Energy internet network topology optimization method and system based on multi-objective evolutionary algorithm

Abstract

The invention discloses an energy internet network topology optimization method and system based on a multi-objective evolutionary algorithm, and belongs to the technical field of energy internet, and the method comprises the steps of geometrically driving evolutionary operation, preferentially connecting adjacent nodes, strengthening a triangular structure, and optimizing a link layout in combination with an energy transmission efficiency weight. An NSGA-III algorithm is improved, a cosine similarity punishment mechanism and a dynamic weight strategy are introduced, and algebraic connectivity and an R index are balanced; a dynamic repair mechanism quickly complements isolated sub-graphs through local evolution and depth-first search, and the response time is shortened; a learning prediction topology adjustment strategy is reinforced, and a sliding window multiplexing historical solution is combined, so that the convergence speed under wind and light fluctuation is improved. According to the method, the problems of target singleness, uneven Pareto solution set distribution and dynamic response lag in the prior art can be solved, algebraic connectivity and node balance are improved, and the method can be widely applied to dynamic energy internet optimization scenes containing wind and light energy.

Description

Technical Field

[0001] The present invention relates to the technical field of energy Internet, and particularly relates to a method and system for optimizing the network topology of the energy Internet based on a multi-objective evolutionary algorithm, which is particularly applicable to the collaborative optimization of the robustness and node balance of the wind-solar output fluctuations and large-scale energy networks. Background Technique

[0002] With the continuous development of distributed energy technologies, the application of distributed energy such as photovoltaic and wind power in the energy system has gradually increased, and the optimization of the energy network topology has become the key to improving the system efficiency and reliability. The current non-dominated genetic algorithm for single-objective optimization mostly focuses on a single objective (such as maximizing algebraic connectivity), resulting in uneven node degree distribution, overloading of some nodes, and causing energy transmission bottlenecks. And multi-objective optimization algorithms such as the NSGA-III algorithm ignore the spatial distribution information of solutions in non-dominated sorting, resulting in the aggregation of Pareto front solutions and uneven distribution of solution sets, which cannot meet diverse requirements. In addition, static topology optimization has insufficient response in dynamic scenarios, takes a long time, even more than 2 seconds, and it is difficult to adapt to dynamic changes such as node movement and energy fluctuations (especially photovoltaic and wind power), resulting in network performance degradation.

[0003] Therefore, how to simultaneously improve the network robustness (algebraic connectivity) and node balance (R index) through multi-objective collaborative optimization in a dynamic and large-scale energy Internet, and achieve efficient real-time response has become a difficult problem that needs to be solved urgently. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for optimizing the network topology of the energy Internet based on a multi-objective evolutionary algorithm, which simultaneously optimizes the algebraic connectivity (network robustness) and node balance (R index), improves the distribution of the Pareto solution set through geometric-driven evolution and dynamic weight strategy, combines reinforcement learning and sliding window optimization, shortens the response time, and solves the technical problems of single objective, uneven distribution of Pareto solution sets, and lag in dynamic response existing in the existing energy Internet network topology optimization.

[0005] To achieve these purposes and other advantages of the present invention, the present invention provides a method for optimizing the network topology of the energy Internet based on a multi-objective evolutionary algorithm, including the following steps:

[0006] 1) Generate an initial adjacency matrix and calculate node coordinates based on geometric distance;

[0007] 2) Perform geometric-driven evolution operations, including:

[0008] In the mutation operation, preferentially connect node pairs with geometric distances less than the threshold, and strengthen the triangle structure;

[0009] Select parental gene fragments according to the weight of energy transfer efficiency during the crossover operation;

[0010] 3) Improve the NSGA-III algorithm, calculate the cosine similarity between solution vectors, and impose a fitness penalty on solutions with a similarity > 0.8; and dynamically adjust the algebraic connectivity C in stages algebraic and the weight of the node balance R index ;

[0011] 4) Detect isolated subgraphs through depth-first search and generate a minimum spanning tree to complete the connection; when the node energy value is lower than the threshold, disconnect its high-energy-consuming links, and select an alternative path based on the weight of energy transfer efficiency; perform local evolution operations on the affected subgraphs;

[0012] 5) Predict the topology adjustment strategy through the DQN model, and train the DQN model through the reward function R = λ1ΔC algebraic + λ2ΔR index - λ3ΔE loss to output the node addition and deletion priority list;

[0013] 6) Retain the historical solutions that satisfy the algebraic connectivity C algebraic ≥ 0.8 × the current optimum and the node degree balance R index ≥ 0.7 × the current optimum in the last K iterations as the initial population to accelerate convergence;

[0014] 7) Output the optimization results and generate the Pareto front solution set.

[0015] Preferably, in step S2, the generation of the triangular structure specifically includes:

[0016] If nodes A, B, and C satisfy the collinear condition d AB + d BC = d AC , then preferentially connect A - B - C to form a triangle;

[0017] Assign a clustering weight to each triangle: C w = 1 / (d AB + d BC + d AC ), and the higher the weight, the greater the probability that the corresponding triangle is selected by the mutation operation.

[0018] Preferably, a fault-tolerant triangle generation rule is also set;

[0019] The triangle generation condition is relaxed to: d AB + d BC ≤ d AC + ε, ε = 0.1 × d AC ;

[0020] The clustering weight is dynamically calculated as: C w = 1 / (d AB + d BC + d AC + γ × ε), where γ = 0.5;

[0021] The higher the weight, the greater the generation probability in the mutation operation.

[0022] Preferably, in the energy Internet network topology optimization method based on the multi-objective evolutionary algorithm, in step S3, the time threshold T of the dynamic weight adjustment mechanism is the total number of evolutionary iterations, specifically:

[0023] When t < T / 3, the algebraic connectivity weight is 0.7 and the R-index weight is 0.3;

[0024] When T / 3 ≤ t < 2T / 3, both weights are 0.5;

[0025] When t ≥ 2T / 3, the algebraic connectivity weight is 0.3 and the R-index weight is 0.7.

[0026] Preferably, a dynamic weight adaptive adjustment mechanism is also set up, including:

[0027] Real-time monitoring of the change rates ΔC algebraic of the algebraic connectivity C index and the node balance R algebraic , ΔR index ;

[0028] If ΔC algebraic continuously drops by more than the threshold θ1 for K consecutive iterations, the weight stage is switched in advance;

[0029] If the variance of ΔR index exceeds θ2, the current stage is extended until ΔR index is stable.

