A computer for monitoring the operation of distribution network

Abstract

The present invention relates to the technical field of power grid operation monitoring, and discloses a computer for monitoring the operation of a distribution network. The computer can execute the following methods: processing the virtual computing resource topology map and the monitoring task demand matrix through a virtual network slicing algorithm to obtain a dedicated monitoring network set; processing the dedicated monitoring network set and the operator scheduling strategy through a distributed collaborative learning system to obtain a resource scheduling plan; processing the resource scheduling plan and real-time monitoring task status data through an adaptive task migration system to obtain a task migration execution plan; processing the task migration execution plan and regional monitoring data through a regional adaptive resource control system to obtain a key area resource allocation plan; the method constructs multiple logically isolated but collaborative dedicated monitoring networks on existing computing resources in a software-defined manner to achieve comprehensive monitoring of the distribution network operation status.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid operation monitoring, and more particularly to a computer for monitoring the operation of a power distribution network. Background Art

[0002] With the rapid development of urban agglomerations, the scale and complexity of distribution networks are constantly increasing. Traditional monitoring methods often rely on a centralized architecture within a single operator, making it difficult to quickly mobilize cross-regional computing resources in the event of emergencies or seasonal load fluctuations. As a result, distribution network monitoring is often restricted by administrative divisions and corporate boundaries, leading to a lack of effective collaboration between different operators, low resource utilization, and limited monitoring efficiency. This makes it difficult to meet large-scale, dynamically changing monitoring needs, and lacks sufficient flexibility and cross-regional collaboration capabilities.

[0003] Therefore, how to monitor the ultra-large-scale distribution network jointly managed by multiple grid operators in urban agglomerations has become an urgent problem to be solved. Summary of the Invention

[0004] The present invention provides a computer for monitoring the operation of a power distribution network, which solves the technical problem of monitoring a super-large-scale power distribution network jointly managed by multiple power grid operators in a city cluster in the related art.

[0005] The present invention provides a computer for monitoring the operation of a power distribution network, wherein the computer can execute a method comprising the following steps:

[0006] Processing heterogeneous computing resource status data through a hierarchical resource-aware network to obtain a virtual computing resource topology map;

[0007] Processing the virtual computing resource topology map and the monitoring task requirement matrix through a virtual network slicing algorithm to obtain a dedicated monitoring network set;

[0008] Processing the dedicated monitoring network set and the operator scheduling strategy through a distributed collaborative learning system to obtain a resource scheduling solution;

[0009] Processing the resource scheduling plan and real-time monitoring of task status data through the adaptive task migration system to obtain a task migration execution plan;

[0010] Processing the task migration execution plan and regional monitoring data through the regional adaptive resource control system to obtain a resource allocation plan for key areas;

[0011] The method is used to realize collaborative scheduling of computing resources and dynamic allocation of monitoring tasks in a distribution network environment jointly managed by multiple grid operators, thereby ensuring the continuity and efficiency of the distribution network monitoring function.

[0012] As a further optimization solution of the present invention, the hierarchical resource perception network includes a hierarchical resource feature extraction network, a graph attention network and a topology fusion module, wherein the hierarchical resource feature extraction network includes a node feature layer, a regional aggregation layer and a global association layer, and processes and calculates node status data according to the following mathematical expression;

[0013] Node feature layer:

[0014] Regional aggregation layer:

[0015] Global association layer:

[0016] Among them, X i represents the original resource indicator vector of the i-th computing node, is the weight matrix of each layer, is the bias vector of each layer, A i represents the set of all nodes in the region to which node i belongs, F i They are node-level, region-level and final feature vectors respectively.

[0017] As a further optimization solution of the present invention, the graph attention network includes an attention calculation unit and a feature update unit, which processes the node feature vector set according to the following mathematical expression;

[0018] Attention calculation unit:

[0019]

[0020] Feature update unit:

[0021] Among them, F i ,F j is the node feature vector, W2 is the feature transformation matrix, a is the attention vector, || represents the vector concatenation operation, N i represents the set of neighbor nodes directly connected to node i, α ij is the attention coefficient, H i is the final topological structure representation of node i;

[0022] The topology fusion module combines the topological structure representations of all nodes into a complete virtual computing

[0023] Resource topology: G = (V, E, H);

[0024] Among them, V is the node set, E is the edge set, and H is the node feature matrix.

[0025] As a further optimization solution of the present invention, the virtual network slicing algorithm includes a multi-objective slicing optimizer and an improved Hungarian mapper, wherein the multi-objective slicing optimizer adopts an improved NSGA-III algorithm.

[0026] The optimization objective function is:

[0027] minJ=ω1J isolation +ω2J resource +ω3J balance ;

[0028] The isolation target is:

[0029] Resource efficiency goals:

[0030] Load balancing target:

[0031] Among them, N s is the number of slices, N r is the number of resource nodes, O ij is the resource overlap between slices i and j, U k is the resource usage of node k, C k is the resource capacity of node k, L i is the load level of slice i, is the average load level, ω1, ω2, ω3 are weight coefficients.

