Electric power communication multi-route intelligent planning method and device based on scene classification

Abstract

The invention discloses an electric power communication multi-route intelligent planning method and device based on scene classification. The multi-route intelligent planning method in the electric power field comprises the steps that S1, multi-source heterogeneous data fusion processing is carried out; s2, classifying three-dimensional scene features; s3, dynamic routing modeling is carried out; s4, a hybrid intelligent optimization algorithm; and S5, real-time verification and feedback optimization are carried out. According to the multi-route intelligent planning method in the electric power field, the problems of static route planning, poor adaptability and insufficient multi-source data fusion capability in a traditional electric power communication network are solved, intelligent perception and dynamic adaptation to a complex communication environment are realized, and the multi-source data fusion capability is improved. And the network resource scheduling efficiency, the flexibility of the routing strategy and the overall service quality are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric power, in particular to a power communication multi-route intelligent planning method and device based on scene classification. BACKGROUND

[0002] In the power communication network, route planning is a key link to ensure the safe and stable operation of the power system and efficient information transmission. With the rapid development of smart grids, the types of services carried by the power communication network are becoming increasingly complex, including daily operation and maintenance, emergency support, fault handling and other scenarios, which puts higher requirements on the flexibility, real-time performance and reliability of the network. However, the traditional power communication route planning method mostly uses static or semi-static models, relies on manual experience to set rules, and is difficult to adapt to dynamic changes in business requirements and network environment. Especially in the face of sudden traffic, device failure or environmental disturbance, the traditional mechanism lacks effective adaptive adjustment capability, which easily leads to unreasonable resource allocation, degradation of service quality and other problems.

[0003] In addition, in terms of data processing, existing systems generally use static weighting or simple splicing to integrate multi-source heterogeneous data from device status, business requirements and environmental monitoring, which is difficult to accurately reflect the dynamic change characteristics of each dimension data and the importance difference in different scenarios, thereby affecting the accuracy of network state perception. At the same time, traditional optimization algorithms such as particle swarm optimization (PSO) or genetic algorithm (GA) have the problems of slow convergence speed or easy to fall into local optimum, which limits their application effect in high-dimensional and multi-objective route optimization. Overall, the existing technology still has obvious limitations in network perception, scene identification, route modeling and optimization solving, and is difficult to meet the development needs of intelligent and fine management of modern power communication networks. SUMMARY

[0004] The purpose of the present application is to overcome the shortcomings of the prior art and provide a power communication multi-route intelligent planning method and device based on scene classification, which not only greatly improves the adaptability of the network to complex and variable environment and the intelligent level of route decision-making, but also realizes the efficient fusion of multi-source heterogeneous data and the closed-loop optimization of dynamic strategy, and solves the problems of static, rigid and insufficient real-time scene response capability in traditional route planning.

[0005] The present application is realized by the following technical solutions:

[0006] A power communication multi-route intelligent planning method based on scene classification, comprising the following steps:

[0007] S1: Collect power communication scene data, establish a dynamic weighted fusion model, generate a fusion feature vector, introduce a spatiotemporal correlation feature extraction mechanism, and obtain spatiotemporal feature parameters;

[0008] S2: The fused feature vector and the spatio-temporal feature parameter are spliced and input into a three-layer fully connected neural network for feature mapping. The mapped features are used for class score calculation by a fully connected layer, and the predicted class probability distribution of the current communication scenario and the predicted class of the current communication scenario are output;

[0009] S3: The classification information of the current scenario is obtained from the result of S2, a multi-objective routing optimization function is defined, the dynamic weights of each optimization objective are dynamically adjusted, the dynamic weights are substituted into the multi-objective routing optimization function, all candidate paths are scored and sorted, and the optimal path is selected as the current routing strategy;

[0010] S4: The particle swarm optimization algorithm and the genetic algorithm are mixed for optimization, and the adaptive inertia weight mechanism is combined to optimize the current routing strategy, and the optimized routing strategy is output;

[0011] S5: The optimized routing strategy is deployed to the digital twin verification platform, the verification result is quantified by the verification evaluation function, the routing modeling parameters are adjusted reversely based on the verification result, the routing model is updated, a new optimization strategy is generated, and the optimization strategy is iterated until convergence, and the final decision result is output.

[0012] S1: The power communication scenario data is collected, a dynamic weighted fusion model is established, a fused feature vector is generated, and a spatio-temporal correlation feature extraction mechanism is introduced to obtain a spatio-temporal feature parameter. The specific content is as follows:

[0013] S1.1: Data collection and preprocessing;

[0014] Collecting equipment state data D s , business demand data D d , and environmental monitoring data D e . All data are standardized and time-aligned;

[0015] S1.2: Establishing a dynamic weighted fusion model;

[0016] The fusion expression is constructed as follows:

[0017] F=W⊙(αD s +βD d +γD e )

[0018] Wherein, F is the fused feature vector, α, β, γ are the initial weight coefficients of each data source, W is a dynamic weight matrix, and represents Hadamard product;

[0019] S1.3: Introducing a spatio-temporal correlation feature extraction mechanism;

[0020] Constructing spatio-temporal correlation features:

[0021]

[0022] Where T(t) is the spatiotemporal characteristic parameter, Φ(τ) represents the characteristic value of the communication traffic at time τ; λ is the attenuation coefficient, which is used to emphasize the impact of recent data on the current state; Δt represents the length of the time window, which controls the retrospective range of historical information.

[0023] S1.4: Output the fused feature vector F and spatiotemporal feature parameters T(t).