[0030] Preferably, in the energy Internet network topology optimization method based on the multi-objective evolutionary algorithm, in step S4, the local evolutionary operation includes:

[0031] Parallelly optimizing subgraphs through distributed computing nodes, and retaining the link with the highest energy transmission efficiency during merging;

[0032] For the uncovered nodes, use the greedy algorithm to complete the shortest path.

[0033] An intelligent microgrid energy scheduling system based on dynamic network topology optimization provided by the present invention includes:

[0034] Multi-source energy access module: integrating photovoltaic, wind power, energy storage batteries and grid interfaces, and nodes are connected to the microgrid DC bus through adaptive connectors;

[0035] Real-time data acquisition module: Collect power generation, energy storage SOC, load demand, and environmental parameters;

[0036] Dynamic topology optimization controller: Invoke any of the energy Internet network topology optimization methods based on the multi-objective evolutionary algorithm to generate an optimal adjacency matrix in real time, and dynamically adjust the connection priority according to the energy storage SOC:

[0037] When SOC < 30%, disconnect the energy storage battery from the high-load link and preferentially connect to renewable energy;

[0038] When SOC > 80%, use the energy storage battery as a voltage-stabilizing node and connect it to critical loads;

[0039] Edge computing acceleration unit: Perform distributed parallel computing on the microgrid subgraph through the GraphSAGE model to predict the objective function value of local topology optimization.

[0040] Preferably, exchange the topology optimization results with the adjacent microgrid controller through the 5G communication module to achieve cross-microgrid energy collaborative scheduling;

[0041] In the scenario of sudden change in photovoltaic output or sudden increase in load, the system completes topology reconstruction within 200 ms, shortening the power supply interruption time.

[0042] Preferably, in the edge computing acceleration unit, the training data of the GraphSAGE model includes:

[0043] The adjacency matrix of the historical topology structure, node energy vectors, algebraic connectivity, and label data of the R index;

[0044] Use the mean square error loss function for supervised training to reduce the prediction error.

[0045] The present invention has at least the following beneficial effects:

[0046] 1. The optimization method of the present invention, through the improved NSGA-III algorithm, synchronously improves the robustness of the energy network and the node balance, and solves the problem of node overload caused by traditional single-objective optimization.

[0047] 2. The method of the present invention introduces a cosine similarity penalty mechanism and a dynamic weight strategy, which improves the uniformity of the Pareto front solution set distribution and meets the diverse needs in the scenario of wind and light output fluctuations.

[0048] 3. The method of the present invention, through the combination of reinforcement learning (DQN) and local incremental optimization, shortens the response time to within 200 ms, adapting to the real-time adjustment requirements of node movement and energy fluctuation.

[0049] 4. The method of the present invention preferentially connects low-loss links through geometric-driven evolution to improve the energy transmission efficiency, and at the same time ensures the network connectivity through a dynamic repair mechanism.

[0050] Other advantages, objectives, and features of the present invention will be partially reflected by the following description, and partially will be understood by those skilled in the art through the research and practice of the present invention. Brief Description of the Drawings

[0051] Figure 1 : Pareto front comparison diagram (NSGA-II, standard NSGA-III vs improved NSGA-III of the present invention);

[0052] Figure 2 : Comparison curve of dynamic topology reconstruction response time;

[0053] Figure 3 : Schematic diagram of triangle structure generation (collinear condition of nodes A-B-C);

[0054] Figure 4 : Flow chart of dynamic repair and incremental optimization;

[0055] Figure 5 : Comparison of iteration times vs objective function value curves. Detailed Description of the Preferred Embodiments

[0056] The following further describes the present invention in detail with reference to the embodiments, so that those skilled in the art can implement it according to the description in the specification.

[0057] Introduction to the R index:

[0058] The calculation method of the node balance R index is as follows:

[0059]

[0060] where n is the total number of nodes in the network, and LCS d represents the size of the largest connected subgraph of the remaining network after deleting the d-th highest-degree node (i.e., the number of nodes included in the largest connected subgraph). The normalization factor makes it easy to compare networks with various sizes and edge densities. For example, in a fully connected network, each node in the network is linked to n - 1 other nodes, and deleting any node will not affect the communication between nodes. At this time, R = 0.5. In a star topology, the failure of the central node (the highest-degree node) will cause the entire network to split into isolated nodes, and at this time, R = 0. For other general networks, 0 ≥ R ≤ 0.5. The larger the value, the more balanced the node degree distribution and the stronger the network fault tolerance. This index needs to be calculated based on the entire network topology information, which can effectively quantify the node load balance and avoid transmission bottlenecks caused by a single high-load node.

[0061] As shown Figures 1 to 5 in the figure, a method for optimizing the network topology of the energy Internet based on a multi-objective evolutionary algorithm according to the present invention includes the following steps:

[0062] S1. Network topology initialization:

[0063] Generate an initial network adjacency matrix based on the BA model or import external network topology data; provide a network structure basis for geometric-driven evolution and node coordinate data for geometric distance calculation;

[0064] Exemplarily, generate a scale-free network based on the Barabási-Albert model, initially including 3 fully connected core nodes, and add new nodes according to the preferential connection probability P(k i ) = k i / ∑ j k j to join the network, form an initial network of 100 nodes, generate an adjacency matrix A N×N , and generate node coordinate data.

[0065] Or import through external topology: support actual microgrid topology data (such as IEEE standard nodes), parse it into the adjacency matrix format, and extract node coordinate data (x i , y i ).

[0066] S2. As shown Figure 3 in the figure, the geometric-driven evolution operation includes:

[0067] During the mutation operation, preferentially select node pairs with a geometric distance less than the threshold d for connection and strengthen the generation of triangular structures. Such triangular structures can provide redundant paths in the energy Internet, enhancing network reliability and fault tolerance.

