nodes is the microgrid topology structure diagram, which is specifically constructed as follows: Divide the microgrid into nodes, then there is , where represents the node set in the microgrid, represents the first node, represents the th node, and sensors are deployed at each node to collect the real-time operation data of each node, including current data, current phase angle, voltage data, and voltage phase angle. Judge the type of the current node according to the data collected by the sensors at each node. Specifically: Based on the real-time operation data of the microgrid collected by the node, calculate the active power and reactive power corresponding to the current node. At the same time, determine the power flow direction of the current node, then there is Active power

[0032] Reactive power

[0033] where and respectively represent the voltage data and current data collected at the current node, and represent the current phase angle and voltage phase angle corresponding to the current node, represents the active power corresponding to the current node, represents the reactive power corresponding to the current node; Judge the power flow direction of the current node according to the power corresponding to the node, then there is For the current node, select the adjacent nodes of the current node, calculate the active power flow and reactive power flow from the current node to the adjacent nodes respectively, and judge the power flow direction of the current node according to the calculation results. Specifically: Active power from the current node to the adjacent node

[0034] Reactive power from the current node to the adjacent node

[0035] wherein represents the voltage data of the current node represents the voltage data of the adjacent node represents the voltage phase angle of the current node represents the voltage phase angle of the adjacent node represents the impedance between the current node and the adjacent node represents the active power flow from the current node to the adjacent node represents the reactive power flow from the current node to the adjacent node; According to the calculation results of the active power flow and the reactive power flow, judge the power flow direction of the current node, then If the active power flow of the current node satisfies the formula , it means that the power of the current node flows from the current node to the adjacent node, otherwise, it means that it flows from the adjacent node to the current node; If the reactive power flow of the current node satisfies the formula , it means that the reactive power of the current node flows from the current node to the adjacent node, otherwise, it means that it flows from the adjacent node to the current node; According to the active power, reactive power and power flow direction corresponding to the current node, judge the node type, specifically as follows: If the active power corresponding to the current node satisfies the formula , and the power flow direction of the current node is from the current node, it means that the current node type is a power generation node; If the active power corresponding to the current node satisfies the formula , and the reactive power satisfies the formula , and at the same time, the power flow direction of the current node is from the adjacent node into the current node, it means that the current node type is a load node; For a node with the node type of an energy storage node, it is judged based on the state of charge corresponding to the node, specifically: Detect whether the current node has a capacitance parameter. If the current node cannot collect the capacitance parameter, it means that the current node is a non-energy storage node; If the current node can collect the capacitance parameter, it means that the current node type is an energy storage node.

[0036] Based on different node types, construct a microgrid topology structure diagram, specifically: According to the nodes divided by the microgrid, and the identified node types, then , wherein represents the set of load nodes, represents the set of discharge nodes, represents the set of energy storage nodes, and the total number of the three node types is pieces, represents the set of nodes in the microgrid; According to the calculated power flow direction, all nodes are connected based on the graph structure form. After the edge relationships between all nodes are determined, it means that the construction of the microgrid topology graph is completed.

[0037] S2: Based on the in-depth analysis of the microgrid topology graph, determine the degree of association between topological nodes.

[0038] Specifically, the in-depth analysis based on the microgrid topology graph includes analyzing the degree of association between nodes based on the deep learning algorithm, and predicting the operating load of nodes based on historical load. The specific implementation is as follows: Analyzing the degree of association between nodes based on the deep learning algorithm is to use the deep learning algorithm to deeply analyze the edges of the topological structure to determine the degree of association between nodes. The specific implementation is as follows: According to the constructed microgrid topology graph , where represents the set of nodes in the microgrid, represents the set of connection edges between nodes, represents the set of influence weights between nodes; Regarding the influence weight between nodes , which represents the influence weight of node on node , conduct a deep analysis of the influence weight between nodes based on the deep learning algorithm. Specifically: For any node in the microgrid topology structure, extract the node features corresponding to the node, and deeply analyze the operating state of node through the convolutional neural network model, and set the analysis result as . Since the convolutional neural network model is a well-known technology, it will not be elaborated here; Regarding the initial analysis result of node , select a node in the microgrid topology structure that is not the node, and splice the two node features. Then,

[0039] where represents the node features extracted by node Represents a node The extracted node features Represents the concatenated features for in-depth analysis of the influence weights between nodes. Then, Based on the concatenated features, perform state analysis through a convolutional neural network model and set the analysis result as , and compare the analysis results of the two times And , and perform in-depth analysis of the influence weights between nodes based on the comparison result. Specifically: If the comparison result satisfies the formula , it means that the node Relative to the node The influence weight is a negative influence, and the weight coefficient value satisfies the formula , and the specific value is set by the implementer according to the actual application scenario; If the comparison result satisfies the formula , it means that the node Relative to the node The influence weight is a positive influence, and the weight coefficient value satisfies the formula , and the specific value is set by the implementer according to the actual application scenario; If the comparison result satisfies the formula , it means that the node Relative to the node The influence weight is no influence, and the weight coefficient value satisfies the formula .

