Energy safety optimization method based on deep learning

Abstract

The invention discloses an energy safety optimization method based on deep learning. The method comprises the steps that S1, multi-source data from an energy system is collected and preprocessed; s2, constructing a graph structure model of the energy system; s3, processing the graph structure by applying a physical perception convolutional neural network, and generating a global feature vector of each node; s4, inputting the global feature vector into an adaptive fuzzy inference system to generate a final scheduling strategy; s5, based on the difference between the final scheduling strategy and the expected target scheduling strategy, performing adaptive adjustment on the membership function of the fuzzy inference rule; s6, adjusting the running state of the energy equipment in real time in combination with the node feature vector and a scheduling strategy; and S7, collecting real-time feedback data, and further optimizing the scheduling strategy of the energy system according to the real-time feedback data. According to the method, an efficient and scientific optimization scheme can be provided in energy safety optimization, and remarkable technical values and economic benefits are brought to practical application.

Description

Technical Field

[0001] The present invention relates to the technical field of energy security optimization, and particularly to an energy security optimization method based on deep learning. Background Art

[0002] The problem of optimal scheduling of energy systems has always been a research focus in the field of energy management and control. Especially in modern power, heating and other energy networks, how to achieve efficient resource allocation and equipment scheduling to improve the stability and energy-saving effect of the system has become a topic widely concerned by the academic and industrial communities. At present, many energy system optimization methods adopt scheduling models based on traditional algorithms, such as linear programming, dynamic programming, genetic algorithms, etc. These methods optimize the configuration of resources and equipment scheduling by constructing mathematical models. However, these methods are usually restricted by factors such as computational complexity, non-linear relationships and environmental changes, resulting in certain deficiencies when dealing with large-scale systems and complex environments.

[0003] In recent years, with the rapid development of artificial intelligence and deep learning technologies, deep learning-based methods have gradually become an effective way to optimize the scheduling of energy systems. As an emerging deep learning model, the physics-informed convolutional neural network has achieved remarkable results in the fields of image recognition, signal processing, etc. It can effectively capture local features through convolutional operations and can handle complex spatial dependencies. Therefore, it is expected to be applied to the scheduling problem of energy systems. On the other hand, as a tool for solving uncertainty and ambiguity problems, the adaptive fuzzy inference system is widely used in the fields of control systems, decision-making support, etc. The adaptive fuzzy inference system can dynamically adjust fuzzy rules and parameters according to the input real-time data, thereby optimizing scheduling decisions and enhancing the adaptability and robustness of the system.

[0004] However, the existing energy system optimization methods based on deep learning and fuzzy inference still face some challenges. Although the physics-informed convolutional neural network has advantages in extracting spatial features, it often lacks sufficient adaptability to dynamic changes when dealing with spatio-temporal dependencies in energy systems. Traditional convolutional neural networks have certain limitations in modeling time-series data. Therefore, how to effectively combine spatio-temporal information to accurately predict and optimize the real-time dynamics of the system is an urgent problem to be solved.

[0005] In addition, there are certain deficiencies in the existing technologies in terms of data processing and feedback mechanisms. Many traditional methods rely on static datasets and preset scheduling strategies, and cannot fully take into account the real-time changes in system states and environmental conditions. In the actual operation of energy systems, fluctuations in load demands, equipment states, and external environments are often unpredictable, and traditional optimization algorithms and fuzzy inference systems often prove inadequate in the face of these dynamic changes. Due to the lack of an effective real-time feedback mechanism, existing optimization methods are difficult to quickly respond to changes in system states, resulting in low scheduling efficiency of energy systems and an inability to make flexible adjustments in a short period of time.

[0006] To address these problems, the present invention proposes an energy security optimization method combining a physics-aware convolutional neural network and adaptive fuzzy inference. By introducing the joint modeling of spatio-temporal features and integrating the advantages of deep learning and fuzzy inference systems, this method effectively overcomes the deficiencies of traditional methods in dealing with spatio-temporal dependence relationships and real-time feedback mechanisms. The physics-aware convolutional neural network can extract node features and capture the spatial dependence relationships in the system, while the adaptive fuzzy inference system ensures the optimal scheduling of the system in different time periods and spatial ranges by adjusting fuzzy rules and parameters in real time. This joint optimization method based on deep learning and fuzzy inference not only improves the scheduling efficiency of energy systems but also enhances their adaptability to uncertainties and complex environmental changes. Summary of the Invention

[0007] An object of the present invention is to propose an energy security optimization method based on deep learning. The present invention can provide an efficient and scientific optimization scheme in energy security optimization, bringing significant technical value and economic benefits to practical applications.

