Building group development intensity and municipal supporting facility capacity matching analysis method

Abstract

The invention provides a building group development intensity and municipal supporting facility capacity matching analysis method, which comprises the steps of multi-source data standardization cleaning, space-time alignment and feature extraction, development behavior trend prediction based on a time convolutional network, key time sequence enhancement in combination with an attention mechanism, and introduction of dynamic confidence gating shunting direct connection and correction flow. Carrying out buffer correction on low-confidence prediction by utilizing a generative adversarial network; an improved fuzzy evaluation model is fused, municipal facility capacity indexes are integrated, regional supply and demand pressure grading early warning is realized, online feedback learning and gating parameter adaptive adjustment capabilities are provided, prediction precision, stability and facility bearing risk pre-judgment effects are improved, and intelligent urban planning decision making is supported.

Description

Technical Field

[0001] This invention relates to the field of urban development supply and demand matching analysis and model robustness optimization technology, and in particular to a method for matching the development intensity of building complexes with the capacity of municipal supporting facilities. Background Technology

[0002] Currently, the supply-demand matching analysis of urban building development intensity and municipal infrastructure capacity is a cutting-edge research direction in the field of urban planning and intelligent assessment. With the acceleration of urbanization, various regions generally adopt a method combining pre-land transfer development demand estimation and infrastructure capacity assessment to guide urban spatial structure layout and municipal facility configuration. Most mainstream technical systems utilize multi-source heterogeneous data, including historical land transfer data, planning documents, and macroeconomic indicators, to quantitatively predict future development intensity through rolling forecasts or machine learning models. This data is then linked with municipal infrastructure capacity models to calculate the supply-demand matching degree or infrastructure carrying capacity pressure within the region, outputting tiered risk warnings and providing data support for planning, transfer, and construction decisions. Simultaneously, the industry is gradually exploring trend prediction analysis based on deep learning, automatic processing of multi-source data, and intelligent matching assessment methods, striving to improve the accuracy, usability, and forward-looking nature of the models.

[0003] However, in practical applications, existing technologies generally suffer from the following prominent problems: (1) Most of the development behavior prediction models and facility matching degree assessment modules are strongly coupled and integrated feedforward structures. The output of development intensity prediction is directly used for matching degree calculation, and there is a lack of flexible error isolation and compensation methods between models. Once the prediction part has local errors, abnormal fluctuations or inaccurate data extrapolation, the assessment results are easily amplified or distorted, the overall robustness of the model decreases, and it may even affect the final early warning decision; (2) Most current methods do not adequately address prediction uncertainty. Mainstream models rely solely on goodness of fit or a single confidence index to assess the reliability of the results, failing to establish an adaptive mechanism to dynamically adjust the weight of the predicted values ​​on the matching evaluation. This directly leads to unstable model outputs and increased error accumulation in situations where development activities are expected to fluctuate sharply, data distribution deviates from statistical assumptions, or uncertainty is high, thus weakening the system's foresight and practicality. (3) Although some technologies attempt to introduce feedback learning to correct prediction bias, they are often limited to periodic data backtracking, making it difficult to perceive real-time changes in model stability in a timely manner, and also unable to perform rapid and fine-grained error correction for low-confidence areas. The lack of diversion and hierarchical processing and local feature reconstruction capabilities makes it difficult for the model to take into account both overall trend grasp and local extreme value correction, resulting in significant shortcomings in highly dynamic and heterogeneous urban development scenarios. Summary of the Invention

[0004] In order to solve the above-mentioned technical problems, this invention provides a method for matching the development intensity of building complexes with the capacity of municipal supporting facilities.

[0005] The technical solution of this invention is implemented as follows: a method for matching the development intensity of building complexes with the capacity of municipal supporting facilities, comprising: S1: Collect historical land transfer data, regional development planning texts and macroeconomic indicators of urban areas, perform standardization cleaning and spatiotemporal alignment preprocessing on the multi-source heterogeneous data, and generate a basic dataset for development behavior prediction. S2: Based on the Temporal Convolutional Network (TCN) architecture, a development behavior trend prediction sub-model is constructed. The multi-source data preprocessed by S1 is input, and the spatial correlation features and temporal evolution laws are extracted through multi-scale dilated convolutional layers. The output is a probability distribution map of building increment in the next 3-5 years and the corresponding spatial unit confidence score matrix. S3: Construct a dynamic confidence gating unit. Based on the confidence score matrix generated in S2, use a piecewise linear function to establish the mapping relationship between the confidence threshold α and the flow weight, and generate a dynamic allocation strategy for the development intensity direct flow and correction flow. The range of α is determined according to the historical prediction backtesting deviation distribution. S4: Design a lightweight adversarial reconstruction buffer correction module, receive the low confidence correction stream data in S3, construct a training set using the historical development trajectory of similar plots in the region, reconstruct local features of the low confidence prediction through a generative adversarial network (GAN) architecture, and output the corrected development intensity estimate and its residual compensation coefficient. S5: Establish an improved fuzzy comprehensive evaluation matching degree fusion layer, weight and merge the outputs of S3 direct flow and S4 correction flow according to confidence weight, construct the supply and demand matching degree feature vector by combining the municipal facility capacity calculation results, calculate the facility bearing pressure index through fuzzy membership function, and generate graded early warning signals; S6: Deploy an online feedback learning mechanism to periodically collect actual development progress data and initial prediction deviation values, update the time attention weight entropy calculation parameters of the TCN model in S2, and optimize the dynamic adjustment rules of the confidence threshold α in S3 based on the deviation distribution characteristics to achieve adaptive calibration of the gating strategy parameters.

