A CIM-based urban information model data calculation method

Abstract

This application proposes a data calculation method for a city information model based on CIM, belonging to the field of data processing. The method includes: semantic parsing of the target test scenario description information to construct a test scenario feature vector; constructing an evolutionary parameter space for at least one evolutionary prediction model based on the CIM city information model; training the probability function mapping relationship between the test scenario feature vector and the evolutionary parameter space, and outputting an adaptive evolutionary parameter combination based on the probability function mapping relationship; importing the data into the evolutionary prediction model to perform data evolution calculation, obtaining a dynamic evolutionary time-series dataset, which is then transmitted as test cases to the test system terminal containing the target test scenario description information to obtain the test performance results of the tested target. By adaptively adjusting the evolutionary parameter combination according to the test scenario characteristics, the dynamic time-series data output by the evolutionary prediction model better matches the needs of various test scenarios, thereby improving the accuracy and computational efficiency of test cases.

Description

Technical Field

[0001] This invention relates to the field of data processing, and in particular to a method for calculating urban information model data based on CIM. Background Technology

[0002] City Information Modeling (CIM) is an organic integration of a three-dimensional urban spatial model and multi-dimensional urban information, including above-ground and underground, indoor and outdoor, historical, current, and future data, based on urban information data. With the deepening of smart city construction, CIM platforms are widely used in urban planning, infrastructure management, emergency response, and urban operation simulation, and are gradually becoming an important computing foundation for city-level system testing, verification, and predictive analysis.

[0003] In the CIM-based urban operation simulation and testing scenario, the evolution prediction model is the core component that drives the dynamic generation of urban state data. It generates a dynamic time-series dataset that meets the testing requirements by modeling and analyzing historical urban data, and uses this as a test case input to the system under test to evaluate the system's response performance under real urban operating conditions.

[0004] However, existing CIM-based data computation methods have significant shortcomings in practical applications. Current technologies typically employ fixed evolutionary parameter configurations for evolutionary prediction models, using uniform parameter settings to handle data evolution computation tasks across all test scenarios. Since different test scenarios have significantly different data requirements—for example, urban traffic simulation scenarios have high requirements for temporal continuity, while emergency response plan verification scenarios have stricter requirements for data accuracy within short periods—fixed parameter configurations struggle to accommodate the data characteristics of different scenarios. This results in poor adaptability of the generated data to the actual needs of the target test scenario, low accuracy of test cases, and consequently, affects the testing and verification effectiveness and reliability of the system under test. Summary of the Invention

[0005] This invention addresses the technical problem in existing technologies where CIM (City Information Model) uses fixed evolution parameters for unified calculations when facing different test scenarios, resulting in poor adaptability of evolution data to actual scenario requirements and low accuracy of test cases. It provides a CIM-based method for calculating city information model data to solve this problem.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0007] This invention provides a method for calculating urban information model data based on CIM, comprising: semantic parsing of the target test scenario description information to construct a test scenario feature vector; constructing an evolutionary parameter space for at least one evolutionary prediction model based on the CIM urban information model, wherein the evolutionary parameter space includes at least a time step parameter, an evolutionary time window, a data smoothing factor, a historical weight decay coefficient, a spatial coupling coefficient, and a data compression rate; training a probability function mapping relationship between the test scenario feature vector and the evolutionary parameter space with prediction error target, data redundancy target, and computational efficiency target as optimization objectives, and outputting an adaptive evolutionary parameter combination based on the probability function mapping relationship; importing the adaptive evolutionary parameter combination into the evolutionary prediction model to perform data evolution calculation to obtain a dynamic evolutionary time series dataset; and transmitting the dynamic evolutionary time series dataset as test cases to the test system terminal of the target test scenario description information to obtain the test performance results of the tested target.

[0008] Optionally, semantic parsing is performed on the target test scenario description information to construct a test scenario feature vector. The method includes: extracting text description information, structured input information, and configuration parameter information from the target test scenario description information; standardizing the text description information, structured input information, and configuration parameter information to obtain initialization scenario description information; performing word segmentation on the initialization scenario description information to extract key semantic elements; performing feature label processing on the key semantic elements to obtain a semantic label set; and performing multi-dimensional feature quantization processing on the semantic label set to output a test scenario feature vector.

[0009] Optionally, the key semantic elements include time-related keywords, behavior-related keywords, precision-related keywords, and constraint-related keywords; the key semantic elements are processed by feature labeling to obtain a set of semantic labels that integrate time scale features, dynamic change intensity features, precision requirement features, spatial range features, and redundancy tolerance features.

[0010] Optionally, the probability function mapping relationship between the test scenario feature vector and the evolution parameter space is trained with prediction error target, data redundancy target, and computational efficiency target as optimization objectives. The method includes: collecting historical test task data to construct a training sample set, wherein each training sample includes a test scenario feature vector sample, an adaptive evolution parameter combination sample, and multiple target evaluation values ​​corresponding to the evolution calculation results of the corresponding combination samples; initializing the probability function mapping model based on meta-learning; and using prediction error target, data redundancy target, and computational efficiency target as optimization objectives, the initialization probability function mapping model performs Bayesian iterative optimization on the training sample set by introducing multi-objective joint loss until the evolution parameter probability distribution corresponding to each test scenario feature vector sample is obtained, and constructing the probability function mapping relationship according to the evolution parameter probability distribution.

[0011] Optionally, the evolution parameter probability distribution is a conditional probability distribution generated based on the optimal solution of the evolution parameter combination under each test scenario feature vector sample; the optimal solution of the evolution parameter combination is an evolution parameter combination where the multi-objective joint loss data is less than a preset loss threshold.

[0012] Optionally, after obtaining the evolutionary parameter probability distribution corresponding to each test scene feature vector sample, the method further includes: calculating the feature vector similarity for each test scene feature vector sample in the training sample set, and obtaining the similarity calculation result; clustering the test scene feature vector samples according to the similarity calculation result to obtain test scene clusters; extracting the evolutionary parameter probability distribution clusters corresponding to the test scene clusters, fusing the evolutionary parameter probability distribution clusters, and updating the evolutionary parameter probability distribution corresponding to the test scene clusters; and re-establishing the probability function mapping relationship between the test scene clusters and the evolutionary parameter probability distributions.

[0013] Optionally, the method for initializing the probability function mapping model based on meta-learning includes: connecting to a historical test scenario knowledge base; retrieving prior test scenario feature vector samples from the historical test scenario knowledge base that have a preset vector similarity to each test scenario feature vector sample; extracting prior evolution parameter combinations of the prior test scenario feature vector samples; and performing preliminary training on the probability function mapping model using the prior test scenario feature vector samples and the prior evolution parameter combinations to establish the initialized probability function mapping model after meta-learning.

[0014] Optionally, after establishing the initial probability function mapping model after meta-learning, the method further includes: configuring a meta-learning guidance factor, traversing the evolution parameter space to calculate the guidance value of each candidate evolution parameter combination based on the meta-learning guidance factor; and executing the candidate evolution parameter combination selected in the current iteration according to the guidance value.

[0015] Optionally, the method for outputting an adaptive evolution parameter combination based on the probability function mapping relationship includes: identifying a matching test scene cluster corresponding to the test scene feature vector based on the probability function mapping relationship; extracting the evolution parameter probability distribution of the matching test scene cluster mapping; and sampling parameters according to the evolution parameter probability distribution to obtain an adaptive evolution parameter combination based on Pareto solution.

[0016] Optionally, the evolution prediction model includes at least one of a traffic evolution model, an energy consumption evolution model, and an environmental evolution model.

[0017] The beneficial effects of this invention are:

[0018] Semantic parsing of the target test scenario description information is performed to construct test scenario feature vectors, thereby transforming the semantic information of the test scenario into a quantifiable feature representation, providing a structured input foundation for subsequent adaptive matching of evolutionary parameters. For at least one evolutionary prediction model based on the CIM city information model, an evolutionary parameter space is constructed, including time step parameters, evolutionary time windows, data smoothing factors, historical weight decay coefficients, spatial coupling coefficients, and data compression rates. This comprehensively covers the key parameter dimensions affecting the computational behavior of the evolutionary prediction model, providing a complete range of parameter candidates for adaptive parameter adjustment under different test scenarios. With prediction error, data redundancy, and computational efficiency as optimization objectives, a probability function mapping relationship between the test scenario feature vectors and the evolutionary parameter space is trained. Based on this probability function mapping relationship, an adaptive evolutionary parameter combination is output, establishing an intelligent mapping mechanism between test scenario features and optimal evolutionary parameters, enabling automatic parameter inference and optimization selection for different test scenarios. Adaptive evolutionary parameters are combined and imported into an evolutionary prediction model to perform data evolution calculations, resulting in a dynamic evolutionary time-series dataset. This allows the evolutionary prediction model to generate dynamic time-series data that highly matches the requirements of the target test scenario, driven by scenario-adaptive parameters. The dynamic evolutionary time-series dataset is then transmitted as test cases to the test system terminal containing the target test scenario description information. The test performance results of the target system are obtained, thus driving the system under test to complete test verification with scenario-adaptive dynamic time-series data, obtaining test results that truly reflect the system's performance.

[0019] Through the above technical solution, the evolutionary parameter combination of the evolutionary prediction model is adaptively adjusted according to the semantic features of different test scenarios, so that the dynamic time series data output by the evolutionary prediction model can better fit the actual data needs of various test scenarios, effectively improving the scenario adaptability and accuracy of test cases, while taking into account data redundancy control and computational efficiency optimization, thereby improving the overall test verification effect and credibility of CIM-based urban information model data calculation. Attached Figure Description

[0020] Figure 1 A flowchart illustrating a method for calculating urban information model data based on CIM provided by this invention;

[0021] Figure 2 This is a schematic diagram illustrating the structure for constructing the probability function mapping relationship provided by the present invention. Detailed Implementation

[0022] 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.

