Electric power distribution network planning system based on artificial intelligence

Abstract

The invention discloses a power distribution network planning system based on artificial intelligence, which relates to the technical field of power system planning and comprises a data preprocessing module, a topological optimization module, a scheme evaluation module and a decision output module. Wherein the data preprocessing module is used for extracting a power distribution network load feature vector set; the topological optimization module constructs a topological optimization model by using a self-adaptive genetic algorithm unit and taking a power supply reliability index and a line loss rate as double objective functions, and generates a topological structure candidate scheme; the scheme evaluation module performs quantitative analysis on the candidate schemes by adopting a deep Q network to obtain an evaluation score matrix; and the decision output module ranks the candidate schemes through the fuzzy comprehensive evaluation unit and outputs an optimal power distribution network planning scheme. According to the method, load characteristics are extracted through a deep learning technology, a genetic algorithm is improved to optimize a topological structure, learning evaluation scheme quality is reinforced, an optimal scheme is determined through fuzzy evaluation, and the accuracy and scientificity of power distribution network planning are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system planning, and in particular to an artificial intelligence-based power distribution network planning system and system. Background Art

[0002] With the continuous expansion of the power grid and the continuous growth of electricity load, distribution network planning faces increasingly complex challenges. Existing distribution network planning methods mainly rely on manual experience and traditional algorithms. They have low computational efficiency when processing large-scale distribution network data and are difficult to fully consider the dynamic characteristics of load distribution. Especially in urban areas where electricity demand is growing rapidly, traditional planning methods often require a lot of manpower for data analysis and scheme comparison, and the planning efficiency is relatively low.

[0003] In addition, the existing distribution network planning system has deficiencies in multi-objective optimization and scheme evaluation. Problems such as a single optimization target and an imperfect evaluation index system often occur during the planning process, making it difficult to simultaneously take into account the economy, reliability, and environmental protection of the power grid planning scheme, which easily leads to insufficient practicality and foresight of the planning scheme. Summary of the invention

[0004] In view of the problems existing in the existing distribution network planning system in terms of data processing efficiency and multi-objective optimization, the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is how to build an intelligent and systematic distribution network planning system to improve planning efficiency and solution quality.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: In the first aspect, an embodiment of the present invention provides an artificial intelligence-based power distribution network planning system, which includes a data preprocessing module, a topology optimization module, a scheme evaluation module, and a decision output module, wherein: a deep convolutional neural network is provided in the data preprocessing module, the deep convolutional neural network receives the historical load data and power demand data of the distribution network and performs feature extraction to obtain a distribution network load feature vector set; an adaptive genetic algorithm unit is provided in the topology optimization module, the adaptive genetic algorithm unit receives the distribution network load feature vector set, and constructs a topology optimization model with power supply reliability index and line loss rate as a dual objective function to obtain a topology structure candidate scheme; a deep reinforcement learning unit is provided in the scheme evaluation module, the deep reinforcement learning unit includes a deep Q network, the deep Q network performs quantitative analysis on the topology structure candidate scheme to obtain an evaluation score matrix; a fuzzy comprehensive evaluation unit is provided in the decision output module, the fuzzy comprehensive evaluation unit sorts the topology structure candidate schemes based on the evaluation score matrix, and outputs the optimal distribution network planning scheme.

[0007] As a preferred solution of the artificial intelligence-based power distribution network planning system of the present invention, the data preprocessing module also includes a filtering preprocessing unit, which performs outlier detection and smoothing on the historical load data and power demand data of the distribution network; the adaptive genetic algorithm unit generates topological structure candidate solutions through crossover mutation operations; the deep Q network quantitatively analyzes the economy, reliability and environmental protection of the topological structure candidate solutions; and the fuzzy comprehensive evaluation unit uses a hierarchical analysis method to establish a set of discrimination criteria.

[0008] As a preferred solution of the artificial intelligence-based power distribution network planning system described in the present invention, the deep convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer; the input layer receives the historical load data and power demand data of the distribution network, and converts the historical load data and power demand data of the distribution network into a standardized input matrix; the convolutional layer uses multiple convolution kernels of different sizes to perform multi-scale feature extraction on the standardized input matrix, and the sizes of the convolution kernels are arranged in an increasing sequence; the pooling layer uses a maximum pooling operation to reduce the dimension of the features extracted by the convolution layer, and the windows of the maximum pooling operation are in a multiple relationship between adjacent layers; the fully connected layer flattens the feature map output by the pooling layer into a one-dimensional vector, and performs nonlinear feature transformation through a ReLU activation function; the output layer performs dimensionality reduction compression on the one-dimensional vector output by the fully connected layer to form a distribution network load feature vector set.