[0006] The method for matching the development intensity of building complexes with the capacity of municipal supporting facilities provided by this invention has the following beneficial effects: (1) This invention introduces a dynamic confidence gating mechanism to achieve differentiated processing and risk isolation of prediction information. It constructs a confidence scoring system based on the joint generation of attention weight entropy value and historical backtesting deviation, and sets a dynamic gating unit accordingly. High-confidence predictions are classified as "direct flow" and directly enter the supply and demand matching stage, while low-confidence predictions are transferred to the buffer correction module for local reconstruction and correction, effectively avoiding interference from unreliable prediction signals to the main evaluation process. This mechanism significantly improves the robustness and adaptability of feedforward evaluation technology without relying on complex integrated modeling, making the evaluation results more in line with actual development flexibility, and is especially suitable for scientific decision support in high uncertainty environments such as the early stage of land transfer. (2) This invention designs a lightweight adversarial reconstruction correction network, which achieves intelligent repair of abnormal development trends while maintaining the simplicity of the overall architecture. It uses the historical development trajectory of similar plots in the region as knowledge guidance, learns the spatial expression features of typical development patterns through a generator-discriminator structure, and performs semantically consistent local adjustments to low-confidence predictions to ensure that the corrected development intensity is both consistent with the regional development trend and interpretable. The corrected signal and the original high-confidence prediction are dynamically weighted and integrated in the fusion layer, and then fed into the improved fuzzy comprehensive evaluation model to complete the multi-dimensional load-bearing pressure classification assessment, forming a closed-loop optimization of the entire chain from "trend identification - confidence judgment - deviation correction - fusion evaluation". The whole process does not require increasing the number of parameters in the backbone model, but significantly improves the stability and spatial sensitivity of the evaluation output, which is significantly better than the traditional static weighting or unified correction strategy. (3) This invention, through the design of a feedback learning mechanism, can continuously absorb deviation data between the actual development progress and the initial prediction, driving the online iterative optimization of the gating threshold and correction network parameters, thereby constructing a feedforward evaluation closed-loop system with adaptive evolution capabilities. Compared with the traditional one-time training and fixed parameter deployment mode, this scheme, through periodic injection of real observation feedback, enables the key control parameters of the model to dynamically adjust with changes in the external environment, effectively responding to structural shifts caused by new policies, new plans, or emergencies in urban development. This mechanism not only enhances the reliability of the method's long-term operation but also reduces the cost of manual intervention and frequent retraining, possessing good engineering application value. Attached Figure Description

[0007] Figure 1 This is a flowchart of the method for matching the development intensity of building complexes with the capacity of municipal supporting facilities according to the present invention. Figure 2 This is a sub-flowchart of the method for matching the development intensity of building complexes with the capacity of municipal supporting facilities according to the present invention. Figure 3 This is another sub-flowchart of the method for matching the development intensity of building complexes with the capacity of municipal supporting facilities in this invention. Detailed Implementation