[0068] Specifically, calculate the Euclidean distance between node pairs Furthermore, preferentially connect node pairs that satisfy d ij < threshold d, for example, d = 0.2 × D, where D is the network diameter;

[0069] Strengthen the triangular structure: If the node triple A - B - C satisfies the collinearity condition d AB + d BC = d AC , assign a clustering weight C w = 1 / (d AB + d BC + d AC ), and allocate the mutation probability according to C w (the higher the weight, the smaller the sum of geometric distances, the closer the nodes, and the greater the probability of being selected for mutation during the evolution process).

[0070] For example Figure 3 As shown, the algorithm preferentially adds an edge 2-3 to the 2-1-3 node triple to form a highly reliable loop; node 4 is not connected yet due to its relatively long distance.

[0071] The smaller the geometric distance, the stronger the spatial compactness between nodes, which means lower cable / line construction costs for short-distance connections; power transmission loss decreases with the shortening of distance, resulting in higher energy efficiency, forming a subnet with a high clustering coefficient in the short-distance area and enhancing local reliability.

[0072] In the crossover operation, the parental gene segments are selected according to the energy transfer efficiency weight W ij between nodes;

[0073]

[0074] where W ij is the energy transfer efficiency weight, d ij is the geometric distance between nodes i and j, and E loss is the link energy loss.

[0075] Specifically, first sort the links in descending order of weight; for example Figure 3 in, link 1-4(43.48)>3-4(29.33)>1-2(32.47)>1-3(23.87)>2-4(20.45)>2-3(13.77), long-distance low-loss links (such as 1-4) are preferentially retained.

[0076] Select the Top 20% of the high-weight links; total number of links = 6, Top 20% ≈ 1.2, take the first 2: for example, 1-4 (weight 43.48), 3-4 (weight 29.33).

[0077] The crossover operation generates offspring; parental network selection, select two parents P1 and P2 from the population; force the retention of gene segments: the offspring must contain the Top 20% of the links (1-4 and 3-4); supplement the links: randomly select the remaining links from the parents to ensure connectivity.

[0078] For example, parental P1 links: {1-4, 3-4, 1-2, 2-4}; parental P2 links: {1-4, 3-4, 1-3, 2-3};

[0079] Offspring C generation: force the retention of 1-4, 3-4;

[0080] Random supplementation: select other links from P1 or P2 (such as selecting 1-2 and 2-4), and the final link set is {1-4, 3-4, 1-2, 2-4}.

[0081] S3. Improved NSGA-III Selection Strategy: including Step S31 and Step S32;

[0082] S31. For all solution pairs (i, j) within the population, calculate the cosine similarity S between the obtained solution vectors ij , if S ij > 0.8, it is considered that the two solutions are too similar in the objective space, then impose a penalty F on the fitness new = F original ·(1 - αS ij ); F original Original fitness value, α is the penalty coefficient; it can take values from 0.05 - 0.2 to dynamically control the dispersion degree of the solution set;

[0083] Specifically, S ij Cosine similarity between solution vectors, and the calculation method is:

[0084]

[0085] where x = [C algebraic , R index , is the solution vector containing the algebraic connectivity C algebraic and the R index R index ;

[0086] Dot product calculation:

[0087] Magnitude calculation:

[0088] Traditional NSGA-III relies on crowding distance to maintain diversity, but in a high-dimensional objective space (such as multi-objective optimization), the crowding distance may not be able to effectively distinguish the distribution differences of solutions, resulting in the aggregation of the solution set. In this embodiment, through cosine similarity penalty, similar solutions are forced to reduce their fitness, prompting the solution set to be evenly distributed on the Pareto front.

[0089] Exemplarily, assume two solution vectors:

[0090] Solution x i :

[0091] Solution x j :

[0092] Calculation steps:

[0093] Dot product: x i ·x j = (0.8 × 0.75) + (0.6 × 0.65) = 0.6 + 0.39 = 0.99

[0094] Magnitude:

[0095] ∥x i ∥=0.82 + 0.62=1.0, ∥x j ∥=0.752 + 0.652≈0.997;

[0096] Cosine similarity:

[0097] S ij =0.991.0×0.997≈0.993;

[0098] Fitness penalty (taking α = 0.1):

[0099]

[0100] The solution x i has its fitness reduced, reducing its competitive advantage in the population, thus encouraging diversity.

[0101] S32. Dynamically adjust the target algebraic connectivity C algebraic and the node balance R index Weights:

[0102] For example, in the initial stage, the weight of algebraic connectivity is 0.7 and the weight of the R index is 0.3, so as to quickly improve the robustness;

[0103] In the middle stage, the weights are both 0.5, achieving the effect of balanced optimization;

[0104] In the later stage, the weight of algebraic connectivity is 0.3 and the weight of the R index is 0.7, achieving the purpose of refined node balance.

[0105] This step solves the problem that traditional NSGA-III ignores the solution distribution in non-dominated sorting, avoids the solution set being biased towards a single objective due to fixed weights, and balances multi-objective conflicts through dynamic weights.

[0106] S4. As Figure 4 shown in the flowchart of dynamic repair and incremental optimization, dynamic repair and incremental optimization specifically include:

[0107] S41. Traverse the network topology through depth-first search (DFS) to detect isolated subgraphs (i.e., isolated regions that cannot reach other nodes through paths). For each isolated subgraph, generate a minimum spanning tree (MST) and complete the least number of high-energy efficiency links to ensure full network connectivity;

[0108] Specifically, DFS traversal to detect isolated subgraphs includes:

[0109] Input: Network adjacency matrix (including node 1-4 coordinate data).

[0110] Output: Identify isolated subgraphs (such as partitions caused by node failures).

[0111] Example: Assume that links 1 - 4 are broken, node 4 becomes an isolated node, and the isolated sub - graph {4} is detected.

[0112] Generate a minimum spanning tree (MST) to complete the links, including:

[0113] MST construction rule: Based on the energy - efficiency weight Select edges.

[0114] Completion strategy: Add the minimum number of high - energy - efficiency links to connect the isolated sub - graph to the main network.

[0115] Example: Possible connection links between the isolated sub - graph {4} and the main network {1, 2, 3}: 1 - 4 (W 1-4 = 43.48); 3 - 4 (W 3-4 = 29.33); 2 - 4 (W 2-4 = 20.45).

[0116] Select the highest - energy - efficiency link: Add 1 - 4 (weight 43.48) to restore global connectivity and avoid partitioning due to node / link failures.

