Planning analysis workflow intelligent arrangement method and device based on natural language instruction

Abstract

The invention provides a planning analysis workflow intelligent arrangement method and device based on a natural language instruction, and the method and device can deeply understand the complex analysis intention of a plan in a natural language through a mixed analysis engine comprising a large-scale language model and a named entity recognition model. And according to the understood intention, a proper operator and a proper model are autonomously selected from a tool library in combination with a planning analysis knowledge graph, so that a complete, logic self-consistent and executable multi-step analysis workflow can be autonomously and dynamically arranged, the workflow is automatically executed, and multi-modal data is collected. According to the method and the device, the technical problems in multiple aspects from keyword matching to composite intention deep understanding, from artificial process design to knowledge-driven intelligent arrangement and from one-way instruction execution to interactive exploration analysis in the prior art are solved, and the method and the device have the advantages that the technical problems in multiple aspects from keyword matching to composite intention deep understanding, from artificial process design to knowledge-driven intelligent arrangement and from one-way instruction execution to interactive exploration analysis are solved; the planning work efficiency can be improved, the technical threshold is reduced, and industry digital transformation is promoted.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of land spatial planning and artificial intelligence, specifically to a method and device for intelligent orchestration of planning analysis workflow based on natural language commands. Background Technology

[0002] Planning, monitoring, evaluation, and early warning processes have placed unprecedented demands on the depth, breadth, and timeliness of data analysis. Spatial analysis technologies, centered on Geographic Information Systems (GIS), Remote Sensing (RS), and various specialized planning models, have become an indispensable toolkit for planners. However, using these powerful tools requires planners to master complex interfaces, cumbersome data processing workflows, and a steep learning curve.

[0003] In their daily work, planners often need to translate macro-level analytical "intent" (e.g., analyzing the impact of a new subway line opening on property prices along the line) into a series of specific and detailed software operation steps. This transformation process highly depends on the planner's professional experience, proficiency with software tools, and data processing capabilities. It not only consumes a lot of time and energy, but also, due to differences in the operator's experience, the analysis path and results may vary, affecting the standardization and scientific nature of the work.

[0004] Currently, there are three main types of analytical methods for land and space planning analysis:

[0005] The first category is analysis methods based on traditional GIS software operation and script programming, such as the method disclosed in CN112905911A. In this type of method, planners manually operate GIS software (such as ArcGIS, QGIS) through a graphical user interface (GUI), sequentially calling tools such as buffer analysis, overlay analysis, and network analysis; or, planners with programming skills will use languages ​​such as Python to write scripts (such as ArcPy, PyQGIS) to solidify a series of analysis steps into code, thereby achieving semi-automation of the process. Planners need to pre-design the complete analysis logic and steps, and then call the underlying tools step by step by clicking the mouse or writing code, manually managing intermediate data, and finally obtaining the analysis results.

[0006] This type of approach forms the basis of current land spatial planning analysis, but its core flaw lies in the fact that it is human-driven rather than intent-driven. This method requires users to be proficient operators of GIS tools, the entire process is time-consuming and labor-intensive, has a high technical threshold, and is difficult to flexibly adapt to changes in analytical needs.

[0007] The second type utilizes analysis methods based on keyword matching and simple natural language query interfaces. Some GIS platforms or data query systems attempt to provide simple natural language query functions. The system typically has a pre-set keyword database, triggering predefined simple queries or operations by matching specific words in the user's input (such as "search," "nearby," and "buffer"). The system backend maintains a "keyword-function" mapping table. When a user enters "find all hospitals within 1 kilometer," the system matches keywords such as "search," "nearby," "kilometer," and "hospital," and then calls a pre-written fixed template program for point-buffer-query.

[0008] These solutions offer some convenience in specific and simple query scenarios, but they are essentially keyword matching rather than intent understanding. Therefore, they cannot handle complex, multi-step planning and analysis tasks that involve complex logic (such as evaluating dynamic impacts). They lack the ability to understand the context and make logical reasoning, and their application scenarios are extremely limited.

[0009] The third category is analysis methods based on visual workflow modeling tools. Some GIS software or data science platforms provide visual workflow modeling tools (such as ArcGIS ModelBuilder / FME Workbench). Users can construct visual analysis flowcharts by dragging and dropping different tool modules and connecting them with lines. Users can drag icons representing different functions (such as "input data", "buffer", "intersection", and "output results") onto the canvas and connect them with arrows to define the flow of data and the processing order, forming a repeatable model.

[0010] These methods make the analysis process explicit, improving its reusability and transparency, and lowering the barrier to entry compared to purely manual operations and script programming. However, the core issue is that the orchestration process for building the workflow still requires manual intervention. Users still need to clearly know in advance which tools to use and the logical order between them. While these methods automate execution, they do not address the intelligentization of the crucial step from intent to process design.