[0008] An energy security optimization method based on deep learning according to an embodiment of the present invention includes the following steps: S1. Collect multi-source data from the energy system and perform preprocessing; S2. Construct a graph structure model of the energy system, taking each energy device, region, and load point as a node in the graph, and taking energy flows and device connections as edges in the graph; S3. Apply a physics-aware convolutional neural network to process the graph structure, extract the spatial features of each node, capture the physical dependence relationships between devices, and generate a global feature vector for each node; S4. Input the global feature vector generated by the physics-aware convolutional neural network into an adaptive fuzzy inference system, and generate a final scheduling strategy according to the real-time energy supply and demand situation and external environmental changes; S5. Based on the difference between the final scheduling strategy and the expected target scheduling strategy, adaptively adjust the membership function of the fuzzy inference rule; S6. Combine the node feature vectors generated by the physical perception convolutional neural network and the scheduling strategy generated by the adaptive fuzzy inference to adjust the operating state of the energy equipment in real time; S7. During the operation of the system, collect real-time feedback data by continuously monitoring the status of energy equipment, load demand, and external environment changes, and further optimize the scheduling strategy of the energy system according to the real-time feedback data.

[0009] Optionally, S1 includes the following steps: S11. Automatically collect real-time data from various links in the energy system through a data acquisition system, including energy production data, power load data, equipment status information, meteorological data, and historical event data; S12. Preprocess the collected raw data. The preprocessing steps include: standardizing the data and filling in the missing values in the data.

[0010] Optionally, S2 includes the following steps: S21. According to the topological structure of the energy system, define each energy equipment, area, and load point as a node in the graph, and define the energy flow and equipment connection as the edges in the graph to construct a graph structure model of the energy system; S22. Calculate the weights of the edges according to the energy flow situation between the equipment. , and the calculation formula is as follows: ; Where, is the edge weight between node i and node j, is the energy flow value from node i to node j, is the set of all nodes connected to node i, is the energy flow value from node i to node k; S23. Assign initial features to each node. The initial features are the status data of each node, including energy production capacity, load demand, and equipment health status. The feature vector of the node is obtained by standardizing the node status data , and a node feature matrix X is constructed based on the feature vectors of each node; S24. Determine the adjacency relationship between nodes through the graph structure model to construct an adjacency matrix A. Each element in the adjacency matrix A is: ; Where, is the weight of the edge between node i and node j.

[0011] Optionally, S3 includes the following steps: S31. Perform convolution processing on the graph structure model. Use the convolution layer in the physics-aware convolutional neural network to process the input node feature matrix: ; Among them, is the node feature matrix after convolution operation, is the normalized adjacency matrix of the adjacency matrix A, is the ReLU activation function, X is the node feature matrix, is the weight matrix of the convolution layer, is the bias term; S32. Based on the adaptive graph convolution mechanism, extract multi-order adjacency information through multiple graph convolution layers: ; Among them, is the output feature matrix of the -th layer convolution operation, representing the multi-step propagation relationship between nodes, is the -th layer's k-th order weight matrix, is the -th layer's bias term, is the output of the previous layer's convolution operation, K is the order used by the convolution layer; S33. Perform pooling on the graph convolution result after L-layer processing to generate the global feature vector H of each node.

[0012] Optionally, the S4 includes the following steps: S41. Input the node global feature vector H generated by the physics-aware convolutional neural network into the adaptive fuzzy inference system, and automatically generate fuzzy inference rules according to the content of the node global feature vector H; S42. Calculate the association degree between nodes based on the similarity of the node global feature vector and : ; Among them, is the similarity between node i and node j, is the feature vector of node i, is the feature vector of node j, is the norm of the node feature vector; S43. Use the calculated similarity to automatically generate fuzzy sets and for each pair of feature vectors, and derive the corresponding fuzzy inference rules according to the similarity; S44. Define fuzzy inference rules, which are based on real-time energy supply and demand situations and external environmental factors and describe the scheduling and load distribution logic between energy devices, and are specifically expressed as: ; where, is the feature vector of node i, is the feature vector of node j, is 's fuzzy set, is 's fuzzy set, Y is the output scheduling strategy, is the corresponding fuzzy result; S45. Use the fuzzy inference method to deduce the rule set, and combine the input node feature vector H to deduce the fuzzy set. The inference process obtains the inference result of each rule by adopting a weighted membership function, and this weighted membership function is adjusted not only based on the input membership degree but also according to the dynamic weights of each fuzzy set: ; where, is the output corresponding to the kth rule, is the input belongs to the fuzzy set 's membership degree, is the dynamic weight associated with the fuzzy set , is the membership degree that the output Y belongs to the fuzzy set , is the dynamic weight associated with the fuzzy set , N is the total number of rules; S46. Fuse the inference results of all rules through a weighted fusion strategy to generate the final scheduling strategy .