[0023] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0024] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0025] Example 1, as Figure 1 As shown, this embodiment of the invention provides a method for calculating urban information model data based on CIM, including:

[0026] S1. Perform semantic parsing on the target test scenario description information to construct the test scenario feature vector.

[0027] Specifically, the target test scenario description information refers to the test task input corresponding to the test system terminal, used to describe the specific requirements of the current test scenario for the evolution data of the urban information model. Its sources can be test requirement documents written in natural language by testers, test configuration parameters submitted in the form of structured forms, or test task configuration files automatically generated by the system. Description information from different sources and in different formats exhibits heterogeneity in its expression, and directly using it for parameter calculations can lead to semantic ambiguity and format incompatibility issues. Therefore, it is necessary to first perform semantic parsing on the target test scenario description information, transforming it into a structured feature representation that can be used for subsequent calculations, i.e., the test scenario feature vector.

[0028] Specifically, when performing semantic parsing on the target test scenario description information, the text description information, structured input information, and configuration parameter information are extracted and standardized to eliminate format differences between information from different sources, resulting in initial scenario description information. Further, the initial scenario description information is segmented to extract key semantic elements, which are then labeled with feature tags to form a semantic tag set reflecting the multi-dimensional requirements of the test scenario. Subsequently, multi-dimensional feature quantization is performed on the semantic tag set, mapping each dimension of semantic features to continuous numerical values, outputting a test scenario feature vector. This test scenario feature vector is used to generate adaptive evolution parameter combinations, achieving precise alignment between the test scenario semantic information and the evolution parameter space.

[0029] By semantically parsing the target test scenario description information, a test scenario feature vector is constructed. This transforms heterogeneous and unstructured target test scenario description information into a unified feature vector with clear semantic dimensions, achieving a standardized expression of test scenario information. On the one hand, multi-dimensional feature quantization ensures that scenario description information from different sources and in different formats can be expressed in a unified feature space, providing a standardized feature foundation for the subsequent generation of adaptive evolution parameter combinations. On the other hand, there is a clear semantic correspondence between each dimension of the feature vector and each parameter in the evolution parameter space, making the adaptive adjustment of parameters have an interpretable scenario basis, thereby improving the overall evolutionary computation efficiency and parameter adaptation accuracy of the scheme.

[0030] S2. For at least one evolutionary prediction model based on the CIM city information model, construct an evolutionary parameter space, wherein the evolutionary parameter space includes at least the time step parameter, evolution time window, data smoothing factor, historical weight decay coefficient, spatial coupling coefficient and data compression rate.

[0031] Specifically, the CIM (City Information Model) serves as the computational foundation for evolutionary prediction models—the fundamental data platform upon which these models rely. It integrates the geometric and attribute information of urban spatial elements, along with their dynamic relationships, providing a unified data organization framework for the time-series extrapolation calculations of these models. Evolutionary prediction models are computational models that use the CIM to extrapolate and predict urban operational status data over time. They include at least one of three types: traffic evolution models, energy consumption evolution models, and environmental evolution models. Different types of evolutionary prediction models correspond to different dimensions of urban operational status. The quality of the computational results of the evolutionary prediction models directly depends on whether the configuration of the evolutionary parameters matches the data requirements of the target test scenario. Therefore, before performing evolutionary calculations, corresponding evolutionary parameter spaces are constructed for different evolutionary prediction models, clarifying the search range and constraint boundaries for subsequent adaptive parameter adjustments.

[0032] The evolution parameter space includes at least the time step parameter, evolution time window, data smoothing factor, historical weight decay coefficient, spatial coupling coefficient, and data compression ratio.

[0033] The time step parameter controls the granularity of each forward movement in the evolutionary prediction model during time-series extrapolation. A smaller time step parameter allows for more precise capture of urban state changes, but also increases computational overhead. Conversely, a larger time step parameter improves computational efficiency but may decrease prediction accuracy for rapidly changing states. For example, traffic evolution models are sensitive to short-term changes in urban traffic conditions, and their time step parameter can be set between 5 and 30 minutes. Energy consumption and environmental evolution models experience relatively gradual state changes, and their time step parameter can be appropriately widened to between 15 and 60 minutes.

[0034] The evolution time window is used to determine the time span covered by the evolution prediction model for state extrapolation, that is, the prediction range extending from the current moment to the future. Different test scenarios have significantly different requirements for the prediction time span. For example, traffic evolution models are mainly used for short-term extrapolation of urban traffic conditions, and their evolution time window can be set to a range of 1 to 12 hours; energy consumption evolution models and environmental evolution models need to make long-term predictions of urban energy consumption trends and environmental changes, and their evolution time window can be set to a range of 24 to 72 hours.

[0035] The data smoothing factor controls the degree of noise suppression in the input data during the evolution process. A larger value results in stronger filtering of abnormal fluctuations but weakens the sensitivity to sudden changes in the actual state. A smaller value retains more detailed features of the original data. For example, the input data of the environmental evolution model comes from sensor data collection and is susceptible to external noise such as measurement errors and signal interference. Its data smoothing factor can be set to a range of 0.5 to 0.9 to effectively filter sensor noise and ensure the stability of the input data for evolution calculation. The traffic evolution model has high requirements for responding to sudden changes in traffic state. An excessively large data smoothing factor will smooth out sudden congestion and other state changes. Its data smoothing factor can be set to a range of 0.1 to 0.5 to retain the details of dynamic changes in traffic state. The input data of the energy consumption evolution model has both periodic regularity and short-term fluctuation characteristics. A balance needs to be struck between noise filtering and detail preservation. Its data smoothing factor can be set to a range of 0.3 to 0.7 to suppress short-term fluctuation noise in energy consumption data while retaining the periodic trend.

[0036] The historical weight decay coefficient is used to adjust the contribution weight of historical data to the current evolution calculation, reflecting the decreasing importance of historical data over time. The larger the coefficient, the higher the model's dependence on recent data; the smaller the coefficient, the longer the influence of historical data on the evolution results lasts. For example, in traffic evolution models, urban traffic conditions change frequently, and recent data has higher reference value. The historical weight decay coefficient can be set to 0.8 to 0.99, thus making the evolution calculation more focused on recent traffic condition data and responding quickly to dynamic changes in traffic conditions. In energy consumption evolution models, energy consumption data has strong long-term regularity, and the reference value of historical data lasts for a long time. The historical weight decay coefficient can be set to 0.5 to 0.85 to fully utilize historical energy consumption patterns to provide effective guidance for the current evolution calculation. In environmental evolution models, environmental conditions are significantly affected by long-term factors such as seasons and climate. Historical data has important reference value for predicting current environmental trends. The historical weight decay coefficient can be set to 0.5 to 0.80 to achieve a reasonable balance between recent data and long-term historical patterns.

[0037] The spatial coupling coefficient is used to quantify the intensity of mutual influence between different spatial elements in the CIM (City Information Model). A larger value indicates stronger cross-spatial unit state transmission, while a smaller value indicates that each spatial unit tends to evolve independently. For example, in the traffic evolution model, there is a significant road network correlation effect between spatial units, with the traffic states of adjacent road segments influencing each other. The spatial coupling coefficient can be set from 0.5 to 1.0 to fully capture the road network spatial transmission effect. In the energy consumption evolution model, there is a certain energy consumption correlation between spatial units such as building clusters and industrial areas, but the correlation strength is weaker than that of the traffic road network. The spatial coupling coefficient can be set from 0.3 to 0.7 to reflect a moderate energy consumption transmission relationship between spatial units. In the environmental evolution model, the environmental state of each spatial unit is mainly affected by factors such as local pollution sources and meteorological conditions, and the direct correlation between spatial units is relatively weak. The spatial coupling coefficient can be set from 0.1 to 0.5 to avoid excessive cross-spatial unit state interference.

[0038] Data compression ratio controls the simplification of the output data from evolutionary calculations. Evolutionary prediction models with high redundancy in their output data can appropriately increase the data compression ratio to reduce data transmission and storage overhead. Conversely, evolutionary prediction models with highly dynamic output data should decrease the data compression ratio to retain more effective information. For example, energy consumption evolution models have output data with gradual changes and high repetition between adjacent time points; the data compression ratio can be set between 30% and 80%. Traffic evolution models have highly dynamic output data with significant differences in state across time points; the data compression ratio can be set between 10% and 40%. Environmental evolution models have output data exhibiting both short-term fluctuations and long-term trends; the data compression ratio can be set between 20% and 60%.

[0039] The evolutionary parameter space is jointly constituted by the time step parameter, evolutionary time window, data smoothing factor, historical weight decay coefficient, spatial coupling coefficient, and data compression ratio. Different types of evolutionary prediction models correspond to different parameter value ranges, forming their own independent parameter search spaces. The evolutionary parameter space clarifies the constraint boundaries of adaptive parameter adjustment, avoids unconstrained search of parameter combinations, and reduces the computational complexity of parameter optimization. At the same time, by constructing evolutionary parameter spaces for each evolutionary prediction model, the parameter search range is matched with the actual data characteristics of each model, providing a structured foundation for the accurate generation of adaptive evolutionary parameter combinations in different test scenarios.

[0040] S3. Train the probability function mapping relationship between the test scenario feature vector and the evolution parameter space with prediction error target, data redundancy target and computational efficiency target as optimization targets, and output an adaptive evolution parameter combination based on the probability function mapping relationship.

[0041] Specifically, the probability function mapping relationship takes the test scenario feature vector as input and the parameter probability distribution in the evolution parameter space as output, establishing an automated mapping between the semantic features of the test scenario and the configuration of evolution parameters. To train this probability function mapping relationship, historical test task data is used as the training basis. With prediction error, data redundancy, and computational efficiency as optimization objectives, iterative optimization ensures that the probability function mapping relationship accurately reflects the correspondence between different test scenario feature vectors and high-quality evolution parameter combinations. This allows for the automatic output of an evolution parameter combination that matches the requirements of any given test scenario feature vector from the corresponding evolution parameter space—that is, an adaptive evolution parameter combination.

