Space-time evolution based intelligent planning method for life cycle of regional energy system

Abstract

The application provides a life cycle intelligent planning method for a regional energy system based on space-time evolution, relates to the technical field of energy planning, and comprises the following steps: constructing and training a space-time evolution model by acquiring historical energy data, analyzing energy facility cost and benefit based on the model, obtaining a Pareto optimal solution by using a multi-objective optimization algorithm, generating a candidate planning scheme set, performing space-time dynamic simulation and calculating a comprehensive score, and selecting a highest-score scheme. The application can improve the scientificity and adaptability of regional energy system planning and realize balanced optimization of energy cost and benefit.

Description

Technical Field

[0001] This invention relates to energy planning technology, and more particularly to a method for intelligent planning of regional energy systems throughout their entire life cycle based on spatiotemporal evolution. Background Technology

[0002] With the acceleration of urbanization and the continuous growth of energy demand, the planning and optimization of regional energy systems have become crucial to ensuring the reliability of energy supply, improving energy efficiency, and reducing environmental pollution. A regional energy system refers to an integrated system of energy production, conversion, transmission, distribution, and consumption constructed within a specific region to meet diverse energy needs. Traditional regional energy system planning mainly relies on static models and empirical judgments, which are insufficient to address the dynamic changes in energy demand and long-term development requirements. With the development of information technology and artificial intelligence, intelligent planning methods based on spatiotemporal evolution have gradually become a research hotspot. This method can consider the spatiotemporal variation characteristics of energy demand, energy prices, and environmental factors, providing more scientific decision support for the full life-cycle planning of regional energy systems.

[0003] Traditional planning methods lack in-depth consideration of the spatiotemporal dynamic characteristics of regional energy systems and mostly adopt static planning models, which cannot effectively reflect the changing patterns of factors such as energy demand, energy prices, and climate conditions over time and space, resulting in planning results that are difficult to adapt to the actual needs of long-term regional development.

[0004] Existing technologies often use weighting or constraint methods to transform multiple objectives into a single objective when dealing with multi-objective optimization problems. This simplified approach makes it difficult to fully capture the complex trade-offs between multiple objectives such as economy, environmental protection, and reliability. This can easily lead to a planning scheme that performs well in some aspects but has poor overall performance.

[0005] Existing planning methods do not adequately consider the entire life cycle of energy systems. They typically focus only on initial investment costs and short-term operational benefits, neglecting the impact of factors such as equipment aging, technological upgrades, and policy changes on system performance during long-term operation. This makes it difficult to guarantee the long-term feasibility and economic viability of planning schemes. Summary of the Invention

[0006] This invention provides a method for intelligent planning of the entire life cycle of regional energy systems based on spatiotemporal evolution, which can solve the problems in the prior art.

[0007] A first aspect of this invention provides a method for intelligent planning of a regional energy system throughout its entire lifecycle based on spatiotemporal evolution, comprising:

[0008] Historical energy data of the target area is obtained, a spatiotemporal evolution model of the regional energy system is constructed, and the spatiotemporal evolution model is trained based on the historical energy data to obtain a trained spatiotemporal evolution model.

[0009] Based on the trained spatiotemporal evolution model, the energy facility costs and energy benefits in the historical energy data are analyzed. The costs and energy benefits are input into the trained spatiotemporal evolution model, and multiple Pareto optimal solutions are obtained by using a multi-objective optimization algorithm.

[0010] Based on the multiple Pareto optimal solutions, a set of candidate planning schemes for the regional energy system is generated using the trained spatiotemporal evolution model. The set of candidate planning schemes is then input into the trained spatiotemporal evolution model, and spatiotemporal dynamic simulation is performed on each planning scheme. Based on the results of the spatiotemporal dynamic simulation, a comprehensive score for each planning scheme is calculated, and the planning scheme with the highest comprehensive score is selected as the final regional energy system planning scheme.

[0011] A spatiotemporal evolution model of a regional energy system is constructed. Based on the historical energy data, the spatiotemporal evolution model is trained to obtain a trained spatiotemporal evolution model, including:

[0012] The set of graph elements in the spatiotemporal evolution model is used to form a dynamic knowledge graph. Temporal features are obtained from historical data of the regional energy system. The correlation strength between nodes is calculated based on the temporal features. The correlation relationship in the dynamic knowledge graph is updated according to the correlation strength.

[0013] The relationships between nodes in the dynamic knowledge graph are analyzed, the degree of mutual influence between nodes is calculated, and the degree of mutual influence is written into the edge set of the dynamic knowledge graph to obtain a dynamic knowledge graph containing the influence relationships.

[0014] Feature extraction is performed on each node in the dynamic knowledge graph containing influence relationships, and a node feature vector is generated based on the node's neighborhood information; the node feature vector is fused with the spatiotemporal features in historical energy data, and the fusion ratio of the node feature vector and the spatiotemporal features is adjusted by weighting coefficients to obtain fused features containing knowledge graph information;

[0015] Based on the fusion features, feature analysis is performed on energy system data from different regions to calculate the feature differences between the data from different regions. Regional feature alignment is achieved by reducing the feature differences. Feature optimization is then performed on the data after regional feature alignment to generate feature data for model training.

[0016] The feature data and the feature differences are weighted and combined to obtain the optimization target. The spatiotemporal evolution model is trained by optimizing the optimization target to obtain a trained spatiotemporal evolution model with transferability.

[0017] Based on the trained spatiotemporal evolution model, the energy facility costs and energy benefits in the historical energy data are analyzed. The costs and energy benefits are input into the trained spatiotemporal evolution model, and a multi-objective optimization algorithm is used to obtain multiple Pareto optimal solutions, including:

[0018] Extract the cycle operating cost of energy facilities from historical energy data, and construct the first input parameter of the optimization function based on the cycle operating cost; extract the input power and output power of energy facilities from the historical energy data, calculate the efficiency index of energy facilities based on the input power and output power, and construct the second input parameter of the optimization function based on the efficiency index;

[0019] The first input parameter and the second input parameter are input into the trained spatiotemporal evolution model to construct a feature vector containing temporal and spatial features; the feature vector is divided into a grid to obtain a feature grid; and a coarse-grained search is performed on the trained spatiotemporal evolution model based on the feature grid to determine the optimal search region.

[0020] Set a search center point in the optimal search area, calculate the search step size of the current search position based on the search center point, calculate the distance from the search center point to the current search position, and determine the search constraints based on the distance.

[0021] Under the search constraints, the optimization function is iteratively optimized based on the search step size to obtain the Pareto optimal solution of the energy system.

[0022] The feature vectors are divided into a grid to obtain a feature grid. Based on the feature grid, a coarse-grained search is performed in the trained spatiotemporal evolution model to determine the optimal search region, including:

[0023] Obtain an n-dimensional feature vector, divide the n-dimensional feature vector into regions in the feature space, initialize the density of the region division result with an initial density value to generate a chaotic sequence, construct a density function by multiplying the chaotic sequence with the initial density value, and dynamically adjust the density of the region based on the density function to obtain a density distribution;

[0024] Based on the density distribution, the results of the region division are restructured to generate adaptive feature regions, and the adaptive feature regions are input into the trained spatiotemporal evolution model.

[0025] In the trained spatiotemporal evolution model, a search direction vector is constructed based on the chaotic sequence, and an adaptive search step size is constructed by multiplying the chaotic sequence with the baseline step size; the adaptive feature region is traversed according to the search direction vector and the adaptive search step size to obtain a search sequence, and the feature value of each position in the search sequence is obtained.

[0026] The convergence of the search sequence, the diversity of the feature values, and the stability of the adaptive search step size are calculated. The weighted sum of the convergence, diversity, and stability is used to construct a search efficiency evaluation index. The search sequence is evaluated based on the search efficiency evaluation index, and the optimal search region is determined in the adaptive feature region.

[0027] Based on the multiple Pareto optimal solutions, a set of candidate planning schemes for the regional energy system is generated using the trained spatiotemporal evolution model; the set of candidate planning schemes is input into the trained spatiotemporal evolution model, and spatiotemporal dynamic simulation is performed on each planning scheme, including:

[0028] A knowledge graph is constructed based on multiple Pareto optimal solutions. Historical operational experience, domain expert knowledge, and technical standards and specifications are extracted from the knowledge graph. A feature mapping function is constructed based on the historical operational experience, the domain expert knowledge, and the technical standards and specifications.

[0029] For each Pareto optimal solution among the plurality of Pareto optimal solutions, the frequency of use and the degree of novelty are calculated, and an evaluation result of the feature mapping function is constructed based on the frequency of use and the degree of novelty;

[0030] The evaluation results of the feature mapping function are used to evaluate multiple Pareto optimal solutions, the value gradient of the feature mapping function is calculated, and the strategy is updated according to the value gradient generation scheme.

[0031] Based on the proposed scheme update strategy, a set of candidate planning schemes is generated, and the set of candidate planning schemes is input into a spatiotemporal evolution model. Spatial and temporal features are extracted from the set of candidate planning schemes, and the spatial and temporal features are input into the spatiotemporal evolution model for spatiotemporal dynamic simulation.

[0032] For each Pareto optimal solution among the plurality of Pareto optimal solutions, the frequency of use and the degree of novelty are calculated, and the evaluation result of the feature mapping function is constructed based on the frequency of use and the degree of novelty, including:

[0033] Establish pheromone concentrations for multiple Pareto optimal solutions, calculate the usage frequency of each Pareto optimal solution based on the pheromone concentrations, and dynamically fuse the usage frequency with the path selection factor through an adaptive threshold processor to form a corrected usage frequency.

[0034] The final usage frequency value of each Pareto optimal solution is generated based on the corrected usage frequency; the similarity between the multiple Pareto optimal solutions is calculated, and the minimum similarity value corresponding to each Pareto optimal solution and the maximum similarity value between the multiple Pareto optimal solutions are identified by the adaptive threshold processor to generate the novelty.

[0035] Calculate the information entropy of the final usage frequency value and the novelty level, and determine the weights of the final usage frequency value and the novelty level based on the information entropy; use the adaptive threshold processor to combine the weights with the final usage frequency value and the novelty level to generate a feature mapping function evaluation result.

[0036] Based on the results of the spatiotemporal dynamic simulation, a comprehensive score is calculated for each planning scheme, and the planning scheme with the highest comprehensive score is selected as the final regional energy system planning scheme, including:

[0037] Based on the results of the spatiotemporal dynamic simulation, the spatiotemporal features of each planning scheme are extracted, and the time decay weight of the spatiotemporal features is calculated. The spatiotemporal features are weighted based on the time decay weight to obtain the weighted values ​​of the spatiotemporal features of the planning scheme, and the correlation coefficient between the weighted values ​​of the spatiotemporal features is calculated.

[0038] The feature membership degree of the planning scheme is determined based on the correlation coefficient. The feature membership degree is multiplied by the corresponding feature value to obtain the basic score. The basic score is dynamically corrected based on the periodic change pattern.

[0039] Evidence enhancement is performed on the basic scores to obtain a comprehensive score for each planning scheme. The planning scheme with the highest comprehensive score is selected as the final regional energy system planning scheme.

[0040] A second aspect of the present invention provides an electronic device, comprising:

[0041] processor;

[0042] Memory used to store processor-executable instructions;

[0043] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0044] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0045] The beneficial effects of this application are as follows:

[0046] The present invention provides a regional energy system lifecycle intelligent planning method based on spatiotemporal evolution. By constructing a spatiotemporal evolution model, it can more accurately characterize the changes in regional energy systems over time and space, thereby improving the rationality and adaptability of planning schemes.

