An eco-driven structural pruning based power network sparse solution method
By employing a hierarchical pruning method using partitioning and elimination trees, the problem of repetitive calculations in power network analysis during the ECO stage is solved, achieving efficient sparse solutions and improving the ECO iteration efficiency of integrated circuit design.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2026-05-12
- Publication Date
- 2026-06-09
AI Technical Summary
In the ECO stage of integrated circuit design, existing technologies suffer from excessive computational overhead in power network analysis due to local modifications. Furthermore, existing incremental sparse direct solution methods are difficult to adapt to changes in matrix dimension or sparse structure, resulting in severe repetitive calculations and impacting design efficiency.
A hierarchical pruning method based on partition trees and elimination trees is adopted. Local changes in the power network are identified by path-aware dirty marking. Only dirty regions of reordering, symbolic decomposition and numerical decomposition are locally recalculated, while other parts reuse the results of the previous round, thus preserving the exact solution.
It significantly reduces redundant calculations in multiple ECO iterations, improves analysis efficiency, achieves an average speedup of 4.02 times compared to the full direct solver, and is widely applicable to local modifications, maintaining the consistency of the exact solution.
Smart Images
Figure CN122174411A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer-aided design and power integrity analysis technology of integrated circuits, and specifically relates to a sparse solution method for power networks based on structural pruning driven by Engineering Change Order (ECO). Background Technology
[0002] As process nodes continue to evolve, supply voltages decrease, and power density increases, voltage drop issues in power networks have an increasingly significant impact on integrated circuit performance and long-term reliability. To ensure power integrity meets design requirements and to promptly identify and mitigate potential voltage drop risks, continuous power network analysis is necessary throughout the integrated circuit design process. This analysis typically involves solving large-scale sparse symmetric positive definite linear equation systems, often employing direct solvers based on sparse direct decomposition to perform reordering, symbolic decomposition, numerical decomposition, and trigonometric solving. However, as network size continues to increase, the computational overhead of these solution processes also increases significantly.
[0003] In the later ECO phase of the design process, designers often need to iterate multiple times on modifications such as local routing, vias, metal widening, and metal striping. After each modification, they must re-execute the power network analysis to determine if any new voltage drop violations have been introduced. Although such modifications are usually localized, if the entire solution process is executed from scratch every time, a large amount of redundant calculations will be generated in consecutive iterations, thus becoming one of the main bottlenecks to rapid design convergence.
[0004] While existing incremental sparse direct solution techniques can reuse existing results to some extent, they typically only cover part of the solution process or impose strong restrictions on the types of modifications, such as requiring unchanged matrix dimensions, supporting only specific topological operations, or relying on a fixed sparse structure. Once an ECO (Extended Objective) causes changes in matrix dimensions or sparse structure, existing methods often struggle to simultaneously maintain applicability, reuse granularity, and accurate solution preservation. Therefore, there is an urgent need for an incremental sparse direct solution method that can handle general local ECO modifications, uniformly cover the three stages of reordering, symbolic decomposition, and numerical decomposition, and maintain the same accuracy as ab initio direct solutions. Summary of the Invention
[0005] To overcome the shortcomings of existing technologies, this invention aims to provide an ECO-driven sparse solution method for power networks based on structural pruning. This method uses a partition tree and an elimination tree as hierarchical structural carriers. It leverages the spatial locality characteristic of ECO modifications and the recursive partitioning of spatially adjacent nodes by nested decompositions. For local changes between old and new matrices, it performs path-aware impact propagation, precisely mapping changes to three levels of dirty labels: reordered dirty regions, symbolically decomposed dirty regions, and numerically decomposed dirty nodes. Based on this, it performs local recalculation only on the least necessary parts, directly reusing the cached results from the previous round for the remaining parts. Figure 1 Looking at the overall framework shown, the baseline rounds correspond to the complete solution process; subsequent rounds first perform change detection to identify five types of changes; then, path-aware dirty tags are generated in three stages: reordering, symbolic decomposition, and numerical decomposition; then, the path-aware structure pruning solution process is entered, and partial reordering, partial symbolic decomposition, and partial numerical decomposition are performed on the affected parts in sequence.