[0032] As a further optimization solution of the present invention, the improved Hungarian mapper includes a cost matrix generation unit and an optimal matching calculation unit.

[0033] The cost matrix generation formula is:

[0034]

[0035] Among them, R ij is the computing resource cost, B ij is the bandwidth resource cost, D ij is the delay cost, R max , B max , D max is the maximum value of each type of cost, α, β, γ are weight coefficients;

[0036] The objective function of optimal matching is: minΣ i，j C ij M ij ;

[0037] Among them, M ij is the mapping matrix element, the value is 0 or 1, M ij=1 means mapping virtual node i to physical node j, M ij =0 means no mapping is performed.

[0038] As a further optimization solution of the present invention, the distributed collaborative learning system includes a local policy learner and a global policy aggregator, wherein the local policy learner adopts a two-layer LSTM policy network, and the mathematical expression of the state encoding layer is:

[0039] f t =σ(W f ·[h t-1 , x t ]+b f );

[0040] i t =σ(W i ·[h t-1 , x t ]+b i );

[0041]

[0042] o t =σ(W o ·[h t-1 , x t ]+b o );

[0043] h t =o t ☉tanh(c t );

[0044] The expression of the strategy generation layer is: π t =softmax(W p ·h t +b p );

[0045] Among them, x t is the input state vector at time t, h t is the hidden layer state vector, c t is the unit state vector, f t ,i t , o t are the activation values ​​of the forget gate, input gate, and output gate, respectively, W f , W i , W c , W o , W p is the weight matrix, b f , b i , b c , b o , bp is the bias vector, σ is the sigmid activation function, ⊙ represents the vector element-wise multiplication, π t Assign a policy vector to the resource.

[0046] As a further optimization solution of the present invention, the global policy aggregator adopts a weighted average algorithm with differential privacy protection, which includes a noise addition unit and a security aggregation unit. Its mathematical expression is:

[0047] Noise adding unit:

[0048] Security Aggregation Unit:

[0049] Among them, W k is the local model parameter of the kth operator, Lap(b) is the Laplace noise with scale parameter b, is the model parameter after adding noise, n k is the effective data volume of the kth operator, n is the total data volume of all operators, K is the number of operators, W global It is the global scheduling policy parameter.

[0050] As a further optimization solution of the present invention, the adaptive task migration system includes a multi-layer task evaluation network and a deep reinforcement learning decision maker, wherein:

[0051] The multi-layer task evaluation network consists of a feature extraction layer, a feature fusion layer, and a score generation layer, and its mathematical expression is:

[0052] Feature extraction layer:

[0053] F R =ReLU(W R ·R i +b R );

[0054] F P =ReLU(W P ·P i +b P );

[0055] F T =ReLU(W T ·T i +b T );

[0056] Feature fusion layer:

[0057] F fused =ReLU(W F ·[F R ||F P ||FT ]+b F );

[0058] Score generation layer: S i =sigmoid(W S ·F fused +b S );

[0059] Among them, R i is the resource occupancy vector of task i, P i is the performance indicator vector of task i, T i is the time feature vector of task i, W R , W P , W T , W F , W S is the weight matrix of each layer, b R , b P , b T , b F , b S is the bias vector of each layer, F R , F P , F T is the extracted feature vector, F fused is the fused feature vector, S i Score the health of task i.

[0060] As a further optimization solution of the present invention, the deep reinforcement learning decision maker adopts a dual DQN architecture;

[0061] The state space is defined as s t =[H t , R t , N t ];

[0062] The action space is defined as a t =[src t , dst t ,pri t ];

[0063] The Q value update formula is: Q(s t , a t )=Q(s t , a t )+α[r t +γmax a Q target (s t+1 ,a)-Q(s t , a t )];

[0064] The instant reward is calculated as: rt =w1ΔH+w2ΔR-w3C mig ;

[0065] Among them, H t is the health score vector of all tasks at the current moment, R t is the state vector of all resource nodes at the current moment, N t is the network state vector, src t is the source node number, dst t is the target node number, pri t is the migration priority, ΔH is the health improvement after migration, ΔR is the improvement of resource utilization, C mig is the migration cost, w1, w2, w3 are the reward weights, α is the learning rate, γ is the discount factor, Q target is the Q-value function of the target network.

[0066] As a further optimization solution of the present invention, the regional adaptive resource control system includes a multidimensional feature fusion network and a temperature adaptive resource allocator, wherein the multidimensional feature fusion network adopts a hierarchical structure of the attention mechanism, including a feature extraction unit, an attention calculation unit and a feature fusion unit, and its mathematical expression is:

[0067] Feature extraction unit: f d (X r )=tanh(W d ·X r +b d );

[0068] The attention calculation unit also includes:

[0069] e d =v T tanh(W a ·f d (X r )+b a );

[0070]

[0071] Feature fusion unit:

[0072] Among them, X r is the original monitoring data matrix of region r, W d , W a is the weight matrix, b d , b a is the bias vector, v is the attention vector, e d is the attention score of the d-th dimension, β dis the attention weight of the dth dimension, D is the total number of feature dimensions, I r Score the final importance of region r;

[0073] The temperature adaptive resource allocator includes a temperature parameter regulator and a resource allocation optimizer.