[0047] By introducing a multi-objective optimization algorithm to generate a Pareto optimal solution set, the balance between cost and energy efficiency is fully considered, avoiding the limitations of traditional single-objective programming methods and providing decision-makers with more diversified planning options.

[0048] This invention employs spatiotemporal dynamic simulation to evaluate candidate planning schemes throughout their entire life cycle, calculates a comprehensive score to select the optimal scheme, and ensures that the final planning scheme has good economic efficiency, reliability and environmental friendliness during long-term operation, effectively improving the overall operating efficiency and sustainable development capability of the regional energy system. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the intelligent planning method for the entire life cycle of a regional energy system based on spatiotemporal evolution, as described in an embodiment of the present invention.

[0050] Figure 2 This is a flowchart illustrating the Pareto optimal solution evaluation process according to an embodiment of the present invention.

[0051] Figure 3 This is a flowchart of the search optimization process based on the genetic algorithm in an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0054] Figure 1 This is a flowchart illustrating the intelligent lifecycle planning method for regional energy systems based on spatiotemporal evolution, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0055] Historical energy data of the target area is obtained, a spatiotemporal evolution model of the regional energy system is constructed, and the spatiotemporal evolution model is trained based on the historical energy data to obtain a trained spatiotemporal evolution model.

[0056] Based on the trained spatiotemporal evolution model, the energy facility costs and energy benefits in the historical energy data are analyzed. The costs and energy benefits are input into the trained spatiotemporal evolution model, and multiple Pareto optimal solutions are obtained by using a multi-objective optimization algorithm.

[0057] Based on the multiple Pareto optimal solutions, a set of candidate planning schemes for the regional energy system is generated using the trained spatiotemporal evolution model. The set of candidate planning schemes is then input into the trained spatiotemporal evolution model, and spatiotemporal dynamic simulation is performed on each planning scheme. Based on the results of the spatiotemporal dynamic simulation, a comprehensive score for each planning scheme is calculated, and the planning scheme with the highest comprehensive score is selected as the final regional energy system planning scheme.

[0058] In one optional implementation, a spatiotemporal evolution model of the regional energy system is constructed. Based on the historical energy data, the spatiotemporal evolution model is trained to obtain a trained spatiotemporal evolution model, including:

[0059] The set of graph elements in the spatiotemporal evolution model is used to form a dynamic knowledge graph. Temporal features are obtained from historical data of the regional energy system. The correlation strength between nodes is calculated based on the temporal features. The correlation relationship in the dynamic knowledge graph is updated according to the correlation strength.

[0060] The relationships between nodes in the dynamic knowledge graph are analyzed, the degree of mutual influence between nodes is calculated, and the degree of mutual influence is written into the edge set of the dynamic knowledge graph to obtain a dynamic knowledge graph containing the influence relationships.

[0061] Feature extraction is performed on each node in the dynamic knowledge graph containing influence relationships, and a node feature vector is generated based on the node's neighborhood information; the node feature vector is fused with the spatiotemporal features in historical energy data, and the fusion ratio of the node feature vector and the spatiotemporal features is adjusted by weighting coefficients to obtain fused features containing knowledge graph information;

[0062] Based on the fusion features, feature analysis is performed on energy system data from different regions to calculate the feature differences between the data from different regions. Regional feature alignment is achieved by reducing the feature differences. Feature optimization is then performed on the data after regional feature alignment to generate feature data for model training.

[0063] The feature data and the feature differences are weighted and combined to obtain the optimization target. The spatiotemporal evolution model is trained by optimizing the optimization target to obtain a trained spatiotemporal evolution model with transferability.

[0064] When assembling a set of graph primitives in a spatiotemporal evolution model into a dynamic knowledge graph, a graph primitive structure is defined. A graph primitive consists of three parts: nodes, edges, and attributes. Nodes represent various entities in the energy system, such as generator sets, substations, power distribution networks, and power loads. Edges represent the relationships between entities, such as power transmission, energy conversion, and control signals. Attributes represent the characteristic parameters of entities and their relationships, such as capacity, efficiency, and distance.

[0065] For the energy system of an industrial park, an initial knowledge graph was constructed, containing 152 nodes and 437 edges. The node set includes 35 power generation nodes, 42 substation nodes, 58 load nodes, and 17 energy storage nodes. For example, node N047 represents distributed photovoltaic power station No. 3, with attributes including installed capacity of 2.5MW, average power generation efficiency of 16.8%, and completion date of 2019; edge E129 connects nodes N047 and N086 (distribution room No. 5), representing the power transmission from the photovoltaic power station to the distribution room, with attributes including maximum transmission power of 2.3MW and average loss rate of 3.2%. The initial knowledge graph is constructed based on the physical structure of the energy system and does not yet include data-driven dynamic relationships.

[0066] When extracting time-series features from historical data of a regional energy system, operational data for each node at different points in time are collected. These time-series features include information on power, efficiency, and status data that change over time. For example, for node N047 (PV Power Station No. 3), hourly power generation data for one year is collected, forming a time series of 8760 points [0, 0,0, ..., 1.85, 1.92, 1.78, ..., 0, 0, 0] MW, where zero values ​​correspond to nighttime periods without sunlight. Similarly, time-series data for all nodes are collected to form a complete time-series feature dataset. These time-series features reflect the dynamic operational characteristics of each component of the energy system, providing a data foundation for subsequent correlation analysis.

[0067] When calculating the correlation strength between nodes based on time-series characteristics, time-series correlation analysis is used to calculate the correlation coefficient between the time-series data of any two nodes. The higher the correlation coefficient, the stronger the correlation between the two nodes. In specific implementation, Pearson correlation coefficient, mutual information, or Granger causality index can be calculated. For example, the correlation coefficient of the power time-series data of nodes N047 (PV Power Station No. 3) and N086 (Distribution Room No. 5) is 0.92, indicating a high correlation between them; the correlation coefficient of N047 and N118 (Data Center No. 2) is 0.15, indicating a weak correlation. By calculating the correlation strength between all node pairs, a correlation strength matrix of size 152×152 is formed.

[0068] When updating the relationships in the dynamic knowledge graph based on the association strength, the edge set of the knowledge graph is adjusted based on the association strength matrix. When the association strength between two nodes exceeds the threshold (e.g., 0.6) and there is no connection in the original knowledge graph, a new edge is added. When the association strength between two nodes is lower than the threshold and there is a connection in the original knowledge graph, the edge may be removed. When the edge already exists, the association strength is written into the edge's attributes.

[0069] For example, nodes N047 and N118 were not originally directly connected, but data analysis revealed that when photovoltaic power generation is high, the data center's power load is also often high (correlation coefficient of 0.73). Therefore, an edge E592 was added between them in the knowledge graph to represent a potential energy scheduling relationship. The updated dynamic knowledge graph contains 152 nodes and 498 edges, an increase of 61 data-driven association edges compared to the initial knowledge graph.

[0070] When analyzing the relationships between nodes in a dynamic knowledge graph, not only directly connected nodes are considered, but also the path structures formed by multi-hop connections. Graph analysis algorithms such as PageRank or node centrality calculation are used to assess the importance and influence of nodes in the entire network. For example, node N021 (main substation) has a PageRank value of 0.083, ranking second among all nodes, indicating its important position in the energy network. Furthermore, the shortest paths and connectivity between nodes are calculated to identify critical paths and potential bottlenecks in the network. For instance, the shortest path length between nodes N086 and N118 is 3, requiring two intermediate nodes to connect, which leads to reduced energy transmission efficiency.

[0071] When calculating the degree of mutual influence between nodes, the propagation effect of disturbances is analyzed based on a network propagation model. By simulating the impact of a node's state change on other nodes in the network, the degree of influence between nodes is quantified. In specific implementations, a linear propagation model or a cascading failure model can be used.

[0072] For example, when the output power of node N021 (main substation) decreases by 10%, the available power of node N086 decreases by 8.7% and the available power of node N118 decreases by 7.2% through network propagation. Based on this, the influence degree of N021 on N086 is calculated to be 0.87, and the influence degree on N118 is calculated to be 0.72. Through systematic analysis, a node influence matrix of size 152×152 is constructed to record the degree of mutual influence between all node pairs.

[0073] When incorporating the degree of mutual influence into the edge set of the dynamic knowledge graph, the values ​​in the node influence matrix are used as new attributes for the edges. For edges already existing in the knowledge graph, the degree of influence attribute is added directly; for node pairs with significant influence but no connection yet established, new influence edges can be added. For example, edge E129 (connecting N047 and N086) has an added degree of influence attribute of 0.92, representing the direct influence of the photovoltaic power station on the substation. The updated knowledge graph not only includes physical connections but also data-driven influence relationships, providing a more comprehensive description of the complex interactions of the energy system.

[0074] When extracting features from each node in a dynamic knowledge graph containing influencing relationships, both attribute features and structural features of the node are considered. Attribute features are directly extracted from the node's attribute set, such as parameters like capacity and efficiency. Structural features are extracted from the node's network location and connection patterns using graph algorithms, such as degree centrality and proximity centrality. For example, the attribute features of node N047 include an installed capacity of 2.5MW and an average power generation efficiency of 16.8%, while its structural features include a degree centrality of 0.072 and a proximity centrality of 0.45. Combining the attribute features and structural features forms a comprehensive feature representation of the node.

[0075] When generating node feature vectors based on node neighborhood information, a neighborhood aggregation mechanism in graph neural networks is adopted. For each node, the feature information of its one-hop neighbors and two-hop neighbors is collected and aggregated through weighted averaging or attention mechanism to generate a feature vector containing neighborhood information.

[0076] For example, node N047's one-hop neighbors include three nodes: N086, N093, and N102, and its two-hop neighbors include four nodes: N118, N125, N134, and N142. Aggregating the feature information of these neighbors yields a feature vector of dimension 64 [0.37, 0.28, 0.15, ..., 0.42, 0.19]. Through neighborhood aggregation, the node feature vector not only contains information about the node itself but also contextual information about its surrounding environment.

[0077] When fusing node feature vectors with spatiotemporal features from historical energy data, it is necessary to process features from two different sources: spatiotemporal features, including time-dimensional features (such as time period, season, temperature, etc.) and spatial-dimensional features (such as geographical location, region type, etc.). For example, time-dimensional features can be represented as [0.25, 0.75, 0.82], representing the time period (0-1), seasonal factor (0-1), and temperature factor (0-1) of a day, respectively; spatial-dimensional features can be represented as [0.42, 0.63, 0.51], representing the X-coordinate (0-1), Y-coordinate (0-1), and region type (0-1), respectively. Concatenating the node feature vector (64 dimensions) with the spatiotemporal features (6 dimensions) forms a preliminary fusion feature with a dimension of 70.

[0078] When adjusting the fusion ratio of node feature vectors and spatiotemporal features by weighting coefficients, an adaptive weighting mechanism is set up to dynamically adjust the importance weights of the two types of features according to different node types and application scenarios. For example, for photovoltaic power generation nodes that are significantly affected by weather, the weight of spatiotemporal features needs to be increased; for stable-operating substation equipment, more attention should be paid to the information in the node feature vectors.