[0024] As described in S2, the fusion feature vector and the spatiotemporal feature parameters are spliced ​​and input into a three-layer fully connected neural network for feature mapping. The mapped features are subjected to category score calculation using a fully connected layer, and the predicted category probability distribution and the predicted category of the current communication scenario are output. The specific contents are as follows:

[0025] S2.1: Input feature vector preparation;

[0026] The output fusion feature vector F and spatiotemporal correlation feature parameter T(t) are used as input and concatenated to form a complete feature representation:

[0027] [F; T(t)]

[0028] Where F is the fused feature vector after fusion of multi-source heterogeneous data; “;” represents the vector concatenation operation, which is used to integrate spatial and temporal information;

[0029] S2.2: Construct feature projection space;

[0030] The concatenated features are input into a three-layer fully connected neural network, and a nonlinear activation function is added for feature mapping:

[0031] Z=ReLU(W z ·[F;T(t)]+b z )

[0032] Among them, W z is the trainable weight matrix; b z is the bias term; ReLU is the activation function used to introduce nonlinear characteristics; Z is the high-dimensional feature representation of the output;

[0033] S2.3: Design a deep reinforcement classification network;

[0034] Based on the feature projection, a fully connected layer with Sigmoid activation is used to calculate the category score, and the Softmax function is combined to output the final classification probability distribution:

[0035] P(c k |F)=softmax(σ(W c• F + b c ))

[0036] where c k ∈ {emergency guarantee, daily operation and maintenance, fault handling}: target scenario category; W c , b c are trainable parameters of the classification layer; σ is a Sigmoid activation function for introducing a nonlinear mapping; softmax is used to convert the output into a probability distribution of each scenario category;

[0037] Through end-to-end training, the parameters are continuously optimized, and finally the predicted category c k of the current communication scenario is output.

[0038] S3 obtains the classification information of the current scenario from the results of S2, defines a multi-objective routing optimization function, dynamically adjusts the dynamic weights of each optimization objective, substitutes the dynamic weights into the multi-objective routing optimization function, scores and sorts all candidate paths, and selects the optimal path as the current routing strategy. The specific content is as follows:

[0039] S3.1: Obtain the classification information of the current scenario;

[0040] Obtain the predicted category c k of the current communication scenario, and extract the corresponding scene sensitivity parameter S j (t);

[0041] S3.2: Construct a multi-objective routing optimization function;

[0042] The multi-objective routing optimization function is defined as follows:

[0043]

[0044] where C i is the cost of the ith candidate path; D i is the end-to-end transmission delay of the ith path; R i is the reliability index of the ith path; indicates that maximizing reliability is converted into a minimization problem; ω1, ω2, ω3 are the dynamic weights of each optimization objective, and n represents the total number of candidate paths;

[0045] S3.3: Implement a dynamic weight adjustment mechanism;

[0046] According to the characteristics of the current scenario, a weighted strategy in the form of softmax is used to dynamically adjust the importance of each objective:

[0047]

[0048] where m is an index variable for traversing all optimization objectives; km is the adjustment coefficient related to the mth optimization objective; S m (t) represents the scenario sensitivity parameter of the mth optimization objective at the current time t; S j (t) is the scenario sensitivity parameter of the jth objective at the current time t; k j is the adjustment coefficient, which controls the response sensitivity of the objective to the scenario change, and the denominator is the normalization term, which ensures that the sum of all weights is 1;

[0049] The dynamic weight is substituted into the multi-objective routing optimization function, the shortest path algorithm is used to score and sort all candidate paths, the path with the best comprehensive performance is selected as the final routing strategy, and the optimal routing path and corresponding strategy parameters that adapt to the current communication scenario are output.

[0050] S4 uses the particle swarm optimization algorithm and genetic algorithm to optimize the current routing strategy, and outputs the optimized routing strategy, as follows:

[0051] S4.1: initialization of population and parameter setting;

[0052] First, a group of initial solutions are randomly generated as population individuals according to the dimension of the optimization problem, and each individual is represented as a vector x i , the following parameters are set: particle number N; maximum iteration number iter max ; learning factors c1, c2; inertia weight range [ω min ,ω max ]; crossover factor range [0.3, 0.7]; Gaussian disturbance variance σ 2 ;

[0053] S4.2: introduction of adaptive inertia weight mechanism;

[0054] In each iteration process, the inertia weight ω is dynamically adjusted in a linear decreasing manner:

[0055]

[0056] where ω max is the initial larger inertia weight; ω min is the smaller inertia weight in the later period; iter is the current iteration number;

[0057] S4.3: execute particle swarm speed and position update

[0058] The speed and position of each particle are updated according to the standard particle swarm optimization algorithm formula:

[0059]

[0060] where, and are the current velocity and position of the ith particle in the dth dimension, respectively; pbest id is the individual historical optimal solution; gbest d is the global optimal solution; r1, r2 are random numbers in the interval [0, 1], and k represents the iteration number;

[0061] S4.4: Introducing genetic algorithm

[0062] After every several iterations, a part of excellent individuals in the current population are selected for genetic operation, including crossover and mutation, to jump out of the local optimal trap;

[0063] According to the multi-objective routing optimization function, the fitness of all particles is evaluated, and the individual optimal solution pbest and the global optimal solution gbest are updated;

[0064] If the maximum iteration number is reached or the preset convergence condition is met, the algorithm is terminated, and the current optimal solution gbest, i.e., the final optimized routing strategy or resource configuration scheme, is output.

[0065] The crossover operation is as follows: two parent individuals x p and x q are randomly selected, and new individuals are generated according to the following formula:

[0066] x new = αx p + (1-α)x q + N(0, σ 2 )

[0067] Wherein: α ∈ [0.3, 0.7] is a dynamic crossover factor, which controls the fusion proportion of parent information; N(0, σ 2 ) is a Gaussian noise disturbance term;

[0068] The mutation operation is as follows: a small disturbance is applied to part of the individuals to further expand the search space;

[0069] The specific operation is as follows:

[0070] Individual selection: in the current population, a part of individuals are randomly selected for mutation operation;

[0071] Disturbance generation: a small disturbance term obeying Gaussian distribution is applied to each dimension variable of the selected individual, and the disturbance amplitude is adjusted according to the search space of the actual problem, and the expression is as follows:

[0072]

[0073] Wherein, x i is the position vector of the original individual, For the disturbance amplitude, N(0, 1) represents a standard normal distribution random number;

[0074] Boundary constraint processing: in order to avoid the new individual after mutation exceeding the variable definition domain range, boundary check and clipping are carried out on the mutation result, so that it is still in the feasible solution space.