[0009] As a preferred solution of the artificial intelligence-based power distribution network planning system described in the present invention, the adaptive genetic algorithm unit includes an encoding subunit, a fitness calculation subunit, a selection subunit, a crossover subunit and a mutation subunit; the encoding subunit adopts a binary encoding method to convert the distribution network load feature vector set into a chromosome group, and each chromosome in the chromosome group represents a topological structure scheme; the fitness calculation subunit constructs a dual-objective fitness function based on the power supply reliability index and the line loss rate, the power supply reliability index includes the power supply interruption time and the power supply interruption frequency, and the line loss rate includes active loss and reactive loss; the selection subunit adopts a roulette wheel method to select high-quality chromosomes from the chromosome group, and the selection probability is positively correlated with the fitness function value; the crossover subunit adopts an adaptive crossover operator to recombine the high-quality chromosomes, and the crossover probability of the adaptive crossover operator changes with the population evolution algebra; the mutation subunit adopts an adaptive mutation operator to perform gene mutation on the chromosomes after crossover recombination, and the mutation probability of the adaptive mutation operator changes with the fitness function value, thereby completing the generation of topological structure candidate schemes.

[0010] As a preferred solution of the artificial intelligence-based power distribution network planning system described in the present invention, the deep Q network includes a state perception layer, an action evaluation layer and a value calculation layer; the state perception layer receives topology candidate solutions and converts the topology candidate solutions into a state space matrix, and the state space matrix includes node connection relationships, line capacity, and load distribution information; the action evaluation layer constructs a reward function based on the state space matrix, and the reward function consists of an economic evaluation sub-function, a reliability evaluation sub-function, and an environmental evaluation sub-function; the value calculation layer uses an experience replay mechanism to iteratively optimize the reward function, and organizes the iteratively optimized reward values ​​to form an evaluation score matrix; the experience replay mechanism includes priority sampling and temporal difference learning.

[0011] As a preferred solution of the artificial intelligence-based power distribution network planning system described in the present invention, the economic evaluation sub-function includes equipment investment cost and operation and maintenance cost, the reliability evaluation sub-function includes voltage deviation and power loss, and the environmental evaluation sub-function includes carbon emissions and land occupation; priority sampling extracts samples based on the size of the reward value, and temporal difference learning dynamically updates the value function; the row vectors of the evaluation score matrix correspond to different evaluation dimensions, and the column vectors of the evaluation score matrix correspond to different candidate solutions.

[0012] As a preferred solution of the artificial intelligence-based power distribution network planning system described in the present invention, the fuzzy comprehensive evaluation unit includes a weight allocation subunit, a membership calculation subunit, and a sorting subunit; the weight allocation subunit adopts a hierarchical analysis method to construct a discrimination criterion system, the discrimination criterion system includes a primary indicator and a secondary indicator, the primary indicator includes an economic indicator, a reliability indicator and an environmental protection indicator, and the secondary indicator corresponds to the evaluation dimension in the evaluation score matrix; the membership calculation subunit normalizes the evaluation score matrix, and adopts a trapezoidal membership function to construct a fuzzy relationship matrix, and the trapezoidal membership function sets upper and lower thresholds based on expert experience; the sorting subunit synthesizes the weight vector of the discrimination criterion system with the fuzzy relationship matrix through a fuzzy synthesis operator to obtain a comprehensive evaluation vector; the sorting subunit sorts the topology structure candidate schemes in descending order according to the element values ​​of the comprehensive evaluation vector, and selects the candidate scheme with the highest score as the optimal distribution network planning scheme.

[0013] In the second aspect, an embodiment of the present invention provides an artificial intelligence-based power distribution network planning method, which includes obtaining historical load data and electricity demand data of the distribution network; using a deep convolutional neural network to receive the historical load data and electricity demand data of the distribution network and perform feature extraction to obtain a distribution network load feature vector set; using an adaptive genetic algorithm unit to receive the distribution network load feature vector set, and constructing a topology optimization model with power supply reliability index and line loss rate as a dual objective function to obtain a topology structure candidate scheme; using a deep Q network in a deep reinforcement learning unit to perform quantitative analysis on the topology structure candidate scheme to obtain an evaluation score matrix; using a fuzzy comprehensive evaluation unit to sort the topology structure candidate schemes based on the evaluation score matrix, and output the optimal distribution network planning scheme.

[0014] The beneficial effects of the present invention are as follows: the present invention applies deep convolutional neural networks to load feature extraction of distribution networks, and realizes accurate analysis of load data through multi-scale convolution kernels and progressive pooling structures; creatively introduces adaptive genetic algorithms in topology optimization, and effectively improves optimization efficiency by designing dynamically adjusted crossover mutation operators and roulette-based selection mechanisms; uniquely combines deep Q networks with fuzzy comprehensive evaluation to construct a multi-dimensional evaluation system that includes economy, reliability and environmental protection, and improves the accuracy of scheme evaluation through priority sampling and temporal difference learning, while using hierarchical analysis method and trapezoidal membership function to realize scientific quantification of evaluation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0016] Figure 1 System architecture diagram of the artificial intelligence-based power distribution network planning system.