[0079] In practice, the initial weights can be set to [0.6, 0.4], corresponding to the importance of the node feature vector and the spatiotemporal features, respectively. These weights are then dynamically adjusted based on performance on the validation dataset. For node N047, the adjusted weights are [0.45, 0.55], indicating that the spatiotemporal features are slightly more important than the node features; for node N086, the weights are [0.7, 0.3], indicating that the node features are more important. The final fused feature vector is obtained through weighted fusion.

[0080] When performing feature analysis on energy system data from different regions based on fusion characteristics, the energy system is divided into multiple functional or geographical regions. The feature distribution of data from different regions is compared. For example, an industrial park can be divided into three regions: A, B, and C, each containing a certain number of energy nodes. The fusion characteristics of the nodes in each region are extracted, and statistical indicators such as mean, variance, and distribution shape are calculated. The feature mean vector for region A is [0.48, 0.35, 0.62, ..., 0.53, 0.41], for region B it is [0.51, 0.38, 0.59, ..., 0.49, 0.44], and for region C it is [0.39, 0.42, 0.57, ..., 0.61, 0.36]. Through feature analysis, the characteristic differences and distribution patterns of data from different regions are identified.

[0081] When calculating the feature differences between data from different regions, a distribution distance metric is used to calculate the KL divergence, JS divergence, or Wasserstein distance of the feature distributions between regions, quantifying the degree of difference in feature distributions. For example, the JS divergence between regions A and B is 0.15, between regions A and C is 0.28, and between regions B and C is 0.22. The greater the feature difference, the more significant the difference in data distribution between regions, and the greater the difficulty of model transfer. A region-pair difference matrix is ​​constructed to record the feature difference values ​​between all region pairs.

[0082] When achieving regional feature alignment by reducing feature differences, feature transformation or domain adaptation techniques are employed. Feature transformation functions are designed to map features from different regions to a common feature space, making their distributions more similar. Specifically, linear transformations, kernel transformations, or neural network transformations can be used. For example, three transformation matrices TA, TB, and TC are designed to transform the features of regions A, B, and C to the common space, respectively. After applying the transformation, the JS divergence of the feature distributions in regions A and B decreases to 0.08, the JS divergence between regions A and C decreases to 0.12, and the JS divergence between regions B and C decreases to 0.10. Feature alignment allows the model to better adapt to data distributions in different regions, improving its transferability.

[0083] When optimizing features on data aligned with regional features, feature selection, dimensionality reduction, or reconstruction techniques are used to improve feature quality. Principal component analysis, autoencoders, or feature importance assessment methods are employed to identify the subset of features that contribute most to model performance. For example, feature importance assessment can select the 35 most important features from a 70-dimensional fused feature set to form an optimized feature set. Feature optimization can reduce model complexity, decrease the risk of overfitting, and improve computational efficiency. The generated feature data includes a feature matrix (number of samples × feature dimension) and corresponding labels (such as energy consumption, efficiency indicators, etc.), which serve as input for model training.

[0084] When weighting and combining feature data and feature differences to obtain the optimization objective, a multi-objective optimization function is constructed. The optimization objective consists of two parts: a primary task objective and a transferability objective. The primary task objective focuses on the model's prediction accuracy on the current data, such as mean squared error or classification accuracy; the transferability objective focuses on the feature differences between different regions, encouraging the model to learn region-invariant feature representations. For example, setting the weight of the primary task objective to 0.7 and the weight of the transferability objective to 0.3 constructs a comprehensive optimization objective. When the model's performance on the validation set improves but its transferability decreases, the weight ratio can be dynamically adjusted to balance the importance of the two objectives.

[0085] When training a spatiotemporal evolution model by optimizing the objective, optimization algorithms such as gradient descent are used to iteratively update the model parameters. The spatiotemporal evolution model can adopt a combined architecture, such as using a neural network to handle spatial dependencies and a recurrent neural network or attention mechanism to handle temporal dependencies. During training, the feature data is divided into training, validation, and test sets, and iterative optimization is performed using the mini-batch gradient descent method.

[0086] For example, the batch size was set to 32, the initial learning rate was 0.001, and training was conducted for 100 epochs. During training, the trends of the main task objective and the transfer capability objective were monitored. When the performance on the validation set stopped improving, the learning rate was reduced or training was stopped early. The final trained model achieved a mean squared error of 0.037 on the test set and an average transfer error of 0.052 between different regions, demonstrating good prediction accuracy and transfer capability.

[0087] This method effectively captures the complex relationships and spatiotemporal characteristics of regional energy systems through dynamic knowledge graphs and graph neural network technology, and constructs a spatiotemporal evolution model with transferability, providing strong technical support for the analysis, prediction and optimization of regional energy systems.

[0088] In one optional implementation, based on the trained spatiotemporal evolution model, the energy facility costs and energy benefits in the historical energy data are analyzed. The costs and energy benefits are input into the trained spatiotemporal evolution model, and a multi-objective optimization algorithm is used to obtain multiple Pareto optimal solutions, including:

[0089] Extract the cycle operating cost of energy facilities from historical energy data, and construct the first input parameter of the optimization function based on the cycle operating cost; extract the input power and output power of energy facilities from the historical energy data, calculate the efficiency index of energy facilities based on the input power and output power, and construct the second input parameter of the optimization function based on the efficiency index;

[0090] The first input parameter and the second input parameter are input into the trained spatiotemporal evolution model to construct a feature vector containing temporal and spatial features; the feature vector is divided into a grid to obtain a feature grid; and a coarse-grained search is performed on the trained spatiotemporal evolution model based on the feature grid to determine the optimal search region.

[0091] Set a search center point in the optimal search area, calculate the search step size of the current search position based on the search center point, calculate the distance from the search center point to the current search position, and determine the search constraints based on the distance.

[0092] Under the search constraints, the optimization function is iteratively optimized based on the search step size to obtain the Pareto optimal solution of the energy system.

[0093] like Figure 2 As shown, the method includes:

[0094] When extracting the cycle operating costs of energy facilities from historical energy data, it is necessary to process the raw operating data and summarize the cost composition of different energy facilities in each operating cycle. For the energy system of an industrial park, operating cost data of various energy facilities over the past three years can be extracted from the historical database. For example, for 10 distributed generator sets in the park, data such as fuel cost, maintenance cost, labor cost, and depreciation cost for each device can be extracted.

[0095] Specifically, the average fuel cost of generator set No. 1 over the past 12 quarters was RMB 425,000 per quarter, maintenance cost was RMB 83,000 per quarter, labor cost was RMB 57,000 per quarter, depreciation cost was RMB 128,000 per quarter, and total operating cost was RMB 693,000 per quarter. Similar data extraction and statistical analysis were performed on all generator sets, heat pump systems, energy storage devices, etc., to obtain a complete dataset of cycle operating costs.

[0096] When constructing the first input parameter of the optimization function based on the cycle operating cost, the extracted cost data is standardized and structured. A cost parameter matrix is ​​constructed, with rows representing different types of energy facilities and columns representing different cost categories. For example, for an energy system containing three types of equipment—generator sets, heat pump systems, and energy storage devices—the constructed cost parameter matrix is ​​3×4 in size, storing four types of cost data for each of the three types of equipment.

[0097] For generator sets, the standardized values ​​for the four costs are 0.85, 0.65, 0.45, and 0.72, respectively; for heat pump systems, the standardized values ​​are 0.52, 0.78, 0.38, and 0.65, respectively; and for energy storage devices, the standardized values ​​are 0.35, 0.82, 0.30, and 0.88, respectively. By weighted summation, the cost parameter matrix is ​​transformed into a cost vector [0.72, 0.60, 0.58], which serves as the first input parameter of the optimization function.

[0098] When extracting the input and output power of energy facilities from historical energy data, power monitoring data of various energy devices under different operating conditions are collected. Taking a heat pump system as an example, its input electrical power and output thermal power under different ambient temperatures and load conditions are extracted from historical data. For heat pump system No. 2, under the condition of an ambient temperature of 5℃ and a load rate of 80%, the input electrical power is 120 kW and the output thermal power is 380 kW. Similarly, power data of all energy facilities under different operating conditions are collected to establish a two-dimensional data table containing input and output power.

[0099] When calculating the efficiency indicators of energy facilities based on input and output power, indicators such as energy conversion efficiency and system performance coefficient are used. Energy conversion efficiency equals output power divided by input power, while the system performance coefficient takes into account factors such as rated power and part-load characteristics. Taking heat pump system No. 2 as an example, its energy conversion efficiency under the above operating conditions is 380 ÷ 120 = 3.17.

[0100] For generator sets, the energy conversion efficiency equals the output electrical power divided by the calorific value of the input fuel; for example, the conversion efficiency of generator set No. 3 is 0.42. For energy storage devices, charge and discharge efficiency must be considered; for example, the round-trip efficiency of battery energy storage system No. 1 is 0.88. After calculating the efficiency indices of all devices, standardization is performed to obtain an efficiency index vector. For example, the standardized efficiency indices for the three types of devices are 0.65, 0.82, and 0.78, respectively.

[0101] When constructing the second input parameter of the optimization function based on the efficiency index, the efficiency index is combined with factors such as system operational stability and equipment lifespan to form a comprehensive performance index. For example, for a generator set, the comprehensive performance index equals the efficiency index multiplied by the stability coefficient (0.92) and then multiplied by the lifespan coefficient (0.88), resulting in 0.65 × 0.92 × 0.88 = 0.53. For heat pump systems and energy storage devices, similar calculations yield comprehensive performance indices of 0.71 and 0.65, respectively. These comprehensive performance indices form a vector [0.53, 0.71, 0.65], which serves as the second input parameter of the optimization function.

[0102] When inputting the first and second input parameters into the trained spatiotemporal evolution model, the two sets of parameters need to be fused and formatted. The cost vector [0.72, 0.60, 0.58] and the performance vector [0.53, 0.71, 0.65] are merged into an input matrix, with timestamps and spatial location information added. For example, if the current planning period is set to T2023Q4 and the planning area is set to the central area of ​​the park, the input data packet is constructed as {T2023Q4, central area, [0.72, 0.60, 0.58], [0.53, 0.71, 0.65]}. The trained spatiotemporal evolution model is a computational model that is pre-trained using historical data and can predict the spatiotemporal evolution characteristics of the regional energy system. It receives the input data packet and performs internal calculations and feature extraction.

[0103] When constructing feature vectors that incorporate both temporal and spatial features, the spatiotemporal evolution model extracts and enhances these features from the input data. Temporal features include seasonal variations, daily load curves, and holiday effects, while spatial features include geographical distribution, network topology, and regional connectivity.

[0104] For example, for the input data packet {T2023Q4, central area, [0.72, 0.60, 0.58], [0.53, 0.71, 0.65]}, the extracted temporal features include winter heating characteristic coefficient 0.82, seasonal peak-to-valley ratio 1.45, and intraday load uniformity 0.73; the extracted spatial features include central area concentration 0.68, network connectivity 0.85, and regional boundary effect 0.37. These features, combined with the original input parameters, form a 12-dimensional feature vector [0.72, 0.60, 0.58, 0.53, 0.71, 0.65, 0.82, 1.45, 0.73, 0.68, 0.85, 0.37].