[0006] On one hand, this invention provides an ECO-driven method for solving sparse power networks based on structure pruning, comprising the following steps: Step 1: Perform ab initio solution for the power network admittance matrix of the baseline round to obtain the partition tree, elimination tree and decomposition factor, and cache them; Step 2: For the current round power network admittance matrix, detect matrix changes based on the mapping relationship from the current round node to the previous round node. The matrix changes include one or more of the following: adding nodes, deleting nodes, adding edges, deleting edges, and value changes. Step 3: Based on the old partition tree, determine the reordered dirty regions according to the matrix changes; Step 4: Perform local nested partitioning and reconstruction only on the subtrees corresponding to the reordered dirty regions, and merge them with the reordering results of the unaffected regions to obtain the current round's partition tree and reordering results; Step 5: Based on the new partition tree, identify the dirty regions for symbolic decomposition; Step 6: Perform local symbolic decomposition based on the nodes corresponding to the dirty regions of the symbolic decomposition, reuse the old elimination tree parent-child relationship and the sparse mode of the decomposition factor in the unaffected regions, and obtain the elimination tree and sparse structure of the decomposition factor in the current round. Step 7: Based on the new elimination tree, identify the dirty nodes in the numerical decomposition; Step 8: Perform local numerical updates on the columns corresponding to the affected dirty nodes, only recalculate the decomposition factor values of the corresponding dirty columns, and directly reuse the old decomposition factor values of the unaffected columns. Step 9: Use the updated decomposition factor to perform trigonometric solutions on the current round of linear equations to obtain the exact nodal voltage solutions of the current round of power network.
[0007] Furthermore, in step 2, matrix change detection identifies five types of changes based on the mapping relationship from the current round node to the previous round node and its corresponding reverse mapping. Among them, newly added and deleted nodes are directly given by the missing items in the mapping; newly added and deleted edges are obtained by comparing the adjacency relationship of the old and new matrices; and value changes are obtained by comparing the values of the corresponding non-zero entries.
[0008] Furthermore, the generation rules for reordered dirty regions in step 3 include: when a structural change falls into a leaf region, the corresponding leaf region is marked as a reordered dirty region; when a new node appears, the region to which the new node belongs is determined according to the lowest common ancestor of the regions to which its neighboring nodes belong; if the region to which it belongs is a partition region, the partition region is marked as a reordered dirty region; when a new edge connects two regions that do not have an ancestor-descendant relationship and destroys the partition tree structure, the region corresponding to the lowest common ancestor of the two regions is marked as a reordered dirty region.
[0009] Furthermore, step 4 specifically includes: first, performing ancestor filtering on the reordered dirty regions, retaining only the highest-level dirty ancestor regions; then, collecting all descendant nodes corresponding to the retained dirty ancestor regions as a set of nodes to be reconstructed; extracting the induced subgraph corresponding to the set of nodes to be reconstructed from the current round matrix, and performing nested subdivision reconstruction only on the induced subgraph.
[0010] Furthermore, step 5 specifically includes: marking all reordered regions in step 4 as symbolic decomposition dirty regions; for structural changes that do not trigger reordering but still change local adjacency relationships, marking their associated regions as symbolic decomposition dirty regions, and extending the symbolic decomposition dirty regions along the old partition tree to the root region to cover the ancestor regions affected by column merging properties.
[0011] Furthermore, in step 6, for nodes located in unaffected regions, the parent-child relationship of the old elimination tree and the sparse mode of the decomposition factor are reused directly through the index mapping between the old round permutation and the current round permutation; for nodes located in the dirty region of symbolic decomposition, their elimination tree parent node and the sparse mode of the decomposition factor are recalculated.