[0042] Among these objectives, the prediction error objective measures the prediction accuracy of the evolutionary prediction model under the current parameter combination; a smaller prediction error indicates a higher degree of adaptability of the parameter combination to the data requirements of the test scenario. The data redundancy objective measures the degree of redundancy in the output data of the evolutionary computation; a lower redundancy degree indicates higher data efficiency of the parameter combination, provided that the accuracy requirements of the test scenario are met. The computational efficiency objective measures the computational resources consumed by the evolutionary computation under the current parameter combination; a lower computational cost indicates stronger engineering practicality of the parameter combination. The prediction error objective, data redundancy objective, and computational efficiency objective jointly constrain the optimal direction of the parameter combination, avoiding parameter bias problems caused by single-objective optimization, and ensuring that the output adaptive evolutionary parameter combination achieves a reasonable balance between prediction accuracy, data efficiency, and computational cost.

[0043] During the inference phase, the test scenario feature vector of the current target test scenario is input into the trained probability function mapping relationship. Based on the output evolution parameter probability distribution, parameter sampling is performed to obtain an adaptive evolution parameter combination that matches the requirements of the current target test scenario, which is then used for the calculation and execution of the subsequent evolution prediction model.

[0044] By outputting adaptive evolution parameter combinations, an automated mapping from semantic features of the test scenario to evolution parameter configurations is achieved. This ensures prediction accuracy while balancing data efficiency and computational overhead, providing adaptive parameter guarantees for the quality of evolution calculation results under different test scenarios.

[0045] S4. Import the adaptive evolution parameter combination into the evolution prediction model to perform data evolution calculation and obtain a dynamic evolution time series dataset.

[0046] Specifically, after obtaining the adaptive evolution parameter combination, it is imported into the corresponding evolution prediction model, driving the model to perform data evolution calculations based on the adaptive evolution parameter combination of the current target test scenario. Specifically, each parameter in the adaptive evolution parameter combination is bound to the corresponding parameter interface in the evolution prediction model, including the time step parameter, evolution time window, data smoothing factor, historical weight decay coefficient, spatial coupling coefficient, and data compression rate. After each parameter is configured according to its value in the adaptive evolution parameter combination, the evolution prediction model uses the currently stored urban operation status data of the CIM city information model as the initial state and gradually advances according to the time step parameter, continuously performing state deduction calculations within the time span defined by the evolution time window.

[0047] During the extrapolation process at each time step, the evolutionary prediction model first performs noise suppression processing on the input data of the current time step based on the data smoothing factor, obtaining smoothed input data for the current time step. Then, it weights the state data of each historical time step according to the historical weight decay coefficient, assigning lower weights to historical data further away from the current time step. The weighted historical state data is then fused with the smoothed input data of the current time step to obtain the state estimate for the current time step. Further, it models the state transfer between adjacent spatial units based on the spatial coupling coefficient, and weights and superimposes the state estimates of adjacent spatial units according to the spatial coupling coefficient to output the final predicted state value of each spatial unit at the current time step. After completing the extrapolation calculations for all time steps within the evolution time window, the output data of each time step is simplified according to the data compression ratio, and the output results of all time steps are organized in time sequence to obtain a dynamic evolution time series dataset.

[0048] The dynamic evolution time-series dataset records the complete prediction sequence of urban operational status by the evolution prediction model within the evolution time window under the current adaptive evolution parameter combination configuration. It includes time-series prediction data for dimensions such as traffic status, energy consumption status, and environmental status, with the specific dimension depending on the type of evolution prediction model used. This dynamic evolution time-series dataset serves as the data source for test cases, transmitted to the test system terminal for testing and verification of the target object.

[0049] By importing the adaptive evolutionary parameter combination into the evolutionary prediction model to perform data evolution calculations, the computational behavior of the evolutionary prediction model can accurately respond to the data requirements of the target test scenario. The resulting dynamic evolutionary time series dataset meets the requirements of the target test scenario in terms of prediction accuracy, data redundancy, and computational overhead, providing a reliable data foundation for the generation of subsequent test cases.

[0050] S5. Transmit the dynamic evolution time series dataset as test cases to the test system terminal of the target test scenario description information to obtain the test performance results of the target under test.

[0051] Specifically, the test system terminal refers to the test execution platform corresponding to the target test scenario description information. It is responsible for receiving test case data and driving the target under test to run under the test case, thereby collecting the response data of the target under test and outputting test performance results. The target under test refers to the specific object that the test system terminal wants to verify. It can be a software system or functional module related to the CIM city information model, such as an urban traffic scheduling algorithm, an energy consumption management system, or an environmental monitoring and early warning module.

[0052] Specifically, the obtained dynamic evolution time-series dataset is encapsulated according to the data format required by the test system terminal to form structured test cases. Each test case includes a time-series index, city operational status prediction data for each time step, and test scenario configuration information. The time-series index identifies the temporal position of the data at each time step. The city operational status prediction data includes numerical sequences of dimensions such as traffic status, energy consumption status, or environmental status, depending on the evolution prediction model type. The test scenario configuration information informs the test system terminal of the target test scenario corresponding to the current test case. After encapsulation, the test cases are transmitted to the test system terminal via a standard data interface or message transmission protocol.

[0053] After receiving the test cases, the test system terminal sequentially inputs the data for each time step into the target under test according to the time sequence index. This drives the target under test to execute response processing step-by-step according to the city operation state sequence defined in the test cases, and collects the output results of the target under test at each time step. After completing the test execution for all time steps, the test system terminal summarizes and analyzes the collected results for each time step, outputting the test performance results of the target under test. The test performance results include at least the performance indicators of the target under test under the current test cases, such as response accuracy, response latency, and exception handling capability, used to evaluate the actual operating capability of the target under test in the target test scenario.

[0054] By transmitting the dynamically evolving time-series dataset as test cases to the test system terminal and obtaining test performance results, an automated testing closed loop based on CIM city information model evolution data was achieved. Compared to the traditional method of manually constructing test cases, the test cases generated in this embodiment are derived from an evolution prediction model with adaptive parameter configuration. Their data characteristics are highly matched with the actual needs of the target test scenario, effectively improving the coverage of test cases to the real-world operating scenario of the target under test, thereby enhancing the reliability and effectiveness of the test results.

[0055] By semantically parsing the description information of the target test scenario to construct a test scenario feature vector, and constructing an evolutionary parameter space for the evolutionary prediction model, the probability function mapping relationship between the test scenario feature vector and the evolutionary parameter space is trained with prediction error target, data redundancy target, and computational efficiency target as optimization guides. Based on the probability function mapping relationship, an adaptive evolutionary parameter combination matching the current target test scenario is output, driving the evolutionary prediction model to perform data evolution calculation to obtain a dynamic evolutionary time series dataset, which is then transmitted as a test case to the test system terminal to obtain test performance results. This achieves the technical effect of adaptively adjusting the evolutionary parameter combination according to the test scenario characteristics, making the dynamic evolutionary time series data output by the evolutionary prediction model more in line with the needs of various test scenarios, thereby improving the accuracy and computational efficiency of test cases.

[0056] Furthermore, semantic parsing is performed on the target test scenario description information to construct a test scenario feature vector. Methods include:

[0057] S11. Extract the text description information, structured input information, and configuration parameter information of the target test scenario description information;

[0058] S12. Standardize the text description information, structured input information, and configuration parameter information to obtain the initial scene description information;

[0059] S13. Perform word segmentation on the initialization scene description information to extract key semantic elements, and perform feature label processing on the key semantic elements to obtain a semantic label set.

[0060] S14. Perform multi-dimensional feature quantization processing on the semantic tag set and output the test scene feature vector.

[0061] In a preferred embodiment, the text description information, structured input information, and configuration parameter information of the target test scenario description information are extracted first. The target test scenario description information comes from diverse sources and has heterogeneous formats in practical applications, therefore it needs to be categorized and extracted first. Text description information refers to test requirement documents written by testers in natural language, such as "Stress testing is required on the traffic scheduling algorithm during the morning rush hour on urban main roads, covering abnormal scenarios such as sudden increases in traffic flow and traffic light malfunctions." This type of information is semantically rich but loosely structured, requiring subsequent semantic extraction using natural language processing techniques. Structured input information refers to test configuration data submitted in structured formats such as forms, key-value pairs, XML, or JSON, such as {"Test Object": "Traffic Scheduling Module", "Test Priority": "High", "Spatial Range": "Urban Main Roads", "Constraints": "Peak Traffic Flow ≥ 3000 vehicles / hour"}. This type of information has a clear structure but limited semantic expression. Configuration parameter information refers to a set of parameters that quantify the constraints on the execution conditions of test tasks. It supplements execution-level requirements that are not explicitly quantified in text descriptions and structured input information. For example, the test execution time window specifies the time range to be covered in this test; the data sampling frequency specifies the time interval for data collection during the test; the performance evaluation threshold specifies the performance standard of the target under test in the current test scenario; and the anomaly triggering condition specifies the specific conditions under which anomalies are triggered during the test. Configuration parameter information is primarily in the form of numerical or enumerated values, representing a quantitative expression of the execution constraints of the target test scenario.