[0013] Optionally, S46 includes the following steps: S461. The weighted fusion strategy dynamically adjusts according to the importance of the rules by introducing an adaptive weight, and the calculation method of the final scheduling strategy is: ; where, is the adaptive weight of rule k, is the input belongs to the fuzzy set 's membership degree, is the output membership degree corresponding to rule k, M is the total number of the rule set.

[0014] Optionally, S5 includes the following steps: S51. Utilize the final scheduling strategy output by the system and the expected target scheduling strategy to calculate the error value : ; S52. According to the error value and the membership function of the fuzzy rule, use the gradient descent optimization algorithm to adjust the parameters in the fuzzy rule; S53. Perform adaptive adjustment on the membership function of the fuzzy inference rule according to the feedback data. The adaptive adjustment includes dynamically weighting each parameter of the membership function: ; Among them, is the dynamic weight of the k-th rule, is the adaptive weight of rule k, is the error value of rule k, gradually decreases as the error increases, and vice versa.

[0015] Optionally, the S52 includes the following steps: S521. The parameter adjustment formula in the fuzzy rule is: ; Among them, is the update amount of the membership function , is the learning rate, is the gradient of the error function with respect to the membership function.

[0016] Optionally, the S6 includes the following steps: S61. Combine the global feature vector H of the nodes generated by the physics-aware convolutional neural network and the final scheduling strategy generated by the adaptive fuzzy inference to construct the spatio-temporal feature vector of each node : ; Among them, is the vector concatenation operation, represents the function of combining node features and scheduling strategies; S62. Model the interdependence relationship between nodes through the spatio-temporal feature vector to construct the global spatio-temporal relationship matrix S. Each element in the matrix S represents the spatio-temporal dependence relationship between node i and node j: ; Among them, is the similarity between node i and node j, and are the spatio-temporal feature vectors of node i and node j, and is the norm of the spatio-temporal feature vector; S63. Perform global optimization of the graph structure through the spatio-temporal relationship matrix S. The optimization objective is to minimize the spatio-temporal differences between nodes to improve the overall efficiency of the energy system. The optimization process is carried out by using a graph smoothing method based on the Laplacian operator to obtain the optimized node feature matrix : ; wherein, is the optimized node feature matrix, L = D - S is the Laplacian matrix, D is the degree matrix, and S is the spatio-temporal relationship matrix; S64. Based on the optimized node feature matrix and the spatio-temporal relationship matrix S, update the feature representations of each node; S65. Combine the spatio-temporally optimized node features with the adaptive fuzzy inference system to further adjust the operating states of the energy devices.

[0017] Optionally, the S7 includes the following steps: S71. During the operation of the system, continuously monitor the states of the energy devices, the load demands, and the changes in the external environment, and collect feedback data in real time; S72. Jointly analyze the collected feedback data with the node feature vectors generated by the physical perception convolutional neural network, calculate the relationship between the node features and the feedback data, and generate a feedback correction factor; S73. Dynamically adjust the node feature vectors according to the feedback correction factor; S74. Use the updated node feature vectors to continue performing the scheduling decisions of the energy devices and transmit them to the adaptive fuzzy inference system for the next round of scheduling strategy optimization.

[0018] The beneficial effects of the present invention are: (1) By combining a physics-informed convolutional neural network with an adaptive fuzzy inference system, the present invention successfully addresses multiple deficiencies in existing energy system optimization technologies, achieving significant beneficial effects. First, the physics-informed convolutional neural network plays an important role in extracting the spatial features of energy system nodes, capable of effectively capturing the physical dependencies between devices. In traditional energy system scheduling methods, the complex spatial and spatio-temporal dependencies between devices are often overlooked, resulting in limitations in scheduling efficiency and accuracy. By introducing the physics-informed convolutional neural network, the present invention can perform more fine-grained feature modeling of the energy system, enhancing the scheduling system's perception of spatial dependencies, thereby effectively optimizing the scheduling decisions between devices.