[0040] It should be noted that for the in-depth analysis of the influence weights between nodes, each time the state analysis is performed for any one node. After the state analysis is completed, the features of the next node are concatenated with those of the previous node (the node for which the analysis is completed), and the concatenated features are subjected to a secondary state analysis. Then, the results of the two state analyses are compared, and based on the analysis results, the in-depth analysis of the influence weights between nodes is carried out, thereby providing a judgment basis for the value of the weight coefficient until the state analyses corresponding to all nodes are completed. The selection of the front and rear nodes is based on the connection edges of the topological structure.

[0041] Furthermore, the running load of the node predicted based on the historical load is to construct the node load sequence features and structural embedding features for each node according to the input historical load data set and the microgrid topological structure diagram, and perform the running load prediction of the node based on the influence weights between nodes. The specific implementation is as follows: For each node, based on its historical load sequence, extract the sliding sequence with a time window length of , and construct the historical state matrix. Then, For any node , set the corresponding historical load sequence, then there is,

[0042] Among them, represents the load data of node at time , including active power and reactive power ; For any node , construct a set of sliding time slices with a time window length of , then there is,

[0043] Among them, represents the historical state matrix corresponding to node , represents the length of the divided time step, represents the total length of the time window, represents one row in the historical state matrix; According to the active power, reactive power of the node and its node type label (generation / load / storage), generate a set of prediction input samples for each node, then there is, Define the node type label , with values of -1, 0, 1 respectively. When the value is -1, it represents a load node. When the value is 0, it represents a storage node. When the value is 1, it represents a generation node; Construct the input sample set , among which, represents the node type label, represents the length of the time step, represents the feature dimension of each time step. If the formula is satisfied, it means that the features of the current time step include active power, reactive power and node type; For the generated set of node historical state samples, use the graph convolutional network to extract the topological features between nodes and perform time series modeling on the extracted topological features to generate an initial prediction result: Input the constructed joint feature tensor into the graph convolutional layer to extract the node graph embedding, then there is,

[0044] Among them, represents the normalized adjacency matrix, represents the graph convolutional weight, represents the activation function, represents the constructed joint feature tensor, Represents the extracted node graph embedding features; Input the graph embedding features into the LSTM network to obtain the hidden state output. Then,

[0045] where, represents the graph embedding feature corresponding to node ; represents the hidden state output corresponding to node ; Generate the initial prediction result based on the hidden state output. Then,

[0046] where, and represent the weight coefficient and bias term of the initial prediction respectively, represents the hidden state output corresponding to node ; represents the initial prediction result of the load corresponding to node ;

[0047] For the generated set of initial prediction results, conduct error analysis on the prediction results, and readjust the graph embedding method based on the prediction error, thereby generating the optimized prediction result (the second-generation prediction value): Perform backpropagation on the error between the first-generation prediction value and the actual value. Then,

[0048] where, represents the initial prediction result of the load corresponding to node ; represents the actual measurement result of the load corresponding to node ; represents the error of the first-generation prediction value; Dynamically update the adjacency weights in the graph structure based on the error of the first-generation prediction value, and correct the influence relationship between nodes. Then,

[0049] where, represents the normalized adjacency matrix, represents the adjusted normalized adjacency matrix, represents the error change of the first-generation prediction value, represents the attention weight, which is set by the implementer according to the actual application scenario; Based on the adjusted normalized adjacency matrix, perform secondary extraction of the graph embedding features of the first prediction value. Then,

[0050] Among them, represents the adjusted normalized adjacency matrix, represents the graph convolution weight, represents the activation function, represents the constructed joint feature tensor, represents the node graph embedding feature extracted for the second time; And based on the second extraction result of the graph embedding feature, the corresponding hidden state output is obtained. Then,

[0051] Among them, represents the node corresponding graph embedding feature extracted for the second time, represents the node hidden state output corresponding to the second extraction result; According to the obtained hidden state data, the second-generation prediction of the node load is performed. Then,