[0008] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0009] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0010] like Figure 1 As shown, this invention provides a method for matching the development intensity of building complexes with the capacity of municipal supporting facilities, specifically including: S1: Collect historical land transfer data, regional development planning texts and macroeconomic indicators of urban areas, perform standardization cleaning and spatiotemporal alignment preprocessing on the multi-source heterogeneous data, and generate a basic dataset for development behavior prediction. S2: Based on the Temporal Convolutional Network (TCN) architecture, a development behavior trend prediction sub-model is constructed. The multi-source data preprocessed by S1 is input, and the spatial correlation features and temporal evolution laws are extracted through multi-scale dilated convolutional layers. The output is a probability distribution map of building increment in the next 3-5 years and the corresponding spatial unit confidence score matrix. S3: Construct a dynamic confidence gating unit. Based on the confidence score matrix generated in S2, use a piecewise linear function to establish the mapping relationship between the confidence threshold α and the flow weight, and generate a dynamic allocation strategy for the development intensity direct flow and correction flow. The range of α is determined according to the historical prediction backtesting deviation distribution. S4: Design a lightweight adversarial reconstruction buffer correction module, receive the low confidence correction stream data in S3, construct a training set using the historical development trajectory of similar plots in the region, reconstruct local features of the low confidence prediction through a generative adversarial network (GAN) architecture, and output the corrected development intensity estimate and its residual compensation coefficient. S5: Establish an improved fuzzy comprehensive evaluation matching degree fusion layer, weight and merge the outputs of S3 direct flow and S4 correction flow according to confidence weight, construct the supply and demand matching degree feature vector by combining the municipal facility capacity calculation results, calculate the facility bearing pressure index through fuzzy membership function, and generate graded early warning signals; S6: Deploy an online feedback learning mechanism to periodically collect actual development progress data and initial prediction deviation values, update the time attention weight entropy calculation parameters of the TCN model in S2, and optimize the dynamic adjustment rules of the confidence threshold α in S3 based on the deviation distribution characteristics to achieve adaptive calibration of the gating strategy parameters.

[0011] Step S1: Collect historical land transfer data, regional development planning documents, and macroeconomic indicators for the urban area; perform standardization cleaning and spatiotemporal alignment preprocessing on the multi-source heterogeneous data to generate a basic dataset for development behavior prediction. Specifically, this includes: S1.1: Based on the database and public data platform of the urban planning management department, obtain historical land transfer data within the urban area. The historical land transfer data includes the plot number, transfer year, land use nature, building area, transferor and transferee information, so as to form a historical record table of plot-level development behavior. Input criteria include structured and semi-structured land transfer records from the city planning and management department's internal database and multiple publicly available data platforms; By using a heterogeneous data interface calling method (parameters: API key, query field set, pagination strategy), basic information such as land parcel number, land grant year, land use nature, building area, and transferor and transferee can be retrieved in batches from the planning management department's database. Furthermore, by using batch crawling and structured parsing methods (parameters: HTML parsing rules, regular expression matching pattern set), the land transfer announcements and transaction lists provided by the public data platform are automatically crawled and their fields mapped to obtain a transaction record dataset that is consistent with the fields in the departmental database. Furthermore, a duplicate detection algorithm based on primary key association (parameters: land parcel number primary key, time field consistency threshold) is adopted to identify and remove duplicate records across data sources and generate a unique land parcel transaction index table; Furthermore, by using field value validation methods (parameters: building area range filtering threshold, land use nature enumeration table), abnormal or illegal values ​​in the collected data are identified and marked, resulting in land transfer record data that has undergone preliminary screening. By using a multi-source data merging process, the cleaned records from different sources are merged by external join according to the plot number and time field to form a plot-level development behavior history table that covers the entire region and is time-series continuous, thus achieving high-quality basic data output that can be called by subsequent prediction models. For example, during the construction process of a municipality, an API key is set to call the urban planning management department's database. The selected query fields include plot number, year of land grant, land use, building area, and information on the transferor and transferee. The pagination strategy is set to 1000 records per batch. Transaction records from the public data platform tables are extracted using HTML parsing rules, and regular expression matching is used for date field parsing. In the primary key association duplicate detection algorithm, plot numbers are consistent and the time field difference does not exceed a certain threshold. Records for days are considered duplicates and are removed. During field value validation, records with a building area lower than [a certain value] are excluded. square