[0019] (2) Based on the adaptive fuzzy inference system, the present invention overcomes the limitations of traditional fuzzy inference methods by introducing a dynamic feedback mechanism. In traditional fuzzy inference systems, the rules and parameters are usually static and lack the ability to respond to real-time feedback data, making it difficult for the fuzzy inference system to cope with complex and rapidly changing energy environments. The present invention automatically corrects the system based on real-time feedback data by dynamically adjusting the fuzzy rules and parameters, ensuring that the scheduling strategy can adapt to changes in energy load demand and fluctuations in device operating states in real time, thereby greatly improving the system's adaptability and robustness.

[0020] (3) By combining the physics-informed convolutional neural network and the adaptive fuzzy inference system, the present invention can achieve more accurate spatio-temporal feature modeling. In energy systems, spatio-temporal features are crucial for scheduling decisions. Traditional methods often overlook spatio-temporal dependencies, resulting in poor performance when dealing with complex changes. The present invention ensures that the system can make optimal decisions in different time periods and spatial ranges by jointly modeling spatio-temporal features, thereby enhancing the overall efficiency and stability of the energy system. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings: Figure 1 is a flowchart of a deep learning-based energy security optimization method proposed by the present invention; Figure 2 is a flowchart of the generation of the node global feature vector in a deep learning-based energy security optimization method proposed by the present invention; Figure 3 is a flowchart of the generation of the final scheduling strategy in a deep learning-based energy security optimization method proposed by the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0023] Reference Figures 1-3 , an energy security optimization method based on deep learning, comprising the following steps: S1. Collect multi-source data from the energy system and perform preprocessing; S2. Construct a graph structure model of the energy system, taking each energy device, region, and load point as nodes in the graph, and taking energy flow and device connections as edges in the graph; S3. Apply a physics-aware convolutional neural network to process the graph structure, extract the spatial features of each node, capture the physical dependencies between devices, and generate a global feature vector for each node; S4. Input the global feature vector generated by the physics-aware convolutional neural network into an adaptive fuzzy inference system, and generate a final scheduling strategy according to the real-time energy supply and demand situation and external environment changes; S5. Based on the difference between the final scheduling strategy and the expected target scheduling strategy, adaptively adjust the membership function of the fuzzy inference rules; S6. Combine the node feature vector generated by the physics-aware convolutional neural network and the scheduling strategy generated by the adaptive fuzzy inference to adjust the operating state of the energy device in real time; S7. During the operation of the system, collect real-time feedback data by continuously monitoring the energy device status, load demand, and external environment changes, and further optimize the scheduling strategy of the energy system according to the real-time feedback data.

[0024] By automatically collecting and preprocessing multi-source data from the energy system, the present invention can ensure the consistency and accuracy of data quality. This process, through techniques such as standardization and missing value filling, avoids the impact of incomplete or inconsistent data on the subsequent optimal scheduling process. By adopting the standardization method, the system can process data from different devices and sensors, ensuring that different types of data are processed on a unified scale, thereby improving the accuracy of subsequent modeling and decision-making. The filling of missing values guarantees data integrity, ensuring that the system can process incomplete inputs during real-time operation, further enhancing the robustness and stability of the system.

[0025] In this embodiment, S1 includes the following steps: S11. Automatically collect real-time data from each link in the energy system through a data acquisition system, including energy production data, power load data, device status information, meteorological data, and historical event data; S12. Preprocess the collected raw data. The preprocessing steps include: standardizing the data and filling in the missing values in the data.

[0026] In the present invention, by constructing a graph structure model of the energy system, the relationships between each energy device, area, and load point are defined, forming a multi-dimensional network model. This method effectively captures the complex spatial dependence relationships between the devices in the system, enabling the system to comprehensively understand the interactions and resource flows between the devices. This graph structure modeling method ensures that the interconnections between the devices can be reflected in subsequent calculations and provides clear spatial relationship data, facilitating feature extraction and scheduling decision-making in the subsequent optimization process.

[0027] In this embodiment, S2 includes the following steps: S21. According to the topological structure of the energy system, define each energy device, area, and load point as a node in the graph, and define the energy flow and device connections as the edges in the graph, to construct a graph structure model of the energy system; S22. Calculate the weights of the edges according to the energy flow conditions between the devices , and the calculation formula is as follows: ; Wherein, is the edge weight between node i and node j, is the energy flow value from node i to node j, is the set of all nodes connected to node i, is the energy flow value from node i to node k; S23. Assign initial features to each node. The initial features are the state data of each node, including energy production capacity, load demand, and device health status. By standardizing the node state data, the feature vector of the node is obtained , and a node feature matrix X is constructed based on the feature vectors of each node; S24. Determine the adjacency relationship between the nodes through the graph structure model, and construct an adjacency matrix A. Each element in the adjacency matrix A is: ; Wherein, is the weight of the edge between node i and node j.