[0105] When dividing the feature vectors into a grid to obtain the feature grid, the 12-dimensional feature space is divided into a regular grid. Each dimension is divided into several intervals according to its value range, forming a hypercube grid. For example, dividing each dimension into 6 intervals generates a total of 6^12 grid cells. In actual implementation, considering computational complexity, adaptive grid or sparse grid techniques can be used, focusing on the effective regions in the feature space. For example, for the first dimension (generator set cost), the interval [0,1] is divided into five sub-intervals: [0, 0.2), [0.2, 0.4), [0.4, 0.6), [0.6, 0.8), and [0.8, 1.0]. The current value of 0.72 falls in the fourth sub-interval. Similar processing is performed on all dimensions to determine the grid position of the feature vector.

[0106] When performing coarse-grained search based on feature grids in the trained spatiotemporal evolution model, an initial search range is determined by expanding the search space outwards from the grid containing the feature vector. For example, expanding by two grids in each positive and negative direction creates a hypercube search space of size 5^12. Within this search space, a jump-point search strategy is employed, jumping a certain step size in each dimension to evaluate the objective function value at the current point. For instance, setting the initial step size to 0.1 generates approximately 5000 evaluation points in the feature space. For each evaluation point, the corresponding energy system cost target value and benefit target value are calculated, forming a point set in the two-dimensional evaluation space.

[0107] When determining the optimal search region, analyze the coarse-grained search results to identify regions containing potential Pareto solutions. In the two-dimensional evaluation space, find the boundaries of the point set along the direction from high to low cost and from low to high benefit. The feature space regions corresponding to these boundary points are the optimal search regions.

[0108] For example, by analyzing the distribution of 5000 evaluation points, three potential optimal search regions R1, R2, and R3 were identified. Among them, region R2 has the highest point density and contains the most boundary points, so R2 was selected as the optimal search region. R2 is represented as a hypercube in the 12-dimensional feature space, with each dimension ranging from [0.65,0.75], [0.55,0.65], [0.50,0.60], [0.50,0.60], [0.65,0.80], [0.60,0.70], [0.75,0.90], [1.30,1.60], [0.65,0.80], [0.60,0.75], [0.80,0.95], to [0.30,0.45].

[0109] When setting the search center point within the optimal search region, you can choose the region's center point or the point within the region with the best evaluation performance as the search center. For example, you can choose the center point of region R2 [0.70, 0.60, 0.55, 0.55, 0.73, 0.65, 0.83, 1.45, 0.73, 0.68, 0.88, 0.38] as the search center point. The search center point represents the core location of the optimal search region and will serve as the starting point for fine-tuning the search.

[0110] When calculating the search step size based on the search center point, an adaptive step size strategy is adopted. The initial step size is set to 5% of the optimal search region size, and it is dynamically adjusted as the search progresses. The step size adjustment rule is as follows: when no better solution is found in three consecutive iterations, the step size is reduced to 80% of the current value; when a better solution is found, the step size is increased to 110% of the current value. For example, for dimension 1 (generator set cost), the optimal search region range is [0.65, 0.75], and the initial step size is (0.75-0.65)×5%=0.005. As the search progresses, the step size is adjusted to values ​​such as 0.004, 0.0032, and 0.0035.

[0111] When calculating the distance from the search center point to the current search position, the Euclidean distance metric is used. The search center point and the current search position are represented as 12-dimensional vectors, and the Euclidean distance between the two points is calculated. For example, if the search center point is [0.70, 0.60, 0.55, 0.55, 0.73, 0.65, 0.83, 1.45, 0.73, 0.68, 0.88, 0.38], and the current search position is [0.69, 0.61, 0.54, 0.56, 0.72, 0.66, 0.84, 1.44, 0.74, 0.67, 0.89, 0.37], the calculated distance is 0.028.

[0112] When determining search constraints based on distance, a distance threshold and a search radius are set. The distance threshold controls the search range to prevent the search from deviating too far from the optimal region. The search radius is dynamically adjusted during iteration, with an initial value set at 50% of the diagonal length of the optimal search region, gradually decreasing to 10%. For example, if the diagonal length of the optimal search region R2 is approximately 0.42, the initial search radius is 0.21, and the search constraint is set as "the distance from the current search position to the search center point does not exceed the search radius." As iterations proceed, the search radius is adjusted to values ​​such as 0.18, 0.15, and 0.12.

[0113] When iteratively optimizing the optimization function based on the search step size under search constraints, multi-objective optimization algorithms such as NSGA-II (Non-dominated sorting genetic algorithm II) or MOPSO (Multi-objective particle swarm optimization algorithm) are employed. Taking NSGA-II as an example, the population size is set to 100, the maximum number of iterations to 200, the crossover probability to 0.85, and the mutation probability to 0.15. In each iteration, new candidate solutions are generated, their corresponding cost and benefit objectives are calculated, non-dominated sorting and crowding calculations are performed, and high-quality individuals are selected to enter the next generation. For example, in the 50th iteration, the number of non-dominated solutions in the current population is 35. These solutions form a curve on the cost-benefit plane, representing the Pareto front found so far.

[0114] When the Pareto optimal solution of the energy system is obtained, all non-dominated solutions are extracted from the final population to form the Pareto optimal solution set. Each Pareto optimal solution corresponds to a point in the feature space, representing an energy system configuration scheme. For example, through 200 iterations, 42 Pareto optimal solutions are finally obtained, among which three typical solutions are P1[0.68, 0.62, 0.53, 0.57, 0.74, 0.67, 0.85, 1.43, 0.75, 0.66, 0.90, 0.36], P2[0.72, 0.59, 0.56, 0.54, 0.75, 0.64, 0.82, 1.46, 0.72, 0.69, 0.87, 0.39] and P3[0.66, 0.63, 0.52, 0.58, 0.71, 0.68, 0.86, 1.42, 0.76, 0.65, 0.91, 0.35]. The target cost for P1 is 0.58, and the target benefit is 0.72; the target cost for P2 is 0.62, and the target benefit is 0.76; the target cost for P3 is 0.54, and the target benefit is 0.68. These Pareto optimal solutions represent the optimal trade-off between cost and benefit, providing diverse options for regional energy system planning.

[0115] This technical method extracts key parameters from historical energy data and combines them with spatiotemporal evolution models and multi-objective optimization algorithms to achieve optimal design of regional energy systems. It obtains a series of Pareto optimal solutions that take into account both cost and benefit, providing a scientific basis for energy system planning.

[0116] In one optional implementation, the feature vectors are divided into a grid to obtain a feature grid. Based on the feature grid, a coarse-grained search is performed in the trained spatiotemporal evolution model to determine the optimal search region, including:

[0117] Obtain an n-dimensional feature vector, divide the n-dimensional feature vector into regions in the feature space, initialize the density of the region division result with an initial density value to generate a chaotic sequence, construct a density function by multiplying the chaotic sequence with the initial density value, and dynamically adjust the density of the region based on the density function to obtain a density distribution;

[0118] Based on the density distribution, the results of the region division are restructured to generate adaptive feature regions, and the adaptive feature regions are input into the trained spatiotemporal evolution model.

[0119] In the trained spatiotemporal evolution model, a search direction vector is constructed based on the chaotic sequence, and an adaptive search step size is constructed by multiplying the chaotic sequence with the baseline step size; the adaptive feature region is traversed according to the search direction vector and the adaptive search step size to obtain a search sequence, and the feature value of each position in the search sequence is obtained.

[0120] The convergence of the search sequence, the diversity of the feature values, and the stability of the adaptive search step size are calculated. The weighted sum of the convergence, diversity, and stability is used to construct a search efficiency evaluation index. The search sequence is evaluated based on the search efficiency evaluation index, and the optimal search region is determined in the adaptive feature region.

[0121] When obtaining an n-dimensional feature vector, key feature parameters are extracted from the original data of the regional energy system. For a certain regional energy system, n key parameters can be extracted, including energy station capacity, pipeline connection method, load distribution characteristics, energy storage device configuration and operation strategy, forming an n-dimensional feature vector. For example, for the energy system of an industrial park, the extracted 7-dimensional feature vector can be represented as [0.72, 0.58, 0.63, 0.45, 0.81, 0.37, 0.69], corresponding to the power generation capacity coefficient, heating capacity coefficient, cooling capacity coefficient, energy storage ratio coefficient, network connectivity, demand response flexibility, and system redundancy, respectively.

[0122] When partitioning an n-dimensional feature vector into regions in the feature space, a hypercube mesh partitioning method is used to uniformly divide the value range of each dimension into m sub-intervals, forming m... n There are 5 grid cells. Taking a 7-dimensional feature vector as an example, if each dimension is divided into 5 sub-intervals, then 5 grid cells are generated. 7 =78125 grid cells. Each grid cell represents a local region in the feature space, with a clearly defined boundary and center point. For example, grid cell G4213 can be represented as [[0.6,0.8], [0.4,0.6], [0.6,0.8], [0.4,0.6], [0.8,1.0], [0.2,0.4], [0.6,0.8]], where each pair of values ​​represents the range of values ​​for that dimension.

[0123] When initializing the density of the region partitioning results using initial density values, an initial density value is assigned to each grid cell based on prior knowledge or a uniform distribution. The initial density value reflects the prior judgment that the region contains a high-quality solution. For energy system planning problems, regions that meet the basic energy balance constraints can be assigned higher initial density values ​​based on experience, while regions that do not meet the basic constraints can be assigned lower initial density values. For example, grid cell G4213, which represents a region with moderate power generation capacity and a reasonable energy storage ratio, can be assigned an initial density value of 0.75. For grid cell G1122, which does not meet the energy balance constraints, an initial density value of 0.15 can be assigned.

[0124] When generating chaotic sequences, a chaotic system such as the Logistic mapping is used to produce sequences with pseudo-randomness and determinism. The generation of chaotic sequences relies on iterative computation, with an initial value set to 0.35 and a parameter value of 3.9. A chaotic sequence of length 1000 is generated through iteration. The first 10 values ​​of the sequence are [0.35, 0.89, 0.38, 0.92, 0.29, 0.81, 0.60, 0.94, 0.22, 0.67], distributed within the (0,1) interval, exhibiting good ergodicity and unpredictability. The chaotic sequence provides the foundational data for subsequent dynamic density adjustment and search direction determination.

[0125] When constructing the density function by multiplying the chaotic sequence by the initial density value, the elements in the chaotic sequence are multiplied by the initial density value of the corresponding grid cell to form the dynamic density function. This function represents the dynamic density value of each grid cell as equal to the initial density value of that cell multiplied by the corresponding chaotic sequence element. For example, for grid cell G4213, the initial density value is 0.75, and the corresponding chaotic sequence element is 0.38, then its dynamic density value is 0.75 × 0.38 = 0.285. For grid cell G1122, the initial density value is 0.15, and the corresponding chaotic sequence element is 0.92, then its dynamic density value is 0.15 × 0.92 = 0.138.

[0126] When dynamically adjusting the density of a region based on a density function, the weight and influence range of each grid cell in the feature space are adjusted according to its dynamic density value. The dynamic adjustment employs a density diffusion mechanism: high-density regions diffuse outwards, while low-density regions absorb the density values ​​from their surroundings. The adjustment process is iterative, updating the density values ​​of all grid cells in each iteration until a stable state is reached. For example, after 10 iterations, the density value of grid cell G4213 increases from 0.285 to 0.412, while the density value of grid cell G1122 decreases from 0.138 to 0.092. The results of the dynamic density adjustment reflect the importance distribution of each region in the feature space, which is helpful for subsequent structure reconstruction and search strategy optimization.