[0074] The expression of the temperature parameter regulator is:

[0075] λ=λ base ·exp(-η·P global );

[0076] The probability allocation expression of the resource allocation optimizer is:

[0077] Among them, λ base is the basic temperature parameter, η is the adjustment coefficient, P global is the global resource pressure indicator, I r is the importance score of region r, λ is the current temperature parameter, P r is the proportion of resources obtained by region r, and R is the total number of regions.

[0078] The beneficial effects of the present invention are as follows: the present invention realizes cross-regional collaboration and efficient resource scheduling by perceiving the status of heterogeneous computing resources in real time and constructing an accurate virtual computing resource topology map. Combined with the virtual network slicing algorithm and the adaptive task migration system, it is possible to dynamically generate a dedicated monitoring network set according to the monitoring task requirements and optimize the resource scheduling scheme, thereby ensuring efficient task allocation and timely response, especially in the event of emergencies or load changes. At the same time, through the regional adaptive resource control system, regional resources are accurately configured to ensure the continuity and stability of monitoring tasks, effectively solving the problem of ultra-large-scale distribution network monitoring that is difficult for multiple power grid operators to jointly manage in traditional methods, so that resources are evenly distributed and monitoring tasks are continuous, greatly improving the efficiency, flexibility and cross-regional collaboration capabilities of distribution network monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 This is a computer flow diagram for monitoring the operation of a power distribution network provided by an embodiment of the present invention;

[0080] Figure 2 is a table of computing resources of participating units provided by an embodiment of the present invention;

[0081] Figure 3 is a monitoring task type and resource requirement table provided by an embodiment of the present invention;

[0082] Figure 4 The embodiment of the present invention provides a regional monitoring importance grading table;

[0083] Figure 5 is a resource feature extraction network configuration table provided by an embodiment of the present invention;

[0084] Figure 6 is a slice optimizer configuration parameter table provided by an embodiment of the present invention;

[0085] Figure 7 is a dedicated network slice configuration table provided by an embodiment of the present invention;

[0086] Figure 8 LSTM strategy network parameter table provided by the embodiment of the present invention;

[0087] Figure 9 is a differential privacy protection parameter table provided by an embodiment of the present invention;

[0088] Figure 10 is a task evaluation network configuration table provided by an embodiment of the present invention;

[0089] Figure 11 is a DQN decision maker parameter table provided by an embodiment of the present invention;

[0090] Figure 12 is a feature fusion network parameter table provided by an embodiment of the present invention;

[0091] Figure 13 is a resource allocator parameter table provided by an embodiment of the present invention;

[0092] Figure 14 is a resource feature extraction accuracy verification table provided by an embodiment of the present invention;

[0093] Figure 15 is a topological structure performance table provided by an embodiment of the present invention;

[0094] Figure 16 This is a slice isolation performance test table provided by an embodiment of the present invention;

[0095] Figure 17 is a resource utilization statistics table provided by an embodiment of the present invention;

[0096] Figure 18 This is a resource scheduling performance test table provided by an embodiment of the present invention;

[0097] Figure 19 This is a task migration effect statistics table provided by an embodiment of the present invention;

[0098] Figure 20 This is a table of key area monitoring effects provided by an embodiment of the present invention;

[0099] Figure 21 This is a table showing the overall system performance improvement effect provided by the embodiments of the present invention. DETAILED DESCRIPTION

[0100] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0101] Implementation Method 1

[0102] This implementation is applicable to monitoring scenarios of ultra-large-scale distribution networks jointly managed by multiple grid operators in urban agglomerations. In this scenario, the following specific technical requirements exist:

[0103] 1. It is necessary to build a logically isolated dedicated monitoring network on the heterogeneous computing resources of multiple grid operators to achieve comprehensive monitoring of the distribution network operation status;

[0104] 2. It is necessary to achieve coordinated scheduling of computing resources among multiple operators and realize efficient sharing of computing power while ensuring resource autonomy;

[0105] 3. It is necessary to support dynamic migration of monitoring tasks to ensure the continuity of monitoring functions when local computing resources are insufficient;

[0106] 4. It is necessary to achieve optimal allocation of global surveillance resources while ensuring the autonomy of each region;

[0107] 5. It is necessary to support the enhancement of elastic monitoring capabilities in key areas and realize dynamic focusing of computing resources.