[0011] As mentioned above, existing analytical methods cannot truly understand the complex planning analysis logic, spatial relationships, and causal inferences inherent in users' natural language, mostly remaining at a superficial keyword matching level. Whether through manual operation, scripting, or visual modeling, the "orchestration and design" process from an analytical idea to a concrete, executable workflow requires deep human intervention, resulting in low levels of intelligence. Planners still need to spend a significant amount of time learning and operating complex GIS software and modeling tools, unable to fully focus their energy on the thinking and decision-making of the planning business itself. Existing technologies cannot support truly natural, flexible, and multi-turn conversational interaction, making it difficult to adapt to the exploratory and iterative nature of planning analysis. In other words, existing technologies lack an intelligent mechanism capable of deeply understanding the complex analytical intentions in planners' natural language and autonomously and dynamically orchestrating a complete, logically consistent, and executable multi-step analytical workflow.

[0012] In recent years, generative artificial intelligence technologies, represented by Large Language Models (LLMs), have made groundbreaking progress. Their powerful natural language understanding and logical reasoning capabilities have made it possible to transform traditional human-computer interaction models and achieve "what you say is what you get" intelligent analysis. Introducing natural interaction capabilities into the field of land and space planning, enabling planners to directly drive complex spatial analyses through everyday language, is a key direction for improving planning efficiency, lowering technical barriers, and promoting the industry's digital transformation. Therefore, developing a technology that can accurately understand planners' natural language intentions and automatically orchestrate them into machine-executable workflows has significant practical implications and application value. Summary of the Invention

[0013] This invention addresses the aforementioned problems by providing a method and apparatus that can completely bridge the significant gap between the high-level analytical intentions of planners and the complex operational tools used at lower levels in the field of land spatial planning. The invention employs the following technical solution:

[0014] This invention provides an intelligent orchestration method for planning analysis workflows based on natural language commands. The method includes the following steps: Step S1, parsing the natural language commands of the target analysis task input by the planner using a hybrid parsing engine, and constructing an analysis task template based on the parsing results, which contains structured analysis intent information; Step S2, querying the planning analysis knowledge graph based on the analysis intent information, and matching the analysis intent information with tool components based on the query results, wherein the planning analysis knowledge graph contains various data types, analysis concepts, tool components, and their relationships related to land spatial planning analysis; Step S3, generating a workflow for the target analysis task based on the matching results of Step S2 and the dependencies between the tool components; Step S4, automatically executing the target analysis task according to the workflow, and generating an intelligent analysis report based on the intermediate and final results during execution. The hybrid parsing engine includes a large language model and a named entity recognition model. In step S1, the large language model performs preliminary parsing on the natural language instruction to obtain preliminary parsing results. Based on the preliminary parsing results, the named entity recognition model scans the natural language instruction to extract structured elements and constructs the analysis task template based on the structured elements.

[0015] The intelligent orchestration method for planning and analysis workflow based on natural language instructions provided by this invention also has the following technical feature: the named entity recognition model is fine-tuned on the territorial spatial planning corpus. Step S1 includes the following sub-steps: Step S1-1, performing preliminary parsing of the natural language instructions using the large-scale language model to identify the task objective, target object, and key entities of the target analysis task as the preliminary parsing result; Step S1-2, scanning the natural language instructions based on the preliminary parsing result using the named entity recognition model to extract the structured elements; Step S1-3, standardizing the structured elements and combining the standardized structured elements into the analysis task template.

[0016] The intelligent orchestration method for planning and analysis workflow based on natural language instructions provided by this invention may also have the following technical features: in steps S1-3, each structured element is filled into a standardized JSON object to obtain the analysis task template in JSON format. The analysis task template contains multiple analysis intent information entries, each of which includes an analysis type and analysis content. The analysis type includes task type, spatial scope, analysis topic, suitability criteria, and limiting factors. The analysis content includes named entities and specific numerical values.

[0017] The intelligent orchestration method for planning and analysis workflow based on natural language instructions provided by this invention may also have the following technical features, wherein step S2 includes the following sub-steps: Step S2-1, constructing corresponding query statements based on the named entities in the analysis content; Step S2-2, querying the planning and analysis knowledge graph using the query statements to obtain one or more data types and one or more corresponding tool components required for analyzing the named entities; Step S2-3, planning the optimal component path based on the data types and tool components obtained from the query, which includes obtaining the input data of the required data types and processing the corresponding input data using each of the tool components; Step S2-4, if the analysis content contains the specific numerical value, binding the specific numerical value to the corresponding parameter of the corresponding tool component.