[0028] In the present invention, by applying a physically aware convolutional neural network to extract features from the graph structure model, the physical dependence relationship between nodes in the energy system can be accurately captured. The application of the convolutional neural network to graph structure data enables the system to extract spatial features from multiple dimensions, thereby enhancing the understanding ability of the complex structure of the system. Through multiple processes of the convolutional layer, the system can deeply explore the potential connections between nodes layer by layer and generate high-quality feature vectors suitable for scheduling decisions. This process improves the accuracy of feature extraction, provides more representative input data for the subsequent adaptive fuzzy inference system, and further improves the optimization effect.

[0029] In this embodiment, S3 includes the following steps: S31. Perform convolutional processing on the graph structure model, and use the convolutional layer in the physically aware convolutional neural network to process the input node feature matrix: ; Among them, is the node feature matrix after convolutional operation, is the normalized adjacency matrix of the adjacency matrix A, is the ReLU activation function, X is the node feature matrix, is the weight matrix of the convolutional layer, is the bias term; S32. Based on the adaptive graph convolution mechanism, extract multi-order adjacency information through multiple graph convolutional layers: ; Among them, is the output feature matrix of the -th convolutional operation, representing the multi-step propagation relationship between nodes, is the k-th order weight matrix of the -th layer, is the bias term of the -th layer, is the output of the previous convolutional operation, and K is the order used in the convolutional layer; S33. Perform weighted graph pooling operation on the graph convolution result after L-layer processing to generate the global feature vector H of each node.

[0030] The present invention generates a scheduling strategy based on the node feature vector through an adaptive fuzzy inference system, realizing the dynamic scheduling and load distribution of energy devices. Different from the traditional static rule inference system, the present invention can dynamically adjust the fuzzy rules according to the real-time energy demand and external environment changes, ensuring that the system can adapt to uncertainty and complexity. This flexibility enables the system to optimize the scheduling strategy in real time, avoiding the problems of low efficiency and insufficient adaptability caused by fixed rules in the traditional method. Through this dynamic adjustment, the system can maintain a high operating efficiency and stability under different working conditions.

[0031] In this embodiment, S4 includes the following steps: S41. Input the node global feature vector H generated by the physical perception convolutional neural network into the adaptive fuzzy inference system, and automatically generate fuzzy inference rules according to the content of the node global feature vector H; S42. Calculate the association degree between nodes based on the similarity of and : ; Among them, is the similarity between node i and node j, is the feature vector of node i, is the feature vector of node j, is the norm of the node feature vector; S43. Use the calculated similarity to automatically generate fuzzy sets and for each pair of feature vectors, and derive the corresponding fuzzy inference rules according to the similarity; S44. Define the fuzzy inference rules, which describe the scheduling and load distribution logic between energy devices based on the real-time energy supply and demand situation and external environmental factors, and are specifically expressed as: ; Among them, is the feature vector of node i, is the feature vector of node j, is 's fuzzy set, is 's fuzzy set, Y is the output scheduling strategy, is the corresponding fuzzy result; S45. Use the fuzzy inference method to derive the rule set, and combine the input node feature vector H to derive the fuzzy set. The inference process obtains the inference result of each rule by adopting a weighted membership function, and this weighted membership function is adjusted not only based on the input membership degree but also according to the dynamic weights of each fuzzy set: ; Among them, is the output corresponding to the k-th rule, is the input belongs to the membership degree of the fuzzy set . is the dynamic weight associated with the fuzzy set , is the membership degree that the output Y belongs to the fuzzy set , is the dynamic weight associated with the fuzzy set , and N is the total number of rules; S46. Fuse the inference results of all rules through a weighted fusion strategy to generate a final scheduling strategy .

[0032] In this embodiment, S46 includes the following steps: S461. The weighted fusion strategy dynamically adjusts according to the importance of the rules by introducing an adaptive weight, and the final scheduling strategy The calculation method is: ; Among them, is the adaptive weight of rule k, is the input belongs to the membership degree of the fuzzy set , is the output membership degree corresponding to rule k, and M is the total number of rule sets.

[0033] The present invention can make the energy system quickly respond to the changing external environment and internal conditions by adjusting the rules and parameters in the fuzzy inference system in real time. Through the error adjustment mechanism based on real-time feedback, the present invention can automatically optimize the membership function and weight of the fuzzy rules, thereby improving the decision-making accuracy of the system. This process not only enhances the system's ability to cope with uncertainty and volatility, but also ensures that in a dynamic environment such as energy load fluctuations and equipment state changes, the system can make optimal adjustments in real time, thereby improving the adaptability and stability of the system.