[0075] Population update: the individual after mutation is reinserted into the population to participate in the next round of fitness evaluation and evolution process.

[0076] Through the mutation operation, the algorithm can introduce random disturbance on the basis of local search, effectively avoid falling into local optimal trap, thereby improving the robustness and adaptability of power communication multi-route intelligent planning method in complex scene.

[0077] S5. The optimized routing strategy is deployed to the digital twin verification platform, the verification result is quantified by the verification evaluation function, the routing modeling parameters are adjusted reversely based on the verification result, the routing model is updated, and a new optimization strategy is generated, the optimization strategy is iterated until convergence, and the final decision result is output, as follows:

[0078] S5.1: Establish a digital twin verification platform

[0079] Based on the current network topology structure, device state, service flow characteristics and the routing strategy output by S4, a virtual simulation environment is constructed;

[0080] S5.2: Define verification evaluation function

[0081] The routing strategy is executed in the digital twin environment and the performance indicators are collected, and the following evaluation function is constructed to quantify the strategy quality:

[0082]

[0083] Wherein, and are the transmission delays of the ith service flow in simulation and actuality respectively; is the energy efficiency value of the path in simulation; E th is the set energy efficiency threshold; N is the number of service flows participating in evaluation;

[0084] The routing strategy output by S4 is deployed to the digital twin platform to simulate its running process in the network, record various performance indicators and compare them with the actual network running data, and calculate the verification evaluation function Q of the current strategy;

[0085] S5.3: Implement parameter dynamic updating mechanism

[0086] According to the verification result, the routing modeling parameters are adjusted reversely, forming a closed loop optimization; the model parameters are updated by using gradient descent method with momentum term:

[0087]

[0088] wherein θ t is the model parameter at the current time; η is the learning rate, controlling the update step size; is the gradient of the validation evaluation function Q with respect to the parameter; ξ is the momentum factor, used to accelerate convergence and suppress oscillation; Δθ t-1 is the previous parameter update;

[0089] The updated routing model is used to generate a new optimization strategy and is fed into the digital twin platform again for verification, forming a continuous iterative closed-loop optimization process.

[0090] When the validation evaluation function Q converges to a stable level or meets the preset performance target, the current optimal routing strategy is output as the final decision result.

[0091] A power communication multi-routing intelligent planning device based on scene classification, comprising:

[0092] A multi-source heterogeneous data fusion processing module: collects power communication scene data, establishes a dynamic weighted fusion model, generates a fusion feature vector, introduces a spatiotemporal correlation feature extraction mechanism, and obtains spatiotemporal feature parameters;

[0093] A three-dimensional scene feature classification module: concatenates the fusion feature vector and the spatiotemporal feature parameters, inputs them into a three-layer fully connected neural network for feature mapping, calculates the class scores using a fully connected layer for the mapped features, and outputs the predicted class probability distribution of the current communication scene and the predicted class of the current communication scene.

[0094] A dynamic routing modeling module: obtains the classification information of the current scene, defines a multi-objective routing optimization function, dynamically adjusts the dynamic weights of each optimization objective, substitutes the dynamic weights into the multi-objective routing optimization function, scores and sorts all candidate paths, and selects the optimal path as the current routing strategy.

[0095] A hybrid intelligent optimization module: uses a particle swarm optimization algorithm and a genetic algorithm for hybrid optimization, combines an adaptive inertia weight mechanism to optimize the current routing strategy, and outputs the optimized routing strategy.

[0096] A real-time verification and feedback optimization module: deploys the optimized routing strategy to a digital twin verification platform, quantifies the verification results using a validation evaluation function, reversely adjusts the routing modeling parameters based on the verification results, updates the routing model, generates a new optimization strategy, iteratively optimizes the strategy until convergence, and outputs the final decision result.

[0097] A computer device comprises a memory and a processor, the memory stores a computer program, and the processor implements the steps of the power communication multi-routing intelligent planning method based on scene classification when executing the computer program.

[0098] A computer readable storage medium stores a computer program, and the computer program implements the steps of the power communication multi-routing intelligent planning method based on scene classification when executed by a processor.

[0099] The advantages of the present application are: the present application significantly improves the intelligent level and adaptability of power communication network management through a series of innovative methods. First, in the multi-source heterogeneous data fusion processing, the dynamic feature extraction matrix and the spatio-temporal correlation feature extraction mechanism are adopted, which solves the problem of insufficient perception ability caused by the traditional system static weighting or simple splicing method, and realizes efficient and accurate perception of complex network state. Secondly, in the three-dimensional scene feature classification stage, a feature projection space with strong expression ability is constructed based on a deep reinforcement classification network, which overcomes the limitations of traditional shallow classification models in high-dimensional feature modeling, greatly improving the accuracy and adaptability of scene recognition. Thirdly, the time-varying constraint modeling method is introduced in the dynamic routing modeling part, and the target weight is adjusted and optimized according to the real-time scene, which breaks through the limitation of static routing model that cannot flexibly respond to different business demands, and enhances the network service quality and resource utilization efficiency. In addition, the hybrid intelligent optimization algorithm combines the advantages of PSO and GA, effectively avoiding the problem that single algorithm is easy to fall into local optimum, ensuring the stability and efficiency of the optimization process. Finally, through the digital twin closed-loop verification mechanism, online simulation evaluation and parameter dynamic updating of the routing scheme are realized, forming a complete closed-loop management system from "prediction - execution - verification - optimization", which significantly improves the practical feasibility and response speed of the strategy. These improvements together provide a more intelligent, more reliable and more forward-looking management solution for power communication networks. BRIEF DESCRIPTION OF DRAWINGS

[0100] Figure 1 The flowchart of the method of the present application. DETAILED DESCRIPTION

[0101] The principles and features of the present application are described below in conjunction with the accompanying drawings, and the examples are only used to explain the present application and not to limit the scope of the present application.