[0012] Furthermore, in step 7, the initial set of dirty nodes for numerical decomposition consists of the following parts: all nodes in the reordered dirty region, endpoint nodes directly affected by structural changes except for the aforementioned reordered failed nodes, and endpoint nodes of matrix entries directly affected by numerical changes; and the initial set is extended along the new elimination tree to the root node to cover the ancestor nodes affected by the column merging property, thereby determining the final set of dirty nodes for numerical decomposition.
[0013] Furthermore, in step 8, for columns that do not fall into the final set of dirty nodes in the numerical decomposition, the corresponding values are directly copied through the index mapping between the old round decomposition factor columns and the current round columns; for columns that fall into the final set of dirty nodes in the numerical decomposition, the column values are recalculated using the left-look method.
[0014] In a preferred embodiment, the solution method is applicable to symmetric positive definite sparse linear systems in power network analysis and maintains an accurate solution consistent with the direct solution from scratch during multiple rounds of ECO-driven iterations.
[0015] On the other hand, the present invention provides an electronic device including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, it runs the above-described ECO-driven sparse solution method for power networks based on structure pruning.
[0016] The present invention also provides a computer-readable storage medium storing computer instructions thereon, which, when executed by a processor, run the above-described ECO-driven sparse power network solution method based on structure pruning.
[0017] Compared with the prior art, the present invention has the following advantages: 1) It uniformly covers the three stages of reordering, symbolic decomposition, and numerical decomposition, avoiding the limitations of incremental updates in only a single stage; 2) It utilizes the spatial locality characteristic of ECO modifications and leverages the path-aware dirty mark propagation mechanism on the partition tree and elimination tree to accurately define the minimum recalculation range in a combination of region-level and node-level methods; 3) It supports general local ECO modifications such as adding nodes, deleting nodes, adding edges, deleting edges, and value changes, making it more applicable; 4) While maintaining the same accurate solution as direct solutions from scratch, it significantly reduces redundant calculations and improves the efficiency of multi-round ECO analysis; 5) Through comparative experiments with 12 test cases, it can be seen that the present invention can achieve an average speedup of about 4.02 times compared to the CHOLMOD full direct solver, and can achieve a higher cumulative speedup ratio compared to the comparative scheme that only performs incremental updates in a single stage, verifying the effectiveness of the three-stage collaborative pruning update. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the overall process framework of the present invention, which includes the complete solution process of the baseline round, the change detection of subsequent rounds, and the path-aware dirty tag generation and path-aware structure pruning solution process that unfolds around the three stages of reordering, symbolic decomposition and numerical decomposition. Figure 2 This is a schematic diagram of the nested partitioning and reordering algorithm of the present invention applied to a sparse graph structure and its corresponding matrix structure. Figure 3This is a schematic diagram of the partition tree obtained by nested decomposition and the elimination tree constructed by the symbol decomposition stage of the present invention; wherein, (a) represents the partition tree and (b) represents the elimination tree; Figure 4 This is a schematic diagram illustrating the structural pruning and updating of the partition tree and elimination tree under partial graph structure modification according to the present invention; wherein, (a) represents the old graph ( Figure 2 (a) represents the new graph obtained by modification, (b) represents the old partition tree with modification marks, (c) represents the partition tree after shrinking, (d) represents the partition tree after region re-partitioning, (e) represents the new partition tree, and (f) represents the new elimination tree; Figure 5 The figure shows the cumulative speedup results of the present invention and incremental solvers a and b relative to the CHOLMOD full direct solver on 12 test cases, varying with the number of ECO iterations. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0020] like Figure 1 As shown, Figure 1 The left side shows the complete solution process for the baseline rounds, the middle side shows the path-aware structure pruning solution process, and the right side shows the generation of path-aware dirty tags based on five types of changes. In subsequent rounds, change detection is performed first; then, path-aware dirty tags are generated for the three stages of reordering, symbolic decomposition, and numerical decomposition; and then the path-aware structure pruning solution process is executed based on these tags.