[0062] Subsequently, the text description information, structured input information, and configuration parameter information are standardized to obtain the initial scene description information. Since the text description information, structured input information, and configuration parameter information differ in format, encoding, and semantic expression, they need to be standardized separately before being unified and integrated. Specifically, when standardizing the text description information, the following operations are performed in sequence: removing redundant symbols, special characters, and format marks from the text; performing a unified encoding format conversion on the text (e.g., unifying it to UTF-8 encoding); and standardizing and expanding technical terms and abbreviations appearing in the text, for example, unifying "morning rush hour" to "07:00-09:00 time period" and unifying "main road" to the corresponding road level identifier. When standardizing the structured input information, operations such as field naming standardization, filling missing fields with default values, and unified conversion of numerical units are performed to ensure that structured data from different sources are consistent in field definitions and numerical units. When standardizing the configuration parameter information, operations such as parameter naming standardization and numerical range normalization are performed to uniformly map the values ​​of each configuration parameter to a predefined value range. After completing the above standardization process, the three types of standardized information are integrated according to a unified data structure to obtain initial scene description information with a unified format and consistent semantics.

[0063] Next, the initial scenario description information is segmented to extract key semantic elements. These key semantic elements are then labeled to obtain a semantic tag set. Specifically, firstly, natural language processing (NLP) is used to segment the text content in the initial scenario description information, dividing the continuous text into a sequence of lexical units with independent semantic meaning. Based on the segmentation, and combined with a predefined dictionary of key semantic elements, key semantic elements directly related to the test scenario requirements are identified and extracted from the lexical unit sequence. For example, from the statement "It needs to cover traffic state changes for 6 consecutive hours during the morning rush hour, with high prediction accuracy requirements and a certain degree of data redundancy allowed," "6 consecutive hours" is extracted as a time-related key semantic element, "traffic state changes" as a behavior-related key semantic element, "high prediction accuracy requirements" as a precision-related key semantic element, and "a certain degree of data redundancy allowed" as a constraint-related key semantic element. Then, the extracted key semantic elements are matched with a predefined semantic tag system. Feature tag processing is performed on each key semantic element, mapping it to the corresponding tag category and tag value in the semantic tag system. For example, "6 consecutive hours" is mapped to a time-scale feature label, with the label value corresponding to a long-term demand within the evolution time window; "traffic status change" is mapped to a dynamic change intensity feature label, with the label value corresponding to high dynamic change intensity; "high prediction accuracy requirement" is mapped to an accuracy requirement feature label, with the label value corresponding to low error tolerance; and "allowing a certain degree of data redundancy" is mapped to a redundancy tolerance feature label, with the label value corresponding to high redundancy tolerance. Finally, the feature labels corresponding to all key semantic elements are aggregated to form a semantic label set reflecting the multi-dimensional demand characteristics of this test scenario.

[0064] Subsequently, multi-dimensional feature quantization processing is performed on the semantic label set to output the test scenario feature vector. Multi-dimensional feature quantization processing refers to converting the semantic labels of each dimension in the semantic label set into continuous numerical representations for subsequent numerical calculations. Specifically, for degree-type semantic labels, a predefined level rating system is used to map them to normalized values. For example, the intensity of dynamic change is mapped to 0.0, 0.5, and 1.0 for low, medium, and high levels, respectively; the precision requirement is mapped to 0.0, 0.5, and 1.0 for low, medium, and high levels, respectively, and so on, ensuring that all degree-type semantic labels can participate in the construction of subsequent feature vectors with a unified numerical scale. For numerical semantic labels, the values ​​are directly extracted and normalized according to a predefined value range. For example, the value 6 is extracted from "6 consecutive hours," normalized according to the evolution time window range of 1 hour to 72 hours, and mapped to the corresponding normalized value, ensuring that numerical semantic labels of different dimensions can participate in the construction of feature vectors under a unified numerical scale. After quantizing the semantic labels for each dimension, the quantized values ​​of each dimension are concatenated into a fixed-length vector according to the predefined feature dimension order, outputting the test scene feature vector. The value of each dimension in this test scene feature vector has a clear semantic correspondence with the configuration requirements of the corresponding parameters in the evolution parameter space, providing standardized numerical input for subsequent training and inference of the probability function mapping relationship.

[0065] Through the above steps, it is ensured that scene description information from different sources and in different formats in the target test scenario can be standardized and transformed into test scenario feature vectors with clear semantic dimensions. This ensures that the values ​​of each dimension in the test scenario feature vector correspond clearly to the configuration requirements of each parameter in the evolution parameter space, providing a high-quality feature foundation for subsequent steps.

[0066] Furthermore, the key semantic elements include time-related keywords, behavior-related keywords, precision-related keywords, and constraint-related keywords;

[0067] S131. The key semantic elements are processed by feature labeling to obtain a set of semantic labels that integrate time scale features, dynamic change intensity features, accuracy requirement features, spatial range features and redundancy tolerance features.

[0068] In a preferred embodiment, the key semantic elements include time-related keywords, behavior-related keywords, precision-related keywords, and constraint-related keywords.

[0069] Among them, time-related keywords refer to words related to time requirements in the description information of the target test scenario, including time span description words, time granularity description words, and time stage description words, such as "6 consecutive hours" or "72 consecutive hours" to describe the test coverage duration, "every 5 minutes" or "statistics by hour" to describe the sampling interval, and "morning peak period" or "all day" to describe the test period, which are used to reflect the test scenario's requirements for evolution time window length and time step parameters.

[0070] Behavioral keywords refer to words in the target test scenario description information that are related to the dynamic changes in the city's operational status. These include words describing the magnitude of change, the frequency of change, and the type of event. For example, words describing the intensity of change include "drastic fluctuations" and "stable changes," words describing the frequency of change include "frequent switching" and "slow evolution," and words describing specific events include "sudden congestion," "peak traffic flow," and "abnormal emissions." These keywords are used to reflect the intensity characteristics of the dynamic changes in the city's status in the test scenario.

[0071] Accuracy-related keywords refer to words in the description information of the target test scenario that are related to the prediction accuracy and data quality requirements. These include accuracy level descriptors, error tolerance descriptors, and data integrity descriptors. For example, words describing accuracy requirements such as "high accuracy" and "strict accuracy requirements"; words describing error tolerance range such as "allow a certain error" and "error not exceeding 5%"; and words describing data integrity requirements such as "no loss of key data" and "complete coverage". These keywords reflect the lower limit of the prediction error tolerance of the test scenario.

[0072] Constraint keywords refer to words in the description information of the target test scenario that are related to the test execution constraints. These include spatial range descriptors, redundancy tolerance descriptors, and resource constraint descriptors. For example, words describing spatial coverage such as "urban main road", "citywide", and "local area", words describing data redundancy tolerance such as "allow data redundancy" and "compressed storage", and words describing computing resource constraints such as "limited computing resources" and "low power mode" are used to reflect the requirements of the test scenario in terms of spatial coverage, data redundancy tolerance, and computing resource constraints.

[0073] When processing key semantic elements to obtain a semantic tag set, the extracted time-related keywords, behavioral keywords, accuracy-related keywords, and constraint-related keywords are mapped to a predefined semantic tag system, resulting in a semantic tag set that integrates time-scale features, dynamic change intensity features, accuracy requirement features, spatial range features, and redundancy tolerance features. Specifically, time-scale features are mapped from time-related keywords, reflecting the test scenario's requirements for evolution time window length and time step parameters; for example, "6 consecutive hours" maps to a medium-time-scale feature, and "72 consecutive hours" maps to a long-time-scale feature. Dynamic change intensity features are mapped from behavioral keywords, reflecting the drastic changes in urban state data within the test scenario; for example, "sudden congestion" and "drastic fluctuations" map to high dynamic change intensity features, while "slow evolution" and "stable change" map to low dynamic change intensity features. Accuracy requirement features are mapped from accuracy-related keywords, reflecting the test scenario's tolerance limit for prediction errors; for example, "high accuracy" and "error not exceeding 5%" map to high accuracy requirement features, while "allowing a certain error" maps to low accuracy requirement features. Spatial extent features are mapped from spatial extent descriptors in the constraint keywords, reflecting the scale of urban spatial coverage involved in the test scenario. For example, "citywide" maps to a large spatial extent feature, while "local area" maps to a small spatial extent feature. Redundancy tolerance features are mapped from redundancy tolerance descriptors in the constraint keywords, reflecting the test scenario's tolerance for duplicate data. For example, "allow data redundancy" maps to a high redundancy tolerance feature, while "compressed storage" maps to a low redundancy tolerance feature.

[0074] By refining key semantic elements into time-related keywords, behavior-related keywords, precision-related keywords, and constraint-related keywords, and mapping them to a set of semantic labels in five dimensions—time scale features, dynamic change intensity features, precision requirement features, spatial range features, and redundancy tolerance features—each dimension in the test scenario feature vector has a clear semantic origin and physical meaning. This improves the accuracy of the feature vector in expressing the requirements of the test scenario and provides more granular feature support for accurately identifying the parameter configuration requirements of different test scenarios in subsequent probability function mapping relationships.

[0075] Furthermore, such as Figure 2 As shown, the method for training the probability function mapping relationship between the test scene feature vector and the evolution parameter space with prediction error target, data redundancy target, and computational efficiency target as optimization objectives includes:

[0076] S31. Collect historical test task data to construct a training sample set, wherein each training sample includes test scene feature vector samples, adaptive evolution parameter combination samples, and multiple target evaluation values ​​corresponding to the evolution calculation results of the corresponding combination samples.

[0077] S32. Meta-learning-based initialization probability function mapping model;

[0078] S33. With prediction error target, data redundancy target and computational efficiency target as optimization targets, the initialization probability function mapping model performs Bayesian iterative optimization on the training sample set by introducing multi-objective joint loss until the evolution parameter probability distribution corresponding to the feature vector sample of each test scene is obtained, and the probability function mapping relationship is constructed according to the evolution parameter probability distribution.