[0017] Figure 2 Workflow diagram of the adaptive genetic algorithm unit for artificial intelligence-based power distribution network planning system.

[0018] Figure 3 Deep Q network structure diagram for artificial intelligence-based power distribution network planning system.

[0019] Figure 4 Workflow diagram of the decision output module of the artificial intelligence-based power distribution network planning system. DETAILED DESCRIPTION

[0020] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0021] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0023] Example 1, reference Figure 1~Figure 4 , which is the first embodiment of the present invention, and this embodiment provides a power distribution network planning system based on artificial intelligence, and the system architecture diagram is as follows Figure 1 As shown in the figure, it includes data preprocessing module, topology optimization module, scheme evaluation module and decision output module. The specific structure and working process of each module are as follows: The data preprocessing module is provided with a deep convolutional neural network and a filter preprocessing unit. The filter preprocessing unit includes a data acquisition unit, a data normalization unit, an outlier processing unit and a smoothing processing unit. The data acquisition unit collects distribution network operation data according to a preset sampling frequency (such as 5 minutes). The data normalization unit performs time series alignment and data format unification on the collected data. The outlier processing unit detects outliers based on the box plot method and replaces outliers with the mean of the previous and next data. The smoothing processing unit uses the sliding average method to smooth the data, where the sliding average window size is set to 12 sampling points and the step size is 1 sampling point. The window size can be adjusted adaptively according to the actual data fluctuation degree. The filter preprocessing unit performs outlier detection and smoothing on the historical load data and power demand data of the distribution network. The deep convolutional neural network receives the historical load data and power demand data of the distribution network after filter preprocessing and performs feature extraction to obtain the distribution network load feature vector set.

[0024] The deep convolutional neural network includes input layer, convolution layer, pooling layer, fully connected layer and output layer. The input layer receives the historical load data and power demand data of the distribution network, and converts the historical load data and power demand data of the distribution network into a standardized input matrix using the z~score standardization method. The convolution layer uses multiple convolution kernels of different sizes to extract multi-scale features from the standardized input matrix. The sizes of the convolution kernels are arranged in an increasing sequence, specifically 3×3 and 5×5 convolution kernel sequences. The number of convolution kernels of each size is set to 64, the convolution step is 1, and zero padding is used to keep the size of the feature map unchanged. The pooling layer uses the maximum pooling operation to reduce the dimension of the features extracted by the convolution layer. The windows of the maximum pooling operation are multiples between adjacent layers, and a 2×2 pooling window is used. The fully connected layer flattens the feature map output by the pooling layer into a one-dimensional vector, and performs nonlinear feature transformation through the ReLU activation function. The dimension of the one-dimensional vector is 512. A Dropout layer is added to the fully connected layer, and the dropout rate is set to 0.5 to prevent overfitting. The output layer performs dimensionality reduction compression on the one-dimensional vector output by the fully connected layer, and uses an autoencoder to compress the feature dimension to 128 dimensions to form a distribution network load feature vector set.

[0025] First, the deep convolutional neural network uses two sizes of convolution kernel sequences, 3×3 and 5×5, for multi-scale feature extraction, effectively capturing the changing characteristics of distribution network load data at different time scales. Secondly, by setting 64 convolution kernels and adopting a zero-filling strategy, the integrity of the feature map is maintained, and the model's ability to recognize load patterns is enhanced. In terms of feature dimensionality reduction, the combination of maximum pooling operation and Dropout technology not only retains key feature information, but also effectively suppresses overfitting. Finally, the autoencoder is innovatively introduced to compress the feature dimension to 128 dimensions, which reduces the computational complexity while maintaining the expressive power of load features, providing a high-quality feature vector set for subsequent topology optimization. This structural design fully considers the temporal characteristics and spatial correlation of distribution network load data, and improves the accuracy and robustness of feature extraction.

[0026] An adaptive genetic algorithm unit is set in the topology optimization module. The adaptive genetic algorithm unit receives the distribution network load feature vector set, constructs a topology optimization model with the power supply reliability index and line loss rate as the dual objective function, and obtains the topology structure candidate scheme. The adaptive genetic algorithm unit generates the topology structure candidate scheme through crossover mutation operation. The initial population size of the adaptive genetic algorithm unit is dynamically set according to the number of distribution network nodes N, specifically 4N, and the maximum number of iterations is set to 8N to ensure that the search space is fully covered.