[0018] The intelligent orchestration method for planning and analysis workflow based on natural language instructions provided by this invention may also have the following technical features, wherein step S3 includes the following sub-steps: Step S3-1, generating a directed acyclic graph (DAG) based on the dependency relationships between the tool components in the optimal component path, which represents the workflow of the target analysis task; Step S3-2, automatically verifying the generated DAG, including integrity verification and precondition verification; Step S3-3, when it is determined in step S3-2 that the precondition verification has failed, generating a corresponding interactive interface based on the verification result to provide the user with a means to complete the missing input data, wherein the completion means at least includes the missing input data input by the user.

[0019] The intelligent orchestration method for planning and analysis workflow based on natural language instructions provided by this invention may also have the following technical features: the completion means further include using relevant historical data for deduction and using relevant tools to generate relevant prediction data.

[0020] The intelligent orchestration method for planning and analysis workflows based on natural language instructions provided by this invention may also have the following technical features, wherein step S4 includes the following sub-steps: Step S4-1, for the validated workflow, the workflow engine debugs the underlying geographic information services, databases, and computing resources according to the information in the workflow to achieve automated execution of the entire analysis process; Step S4-2, after the workflow is executed, all intermediate and final results generated during the workflow execution are collected to form a multimodal result set; Step S4-3, an intelligent analysis report is generated based on the multimodal result set using a large language model.

[0021] This invention provides an intelligent orchestration device for planning and analysis workflows based on natural language commands. The device comprises: an intent-structured parsing module, used to parse the natural language commands of the target analysis task input by the planner using a hybrid parsing engine, and construct an analysis task template based on the parsing results, which includes structured analysis intent information; a component matching and numerical binding module, used to query a planning analysis knowledge graph based on the analysis intent information, and match the analysis intent information with tool components based on the query results, wherein the planning analysis knowledge graph includes various data types, analysis concepts, tool components, and their relationships related to land and space planning analysis; a workflow generation and interaction verification module, used to generate a workflow for the target analysis task based on the matching results and the dependencies between the tool components; and an orchestration execution and report synthesis module, used to automatically execute the target analysis task according to the workflow, and generate an intelligent analysis report based on intermediate and final results during execution. The hybrid parsing engine includes a large language model and a named entity recognition model. The intent structured parsing module performs preliminary parsing of the natural language instruction using the large language model to obtain preliminary parsing results. Based on the preliminary parsing results, the named entity recognition model scans the natural language instruction to extract structured elements and constructs the analysis task template based on the structured elements.

[0022] The role and effect of invention

[0023] The intelligent orchestration method and apparatus for planning and analysis workflow based on natural language commands provided by the present invention have the following advantages compared with existing analysis methods:

[0024] First, it solves the technical challenge of moving from keyword matching to deep understanding of complex intents. Existing analysis methods cannot parse the multi-dimensional analytical logic contained in natural language instructions, such as time series, causal relationships, spatial coupling, and multi-scenario comparisons. In contrast, the method and apparatus of this invention employ a hybrid parsing engine that includes a large language model and a named entity recognition model. This allows the analysis system to fully understand the internal structure and objectives of complex analysis tasks, much like a seasoned planner, while avoiding the illusions created by large language models that affect the accuracy of the analysis.

[0025] Secondly, it solves the technical challenge of transitioning from manual process design to knowledge-driven intelligent orchestration. As mentioned above, regardless of the form, existing analytical methods require manually pre-defined analytical processes. In contrast, the method and apparatus of this invention, based on hybrid model parsing, utilize a planning analysis knowledge graph to match analytical intent, required data, and tool components. Therefore, it can autonomously select appropriate operators and models from the tool library according to the understood intent and weave them into a complete analytical workflow in the optimal logical order.

[0026] Third, it solves the technical challenge of moving from one-way command execution to interactive exploratory analysis. Traditional spatial analysis processes are unidirectional and rigid; if a command is ambiguous or lacks necessary conditions, the process will be interrupted or erroneous. In contrast, this invention establishes a more resilient and intelligent interactive mechanism, enabling the analysis system to proactively ask questions and provide options when encountering insufficient information. This supports planners in conducting exploratory and iterative analysis, elevating the analysis system from a tool to an intelligent analysis partner. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the intelligent orchestration method for planning and analysis workflows based on natural language commands in an embodiment of the present invention.

[0028] Figure 2 This is a flowchart of the intelligent orchestration method for planning and analysis workflow based on natural language instructions in an embodiment of the present invention.

[0029] Figure 3 This is a schematic diagram of the intelligent orchestration device for planning and analysis workflow based on natural language instructions in an embodiment of the present invention. Detailed Implementation

[0030] To make the technical means, creative features, objectives and effects of the present invention easy to understand, the following describes in detail the intelligent orchestration method and device for planning and analysis workflow based on natural language instructions.

[0031] Figure 1 This is a flowchart illustrating the intelligent orchestration method for planning and analysis workflows based on natural language commands in this embodiment. Figure 2 This is a flowchart of the intelligent orchestration method for planning and analysis workflow based on natural language instructions in this embodiment.