[0108] like Figure 1 As shown, a computer for monitoring the operation of a power distribution network includes:

[0109] Step 1: Process heterogeneous computing resource status data through a hierarchical resource-aware network to obtain a virtual computing resource topology map;

[0110] Specifically, the hierarchical resource awareness network consists of three core components: a hierarchical resource feature extraction network, a graph attention network, and a topology fusion module. The output of the hierarchical resource feature extraction network serves as the input to the graph attention network, which is then processed by the topology fusion module to generate the final virtual computing resource topology map.

[0111] Sub-step 1.1: Process the computing node status data through the hierarchical resource feature extraction network to obtain a computing resource feature vector set;

[0112] Specifically, the hierarchical resource feature extraction network consists of a three-layer serial structure: the node feature layer: receives the original resource indicator vector (including CPU usage, memory occupancy, storage capacity, network bandwidth and other indicators), and outputs the node-level feature vector; the regional aggregation layer: receives the node-level feature vector, generates the regional-level feature vector by aggregating the node features within the region; the global association layer: receives the regional-level feature vector, establishes the association between regions and outputs the final feature vector.

[0113] The mathematical expressions of each layer are:

[0114] 1. Node feature layer:

[0115] 2. Regional aggregation layer:

[0116] 3. Global association layer:

[0117] Among them, X i ∈R d represents the d-dimensional raw resource indicator vector of the i-th computing node, are the weight matrices of each layer, are the bias vectors of each layer, A i represents the set of all nodes in the region to which node i belongs, F i They are node-level, region-level and final feature vectors respectively.

[0118] Sub-step 1.2: Process the node feature vector set and node connection relationship through the graph attention network to obtain the resource topology structure representation;

[0119] Specifically, the graph attention network consists of an attention calculation unit and a feature update unit. The attention calculation unit first calculates the attention coefficient between nodes, and then the feature update unit updates the node features based on the attention coefficient.

[0120] Attention calculation unit:

[0121]

[0122] The expression of the feature update unit is:

[0123] Among them, F i , F j is the node feature vector output by sub-step 1.1, W2 is the feature transformation matrix, a is the attention vector, ‖ represents the vector concatenation operation, N i represents the set of neighbor nodes directly connected to node i, α ij is the attention coefficient, which represents the influence weight of node j on node i, Hi is the final topological structure representation of node i. The topological fusion module represents the topological structure of all nodes {H i} are combined into a complete virtual computing resource topology graph G = (V, E, H), where V is the node set, E is the edge set, and H is the node feature matrix.

[0124] Step 2: Process the virtual computing resource topology map and the monitoring task requirement matrix through the virtual network slicing algorithm to obtain a dedicated monitoring network set;

[0125] Specifically, the virtual network slicing algorithm consists of two main modules: a slice optimizer and a mapping executor. The slice optimizer is responsible for generating a slicing scheme that meets isolation and resource efficiency requirements, while the mapping executor is responsible for mapping the slicing scheme to physical resources. The output of the slice optimizer serves as the input of the mapping executor.

[0126] Sub-step 2.1: Process the virtual computing resource topology map and monitoring task requirement matrix in step 1 through the multi-objective slicing optimizer to obtain a set of network slicing solutions.

[0127] Specifically, the multi-objective slice optimizer uses the improved NSGA-III algorithm and includes three evaluation modules for optimization objectives:

[0128] 1. Isolation evaluation module: calculates the degree of resource isolation between slices;

[0129] 2. Resource efficiency evaluation module: calculates resource utilization;

[0130] 3. Load balancing evaluation module: evaluates the load distribution between slices;

[0131] The optimization objective function is:

[0132] minJ=ω1J isolation +ω2J resource +ω3J balance ;

[0133] The isolation target is:

[0134] Resource efficiency goals:

[0135] Load balancing target:

[0136] Among them, N s is the number of slices, N r is the number of resource nodes, O ij is the resource overlap between slices i and j, U k is the resource usage of node k, C k is the resource capacity of node k, Li is the load level of slice i, is the average load level, ω1, ω2, ω3 are weight coefficients.

[0137] Sub-step 2.2: Obtain a set of dedicated monitoring network instances by processing the set of network slicing schemes through the improved Hungarian mapper.

[0138] The improved Hungarian mapper includes a cost matrix generation unit and an optimal matching calculation unit. The cost matrix generation unit calculates the mapping cost from virtual nodes to physical nodes, and the optimal matching calculation unit solves the optimal mapping relationship based on the cost matrix.

[0139] Cost matrix generation formula:

[0140] Among them, R ij is the computing resource cost, B ij is the bandwidth resource cost, D ij is the delay cost, R max , B max , D max is the maximum value of each cost, α, β, γ are weight coefficients.

[0141] Objective function for optimal matching: min∑ i，j C ij M ij

[0142] Constraints: (Each virtual node can only be mapped to one physical node); (Each physical node accepts the mapping of at most one virtual node).

[0143] Among them, M ij is the mapping matrix element, the value is 0 or 1, M ij =1 means mapping virtual node i to physical node j, M ij =0 means no mapping is performed.