[0027] The workflow diagram of the adaptive genetic algorithm unit is as follows: Figure 2As shown in the figure, it includes a coding subunit, a fitness calculation subunit, a selection subunit, a crossover subunit and a mutation subunit. The coding subunit uses binary coding to convert the distribution network load feature vector set into a chromosome group. Each chromosome in the chromosome group represents a topological structure scheme. The chromosome length is equal to the total number of lines that may be connected in the distribution network. Each gene in the chromosome takes a value of 0 or 1, indicating that the line is not connected or connected, respectively. When encoding, the radiation constraints unique to the power system must be considered to ensure that the encoded topological structure meets the requirements of acyclicity.

[0028] The fitness calculation subunit constructs a dual-objective fitness function based on the power supply reliability index and the line loss rate. The power supply reliability index includes the power supply interruption time and the power supply interruption frequency. The line loss rate includes the active loss and the reactive loss. The dual-objective fitness function uses the weighted summation method for target combination. The weight ratio of the power supply reliability index to the line loss rate is 0.6:0.4. The selection subunit uses the roulette method to select high-quality chromosomes from the chromosome group. The selection probability is positively correlated with the fitness function value, and the selection probability is calculated by normalizing the fitness value.

[0029] The crossover subunit uses an adaptive crossover operator to reorganize high-quality chromosomes. The crossover probability of the adaptive crossover operator changes with the population evolution algebraically, specifically: ,in, , , g is the current generation, G is the maximum generation. The mutation subunit uses the adaptive mutation operator to perform gene mutation on the chromosome after crossover recombination. The mutation probability of the adaptive mutation operator changes with the value of the fitness function. Specifically, when the individual fitness value is greater than the average fitness of the population, ; When the individual fitness value is less than or equal to the average fitness of the population, ,in , , is the individual fitness value, is the maximum fitness value of the population, is the average fitness value of the population.

[0030] The generation process of topological structure candidate solutions includes the following steps: the coding subunit generates the initial chromosome group; the fitness calculation subunit calculates the fitness value of each individual; according to the individual fitness value, the selection subunit selects the high-quality chromosome through the roulette selection operation; the crossover subunit performs a single-point crossover operation on the selected chromosome according to the adaptive crossover probability; the mutation subunit mutates the chromosome according to the adaptive mutation probability; the fitness value of the new population is evaluated and the historical best individual is retained; it is determined whether the termination condition is met: if the maximum number of iterations is reached or the change in the optimal fitness value for 30 consecutive generations is less than 0.001, the algorithm is terminated and the topological structure corresponding to the optimal chromosome is output as a candidate solution; if the termination condition is not met, it returns to the selection subunit to continue iteration.

[0031] First, the topology optimization module innovatively dynamically associates the population size and the number of iterations with the number of distribution network nodes N, and adopts the configuration schemes of 4N and 8N, which not only ensures the coverage of the search space, but also avoids excessive consumption of computing resources. Secondly, in terms of encoding strategy, by considering the radiation constraints unique to the power system, it ensures that the topological structure meets the requirements of acyclicity and improves the feasibility of the solution. In terms of fitness evaluation, the power supply reliability index and line loss rate are used to construct a dual-objective fitness function, and the weight ratio of 0.6:0.4 reflects the importance of power supply reliability in actual engineering. The most distinctive feature is the design of the adaptive crossover mutation mechanism. The crossover probability is dynamically adjusted with the evolutionary generation, and the mutation probability is associated with the individual fitness value. This adaptive mechanism significantly improves the convergence performance of the algorithm and the quality of the solution. At the same time, a refined convergence condition of less than 0.001 change in the optimal fitness value for 30 consecutive generations is set to ensure the stability of the optimization results.

[0032] A deep reinforcement learning unit is set in the scheme evaluation module, and the deep reinforcement learning unit includes a deep Q network. The deep Q network quantitatively analyzes the economy, reliability, and environmental protection of the candidate topological structure scheme to obtain an evaluation score matrix. The deep Q network adopts a dual network structure, including a target network and an evaluation network. The parameters of the target network are updated every 100 iterations. The target network and the evaluation network use the same network structure, including 3 convolutional layers and 2 fully connected layers. The convolution kernel sizes of the convolution layers are 3×3, 5×5, and 7×7, respectively. Each layer uses 64 convolution kernels, and the number of neurons in the fully connected layers is 512 and 256, respectively.

[0033] Specifically, the three-layer structure of the deep Q network includes a state perception layer, an action evaluation layer, and a value calculation layer, such as Figure 3As shown. The state perception layer receives the topology candidate scheme and converts the topology candidate scheme into a state space matrix. The state space matrix includes node connection relationship, line capacity, and load distribution information. The state space matrix uses an adjacency matrix to represent the node connection relationship, a diagonal matrix to represent the line capacity, and a vector to represent the load distribution. The values ​​in the state space matrix are normalized using Min~Max to ensure that the value range is between [0,1]. At the same time, for the adjacency matrix of the node connection relationship, additional node degree information and shortest path information are introduced to enhance the expression of topological features.