[0052] Among them, , respectively represent the weight coefficient and bias term of the initial prediction, which are set by the implementer according to the actual application scenario, represents the node hidden state output corresponding to the second extraction result, represents the node second prediction result of the load; According to the second-generation prediction value set, by introducing a fitting scoring mechanism, the comparison result between the fitting score corresponding to the second-generation prediction value and the fitting score threshold is judged, and the adjacency structure is readjusted. Based on the reconstructed adjacency structure, the load prediction of the node (third-generation prediction value) is performed. Specifically: Compare the fitting score between the second-generation prediction value and the target load curve. Then,

[0053] Among them, represents the actual load measurement value of the node at time, represents the second-generation prediction value of the node at time, represents the time window length, represents the fitting score of the second-generation prediction value, which is used for reconstructing the adjacency structure. Specifically: Set the fitting score threshold , if the calculated fit score satisfies the formula , it means that the adjacency matrix needs to be readjusted to achieve the reconstruction of the adjacency structure until the fitting score is higher than the set score threshold. Otherwise, it means that there is no need to readjust the adjacency structure. The specific value of the adjustment of the adjacency matrix is ​​set by the implementer according to the actual application scenario. For the adjusted prediction model, the third-generation prediction value is regenerated according to the prediction method of the second-generation prediction value, and the fitting score corresponding to the third-generation prediction value is calculated , compare the fitting score corresponding to the third-generation prediction value with the fitting score corresponding to the second-generation prediction value until the comparison result satisfies the formula So far, it means that the load forecast for the node is accurate, and the third-generation prediction value is used as the load forecast result of the node.

[0054] S3: Based on the influence weights between nodes and the load forecast results of the nodes, the microgrid load forecasting is optimized.

[0055] Specifically, the microgrid load prediction optimization is to coordinately optimize and predict the global load of the microgrid based on the influence weights between nodes in the microgrid topology structure and the load prediction results of the nodes. The specific implementation is as follows: According to the influence weights between nodes and the topological structure of nodes, a global influence weight matrix is ​​constructed , represents the total number of nodes in the microgrid node set, where , positive values ​​indicate positive impact, negative values ​​indicate negative impact, and 0 indicates no impact; According to the node load prediction results, obtain the load of each node at the prediction time. The third generation forecast value of the load is constructed, and the initial forecast set of the global load is constructed.

[0056] in, The initial global load forecast set constructed, represents the third-generation predicted value of the first node, Indicates The third-generation predicted value of nodes; For nodes The prediction result is based on its adjacent node set , combining the prediction outputs of adjacent nodes and making collaborative corrections to the prediction values, we have:

[0057] in, Representation Node The predicted value after integrating the adjacent nodes, The third-generation predicted value of node , The third-generation predicted value of node , The third-generation predicted value of node represents the influence weight between node and node The set of adjacent nodes of node is the set of nodes that can form an edge relationship with node in the topological structure; The collaborative correction coefficient is set by the implementer according to the actual application scenario. Specifically: If the predicted result of node meets the formula when compared with the predicted results of the set of adjacent nodes , it means that the node prediction is perturbed by the adjacent nodes, and the correction coefficient is increased to achieve the correction of the adjacent nodes; For node the predicted value after integrating the adjacent nodes and the error between the actual measured value, calculate the weight adjustment factor between each pair of nodes, then there is

[0058] where represents the adjustment coefficient, represents the predicted value of node after integrating the adjacent nodes, represents the third-generation predicted value of node , The actual measured result of the load of node , represents the weight adjustment factor between node and node for realizing the dynamic adjustment of the influence weight between nodes, then there is

[0059] where represents the influence weight between node and node , represents the weight adjustment factor between node and node , represents the adjusted influence weight between node and node ; Repeat the collaborative correction of the predicted values of comprehensive adjacent nodes and the dynamic adjustment process of influence weights until the calculated weight adjustment factor is less than the set weight adjustment factor threshold (set by the implementer according to the actual application scenario). At this time, the set of prediction results generated in the current iteration:

[0060] Among them, represents the optimized result of microgrid load prediction, represents the optimized result of load prediction for the first node, represents the th optimized result of load prediction for the node.

[0061] Furthermore, if the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage medium includes: USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical disks, etc., which can store program codes.

[0062] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or in combination with these instruction execution systems, apparatus, or devices. For the purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device.

[0063] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection (electronic device) having one or more wirings, a portable computer diskette (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CDROM). Additionally, the computer-readable media can even be paper or other suitable media on which the program can be printed, as the program can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpretation, or, if necessary, other suitable processing, and then storing it in a computer memory.