[0144] Step 3: Process the dedicated monitoring network instance set and operator scheduling strategy in step 2 through the distributed collaborative learning system to obtain a resource scheduling solution;

[0145] The distributed collaborative learning system consists of a local policy learner and a global policy aggregator. The local policy learner runs within each operator and is responsible for learning the locally optimal scheduling policy. The global policy aggregator runs at the coordination center and is responsible for integrating the policies of each operator and generating a global scheduling solution. The output of the local policy learner is transmitted to the global policy aggregator via a secure channel.

[0146] Sub-step 3.1: Process the local resource state sequence and monitoring task sequence through a two-layer LSTM policy network to obtain the operator's local scheduling policy set;

[0147] The two-layer LSTM policy network consists of a state encoding layer and a policy generation layer:

[0148] 1. State encoding layer: encodes the temporal state information into a hidden layer representation;

[0149] 2. Strategy generation layer: Generates resource allocation strategies based on the hidden layer representation;

[0150] The mathematical expression of the state encoding layer is:

[0151] f t =σ(W f ·[h t-1 , x t ]+b f )

[0152] i t =σ(W i ·[h t-1 , x t ]+b i )

[0153]

[0154] o t =σ(W o ·[h t-1 , x t ]+b o )

[0155] h t =o t ☉tanh(c t )

[0156] Among them, x t is the input state vector at time t, including features such as resource utilization and task queue length, h t is the hidden layer state vector, c t is the unit state vector, f t ,i t , o t are the activation values ​​of the forget gate, input gate, and output gate, respectively, W f , W i , W c , W o is the weight matrix, b f , b i , b c , b ois the bias vector, b is the sigmid activation function, and ⊙ represents vector element-wise multiplication.

[0157] The expression of the strategy generation layer: π t =softmax(W p ·h t +b p )

[0158] Among them, W p Generate a weight matrix for the policy, b p Generate a bias vector for the policy, π t is the resource allocation strategy vector, which represents the probability distribution of allocating tasks to each resource node.

[0159] Sub-step 3.2: Process the operator's local scheduling policy set through the security aggregator to obtain a global resource scheduling solution;

[0160] The noise adding unit adds Laplace noise to the local model parameters:

[0161]

[0162] Among them, W k is the local model parameter of the kth operator, Lap(b) is the Laplace noise with scale parameter b, are the model parameters after adding noise.

[0163] The security aggregation unit performs a weighted average:

[0164] Among them, n k is the effective data volume of the kth operator, n is the total data volume of all operators, K is the number of operators, W global It is the global scheduling policy parameter.

[0165] Step 4: The adaptive task migration system processes the resource scheduling plan in step 3 and monitors the task status data in real time to obtain the task migration execution plan.

[0166] The adaptive task migration system consists of two core components: a task status evaluator and a migration decision generator. The task status evaluator is responsible for real-time assessment and monitoring of task health. The migration decision generator generates specific migration plans based on the assessment results and resource status. The output of the task status evaluator serves as an important input to the migration decision generator.

[0167] Sub-step 4.1: Process the runtime data of the monitoring task through the multi-layer task evaluation network to obtain the task health score set;

[0168] The multi-layer task evaluation network consists of a feature extraction layer, a feature fusion layer, and a score generation layer:

[0169] 1. Feature extraction layer: processes resource usage, performance indicators, and time characteristics respectively;

[0170] 2. Feature fusion layer: fuses multi-dimensional features;

[0171] 3. Score generation layer: calculates the final health score;

[0172] Mathematical expression of the feature extraction layer:

[0173] F R =ReLU(W R ·R i +b R )

[0174] F P =ReLU(W P ·P i +b P )

[0175] F T =ReLU(W T ·T i +b T )

[0176] Among them, R i is the resource occupancy vector of task i, including CPU, memory, bandwidth and other indicators, P i is the performance indicator vector of task i, including response time, throughput and other indicators, T i is the time feature vector of task i, including indicators such as running time and scheduling delay, W R , W P , W T is the weight matrix of each feature extraction layer, b R , b P , b T is the bias vector of each feature extraction layer, F R , F P , F T is the extracted feature vector.

[0177] The expression of the feature fusion layer:

[0178] F fused =ReLU(W F ·[F R ||F P ||F T ]+b F )

[0179] Where: ‖ represents the vector concatenation operation, W F is the feature fusion weight matrix, bF is the feature fusion bias vector, F fused is the fused feature vector.

[0180] The expression of the score generation layer:

[0181] S i =sigmoid(W S ·F fused +b S )

[0182] Among them, W S Generate weight matrix for scoring, b S Generate bias vector for scoring, S i is the health score of task i, ranging from [0,1].

[0183] Sub-step 4.2: Process the task health score set and available resource status data through the deep reinforcement learning decision maker to obtain the task migration execution plan;

[0184] The deep reinforcement learning decision maker uses a dual DQN architecture, consisting of an online network and a target network. The online network is responsible for generating the current decision, while the target network is used to stabilize the training process.