[0032] like Figure 1 and Figure 2 As shown, the intelligent orchestration method for planning and analysis workflows based on natural language instructions includes the following steps:

[0033] Step S1: Intent structure parsing based on the hybrid model.

[0034] Step S2: Component matching and numerical binding based on the planning analysis knowledge graph.

[0035] Step S3: Dynamic workflow generation and interactive verification.

[0036] Step S4: Arrange execution and synthesize multimodal analysis reports.

[0037] The steps described above will be explained in detail below.

[0038] Step S1: Intent structure parsing based on the hybrid model.

[0039] This step is fundamental to accurately understanding user instructions. In this step, a hybrid parsing engine parses the user's natural language instructions and constructs an analysis task template containing information about the analysis intent based on the parsing results. The hybrid parsing engine includes a general-purpose generative AI Large Language Model (LLM) and a Named Entity Recognition (NER) model. The NER model has been specifically trained (fine-tuned) on a corpus of land spatial planning data.

[0040] Step S1 specifically includes the following sub-steps:

[0041] Step S1-1: The received natural language instructions are initially parsed using LLM to identify the core objectives, target objects, and key entities of the target analysis task, which serve as the initial parsing results.

[0042] For example, if a user's natural language instruction is received as "Comparative analysis of the spatiotemporal evolution characteristics of the agglomeration of high-tech enterprises within a 5-kilometer radius of Shanghai's two major hubs, Hongqiao and Pudong, from 2015 to 2025," LLM will perform a preliminary analysis to identify the core objective (comparative analysis), the target object (enterprise agglomeration), and key entities. The preliminary analysis result text will then serve as the input to the NER model.

[0043] Step S1-2 involves using the NER model to perform a refined scan of the natural language instructions based on the preliminary analysis results, extracting structured elements related to the target analysis task.

[0044] For example, based on the preliminary analysis results above, the NER model performs a refined scan of natural language instructions, extracting the following structured elements:

[0045] Spatial entities: Shanghai Hongqiao Hub, Pudong Hub

[0046] Spatial operator: 5-kilometer radius (analyzed as a buffer zone)

[0047] Time entity: 2015 to 2025 (parsed into a time series)

[0048] Analysis Indicator: High-tech Enterprise Cluster

[0049] That is, each extracted structured element contains a key name and a key value.

[0050] Steps S1-3 involve standardizing the extracted structured elements and combining the standardized structured elements into an analysis task template.

[0051] Specifically, the extracted structured elements are automatically populated into a standardized JSON object, forming a complete, unambiguous, machine-readable description of the target analysis task, which serves as the analysis task template. The automatic population process is completed by a program script reading the output of the structured element scan and matching it according to the entity type.

[0052] The analysis task template contains multiple structured analysis intent messages parsed from the natural language instructions of the target analysis task. Each structured analysis intent message includes an analysis type and analysis content. The analysis type (key name) can include task type, spatial scope, analysis topic, suitability criteria, and limiting factors. The analysis content (key value) must at least contain named entities and may further include corresponding specific numerical values.

[0053] Step S2: Component matching and numerical binding based on the planning analysis knowledge graph.

[0054] In this step, the pre-built planning analysis knowledge graph is queried based on the analysis intent information in the analysis task template, and the analysis intent information is matched with the tool components based on the query results. If the analysis intent information contains specific values, the specific values ​​are bound to the corresponding parameters of the tool components based on the query results.

[0055] The planning analysis knowledge graph contains multiple nodes and edges. The nodes include various data types, analysis concepts, and tool components typically involved in land and space planning analysis. The edges represent the relationships between these nodes, that is, the dependencies between different data types (for example, the calculation of one data type requires data of another data type as input).

[0056] Specifically, data types include POI data and road network data; analytical concepts include reachability and clustering; and tool components include Buffer_Analysis_Tool and Kernel_Density_Tool. The edge between the clustering node and its corresponding tool component node contains the information: is_calculated_by Kernel_Density_Tool, meaning that the clustering is calculated using the Kernel_Density_Tool tool component. The edge between the Kernel_Density_Tool tool component node and the POI data data type node contains the information: the input of Kernel_Density_Tool, requires_input, is POI data.

[0057] Step S2 specifically includes the following sub-steps:

[0058] Step S2-1: Construct corresponding query statements based on the named entities in the analysis intent information in the analysis task template.

[0059] Step S2-2: Use a query statement to query the planning analysis knowledge graph to obtain the nodes and edges related to the named entity, that is, to determine the analysis concept that the named entity conforms to, and based on the analysis concept, obtain one or more data types and one or more tool components required for analyzing the named entity.

[0060] For example, based on the named entity "High-tech Enterprise Clustering Degree" in the analysis task template, a corresponding query statement is constructed. This query statement is then used to query the planning analysis knowledge graph to obtain the relevant nodes "POI Data" and "Kernel Density Analysis Tool". In other words, the analysis of this named entity requires POI data as input, and the kernel density analysis tool is used to analyze the POI data.