[0021] An ECO-driven sparse solution method for power networks based on structure pruning includes the following nine steps: Step 1: Perform the complete solution process for the power network admittance matrix of the baseline round, i.e., solve from scratch to obtain the partition tree, elimination tree and decomposition factor, and cache them.
[0022] In the baseline round, preprocessing, nested partitioning and reordering, symbolic decomposition, numerical decomposition, and triangulation are performed sequentially, and the partition tree structure, permutation vectors, elimination of parent-child relationships in the trees, sparse structure of the decomposition factors, and numerical values of the decomposition factors are cached. Specifically, preprocessing includes reading in the power network admittance matrix and right-hand side terms, converting the matrix into the CSC storage format required for subsequent solutions, and generating the graph format required for the nested partitioning and reordering algorithm METIS; the nested partitioning and reordering stage calls METIS to recursively partition the graph structure. Figure 2 This demonstrates the process of recursively partitioning the original graph to form separating nodes and their corresponding matrix structures, and from this, we obtain... Figure 3The partition tree shown in (a) is constructed based on the reordering matrix during the symbolic decomposition stage. Figure 3 The elimination tree shown in (b) is used to determine the sparse structure of the decomposition factor; the numerical decomposition stage calculates the value of the decomposition factor; and the triangular solution stage uses the decomposition factor to obtain the baseline round accurate node voltage solution.
[0023] Step 2: For the current round power network admittance matrix, detect matrix changes based on the mapping relationship between the current round node and the previous round node. The matrix changes include one or more of the following: adding nodes, deleting nodes, adding edges, deleting edges, and value changes.
[0024] Change detection simultaneously generates the node mapping relationships between the old and new matrices required for subsequent updates. The input node mapping is... This represents the correspondence between nodes in the current round and nodes in the previous round, where a mapping value of -1 indicates that the node is a newly added node in the current round; based on this mapping, a reverse mapping can be further obtained. The mapping value of -1 indicates that the old node has been deleted in the current round. Based on the above bidirectional mapping, newly added and deleted nodes can be directly identified by the missing mapping item. For newly added nodes, since they have no historical position in the old partition tree, the lowest common ancestor of the regions to which the neighboring nodes belong in the old partition tree is determined by checking their neighboring nodes in the current matrix and finding the regions to which the neighboring nodes belong. Newly added and deleted edges are identified by comparing the adjacency relationships of the new and old matrices, and changes in values are identified by comparing the values of the corresponding entries. Figure 4 Is Figure 2 and Figure 3 The example shown is a local modification instance constructed based on the example shown. In region 3, the edge (0, 1) of the original graph is deleted and nodes 20 and 25 are added. In region 4, node 20 of the original graph and its associated edge are deleted, and edges (5, 15) and (11, 15) are added. Figure 4 Examples of value changes are omitted because value changes do not alter the structure of the split tree and the elimination tree.
[0025] Step 3: Based on the old partition tree, determine the reordered dirty regions according to the matrix changes.
[0026] The reordering of dirty regions is determined according to the following rules: When a structural change falls into a leaf region, the corresponding leaf region is marked as a reordered dirty region; when a new node appears, its region is first determined based on the lowest common ancestor of its neighboring regions. If the region is a partitioned region, it is marked as a reordered dirty region; when a new edge connects two regions without an ancestor-descendant relationship and disrupts the partitioned tree structure, the region corresponding to the lowest common ancestor of the two regions is marked as a reordered dirty region; when a new edge connects regions with an ancestor-descendant relationship, cross-branch reordering is not triggered. Figure 4 (a) and Figure 4 In example (b), the structural changes first fall within regions 3 and 4. In region 3, the edge (0, 1) of the original graph is deleted, and nodes 20 and 25 are added. For the newly added nodes 20 and 25, their regions can be determined by the lowest common ancestor of their neighboring nodes' regions in the old partition tree. Figure 4 In the example shown, the region to which this belongs is region 3, and it participates in the subsequent dirty region determination accordingly. In region 4, the original graph node 20 and its associated edges are deleted. Based on the above rules, regions 3 and 4 are marked first; the newly added edge (5, 15) crosses different branches and disrupts the partition tree structure, triggering its lowest common ancestor corresponding region 1 to be further marked; the newly added edge (11, 15) is located within the ancestor-descendant path and does not trigger cross-branch reordering.