[0185] State space definition: s t =[H t , R t , N t ]

[0186] Among them, H t is the health score vector of all tasks at the current moment, R t is the state vector of all resource nodes at the current moment, N t It is the network state vector, including indicators such as bandwidth and delay.

[0187] Action space definition: a t =[src t , dst t ,pri t ]

[0188] Among them, src t is the source node number, dst t is the target node number, pri t Migration priority.

[0189] The Q value update formula is: Q(s t , a t )=Q(s t , a t )+α[r t +γmax a Qtarget (s t+1 ,a)-Q(s t ,a t )]

[0190] Instant reward calculation: r t =w1ΔH+w2ΔR-w3C mig

[0191] Among them, H t is the health score vector of all tasks at the current moment, R t is the state vector of all resource nodes at the current moment, N t is the network state vector, src t is the source node number, dst t is the target node number, pri t is the migration priority, ΔH is the health improvement after migration, ΔR is the improvement of resource utilization, C mig is the migration cost, w1, w2, w3 are the reward weights, α is the learning rate, γ is the discount factor, Q target is the Q-value function of the target network.

[0192] Step 5: The regional adaptive resource control system processes the task migration execution plan and regional monitoring data from step 4 to obtain the resource allocation plan for key areas;

[0193] The regional adaptive resource control system consists of two main modules: a regional importance evaluator and a dynamic resource allocator. The regional importance evaluator is responsible for real-time assessment of the monitoring demand intensity of each region, while the dynamic resource allocator optimizes resource allocation based on the assessment results. The output of the regional importance evaluator directly determines the resource allocation strategy of the dynamic resource allocator.

[0194] Sub-step 5.1: Process the region monitoring data stream through a multi-dimensional feature fusion network to obtain a set of region importance scores;

[0195] The multi-dimensional feature fusion network adopts a hierarchical structure of the attention mechanism, which includes a feature extraction unit, an attention calculation unit, and a feature fusion unit:

[0196] 1. Feature extraction unit: processes the raw data of each dimension separately;

[0197] 2. Attention calculation unit: calculates the importance weight of each dimension feature;

[0198] 3. Feature fusion unit: weighted fusion of multi-dimensional features;

[0199] Mathematical expression of feature extraction unit:

[0200] f d (Xr )=tanh(W d ·X r +b d )

[0201] Among them, X r is the original monitoring data matrix of region r, W d is the feature extraction weight matrix of the dth dimension, b d Extract the bias vector for the feature of the dth dimension, f d (X r ) is the feature vector extracted from the dth dimension.

[0202] The expression of the attention calculation unit also includes:

[0203] e d =v T tanh(W a ·f d (X r )+b a )

[0204]

[0205] Among them, W a Calculate the weight matrix for attention, b a Calculate the bias vector for attention, v is the attention vector, e d is the attention score of the d-th dimension, β d is the attention weight of the d-th dimension, and D is the total number of feature dimensions.

[0206] The expression of feature fusion unit:

[0207]

[0208] Among them, I r Score the final importance of region r.

[0209] Sub-step 5.2: Process the regional importance score set and current resource distribution data through the temperature adaptive resource allocator to obtain a resource reallocation plan.

[0210] The temperature-adaptive resource allocator consists of a temperature parameter regulator and a resource allocation optimizer. The temperature parameter regulator dynamically adjusts the allocation concentration according to the global resource pressure, and the resource allocation optimizer generates the final allocation plan based on the adjusted temperature parameters.

[0211] The expression of the temperature parameter regulator:

[0212] λ=λ base ·exp(-η·P global )

[0213] Among them, λ base is the basic temperature parameter, η is the adjustment coefficient, P global It is the global resource pressure indicator.

[0214] The probability allocation expression of the resource allocation optimizer is:

[0215]

[0216] Among them, I r is the importance score of region r, λ is the current temperature parameter, P r is the proportion of resources obtained by region r, and R is the total number of regions.

[0217] Resource allocation optimization goals:

[0218]

[0219] Constraints:

[0220] 1. (total resource constraints);

[0221] 2. (minimum resource guarantee);

[0222] 3. (Adjustment range constraints);

[0223] Among them, P0 is the initial resource distribution vector, D KL is the KL divergence, used to balance the resource adjustment range, η is the balance factor, P min To ensure the minimum resource ratio, is the resource allocation ratio at the previous moment, Δ max The maximum adjustment range is limited.

[0224] Technical effects of this embodiment

[0225] This embodiment achieves the following specific technical effects through the synergistic effect of the above five steps:

[0226] 1. Through the combination of a hierarchical resource-aware network and a graph attention network, unified modeling and representation of heterogeneous computing resources is achieved, specifically:

[0227] The accuracy of resource feature extraction has increased from 85% of traditional methods to over 95%;

[0228] The data compression rate of the topological structure representation reaches 85%, while the information loss rate is controlled below 5%;

[0229] The resource status update delay is reduced from the original 500ms to less than 100ms.