[0061] Step S2-3: Based on the data types and tool components obtained from the query, an optimal component path is automatically planned. This optimal component path includes obtaining the input data of the required data type and processing the corresponding input data using each tool component.

[0062] For example, based on the above query results, the system automatically plans an optimal component path from "obtaining POI data of high-tech enterprises" to "performing kernel density analysis using kernel density analysis tools".

[0063] Step S2-4: If the content of the analysis intent information contains a specific value corresponding to the named entity, bind the specific value to the corresponding parameter of the corresponding tool component.

[0064] Step S3: Dynamic workflow generation and interactive verification.

[0065] Step S3 specifically includes the following sub-steps:

[0066] Step S3-1: Based on the dependencies between the tool components in the optimal component path (defined by the edges between the tool components in the planning analysis knowledge graph above), a directed acyclic graph (DAG) is automatically generated, representing the workflow of the target analysis task. The generated DAG may contain parallel logic and / or serial logic.

[0067] In this context, DAG is used to describe the sequential order of data flow and computation in a workflow. It contains multiple nodes and directed edges, where a node represents a processing step in the workflow, containing the input data required for that processing and information about the corresponding tool components; and directed edges represent the sequential order between processing steps.

[0068] Step S3-2: Automatically verify the generated DAG. If the verification fails, proceed to step S3-3; if the verification passes, proceed to step S4.

[0069] In this step, automatic verification includes integrity verification and prerequisite verification. For integrity verification, it checks whether the data types and tool components appearing in the optimal component path are all included in the generated DAG; for prerequisite verification, it checks whether the necessary information, such as the input data required by each tool component in the optimal component path, has been located.

[0070] Step S3-3: When it is determined in step S3-2 that the precondition verification has failed, that is, when the required input data is missing, a corresponding interactive interface is generated based on the verification result to provide the user with a means to complete the missing input data.

[0071] The completion methods include at least the missing data input by the user. Optionally, the completion methods may further include extrapolation using relevant historical data and extrapolation using appropriate tools to generate missing data (e.g., through prediction by machine learning models).

[0072] For example, if the automatic verification reveals that the planning analysis knowledge graph lacks enterprise data for 2025, the system will not immediately exit with an error message. Instead, it will activate its interactive module and provide feedback to the user through the interface: "The 2025 high-tech enterprise data required for the analysis is temporarily unavailable. You can (A) use 2024 data for prediction and deduction, or (B) the user needs to provide the data file. Please select." The user can make a choice through the interactive interface and upload the data file when selecting option B. This dynamic interactive capability in the analysis process greatly improves the robustness of the method and the user experience.

[0073] Step S4: Arrange execution and synthesize multimodal analysis reports.

[0074] Step S4 includes the following sub-steps:

[0075] Step S4-1: For workflows that pass verification, the background workflow engine schedules underlying geographic information (GIS) services, databases, and computing resources based on the information in the workflow to automate the entire analysis process.

[0076] Step S4-2: After the workflow is completed, collect all intermediate and final results generated during the workflow execution to form a multimodal result set. These intermediate and final results may include, for example, kernel density analysis plots for each year, clustering change curves, and key statistical data.

[0077] Step S4-3: Generate an intelligent analysis report based on the multimodal result set using a generative AI large-scale language model. That is, the generative AI large-scale language model is invoked again, and prompted with keywords containing the aforementioned multimodal result set. The model then automatically writes an intelligent analysis report that is logically clear and richly illustrated, summarizing the analysis process and results using natural language.

[0078] For example, the intelligent analysis report generated by the large language model contains the following: "The results show that since 2015, the annual growth rate of high-tech enterprise agglomeration around Hongqiao Hub has reached 12%, which is significantly higher than the 7% of Pudong Hub. Especially after 2020, when the traffic lines were densified, a clear industrial agglomeration hotspot was formed."

[0079] Figure 3 This is a schematic diagram of the intelligent orchestration device for planning and analysis workflow based on natural language commands in this embodiment.

[0080] like Figure 3 As shown, this embodiment also provides an intelligent orchestration device 10 for planning and analysis workflow based on natural language commands, corresponding to the above method. It includes an intent structure parsing module 11, a component matching and numerical binding module 12, a workflow generation and interaction verification module 13, an orchestration execution and report synthesis module 14, and a control module 15. This device can be part of a land spatial planning and analysis system or communicate with the analysis system.

[0081] The intent structured parsing module 11 includes a hybrid parsing engine 111 and an analysis task template generation module 112, which uses the method described in step S1 to parse natural language instructions and construct analysis task templates.