[0034] The action evaluation layer constructs a reward function based on the state space matrix. The reward function consists of an economic evaluation sub-function, a reliability evaluation sub-function, and an environmental evaluation sub-function. The economic evaluation sub-function includes equipment investment cost and operation and maintenance cost, the reliability evaluation sub-function includes voltage deviation and power loss, and the environmental evaluation sub-function includes carbon emissions and land occupation. The weight of each evaluation sub-function is set by expert experience, and the weight ratio of economy: reliability: environmental protection is 0.4:0.4:0.2. When calculating the comprehensive reward value, a weighted summation method is used for combination.

[0035] Among them, the equipment investment cost is calculated according to the unit cost and length of the line; the operation and maintenance cost is estimated according to the number of equipment and maintenance frequency; the voltage deviation is obtained by flow calculation to obtain the deviation between the voltage of each node and the rated value; the power loss is calculated through network topology and load distribution; the carbon emissions are based on line loss and equipment life cycle assessment; and the land occupation is calculated according to the width and length of the line corridor.

[0036] The calculation results of the evaluation indicators need to be standardized, and the fuzzy measurement method is used to convert each indicator into a membership value in the interval [0,1]. For cost indicators (such as investment cost, operation and maintenance cost, etc.), a smaller fuzzy measurement is used, and for benefit indicators (such as power supply reliability, etc.), a larger fuzzy measurement is used. The standardized indicator values ​​are used as input features of the neural network.

[0037] The value calculation layer uses the experience replay mechanism to iteratively optimize the reward function. The experience replay mechanism includes priority sampling and temporal difference learning, and organizes the iteratively optimized reward values ​​into an evaluation score matrix. Among them, priority sampling extracts samples based on the size of the reward value, and uses a sum tree data structure to store priority values. The priority value is calculated according to the temporal difference error: ,in is the timing difference error, It is a small positive number with a value of 0.01. The probability calculation formula for priority sampling is: ,in is the priority index, with a value of 0.6, which is used to control the degree of priority sampling.

[0038] Temporal difference learning dynamically updates the value function using the following formula: ,in is the learning rate, with a value of 0.01, is the discount factor, with a value of 0.9, is the expected long-term return, s is the current state, a is the current action, r is the immediate reward, For the next state, is the possible action for the next state, is the best estimate of future returns, is the time series difference error. The row vectors of the evaluation score matrix correspond to different evaluation dimensions, and the column vectors of the evaluation score matrix correspond to different candidate solutions. The evaluation scores are normalized using min~max so that the scores of each dimension are normalized to the interval [0,1].

[0039] The training process of the deep Q network includes: first, initializing the experience replay pool, setting the capacity to 10,000, and initializing the parameters of the target network and the evaluation network. For each topology candidate, the training iteration process is performed: The greedy strategy selects actions, The initial value is 1.0; input the current state into the evaluation network and select the action corresponding to the maximum Q value; execute the selected action to obtain the reward and the next state; store the transfer sample in the experience replay pool; when the number of samples in the experience replay pool reaches the batch size of 32, perform priority sampling from the replay pool and perform training; update the target network parameters every 100 steps. Continue to execute the above training iteration process until the network converges or the maximum training round is 1000.

[0040] A fuzzy comprehensive evaluation unit is set in the decision output module. The fuzzy comprehensive evaluation unit prioritizes the topology candidate schemes based on the evaluation score matrix and outputs the optimal distribution network planning scheme. The work flow diagram is as follows: Figure 4 This unit uses the hierarchical analysis method to establish a set of discrimination criteria, which specifically includes three core components: the weight allocation subunit, the membership calculation subunit and the sorting subunit.

[0041] The judgment criterion system constructed by the weight allocation subunit using the hierarchical analysis method includes primary indicators and secondary indicators. The primary indicators cover economic indicators, reliability indicators and environmental indicators, and the secondary indicators correspond to the evaluation dimensions in the evaluation score matrix. In terms of weight allocation, the 9-level scaling method is used to construct the judgment matrix between the primary indicators. The judgment matrix elements of economy, reliability and environmental protection are [1,2,3; 1 / 2,1,2; 1 / 3,1 / 2,1] respectively, and the weight vector is calculated by the eigenvalue method. The weights of the secondary indicators are determined according to the expert scoring method: the economy is divided into equipment investment (0.6) and operation and maintenance cost (0.4); the reliability is divided into voltage deviation (0.5) and power loss (0.5); the environmental protection is divided into carbon emissions (0.7) and land occupation (0.3). All weight allocations must meet the consistency test requirements to ensure that the consistency ratio CR is less than 0.1.