[0064] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

Claims

1. A microgrid load forecasting optimization algorithm based on deep learning, characterized by: The following steps are included: Using sensors to collect real-time operation data of the microgrid, and based on the collected real-time operation data, dividing the microgrid into different types of nodes, and constructing a microgrid topology diagram based on the connection relationship between all nodes; The influence weights between nodes are analyzed through deep learning algorithms, and the operating load of nodes is predicted using historical loads. Specifically: According to the input historical load data set and microgrid topology diagram, the node load sequence characteristics and structural embedding characteristics are constructed for each node, and the node operation load is predicted based on the influence weights between nodes; According to the influence weights between nodes in the microgrid topology and the load prediction results of the nodes, the global load of the microgrid is collaboratively optimized and predicted, specifically: According to the prediction results of the node, the prediction outputs of the adjacent nodes are integrated to make collaborative corrections to the prediction values. At the same time, according to the errors between the prediction outputs of the adjacent nodes and the actual measured values, the influence weights between the nodes are dynamically adjusted, and the microgrid load is predicted based on the adjusted influence weights.

2. The microgrid load prediction optimization algorithm based on deep learning according to claim 1 is characterized in that: The use of sensors to collect real-time operation data of the microgrid is specifically as follows: Calculate the active power and reactive power corresponding to the current node, and determine the flow direction of the current node, then we have: Active power, Reactive power, in, , Respectively represent the voltage data and current data collected at the current node, , Indicates the current phase angle and voltage phase angle corresponding to the current node, Indicates the active power corresponding to the current node, Indicates the reactive power corresponding to the current node.

3. The microgrid load prediction optimization algorithm based on deep learning according to claim 2 is characterized in that: The flow direction of the current node is determined as follows: For the current node, select the adjacent nodes of the current node, calculate the active power flow and reactive power flow from the current node to the adjacent nodes respectively, and judge the power flow direction of the current node according to the calculation results, specifically: The active power from the current node to the adjacent node, Reactive power from the current node to the adjacent node, in, Represents the voltage data of the current node, Represents the voltage data of adjacent nodes, represents the voltage phase angle of the current node, represents the voltage phase angle of adjacent nodes, represents the impedance between the current node and the adjacent nodes, Represents the active power flow from the current node to the adjacent node, Represents the reactive power flow from the current node to the adjacent node; According to the calculation results of active power flow and reactive power flow, the power flow direction of the current node is determined, and then, If the active power flow of the current node satisfies the formula , indicating that the power of the current node flows from the current node to the adjacent node, and vice versa, indicating that the power flows from the adjacent node to the current node; If the reactive power flow of the current node satisfies the formula , indicating that the reactive power of the current node flows from the current node to the adjacent node, and vice versa, it means that it flows from the adjacent node to the current node.

4. The microgrid load prediction optimization algorithm based on deep learning according to claim 3 is characterized in that: The microgrid is divided into different types of node types as follows: If the active power corresponding to the current node satisfies the formula , and the current node's power flow direction is outflowing from the current node, which means that the current node type is a power generation node; If the active power corresponding to the current node satisfies the formula , and the reactive power satisfies the formula ,At the same time, if the flow direction of the current node is flowing from the adjacent node to the current node, it means that the current node type is a load node; For nodes whose node type is energy storage node, the judgment is based on the charge state of the node, specifically: Detect whether the current node has capacitance parameters. If the current node cannot collect capacitance parameters, it means that the current node is a non-energy storage node; If the current node can collect capacitance parameters, it means that the current node type is an energy storage node.

5. The microgrid load prediction optimization algorithm based on deep learning according to claim 4 is characterized in that: The influence weights between nodes analyzed by deep learning algorithm are as follows: For any node in the microgrid topology , extract the node features corresponding to the node , through the convolutional neural network model to the node The running status of the ; For nodes The initial analysis results , in the microgrid topology, choose non Node of Node , and concatenate the two node features, then we have, in, Representation Node The extracted node features, Representation Node The extracted node features, Represents the concatenated features, which are used for in-depth analysis of the influence weights between nodes. Then, According to the spliced ​​features, the state analysis is performed through the convolutional neural network model, and the analysis results are set as , comparing the two analysis results as well as , based on the comparison results Conduct an in-depth analysis of the influence weights between nodes, specifically: If the comparison result satisfies the formula , then it means the node Relative to the node The impact weight is negative; If the comparison result satisfies the formula , then it means the node Relative to the node The influence weight is positive; If the comparison result satisfies the formula , then it means the node Relative to the node The influence weight is no influence.