[0102] EMBODIMENT

[0103] The present application provides a power communication multi-routing intelligent planning method based on scene classification in the field of power, as shown in Figure 1 The power communication multi-routing intelligent planning method comprises:

[0104] S1: multi-source heterogeneous data fusion processing;

[0105] In the traditional power communication network management system, the integration capability of multi-source heterogeneous data has obvious limitations. The existing technology often uses static weighting or simple splicing to process data from different dimensions such as device status, service demand and environmental monitoring, which is difficult to effectively reflect the dynamic change characteristics of each data source and its importance difference in different space-time scenes, resulting in insufficient system perception ability, response lag, and affecting the overall decision efficiency and accuracy.

[0106] The present application proposes a multi-source heterogeneous data fusion processing method based on dynamic feature extraction matrix. By introducing a dynamic weighted fusion mechanism, a unified data representation model is constructed to achieve efficient and accurate perception of the power communication network state. This method not only improves the flexibility and adaptability of data fusion, but also provides high-quality input for subsequent scene recognition and intelligent scheduling.

[0107] S1.1: Data acquisition and preprocessing;

[0108] Data acquisition and preprocessing, first, three types of key data are collected from multiple sources respectively: device status data (D s ): such as bandwidth, CPU utilization, etc.; service demand data (D d ): such as service priority, QoS level; environmental monitoring data (D e ): such as temperature and humidity, interference intensity, etc. All data are standardized and time-aligned to ensure their comparability and consistency in subsequent fusion process.

[0109] S1.2: Establish a dynamic weighted fusion model;

[0110] The fusion expression is constructed as follows:

[0111] F=W⊙(αD s +βD d +γD e )

[0112] Where, α, β, γ are the initial weight coefficients of each data source, representing its basic importance in the system; W is a dynamic weight matrix, which is self-adaptively adjusted according to real-time network state and task demand, used to highlight the most critical data dimension in the current scene; ⊙ represents Hadamard product (i.e. element-wise multiplication), realizing fine-grained weighting control of the fusion result.

[0113] S1.3: Introduce spatiotemporal correlation feature extraction mechanism;

[0114] After completing the preliminary fusion, further construct spatiotemporal correlation features to enhance global perception ability:

[0115]

[0116] where φ(τ) represents the communication traffic feature value at time τ; λ is the decay coefficient, used to emphasize the influence of recent data on the current state; Δt represents the length of the time window, controlling the backtracking range of historical information.

[0117] The integral form of the time weighting mechanism enables the system to capture the trend of the evolution of the communication network state over time and suppress noise interference, improving the robustness of the model.

[0118] S1.4: Output the fusion feature vector;

[0119] The final output fusion feature vector F and the spatiotemporal feature T(t) will serve as key inputs for the next stage (such as scene classification, resource scheduling), supporting more intelligent network management decisions.

[0120] S2: Three-dimensional scene feature classification;

[0121] In traditional power communication network management, the scene recognition link generally relies on manually set rules or shallow classification models, making it difficult to effectively cope with complex and variable communication environments. In particular, when facing diverse scenarios such as emergency support, daily operation and maintenance, and fault handling, traditional methods often lack effective modeling capabilities for high-dimensional features, resulting in low classification accuracy and poor adaptability, which in turn affects the intelligent level of subsequent routing strategies and resource scheduling.

[0122] The present invention proposes a three-dimensional scene feature classification method based on a deep reinforcement classification network, which introduces neural network structures and feature enhancement mechanisms to construct a feature projection space with strong expression ability, thereby achieving accurate recognition and efficient classification of communication scenes. This method not only improves the accuracy of classification, but also enhances the system's adaptability to complex environmental changes, providing a reliable basis for intelligent decision-making.

[0123] S2.1: Input feature vector preparation;

[0124] The fusion feature vector F and the spatiotemporal correlation feature T(t) output in step S1 are used as inputs to form a complete feature representation:

[0125] [F; T(t)]

[0126] where F is the feature vector after fusion of multi-source heterogeneous data; T(t) is the communication traffic evolution feature in the time dimension; ";" represents the vector concatenation operation, used to integrate spatial and temporal information and improve the perception dimension of the classification model.

[0127] S2.2: Constructing a feature projection space;

[0128] The concatenated features are input into a three-layer fully connected neural network with a nonlinear activation function for feature mapping:

[0129] Z = ReLU(W z · [F; T(t)] + b z )

[0130] where W z is a trainable weight matrix; b z is a bias term; ReLU is an activation function to introduce nonlinearity and enhance model expressiveness; Z is the output high-dimensional feature representation for subsequent classification tasks.

[0131] This process realizes dimensionality reduction and semantic enhancement of the original features, making the feature differences between different scenarios more obvious and facilitating subsequent classifier discrimination.

[0132] S2.3: Design a deep reinforcement classification network;

[0133] Based on feature projection, a fully connected layer with Sigmoid activation is used to calculate class scores, and a Softmax function is used to output the final classification probability distribution:

[0134] P(c k | F) = softmax(σ(W c · F + b c ))

[0135] where c k ∈ {emergency support, daily operation and maintenance, fault handling}: target scenario category; W c , b c are trainable parameters of the classification layer; σ is a Sigmoid activation function to introduce non-linear mapping; softmax is used to convert the output into a probability distribution of each scenario category.

[0136] The classification network continuously optimizes parameters through end-to-end training, combined with actual feedback signals (such as misclassification penalty, response delay, etc.), gradually improving classification accuracy and robustness, embodying the idea of "reinforcement learning".

[0137] The final output is the predicted category c k of the current communication scenario, which serves as an important basis for the next stage (such as route strategy generation) to support differentiated service support and dynamic resource scheduling.

[0138] The end-to-end training refers to the entire process from the input feature vector (fusion feature F and spatiotemporal feature T(t)) to the final output classification probability distribution, forming a differentiable whole model. By defining a loss function, the backpropagation algorithm is used to automatically calculate the gradients of all trainable parameters in the model, and the optimizer is used to update the parameters to minimize the prediction error.