[0117] S42. When the node energy value is lower than the threshold, disconnect the high - energy - consumption links of the low - energy nodes, and preferentially select the link with the highest energy - transfer efficiency to complete the alternative path;

[0118] For example, when the node energy E i < 30% of the rated capacity (such as when the state of charge (SOC) of the energy storage battery is too low), automatically disconnect the high - energy - consumption links it is connected to (links with loss E loss > 15%);

[0119] Example: Links of node 2: 2 - 1 (E loss = 3.08%), 2 - 3 (E loss = 7.26%), 2 - 4 (E loss = 4.89%). No link meets the disconnection condition (all E loss < 15%), so the disconnection operation is not performed. For example: If the Eloss of link 2 - 3 = 16%, then disconnect 2 - 3.

[0120] Based on the energy - transfer efficiency weight W ij , select the link with the highest energy - efficiency to reconnect.

[0121] Example (assuming 2 - 3 is disconnected): Alternative links available: 2 - 1 (W = 32.47), 2 - 4 (W = 20.45); Select the 2 - 1 link (higher weight) to ensure efficient energy transfer of node 2.

[0122] S43. In response to the change in node status, only perform local evolution operations on the affected subgraphs (such as nodes within 3 hops centered around the faulty node).

[0123] Specifically, it includes:

[0124] Define the affected subgraph: Centered around the faulty node, extract the nodes and links within 3 hops.

[0125] Example: If node 3 fails, the nodes within 3 hops are {1, 2, 4} (links 1 - 3, 2 - 3, and 3 - 4 all fail).

[0126] Local evolution operation: Only perform mutation and crossover on the subgraph {1, 2, 4}.

[0127] Optimization objective: Maximize the algebraic connectivity C algebraic and the node balance R index .

[0128] The example optimization process includes:

[0129] Mutation: Preferentially connect high - energy - efficient links (such as 1 - 4) within the subgraph.

[0130] Crossover: Retain the high - energy - efficient gene fragments of the parent generation (such as 1 - 4 and 2 - 4).

[0131] Generate offspring: The new topology {1 - 2, 1 - 4, 2 - 4}, ensuring that the subgraph is connected and the energy efficiency is improved.

[0132] Global merge: Merge the optimized subgraph with the unaffected part of the main network to form the final topology.

[0133] As Figure 2 shown, it is the comparison curve of the dynamic topology reconstruction response time of different algorithms, where nsga3optimize is the improved NSGA - III method of the present invention, nsga2 is the traditional NSGA - II, and nsga3 is the standard NSGA - III algorithm. Through dynamic repair and incremental optimization, the present invention realizes the real - time adaptive adjustment of the energy Internet topology. While ensuring network connectivity and energy efficiency, it significantly reduces the computational complexity, provides an efficient solution for dynamic scenarios such as wind and solar power output fluctuations and node movement, and solves the problem of network performance degradation in dynamic scenarios.

[0134] S5. Reinforcement learning - assisted decision - making:

[0135] Predict the topology adjustment strategy through the DQN (Deep Q - Network) model and output the node addition and deletion priority list;

[0136] The state space is defined as the adjacency matrix of the current topology and the node energy vector, including:

[0137] Input adjacency matrix A ∈ {0, 1} N×N : representing the node connection relationship (1 for connected, 0 for disconnected);

[0138] Input node energy vector E ∈ R N : the real-time energy value of each node (such as PV output, energy storage SOC).

[0139] Example: The adjacency matrix of a 54-node network is a 54×54 sparse matrix, and the node energy vector E = [0.85, 0.92, 0.45, …, 0.78] (normalized value, 1 represents the rated capacity); assuming the energy of node 3 is 45% (the PV output may decrease due to insufficient light), feature extraction: using a graph convolutional network (GCN), encoding the adjacency matrix and node features (energy, geographical location) into a low-dimensional vector to capture the joint features of topology and energy; output state vector: st ∈ R 128 (feature representation after compression by GCN).

[0140] The action space is a binary operation of node connection / disconnection;

[0141] Specifically, the action space (Action): is defined as a binary operation, performing connection (1) or disconnection (0) on candidate node pairs, such as "connect node 1-2" or "disconnect node 3-4". To reduce the dimension of the action space, only consider node pairs that meet the following conditions: geometric distance < dynamic threshold d (such as d = 0.2×D, D is the current network diameter); energy transfer efficiency W ij The top 20% of the links (disconnecting inefficient links or complementing efficient links).

[0142] Example, the action space for a 54-node network:

[0143] Candidate node pairs: Approximately 200 pairs are retained through screening (the full combination C(54, 2) = 1431 pairs, reducing by 86%).

[0144] Action encoding: a t = [connect 1-4 , disconnect 2-3 ,…] (one-hot encoding).

[0145] The reward function is:

[0146] R = λ1ΔC algebraic + λ2ΔR index - λ3ΔE loss

[0147] ΔC algebraic It is the change in algebraic connectivity (after action execution - before execution), which measures the network robustness;

[0148] ΔR index It is the change in node degree balance. The larger the value, the more balanced;

[0149] ΔE loss It is the change in energy loss (after action execution - before execution);

[0150] λ1, λ2, and λ3 are weight coefficients. For example, the initial weight coefficients satisfy λ1:λ2:λ3 = 3:2:1, and it is preferred to ensure robustness;

[0151] After every 100 iterations, the distribution range of each objective in the Pareto front solution set is statistically analyzed, and the weights are adjusted inversely proportional to the objective variances.

[0152] For example, if the variance of C in the front solution set algebraic is small (tending to converge), then reduce λ1 and increase λ2.

[0153] S54. Decision-making process: DQN learns how to select actions that can maximize the long-term reward according to the current topology and energy state (such as preferentially disconnecting high-loss links and connecting to efficient paths) through training; it outputs a list of node addition and deletion priorities to guide real-time topology adjustment.

[0154] Specifically, the training process includes:

[0155] Experience replay: Store state transition samples (st, at, rt, st+1) to break data correlation.

[0156] Dual-network structure:

[0157] Online network (Q-network): Update the policy and output Q values.

[0158] Target network (Target Q-network): Stabilize the training and synchronize parameters regularly.