[0079] In a preferred embodiment, firstly, historical test task data is collected to construct a training sample set. Historical test task data refers to test record data accumulated during the execution of past test tasks, including descriptions of historical test scenarios, combinations of evolutionary parameters used, and corresponding evaluation data of evolutionary calculation results. The historical test scenario descriptions in the historical test task data have the same data format as the target test scenario descriptions, both containing text descriptions, structured input information, and configuration parameter information. Therefore, they can be processed according to the semantic parsing method in S1 to obtain test scenario feature vector samples corresponding to each historical test scenario. The evolutionary parameter combinations used in each historical test task are extracted to obtain adaptive evolutionary parameter combination samples. For the evolutionary calculation results corresponding to each adaptive evolutionary parameter combination sample, the prediction error target evaluation value, data redundancy target evaluation value, and computational efficiency target evaluation value are calculated respectively to obtain multiple target evaluation values. Among them, the prediction error target evaluation value refers to the error metric between the prediction result of the evolution prediction model under the current evolution parameter combination and the actual urban operation status. The smaller the error metric value, the higher the prediction accuracy of the parameter combination. The data redundancy target evaluation value refers to the redundancy metric of the evolution calculation output data under the current evolution parameter combination. The lower the redundancy metric, the higher the data efficiency of the parameter combination. The computational efficiency target evaluation value refers to the computational resource consumption metric of the evolution calculation under the current evolution parameter combination. The smaller the computational resource consumption, the higher the computational efficiency of the parameter combination. The test scenario feature vector samples, adaptive evolution parameter combination samples, and the corresponding multiple target evaluation values ​​are combined into a training sample. The training samples corresponding to all historical test task data are aggregated to construct a training sample set.

[0080] Then, a probability function mapping model is initialized based on meta-learning. Specifically, a historical test scenario knowledge base is connected, and prior test scenario feature vector samples similar to the feature vector samples of each test scenario in the training sample set are retrieved from the historical test scenario knowledge base. Prior evolution parameter combinations corresponding to the prior test scenario feature vector samples are extracted. The probability function mapping model is initially trained using the prior test scenario feature vector samples and the prior evolution parameter combinations as training data, resulting in an initialized probability function mapping model. This enables the probability function mapping model to have a preliminary understanding of the mapping rules between different test scenarios and evolution parameters before formal Bayesian iterative optimization, thereby accelerating the convergence speed of subsequent Bayesian iterative optimization and reducing the number of samples required for training.

[0081] Subsequently, with prediction error, data redundancy, and computational efficiency as optimization objectives, the probability function mapping model is initialized and Bayesian iterative optimization is performed on the training sample set by introducing a multi-objective joint loss until the evolutionary parameter probability distribution corresponding to the feature vector sample of each test scenario is obtained. The probability function mapping relationship is then constructed according to the evolutionary parameter probability distribution. Specifically, the multi-objective joint loss consists of a weighted sum of prediction error loss, data redundancy loss, and computational efficiency loss. The prediction error loss is measured by the mean squared error between the prediction results of the evolutionary prediction model under the current parameter combination and the actual urban operating status. Specifically, the difference between the predicted state value output by the evolutionary prediction model at each time step under the current parameter combination and the actual urban operating status value at the corresponding time step is squared. The average of the squared differences of all time steps within the evolutionary time window is then calculated to obtain the mean squared error value. A larger mean squared error value indicates lower prediction accuracy for the current parameter combination, and a higher corresponding prediction error loss. The data redundancy loss is measured by the average repetition of data from adjacent time steps in the evolutionary calculation output data. Specifically, the cosine similarity between the output data of adjacent time steps is calculated, and the average repetition of data from adjacent time steps within the evolutionary time window is then calculated. The average cosine similarity of all adjacent time steps is taken to obtain the average repetition value. A higher average repetition value indicates greater redundancy in the output data, and the corresponding data redundancy loss is higher. Computational efficiency loss is measured by the normalized computation time consumed in the evolutionary computation under the current parameter combination. A longer normalized computation time indicates greater computational overhead, and the corresponding computational efficiency loss is higher. Normalized computation time is calculated by dividing the actual computation time under the current parameter combination by the maximum computation time of all candidate parameter combinations in the evolutionary parameter space, mapping the computation time to the range of 0 to 1. This allows for comparison of computational overhead between different parameter combinations on a uniform numerical scale, avoiding dimensional interference from absolute differences in computation time in the joint loss calculation. The three types of loss values ​​are weighted and summed according to predefined weights to obtain the multi-objective joint loss value. The weight configuration is set according to the emphasis of the three optimization objectives based on the test scenario. For example, for test scenarios with high accuracy requirements, the weight of prediction error loss can be set to 0.6, the weight of data redundancy loss to 0.2, and the weight of computational efficiency loss to 0.2. For test scenarios with limited computational resources, the weight of prediction error loss can be set to 0.3, the weight of data redundancy loss to 0.2, and the weight of computational efficiency loss to 0.5. The sum of the weights of the three types of losses is 1.0. The smaller the joint loss value, the better the overall performance of the current evolution parameter combination in terms of prediction accuracy, data efficiency, and computational cost.

[0082] The initialization probability function mapping model iteratively trains the training sample set according to the following Bayesian iterative optimization process. In each iteration, candidate parameter combinations in the evolutionary parameter space are sampled based on the current probability function mapping model. The sampled candidate parameter combinations are then imported into the evolutionary prediction model for evolutionary computation. Prediction error loss, data redundancy loss, and computational efficiency loss are calculated separately, and a weighted sum is obtained to obtain the multi-objective joint loss value. The parameters of the probability function mapping model are updated based on the multi-objective joint loss value, so that the model gradually improves the probability estimation accuracy of high-quality parameter combination regions. The above sampling, computation, and update process is repeated until the convergence condition is met, that is, the change in the multi-objective joint loss value is lower than the preset loss threshold after several consecutive iterations, or the number of iterations reaches the preset iteration limit. After the model converges, the probability distribution of the evolutionary parameters corresponding to each test scenario feature vector sample of the probability function mapping model tends to stabilize. The probability distribution of the evolutionary parameters reflects the probability that each parameter combination in the evolutionary parameter space is rated as a high-quality parameter combination under the given test scenario feature vector conditions. The higher the probability, the better the overall performance of the parameter combination in the corresponding test scenario. Finally, a probability function mapping relationship is constructed based on the mapping relationship between the feature vector samples of each test scenario and their corresponding evolution parameter probability distribution.

[0083] Through the training process described above, the probability function mapping relationship learns the correspondence between different test scenario features and high-quality evolution parameter combinations from historical test task data. On the one hand, meta-learning initialization allows the model to fully utilize historical prior knowledge, reducing the sample size required for training from scratch. On the other hand, Bayesian iterative optimization under multi-objective joint loss constraints ensures that the probability function mapping relationship can output evolution parameter combinations with excellent overall performance in terms of prediction accuracy, data efficiency, and computational cost, providing a reliable mapping foundation for the accurate generation of adaptive evolution parameters under different test scenarios.

[0084] Furthermore, the evolution parameter probability distribution is a conditional probability distribution generated based on the optimal solution of the evolution parameter combination under each test scenario feature vector sample;

[0085] The optimal solution for the evolution parameter combination is the evolution parameter combination where the joint loss data of multiple objectives is less than a preset loss threshold.

[0086] In a preferred embodiment, the evolutionary parameter probability distribution is a conditional probability distribution generated based on the optimal solution of the evolutionary parameter combination under each test scenario feature vector sample. The optimal solution of the evolutionary parameter combination refers to the combination of evolutionary parameters whose joint loss value during Bayesian iterative optimization is less than a preset loss threshold; that is, a parameter combination whose comprehensive performance across the three optimization objectives of prediction error, data redundancy, and computational efficiency meets a preset quality standard. The preset loss threshold is set according to the quality requirements of the actual test scenario; for example, it can be set to 0.15. For example, in the Bayesian iterative optimization process, if the prediction error loss of a candidate evolution parameter combination is 0.06, the data redundancy loss is 0.04, and the computational efficiency loss is 0.03, and the multi-objective joint loss value after being weighted and summed with weights of 0.6, 0.2, and 0.2 is 0.054, which is lower than the preset loss threshold of 0.15, then this parameter combination is identified as a good solution. If the multi-objective joint loss value of another candidate parameter combination is 0.18, which is higher than the preset loss threshold, then this parameter combination is not identified as a good solution and is not included in the fitted data of the conditional probability distribution.

[0087] For each test scenario feature vector sample, all evolutionary parameter combinations evaluated as optimal solutions during the Bayesian iterative optimization process are summarized. Based on the distribution of these optimal solutions in the evolutionary parameter space, a probability density fit is performed to generate a conditional probability distribution conditioned on the current test scenario feature vector sample. This conditional probability distribution describes the clustering location and degree of historical optimal solution parameter combinations in the evolutionary parameter space under a given test scenario feature vector. Regions with higher probability density indicate that parameter combinations within that region are more likely to meet the optimal solution criteria in the corresponding test scenario. Subsequent parameter sampling prioritizes sampling from high probability density regions, thereby improving the output quality of adaptive evolutionary parameter combinations. For example, for the feature vector of a traffic test scenario with high dynamic change intensity and medium time scale, 47 sets of evolutionary parameter combinations that meet the optimal solution conditions were selected during the historical Bayes iterative optimization process. After performing probability density fitting on the distribution of these 47 optimal solutions in the evolutionary parameter space, it was found that the time step parameter is concentrated in the range of 5 minutes to 15 minutes, the historical weight decay coefficient is concentrated in the range of 0.88 to 0.96, and the data smoothing factor is concentrated in the range of 0.1 to 0.3. The probability density corresponding to the above ranges is significantly higher than that of other regions in the evolutionary parameter space, forming a high probability density region under the feature vector of this test scenario. Subsequent parameter sampling will be preferentially carried out in the above high probability density region.