[0139] During the training process, the label of each sample is the actual communication scenario category (emergency support, daily operation and maintenance, fault handling), and the model adjusts the parameters to make the output probability distribution as close as possible to the real label distribution. Through multiple rounds of iterative training, the classification network has good scene recognition ability, thereby providing accurate scene classification basis for subsequent route optimization.

[0140] The end-to-end training mechanism realizes the joint optimization of feature extraction, feature mapping and classification decision, and improves the overall generalization performance and adaptability of the model.

[0141] S3: Dynamic route modeling;

[0142] In the traditional power communication network, the route planning generally adopts a static or semi-static model, and the optimization target and weight setting are often fixed, which is difficult to adapt to complex and changeable business scenarios and network environment. This static design makes the system unable to effectively balance the relationship between path cost, transmission delay and link reliability when facing different scenarios such as emergency support, daily operation and maintenance, fault handling, thereby affecting the overall service quality and resource utilization efficiency.

[0143] The application proposes a dynamic route modeling mechanism based on time-varying constraint modeling method, which adjusts the weight distribution of the route optimization target in real time by combining the scene classification results output in step S2, thereby realizing intelligent adaptation of the route strategy under different scenarios. This method breaks through the limitations of traditional static route models and improves the response capability and adaptability of the network under multi-dimensional constraints.

[0144] S3.1: Obtain current scene classification information;

[0145] Obtain the current communication scenario category c from the output of step S2 k (e.g. emergency support, daily operation and maintenance, or fault handling), and extract the corresponding scene sensitivity parameter S j (t) according to the scene type, which is used to guide the subsequent weight dynamic adjustment. The scene sensitivity parameter Sj(t) is a sensitivity template defined based on historical experience and scene category, which is matched and extracted through the classification result ck, and is used to guide the dynamic priority adjustment of the route optimization target. This method can realize the adaptive response of the route strategy to different scenarios, and improve the decision flexibility and performance of the system in complex communication environment.

[0146] S3.2: Construct a multi-objective optimization function;

[0147] Based on the actual network performance indicators, the following multi-objective route optimization function is defined:

[0148]

[0149] Wherein: C i is the cost (such as hop count, energy consumption) of the ith candidate path; D i is the end-to-end transmission delay of the ith path; R i is the reliability index (such as link stability) of the ith path; represents converting maximizing reliability into a minimization problem; ω1, ω2, ω3 are dynamic weights of each optimization objective, calculated in the next step, and n represents the total number of candidate paths, which means that there are n different paths to choose from.

[0150] This function comprehensively considers the economy, timeliness and stability in path selection, and is the core basis for dynamic routing decision.

[0151] S3.3: Implement a dynamic weight adjustment mechanism;

[0152] According to the current scene characteristics, a weighted strategy in the form of softmax is used to dynamically adjust the importance of each objective:

[0153]

[0154] Wherein, m is an index variable for traversing all optimization objectives; k m is the adjustment coefficient related to the mth optimization objective; S m (t) represents the scene sensitivity parameter of the mth optimization objective at the current time t; S j (t) is the scene sensitivity parameter of the jth objective at the current time t; k j is the adjustment coefficient, which controls the response sensitivity of the objective to scene changes, and the denominator is the normalization term, which ensures that the sum of all weights is 1;

[0155] Through this mechanism, the routing strategy can flexibly highlight key indicators according to different scenes. For example, increase the reliability weight in the "emergency guarantee" scene, focus on cost control in the "daily operation and maintenance", and prioritize reducing delay in "fault handling".

[0156] Substitute the dynamic weights into the multi-objective optimization function to score and sort all candidate paths, and select the path with the best comprehensive performance as the final routing scheme. This process can be combined with existing shortest path algorithms (such as Dijkstra or A*) for efficient solution.

[0157] Finally, output the optimal routing path and corresponding strategy parameters that adapt to the current communication scene, for the network scheduling module to call, in order to realize differentiated service guarantee and dynamic resource allocation.

[0158] In this step, the dynamic weights of each optimization objective (such as cost, latency, reliability) are first determined. These weights are calculated based on the current communication scenario and corresponding sensitivity parameters. Then, these weights are used in a multi-objective routing optimization function that considers all candidate paths in terms of cost, latency, and reliability, and generates a comprehensive score for each path through weighted summation.

[0159] Once the comprehensive score of each path is obtained, the next step is to sort all candidate paths according to their scores.

[0160] In this step, according to the sorting result obtained in the previous step, the path with the lowest score is selected as the optimal routing strategy under the current scenario. This means that the best path that can meet the cost control requirements, ensure low latency, and have high reliability under the current communication scenario is selected.

[0161] The corresponding strategy parameters refer to the specific configuration or setting values associated with the selected optimal path. This may include but is not limited to: path selection: that is, which path or paths are actually selected as the channel for data transmission. Resource allocation: such as bandwidth allocation, priority setting, etc., to ensure that the selected path can provide the expected quality of service in actual operation. Dynamic adjustment mechanism: may also include some dynamic adjustment parameters, such as how to automatically adjust the routing strategy to maintain service performance when traffic increases or network conditions change. In summary, "corresponding strategy parameters" refer to specific numerical values or settings that are directly related to the implementation and operation of the selected optimal routing path.

[0162] S4: Hybrid intelligent optimization algorithm;

[0163] In traditional intelligent optimization algorithms, although the particle swarm optimization algorithm (PSO) has a faster convergence speed, it is easy to fall into local optimum in complex search space; while the genetic algorithm (GA) has strong global search ability, but its convergence speed is slow and the computational cost is large. These limitations are particularly prominent in high-dimensional, multi-objective problems such as power communication network resource scheduling and path optimization, seriously affecting system efficiency and stability.

[0164] The present application proposes a hybrid intelligent optimization method based on PSO-GA fusion algorithm, which combines the fast convergence characteristics of particle swarm optimization algorithm and the strong global exploration ability of genetic algorithm, and constructs an efficient optimization framework with adaptive adjustment mechanism, thereby realizing high-quality solution to power communication network scheduling problems. This method not only improves the stability and search efficiency of the algorithm, but also effectively avoids the problem of single algorithm falling into local optimum.