[0159] Loss function:

[0160]

[0161] θ is the parameter of the online network, θ - is the parameter of the target network, γ = 0.99 is the discount factor, r is the immediate reward, st is the state at the current moment, at is the action (Action) executed at the current moment, st+1 is the state at the next moment (NextState), Q(s,a;θ) is the Q value predicted by the online network (action value function), and maxa′Q(st+1,a′;θ-) is the maximum Q value prediction of the target network for the next state.

[0162] For example, when the wind and light fluctuate, the energy of Node 3 drops to 45%, and the E of its connection link 3-4 loss increases to 18%. State input: Update the adjacency matrix (disconnect 3-4) and the node energy vector. The DQN decision includes: Feature extraction: GCN encodes the current state st. Q-value prediction: Output the Q-values of all candidate actions. Action selection (ε-greedy policy, ∈ = 0.1): Select the action with the maximum Q-value with a 90% probability: Connect 1-4 (W 1-4 = 43.48); Explore randomly with a 10% probability: Connect 2-4 (W 2-4 = 20.45). Execute the action: Add the link 1-4 and update the topology. Reward calculation: ΔC algebraic = +0.1 (robustness improved). ΔR index = +0.05 (more balanced node degree). ΔE loss = -0.08 (reduction of the total network loss).

[0163] R = 3×0.1 + 2×0.05 - 1×(-0.08) = 0.48. Model update: Store the samples in the experience pool and perform batch training every 10 steps.

[0164] Among them, for example, if the Q-value of the action "connect 1-4" is high, it means that this action can improve the robustness, balance, and energy efficiency in the long run. The specific example is as follows. As mentioned above, the sudden change of wind and light causes the energy of Node 3 to drop, and the topology needs to be adjusted. The steps are as follows:

[0165] 1. Current state st: The energy of Node 3 = 45%, and the link 3-4 is disconnected in the adjacency matrix.

[0166] 2. Action at: Select "connect 1-4" (with the highest Q-value).

[0167] 3. Reward r:

[0168] ΔC algebraic = +0.1, ΔR index = +0.05, ΔE loss = -0.08

[0169] R = 3×0.1 + 2×0.05 - 1×(-0.08) = 0.48.

[0170] 4. Next state st+1: After connecting 1-4, the algebraic connectivity of the entire network is increased to 0.65.

[0171] 5. Loss calculation:

[0172] L(θ) = (0.48 + 0.99×5.8 - 5.2) 2=(0.48 + 5.742 - 5.2) 2 ≈1.02 2 ≈1.04

[0173] Update θ through backpropagation to make the online network more accurately predict the Q value.

[0174] Due to the high dimensions of the adjacency matrix and the energy vector, the DQN in this step can effectively extract features; when the wind and light output fluctuate and the node energy changes, the DQN quickly adjusts the strategy through online learning, making the response time < 200ms; the reward function integrates robustness, balance, and energy efficiency goals, and the DQN automatically learns the balance strategy of the three, avoiding the limitations of manual rule design.

[0175] S6. Sliding time window optimization:

[0176] Retain the historical high-quality solutions as the initial population, with the window size being the solution sets of the most recent K iterations, and the screening conditions are:

[0177] Algebraic connectivity C algebraic ≥0.8 × current optimal value;

[0178] Node degree balance R index ≥0.7 × current optimal value.

[0179] Specifically, it includes the following steps:

[0180] 1. Define the sliding window: Retain the solution sets of the most recent K iterations (e.g., K = 10).

[0181] 2. Screen high-quality solutions: Select the solutions that simultaneously meet the following conditions from within the window:

[0182] Algebraic connectivity: C algebraic ≥0.8 × C current_best

[0183] Node degree balance: R index ≥0.7 × R current_best

[0184] Merge the populations: The historical solutions selected account for 30% - 50%, and are merged with the current population to form the initial population of the next generation.

[0185] Priority improvement: The weight coefficient of the historical solutions is increased to 1.2 times during the crossover and mutation operations.

[0186] For example, when an energy Internet is optimized to the 50th generation, the current optimal solution C current_best = 0.9, R current_best = 0.8.

[0187] Sliding window (K = 10): Keep the solution sets of the last 10 generations, a total of 200 solutions (population size of 20 per generation).

[0188] Screening conditions:

[0189] C algebraic ≥0.8 × 0.9 = 0.72;

[0190] R index ≥0.7 × 0.8 = 0.56;

[0191] Screening results: 30 solutions meet the conditions (accounting for 15%).

[0192] Merger operation: Current population (20) + historical solutions (6, at a ratio of 30%) = 26, truncated to 20.

[0193] Priority effect: The probability of historical solutions being selected during crossover is increased, and high-quality genes are transmitted faster.

[0194] In this way, the convergence speed is increased, the coverage rate of the Pareto front is improved, which is especially suitable for the long-term operation optimization of large-scale energy Internet, and can significantly reduce the consumption of computing resources and improve the topology adjustment efficiency.

[0195] S7. Output the optimization result: Generate the Pareto front solution set.

[0196] In this embodiment, the bi-objective optimization of algebraic connectivity (network robustness) and node balance (R index) is achieved through the improved NSGA-III algorithm. Compared with the traditional NSGA-II that only relies on crowding distance to maintain diversity, this method introduces a cosine similarity penalty mechanism (see step S3). For example, when the similarity between solution vectors > 0.8, a dynamic penalty coefficient (α = 0.05 - 0.2) is imposed on the fitness, effectively avoiding the problem of Pareto front solution aggregation. At the same time, a phased dynamic weight adjustment strategy is adopted: focusing on algebraic connectivity in the initial stage (weight 0.7), balanced optimization in the middle stage (weight 0.5), and focusing on node balance in the later stage (weight 0.7), adapting to the focus of different optimization stages.

[0197] As Figure 1 shown is the Pareto front comparison chart (NSGA-II, standard NSGA-III vs the improved NSGA-III of the present invention).

[0198] Among them, the C of the traditional NSGA-II algebraic stagnates at about 0.4, and the C of the standard NSGA-III algebraic is only about 0.55. For the improved NSGA-III of the present invention: In the initial stage (iterations 0 - T / 3): The dynamic weight w1 = 0.7 focuses on robustness, and geometrically driven evolution preferentially connects adjacent nodes and generates triangles( Figure 3), C algebraic Quickly climbs to around 0.42. In the later stage (iteration ≥ 2T / 3): The weight is reduced to w1 = 0.3, but the historical solutions are reused through a sliding window (high C algebraic solutions), and the curve still slowly rises to around 0.55.