[0127] When reconstructing the structure based on the density distribution of the region partitioning, the feature space is re-divided according to the adjusted density values, forming an adaptive grid of uneven size. The grid in high-density regions is finer, while the grid in low-density regions is relatively coarser. The structural reconstruction uses hierarchical data structures such as quadtrees (2D) or octrees (3D), and determines whether to subdivide the grid based on a density threshold. For example, setting the density threshold to 0.3, grid cells with a density value higher than 0.3 are further subdivided, while grid cells with a density value lower than 0.3 remain unchanged or are merged. After structural reconstruction, the original grid cell G4213 (density value 0.412) is subdivided into 8 subgrids, while multiple grids in the low-density region are merged.

[0128] When generating adaptive feature regions, the reconstructed grid structure is converted into a data representation suitable for searching. The adaptive feature region contains information such as the spatial location, density value, and subdivision level of the grid, forming a structured dataset. For example, the reconstructed feature space can be represented as a four-level octree structure containing approximately 12,000 effective grid cells. Each cell stores its spatial boundary, center point coordinates, density value, and neighbor information. The adaptive feature region A5 formed by the high-density region contains approximately 4,000 fine grid cells, covering about 15% of the most promising volume in the feature space.

[0129] When inputting the adaptive feature region into the trained spatiotemporal evolution model, the spatial structure data needs to be converted into an input format acceptable to the model. The trained spatiotemporal evolution model is a computational model that can predict the spatiotemporal evolution characteristics of a regional energy system, pre-trained using historical data. The input process includes data format conversion, feature normalization, and batch processing preparation. For example, 4000 grid points in the adaptive feature region A5 are used as input samples. Each sample contains values ​​for 7 feature dimensions and grid density information, organized into a 4000×8 input matrix, and input into the spatiotemporal evolution model in batches.

[0130] When constructing search direction vectors based on chaotic sequences, the pseudo-randomness of the chaotic sequences is utilized to generate search directions with good coverage. The search direction vector consists of n components, each corresponding to one dimension of the feature space. The construction method involves mapping n consecutive values ​​from the chaotic sequence to the interval [-1, 1] to form a unit direction vector. For example, taking seven consecutive values ​​from the chaotic sequence [0.60, 0.94, 0.22, 0.67, 0.87, 0.44, 0.96], the resulting search direction vector is [0.20, 0.88, -0.56, 0.34, 0.74, -0.12, 0.92]. By continuously selecting different segments from the chaotic sequence, multiple different search direction vectors can be generated, covering various directions in the feature space.

[0131] When constructing the adaptive search step size by multiplying the chaotic sequence by the baseline step size, the search step size is dynamically adjusted based on the changing characteristics of the chaotic sequence. The baseline step size is set to 10% of the grid cell edge length, and the adaptive search step size is equal to the baseline step size multiplied by the element value in the chaotic sequence. For example, if the baseline step size is 0.05 and the corresponding chaotic sequence element is 0.67, then the adaptive search step size is 0.05 × 0.67 = 0.0335. The adaptive search step size is dynamically adjusted according to the changes in the chaotic sequence, enabling the search process to simultaneously accommodate large-scale exploration and local fine-grained search.

[0132] When traversing the adaptive feature region based on the search direction vector and the adaptive search step size, the search begins in a high-density area and proceeds along the search direction vector, with the step size being the adaptive search step size. At each step forward, the current position is recorded and added to the search sequence. When the search reaches the feature region boundary or meets the termination condition, a new search direction vector is selected to continue the search. For example, in the adaptive feature region A5, starting from the highest density point [0.72, 0.58, 0.63, 0.45, 0.81, 0.37, 0.69], the search proceeds along the search direction vector [0.20, 0.88, -0.56, 0.34, 0.74, -0.12, 0.92] for 15 steps, generating a search sequence containing 15 points. Throughout the entire search process, a total of 500 search sequences are generated, containing approximately 20,000 search points.

[0133] When obtaining the feature values ​​for each position in the search sequence, the coordinates of the search point are input into the spatiotemporal evolution model to calculate the corresponding energy system performance indicators. The feature values ​​include evaluation indicators across multiple dimensions, such as energy supply and demand balance, system economy, environmental friendliness, and reliability. For example, for the search point [0.74, 0.64, 0.59, 0.47, 0.86, 0.36, 0.72], the calculated feature values ​​are [0.92, 0.87, 0.75, 0.83], representing energy balance, economic indicators, environmental indicators, and reliability indicators, respectively. The feature values ​​reflect the performance of the energy system configuration scheme represented by the search point.

[0134] When calculating the convergence of a search sequence, the trend of eigenvalue changes between adjacent points is analyzed. Convergence is defined as the ratio of the average eigenvalue change in the latter half of the search sequence to that in the first half. For example, for a search sequence containing 15 points, if the average eigenvalue changes for the first 8 points and the last 7 points are 0.047 and 0.023 respectively, the convergence is 0.023 / 0.047 = 0.49. A smaller convergence indicates a more stable search sequence and better search performance.

[0135] When calculating eigenvalue diversity, we analyze the distribution range and variation patterns of eigenvalues ​​in the search sequence. Diversity is defined as the sum of the standard deviations of eigenvalues ​​across all dimensions. For example, for a search sequence containing 15 points, if the standard deviations of the eigenvalues ​​in the four dimensions are 0.08, 0.12, 0.06, and 0.09, then the diversity is 0.08 + 0.12 + 0.06 + 0.09 = 0.35. Higher diversity indicates that the search process has explored a wider solution space and avoided local optima.

[0136] When calculating the stability of the adaptive search step size, the changes in the step size during the search process are analyzed. Stability is defined as the ratio of the average step size change to the average step size. For example, for a search sequence containing 15 points, with an average step size of 0.028 and an average step size change of 0.005, the stability is 1 - 0.005 / 0.028 = 0.82. Higher stability indicates smoother step size adjustments during the search process and higher search efficiency.

[0137] When constructing a search efficiency evaluation index by weighting convergence, diversity, and stability, the weights of the three are assigned according to the characteristics of the search task. For example, when focusing on the accuracy of search results, the convergence weight can be set to 0.5, the diversity weight to 0.3, and the stability weight to 0.2; when focusing on the breadth of the search scope, the diversity weight can be increased. For a certain search sequence, its convergence is 0.49, diversity is 0.35, and stability is 0.82. If the weight allocation is [0.4, 0.4, 0.2], then the search efficiency evaluation index is 0.49×0.4+0.35×0.4+0.82×0.2=0.196+0.14+0.164=0.5.

[0138] When evaluating search sequences based on search efficiency metrics, an evaluation metric is calculated for all search sequences, and they are then sorted according to their metric values. A higher metric indicates higher search efficiency and a more valuable corresponding feature region. For example, if evaluation metrics are calculated for 500 search sequences, the top 10% of the sequences with the highest metric values ​​are considered efficient search sequences, and the feature space regions covered by these sequences are marked as candidate optimal search regions.

[0139] When determining the optimal search region in the adaptive feature region, the spatial distribution of efficient search sequences is analyzed to identify their overlapping and dense regions. These regions are then combined to form the final optimal search region. For example, 50 efficient search sequences are mainly distributed in three sub-regions R1, R2, and R3 of the feature space. Region R2 contains the most efficient search sequences (28) and has the highest average eigenvalue, therefore R2 is selected as the optimal search region. The optimal search region is represented as a hypercube in the seven-dimensional feature space [[0.68,0.78], [0.54,0.68], [0.58,0.66], [0.42,0.52], [0.76,0.88], [0.32,0.42], [0.65,0.75]], with a volume of approximately 2.3% of the original feature space, but containing the most promising energy system configuration schemes.

[0140] This technical method achieves efficient search of the feature space through chaos theory and adaptive grid technology, providing a high-quality set of candidate solutions for regional energy system planning, and significantly improving planning efficiency and solution quality.

[0141] In one optional implementation, based on the plurality of Pareto optimal solutions, a set of candidate planning schemes for the regional energy system is generated using the trained spatiotemporal evolution model; the set of candidate planning schemes is input into the trained spatiotemporal evolution model, and spatiotemporal dynamic simulation is performed on each planning scheme, including:

[0142] A knowledge graph is constructed based on multiple Pareto optimal solutions. Historical operational experience, domain expert knowledge, and technical standards and specifications are extracted from the knowledge graph. A feature mapping function is constructed based on the historical operational experience, the domain expert knowledge, and the technical standards and specifications.

[0143] For each Pareto optimal solution among the plurality of Pareto optimal solutions, the frequency of use and the degree of novelty are calculated, and an evaluation result of the feature mapping function is constructed based on the frequency of use and the degree of novelty;

[0144] The evaluation results of the feature mapping function are used to evaluate multiple Pareto optimal solutions, the value gradient of the feature mapping function is calculated, and the strategy is updated according to the value gradient generation scheme.

[0145] Based on the proposed scheme update strategy, a set of candidate planning schemes is generated, and the set of candidate planning schemes is input into a spatiotemporal evolution model. Spatial and temporal features are extracted from the set of candidate planning schemes, and the spatial and temporal features are input into the spatiotemporal evolution model for spatiotemporal dynamic simulation.

[0146] When constructing a knowledge graph based on multiple Pareto optimal solutions, each Pareto optimal solution is represented as a multi-dimensional vector, with each dimension corresponding to the optimization objective. For a regional energy system planning problem, there are 10 Pareto optimal solutions P1 to P10, each containing optimization objective values ​​in 5 dimensions. The optimization objective values ​​for P1 are [0.82, 0.65, 0.73, 0.58, 0.91], representing economic indicators, environmental indicators, reliability indicators, flexibility indicators, and comprehensive benefit indicators, respectively. The similarity between solutions is calculated as the weight of edges in the knowledge graph. The similarity between P1 and P2 is 0.87, indicating that the characteristics of these two solutions are relatively similar. The knowledge graph is visualized as a network structure, intuitively showing the relationships between Pareto optimal solutions.

[0147] When extracting historical operational experience from the knowledge graph, the high-frequency nodes and their common characteristics were analyzed. It was found that solutions P3, P5, and P8 all had high reliability indices (0.81, 0.79, and 0.85 respectively) and were frequently used in historical planning, leading to the conclusion that "high reliability has priority value in regional energy system planning." When extracting domain expert knowledge, the knowledge graph structure was interpreted by experts, who identified valuable planning patterns based on node distribution and connection patterns. Expert analysis revealed that the subgraph formed by P2, P4, and P9 exhibits the characteristic of "combining distributed energy with centralized energy supply," which was incorporated into the feature mapping function as domain knowledge. When extracting technical standards and specifications, knowledge graph nodes were compared with external standard libraries to select solutions that conformed to relevant technical specifications. The environmental protection indicators of all Pareto optimal solutions were examined, revealing that the environmental protection indicators of P1, P3, P5, P7, and P10 were all higher than 0.7, meeting high-standard environmental protection requirements.

[0148] When constructing the feature mapping function, a multi-layered structure is designed: the input layer receives the original features of the Pareto optimal solution, the intermediate layer performs feature transformation based on extracted knowledge, and the output layer generates the evaluation result. Historical operating experience is transformed into feature weights: reliability index weight is set to 0.3, economic index weight to 0.25, environmental index weight to 0.2, flexibility index weight to 0.15, and comprehensive benefit index weight to 0.1. Domain expert knowledge is transformed into feature combination rules: "When the reliability index is greater than 0.8 and the economic index is greater than 0.7, an additional 0.1 is added to the comprehensive score." Technical standards and specifications are transformed into constraints: "The environmental index must be greater than 0.6, otherwise the score is reduced by 50%." The constructed feature mapping function receives feature inputs from any Pareto optimal solution and outputs the evaluation value of that solution.