[0088] By limiting the probability distribution of evolutionary parameters to a conditional probability distribution generated based on optimal solutions for the combination of evolutionary parameters, the output of the probability function mapping relationship focuses on the clustering distribution of high-quality parameter combinations in the evolutionary parameter space. During the inference phase, based on the conditional probability distribution corresponding to the feature vector of the current test scenario, parameter sampling is prioritized in high-probability-density regions. Specifically, adaptive evolutionary parameter combinations are selected from clusters of parameter combinations that have been repeatedly evaluated as optimal solutions in similar test scenarios historically. This avoids the problem of consuming a large amount of computational resources in low-quality parameter combination regions due to uniform random sampling across the entire evolutionary parameter space, significantly improving output efficiency while ensuring the quality of the adaptive evolutionary parameter combination output.

[0089] Furthermore, after obtaining the probability distribution of evolutionary parameters corresponding to each test scenario feature vector sample, the method also includes:

[0090] S341. Calculate the feature vector similarity for each test scenario feature vector sample in the training sample set, and obtain the similarity calculation results.

[0091] S342. Cluster the test scene feature vector samples according to the similarity calculation results to obtain test scene clusters;

[0092] S343. Extract the evolution parameter probability distribution clusters corresponding to the test scenario clusters, fuse the evolution parameter probability distribution clusters, and update the evolution parameter probability distributions corresponding to the test scenario clusters.

[0093] S344. Re-establish the probability function mapping relationship between the test scenario cluster and the probability distribution of the evolution parameters.

[0094] In a preferred embodiment, after obtaining the evolution parameter probability distribution corresponding to each test scenario feature vector sample, the test scenario feature vector samples are further subjected to cluster analysis, and the evolution parameter probability distributions within the same test scenario cluster are fused and updated to improve the generalization ability of the probability function mapping relationship to new test scenarios that have not appeared in the training sample set.

[0095] First, feature vector similarity is calculated for each test scenario feature vector sample in the training sample set to obtain the similarity calculation results. Specifically, cosine similarity is calculated between every two test scenario feature vector samples in the training sample set. A higher cosine similarity indicates that the two test scenario feature vector samples are more similar in semantic features, that is, the two test scenarios are more similar in overall requirements in dimensions such as time scale requirements, intensity of dynamic changes, accuracy requirements, spatial range, and redundancy tolerance. The cosine similarity is calculated between all pairwise test scenario feature vector samples in the training sample set to obtain a similarity calculation result matrix. As the similarity calculation result, the value in the i-th row and j-th column of the matrix represents the similarity between the i-th test scenario feature vector sample and the j-th test scenario feature vector sample.

[0096] Then, the test scene feature vector samples are clustered according to the similarity calculation results to obtain test scene clusters. Specifically, using the similarity calculation result matrix as input, a hierarchical clustering method is used to cluster the test scene feature vector samples. Hierarchical clustering adopts a bottom-up merging strategy. Initially, each test scene feature vector sample is treated as an independent cluster. Then, in each round, the two clusters with the highest similarity in the similarity calculation result matrix are found and merged into a new cluster. The similarity value related to the new cluster in the similarity calculation result matrix is ​​updated. The above merging operation is repeated until the similarity between all remaining clusters is lower than the preset similarity threshold, at which point the merging stops, and the final test scene clusters are obtained. The preset similarity threshold is set according to the size of the training sample set and the diversity of the scenes. For example, the preset similarity threshold can be set to 0.85, that is, test scene feature vector samples with a cosine similarity higher than 0.85 are classified into the same test scene cluster. After clustering is completed, the arithmetic mean of the values ​​of each dimension of all test scenario feature vector samples within each test scenario cluster is taken to obtain the cluster center vector of that test scenario cluster. The cluster center vector represents the average demand feature of all samples within that test scenario cluster in each feature dimension.

[0097] Subsequently, the evolutionary parameter probability distribution clusters corresponding to the test scene clusters are extracted, and these clusters are fused to update the evolutionary parameter probability distributions corresponding to the test scene clusters. Specifically, for each test scene cluster, the evolutionary parameter probability distributions corresponding to all test scene feature vector samples within that cluster are extracted to form an evolutionary parameter probability distribution cluster. When fusing the evolutionary parameter probability distribution clusters, the cosine similarity between each test scene feature vector sample within the cluster and the cluster center vector is first calculated. The cosine similarity value of each sample is divided by the sum of the cosine similarity values ​​of all samples within the cluster to obtain the normalized fusion weight corresponding to each sample. After normalization, the sum of the fusion weights of each sample is 1.0, and the sample that is more similar to the cluster center has a higher fusion weight. For example, a test scene cluster contains 3 samples whose cosine similarities to the cluster center are 0.92, 0.87, and 0.79, respectively, with a sum of 2.58. The corresponding normalized fusion weights are 0.357, 0.337, and 0.306, respectively. The evolution parameter probability distributions of each sample within the cluster are weighted and averaged according to the normalized fusion weights. This means that the probability density functions of each sample are superimposed with weights and then normalized to obtain a unified evolution parameter probability distribution for the test scenario cluster. This unified evolution parameter probability distribution replaces the original independent evolution parameter probability distributions of each sample, thus completing the update of the evolution parameter probability distribution.

[0098] Next, a new probability function mapping relationship is established between test scenario clusters and the probability distributions of evolutionary parameters. Specifically, taking the cluster center vector of each test scenario cluster as a representative, a mapping relationship is established between the test scenario cluster and its corresponding unified evolutionary parameter probability distribution, replacing the original probability function mapping relationship constructed based on a single test scenario feature vector sample, resulting in an updated probability function mapping relationship. During the inference phase, the test scenario feature vector of the current target test scenario is matched with the cluster center vectors of each test scenario cluster based on similarity. The most matching test scenario cluster is identified, and the unified evolutionary parameter probability distribution corresponding to that test scenario cluster is extracted. Based on this probability distribution, the adaptive evolutionary parameter combination is generated.

[0099] By clustering the feature vector samples of the test scenarios and fusing and updating the evolution parameter probability distributions within the same cluster, the original independent probability distributions built based on individual historical samples are integrated into a unified probability distribution based on scenario clusters. On the one hand, this eliminates the noise influence caused by sample bias in the probability distribution of individual historical samples, making the probability function mapping relationship smoother and more stable. On the other hand, the probability function mapping relationship built based on test scenario clusters can cover the parameter requirements of all similar scenarios within the cluster. For new test scenarios that have not appeared in the training sample set, as long as the similarity between its feature vector and the cluster center of a certain test scenario cluster reaches a threshold, the corresponding unified evolution parameter probability distribution can be matched to complete the generation of adaptive evolution parameter combinations, thereby improving the parameter adaptability of the probability function mapping relationship when facing different test scenarios.

[0100] Furthermore, methods for initializing probability function mapping models based on meta-learning include:

[0101] S321. Connect to the historical test scenario knowledge base, retrieve prior test scenario feature vector samples that have reached a preset vector similarity with each test scenario feature vector sample from the historical test scenario knowledge base, and extract the prior evolution parameter combination of the prior test scenario feature vector samples.

[0102] S322. The probability function mapping model is initially trained using the prior test scene feature vector samples and the prior evolution parameter combination to establish the initial probability function mapping model after meta-learning.

[0103] In a preferred embodiment, when initializing the probability function mapping model based on meta-learning, firstly, a historical test scenario knowledge base is connected. Prior test scenario feature vector samples that achieve a preset similarity with each test scenario feature vector sample are retrieved from the historical test scenario knowledge base. Prior evolutionary parameter combinations of these prior test scenario feature vector samples are then extracted. The historical test scenario knowledge base refers to a previously accumulated database of historical test scenarios, storing test scenario feature vectors for various test scenarios in past test tasks and their corresponding, practically verified, evolutionary parameter combinations—i.e., prior evolutionary parameter combinations. Unlike the adaptive evolutionary parameter combination samples in the training sample set, the prior evolutionary parameter combinations are evolutionary parameter combinations that have been practically verified and confirmed as valid in historical test tasks. They have high reliability and can serve as a reliable source of prior knowledge for initializing the probability function mapping model. Specifically, for each test scenario feature vector sample in the training sample set, the cosine similarity between it and all prior test scenario feature vector samples in the historical test scenario knowledge base is calculated. The samples are sorted from high to low cosine similarity, and the top K prior test scenario feature vector samples with cosine similarity reaching a preset vector similarity are extracted as prior scenarios similar to the current test scenario feature vector sample. Here, K is the preset number of searches. The preset vector similarity is used to ensure that the retrieved prior scenarios and the current test scenario have sufficient semantic similarity. The preset vector similarity is set based on the distribution and diversity of prior test scenario feature vector samples in the historical test scenario knowledge base. A higher preset vector similarity results in a more similar prior scenario to the current test scenario, but may lead to fewer searches. A lower preset vector similarity results in a wider search range, but may introduce prior scenarios with significant semantic differences, affecting the initialization quality. For example, K can be set to 5, and the preset vector similarity threshold can be set to 0.75. That is, for each test scenario feature vector sample, at most 5 samples with the highest cosine similarity are retrieved from the prior test scenario feature vector samples that have reached the preset vector similarity. The prior evolution parameter combination corresponding to the above prior test scenario feature vector samples in the historical test scenario knowledge base is extracted as the prior parameter reference set for the test scenario feature vector sample.