meters or higher Records of square meters were marked as abnormal, and the land use was limited to three categories: residential, commercial, and industrial. The final record was consolidated to form a table covering all land transfer events in the city from 2010 to 2023, containing approximately... A unique transaction record satisfies the spatiotemporal integrity and reliability requirements of the subsequent temporal convolutional network prediction process; S1.2: The regional development planning text is extracted in a structured manner. Natural language processing technology is used to identify and semantically analyze the development goals, land use functional zoning, plot ratio restrictions and infrastructure layout descriptions in the planning text, and generate a structured planning element table as a text input feature for predicting development behavior. For regional development planning text data, a deep semantic parsing algorithm (parameters: BERT-based text encoder, word segmentation granularity set to double-byte Chinese character unit) is used to realize the vectorized representation of the original planning documents and ensure that the encoding results can capture syntactic dependencies and contextual logic; Furthermore, by using a keyword recognition method (parameters: TF-IDF score threshold set to 0.15, dictionary covering four categories of terms: development goals, land use functions, plot ratio, and infrastructure), significant planning elements in the text are accurately extracted, and a preliminary keyword index table is obtained. Furthermore, by using a named entity recognition algorithm (parameter: BiLSTM-CRF, the label system includes functional areas, transportation facilities, water and electricity facilities, etc.), phrases involving specific geographical locations or infrastructure types are identified, and a list of entities with category labels is generated. Furthermore, through dependency parsing (parameters: Arc-standard transfer system, syntactic relation weights set according to the planning dictionary standard), the structural relationships between keywords are analyzed, and a semantic dependency graph is formed to reflect the logical constraints between planning elements; Furthermore, through feature fusion processing, the keyword index table, entity list, and semantic dependency graph are jointly encoded into structured planning element table row records. The fields include element category, element value or description, constraints, and location of occurrence, thereby achieving standardized and structured output of planning text features. By using deep semantic parsing and multi-stage natural language processing, the unstructured information of regional development planning texts is transformed into a structured planning element table containing core attributes such as development goals, land use functional zoning, plot ratio restrictions, and infrastructure layout, thereby enabling unified access of multi-source heterogeneous data in subsequent prediction models. For example, using the development plan text data of a new urban area, a BERT-based Chinese semantic encoder was employed, with a maximum sequence length of 512, resulting in sentence vectors of 768 dimensions. TF-IDF scoring extracted 48 keywords, including "comprehensive commercial area," "residential land," "plot ratio 2.5," and "Metro Line 3." Low-relevance words were removed at a scoring threshold of 0.15. The BiLSTM-CRF model identified named entities such as "riverside park" (functional zoning category), "water supply network (ring)" (water facility category), and "110 kV substation" (power facility category), achieving a significant improvement in label accuracy. Dependency parsing revealed a modification relationship between "comprehensive commercial area—plot ratio 2.5" and a connection relationship between "Metro Line 3—central square." The final fusion yielded approximately 100 records in the structured planning element table. Each record contained category fields (such as functional zoning and transportation facilities), quantitative fields (such as area and plot ratio), constraint fields (such as height limit and impermeability limit), and location information fields (latitude and longitude coordinates or plot number). This table was used for the spatiotemporal alignment processing of S1.5 and input into the S2 prediction model, which significantly improved the spatial semantic sensitivity and element analysis depth of the model in predicting the future development trend of the new area. S1.3: Collect macroeconomic indicator data, including but not limited to regional GDP growth rate, total fixed asset investment, real estate price index, population migration trend and industrial layout change data, and perform standardization and normalization processing on the macroeconomic indicators to obtain a time-series input feature vector reflecting the overall development trend of the city. In the process of collecting and processing macroeconomic indicator data, the input object is a set of economic operation statistics covering the target city area, including multi-dimensional economic characteristics such as regional GDP growth rate, total fixed asset investment, real estate price index, population migration trend and industrial layout changes. This data comes from the National Statistical Database, local statistical annual reports and public data interfaces of authoritative industry research institutions. A multi-source data aggregation method (parameters: data source type, sampling period, spatial region coding) is adopted to achieve unified extraction and field mapping of economic indicators from different sources, ensuring the consistency of each economic sub-item in semantics and statistical definition; Furthermore, through a time series resampling algorithm (parameter: time step size) t, spline interpolation method), to align the sampling frequencies of different macroeconomic indicators and obtain the original sequence set of economic indicators with standardized time index; Furthermore, a standardization method (parameters: mean μ, standard deviation σ) is adopted, and each economic indicator series is standardized item by item using the following formula:

[0012] in, These are the original index values. This is the average value of the indicator over the entire period. This is the full-cycle standard deviation of the indicator, and this processing eliminates dimensional differences. Furthermore, a normalization transformation method is adopted (parameter: minimum value). Maximum value The standardized index is mapped to the [0,1] interval using the following formula:

[0013] in, and These are the minimum and maximum values ​​of the indicator over the entire period, respectively. Furthermore, using the feature vector construction method (parameters: indicator dimension n, time length T), all normalized economic indicators are arranged in time order to form a time-series input feature vector matrix with T rows and n columns, which can be used by the subsequent development behavior trend prediction sub-model. Through the above algorithm processing method, the original macroeconomic data in the previous step is transformed into a multi-dimensional economic indicator feature matrix under a unified time series benchmark and after standardization and normalization, so as to realize high-quality input features that reflect the overall development trend of the city. For example, in a scenario where economic indicators for a certain region are collected, the data sources include the National Bureau of Statistics' annual GDP report, regional fixed asset investment monitoring data, monthly real estate index reports, population migration registration data, and dynamic data on industrial park layout. The sampling period is configured as quarterly. For a time series period of 3 months, cubic spline interpolation was used for resampling to align all economic indicators into a quarterly series. During standardization, the mean μ of the regional GDP growth rate was set to 4.5, and the standard deviation σ was set to 1.2. Substituting these values ​​into the standardization formula, the standardized quarterly result for a GDP growth rate of 6.0 was... =1.25. The minimum net increase in population migration during the normalization process. -2000 people, maximum value With a population of 5000, when the net increase is 3000, substituting into the normalization formula, the result is... =0.714. The final constructed quarterly economic indicator feature matrix contains 20 quarters and 5 economic indicator dimensions. After alignment, standardization and normalization, the values ​​of each column of the matrix are in the range of [0,1], and the differences in the dimensions of different economic indicators are eliminated, which provides a guarantee for the stable input of the subsequent model. S1.4: Perform data cleaning operations on the three types of data obtained in S1.1 to S1.3, including missing value imputation, outlier removal, duplicate record deduplication and field format unification, and use sliding window interpolation to imput missing parts in time series data to improve data integrity and consistency, and generate cleaned land parcel data; S1.5: Based on the Geographic Information System (GIS) platform, spatial coordinate alignment is performed on the cleaned land parcel data. The timestamp field is used to perform time-series alignment of historical land transfer data, planning text parsing results and macroeconomic indicators to generate a multi-source heterogeneous dataset with a unified spatiotemporal benchmark, which serves as input data for the development behavior trend prediction sub-model. The cleaned land parcel data is input into the GIS platform, and a spatial coordinate standardization algorithm is used (parameters: coordinate system set to WGS84, spatial resolution set to 1 meter) to achieve coordinate consistency processing of data from various sources under a unified geographic reference system; Furthermore, by using a spatial topology verification method (parameter: tolerance radius set to 0.5 meters), the precise matching between the plot boundary and the facility layout vector map is achieved, and the plot boundary vector set after spatial coordinate correction is obtained; Furthermore, a timestamp normalization algorithm is adopted (parameters: unified time format is YYYY-MM-DD, time zone is set to GMT+8) to unify the time field format of historical land transfer data, planning text parsing results and macroeconomic indicators, and generate a comparable time series index table; Furthermore, by using a spatiotemporal joint alignment algorithm (parameters: time matching window is ±15 days, spatial matching radius is 50 meters), the accurate correspondence between multi-source data under the same spatiotemporal benchmark is realized, and a set of successfully matched spatiotemporal joint feature records is generated. Furthermore, a missing spatiotemporal unit interpolation algorithm (parameters: linear interpolation in the temporal direction and kriging interpolation in the spatial direction) is adopted to fill in the spatiotemporal units without matching records and generate a complete multi-source heterogeneous spatiotemporal feature matrix. By using the above-mentioned GIS spatial coordinate alignment and timestamp normalization processing methods, the cleaning results of the previous step are transformed into a multi-source heterogeneous dataset with a unified spatiotemporal benchmark, achieving seamless integration of multi-source data and high consistency of subsequent model input. For example, in the early stage of development prediction for a new urban area, land parcel data cleaned using S1.4 (containing coordinates of 500 parcels, all represented in the China Geodetic 2000 coordinate system) was input. The GIS platform coordinate system transformation parameters were set to WGS84, with a spatial resolution of 1 meter. After coordinate standardization, the spatial deviation between all parcel boundaries and the municipal pipeline map was reduced to within 0.3 meters. Using a spatial topology verification method with a tolerance radius of 0.5 meters, polygon closure errors in the boundaries of 5 parcels were identified and corrected. In timestamp normalization, the time fields of all historical land transfer records were converted to YYYY-MM-DD format and unified to the GMT+8 time zone. After processing, macroeconomic indicators and planning text element tables could be directly compared in time series. Using a spatiotemporal joint alignment algorithm, with a time matching window set to ±15 days and a spatial matching radius set to 50 meters, the success rate reached over 95%. For the 5% of missing spatiotemporal units, linear interpolation and kriging interpolation were used to fill in the gaps, resulting in 500 spatial units, each containing a complete record of historical land transfer information, planning elements, and economic indicators. In this example, the output multi-source heterogeneous spatiotemporal feature matrix directly serves as the basis for constructing the input tensor of the temporal convolutional network in step S2, significantly improving the spatial correlation and temporal continuity of the prediction sub-model.