[0149] To calculate usage frequency, the frequency of each Pareto optimal solution in the knowledge graph is counted and divided by the total frequency to obtain the normalized usage frequency. In 100 historical programming iterations, P1 was used 25 times, with a usage frequency of 0.25. To calculate novelty, the average similarity of a solution to all other solutions is calculated, and this value is subtracted from 1 to obtain the novelty level. P7 has an average similarity of 0.65 with other solutions, and its novelty level is 1 - 0.65 = 0.35.

[0150] When constructing the evaluation results of the feature mapping function, a weighted average method is used, with the weights dynamically adjusted according to planning needs. When the planning tends to select validated mature solutions, the usage frequency weight is set to 0.7 and the novelty weight to 0.3; when the planning seeks innovative solutions, the usage frequency weight is set to 0.3 and the novelty weight to 0.7. P2 has a usage frequency of 0.18 and a novelty of 0.42, with weights of 0.6 and 0.4 respectively, and its evaluation result is 0.18×0.6+0.42×0.4=0.108+0.168=0.276. Evaluation results are calculated for all Pareto optimal solutions, forming an evaluation result sequence.

[0151] When evaluating multiple Pareto optimal solutions, the evaluation results of each solution are ranked, and the solutions with the highest evaluation results are selected as the core solution set. The evaluation results of the 10 Pareto optimal solutions are ranked as [0.42(P5), 0.38(P3), 0.35(P8), 0.32(P1), 0.30(P7), 0.28(P4), 0.276(P2), 0.25(P6), 0.22(P10), 0.20(P9)], and the top 5 solutions [P5, P3, P8, P1, P7] are selected as the core solution set.

[0152] When calculating the value gradient of the feature mapping function, the difference between the evaluation results of each solution and its k nearest neighbor solutions is calculated to form a local gradient vector. The three nearest neighbor solutions of P5 are P3, P8, and P1, with evaluation result differences of 0.42-0.38=0.04, 0.42-0.35=0.07, and 0.42-0.32=0.1, respectively, and the local gradient vector is [0.04, 0.07, 0.1].

[0153] When generating the update strategy, the average gradient of each dimension is calculated to determine the dominant gradient direction. The average gradient of the five dimensions of P5 is [0.03, 0.05, 0.02, 0.04, 0.01], and the dominant gradient direction is the second dimension (environmental indicator). The update strategy is "increase the value of the environmental indicator, keep other indicators unchanged or make slight adjustments". An update strategy is generated for each solution in the core solution set, forming a strategy set.

[0154] When generating the candidate planning scheme set, a directed mutation and combination method is used to adjust and combine the solutions in the core solution set according to an update strategy. For P5, the strategy of "increasing environmental protection index values" is applied, generating the mutated solution P5'=[0.81, 0.70, 0.73, 0.58, 0.90] (the environmental protection index increases from 0.65 to 0.70). High-value solution pairs are combined using features; the features of P3 and P8 are combined in a 6:4 ratio to generate the combined solution P3-8=[0.79, 0.68, 0.80, 0.61, 0.88]. Through mutation and combination, 15 candidate planning schemes are finally generated.

[0155] When inputting candidate planning schemes into the spatiotemporal evolution model, the schemes are converted into a format acceptable to the model, and simulation parameters and boundary conditions are set. Candidate planning scheme P5' is represented by specific parameters such as energy equipment configuration, network topology, and control strategy. The simulation duration is set to 8760 hours (one year), the spatial scope covers the entire planning area, and load data uses typical year data. The spatiotemporal evolution model independently simulates each candidate planning scheme, generating a spatiotemporal evolution trajectory.

[0156] When extracting spatial features, attention is paid to the spatial distribution of energy equipment, the topology of the energy network, and the location distribution of load centers. For P5', the extracted spatial features include the geographical coordinates of 10 energy stations, the connection relationships of 15 pipelines, and the locations of 5 load centers. When extracting temporal features, attention is paid to the temporal variation patterns of energy supply and demand, the dynamic characteristics of equipment operation, and the time-series performance of system response. For P5', the extracted temporal features include 8760-hour electricity load curves, heat load curves, equipment output curves, and system efficiency curves.

[0157] Spatiotemporal dynamic simulations simulate the operational status and performance of the proposed scheme during the planning period, generating detailed spatiotemporal dynamic data. The simulation content includes energy supply and demand balance, system response characteristics, economic indicators, environmental indicators, and reliability indicators. Simulations of P5' yield data on hourly energy balance rate, system efficiency, operating costs, carbon emissions, and failure rate throughout the year. Simulation results are visualized as time series graphs, spatial distribution maps, and indicator radar charts, intuitively displaying the spatiotemporal dynamic characteristics and overall performance of the proposed scheme. By conducting spatiotemporal dynamic simulations on all candidate planning schemes, the advantages and disadvantages of each scheme are comprehensively evaluated, providing a scientific basis for the final scheme selection.

[0158] This technical approach integrates multiple technologies such as knowledge graphs, feature mapping, gradient analysis, and spatiotemporal simulation to ensure that the generated planning scheme is both theoretically optimal and practically feasible, providing effective technical support for regional energy system planning.

[0159] In one optional implementation, for each of the plurality of Pareto optimal solutions, the frequency of use and the degree of novelty are calculated, and the evaluation result of the feature mapping function is constructed based on the frequency of use and the degree of novelty, including:

[0160] Establish pheromone concentrations for multiple Pareto optimal solutions, calculate the usage frequency of each Pareto optimal solution based on the pheromone concentrations, and dynamically fuse the usage frequency with the path selection factor through an adaptive threshold processor to form a corrected usage frequency.

[0161] The final usage frequency value of each Pareto optimal solution is generated based on the corrected usage frequency; the similarity between the multiple Pareto optimal solutions is calculated, and the minimum similarity value corresponding to each Pareto optimal solution and the maximum similarity value between the multiple Pareto optimal solutions are identified by the adaptive threshold processor to generate the novelty.

[0162] Calculate the information entropy of the final usage frequency value and the novelty level, and determine the weights of the final usage frequency value and the novelty level based on the information entropy; use the adaptive threshold processor to combine the weights with the final usage frequency value and the novelty level to generate a feature mapping function evaluation result.

[0163] like Figure 3 As shown, the method includes:

[0164] Multiple Pareto optimal solutions are established, with pheromone concentrations reflecting the historical selection history of solutions. This can be achieved by accumulating the number of times each Pareto optimal solution is selected during the iteration process. Specifically, for each Pareto optimal solution, the total number of times it is selected in all iterations is recorded, and this total number is divided by the total number of iterations to obtain the initial pheromone concentration value. For example, in a multi-objective resource scheduling optimization problem, there are 5 Pareto optimal solutions P1, P2, P3, P4, and P5, with a total of 100 iterations. They are selected 35, 25, 20, 15, and 5 times respectively, so the initial pheromone concentrations are 0.35, 0.25, 0.20, 0.15, and 0.05. The initial pheromone concentration is dynamically updated during the algorithm iteration process. After each iteration, the pheromone concentration of selected solutions increases, while the pheromone concentration of unselected solutions decreases, reflecting a positive feedback mechanism.

[0165] Calculating the usage frequency of each Pareto optimal solution based on pheromone concentration involves converting pheromone concentration into a comparable usage frequency index. This calculation considers the relative magnitude of the pheromone concentration, rather than its absolute value. Specifically, the pheromone concentration of each Pareto optimal solution is divided by the sum of the pheromone concentrations of all Pareto optimal solutions to obtain a normalized usage frequency. Continuing the example above, the sum of the pheromone concentrations of the five Pareto optimal solutions is 0.35 + 0.25 + 0.20 + 0.15 + 0.05 = 1.0. Therefore, their usage frequencies are 0.35 / 1.0 = 0.35, 0.25 / 1.0 = 0.25, 0.20 / 1.0 = 0.20, 0.15 / 1.0 = 0.15, and 0.05 / 1.0 = 0.05, respectively. The usage frequency reflects the acceptability of the Pareto optimal solution; a higher value indicates that the solution was selected more frequently in historical iterations.

[0166] The adaptive threshold processor dynamically fuses usage frequency and path selection factor to prevent the algorithm from prematurely converging to a local optimum. The path selection factor is a dynamically adjusted parameter used to balance the impact of usage frequency and encourage the algorithm to explore the insufficiently explored solution space. The adaptive threshold processor automatically adjusts the fusion weights based on the current iteration state, increasing the weight of the path selection factor when the algorithm tends to converge and decreasing it when it does not. In specific implementation, an adaptive threshold t can be set based on the ratio of the current iteration count to the maximum iteration count. When the usage frequency exceeds the threshold t, its impact on the final result is reduced; when the usage frequency is below the threshold t, its impact on the final result is increased. For example, in the 50th iteration (out of a total of 100 iterations), the adaptive threshold t can be set to 0.5. For P1, which has a usage frequency of 0.35, its corrected usage frequency is 0.35 × (1 - 0.5 × (0.35 - 0.5)) = 0.33; for P5, which has a usage frequency of 0.05, its corrected usage frequency is 0.05 × (1 + 0.5 × (0.5 - 0.05)) = 0.06. This correction mechanism somewhat suppresses solutions with high usage frequency and somewhat enhances solutions with low usage frequency, thus helping to maintain the diversity of solutions.

[0167] The final usage frequency value generated for each Pareto optimal solution based on the corrected usage frequency needs to be normalized to ensure that the sum of all final usage frequency values ​​is 1. Normalization is achieved by dividing the corrected usage frequency of each Pareto optimal solution by the sum of all corrected usage frequencies. Continuing the example above, assuming the corrected usage frequencies of the five Pareto optimal solutions are 0.33, 0.24, 0.20, 0.17, and 0.06, with a sum of 1.0, their final usage frequency values ​​remain unchanged at 0.33, 0.24, 0.20, 0.17, and 0.06, respectively. These final usage frequency values ​​will be used for subsequent feature mapping function evaluation.

[0168] The novelty of a solution is assessed by calculating the similarity between multiple Pareto optimal solutions. Similarity is calculated by comparing the distances between Pareto optimal solutions in the decision space. For two Pareto optimal solutions, their differences across each dimension are calculated, and these differences are combined to obtain the overall similarity. Specifically, methods such as Euclidean distance, Manhattan distance, or cosine similarity can be used. For example, for two Pareto optimal solutions P1=[0.7, 0.3, 0.5] and P2=[0.6, 0.4, 0.5] in a resource scheduling problem (where each dimension represents the allocation ratio of a resource), their similarity is calculated using Euclidean distance. The higher the similarity, the more similar the two solutions are; the lower the similarity, the greater the difference between the two solutions.