[0104] Then, the probability function mapping model is initially trained using prior test scene feature vector samples and prior evolution parameter combinations to establish an initial probability function mapping model after meta-learning. Specifically, the obtained prior test scene feature vector samples and corresponding prior evolution parameter combinations constitute the initial training dataset. The probability function mapping model is initially trained using the prior test scene feature vector samples as input and the prior evolution parameter combinations as supervision signals. During the initial training, a Gaussian distribution is constructed with each prior evolution parameter combination as the mean and a preset standard deviation as the variance. This Gaussian distribution is used as the initial probability density distribution for the corresponding evolution parameter space region, ensuring that the initial shape of the probability distribution of the evolution parameter is tilted towards the historical high-quality parameter region before the formal Bayesian iterative optimization begins. The preset standard deviation is set according to the value range of each parameter in the evolution parameter space. For example, the preset standard deviation can be set to 10% of the value range of the corresponding parameter to ensure that the initial probability density distribution has a reasonable coverage range around the prior evolution parameter combinations. The initial training employs the same multi-objective joint loss as the Bayesian iterative optimization in S33. The number of training epochs is set to 10% to 20% of the preset iteration limit for formal Bayesian iterative optimization. This proportion has been verified to achieve a reasonable balance between enabling the model to establish initial mapping cognition and avoiding overfitting to prior data. If the number of initial training epochs is too small, the model cannot effectively absorb prior knowledge; if the number of initial training epochs is too large, the model may over-rely on prior evolutionary parameter combinations, weakening the exploration ability of subsequent Bayesian iterative optimization. After completing the initial training, a meta-learned initial probability function mapping model is obtained. Compared to the randomly initialized probability function mapping model, the initial shape of its evolutionary parameter probability distribution is tilted towards the historical high-quality parameter region. This allows it to converge to the high-quality parameter combination region more quickly in subsequent Bayesian iterative optimization, reducing the number of iterations and lowering the overall training cost.

[0105] By performing meta-learning initialization based on a knowledge base of historical test scenarios, the probability function mapping model has a preliminary understanding of the mapping rules between test scenarios and evolution parameters before formal training. This avoids the problems of slow convergence speed and large sample size required caused by training from a random initial state, and improves both training efficiency and parameter adaptation accuracy.

[0106] Furthermore, after establishing the initial probability function mapping model after meta-learning, the method also includes:

[0107] S323. Configure meta-learning guidance factors, and traverse the evolution parameter space to calculate the guidance value of each candidate evolution parameter combination based on the meta-learning guidance factors;

[0108] S324. Execute the candidate evolution parameter combination selected in the current iteration round according to the guiding value.

[0109] In a preferred embodiment, after establishing the initial probability function mapping model after meta-learning, in each iteration of Bayesian iterative optimization, the selection process of candidate evolution parameter combinations is guided by configuring a meta-learning guiding factor, thereby increasing the probability of selecting parameter regions that have historically performed well in similar test scenarios, and further accelerating the convergence speed of Bayesian iterative optimization.

[0110] First, a meta-learning guidance factor is configured, and the guidance value based on the meta-learning guidance factor is calculated for each candidate evolutionary parameter combination by traversing the evolutionary parameter space. The meta-learning guidance factor is a weight vector with the same dimensions as the evolutionary parameter space, used to quantify the prior quality of each region in the evolutionary parameter space. The weight value of each dimension in the meta-learning guidance factor reflects the frequency of occurrence of the corresponding evolutionary parameter when it was rated as a good solution in similar historical test scenarios. The higher the frequency of occurrence, the better the performance of the parameter value region in similar historical scenarios, and the larger the corresponding guidance factor weight. Specifically, the configuration process of the meta-learning guidance factor is as follows: For the feature vector samples of the current test scenario, the top K prior test scenario feature vector samples with the highest cosine similarity and their corresponding prior evolutionary parameter combinations are retrieved from the historical test scenario knowledge base. The value distribution of six types of evolutionary parameters—time step parameter, evolution time window, data smoothing factor, historical weight decay coefficient, spatial coupling coefficient, and data compression rate—in these K prior evolutionary parameter combinations is statistically analyzed, and the frequency of occurrence of prior evolutionary parameter combinations falling within each value interval for each type of evolutionary parameter is calculated. For example, in the combination of K=5 prior evolutionary parameters, four values ​​of the time step parameter fall within the range of 5 to 15 minutes, with a frequency of 0.8; three values ​​of the evolutionary time window fall within the range of 1 to 6 hours, with a frequency of 0.6; two values ​​of the data smoothing factor fall within the range of 0.1 to 0.3, with a frequency of 0.4; three values ​​of the historical weight decay coefficient fall within the range of 0.88 to 0.96, with a frequency of 0.6; and four values ​​of the spatial coupling coefficient are... The frequency of occurrence is 0.8, falling within the range of 0.6 to 0.9. The data compression rate has three values ​​falling within the range of 10% to 30%, with a frequency of 0.6. The sum of the frequencies of occurrence of the six evolutionary parameters is 3.8. Dividing the frequency of occurrence of each evolutionary parameter by 3.8 for normalization, the weight values ​​of each dimension of the meta-learning guidance factor are 0.211, 0.158, 0.105, 0.158, 0.211, and 0.158, respectively. After normalization, the sum of the weight values ​​of each dimension is 1.0.

[0111] After configuring the meta-learning guidance factor, all candidate evolutionary parameter combinations in the evolutionary parameter space are traversed, and the guidance value is calculated for each candidate evolutionary parameter combination. The guidance value of a candidate evolutionary parameter combination consists of two parts: the first part is the probability density value of the candidate evolutionary parameter combination under the current initial probability function mapping model, reflecting the possibility that the parameter combination is a good solution under the current model estimation; the second part is the weighted matching score of the parameter values ​​of each dimension of the candidate evolutionary parameter combination and the corresponding dimension weights of the meta-learning guidance factor. Specifically, the normalized value of each dimension parameter is obtained by subtracting the minimum value of the parameter's value range from the actual value of the parameter and then dividing by the value range. That is, the normalized value is equal to the actual value minus the minimum value divided by the difference between the maximum and minimum values. After normalization, the values ​​of each dimension parameter are mapped to the range of 0 to 1. For example, the time step parameter ranges from 5 to 30 minutes. If the time step parameter of a candidate evolution parameter combination is 10 minutes, its normalized value is (10-5) / (30-5) = 0.2. The data compression rate ranges from 10% to 40%. If the data compression rate of a candidate evolution parameter combination is 20%, its normalized value is (20-10) / (40-10) ≈ 0.333. The normalized values ​​of each dimension parameter are multiplied by the corresponding dimension weights of the meta-learning guidance factor and then summed to obtain the weighted matching score. A higher weighted matching score indicates a higher degree of agreement between the candidate evolution parameter combination and the historical high-quality parameter range. The probability density value and the weighted matching score are weighted and summed according to a preset ratio to obtain the final guidance value for each candidate evolution parameter combination. For example, the weight of the probability density value can be set to 0.6, and the weight of the weighted matching score can be set to 0.4, ensuring that the guidance value reflects both the current model estimation and historical prior knowledge.

[0112] Next, the candidate evolutionary parameter combinations selected in the current iteration are executed according to the guiding value. Specifically, all candidate evolutionary parameter combinations in the evolutionary parameter space are sorted from high to low according to the guiding value. In the current iteration, the candidate evolutionary parameter combination with the highest guiding value is selected first and imported into the evolutionary prediction model to perform evolutionary calculations. The corresponding multi-objective joint loss value is calculated, and the parameters of the probability function mapping model are updated based on the multi-objective joint loss value. By prioritizing the selection of candidate evolutionary parameter combinations with high guiding values ​​for iteration, Bayesian iterative optimization concentrates resources on exploring historically high-performing parameter regions when exploring the evolutionary parameter space, avoiding wasting iterations in low-quality parameter regions, thereby accelerating model convergence.

[0113] By configuring meta-learning guidance factors and selecting candidate evolutionary parameter combinations based on guidance values, historical prior knowledge is injected into the parameter exploration process in the form of quantified weights during Bayesian iterative optimization. This allows the parameter selection in each iteration to take into account both the probability estimation of the current model and the prior guidance of historical high-quality parameter regions. While ensuring the diversity of parameter exploration, this significantly improves the convergence speed of iterative optimization and reduces the number of iterations and computational overhead required to reach the convergence condition.

[0114] Furthermore, the method for outputting an adaptive evolution parameter combination based on the probability function mapping relationship includes:

[0115] S35. Identify the matching test scene clusters corresponding to the test scene feature vectors based on the probability function mapping relationship;

[0116] S36. Extract the probability distribution of the evolution parameters of the matching test scene cluster mapping;

[0117] S37. Perform parameter sampling according to the probability distribution of the evolution parameters to obtain an adaptive evolution parameter combination based on Pareto solution.

[0118] In a preferred embodiment, when outputting the adaptive evolution parameter combination, firstly, matching test scene clusters corresponding to the test scene feature vectors are identified based on the probability function mapping relationship. Specifically, cosine similarity is calculated between the test scene feature vector of the current target test scene and the cluster center vectors of each test scene cluster in the probability function mapping relationship, obtaining a similarity score between the current test scene feature vector and each test scene cluster. The test scene cluster with the highest similarity score is selected as the matching test scene cluster corresponding to the current test scene feature vector. For example, if there are three test scene clusters, and the cosine similarity calculated between the test scene feature vector of the current target test scene and the cluster center vectors of the three test scene clusters is 0.96, 0.72, and 0.45 respectively, then the first test scene cluster with the highest similarity score is selected as the matching test scene cluster.

[0119] Then, the evolutionary parameter probability distribution of the matching test scenario cluster mapping is extracted. Specifically, the unified evolutionary parameter probability distribution corresponding to the identified matching test scenario cluster is extracted from the probability function mapping relationship. This evolutionary parameter probability distribution is a conditional probability distribution obtained by weighted fusion of the evolutionary parameter probability distributions corresponding to all test scenario feature vector samples within the matching test scenario cluster. It describes the probability density distribution of each region in the evolutionary parameter space being covered by the optimal solution under the test scenario type represented by the current matching test scenario cluster. The higher the probability density of a region, the more likely the parameter combination in that region is to reach the optimal solution standard under the corresponding test scenario type.