[0034] In this embodiment, S5 includes the following steps: S51. Use the difference between the final scheduling strategy output by the system and the expected target scheduling strategy to calculate the error value : ; S52. According to the error value The membership functions of the fuzzy rules, and use the gradient descent optimization algorithm to adjust the parameters in the fuzzy rules; S53. Adaptively adjust the membership functions of the fuzzy inference rules according to the feedback data, and the adaptive adjustment includes dynamically adjusting the weights of each parameter of the membership functions: ; Among them, is the dynamic weight of the k-th rule, is the adaptive weight of rule k, is the error value of rule k, gradually decreases as the error increases, and vice versa.

[0035] In this embodiment, S52 includes the following steps: S521. The parameter adjustment formula in the fuzzy rules is: ; Among them, is the update amount of the membership function , is the learning rate, is the gradient of the error function with respect to the membership function.

[0036] Through spatio-temporal feature modeling, the present invention combines the node feature vectors generated by the physics-aware convolutional neural network with the scheduling strategies generated by the adaptive fuzzy inference system, improving the ability to capture spatio-temporal relationships. Spatio-temporal feature modeling enables the system to more accurately describe the changes in energy demand and supply in different time periods and spatial ranges. By introducing spatio-temporal optimization, the system can comprehensively consider the dynamic changes in time and space in the scheduling decision-making, effectively avoiding the problems caused by ignoring spatio-temporal dependencies in traditional methods, which significantly improves the overall scheduling efficiency and stability of the energy system.

[0037] In this embodiment, S6 includes the following steps: S61. Combine the global feature vector H of the nodes generated by the physics-aware convolutional neural network and the final scheduling strategy generated by the adaptive fuzzy inference to construct the spatio-temporal feature vector of each node: Among them, is the vector concatenation operation, represents a function that combines node features and scheduling strategies; S62. Model the interdependent relationships between nodes through the spatio-temporal feature vector to construct the global spatio-temporal relationship matrix S, and each element Represents the spatio-temporal dependence relationship between node i and node j: ; where is the similarity between node i and node j, and are the spatio-temporal feature vectors of node i and node j, and are the norms of the spatio-temporal feature vectors; S63. Perform global optimization of the graph structure through the spatio-temporal relationship matrix S. The optimization objective is to minimize the spatio-temporal differences between nodes to improve the overall efficiency of the energy system. The optimization process is carried out by using a graph smoothing method based on the Laplacian operator to obtain the optimized node feature matrix : ; where is the optimized node feature matrix, L = D - S is the Laplacian matrix, D is the degree matrix, and S is the spatio-temporal relationship matrix; S64. Based on the optimized node feature matrix and the spatio-temporal relationship matrix S, update the feature representations of each node; S65. Combine the spatio-temporally optimized node features with the adaptive fuzzy inference system to further adjust the operating states of the energy devices.

[0038] Through real-time collection and feedback of system operation data, the present invention further adjusts the node feature vectors, effectively ensuring the flexibility and adaptability of the system during operation. The real-time feedback mechanism enables the system to timely adjust the node feature vectors and re-optimize the scheduling strategy when detecting equipment failures, load fluctuations, or external environmental changes. This data-driven method based on feedback greatly improves the response speed and decision-making accuracy of the system in a dynamic environment. Through this dynamic adjustment, the system can quickly make adaptive adjustments when facing complex and uncertain situations, ensuring the efficient and safe operation of the energy system.

[0039] In this embodiment, S7 includes the following steps: S71. During the operation of the system, continuously monitor the states of energy devices, load demands, and external environmental changes, and collect feedback data in real time; S72. Jointly analyze the collected feedback data with the node feature vectors generated by the physical perception convolutional neural network, calculate the relationship between the node features and the feedback data, and generate a feedback correction factor; S73. Dynamically adjust the node feature vectors according to the feedback correction factor; S74. Continue to execute the scheduling decision of the energy device using the updated node feature vector and transmit it to the adaptive fuzzy inference system for the optimization of the next-round scheduling strategy.

[0040] Embodiment: In this embodiment, taking the smart grid system of a certain city as the scenario, it details how to optimize the power dispatching of this city, improve the operation efficiency and stability of the system by applying the physical-aware convolutional neural network and the adaptive fuzzy inference system in the energy security optimization method of the present invention. The power grid of this city includes multiple substations, distribution equipment, and millions of end-users. Due to large fluctuations in energy demand, uneven loads on power equipment, and unpredictable external environmental factors (weather changes), traditional power grid dispatching methods often face problems such as slow response, uneven resource allocation, and poor power supply stability.