[0199] The R of the traditional NSGA-II index The final value is less than 0.16, and the final value of the standard NSGA-III is also less than 0.18. For the improved NSGA-III of the present invention: In the initial stage (iteration 0 to T / 3): R index grows slowly (weight w2 = 0.3), but the cosine similarity penalty mechanism avoids the aggregation of the solution set and maintains diversity. In the later stage (iteration ≥ 2T / 3): The weight is increased to w2 = 0.7, combined with the node priority adjustment of reinforcement learning (DQN) (step S5), disconnect the inefficient links of the high-load nodes, and R index jumps to nearly 0.22, achieving load balancing.

[0200] Further, in step S2, as Figure 3 shown, the generation of the triangular structure specifically includes:

[0201] If nodes A, B, and C satisfy the collinear condition d AB + d BC = d AC , then preferentially connect A - B - C to form a triangle;

[0202] Assign a clustering weight to each triangle: C w = 1 / (d AB + d BC + d AC ), the higher the weight, the greater the mutation probability and the greater the probability of being selected for generation. For example, if nodes A - B - C satisfy the collinear condition and their total distance d AB + d BC + d AC is large, resulting in a low C w , thus reducing the generation probability of redundant long links.

[0203] The centralized generation of short-distance triangles in this embodiment improves the local transmission efficiency and reduces the global energy loss; the compact triangular structure reduces the dependence on key links, and the network survivability (algebraic connectivity) is significantly improved; cooperating with mechanisms such as sliding windows and local evolution, it realizes the real-time optimization of complex networks, and performs outstandingly in scenarios such as the energy Internet and 5G base stations, especially in dynamic environments with fluctuating wind and light output and frequent node movement.

[0204] Further, in another embodiment, a fault-tolerant triangular generation rule is also set;

[0205] The triangle generation condition is relaxed to: d AB +d BC ≤d AC +ε,ε=0.1×d AC ;

[0206] The clustering weight is dynamically calculated as: C w =1 / (d AB +d BC +d AC +γ×ε), γ=0.5;

[0207] The higher the weight, the greater the probability of generation in the mutation operation.

[0208] For example, when d AC =200km,ε=20km. If d AB =80km,d BC =110km, then d AB +d BC =190km≤200+20=220km, which meets the fault tolerance condition. Calculate C w =1 / (80+110+200+0.5×20)=1 / 400=0.0025, which increases the priority of the triangle in the mutation.

[0209] In a microgrid with dispersed wind and solar power stations, nodes are restricted by terrain and cannot be strictly collinear, so redundant loops cannot be formed. Failures can easily cause regional power outages. In addition, strict collinearity conditions exclude non-collinear but energy-efficient long-distance links (such as cross-regional photovoltaic-energy storage direct connections), resulting in increased global energy losses. This implementation allows ε = 0.1 × d AC The error is reduced to make more node combinations meet the conditions, tolerate measurement errors or node position fluctuations, and avoid priority mutations caused by small deviations. The γ×ε term (γ=0.5) is introduced to smooth the weight calculation, give priority to links with shorter total distances, and adapt to different network scales and energy efficiency requirements by adjusting ε and γ.

[0210] Furthermore, in step S3, the time threshold T of the dynamic weight adjustment mechanism is the total number of evolutionary iterations, specifically:

[0211] t <T / 3时,代数连通性权重为0.7,R指标权重为0.3;

[0212] When T / 3≤t<2T / 3, the weights of both are 0.5;

[0213] When t≥2T / 3, the algebraic connectivity weight is 0.3 and the R index weight is 0.7.

[0214] In the topological optimization of the energy Internet, there is usually a trade-off between algebraic connectivity (reflecting network robustness) and node balance (reflecting load distribution); maximizing algebraic connectivity tends to increase the degree of core nodes (Hub nodes), resulting in uneven load, and maximizing node balance, forcing uniform distribution of node degrees, may reduce the network's resilience. This embodiment divides the optimization process into three stages, and realizes the phased coordination between the objectives by adjusting the weight coefficients of algebraic connectivity and node balance, effectively balancing the multi-objective conflicts in the topological optimization of the energy Internet. As Figure 5 shown in the comparison of the curve of the number of iterations vs the objective function value.

[0215] Furthermore, in another embodiment, a dynamic weight adaptive adjustment mechanism is also set up, including:

[0216] Real-time monitoring of the change rate ΔC algebraic of algebraic connectivity C index and the change rate ΔR algebraic of node balance R index ;

[0217] If ΔC algebraic decreases by more than the threshold θ1 for K consecutive iterations, the weight stage is switched in advance;

[0218] If the variance of ΔR index exceeds θ2, the current stage is extended until ΔR index is stable.

[0219] Exemplarily, in the scenario of sudden changes in wind and light output, when ΔC algebraic decreases by 15% within 10 iterations (θ1 = 15%), the system automatically switches the weight stage from the initial stage (C algebraic weight 0.7) to the middle stage (C / R weight 0.5) to avoid overfitting of robustness optimization. At the same time, if the fluctuation of ΔR index exceeds θ2 = 0.1, the middle stage is extended by 5 iterations until R index steadily improves.

[0220] Weight allocation (such as the weights of algebraic connectivity C and node balance R) is usually fixed and switched according to preset stages (such as C weight 0.7 in the initial stage, 0.5 in the middle stage, and 0.3 in the later stage), and it cannot adapt to dynamic scenarios such as sudden changes in wind and solar power output and sudden increases in load. For example, when the wind and solar power output drops suddenly: if the C weight is still emphasized in the initial stage, it may lead to a lag in robust optimization and an inability to quickly respond to the risk of network splitting; when the node load fluctuates violently: the fixed stage switching may ignore the instantaneous imbalance of the R index and exacerbate local overload. In this embodiment, when C continuously decreases (such as the deterioration of robustness caused by sudden changes in wind and solar power output), the weight is immediately adjusted to balance multiple objectives and avoid overfitting. The ΔR variance trigger stage is extended: if R fluctuates too much (such as the imbalance of node degrees caused by load mutation), the current stage is extended until it is stable to prevent optimization oscillation.