[0082] The component matching and value binding module 12 includes a query statement construction submodule 121, a component matching submodule 122, a path planning submodule 123, and a value binding submodule 124. It uses the method described in step S2 above to complete the matching and value binding of various data types and tool components based on the analysis task template and the planning analysis knowledge graph, and generates the optimal component path.

[0083] The workflow generation and interaction verification module 13 includes a directed graph generation submodule 131, an automatic verification submodule 132, and a user interaction submodule 133. It uses the method described in step S3 above to generate a DAG of the workflow based on the optimal component path and performs automatic verification on it. When the verification fails, it interacts with the user to try to complete the required data.

[0084] The orchestration execution and report synthesis module 14 includes a workflow engine 141, a multimodal result collection submodule 142, and a report synthesis submodule 143 based on generative AI. It uses the method described in step S4 above to automate the execution of the analysis process according to the workflow, collects multimodal result sets and provides them to a large language model, which then generates an intelligent analysis report.

[0085] To better understand the above methods and apparatus, a more detailed example will be given below.

[0086] For example, the planner input the following typical natural language command (analysis command) containing complex logic through the system's natural language interactive interface: "In Chaoyang District, Beijing, taking into account the residential population density, the service radius within a 15-minute walk, and the service gap area, conduct a site selection potential analysis for the construction of a new primary school. The analysis results must exclude all industrial development areas."

[0087] Step S1: Intent structure parsing based on the hybrid model.

[0088] In step S1, after receiving the natural language instruction, the system activates the hybrid parsing engine. The LLM first identifies this as a typical "Site Selection" task under "Multi-Criterion Evaluation (MCE)," obtaining the core objective, target object, and key entities of the task. Subsequently, the specially trained NER model describes the natural language instruction based on the parsing results, accurately extracting the following structured elements:

[0089] Spatial Scope: Chaoyang District, Beijing

[0090] Analysis Topic: Newly Built Primary School

[0091] Suitability criteria: Residential population density (positive factor), service radius within a 15-minute walk (used to define the service area), and service gaps (areas to be considered).

[0092] Restrictive factors: Industrial Development Zones (must be excluded)

[0093] After the structured element extraction is completed, the system automatically fills the above structured elements into a standardized JSON object to form a complete and unambiguous machine-readable description of the analysis task as an analysis task template. The automatic filling can be implemented through corresponding scripts, or can be generated by a large language model through corresponding prompts. An example of the analysis task template is as follows:

[0094] {

[0095] "task_type": "SiteSelectionPotentialAnalysis",

[0096] "spatial_scope": {"city": "Beijing", "district": "Chaoyang District"},

[0097] "analysis_theme": "PrimarySchool",

[0098] "favorable_criteria":

[0099] {"name": "PopulationDensity", "weight": "auto"},

[0100] {"name": "ServiceGapArea", "derived_from": {"source": "ExistingSchools", "method": "WalkingIsochrone", "threshold": 15, "unit": "minutes"}}

[0101] ,

[0102] "exclusion_criteria":

[0103] {"name": "IndustrialZone"}

[0105] }

[0106] Step S2, component matching and numerical binding based on the planning analysis knowledge graph.

[0107] ​In step S2, the system uses the JSON-formatted analysis task template generated in the previous step as a query and inputs it into the planning analysis knowledge graph to obtain the data types and corresponding tool components required for the analysis.

[0108] For example, when querying PopulationDensity, the Planning Analysis Knowledge Graph returns that this indicator requires raster data of the spatial distribution of the population, which can be calculated using the Kernel Density Tool or the Zonal Statistics Tool.

[0109] Querying ServiceGapArea and the planning analysis knowledge graph returns a more complex path. First, it requires existing primary school POI data and urban road network data. Then, it uses the service area analysis tool (Service_Area_Tool) to generate existing service areas based on this data. Finally, it uses the erase tool (Erase_Tool) to subtract the existing service areas from the Chaoyang District area to obtain the blank area.

[0110] Querying IndustrialZone and the planning analysis knowledge graph returns the required vector data for the industrial development zone. The vector data is then clipped using the Clip_Tool to limit the clipped vector data to the Chaoyang District.

[0111] Querying SiteSelectionPotentialAnalysis and planning the analysis knowledge graph returns the results of the composite task identification. Finally, the weighted overlay tool needs to be called. The input of the overlay tool is multiple standardized suitability factor raster layers.

[0112] After the matching is completed, the system binds the specific values ​​(such as 15 minutes, Chaoyang District vector boundary, etc.) in the analysis task template generated in step one to the corresponding parameters of each matched tool component.

[0113] Step S3: Dynamic workflow generation and interactive verification.

[0114] In step S3, the system automatically generates a directed acyclic graph (DAG) containing parallel and serial logic, representing the corresponding workflow, based on the dependencies between tool components (e.g., a service area must be generated before a blank area can be calculated).