[0027] Step 4: In the path-aware structure pruning solution process, only the subtrees corresponding to the reordered dirty regions are partially reordered, and the reordering results are merged with those of the unaffected regions to obtain the current round's separator tree and reordering results.
[0028] like Figure 4 (c) and Figure 4 As shown in (d), ancestor filtering is first performed on the reordered dirty regions. If the ancestor region of a certain dirty region has already been selected as the region to be reconstructed, the local reconstruction process of the subtree corresponding to that ancestor region has already covered all its descendant dirty regions, and there is no need to repeat the reconstruction of the descendant dirty regions. Therefore, only the highest-level dirty ancestor is retained; in this example, only region 1 is retained. Subsequently, all nodes corresponding to region 1 are extracted to form an induced subgraph, which can be understood as shrinking and merging regions 3 and 4 into the subgraph to be reconstructed of region 1, and then re-performing nested partitioning on this subgraph. After reconstruction, it is merged with the clean subtree, and the unaffected regions 0, 2, 5, and 6 remain unchanged, resulting in the current round's split tree.
[0029] Step 5: Based on the new partition tree, identify the dirty regions of symbol decomposition.
[0030] like Figure 4As shown in (e), regions 1, 3, and 4 automatically become symbolic decomposition dirty regions after step 4. In addition, structural changes that do not trigger reordering but still alter local adjacency relationships must be considered. If a node in a partition region is deleted, the shrunk partition region still satisfies the nested partitioning objective, and local nested partitioning does not need to be re-executed. However, the local elimination order of nodes within this region has changed, so this partition region should still be marked as a symbolic decomposition dirty region. If adding or deleting an edge involves a partition region, and this change does not alter the partition tree structure, although it does not trigger reordering updates, it will still change the adjacency relationships of the relevant endpoint nodes. Therefore, the regions at both ends of the edge should be marked as symbolic decomposition dirty regions. Furthermore, considering the column merging property, symbolic decomposition dirty regions should extend along the old partition tree towards the root region. Therefore, in Figure 4 In the example shown, region 0 is synchronously included in the dirty region of symbol decomposition; regions 2, 5, and 6 remain clean and directly reuse the old symbol decomposition results.
[0031] Step 6: In the path-aware structure pruning solution process, partial symbol decomposition is performed only based on the nodes corresponding to the dirty regions of the symbol decomposition. The parent-child relationship of the old elimination tree and the sparse mode of the decomposition factor are reused in the unaffected regions to obtain the elimination tree and the sparse structure of the decomposition factor in the current round.
[0032] like Figure 4 As shown in (f), for clean region nodes, based on the node mapping relationship between the current round and the previous round, the current column index is mapped to the corresponding column in the old decomposition structure, and the old elimination tree parent-child relationship and decomposition factor sparsity mode are directly reused; for dirty region nodes, the parent-child relationship and column sparsity mode are reconstructed according to the current matrix structure, thereby avoiding the reconstruction of the entire elimination tree.
[0033] Step 7: Based on the new elimination tree, identify the dirty nodes in the numerical decomposition.
[0034] The initial set of dirty nodes for numerical decomposition includes: nodes within the reordered dirty region, endpoint nodes directly affected by structural changes other than the aforementioned reordered failed nodes, and endpoint nodes that only undergo value changes; then, it expands along the new elimination tree towards the root node to obtain the final set of dirty nodes for numerical decomposition. Figure 4 Taking the next scenario shown in (f) as an example, when the value of the relevant entry of node 1 changes, the dirty mark propagates along the path of node 1→node 6→node 2→node 7→node 12→node 17→node 22, and only the above nodes are marked as dirty nodes in numerical decomposition.