[0169] An adaptive threshold processor identifies the minimum similarity value corresponding to each Pareto optimal solution and the maximum similarity value among multiple Pareto optimal solutions, generating a novelty score. For each Pareto optimal solution, it calculates its similarity to all other Pareto optimal solutions, finds the minimum similarity value, representing the maximum difference between this solution and other solutions. Simultaneously, it finds the maximum similarity value among all pairs of Pareto optimal solutions, representing the upper limit of similarity across the entire solution set. The novelty score is normalized by subtracting the minimum similarity value from 1 and then dividing by (1 - maximum similarity value), resulting in a value between [0,1]. Continuing the example above, assuming the minimum similarity of P1 to other solutions is 0.65 and the maximum similarity among all solution pairs is 0.95, then the novelty score of P1 is (1 - 0.65) / (1 - 0.95) = (0.35 / 0.05) = 7, which becomes 0.88 after normalization to the [0,1] interval. A higher novelty score indicates a more unique solution and a greater difference from other solutions.

[0170] The information entropy of the final usage frequency and novelty is calculated to determine the weights of these two indicators in the feature mapping function. Information entropy reflects the degree of uncertainty of the data; the higher the entropy value, the more uniform the data distribution and the greater the amount of information contained. For the two sets of data—final usage frequency and novelty—the information entropy is calculated separately. The information entropy calculation process is as follows: the data is normalized to the [0,1] interval; for each data point, its probability is calculated (i.e., the value is divided by the sum of all values); then the logarithm of the probability is multiplied by the negative of the probability itself; finally, all results are summed. For example, for the final usage frequency values ​​[0.33, 0.24, 0.20, 0.17, 0.06], the information entropy is calculated as -(0.33×log(0.33)+0.24×log(0.24)+0.20×log(0.20)+0.17×log(0.17)+0.06×log(0.06))=1.53; for the novelty level [0.88, 0.72, 0.65, 0.58, 0.42], the information entropy is calculated as -(0.27×log(0.27)+0.22×log(0.22)+0.20×log(0.20)+0.18×log(0.18)+0.13×log(0.13))=1.58 (here, the novelty level is first normalized to a probability distribution).

[0171] The weights of the final usage frequency and novelty are determined based on information entropy to balance the influence of these two indicators in the feature mapping function. The weight allocation uses the entropy weighting method, determining the weight based on the relative magnitude of the information entropy. Specifically, the calculation method is as follows: subtract 1 from the information entropy of each indicator to obtain its information utility value. Then, divide the information utility value of each indicator by the sum of the information utility values ​​of all indicators to obtain the weight of that indicator. Continuing the example above, the information entropy of the final usage frequency is 1.53, and the information entropy of novelty is 1.58. Therefore, their information utility values ​​are 1-1.53=-0.53 and 1-1.58=-0.58 respectively, and the sum of their information utility values ​​is -1.11. Thus, the weight of the final usage frequency is (-0.53) / (-1.11)=0.48, and the weight of novelty is (-0.58) / (-1.11)=0.52. This weight allocation reflects the principle that indicators with higher information content should receive lower weights, because high entropy implies weaker discriminative power.

[0172] The adaptive threshold processor combines weights with the final usage frequency and novelty to generate a feature mapping function evaluation result. The feature combination uses a weighted average method, multiplying the final usage frequency and novelty of each Pareto optimal solution by their respective weights, and then summing them to obtain a comprehensive score. To adapt to the characteristics of different problems, the adaptive threshold processor can dynamically adjust the weight allocation based on performance during the iteration process. Specifically, an adaptive coefficient α based on the convergence speed can be set. When the algorithm converges quickly, the weight of novelty is increased to promote exploration of unexplored regions; when the algorithm converges slowly, the weight of usage frequency is increased to accelerate convergence. For example, in the middle of the iteration process (α=0.2), the adjusted weights of the final usage frequency and novelty are 0.48×(1-0.2)=0.384 and 0.52×(1+0.2)=0.624, respectively. Using these adjusted weights, the feature mapping function evaluation result for P1 is 0.33×0.384+0.88×0.624=0.675. The evaluation results of the feature mapping function provide a comprehensive evaluation index for each Pareto optimal solution. The higher the value, the better the overall performance of the solution in terms of usage frequency and novelty.

[0173] Through the above technical steps, a comprehensive evaluation of multiple Pareto optimal solutions is achieved, yielding a feature mapping function evaluation result that balances usage frequency and novelty. This method considers not only the frequency with which solutions are historically selected but also the uniqueness of solutions, providing an effective solution for multi-objective optimization problems. In practical applications, the feature mapping function evaluation result can guide the algorithm to select more valuable Pareto optimal solutions, improving the efficiency and quality of multi-objective optimization.

[0174] In one optional implementation, based on the results of the spatiotemporal dynamic simulation, a comprehensive score is calculated for each planning scheme, and the planning scheme with the highest comprehensive score is selected as the final regional energy system planning scheme, including:

[0175] Based on the results of the spatiotemporal dynamic simulation, the spatiotemporal features of each planning scheme are extracted, and the time decay weight of the spatiotemporal features is calculated. The spatiotemporal features are weighted based on the time decay weight to obtain the weighted values ​​of the spatiotemporal features of the planning scheme, and the correlation coefficient between the weighted values ​​of the spatiotemporal features is calculated.

[0176] The feature membership degree of the planning scheme is determined based on the correlation coefficient. The feature membership degree is multiplied by the corresponding feature value to obtain the basic score. The basic score is dynamically corrected based on the periodic change pattern.

[0177] Evidence enhancement is performed on the basic scores to obtain a comprehensive score for each planning scheme. The planning scheme with the highest comprehensive score is selected as the final regional energy system planning scheme.

[0178] The extraction process of spatiotemporal dynamic simulation results involves exporting detailed data for each planning scheme from the energy system simulation platform. Specifically, for alternative planning schemes A, B, and C of an industrial park, the spatiotemporal features to be extracted include five key categories: power load balance, thermal load balance, renewable energy output, energy storage equipment status, and energy conversion efficiency. Each category of features contains 8760 hours of time-series data for a complete year, forming a multidimensional spatiotemporal feature matrix. For example, for power load balance, the 24-hour data for planning scheme A on a typical working day are 92%, 91%, 90%, 89%, 88%, 87%, 88%, 90%, 93%, 95%, 96%, 97%, 98%, 97%, 96%, 95%, 94%, 93%, 94%, 95%, 94%, 93%, 92%, and 91%. The extracted spatiotemporal feature data is saved as a structured data table for subsequent analysis and processing.

[0179] The time decay weight calculation aims to reflect the varying importance of different time points to the system evaluation. For each time point within the planning period, its weight decreases as the distance from the baseline time point increases. An exponential decay function is used, with the baseline time point's weight set to 1. The weight of each subsequent time point is obtained by multiplying the baseline weight by a power of the decay factor. For a 15-year planning period, if the decay factor is set to 0.95, the weight value for year n is equal to 1 multiplied by 0.95 to the power of n-1. This weight design ensures that the effects of near-term planning have a greater impact on the evaluation results, aligning with the actual needs of planning. In practical implementation, a time weight vector can be established to store the weight value of each time point within the planning period for subsequent spatiotemporal feature weighting processing.

[0180] Spatiotemporal feature weighting involves multiplying the original spatiotemporal feature data by their corresponding time weights to obtain weighted feature values ​​that consider time importance. For each feature category of each planning scheme, its feature value at each time point is multiplied by the corresponding time point's weight value, forming a weighted feature sequence. Taking power load balance as an example, if the annual average values ​​of planning scheme A in years 1, 5, 10, and 15 of the planning period are 92%, 89%, 85%, and 80%, respectively, with corresponding time weights of 1.0, 0.8145, 0.6302, and 0.4877, then the weighted feature values ​​are 92%, 72.49%, 53.57%, and 39.02%, respectively. The weighted feature data more realistically reflects the performance of each planning scheme at different time points and its impact on the final evaluation.

[0181] The correlation coefficient is calculated to quantify the similarity between the characteristics of different planning schemes and the ideal reference. A feature sequence of the ideal reference scheme is established, which can be based on historical best data and theoretical best performance. Then, the correlation coefficient between the weighted feature sequence of each actual planning scheme and the ideal reference sequence is calculated. The correlation coefficient is calculated using the Pearson correlation coefficient method, obtained by dividing the covariance of the two sequences by the product of their respective standard deviations. The result value is between -1 and 1, with a value closer to 1 indicating a higher similarity. For example, the correlation coefficients between the weighted feature sequences of power load balance of planning schemes A, B, and C and the ideal reference sequence are 0.92, 0.88, and 0.76, respectively, indicating that scheme A is closest to the ideal scheme in this feature.

[0182] Determining feature membership degrees involves converting correlation coefficients into the degree of membership of planning schemes on each feature, establishing a mapping relationship between correlation coefficients and membership degrees, and dividing the correlation coefficient interval [0, 1] into multiple levels, each level corresponding to a membership degree value. Specifically, five levels can be set: correlation coefficients between 0.9 and 1.0 have a membership degree of 0.95; correlation coefficients between 0.8 and 0.9 have a membership degree of 0.85; correlation coefficients between 0.7 and 0.8 have a membership degree of 0.75; correlation coefficients between 0.6 and 0.7 have a membership degree of 0.65; and correlation coefficients between 0 and 0.6 have a membership degree of 0.5. Based on this mapping rule, the correlation coefficients of planning schemes on each feature can be converted into corresponding membership degree values, forming a membership degree matrix.

[0183] The basic score is calculated by weighting the feature membership degree and feature weight to determine the importance weight of each feature in the evaluation system. For example, energy supply and demand balance is 30%, system response time is 20%, peak-valley difference is 15%, energy utilization efficiency is 25%, and equipment lifespan is 10%. Then, the membership degree of the planning scheme on each feature is multiplied by the corresponding feature weight, and all products are summed to obtain the basic score of the planning scheme. For example, if the membership degrees of planning scheme A on the five features are 0.95, 0.85, 0.75, 0.85, and 0.85 respectively, its basic score is 0.95×30% + 0.85×20% + 0.75×15% + 0.85×25% + 0.85×10% = 0.865, which is 86.5 points on a percentage scale. This scoring method comprehensively considers the differences in importance of each feature and can more objectively reflect the overall performance of the planning scheme.

[0184] The dynamic correction of periodic variation patterns aims to consider the impact of the periodic variation characteristics of energy system load on the evaluation of planning schemes, introducing two correction factors: seasonal adaptability and intraday adaptability. Seasonal adaptability reflects the planning scheme's ability to adapt to seasonal load changes, while intraday adaptability reflects the scheme's responsiveness to intraday load fluctuations. The two factors are assigned different levels of correction coefficients based on the specific performance of the planning scheme. The seasonal adaptability correction coefficient is divided into five levels: Excellent 1.15, Good 1.05, Average 1.0, Poor 0.9, and Very Poor 0.8. The intraday adaptability correction coefficient is also divided into five levels: Excellent 1.1, Good 1.05, Average 1.0, Poor 0.95, and Very Poor 0.9. The two correction coefficients are multiplied to obtain the comprehensive correction coefficient, and then the base score is multiplied by the comprehensive correction coefficient to obtain the corrected base score. For example, if the seasonal adaptability rating of planning scheme B is excellent (1.15) and the intraday adaptability rating is good (1.05), then the comprehensive correction coefficient is 1.15×1.05=1.21, and the corrected base score is 87.5×1.21=105.88 points.