[0165] S4.1: Initialization of population and parameter setting;

[0166] First, according to the dimension of the optimization problem (such as route path selection, resource allocation strategy, etc.), a set of initial solutions is randomly generated as a population individual, each individual is represented as a vector x i . At the same time, set the following key parameters: particle number N; maximum iteration number iter max ; learning factor c1, c2; inertia weight range [ω min ,ω max ]; crossover factor range [0.3, 0.7]; Gaussian disturbance variance σ 2 .

[0167] S4.2: Introduce adaptive inertia weight mechanism to improve convergence performance;

[0168] In each iteration process, the inertia weight ω is dynamically adjusted in a linear decreasing manner to balance the global exploration and local development ability of the algorithm:

[0169]

[0170] where ω max is the initial larger inertia weight, encouraging extensive search in the early stage; ω min is the later smaller inertia weight, improving local accuracy; iter is the current iteration number.

[0171] This mechanism makes the algorithm have stronger global search ability in the early stage, and gradually focuses on the optimal solution region in the later stage.

[0172] S4.3: Perform particle swarm velocity and position update

[0173] Update the velocity and position of each particle according to the standard PSO formula:

[0174]

[0175] where, and are the current velocity and position of the i-th particle in the d-th dimension; pbest id is the individual historical optimal solution; gbest d is the global optimal solution; r1, r2 are random numbers in the interval [0, 1].

[0176] Through this step, the local optimal region is quickly approached, and the overall convergence efficiency of the algorithm is improved.

[0177] S4.4: Introduce genetic operation to enhance global search ability

[0178] After every several iterations, select some excellent individuals from the current population for genetic operation, including crossover and mutation, to escape from the local optimal trap:

[0179] A. Crossover operation: randomly select two parent individuals x p and x q , generate a new individual according to the following formula:

[0180] x new = αx p + (1-α)x q + N(0, σ 2 )

[0181] where: α ∈ [0.3, 0.7] is a dynamic crossover factor, controlling the fusion ratio of parent information; N(0, σ 2 ) is a Gaussian noise disturbance term, used to increase diversity.

[0182] B. Mutation operation: apply a small disturbance to some individuals, further expand the search space, and prevent premature convergence.

[0183] Evaluate the fitness of all particles (including newly generated individuals) (usually according to the multi-objective function defined in step S3), update the individual optimal solution pbest and the global optimal solution gbest.

[0184] If the maximum number of iterations is reached or the preset convergence condition (such as no significant change in the optimal solution for several consecutive times) is met, terminate the algorithm, and output the current optimal solution gbest, which is the final optimized routing strategy or resource allocation scheme.

[0185] S5: Real-time verification and feedback optimization;

[0186] In traditional power communication networks, the design and deployment of routing schemes often rely on offline simulation or empirical setting, lacking dynamic feedback mechanisms for actual operating environments. This "open-loop" design cannot accurately reflect the changes in complex network states, leading to performance deviations, resource waste, and even degradation of service quality in actual applications. Especially in the face of sudden traffic, equipment failure or environmental disturbance, traditional methods lack the ability to respond quickly and adaptively optimize.

[0187] The present application proposes a real-time verification and feedback optimization method based on a digital twin closed-loop verification mechanism, which builds a high-fidelity network digital twin platform to realize online simulation evaluation of routing schemes, and dynamically adjusts model parameters combined with feedback information, forming a "modeling-verification-optimization" closed-loop system. This method significantly improves the feasibility and stability of routing strategies, providing continuous evolution capabilities for intelligent scheduling.

[0188] S5.1: Establish a digital twin verification platform

[0189] Based on the current network topology, device state, traffic flow characteristics and the routing strategy output in step S4, a virtual simulation environment highly consistent with the physical network is constructed. The platform can synchronize the running state of the physical network in real time and support the rapid deployment and effect preview of the routing strategy.

[0190] S5.2: Define verification evaluation function

[0191] The routing strategy is executed in the digital twin environment and the key performance indicators are collected. The following evaluation function is constructed to quantify the quality of the strategy:

[0192]

[0193] wherein, and is the transmission delay of the i-th traffic flow in simulation and reality; is the energy efficiency value (such as unit data transmission energy consumption) of the path in simulation;E th is the set energy efficiency threshold; N is the number of traffic flows participating in evaluation.

[0194] The function comprehensively considers the simulation error of the routing strategy in delay control and energy efficiency performance. The smaller the value, the closer the strategy is to the real demand.

[0195] The routing strategy recommended in step S4 is deployed to the digital twin platform to simulate its running process in the network. The performance indicators are recorded and compared with the actual network running data. The verification score Q of the current strategy is calculated.

[0196] S5.3: Implement parameter dynamic update mechanism

[0197] According to the verification results, the routing modeling parameters are adjusted reversely to form a closed loop optimization. The gradient descent method with momentum term is used to update the model parameters:

[0198]

[0199] wherein, θ t is the model parameter (such as routing weight, constraint condition, etc.) at the current time; η is the learning rate, which controls the update step size; is the gradient of the evaluation function Q to the parameter; ξ is the momentum factor, which is used to accelerate convergence and suppress oscillation; Δθ t-1 is the previous parameter update amount.

[0200] The mechanism ensures that the routing model can continuously correct itself according to the differences between simulation and reality, improving the robustness and adaptability of the strategy.

[0201] The updated routing model will be used to generate new optimization strategies and sent into the digital twin platform again for verification, forming a continuous iterative closed-loop optimization process. This process can be executed periodically or triggered when a network state mutation is detected to maintain the timeliness and accuracy of the strategy.

[0202] When the verification evaluation function Q converges to a stable level or meets the preset performance target, the current optimal routing strategy is output as the final decision result for the network scheduling module to call, completing the complete closed-loop management from "prediction - execution - verification - optimization".