[0221] Further, in step S4, the local evolution operation includes:

[0222] Parallelly optimize the subgraph through distributed computing nodes, and retain the link with the highest energy transmission efficiency when merging;

[0223] For the nodes not covered, use the greedy algorithm (prim algorithm) to complete the shortest path.

[0224] Specifically, the distributed parallel optimization of the subgraph includes the following steps:

[0225] Subgraph partitioning: Centered on the faulty node, extract the nodes within 3 hops to form a subgraph (ensuring local relevance).

[0226] Parallel optimization: Each computing node independently runs the improved NSGA-III algorithm to optimize the subgraph topology.

[0227] Merging strategy: Retain the cross-subgraph link with the highest energy transmission efficiency (W ij ) to ensure the optimal global energy efficiency.

[0228] The greedy algorithm to complete the uncovered nodes includes the following steps:

[0229] Isolated node identification: Traverse all nodes and mark the nodes not covered by the subgraph.

[0230] Shortest path selection: Based on the energy transmission efficiency W ij , select the locally optimal link hop by hop from the isolated node to the main network.

[0231] Path completion: Add links in sequence until the isolated node is connected to the main network.

[0232] Exemplarily, node 1 is not covered:

[0233] Isolated node: Node 1 (initial connection: 1-2 is disconnected).

[0234] Candidate connection:

[0235] Link 1-2 (W ij = 15.8), Link 1-3 (W ij = 22.4), Link 1-4 (W ij = 27.9)

[0236] Greedy selection: Preferentially connect 1-4 (W = 27.9).

[0237] Verify connectivity: Node 1 accesses Subgraph A through 1-4, and the whole network resumes connectivity.

[0238] This embodiment improves the efficiency and quality of the topology optimization of the energy Internet through a distributed parallel optimization and greedy completion mechanism.

[0239] An intelligent microgrid energy scheduling system based on dynamic network topology optimization provided by the present invention includes:

[0240] Multi-source energy access module: Integrates photovoltaic, wind power, energy storage batteries and grid interfaces. Nodes access the microgrid DC bus through adaptive connectors; the adaptive connectors dynamically adjust the access impedance and voltage matching according to the energy type and power characteristics (such as photovoltaic MPPT control, energy storage bidirectional DC / DC conversion).

[0241] Exemplarily, in the photovoltaic fluctuation scenario: At noon on a sunny day: The photovoltaic output suddenly increases to 120% of the rated power, and the adaptive connector automatically reduces the impedance and raises the bus voltage to 820V to preferentially consume photovoltaic power. In the evening on a cloudy day: The photovoltaic output drops to 30%, and the connector switches to the energy storage battery power supply mode to maintain the bus voltage stable at 800V.

[0242] Real-time data acquisition module: Collects power generation power, energy storage SOC, load demand and environmental parameters (temperature and humidity);

[0243] Dynamic topology optimization controller: Invokes the energy Internet network topology optimization method based on the multi-objective evolutionary algorithm of the present invention to generate the optimal adjacency matrix in real time (for example, once every 5 seconds), and dynamically adjusts the connection priority according to the energy storage SOC:

[0244] When SOC < 30%, disconnect the energy storage battery from the high-load link and preferentially connect renewable energy; Exemplarily, disconnect the energy storage - motor link and connect the photovoltaic - lighting load.

[0245] When SOC is 30% - 80%, balance charging and discharging and act as a frequency modulation node; Exemplarily, maintain the existing topology.

[0246] When SOC > 80%, connect the energy storage battery as a voltage stabilizing node to the critical load; Exemplarily, the energy storage is connected to the ICU power supply link, and THD < 1%.

[0247] Edge computing acceleration unit: Through the GraphSAGE model, it performs distributed parallel computing on the microgrid subgraph to predict the objective function value of local topology optimization.

[0248] Specifically, it includes:

[0249] The GraphSAGE model abstracts the microgrid as a graph structure G=(V, E). The characteristics of the node set V include power, SOC, load type, etc., and the characteristics of the edge set E are impedance and energy transfer efficiency weight (energy efficiency weight). The node characteristics (power, SOC, load type) are directly related to the real-time state of the microgrid, and the edge characteristics (impedance, energy efficiency weight) reflect the energy transfer loss. The combination of the two provides input data with clear physical meaning for GraphSAGE. The GraphSAGE model predicts the objective function value of local topology optimization by aggregating the characteristics of nodes and edges.

[0250] Distributed prediction:

[0251] 1. Subgraph segmentation: Divide the power supply area centered on the transformer (e.g., subgraph A: PV + energy storage, subgraph B: wind power + load).

[0252] 2. Parallel inference: Each edge node independently predicts C algebraic and E loss .

[0253] 3. Global aggregation: Integrate the subgraph prediction results to generate the network-wide optimization objective value.

[0254] Example (subgraph A optimization prediction):

[0255] Input features: PV output 0.8 MW, energy storage SOC = 65%, load demand 1.2 MW.

[0256] GraphSAGE output: Predicts connecting the PV - energy storage - load triangular link to improve C algebraic , reduce E loss .

[0257] Furthermore, in the described system, the dynamic topology optimization controller further includes:

[0258] Exchanges the topology optimization results with the neighboring microgrid controller through the 5G communication module to achieve cross-microgrid energy collaborative scheduling;

[0259] Example:

[0260] Microgrid A: Excessive PV output (1.2 MW), energy storage SOC = 95%.

[0261] Microgrid B: Surge in load (shortage 0.8 MW), energy storage SOC = 25%.

[0262] Coordinated action: Microgrid A sends the available power capacity (1.0 MW) and topological information (critical links 1-4, 2-5) via 5G. The controller of Microgrid B calculates the optimal cross-grid link (from B-3 to A-2, W ij = 38.7). Both parties synchronously update the topology, establish a cross-microgrid DC connection, and shorten the power supply interruption time to within 200 ms.

[0263] In the scenario of sudden change in photovoltaic output or sudden increase in load, the system completes topology reconstruction within 200 ms.

[0264] In this embodiment, based on historical data and reinforcement learning, topology plans for high-frequency scenarios (such as sudden change in photovoltaic power, sudden increase in load) are generated offline. After detecting a mutation event, the plan is directly called and fine-tuned online to avoid the time-consuming of full-process optimization.