[0115] After generation, the system starts the automatic verification submodule. After automatic verification, a potential problem was found: the metadata of the population spatial distribution raster data in the knowledge base shows that its update time is 2020, while the current system time is 2025, which poses a risk of data timeliness.

[0116] At this point, the interactive clarification mechanism is activated. The system will not interrupt the process, but will instead present the planner with the following questions through the interactive interface:

[0117] "The system detected that the population data used is from 2020, which may affect the accuracy of site selection analysis. How can this be handled?"

[0118] [Option A]: Continue using 2020 data for analysis.

[0119] [Option B]: Use the built-in population distribution prediction model to project the population distribution up to 2025.

[0120] [Option C]: Pause the process and manually upload the updated population data.

[0121] If the planner selects [Option B], the system will automatically insert a population prediction model component into the DAG and place it before the kernel density analysis component, thereby completing the dynamic improvement of the workflow.

[0122] Step S4: Arrange execution and synthesize multimodal analysis reports.

[0123] After user confirmation and verification, the DAG is submitted to the workflow engine. The workflow engine then automatically executes a series of spatial calculation and analysis tools in the background, calculating various factor layers in parallel: population density raster, service gap raster, and geological hazard exclusion zone raster; each factor layer is standardized (weighted from 1 to 9 points). Finally, a weighted overlay tool is invoked to overlay the suitability layers and mask them with a constraint layer, ultimately generating a raster data set for a "Comprehensive Evaluation Map of Primary School Site Selection Potential." After analysis, the system collects all key outputs: the final potential distribution map, intermediate factor maps, and area statistics tables for high-potential areas. Finally, the multimodal results are input into the generative AI report synthesis submodule, automatically generating an intelligent analysis report.

[0124] Generative AI can automatically write and generate a richly illustrated intelligent analysis report, which includes the following professional description: "Abstract: This analysis of the site selection potential for a new primary school in Chaoyang District, Beijing, comprehensively assessed the two core suitability criteria of population density and service gaps, and rigidly excluded industrial development areas. Analysis results: such as..." Figure 1 As shown, Changying and Dougezhuang in the eastern part of Chaoyang District, and Laiguangying in the northern part, exhibit significant high site selection potential. These areas generally exhibit the dual characteristics of 'high population density' and 'lack of 15-minute walking service coverage'. The total area of ​​high-potential plots is approximately 8.5 square kilometers (see table below), and these are recommended as key areas for the next stage of on-site investigation and specific plot comparison…

[0125] The role and effect of the embodiments

[0126] The intelligent orchestration method and apparatus for planning and analysis workflow based on natural language commands provided in this embodiment have the following advantages compared with existing analysis methods:

[0127] First, it solves the technical challenge of moving from keyword matching to deep understanding of complex intents. Existing analysis methods cannot parse the multi-dimensional analytical logic contained in natural language instructions, such as time series, causal relationships, spatial coupling, and multi-scenario comparisons. In contrast, the method and apparatus of this embodiment employ a hybrid parsing engine that includes a large language model and a named entity recognition model. This allows the analysis system to fully understand the internal structure and objectives of complex analysis tasks, much like a seasoned planner, while avoiding the problem of large language models creating illusions and affecting the accuracy of the analysis.

[0128] Secondly, it solves the technical challenge of transitioning from manual process design to knowledge-driven intelligent orchestration. As mentioned above, regardless of the form, the analysis process design in existing technologies requires manual pre-definition. In contrast, the method and apparatus of this embodiment, based on hybrid model parsing, utilizes a planning analysis knowledge graph to match the analysis intent, required data, and tool components. Therefore, it can autonomously select appropriate operators and models from the tool library according to the understood intent and weave them into a complete analysis workflow in the optimal logical order.

[0129] Third, it solves the technical challenge of moving from one-way command execution to interactive exploratory analysis. Traditional spatial analysis processes are unidirectional and rigid; if a command is ambiguous or lacks necessary conditions, the process will be interrupted or erroneous. In contrast, this embodiment establishes a more resilient and intelligent interactive mechanism, enabling the analysis system to proactively ask questions and provide options when encountering insufficient information. This supports planners in conducting exploratory and iterative analysis, elevating the analysis system from a tool to an intelligent analysis partner.