[0203] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A scenario-based multi-route intelligent planning method for power communication, characterized by: The following steps are involved: S1: Collect power communication scenario data, establish a dynamic weighted fusion model, generate fusion feature vectors, introduce a spatiotemporal correlation feature extraction mechanism, and obtain spatiotemporal feature parameters; S2: The fused feature vector and spatiotemporal feature parameters are concatenated and input into a three-layer fully connected neural network for feature mapping. The mapped features are then classified using a fully connected layer to calculate the category score. The predicted category probability distribution and the predicted category of the current communication scenario are output. S3: Obtain the classification information of the current scenario from the results of S2, define a multi-objective routing optimization function, dynamically adjust the dynamic weights of each optimization objective, substitute the dynamic weights into the multi-objective routing optimization function, score and rank all candidate paths, and select the optimal path as the current routing strategy; S4: Use the particle swarm optimization algorithm and genetic algorithm hybrid optimization, combined with the adaptive inertia weight mechanism to optimize the current routing strategy and output the optimized routing strategy; S5: Deploy the optimized routing strategy to the digital twin verification platform, quantify the verification results through the verification evaluation function, reversely adjust the routing modeling parameters based on the verification results, update the routing model, and generate a new optimization strategy. Iterate the optimization strategy until convergence and output the final decision result.

2. The method for intelligent planning of multi-routes for power communication based on scenario classification according to claim 1, characterized in that: As described in S1, the power communication scenario data is collected, a dynamic weighted fusion model is established, a fusion feature vector is generated, and a spatiotemporal correlation feature extraction mechanism is introduced to obtain spatiotemporal feature parameters. The specific contents are as follows: S1.1: Data collection and preprocessing; Collect device status data D s , business demand data D d , environmental monitoring data D e Three types of data, all data are normalized and time-aligned; S1.2: Establish a dynamic weighted fusion model; Construct the fusion expression as follows: F=W⊙(αD s +βD d +γD e ) Where F is the fusion feature vector, α, β, and γ are the initial weight coefficients of each data source, W is a dynamic weight matrix, and ⊙ represents the Hadamard product; S1.3: Introduce spatiotemporal correlation feature extraction mechanism; Construct spatiotemporal correlation features: Where T(t) is the spatiotemporal characteristic parameter, Φ(τ) represents the characteristic value of the communication traffic at time τ; λ is the attenuation coefficient, which is used to emphasize the impact of recent data on the current state; Δt represents the length of the time window, which controls the retrospective range of historical information. S1.4: Output fusion feature vector F and spatiotemporal feature parameters T(t).

3. The method for intelligent planning of multi-routes for power communication based on scenario classification according to claim 2, characterized in that: As described in S2, the fusion feature vector and the spatiotemporal feature parameters are spliced ​​and input into a three-layer fully connected neural network for feature mapping. The mapped features are subjected to category score calculation using a fully connected layer, and the predicted category probability distribution and the predicted category of the current communication scenario are output. The specific contents are as follows: S2.1: Input feature vector preparation; The output fusion feature vector F and spatiotemporal correlation feature parameter T(t) are used as input and concatenated to form a complete feature representation: [F; T(t)] Where F is the fusion feature vector after fusion of multi-source heterogeneous data; ";" represents the vector splicing operation, which is used to integrate spatial and temporal information; S2.2: Construct feature projection space; The concatenated features are input into a three-layer fully connected neural network, and a nonlinear activation function is added for feature mapping: Z=ReLU(W z [F;T(t)]+b z ) Among them, W z is the trainable weight matrix; b z is the bias term; ReLU is the activation function used to introduce nonlinear characteristics; Z is the high-dimensional feature representation of the output; S2.3: Design a deep reinforcement classification network; Based on the feature projection, a fully connected layer with Sigmoid activation is used to calculate the category score, and the Softmax function is combined to output the final classification probability distribution: P(c k |F)=softmax(σ(W c ·F+b c )) Among them, c k ∈{emergency support, daily operation and maintenance, fault handling}: target scenario category; W c , b c is the trainable parameter of the classification layer; σ is the Sigmoid activation function, which is used to introduce nonlinear mapping; softmax is used to convert the output into the probability distribution of each scene category; Parameters are continuously optimized through end-to-end training, and the predicted category c of the current communication scenario is finally output. k .

4. The method for intelligent planning of multi-routes for power communication based on scenario classification according to claim 3 is characterized in that: S3 obtains the classification information of the current scenario from the results of S2, defines a multi-objective routing optimization function, dynamically adjusts the dynamic weights of each optimization objective, substitutes the dynamic weights into the multi-objective routing optimization function, scores and ranks all candidate paths, and selects the optimal path as the current routing strategy. The specific contents are as follows: S3.1: Get the current scene classification information; Get the predicted category c of the current communication scenario k , extract the corresponding scene sensitivity parameter S j (t); S3.2: Construct a multi-objective routing optimization function; Define the following multi-objective routing optimization function: Where: C i is the cost of the i-th candidate path; D i is the end-to-end transmission delay of the i-th path; R i is the reliability index of the i-th path; Indicates that maximizing reliability is transformed into a minimization problem; ω1, ω2, ω3 are the dynamic weights of each optimization objective, and n represents the total number of candidate paths; S3.3: Implement a dynamic weight adjustment mechanism; According to the current scene characteristics, a softmax weighting strategy is used to dynamically adjust the importance of each target: Among them, m is an index variable used to traverse all optimization objectives; k m is the adjustment coefficient related to the mth optimization objective; m (t) represents the scene sensitivity parameter of the mth optimization target at the current time t; S j (t) is the scene sensitivity parameter of the jth target at the current time t; k j is the adjustment coefficient, which controls the target's sensitivity to scene changes. The denominator is a normalization term to ensure that the sum of all weights is 1. The dynamic weights are substituted into the multi-objective routing optimization function, and the shortest path algorithm is used to score and sort all candidate paths. The path with the best comprehensive performance is selected as the final routing strategy, and the optimal routing path and corresponding strategy parameters adapted to the current communication scenario are output.