[0230] 2. Based on the improved NSGA-III algorithm and Hungarian mapping algorithm, efficient virtual network slicing is achieved, achieving:

[0231] The resource isolation between slices has been increased from 95% to 99.9%;

[0232] Resource utilization increased from 45% to 85%, an increase of more than 40%;

[0233] Slice build time was reduced from an average of 4 minutes to 90 seconds, a 60% reduction;

[0234] Under the same hardware conditions, the number of supported parallel dedicated monitoring networks increases from 8 to 24.

[0235] 3. A two-layer LSTM strategy network and a federated learning framework with differential privacy protection enable secure and efficient resource collaborative scheduling.

[0236] The efficiency of computing resource sharing between operators increased from 50% to 80%;

[0237] Scheduling decision delay reduced from 2 seconds to less than 1 second;

[0238] Under the condition of differential privacy protection strength ε = 0.1, the accuracy loss of the scheduling model is controlled within 3%;

[0239] The convergence time of the global scheduling solution is reduced from 20 rounds to 6 rounds.

[0240] 4. Accurate task migration decisions are achieved through a multi-layer task evaluation network and dual DQN architecture:

[0241] The accuracy of task health assessment increased from 80% to 92%;

[0242] The task migration success rate increased from 95% to 99.99%;

[0243] Service interruption time during the migration process was reduced from 500ms to 100ms;

[0244] Reduced mission full recovery time from 4 seconds to 1 second.

[0245] 5. Based on the multi-dimensional feature fusion network and temperature-adaptive resource allocator, intelligent resource regulation in key areas is achieved:

[0246] The monitoring accuracy of key areas has been improved by 50% from the original baseline;

[0247] The response time of dynamic resource adjustment is reduced from 1 second to less than 200ms;

[0248] The emergency event handling capacity has been increased from 100 events per second to 300 events per second;

[0249] The imbalance in resource distribution among regions was reduced from 35% to 10%.

[0250] The comprehensive technical effects of this implementation significantly enhance the overall monitoring capabilities of the distribution network in urban agglomerations. The overall availability of the monitoring system has increased from 99.9% to 99.999%, the end-to-end delay of monitoring data has been reduced from 5 seconds to less than 1 second, the average utilization of computing resources has increased from 40% to 85%, the system operation and maintenance costs have been reduced by 45%, and the annual cost savings have exceeded 2 million yuan. The average response time to distribution network failures has been shortened from 10 minutes to 3 minutes.

[0251] Real application example of implementation method 1

[0252] like Figures 2 to 13 As shown in Figure 2, this implementation has been put to practical use in the distribution network operation monitoring system of a city cluster in the Yangtze River Delta region. This city cluster includes three provincial power grid companies and 15 prefecture-level power grid companies, managing over 1,000 substations and 50,000 kilometers of distribution lines. Before the system was deployed, the computing resources of the various operators were dispersed, and monitoring capabilities were uneven, resulting in inconsistent monitoring results.

[0253] like Figures 14 to 21 As shown, this implementation method was actually deployed and verified in the distribution network operation monitoring system of a city cluster in the Yangtze River Delta region for 6 months, and a large amount of measured data was collected.

[0254] Verification data indicates that this implementation method has achieved significant technical results in practical applications, with all indicators meeting or exceeding expectations. In particular, significant improvements have been achieved in key indicators such as system availability, resource utilization efficiency, and fault response time, providing a strong guarantee for the safe and stable operation of the urban agglomeration's distribution network.

[0255] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make many forms based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. A computer for monitoring the operation of a power distribution network, characterized in that: The computer-implemented method comprises the following steps: Processing heterogeneous computing resource status data through a hierarchical resource-aware network to obtain a virtual computing resource topology map; Processing the virtual computing resource topology map and the monitoring task requirement matrix through a virtual network slicing algorithm to obtain a dedicated monitoring network set; Processing the dedicated monitoring network set and the operator scheduling strategy through a distributed collaborative learning system to obtain a resource scheduling solution; Processing the resource scheduling plan and real-time monitoring of task status data through the adaptive task migration system to obtain a task migration execution plan; Processing the task migration execution plan and regional monitoring data through the regional adaptive resource control system to obtain a resource allocation plan for key areas; The hierarchical resource perception network includes a hierarchical resource feature extraction network, a graph attention network, and a topology fusion module. The hierarchical resource feature extraction network includes a node feature layer, a regional aggregation layer, and a global association layer, and processes and calculates node status data according to the following mathematical expression: Node feature layer: ; Regional aggregation layer: ; Global association layer: ; in, represents the original resource indicator vector of the i-th computing node, is the weight matrix of each layer, is the bias vector of each layer, represents the set of all nodes in the region to which node i belongs, They are node-level, region-level and final feature vectors respectively; The graph attention network includes an attention calculation unit and a feature update unit, which processes the node feature vector set according to the following mathematical expression: Attention calculation unit: ； Feature update unit: ; in, is the node feature vector, is the feature transformation matrix, is the attention vector, Represents vector concatenation operation, represents the set of neighbor nodes directly connected to node i, is the attention coefficient, is the final topological structure representation of node i; The topology fusion module combines the topological structure representations of all nodes into a complete virtual calculation; Resource topology diagram: ; in, is a node set, is the edge set, is the node feature matrix.