6. The microgrid load prediction optimization algorithm based on deep learning according to claim 5 is characterized in that: The operation load of the node predicted by using the historical load is specifically as follows: For each node, based on its historical load sequence, the length of the extraction time window is The sliding sequence and construct the historical state matrix; Define node type labels and construct input sample sets. Use graph convolutional networks to extract topological features between nodes and perform time series modeling on the extracted topological features to generate initial prediction results. According to the generated initial prediction result set, the prediction results are analyzed for errors and the prediction errors are calculated, and the graph embedding method is readjusted based on the prediction errors to generate the second generation prediction values; According to the second-generation prediction value set, by introducing the fitting scoring mechanism, the comparison result between the fitting score corresponding to the second-generation prediction value and the fitting score threshold is judged, the adjacency structure is readjusted, and the node load prediction is performed based on the reconstructed adjacency structure.

7. The microgrid load prediction optimization algorithm based on deep learning according to claim 6 is characterized in that: The initial prediction results are generated as follows: Input the constructed joint feature tensor into the graph convolution layer to extract the node graph embedding, then we have, in, represents the normalized adjacency matrix, represents the graph convolution weight, represents the activation function, represents the constructed joint feature tensor, represents the extracted node graph embedding features; Embedding graph into features Input to the LSTM network and obtain the hidden state output, then we have, in, Representation Node The corresponding graph embedding features, Representation Node The corresponding hidden state output; Based on the hidden state output, the initial prediction result is generated, then, in, , Respectively represent the weight coefficient and bias term of the initial prediction, Representation Node The corresponding hidden state output is, Representation Node Initial load forecast results.

8. The microgrid load forecasting optimization algorithm based on deep learning according to claim 7 is characterized in that: The load forecast of the node is as follows: Comparing the fitting scores between the second generation prediction value and the target load curve, we have: in, Representation Node exist The actual load measurement value at the moment, Representation Node exist The second generation prediction value at time, represents the time window length, Represents the fitting score of the second-generation prediction value, which is used to reconstruct the adjacency structure, specifically: Setting the Fit Score Threshold , if the calculated fit score satisfies the formula , it means that the adjacency matrix needs to be readjusted to achieve the reconstruction of the adjacency structure until the fitting score is higher than the set score threshold. Otherwise, it means that there is no need to readjust the adjacency structure. For the adjusted prediction model, the third-generation prediction value is regenerated according to the prediction method of the second-generation prediction value, and the fitting score corresponding to the third-generation prediction value is calculated , compare the fitting score corresponding to the third-generation prediction value with the fitting score corresponding to the second-generation prediction value until the comparison result satisfies the formula So far, it means that the load forecast for the node is accurate, and the third-generation prediction value is used as the load forecast result of the node.

9. The microgrid load forecasting optimization algorithm based on deep learning according to claim 8 is characterized in that: The collaborative correction of the predicted value is specifically as follows: For nodes The prediction result is based on its adjacent node set , combining the prediction outputs of adjacent nodes and making collaborative corrections to the prediction values, we have: in, Representation Node The predicted value after integrating the adjacent nodes, Representation Node The third generation prediction value of Representation Node The third generation prediction value of Representation Node With Node The influence weight between Representation Node The set of adjacent nodes is the node in the topological structure. A set of nodes that can form edge relationships between them. represents the synergy correction coefficient, specifically: If the node The prediction result of , compared with the prediction result of the adjacent node set, satisfies the formula , indicating that the node prediction is disturbed by the adjacent nodes, and the correction coefficient is increased to achieve the correction of the adjacent nodes.

10. The microgrid load prediction optimization algorithm based on deep learning according to claim 9 is characterized in that: The microgrid load forecasting based on the adjusted impact weight is specifically as follows: For nodes The predicted value after integrating adjacent nodes The error between the actual measured value and the weight adjustment factor between each pair of nodes is calculated, then: in, represents the adjustment factor, Representation Node The predicted value after integrating the adjacent nodes, Representation Node The third generation prediction value of node Actual measurement results of load, Representation Node With Node The weight adjustment factor between nodes is used to realize the dynamic adjustment of the weights between nodes, then, in, Representation Node With Node The influence weight between Representation Node With Node The weight adjustment factor between Representation Node With Node The adjusted impact weights between Repeat the coordinated correction of the predicted value by the integrated adjacent nodes and the dynamic adjustment process of the weights until the calculated weight adjustment factor is less than the set weight adjustment factor threshold. At this time, the set of prediction results generated in the current iteration is: in, represents the optimization result of microgrid load forecasting, represents the load forecast optimization result of the first node, Indicates Load forecasting optimization results for each node.

Images