[0042] The membership calculation subunit normalizes the evaluation score matrix and constructs the fuzzy relationship matrix using a trapezoidal membership function, which sets upper and lower thresholds based on expert experience.

[0043] For example, the trapezoidal membership function is set as follows: equipment investment cost is divided into excellent (≤5 million), good (5 million to 8 million), general (8 million to 12 million), and poor (>12 million); operation and maintenance cost is divided into excellent (≤500,000 / year), good (500,000 to 800,000 / year), general (800,000 to 1.2 million / year), and poor (>1.2 million / year); voltage deviation is divided into excellent (≤3%), good (3 to 5%), general (5 to 7%), and poor (>7%); power loss is divided into excellent (≤2%), good (2 to 3%), general (3 to 5%), and poor (>5%); carbon emissions are divided into excellent (≤100 tons / year), good (100 to 150 tons / year), general (150 to 200 tons / year), and poor (>200 tons / year); land occupation is divided into excellent (≤1000m 2 )、Good(1000~1500m 2 ), General (1500~2000m 2 ), poor (>2000m 2 ). Linear transition is used at the interval endpoints of the membership function, and the width of the transition interval is set to 10% of the interval length.

[0044] The sorting subunit uses the fuzzy synthesis operator to synthesize the weight vector of the discrimination criterion system and the fuzzy relationship matrix to obtain a comprehensive evaluation vector. The sorting subunit sorts the topology candidate schemes in descending order according to the element values ​​of the comprehensive evaluation vector and selects the candidate scheme with the highest score as the optimal distribution network planning scheme.

[0045] To ensure the reliability of the decision-making results, the system also sets up multiple guarantee mechanisms. For candidate topological structures with similar scores (difference less than 5%), the system will comprehensively consider factors such as the difficulty of engineering implementation and compatibility with the existing grid. At the same time, by adjusting the weight by ±10%, sensitivity analysis is performed to evaluate the stability of the decision results. The system has established a pool of alternative solutions. In addition to the optimal solution, the top three solutions with the highest comprehensive scores are retained as alternatives. In addition, the system sets a decision veto indicator. When a certain indicator exceeds the safety threshold, the solution will be directly rejected. The final selected solution needs to be confirmed by expert review and a detailed decision description document is formed.

[0046] Furthermore, this embodiment also provides an artificial intelligence-based power distribution network planning method, including obtaining historical load data and electricity demand data of the distribution network; using a deep convolutional neural network to receive the historical load data and electricity demand data of the distribution network and perform feature extraction to obtain a distribution network load feature vector set; using an adaptive genetic algorithm unit to receive the distribution network load feature vector set, and constructing a topology optimization model with power supply reliability index and line loss rate as a dual objective function to obtain a topology structure candidate scheme; using a deep Q network in a deep reinforcement learning unit to perform quantitative analysis on the topology structure candidate scheme to obtain an evaluation score matrix; using a fuzzy comprehensive evaluation unit to sort the topology structure candidate schemes based on the evaluation score matrix, and output the optimal distribution network planning scheme.

[0047] In summary, the present invention applies deep convolutional neural networks to extract load characteristics of distribution networks, and realizes accurate analysis of load data through multi-scale convolution kernels and progressive pooling structures; creatively introduces adaptive genetic algorithms in topology optimization, and effectively improves the optimization efficiency by designing dynamically adjusted crossover mutation operators and roulette-based selection mechanisms; uniquely combines deep Q networks with fuzzy comprehensive evaluation to construct a multi-dimensional evaluation system that includes economy, reliability and environmental protection, and improves the accuracy of scheme evaluation through priority sampling and temporal difference learning, while using hierarchical analysis method and trapezoidal membership function to realize scientific quantification of evaluation results.

[0048] Example 2, reference Figure 1~Figure 4 , which is the second embodiment of the present invention, provides an artificial intelligence-based power distribution network planning system. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0049] In order to verify the effectiveness of the artificial intelligence-based power distribution network planning system proposed in this paper, a new industrial park in a certain city development zone was selected as the experimental research object. The total planned land area of ​​the area is 5.8 square kilometers, including 28 load nodes, and the maximum power load is expected to reach 42MW. The system operating environment is Intel Xeon Gold 5220R CPU, 256GB memory, and 2 NVIDIA A5000 graphics cards.

[0050] In the data preprocessing stage, the electricity consumption data of the area for the past two years was collected, the sampling frequency was set to 5 minutes, and a total of about 210,000 data samples were collected. After processing by the filter preprocessing unit, the outlier detection rate was 1.8%, and the data smoothness was improved by 31.2%. The training of the deep convolutional neural network used a batch size of 32, a learning rate of 0.001, and the number of training rounds was set to 150. Finally, the mean square error on the validation set reached 0.0246, and the mean absolute percentage error was 4.85%.