[0185] The evidence enhancement stage improves the reliability and persuasiveness of the score by introducing various auxiliary evaluation data. Four types of auxiliary evaluation data are collected: historical data verification, expert evaluation, case reference, and user feedback. Historical data verification compares the proposed plan with the best historical plan to obtain a similarity score. Expert evaluation solicits opinions from industry experts to obtain expert scores. Case reference compares the plan with successful cases to obtain reference scores. User feedback collects potential users' acceptance of each plan through questionnaires to obtain user scores. The weights of the four data types can be set at 25%, 30%, 25%, and 20%, respectively. The weighted average of the scores for each data type yields the evidence enhancement score. For example, if the four auxiliary evaluation scores for plan A are 85, 88, 82, and 90, its evidence enhancement score is 85×25% + 88×30% + 82×25% + 90×20% = 86.15 points.

[0186] The comprehensive score is calculated by integrating the adjusted baseline score and the evidence enhancement score, determining the weighting ratio of the two parts (e.g., 60% for the adjusted baseline score and 40% for the evidence enhancement score), and then calculating the weighted average to obtain the final comprehensive score. For example, if the adjusted baseline score for planning scheme B is 105.88 and the evidence enhancement score is 89.3, then its comprehensive score is 105.88 × 60% + 89.3 × 40% = 99.25. After calculating the comprehensive scores for all planning schemes, the scheme with the highest score is selected as the final regional energy system planning scheme. In this example, planning scheme B has the highest comprehensive score, therefore scheme B is selected as the final scheme.

[0187] Sensitivity analysis was used to verify the stability of the scoring method. This was achieved by making small adjustments to key parameters and observing the degree of change in the final score. Fluctuations of ±10% were applied to the time decay weight, feature weight, correction coefficient, and evidence weight, and the changes in the overall score of each planning scheme were recorded. If the score changes were small and did not alter the final scheme selection, the scoring method demonstrated good stability. For example, if the overall score changes of the three schemes by ±2.1%, ±2.4%, and ±1.9% respectively when the time decay weight fluctuated by ±10%, and scheme B still received the highest score, then the scoring method was not sensitive to changes in the time decay weight and exhibited strong stability.

[0188] Robustness testing assesses the adaptability of selected schemes under uncertain conditions. Multiple disturbance scenarios are designed, and the performance changes of each planning scheme are observed. Common disturbance scenarios include energy price fluctuations (±25%), load forecasting deviations (±20%), and extreme weather conditions. For each scenario, the rate of change of key performance indicators of the planning scheme is calculated, such as economic indicators, system adaptability indicators, and system reliability indicators. The lower the rate of change, the stronger the robustness of the scheme. For example, under the energy price fluctuation scenario, if the rate of change of the economic indicator of Scheme B is 13.5%, lower than that of Scheme A (17.8%) and Scheme C (19.2%), then Scheme B has stronger economic robustness.

[0189] Through the detailed technical steps described above, the evaluation and selection process of regional energy system planning schemes becomes systematic, transparent, and quantifiable, providing reliable technical support for energy system planning decisions and effectively improving the scientific rigor and applicability of planning schemes. The entire method begins with spatiotemporal feature extraction, proceeds through time weight calculation, feature weighting processing, correlation coefficient calculation, membership degree determination, basic score calculation, dynamic correction, and evidence enhancement, ultimately yielding a comprehensive score. The reliability of the results is verified through sensitivity analysis and robustness testing, forming a complete method for selecting regional energy system planning schemes.

[0190] A second aspect of the present invention provides an electronic device, comprising:

[0191] processor;

[0192] Memory used to store processor-executable instructions;

[0193] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0194] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0195] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0196] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent planning of the entire life cycle of regional energy systems based on spatiotemporal evolution, characterized in that, include: Acquire historical energy data for the target region, construct a spatiotemporal evolution model of the regional energy system, and train the spatiotemporal evolution model based on the historical energy data to obtain a trained spatiotemporal evolution model, including: The set of graph elements in the spatiotemporal evolution model is used to form a dynamic knowledge graph. Temporal features are obtained from historical data of the regional energy system. The correlation strength between nodes is calculated based on the temporal features. The correlation relationship in the dynamic knowledge graph is updated according to the correlation strength. The relationships between nodes in the dynamic knowledge graph are analyzed, the degree of mutual influence between nodes is calculated, and the degree of mutual influence is written into the edge set of the dynamic knowledge graph to obtain a dynamic knowledge graph containing the influence relationships. Feature extraction is performed on each node in the dynamic knowledge graph containing influence relationships, and a node feature vector is generated based on the node's neighborhood information; the node feature vector is fused with the spatiotemporal features in historical energy data, and the fusion ratio of the node feature vector and the spatiotemporal features is adjusted by weighting coefficients to obtain fused features containing knowledge graph information; Based on the fusion features, feature analysis is performed on energy system data from different regions to calculate the feature differences between the data from different regions. Regional feature alignment is achieved by reducing the feature differences. Feature optimization is then performed on the data after regional feature alignment to generate feature data for model training. The feature data and the feature differences are weighted and combined to obtain the optimization target. The spatiotemporal evolution model is trained by optimizing the optimization target to obtain a trained spatiotemporal evolution model with transferability. Based on the trained spatiotemporal evolution model, the energy facility costs and energy benefits in the historical energy data are analyzed. The costs and energy benefits are input into the trained spatiotemporal evolution model, and multiple Pareto optimal solutions are obtained by using a multi-objective optimization algorithm. Based on the multiple Pareto optimal solutions, a set of candidate planning schemes for the regional energy system is generated using the trained spatiotemporal evolution model. The set of candidate planning schemes is then input into the trained spatiotemporal evolution model, and spatiotemporal dynamic simulation is performed on each planning scheme. Based on the results of the spatiotemporal dynamic simulation, a comprehensive score for each planning scheme is calculated, and the planning scheme with the highest comprehensive score is selected as the final regional energy system planning scheme.

2. The method according to claim 1, characterized in that, Based on the trained spatiotemporal evolution model, the energy facility costs and energy benefits in the historical energy data are analyzed. The costs and energy benefits are input into the trained spatiotemporal evolution model, and a multi-objective optimization algorithm is used to obtain multiple Pareto optimal solutions, including: Extract the cycle operating cost of energy facilities from historical energy data, and construct the first input parameter of the optimization function based on the cycle operating cost; extract the input power and output power of energy facilities from the historical energy data, calculate the efficiency index of energy facilities based on the input power and output power, and construct the second input parameter of the optimization function based on the efficiency index; The first input parameter and the second input parameter are input into the trained spatiotemporal evolution model to construct a feature vector containing temporal and spatial features; the feature vector is divided into a grid to obtain a feature grid; and a coarse-grained search is performed on the trained spatiotemporal evolution model based on the feature grid to determine the optimal search region. Set a search center point in the optimal search area, calculate the search step size of the current search position based on the search center point, calculate the distance from the search center point to the current search position, and determine the search constraints based on the distance. Under the search constraints, the optimization function is iteratively optimized based on the search step size to obtain the Pareto optimal solution of the energy system.

3. The method according to claim 2, characterized in that, The feature vectors are divided into a grid to obtain a feature grid. Based on the feature grid, a coarse-grained search is performed in the trained spatiotemporal evolution model to determine the optimal search region, including: Obtain an n-dimensional feature vector, divide the n-dimensional feature vector into regions in the feature space, initialize the density of the region division result with an initial density value to generate a chaotic sequence, construct a density function by multiplying the chaotic sequence with the initial density value, and dynamically adjust the density of the region based on the density function to obtain a density distribution. Based on the density distribution, the results of the region division are restructured to generate adaptive feature regions, and the adaptive feature regions are input into the trained spatiotemporal evolution model. In the trained spatiotemporal evolution model, a search direction vector is constructed based on the chaotic sequence, and an adaptive search step size is constructed by multiplying the chaotic sequence with the baseline step size; the adaptive feature region is traversed according to the search direction vector and the adaptive search step size to obtain a search sequence, and the feature value of each position in the search sequence is obtained. The convergence of the search sequence, the diversity of the feature values, and the stability of the adaptive search step size are calculated. The weighted sum of the convergence, diversity, and stability is used to construct a search efficiency evaluation index. The search sequence is evaluated based on the search efficiency evaluation index, and the optimal search region is determined in the adaptive feature region.

4. The method according to claim 1, characterized in that, Based on the multiple Pareto optimal solutions, a set of candidate planning schemes for the regional energy system is generated using the trained spatiotemporal evolution model. Inputting the set of candidate planning schemes into the trained spatiotemporal evolution model, and performing spatiotemporal dynamic simulation for each planning scheme includes: A knowledge graph is constructed based on multiple Pareto optimal solutions. Historical operational experience, domain expert knowledge, and technical standards and specifications are extracted from the knowledge graph. A feature mapping function is constructed based on the historical operational experience, the domain expert knowledge, and the technical standards and specifications. For each Pareto optimal solution among the plurality of Pareto optimal solutions, the frequency of use and the degree of novelty are calculated, and an evaluation result of the feature mapping function is constructed based on the frequency of use and the degree of novelty; The evaluation results of the feature mapping function are used to evaluate multiple Pareto optimal solutions, the value gradient of the feature mapping function is calculated, and the strategy is updated according to the value gradient generation scheme. Based on the proposed scheme update strategy, a set of candidate planning schemes is generated, and the set of candidate planning schemes is input into a spatiotemporal evolution model. Spatial and temporal features are extracted from the set of candidate planning schemes, and the spatial and temporal features are input into the spatiotemporal evolution model for spatiotemporal dynamic simulation.

5. The method according to claim 4, characterized in that, For each Pareto optimal solution among the plurality of Pareto optimal solutions, the frequency of use and the degree of novelty are calculated, and the evaluation result of the feature mapping function is constructed based on the frequency of use and the degree of novelty, including: Establish pheromone concentrations for multiple Pareto optimal solutions, calculate the usage frequency of each Pareto optimal solution based on the pheromone concentrations, and dynamically fuse the usage frequency with the path selection factor through an adaptive threshold processor to form a corrected usage frequency. The final usage frequency value of each Pareto optimal solution is generated based on the corrected usage frequency; the similarity between the multiple Pareto optimal solutions is calculated, and the minimum similarity value corresponding to each Pareto optimal solution and the maximum similarity value between the multiple Pareto optimal solutions are identified by the adaptive threshold processor to generate the novelty. Calculate the information entropy of the final usage frequency value and the novelty level, and determine the weights of the final usage frequency value and the novelty level based on the information entropy; use the adaptive threshold processor to combine the weights with the final usage frequency value and the novelty level to generate a feature mapping function evaluation result.

6. The method according to claim 1, characterized in that, Based on the results of the spatiotemporal dynamic simulation, a comprehensive score is calculated for each planning scheme, and the planning scheme with the highest comprehensive score is selected as the final regional energy system planning scheme, including: Based on the results of the spatiotemporal dynamic simulation, the spatiotemporal features of each planning scheme are extracted, and the time decay weight of the spatiotemporal features is calculated. The spatiotemporal features are weighted based on the time decay weight to obtain the weighted values ​​of the spatiotemporal features of the planning scheme, and the correlation coefficient between the weighted values ​​of the spatiotemporal features is calculated. The feature membership degree of the planning scheme is determined based on the correlation coefficient. The feature membership degree is multiplied by the corresponding feature value to obtain the basic score. The basic score is dynamically corrected based on the periodic change pattern. Evidence enhancement is performed on the basic scores to obtain a comprehensive score for each planning scheme. The planning scheme with the highest comprehensive score is selected as the final regional energy system planning scheme.

7. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images