[0120] Subsequently, parameter sampling is performed according to the evolution parameter probability distribution to obtain adaptive evolution parameter combinations based on Pareto optimization. Pareto optimization refers to finding the set of parameter combinations in a multi-objective optimization problem that are not inferior to other solutions across all optimization objectives; this is the Pareto optimal solution set. Each solution in the Pareto optimal solution set improves a particular objective at the expense of at least one other objective. Therefore, the Pareto optimal solution set represents the set of parameter combinations that achieve the optimal trade-off among the three optimization objectives: prediction error, data redundancy, and computational efficiency. Specifically, firstly, according to the extracted evolution parameter probability distribution, a Monte Carlo sampling method is used to generate several candidate evolution parameter combinations from the evolution parameter space. That is, the probability density values ​​of each region in the evolution parameter probability distribution are used as sampling weights. Candidate evolution parameter combinations are randomly generated at a higher frequency in high probability density regions and at a lower frequency in low probability density regions, thus concentrating the generated candidate evolution parameter combinations within historically high-quality parameter regions. The number of candidate evolution parameter combinations is set according to actual computational resources, for example, 50. For each candidate evolutionary parameter combination, its target evaluation values ​​are calculated in three dimensions: prediction error, data redundancy, and computational efficiency, resulting in a three-dimensional target evaluation vector for each candidate evolutionary parameter combination. Subsequently, Pareto dominance is determined for all candidate evolutionary parameter combinations: if candidate evolutionary parameter combination A's target evaluation values ​​in all three dimensions (prediction error, data redundancy, and computational efficiency) are not inferior to candidate evolutionary parameter combination B, and is superior to candidate evolutionary parameter combination B in at least one dimension, then A is said to dominate B. All candidate evolutionary parameter combinations not dominated by any other candidate evolutionary parameter combinations are selected to form the Pareto optimal solution set. For example, the three-dimensional objective evaluation vector of candidate evolutionary parameter combination A is [0.05, 0.12, 0.08], and the three-dimensional objective evaluation vector of candidate evolutionary parameter combination B is [0.07, 0.10, 0.09]. A is better than B in terms of prediction error and computational efficiency, while B is better than A in terms of data redundancy. They do not dominate each other and can both enter the Pareto optimal solution set. If the three-dimensional objective evaluation vector of candidate evolutionary parameter combination C is [0.08, 0.13, 0.10], then A is better than C in all three dimensions, A dominates C, and C does not enter the Pareto optimal solution set. Finally, from the Pareto optimal solution set, each candidate evolutionary parameter combination is weighted and comprehensively scored according to the same preset weight configuration as the multi-objective joint loss. That is, the objective evaluation values ​​of each dimension in the three-dimensional objective evaluation vector are multiplied by their corresponding preset weights and then summed to obtain the comprehensive score of each Pareto optimal solution. The candidate evolutionary parameter combination with the highest comprehensive score is selected as the adaptive evolutionary parameter combination.

[0121] By identifying matching test scenario clusters based on probability function mapping relationships, extracting the corresponding evolutionary parameter probability distributions, and combining Pareto solutions to output adaptive evolutionary parameter combinations, accurate adaptive configuration of evolutionary parameters under multi-objective optimization constraints is achieved. Monte Carlo sampling based on the probability distribution of matching test scenario clusters ensures that candidate evolutionary parameter combinations are generated within historically high-quality parameter regions, improving the overall quality of candidate solutions. Simultaneously, Pareto solutions ensure that the final output adaptive evolutionary parameter combinations achieve optimal trade-offs among prediction error, data redundancy, and computational efficiency, avoiding parameter bias problems caused by single-objective optimization and providing a high-quality parameter configuration foundation for subsequent evolutionary prediction model computation and execution.

[0122] Furthermore, the evolution prediction model includes at least one of the following: a traffic evolution model, an energy consumption evolution model, and an environmental evolution model.

[0123] In one feasible implementation, the evolutionary prediction model includes at least one of a traffic evolution model, an energy consumption evolution model, and an environmental evolution model. The traffic evolution model is used to perform time-series extrapolation and prediction of traffic operation indicators such as traffic flow, vehicle speed, and congestion status in the urban road network; the energy consumption evolution model is used to perform time-series extrapolation and prediction of energy consumption status in urban building clusters, industrial areas, and other regions; and the environmental evolution model is used to perform time-series extrapolation and prediction of environmental status indicators such as urban air pollutant concentration and noise levels. The three evolutionary prediction models correspond to different dimensions of urban operation status. One model can be used individually according to the specific needs of the target test scenario, or multiple evolutionary prediction models can be combined to simultaneously perform evolutionary calculations on multiple dimensions of urban operation status, generating a dynamic evolutionary time-series dataset containing multi-dimensional time-series prediction data to meet the diverse needs of different test scenarios for urban operation status data.

[0124] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0125] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0126] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0127] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0128] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0129] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0130] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for calculating urban information model data based on CIM, characterized in that, The method includes: Semantic parsing is performed on the description information of the target test scenario to construct a test scenario feature vector; For at least one evolutionary prediction model based on the CIM city information model, an evolutionary parameter space is constructed. The evolutionary parameter space includes at least the time step parameter, evolution time window, data smoothing factor, historical weight decay coefficient, spatial coupling coefficient, and data compression rate. The probability function mapping relationship between the feature vector of the test scenario and the evolution parameter space is trained with prediction error target, data redundancy target and computational efficiency target as optimization targets, and an adaptive evolution parameter combination is output based on the probability function mapping relationship; The adaptive evolution parameter combination is imported into the evolution prediction model to perform data evolution calculation, thereby obtaining a dynamic evolution time series dataset; The dynamically evolving time-series dataset is transmitted as test cases to the test system terminal containing the target test scenario description information to obtain the test performance results of the target under test.

2. The method as described in claim 1, characterized in that, Semantic parsing of the target test scenario description information to construct test scenario feature vectors includes the following methods: Extract the text description information, structured input information, and configuration parameter information of the target test scenario description information; The text description information, structured input information, and configuration parameter information are standardized to obtain the initial scene description information; The initial scene description information is segmented to extract key semantic elements, and the key semantic elements are processed with feature labels to obtain a set of semantic labels. Multidimensional feature quantization is performed on the semantic tag set to output the test scene feature vector.

3. The method as described in claim 2, characterized in that, The key semantic elements include time-related keywords, behavior-related keywords, precision-related keywords, and constraint-related keywords; The key semantic elements are processed by feature labeling to obtain a set of semantic labels that integrate time scale features, dynamic change intensity features, accuracy requirement features, spatial range features, and redundancy tolerance features.

4. The method as described in claim 1, characterized in that, The method for training the probability function mapping relationship between the test scenario feature vector and the evolution parameter space with prediction error target, data redundancy target, and computational efficiency target as optimization objectives includes: Collect historical test task data to construct a training sample set, where each training sample includes test scene feature vector samples, adaptive evolution parameter combination samples, and multiple target evaluation values ​​corresponding to the evolution calculation results of the corresponding combination samples; Meta-learning-based initialization probability function mapping model; With prediction error, data redundancy, and computational efficiency as optimization objectives, the initialization probability function mapping model performs Bayesian iterative optimization on the training sample set by introducing a multi-objective joint loss until the evolution parameter probability distribution corresponding to the feature vector sample of each test scenario is obtained, and the probability function mapping relationship is constructed according to the evolution parameter probability distribution.

5. The method as described in claim 4, characterized in that, The evolution parameter probability distribution is a conditional probability distribution generated based on the optimal solution of the evolution parameter combination under each test scenario feature vector sample; The optimal solution for the evolution parameter combination is the evolution parameter combination where the joint loss data of multiple objectives is less than a preset loss threshold.

6. The method as described in claim 4, characterized in that, The method further includes, until the probability distribution of evolutionary parameters corresponding to the feature vector samples of each test scenario is obtained: For each test scenario feature vector sample in the training sample set, the feature vector similarity is calculated to obtain the similarity calculation results; Cluster the test scene feature vector samples according to the similarity calculation results to obtain test scene clusters; Extract the evolution parameter probability distribution clusters corresponding to the test scenario clusters, fuse the evolution parameter probability distribution clusters, and update the evolution parameter probability distributions corresponding to the test scenario clusters; Re-establish the probability function mapping relationship between the test scenario cluster and the probability distribution of the evolution parameters.

7. The method as described in claim 4, characterized in that, The method for initializing probability function mapping models based on meta-learning includes: Connect to the historical test scenario knowledge base, retrieve prior test scenario feature vector samples that have reached a preset vector similarity with each test scenario feature vector sample from the historical test scenario knowledge base, and extract the prior evolution parameter combination of the prior test scenario feature vector samples; The probability function mapping model is initially trained using the prior test scenario feature vector samples and the prior evolution parameter combination to establish the initial probability function mapping model after meta-learning.

8. The method as described in claim 7, characterized in that, After establishing the initial probability function mapping model after meta-learning, the method also includes: Configure a meta-learning guidance factor, and traverse the evolution parameter space to calculate the guidance value of each candidate evolution parameter combination based on the meta-learning guidance factor; The candidate evolution parameter combination selected for the current iteration is executed according to the guiding value.

9. The method as described in claim 6, characterized in that, The method for outputting an adaptive evolution parameter combination based on the probability function mapping relationship includes: Based on the probability function mapping relationship, identify the matching test scenario clusters corresponding to the test scenario feature vectors; Extract the probability distribution of the evolution parameters of the matching test scene cluster mapping; Parameter sampling is performed according to the probability distribution of the evolution parameters to obtain an adaptive combination of evolution parameters based on Pareto solution.

10. The method as described in claim 1, characterized in that, The evolution prediction model includes at least one of the following: traffic evolution model, energy consumption evolution model, and environmental evolution model.

Images