[0051] In the topology optimization stage, the initial population size of the adaptive genetic algorithm was set to 112 (4×28) and the maximum number of iterations was 224 (8×28). After optimization, the algorithm reached convergence at 156 generations and finally obtained 12 representative topology candidate solutions. Among them, the average value of the power supply reliability index reached 99.91%, and the line loss rate dropped to 2.83%.

[0052] In the deep reinforcement learning evaluation phase, an experience replay pool with a capacity of 8000 was used, and the training batch size was 32. After 800 rounds of training, the loss function value of the evaluation network converged to 0.0125, and the standard deviation of the evaluation result was 0.0187. The comprehensive scoring of the 12 candidate solutions showed that the economic scores were distributed between 0.650.82, the reliability scores were distributed between 0.720.85, and the environmental scores were distributed between 0.68 and 0.81.

[0053] The performance comparison between the solution of the present invention and the prior art is shown in Table 1: Table 1 Performance comparison between the solution of the present invention and the prior art Evaluation indicators Traditional planning methods General intelligent optimization method The present invention Planning cycle 40 days 28 days 21 days Power supply reliability 99.85% 99.88% 99.91% Line loss rate 3.56% 3.12% 2.83% Investment cost reduction rate - 9.3% 11.6% Carbon emission reduction rate - 9.2% 13.5% Operation and maintenance cost reduction rate - 7.5% 10.8% From the comparative data in Table 1, it can be seen that compared with the traditional planning method and the ordinary intelligent optimization method, the scheme of the present invention has achieved significant improvements in various key indicators. The planning cycle was shortened by 47.5%, the power supply reliability was improved by 0.06 percentage points, and the line loss rate was reduced by 0.73 percentage points. In terms of economic benefits, the investment cost reduction rate reached 11.6%, which was 3.3 percentage points higher than the ordinary intelligent optimization method; in terms of environmental benefits, the carbon emission reduction rate reached 13.5%, an increase of 4.3 percentage points; in terms of operation and maintenance costs, the reduction rate reached 10.8%, an increase of 3.3 percentage points.

[0054] The optimal distribution network planning scheme finally determined adopts a single ring network and radial hybrid structure, and the main line adopts a cross-section of 185mm 2 The branch line adopts a cable with a cross-section of 95mm 2 The investment cost of this solution is 4.85 million yuan, the annual operation and maintenance cost is 426,000 yuan, the voltage deviation is controlled within 3.2%, the network power loss is 2.65%, the annual carbon emissions are 142 tons, and the land occupation area is 1,125 square meters. Through sensitivity analysis, the weight is within the range of ±10%, and the optimal solution remains stable. In the actual project implementation process, the solution passed the expert group review, and the preliminary design and construction drawing design have been completed.

[0055] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An artificial intelligence-based power distribution network planning system, characterized by: It includes data preprocessing module, topology optimization module, scheme evaluation module and decision output module, among which: The data preprocessing module is provided with a deep convolutional neural network, which receives the historical load data and power demand data of the distribution network and performs feature extraction to obtain a distribution network load feature vector set; The topology optimization module is provided with an adaptive genetic algorithm unit, which receives the distribution network load characteristic vector set, constructs a topology optimization model with a power supply reliability index and a line loss rate as a dual objective function, and obtains a candidate topology structure solution; The scheme evaluation module is provided with a deep reinforcement learning unit, and the deep reinforcement learning unit includes a deep Q network, and the deep Q network performs quantitative analysis on the topological structure candidate scheme to obtain an evaluation score matrix; The decision output module is provided with a fuzzy comprehensive evaluation unit, which ranks the topology candidate schemes based on the evaluation score matrix and outputs an optimal distribution network planning scheme.

2. The power distribution network planning system based on artificial intelligence as claimed in claim 1, characterized in that: The data preprocessing module further includes a filtering preprocessing unit, which performs abnormal value detection and smoothing processing on the historical load data and power demand data of the distribution network; The adaptive genetic algorithm unit generates the topology structure candidate solution through crossover and mutation operation; The deep Q network performs quantitative analysis on the economy, reliability, and environmental friendliness of the candidate topology solutions; The fuzzy comprehensive evaluation unit adopts the hierarchical analysis method to establish a set of discrimination criteria.

3. The power distribution network planning system based on artificial intelligence as claimed in claim 1, characterized in that: The deep convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer; The input layer receives the distribution network historical load data and the power demand data, and converts the distribution network historical load data and the power demand data into a standardized input matrix; The convolution layer uses a plurality of convolution kernels of different sizes to perform multi-scale feature extraction on the standardized input matrix, and the sizes of the convolution kernels are arranged in an increasing sequence; The pooling layer uses a maximum pooling operation to reduce the dimension of the features extracted by the convolutional layer, and the windows of the maximum pooling operation are multiples between adjacent layers; The fully connected layer flattens the feature map output by the pooling layer into a one-dimensional vector and performs nonlinear feature transformation through a ReLU activation function; The output layer performs dimensionality reduction compression on the one-dimensional vector output by the fully connected layer to form the distribution network load feature vector set.