[0041] The power grid system of this city covers the urban area and surrounding industrial parks and commercial areas. The power demand fluctuates significantly at different times. Especially in summer, due to the sharp increase in air conditioner usage, the load demand rises substantially, resulting in an overly high power grid load and certain power supply instability. At the same time, there are multiple high-rise buildings and large commercial centers in the urban area, with high requirements for the timeliness of power supply. The power grid management department hopes to improve the resource utilization rate of the power grid, reduce the risk of equipment overload, and ensure stable and reliable power supply under high-load conditions through real-time optimization of dispatching.

[0042] To optimize the operation of the power grid in this city, the method of the present invention is introduced into the power grid dispatching system. The physical-aware convolutional neural network and the adaptive fuzzy inference system are used together to predict, optimize dispatching, and provide real-time feedback on the load demand of the power grid. First, the system collects data of each node in the power grid through various sensors, including voltage, current, and load data of substations, as well as power consumption demands of users and meteorological data. Then, these data are preprocessed and input into the physical-aware convolutional neural network to extract spatial dependence features and spatio-temporal features between power grid nodes. The adaptive fuzzy inference system automatically adjusts the dispatching strategy and fuzzy rules based on these features and real-time feedback data, so as to optimize the load distribution and ensure the stable operation of the power grid during high-load periods.

[0043] The following is the specific data comparison table before and after implementation: Table 1 Comparison of the load rates of power grid equipment before and after implementation ; By implementing the method of the present invention, the actual operation data shows that the power grid load forecasting is more accurate, the equipment load is more balanced, the number of power supply interruptions is significantly reduced, and the power supply stability of the power grid is significantly improved. By dynamically adjusting the power dispatching strategy, the power grid can flexibly respond to external environmental changes, load fluctuations, and equipment status changes, thereby effectively improving the operation efficiency, stability, and adaptability of the power grid. This shows that the technology of the present invention has significant beneficial effects in improving the dispatching optimization of the energy system.

[0044] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent substitutions or changes should be covered within the protection scope of the present invention.

Claims

1. A deep learning-based energy security optimization method, characterized in that: The steps include: S1, collect and pre-process multi-source data from energy systems; S2. Build a graph structure model of the energy system, taking each energy device, region and load point as a node in the graph, and energy flow and device connection as edges in the graph; S3. Apply a physical-aware convolutional neural network to process the graph structure, extract the spatial features of each node, capture the physical dependencies between devices, and generate a global feature vector for each node; S4, input the global feature vector generated by the physical perception convolutional neural network into the adaptive fuzzy inference system to generate the final scheduling strategy according to the real-time energy supply and demand and external environment changes; S5. Based on the difference between the final scheduling strategy and the expected target scheduling strategy, the membership function of the fuzzy inference rule is adaptively adjusted; S6, combining the node feature vector generated by the physical perception convolutional neural network and the scheduling strategy generated by the adaptive fuzzy reasoning to adjust the operating status of the energy equipment in real time; S7. During system operation, by continuously monitoring the status of energy equipment, load demand and external environmental changes, collect real-time feedback data, and further optimize the scheduling strategy of the energy system based on the real-time feedback data.

2. The energy security optimization method based on deep learning according to claim 1, characterized in that: The S1 comprises the following steps: S11. Automatically collect real-time data from all aspects of the energy system through the data acquisition system, including energy production data, power load data, equipment status information, meteorological data and historical event data; S12. Preprocessing the collected raw data, wherein the preprocessing step includes: standardizing the data and filling in missing values ​​in the data.

3. The energy security optimization method based on deep learning according to claim 1, characterized in that: The S2 comprises the following steps: S21. According to the topological structure of the energy system, define each energy device, region and load point as a node in the graph, define energy flow and device connection as edges in the graph, and construct a graph structure model of the energy system; S22. Calculate edge weights based on energy flows between devices ; S23, assigning initial features to each node, the initial features are the status data of each node, including energy production capacity, load demand, and equipment health status, and obtaining the node feature vector by standardizing the node status data , construct the node feature matrix X based on the feature vector of each node; S24. Determine the adjacency relationship between nodes through the graph structure model and construct the adjacency matrix A.