[0265] Taking the sudden drop in photovoltaic output as an example:

[0266] Cloud cover causes the photovoltaic output to drop from 1.0 MW to 0.2 MW (lasting for 10 seconds). Match the "photovoltaic output < 30%" plan (disconnect non-critical loads, connect energy storage - core loads).

[0267] Reconstruction process: t = 0 ms: Detect a drop in output and trigger the plan. t = 50 ms: Disconnect the photovoltaic - air conditioner load link. t = 150 ms: Connect the energy storage - IT load link. t = 200 ms: Complete the reconstruction, and the power supply interruption time is shortened to 0.2 seconds.

[0268] Furthermore, in the system described above, in the edge computing acceleration unit, the training data of the GraphSAGE model includes:

[0269] The adjacency matrix, node energy vector, algebraic connectivity, and label data of the R index of the historical topological structure;

[0270] Supervised training is carried out using the mean square error loss function to reduce the prediction error.

[0271] Specifically, it includes the adjacency matrix A ∈ {0, 1} N×N . The node energy vector E ∈ R N (normalized to [0, 1]). Label data: algebraic connectivity C algebraic ∈ [0, 1]. Node balance R index ∈ [0, 1]. These historical operation data cover various scenarios such as photovoltaic, wind power, and energy storage.

[0272] Loss function: mean square error (MSE)

[0273]

[0274] Example prediction

[0275] Input:

[0276] Adjacency matrix: Ring topology (1 - 2 - 3 - 4 - 1).

[0277] Node energy: Photovoltaic node (0.9), energy storage node (0.6), load node (0.3).

[0278] Output:

[0279] C pred = 0.68 (true value 0.71, error 4.2%).

[0280] R pred = 0.62 (true value 0.65, error 4.6%).

[0281] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and embodiments. It can be fully applied to various fields suitable for the present invention. For those skilled in the art, additional modifications can be easily achieved.

Claims

1. A method for optimizing energy internet network topology based on a multi-objective evolutionary algorithm, characterized in that: It includes the following steps: 1) Generate an initial adjacency matrix and calculate node coordinates based on geometric distance; 2) Perform geometric-driven evolution operations, including: In the mutation operation, preferentially connect node pairs with geometric distance less than the threshold, and strengthen the triangle structure; In the crossover operation, select parental gene fragments according to the energy transfer efficiency weight; 3) Improve the NSGA-III algorithm, calculate the cosine similarity between solution vectors, impose fitness penalties on solutions with similarity > 0.8, and dynamically adjust the algebraic connectivity C in stages algebraic and node balance R index The weight of 4) Detect isolated subgraphs through depth-first search and generate a minimum spanning tree to complete the connection; when the node energy value is lower than the threshold, disconnect its high-energy-consuming link, and select an alternative path based on the energy transfer efficiency weight; perform local evolution operations on the affected subgraphs; 5) Predict the topology adjustment strategy through the DQN model, and use the reward function R = λ1ΔC algebraic +λ2ΔR index -λ3ΔE loss Train the DQN model and output a priority list of node additions and deletions; 6) Keep the algebraic connectivity C in the last K iterations algebraic ≥0.8×current optimal, and node degree balance R index ≥0.7×the current optimal historical solution, used as the initial population to accelerate convergence.

2. The energy internet network topology optimization method based on a multi-objective evolutionary algorithm according to claim 1, characterized in that: In step S2, the generation of the triangle structure specifically includes: If nodes A, B, and C meet the collinearity condition d AB +d BC =d AC , then ABC is connected first to form a triangle; Assign a cluster weight to each triangle: C w =1 / (d AB +d BC +d AC ), the higher the weight, the greater the probability that the corresponding triangle is selected by the mutation operation.

3. The energy internet network topology optimization method based on a multi-objective evolutionary algorithm according to claim 1, characterized in that: In step S3, the time threshold T of the dynamic weight adjustment mechanism is the total number of evolutionary iterations, specifically: When t < T / 3, the algebraic connectivity weight is 0.7 and the R-index weight is 0.3; When T / 3 ≤ t < 2T / 3, the weights of both are 0.5; When t ≥ 2T / 3, the algebraic connectivity weight is 0.3 and the R-index weight is 0.

7.

4. The energy internet network topology optimization method based on a multi-objective evolutionary algorithm according to claim 1, characterized in that: In step S4, the local evolution operations include: Parallelly optimize the subgraph through distributed computing nodes, and retain the link with the highest energy transfer efficiency during merging; For the uncovered nodes, use the greedy algorithm to complete the shortest path.

5. An intelligent microgrid energy dispatching system based on dynamic network topology optimization, characterized in that: It includes: Multi-source energy access module: Integrate photovoltaic, wind power, energy storage batteries and grid interfaces, and nodes access the microgrid DC bus through adaptive connectors; Real-time data acquisition module: Collect power generation power, energy storage SOC, load demand and environmental parameters; Dynamic topology optimization controller: Invoke the energy Internet network topology optimization method based on the multi-objective evolutionary algorithm described in any one of claims 1 to 4, generate the optimal adjacency matrix in real time, and dynamically adjust the connection priority according to the energy storage SOC: When SOC < 30%, disconnect the energy storage battery from the high-load link and preferentially access renewable energy; When SOC > 80%, use the energy storage battery as a voltage-stabilizing node to connect to the critical load; Edge computing acceleration unit: Perform distributed parallel computing on the microgrid subgraph through the GraphSAGE model to predict the objective function value of local topology optimization.

6. The intelligent microgrid energy dispatching system based on dynamic network topology optimization according to claim 5, characterized in that: Exchange topology optimization results with adjacent microgrid controllers through the 5G communication module to achieve cross-microgrid energy collaborative scheduling; In the scenario of sudden change in photovoltaic output or sudden increase in load, the system quickly completes topology reconstruction and shortens the power supply interruption time.

7. The intelligent microgrid energy dispatching system based on dynamic network topology optimization according to claim 5, characterized in that: In the edge computing acceleration unit, the training data of the GraphSAGE model includes: The adjacency matrix of the historical topology structure, the node energy vector, and the label data of the algebraic connectivity and R-index; Adopt the mean square error loss function for supervised training to reduce the prediction error.

Images