4. The power distribution network planning system based on artificial intelligence as claimed in claim 1, characterized in that: The adaptive genetic algorithm unit includes an encoding subunit, a fitness calculation subunit, a selection subunit, a crossover subunit and a mutation subunit; The encoding subunit uses a binary encoding method to convert the distribution network load characteristic vector set into a chromosome group, and each chromosome in the chromosome group represents a topological structure scheme; The fitness calculation subunit constructs a dual-objective fitness function based on a power supply reliability index and a line loss rate, wherein the power supply reliability index includes a power supply interruption time and a power supply interruption frequency, and the line loss rate includes an active loss and a reactive loss; The selection subunit selects high-quality chromosomes from the chromosome population by a roulette wheel method, and the selection probability is positively correlated with the fitness function value; The crossover subunit reorganizes the high-quality chromosomes using an adaptive crossover operator, and the crossover probability of the adaptive crossover operator changes with the population evolution generation; The mutation subunit uses an adaptive mutation operator to perform gene mutation on the chromosome after crossover recombination, and the mutation probability of the adaptive mutation operator changes with the fitness function value to complete the generation of the topological structure candidate solution.

5. The power distribution network planning system based on artificial intelligence according to claim 1, characterized in that: The deep Q network includes a state perception layer, an action evaluation layer and a value calculation layer; The state perception layer receives the candidate topology scheme and converts the candidate topology scheme into a state space matrix, wherein the state space matrix includes node connection relationship, line capacity, and load distribution information; The action evaluation layer constructs a reward function based on the state space matrix, wherein the reward function is composed of an economic evaluation sub-function, a reliability evaluation sub-function and an environmental evaluation sub-function; The value calculation layer uses an experience replay mechanism to iteratively optimize the reward function, and organizes the iteratively optimized reward values ​​to form the evaluation score matrix; The experience replay mechanism includes priority sampling and temporal difference learning.

6. The power distribution network planning system based on artificial intelligence as claimed in claim 5, characterized in that: The economic evaluation sub-function includes equipment investment cost and operation and maintenance cost, the reliability evaluation sub-function includes voltage deviation and power loss, and the environmental evaluation sub-function includes carbon emissions and land occupation; The priority sampling extracts samples based on the size of the reward value, and the temporal difference learning dynamically updates the value function; The row vectors of the evaluation score matrix correspond to different evaluation dimensions, and the column vectors of the evaluation score matrix correspond to different candidate solutions.

7. The power distribution network planning system based on artificial intelligence according to claim 1, characterized in that: The fuzzy comprehensive evaluation unit includes a weight allocation subunit, a membership calculation subunit, and a sorting subunit; The weight allocation subunit adopts the hierarchical analysis method to construct a discrimination criterion system, wherein the discrimination criterion system includes primary indicators and secondary indicators, wherein the primary indicators include economic indicators, reliability indicators and environmental indicators, and the secondary indicators correspond to the evaluation dimensions in the evaluation score matrix; The membership calculation subunit normalizes the evaluation score matrix and constructs a fuzzy relationship matrix using a trapezoidal membership function, wherein the trapezoidal membership function sets upper and lower thresholds based on expert experience; The sorting subunit performs a synthesis operation on the weight vector of the discrimination criterion system and the fuzzy relationship matrix through a fuzzy synthesis operator to obtain a comprehensive evaluation vector; The sorting subunit sorts the topology structure candidate solutions in descending order according to the element values ​​of the comprehensive evaluation vector, and selects the candidate solution with the highest score as the optimal distribution network planning solution.

8. An artificial intelligence-based power distribution network planning method, based on the artificial intelligence-based power distribution network planning system according to any one of claims 1 to 7, characterized in that: Also includes, Obtain historical load data and electricity demand data of the distribution network; Using a deep convolutional neural network to receive the distribution network historical load data and the power demand data and perform feature extraction to obtain a distribution network load feature vector set; Adopting an adaptive genetic algorithm unit to receive the distribution network load characteristic vector set, constructing a topology optimization model with a power supply reliability index and a line loss rate as a dual objective function, and obtaining a candidate topology structure solution; A deep Q network in a deep reinforcement learning unit is used to quantitatively analyze the candidate topology solutions to obtain an evaluation score matrix; A fuzzy comprehensive evaluation unit is used to sort the topology structure candidate solutions based on the evaluation score matrix, and an optimal distribution network planning solution is output.

Images