4. The energy security optimization method based on deep learning according to claim 1, characterized in that: The S3 comprises the following steps: S31. Perform convolution processing on the graph structure model, use the convolution layer in the physical perception convolutional neural network to process the input node feature matrix X, and obtain the node feature matrix after the convolution operation ; S32. Based on the adaptive graph convolution mechanism, multiple graph convolution layers are used to extract multi-order adjacency information: ; in, is the output feature matrix of the l-th layer convolution operation, is the normalized adjacency matrix of the adjacency matrix A, is the k-th order weight matrix of the l-th layer, is the bias term of the lth layer, is the output of the previous convolution operation, and K is the order used by the convolution layer; S33, use weighted graph pooling operation to convolve the graph after L layers of processing Pooling is performed to generate the global feature vector H for each node.

5. The energy security optimization method based on deep learning according to claim 1, characterized in that: The S4 comprises the following steps: S41, inputting the node global feature vector H generated by the physical perception convolutional neural network into the adaptive fuzzy inference system, and automatically generating fuzzy inference rules according to the content of the node global feature vector H; S42, based on the node global feature vector and The similarity of the calculation node is the degree of association between nodes; S43. Using the calculated similarity, automatically generate fuzzy sets for each pair of feature vectors and , and derive the corresponding fuzzy reasoning rules based on the similarity; S44, define fuzzy reasoning rules, which are based on real-time energy supply and demand and external environmental factors to describe the scheduling and load distribution logic between energy devices, and are specifically expressed as: ; in, is the feature vector of node i, is the feature vector of node j, for The fuzzy set of for fuzzy set, Y is the output scheduling strategy, is the corresponding fuzzy result; S45, deriving the rule set by using the fuzzy reasoning method, and deriving the fuzzy set by combining the input node feature vector H. The reasoning process obtains the reasoning result of each rule by using the weighted membership function: ; in, is the output corresponding to the kth rule, For input Fuzzy set The membership degree of For fuzzy sets The dynamic weight of the association, The output Y belongs to the fuzzy set The membership degree of For fuzzy sets The dynamic weight of the association, N is the total number of rules; S46: Fusion the inference results of all rules through weighted fusion strategy to generate the final scheduling strategy .

6. The energy security optimization method based on deep learning according to claim 5 is characterized in that: The S46 comprises the following steps: S461, the weighted fusion strategy is dynamically adjusted according to the importance of the rules by introducing adaptive weights, and the final scheduling strategy The calculation method is: ; in, is the adaptive weight of rule k, For input Fuzzy set The membership degree of is the output membership corresponding to rule k, and M is the total number of rule sets.

7. The energy security optimization method based on deep learning according to claim 1, characterized in that: The S5 comprises the following steps: S51. Final scheduling strategy using system output Scheduling strategy with expected goals The difference between ; S52, according to the error value and the membership function of the fuzzy rules, and use the gradient descent optimization algorithm to adjust the parameters in the fuzzy rules; S53, adaptively adjusting the membership function of the fuzzy inference rule according to the feedback data, wherein the adaptive adjustment includes dynamically adjusting the weights of various parameters of the membership function.

8. The energy security optimization method based on deep learning according to claim 7 is characterized in that: The S5 comprises the following steps: S521, the parameter adjustment formula in the fuzzy rule is: ; in, is the membership function The update amount, is the learning rate, is the gradient of the error function with respect to the membership function.

9. The energy security optimization method based on deep learning according to claim 1, characterized in that: The S6 comprises the following steps: S61, the final scheduling strategy generated by combining the global feature vector H of the node generated by the physical perception convolutional neural network and the adaptive fuzzy reasoning , construct the spatiotemporal feature vector of each node ; S62, through the space-time feature vector To model the interdependence between nodes, we construct a global spatiotemporal relationship matrix S. Each element in the matrix S Represents the spatiotemporal dependency between node i and node j; S63. Perform global optimization of the graph structure through the spatiotemporal relationship matrix S. The optimization goal is to minimize the spatiotemporal differences between nodes and obtain the optimized node feature matrix ; S64, based on the optimized node feature matrix and the spatiotemporal relationship matrix S, update the feature representation of each node; S65. Optimize the node features after time and space Combined with the adaptive fuzzy inference system, the operating status of energy equipment can be further adjusted.

10. The energy security optimization method based on deep learning according to claim 1, characterized in that: The S7 comprises the following steps: S71. During system operation, continuously monitor the status of energy equipment, load demand and external environment changes, and collect feedback data in real time; S72, jointly analyzing the collected feedback data and the node feature vector generated by the physical perception convolutional neural network, calculating the relationship between the node feature and the feedback data, and generating a feedback correction factor; S73, dynamically adjusting the node feature vector according to the feedback correction factor; S74. Continue to execute the scheduling decision of energy equipment using the updated node feature vector, and pass it to the adaptive fuzzy reasoning system for the next round of scheduling strategy optimization.

Images