[0035] Step 8: In the path-aware structure pruning solution process, only the columns corresponding to the affected dirty nodes are partially updated with numerical decomposition. Only the decomposition factor values of the corresponding dirty columns are recalculated, and the old decomposition factor values of the unaffected columns are directly reused.
[0036] Step 8, the local numerical update, is performed using the left-look method. For each dirty column, the corresponding column of the current matrix is first loaded into the working vector. Then, based on the sparse structure of the current decomposition factor, its left-hand related columns are collected, and the numerical contribution of previous columns to the current column is eliminated column by column. Subsequently, the square root of the diagonal elements and the normalization of the off-diagonal elements of the current column are performed to obtain the updated decomposition factor values. For unaffected columns, the corresponding column values in the previous round of decomposition factors are directly mapped to the current round based on the mapping relationship between the old and new nodes, thus skipping the column modification and column normalization steps.
[0037] Step 9: Use the updated decomposition factor to perform trigonometric solutions on the current round of linear equations to obtain the exact nodal voltage solutions of the current round of power network.
[0038] In a multi-round ECO scenario, after completing steps 2 to 9 of the current round, the partition tree, elimination tree, decomposition factor and node mapping relationship obtained in the current round are cached as the old results for the next round, and the above steps are repeated for the matrix of the next round. This significantly reduces redundant calculations while maintaining the same accuracy as solving directly from scratch.
[0039] To verify the effectiveness of this invention, the solver was implemented in C++, and a partition tree was constructed using the METIS graph partitioning tool. Twelve power network test cases were selected for experiments. Each case first underwent one round of baseline full solution, followed by ten rounds of ECO iteration, with the CHOLMOD full direct solver used as the normalization benchmark. Figure 5 The cumulative speedup of this invention compared to two contrasting incremental solver schemes is presented. Incremental solver a represents a scheme that performs incremental updates only in the reordering stage while solving from scratch in the remaining stages, and incremental solver b represents a scheme that performs incremental updates only in the numerical decomposition stage while reordering and symbolic decomposition are solved from scratch. Based on 12 test cases, this invention achieves an average speedup of approximately 4.02 times compared to the CHOLMOD full direct solver. Figure 5 As shown, all schemes started from 1.0 in round 0. With increasing rounds, this invention maintained the highest cumulative speedup ratio in all 12 cases. At the end of round 10, this invention achieved an average additional speedup advantage of approximately 1.97 times and 3.08 times compared to incremental solver a and incremental solver b, respectively. These results demonstrate that applying path-aware structure pruning simultaneously to partial reordering, partial symbolic decomposition, and partial numerical decomposition can significantly reduce redundant computations in multi-round ECO iterations.
[0040] 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. An ECO-driven sparse solution method for power networks based on structure pruning, characterized in that, Includes the following steps: Step 1: Perform ab initio solution for the power network admittance matrix of the baseline round to obtain the partition tree, elimination tree and decomposition factor, and cache them; Step 2: For the current round power network admittance matrix, detect matrix changes based on the mapping relationship from the current round node to the previous round node. The matrix changes include one or more of the following: adding nodes, deleting nodes, adding edges, deleting edges, and value changes. Step 3: Based on the old partition tree, determine the reordered dirty regions according to the matrix changes; Step 4: Perform local nested partitioning and reconstruction only on the subtrees corresponding to the reordered dirty regions, and merge them with the reordering results of the unaffected regions to obtain the current round's partition tree and reordering results; Step 5: Based on the new partition tree, identify the dirty regions for symbolic decomposition; Step 6: Perform local symbolic decomposition based on the nodes corresponding to the dirty regions of the symbolic decomposition, reuse the old elimination tree parent-child relationship and the sparse mode of the decomposition factor in the unaffected regions, and obtain the elimination tree and sparse structure of the decomposition factor in the current round. Step 7: Based on the new elimination tree, identify the dirty nodes in the numerical decomposition; Step 8: Perform local numerical updates on the columns corresponding to the affected dirty nodes, only recalculate the decomposition factor values of the corresponding dirty columns, and directly reuse the old decomposition factor values of the unaffected columns. Step 9: Use the updated decomposition factor to perform trigonometric solutions on the current round of linear equations to obtain the exact nodal voltage solutions of the current round of power network.