[0130] The above embodiments are merely illustrative of specific implementations of the present invention, and the present invention is not limited to the scope of the description of the above embodiments. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are only for illustrating the principles of the present invention. Various changes and modifications can be made to the present invention without departing from the spirit and scope thereof, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent orchestration of planning and analysis workflows based on natural language commands, characterized in that, Includes the following steps: Step S1: Use the hybrid parsing engine to parse the natural language instructions of the target analysis task input by the planner, and construct an analysis task template based on the parsing results, which contains structured analysis intent information; Step S2: Query the planning analysis knowledge graph based on the analysis intent information, and match the analysis intent information with the tool components based on the query results. The planning analysis knowledge graph includes various data types, analysis concepts, tool components and their relationships related to land and space planning analysis. Step S3: Generate the workflow for the target analysis task based on the matching results of step S2 and the dependencies between the tool components; Step S4: Automatically execute the target analysis task according to the workflow, and generate an intelligent analysis report based on the intermediate and final results during the execution process. The hybrid parsing engine includes a large language model and a named entity recognition model. In step S1, the natural language instruction is initially parsed using the large language model to obtain a preliminary parsing result. Based on the preliminary parsing result, the named entity recognition model scans the natural language instruction to extract structured elements, and constructs the analysis task template based on the structured elements.

2. The intelligent orchestration method for planning and analysis workflow based on natural language instructions as described in claim 1, Its features are: The named entity recognition model has been fine-tuned on the corpus of land and space planning. Step S1 includes the following sub-steps: Step S1-1: The natural language instruction is initially parsed using the large language model to identify the task objective, target object, and key entities of the target analysis task as the initial parsing result; Step S1-2: The named entity recognition model scans the natural language instruction based on the preliminary parsing results to extract the structured elements. Steps S1-3: Standardize the structured elements and combine the standardized structured elements into the analysis task template.

3. The intelligent orchestration method for planning and analysis workflow based on natural language instructions according to claim 2, characterized in that: in, In steps S1-3, each structured element is populated into a standardized JSON object to obtain the analysis task template in JSON format. The analysis task template contains multiple analysis intent information entries. Each analysis intent information entry includes an analysis type and analysis content. The analysis type includes task type, spatial scope, analysis topic, suitability criteria, and limiting factors. The analysis content includes named entities and specific numerical values.

4. The intelligent orchestration method for planning and analysis workflow based on natural language instructions as described in claim 3, Its features are: Step S2 includes the following sub-steps: Step S2-1: Construct corresponding query statements based on the named entities in the analysis content; Step S2-2: Use the query statement to query the planning analysis knowledge graph to obtain one or more data types and one or more corresponding tool components required for analyzing the named entities; Step S2-3 involves planning the optimal component path based on the data type obtained from the query and the tool components, which includes obtaining the required input data of the data type and processing the corresponding input data using each of the tool components. Step S2-4: If the analysis content includes the specific value, bind the specific value to the corresponding parameter of the corresponding tool component.

5. The intelligent orchestration method for planning and analysis workflow based on natural language instructions according to claim 4, Its features are: Step S3 includes the following sub-steps: Step S3-1: Generate a directed acyclic graph based on the dependencies between the tool components in the optimal component path, which represents the workflow of the target analysis task; Step S3-2: Automatically verify the generated directed acyclic graph, including integrity verification and precondition verification. Step S3-3: When it is determined in step S3-2 that the precondition verification has failed, a corresponding interactive interface is generated based on the verification result to provide the user with a means to complete the missing input data. The means to complete the missing input data are at least input by the user.

6. The intelligent orchestration method for planning and analysis workflow based on natural language instructions according to claim 5, characterized in that: in, The completion methods also include using relevant historical data for extrapolation and using relevant tools to generate relevant predictive data.

7. The intelligent orchestration method for planning and analysis workflow based on natural language instructions according to claim 1, Its features are: Step S4 includes the following sub-steps: Step S4-1: For the workflow that has passed the verification, the workflow engine debugs the underlying geographic information service, database and computing resources according to the information in the workflow to realize the automated execution of the entire analysis process. Step S4-2: After the workflow is completed, collect all intermediate and final results generated during the workflow execution to form a multimodal result set; Step S4-3: Generate an intelligent analysis report based on the multimodal result set using a large language model.

8. A planning and analysis workflow intelligent orchestration device based on natural language commands, characterized in that, include: The intent structure parsing module is used to parse the natural language instructions of the planner's target analysis task using a hybrid parsing engine, and to build an analysis task template based on the parsing results, which contains structured analysis intent information. The component matching and numerical binding module is used to query the planning analysis knowledge graph based on the analysis intent information, and match the analysis intent information with tool components based on the query results. The planning analysis knowledge graph includes various data types, analysis concepts, tool components and their relationships related to land and space planning analysis. The workflow generation and interaction verification module is used to generate the workflow for the target analysis task based on the matching results and the dependencies between the tool components. The orchestration, execution, and report synthesis module is used to automatically execute the target analysis task according to the workflow, and generate an intelligent analysis report based on the intermediate and final results during the execution process. The hybrid parsing engine includes a large language model and a named entity recognition model. The intent structure parsing module performs preliminary parsing of the natural language instruction using the large language model to obtain preliminary parsing results. Based on the preliminary parsing results, the named entity recognition model scans the natural language instruction to extract structured elements and constructs the analysis task template based on the structured elements.

Images