5. The method for intelligent planning of multi-routes for power communication based on scenario classification according to claim 4, characterized in that: As described in S4, the particle swarm optimization algorithm and the genetic algorithm are hybrid optimized, combined with the adaptive inertia weight mechanism to optimize the current routing strategy, and the optimized routing strategy is output as follows: S4.1: Initialize population and parameter setting; First, according to the dimension of the optimization problem, a set of initial solutions are randomly generated as population individuals, each of which is represented by a vector x i , and set the following parameters: number of particles N; maximum number of iterations iter max ; learning factors c1, c2; inertia weight range [ω min ,ω max ]; crossover factor range [0.3, 0.7]; Gaussian perturbation variance σ 2 ; S4.2: Introduce adaptive inertia weight mechanism; In each iteration, the inertia weight ω is dynamically adjusted in a linear decreasing manner: Among them, ω max is the initial larger inertia weight; ω min is the smaller inertia weight in the later stage; iter is the current iteration number; S4.3: Perform particle swarm velocity and position updates The speed and position of each particle are updated according to the standard particle swarm optimization algorithm formula: in, and are the current velocity and position of the i-th particle in the d-th dimension; pbest id is the individual historical optimal solution; gbest d is the global optimal solution; r1, r2 are random numbers in the interval [0,1], and k represents the number of iterations; S4.4: Introducing Genetic Algorithms After every several iterations, some excellent individuals are selected from the current population for genetic operations, including crossover and mutation, to escape the local optimal trap; According to the multi-objective routing optimization function, the fitness of all particles is evaluated, and the individual optimal solution pbest and the global optimal solution gbest are updated; If the maximum number of iterations is reached or the preset convergence condition is met, the algorithm is terminated and the current optimal solution gbest is output, which is the final optimized routing strategy or resource allocation plan.

6. The method for intelligent planning of multi-routes for power communication based on scenario classification according to claim 5, characterized in that: The crossover operation is as follows: Randomly select two parent individuals x p and x q , generate new individuals according to the following formula: x new =αx p +(1-a)x q +N(0,σ 2 ) Among them: α∈[0.3,0.7] is the dynamic cross factor, which controls the proportion of parent information fusion; N(0,σ 2 ) is the Gaussian noise disturbance term; The mutation operation is as follows: applying small perturbations to some individuals to further expand the search space; The specific operations are as follows: Individual selection: Randomly select some individuals from the current population for mutation operations; Perturbation generation: A small perturbation term that follows a Gaussian distribution is applied to each dimension of the selected individual. The perturbation amplitude is adjusted according to the search space of the actual problem. The expression is as follows: Among them, x i is the position vector of the original individual, is the disturbance amplitude, N(0,1) represents the standard normal distribution random number; Boundary constraint processing: perform boundary checking and clipping on the mutation results; Population update: The mutated individuals are reinserted into the population to participate in the next round of fitness evaluation and evolution.

7. The method for intelligent planning of multi-routes for power communication based on scenario classification according to claim 5, characterized in that: As described in S5, the optimized routing strategy is deployed to the digital twin verification platform. The verification results are quantified by the verification evaluation function. Based on the verification results, the routing modeling parameters are reversely adjusted, the routing model is updated, and a new optimization strategy is generated. The optimization strategy is iterated until convergence, and the final decision result is output. The details are as follows: S5.1: Establish a digital twin verification platform Build a virtual simulation environment based on the current network topology, device status, service flow characteristics and routing strategy output by S4; S5.2: Define the validation evaluation function Execute the routing strategy in the digital twin environment and collect performance indicators. Construct the following evaluation function to quantify the quality of the strategy: in, and are the transmission delays of the i-th service flow in simulation and reality respectively; is the energy efficiency value of the path in the simulation; E th is the set energy efficiency threshold; N is the number of business flows participating in the evaluation; Deploy the routing strategy output by S4 to the digital twin platform, simulate its operation process in the network, record various performance indicators and compare them with the actual network operation data, and calculate the verification evaluation function Q of the current strategy; S5.3: Implement a dynamic parameter update mechanism According to the verification results, the routing modeling parameters are reversely adjusted to form a closed-loop optimization; the model parameters are updated using the gradient descent method with momentum: Among them, θ t is the model parameter at the current moment; η is the learning rate, which controls the update step size; is to verify the gradient of the evaluation function Q with respect to the parameters; ξ is the momentum factor, which is used to accelerate convergence and suppress oscillation; Δθ t-1 is the previous parameter update amount; The updated routing model is used to generate new optimization strategies and then sent to the digital twin platform for verification, forming a continuously iterative closed-loop optimization process. When the verification evaluation function Q converges to a stable level or meets the preset performance target, the current optimal routing strategy is output as the final decision result.

8. A scenario-classified multi-route intelligent planning device for power communication, characterized by: Includes: Multi-source heterogeneous data fusion processing module: collects power communication scenario data, establishes a dynamic weighted fusion model, generates fusion feature vectors, introduces a spatiotemporal correlation feature extraction mechanism, and obtains spatiotemporal feature parameters; Scene feature classification module: The fused feature vector and spatiotemporal feature parameters are concatenated and input into a three-layer fully connected neural network for feature mapping. The mapped features are then classified using a fully connected layer to calculate the category score. The module then outputs the predicted category probability distribution and the predicted category of the current communication scenario. Dynamic routing modeling module: obtains classification information of the current scenario, defines a multi-objective routing optimization function, dynamically adjusts the dynamic weights of each optimization objective, substitutes the dynamic weights into the multi-objective routing optimization function, scores and ranks all candidate paths, and selects the optimal path as the current routing strategy; Hybrid intelligent optimization module: uses a hybrid optimization algorithm of particle swarm optimization and genetic algorithm, combined with an adaptive inertia weight mechanism to optimize the current routing strategy and output the optimized routing strategy; Real-time verification and feedback optimization module: Deploy the optimized routing strategy to the digital twin verification platform, quantify the verification results through the verification evaluation function, reversely adjust the routing modeling parameters based on the verification results, update the routing model, and generate a new optimization strategy. Iterate the optimization strategy until convergence and output the final decision result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the scene classification-based power communication multi-route intelligent planning method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for intelligent planning of multi-routes of power communication based on scene classification are implemented.

Images