2. The ECO-driven sparse solution method for power networks based on structure pruning according to claim 1, characterized in that, The generation rules for reordered dirty regions in step 3 include: when a structural change falls into a leaf region, the corresponding leaf region is marked as a reordered dirty region; when a new node appears, the region to which the new node belongs is determined according to the lowest common ancestor of the regions to which its neighboring nodes belong; if the region to which it belongs is a partition region, the partition region is marked as a reordered dirty region; when a new edge connects two regions that do not have an ancestor-descendant relationship and destroys the partition tree structure, the region corresponding to the lowest common ancestor of the two regions is marked as a reordered dirty region.
3. The ECO-driven sparse solution method for power networks based on structure pruning according to claim 1, characterized in that, The specific content of step 4 includes: first, performing ancestor filtering on the reordered dirty regions, retaining only the highest-level dirty ancestor regions; then, collecting all descendant nodes corresponding to the retained dirty ancestor regions as a set of nodes to be reconstructed; extracting the induced subgraph corresponding to the set of nodes to be reconstructed from the current round matrix, and performing nested subdivision reconstruction only on the induced subgraph.
4. The ECO-driven sparse solution method for power networks based on structure pruning according to claim 1, characterized in that, The specific content of step 5 includes: marking all regions that have been reordered in step 4 as symbolic decomposition dirty regions; for structural changes that have not triggered reordering but still change local adjacency relationships, marking their associated regions as symbolic decomposition dirty regions, and extending the symbolic decomposition dirty regions along the old partition tree to the root region to cover the ancestor regions affected by column merging properties.
5. The ECO-driven sparse solution method for power networks based on structure pruning according to claim 1, characterized in that, In step 6, for nodes located in unaffected regions, the parent-child relationship of the old elimination tree and the sparse mode of the decomposition factor are reused directly through the index mapping between the old round arrangement and the current round arrangement; for nodes located in the dirty region of symbol decomposition, their elimination tree parent node and the sparse mode of the decomposition factor are recalculated.
6. The ECO-driven sparse solution method for power networks based on structure pruning according to claim 1, characterized in that, In step 7, the initial set of dirty nodes for numerical decomposition consists of the following parts: all nodes in the reordered dirty region, endpoint nodes directly affected by structural changes except for the aforementioned reordered failed nodes, and endpoint nodes of matrix entries directly affected by numerical changes; and the initial set is extended along the new elimination tree to the root node to cover the ancestor nodes affected by the column merging property, thereby determining the final set of dirty nodes for numerical decomposition.
7. The ECO-driven sparse solution method for power networks based on structural pruning according to claim 1, characterized in that, In step 8, for columns that do not fall into the final set of dirty nodes in the numerical decomposition, the corresponding values are directly copied through the index mapping between the old round decomposition factor columns and the current round columns; for columns that fall into the final set of dirty nodes in the numerical decomposition, the column values are recalculated using the left-look method.
8. The ECO-driven sparse solution method for power networks based on structure pruning according to any one of claims 1-7, characterized in that, The proposed solution method is applicable to symmetric positive definite sparse linear systems in power network analysis, and maintains an accurate solution consistent with the ab initio solution during multiple rounds of ECO-driven iterations.
9. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, it runs an ECO-driven sparse solution method for power networks based on structure pruning as described in any one of claims 1-7.
10. A computer-readable storage medium having stored thereon computer instructions, which, when executed by a processor, perform an ECO-driven, structure-pruning-based sparse power network solution method according to any one of claims 1-7.