2. A computer for monitoring the operation of a power distribution network according to claim 1, characterized in that: The virtual network slicing algorithm includes a multi-objective slicing optimizer and an improved Hungarian mapper. The multi-objective slicing optimizer adopts an improved NSGA-III algorithm, and the optimization objective function is: ； The isolation target is: ; Resource efficiency goals: ; Load balancing target: ; in, is the number of slices, is the number of resource nodes, is the resource overlap between slices i and j, is the resource usage of node k, is the resource capacity of node k, is the load level of slice i, is the average load level, is the weight coefficient.

3. A computer for monitoring the operation of a power distribution network according to claim 2, characterized in that: The improved Hungarian mapper includes a cost matrix generation unit and an optimal matching calculation unit. The cost matrix generation formula is: ； in, To calculate resource costs, is the bandwidth resource cost, is the delay cost, is the maximum value of each cost, is the weight coefficient; The objective function of optimal matching is: ; in, is the mapping matrix element, which takes the value of 0 or 1. Indicates mapping virtual node i to physical node j, Indicates that no mapping is performed.

4. A computer for monitoring the operation of a power distribution network according to claim 1, characterized in that: The distributed collaborative learning system includes a local policy learner and a global policy aggregator, wherein the local policy learner adopts a two-layer LSTM policy network, and the mathematical expression of the state encoding layer is: ； ； ； ； ； ； The expression of the strategy generation layer is: ; in, is the input state vector at time t, is the hidden layer state vector, is the cell state vector, are the activation values ​​of the forget gate, input gate, and output gate, respectively. is the weight matrix, is the bias vector, is the sigmid activation function, represents vector element-wise multiplication, Assign a policy vector to the resource.

5. A computer for monitoring the operation of a power distribution network according to claim 4, characterized in that: The global policy aggregator adopts a weighted average algorithm with differential privacy protection, which includes a noise addition unit and a security aggregation unit. Its mathematical expression is: Noise adding unit: ; Security Aggregation Unit: ; in, are the local model parameters of the k-th operator, is the Laplace noise with scale parameter b, are the model parameters after adding noise, is the effective data volume of the kth operator, is the total data volume of all operators, is the number of operators, It is the global scheduling policy parameter.

6. A computer for monitoring the operation of a power distribution network according to claim 1, characterized in that: The adaptive task migration system includes a multi-layer task evaluation network and a deep reinforcement learning decision maker, wherein the multi-layer task evaluation network consists of a feature extraction layer, a feature fusion layer, and a score generation layer, and its mathematical expression is: Feature extraction layer: ； ； ； Feature fusion layer: ； Rating generation layer: ; in, is the resource occupancy vector of task i, is the performance indicator vector of task i, is the time feature vector of task i, is the weight matrix of each layer, is the bias vector of each layer, is the extracted feature vector, is the fused feature vector, Score the health of task i.

7. A computer for monitoring the operation of a power distribution network according to claim 6, characterized in that: The deep reinforcement learning decision maker adopts a dual DQN architecture; The state space is defined as: ; The action space is defined as: ; The Q value update formula is: ; The instant reward is calculated as: ; in, is the health score vector of all tasks at the current moment, is the state vector of all resource nodes at the current moment, is the network state vector, is the source node number, is the target node number, For migration priority, To improve health after migration, The degree of improvement in resource utilization, For migration overhead, is the reward weight, is the learning rate, is the discount factor, is the Q-value function of the target network.

8. A computer for monitoring the operation of a power distribution network according to claim 1, characterized in that: The regional adaptive resource control system includes a multidimensional feature fusion network and a temperature adaptive resource allocator. The multidimensional feature fusion network adopts a hierarchical structure of the attention mechanism, including a feature extraction unit, an attention calculation unit and a feature fusion unit. Its mathematical expression is: Feature extraction unit: ; The attention calculation unit also includes: ； ； Feature fusion unit: ; in, is the original monitoring data matrix of region r, is the weight matrix, is the bias vector, is the attention vector, is the attention score of the d-th dimension, is the attention weight of the d-th dimension, is the total number of feature dimensions, Score the final importance of region r; The temperature adaptive resource allocator includes a temperature parameter regulator and a resource allocation optimizer. The expression of the temperature parameter regulator is: ； The probability allocation expression of the resource allocation optimizer is: ; in, is the basic temperature parameter, is the adjustment coefficient, is the global resource pressure indicator, Score the importance of region r, is the current temperature parameter, is the proportion of resources obtained by region r, The total number of regions.

Images