[0014] Step S2: Construct a development behavior trend prediction sub-model based on the Temporal Convolutional Network (TCN) architecture. Input the multi-source data preprocessed in S1, extract spatial correlation features and temporal evolution patterns through multi-scale dilated convolutional layers, and output a probability distribution map of building increments for the next 3-5 years and the corresponding spatial unit confidence score matrix. Specifically, this includes: S2.1: Based on the basic dataset for predicting development behavior generated in S1, a sub-model architecture for predicting development behavior trends is constructed using a Temporal Convolutional Network (TCN). The input layer receives standardized multi-source heterogeneous data, including historical land transfer records, regional development planning text feature vectors, and macroeconomic indicator sequences, to construct a spatiotemporal joint feature input tensor. The input conditions are a multi-source heterogeneous dataset preprocessed in step S1, including the plot attribute field matrix of historical land transfer records, the feature vector of regional development planning text, and the time series feature vector of macroeconomic indicators. All data have been aligned with the unified benchmark of spatial coordinates and timestamps on the GIS platform. The Temporal Convolutional Network (TCN) model architecture design method is adopted (parameters: the number of input channels corresponds to the sum of the dimensions of multi-source features, and the time window length is set to 3 to 5 years according to the prediction period) to realize the joint input modeling of the above multi-source spatiotemporal data. The historical land transfer record matrix, planning text feature vector and macroeconomic indicator feature vector are matrix concatenated in the input layer to generate a unified spatiotemporal feature input tensor. Furthermore, by using the input tensor channel grouping embedding method (parameter: channel grouping mapping rule is defined based on data source type), different types of features are isolated and encoded in the input layer, so that the TCN convolution kernel can extract the spatial correlation features and temporal evolution rules within the source in subsequent multi-scale dilated convolutions. Furthermore, by using the input layer normalization method (parameter: mean-variance standardization is adopted, and the mean and variance are calculated based on the past 5 years of sample statistics of each data source), the numerical range of multi-source data is unified, the interference of differences in the magnitude of different features on temporal convolution feature extraction is reduced, and a normalized spatiotemporal feature input tensor is generated. Furthermore, through temporal padding and masking mechanisms (parameter: mask weight matrix is ​​generated based on missing data markers in S1), in-model compensation for missing parts of the input tensor is achieved, so that convolution operations do not produce feature leakage at sequence boundaries and missing positions; Using the above method, multi-source heterogeneous and standardized data are mapped into high-dimensional spatiotemporal feature tensors that meet the input requirements of TCN convolution, thereby achieving stable feature extraction results in subsequent multi-scale dilated convolutional layers. For example, in a development behavior prediction task in a central urban area of ​​a city, the input dataset consists of three types of features processed in step S1: a historical land transfer record matrix with a dimension of 200×8, a planning text feature vector with a dimension of 200×20, and a macroeconomic indicator feature vector with a dimension of 200×5. After GIS alignment, these are merged into a 200×33 multi-source feature input tensor. The TCN input layer has a time window length of 5 years, corresponding to a time step of 60 months, and 33 input channels. Channel grouping embedding is used, with 8 channels for land transfer records, 20 channels for planning text, and 5 channels for macroeconomic data, ensuring that the convolutional kernel extracts features independently within each group. During normalization, the mean of the building area field in land transfer records was set to 15000, and the variance to 2500; the mean of the plot ratio field in planning text was set to 2.5, and the variance to 0.5; and the mean of the GDP growth rate in macroeconomic indicators was set to 6.2, and the variance to 1.1. After mean-variance standardization at the input layer, the variance of the data within the feature dimensions tended to be consistent, significantly improving the response balance of the convolution kernel in different channels. The missing data mask matrix accounted for 2.3% of the dataset, and the mask weights were set to zero at the missing marker positions and one at the remaining positions. Through the above processing, the high-dimensional spatiotemporal feature tensor generated at the TCN input end showed stable spatial correlation patterns and temporal evolution capture capabilities in the subsequent multi-scale dilated convolution extraction in S2.2. The boundary effect of the predicted output spatial unit building increment probability distribution map was significantly weakened, and the volatility of the confidence score was reduced, achieving robustness and consistency in the data input stage. S2.2: Perform multi-scale dilated convolution operation on the input tensor. By setting convolution kernel stacking layers with different dilation coefficients, extract the multi-scale evolution law of development behavior in the time dimension and the neighborhood association features between spatial units to obtain a high-dimensional spatiotemporal feature embedding representation. S2.3: Perform time-attention mechanism weighted fusion processing on the feature representation output of the multi-scale dilated convolutional layer. Calculate the attention weight matrix based on the similarity between feature vectors at each time step to enhance the contribution of key time nodes to the prediction of development trends and obtain attention-enhanced temporal feature representation. S2.4: Based on attention-enhanced temporal feature representation, a fully connected network is used to generate the probability distribution of building increment in each spatial unit over the next 3 to 5 years. The output layer uses the Softmax activation function to normalize and model the building increment in each year to obtain the probability distribution matrix of the future annual development intensity of each spatial unit and its corresponding temporal attention weight matrix. S2.5: Based on the entropy value of the time attention weight matrix and the historical backtesting deviation data, construct a spatial unit confidence scoring model, calculate the confidence score of the development intensity prediction results of each spatial unit, where the higher the entropy value, the greater the uncertainty of the model's attention to the corresponding spatial unit, and thus the lower the confidence score. Finally, output a development behavior prediction result matrix containing confidence information. The input consists of the probability distribution matrix of building increments for each spatial unit over the next 3 to 5 years and its corresponding time attention weight matrix, which is the output of step S2.4. It also includes the prediction bias dataset for each spatial unit, which is compiled from the historical prediction backtesting process. An information theory-based entropy calculation method (parameter: the row vector of the attention weight matrix represents the contribution of time step features) is adopted to quantify the attention uncertainty of each spatial unit in the time series prediction process and output the spatial unit entropy matrix. Furthermore, through statistical analysis methods (parameters: root mean square error (RMSE) and skewness coefficient of historical prediction backtesting deviation data), the robustness of the prediction results of each spatial unit is evaluated, and the prediction stability score matrix is ​​obtained. Furthermore, a weighted fusion algorithm (parameters: entropy matrix and stability score matrix, with weights set based on historical verification correlation) is used to jointly evaluate entropy and prediction error, generating a comprehensive uncertainty index matrix. Furthermore, a linear normalization method (parameters: minimum and maximum values ​​of the comprehensive uncertainty index) is used to achieve a reverse mapping of confidence scores under a unified scale, resulting in a preliminary confidence score matrix, in which spatial units with high entropy and large prediction errors are mapped to lower scores. The confidence score adjustment algorithm transforms the score matrix from the previous step into a development behavior prediction result matrix containing confidence information, thereby providing a directly usable score input for the S3 dynamic confidence gating unit. For example, this step is performed on the development behavior trend prediction results of a certain urban area containing 200 spatial units. Let the element in the i-th row of the attention weight matrix represent the weight distribution of the i-th spatial unit at different time steps, and use the entropy calculation formula:

[0015] in This represents the sum of the weights of the spatial unit. Let be the attention weight for the i-th spatial unit at time step t. The entropy matrix of each spatial unit and the historical backtesting bias RMSE matrix are then processed using a weighted fusion formula:

[0016] in For spatial unit i, the comprehensive uncertainty index is... To integrate weights, Let i be the entropy value of spatial unit i. Let be the prediction error value for spatial cell i. The values ​​are linearly normalized to obtain confidence scores on a uniform scale. In this real-world scenario, Set it to 0.6. The normalization results show that some areas with scarce infrastructure and frequent planning adjustments scored significantly lower than the overall average. It was verified that these areas were included in the correction stream in the S3 gating unit, which effectively avoided the impact of low-confidence prediction on the matching evaluation results.

[0017] like Figure 2 As shown, step S3 involves constructing a dynamic confidence gating unit. Based on the confidence score matrix generated in S2, a piecewise linear function is used to establish a mapping relationship between the confidence threshold α and the flow weight, generating a dynamic allocation strategy for the development intensity direct flow and correction flow. The value range of α is determined based on the historical prediction backtesting deviation distribution. Specifically, this includes: S3.1: Normalize the spatial unit confidence score matrix output by the development behavior trend prediction sub-model to eliminate the influence of uneven distribution of prediction confidence in different regions and obtain a standardized confidence matrix as the input basis for dynamic gating decision-making. S3.2: Based on historical prediction backtesting deviation distribution data, statistical analysis methods are used to determine the initial range of the confidence threshold α, and confidence level intervals are divided in combination with the characteristics of regional development stages to generate a multi-segment linear function mapping rule table for subsequent dynamic calculation of diversion weights; A correlation analysis was performed on the standardized confidence matrix and the historical prediction backtesting deviation distribution data. The frequency statistics method (parameters: prediction period Δt = 1 year, number of intervals m = 10) was used to calculate the frequency distribution of each deviation interval in order to establish a quantitative relationship between the deviation value and the prediction stability. The biased distribution fitting method (parameter: the distribution model is a mixture of normal and log-normal distributions) is adopted. The mixed distribution parameters are solved by maximum likelihood estimation, and the lower and upper limits of the confidence intervals for each interval are calculated as the initial constraint range for the confidence threshold α. Furthermore, by using a regional development stage feature annotation method (parameters: stage type is planning period, construction period, maturity period), the deviation distribution data are grouped according to stage characteristics, and variance analysis is performed within each group to determine the differences in confidence distribution corresponding to different development stages, so as to form a stage-based confidence level interval mapping table. Using a piecewise linear function construction method, the mapping relationship between the confidence score c and the splitting weight w is established through the following formula:

[0018] in, For standardized confidence scoring, This is the confidence threshold applied at the current stage. For piecewise slope parameters, Based on the basic weight intercept, adjusted using segmented conditions. The value of is determined to achieve dynamic changes in the weights of the direct flow and the correction flow; Based on the constructed piecewise linear function mapping results, a complete multi-piece linear function rule table is generated. The rule table contains the slope parameter, intercept parameter and threshold α range corresponding to each stage feature, providing structured input for subsequent dynamic calculation of the split weight. Through the above statistical analysis and piecewise linear function construction, the historical prediction error distribution is transformed into parameterized confidence thresholds and diversion weight rules that can be used in the gating unit, thereby realizing an adaptive allocation strategy for prediction data under different stability conditions. For example, using historical prediction backtesting deviation distribution data of land transfer and development behavior in a certain city over the past 10 years, when performing frequency statistical analysis, a prediction period Δt = 1 year and the number of interval divisions m = 10 were selected. The results showed that the deviation occurred most frequently in the range [0.02, 0.05], corresponding to a confidence score range of [0.85, 0.92]. When fitting a mixed distribution model, the mean μ of the normal component was 0.04, the variance σ² was 0.0004, and the mean of the log-normal component was... = 3.2 Variance With σ'² = 0.05, the best-fit combination was selected using the AIC criterion, and the initial confidence threshold range was calculated to be [0.80, 0.94]. Considering the characteristics of regional development stages, the confidence level ranges for the planning, construction, and maturity stages were set to [0.82, 0.94], [0.80, 0.90], and [0.88, 0.96], respectively. In the construction of the piecewise linear function, a slope was set for the planning stage. =0.5, intercept =0.2, slope during construction period =0.4, intercept =0.25, Maturity Slope =0.6, intercept =0.15. After inputting the rule table into the gating unit, during actual operation, the proportion of high-confidence direct flow in the predicted data at different stages was significantly improved, the proportion of low-confidence correction flow was optimized, and the prediction stability of the system was significantly improved in all three stages. S3.3: Based on the piecewise linear function mapping rule table generated in S3.2, perform confidence-splitting weight mapping calculation on each spatial unit in the standardized confidence matrix to generate a splitting weight matrix in the spatial dimension, which is used to distinguish between high-confidence and low-confidence prediction regions. For each spatial cell of the standardized confidence matrix, a piecewise linear function mapping algorithm (parameters: confidence level interval, slope coefficient set) is used to convert the normalized confidence value of the cell into a split weight coefficient. Furthermore, by using interval judgment and linear interpolation calculation (parameters: segment start and end values, slope), the weight of the input confidence value within the corresponding level interval is continuously estimated, and the weight output matrix for different confidence intervals is obtained. Furthermore, a matrix-based batch processing calculation method (parameters: total number of spatial units, batch size) is adopted to realize the synchronous mapping and processing of the diversion weights of the confidence values ​​of all spatial units, and to generate a two-dimensional spatial diversion weight matrix; Furthermore, by using a normalized weight adjustment algorithm (parameters: constraints on maximum and minimum weights), the numerical range of the weight matrix is ​​compressed and boundary constraints are applied to ensure that the subsequent ratio of direct flow to correction flow meets the system stability requirements. Furthermore, by utilizing the weight matrix structure analysis method (parameter: threshold for weight difference between adjacent spatial units), automatic partitioning and labeling of high-confidence and low-confidence prediction regions are achieved, and a spatial index mapping table that can be called within the gated unit is formed; By using the above-mentioned piecewise linear function mapping and weight matrix generation process, the standardized confidence matrix of the previous step is transformed into quantitative data of spatial diversion weights, which enables accurate differentiation of high and low confidence prediction regions and provides a basis for dynamic diversion operations. For example, in an urban area assessment scenario containing 200 spatial units, the standardized confidence matrix ranges from 0.2 to 0.95. The piecewise linear function mapping rule table sets three intervals: 0.2–0.5, 0.5–0.8, and 0.8–1.0, with corresponding slopes of 1.2, 0.8, and 0.5, respectively. For the confidence value c, the formula for calculating the diversion weight w is:

[0019] in , The weights are determined based on the interval mapping rule table. During batch processing, each batch contains 50 spatial units, and the matrix operation outputs a 200×1 two-dimensional splitting weight matrix. The weight range is normalized to 0.1 to 0.9 to avoid extreme values ​​affecting the allocation ratio. A weight difference threshold of 0.25 is used for spatial adjacency analysis, identifying 72 regions with confidence levels above 0.8 for the direct flow, 41 regions with confidence levels below 0.5 for the corrective flow, and the remaining 87 regions in the intermediate range. In subsequent dynamic splitting operations, this weight matrix serves as input to directly allocate direct and corrective flows, significantly improving the gating unit's precise control over the predicted data stream and reducing volatility. S3.4: Based on the splitting weight matrix, perform dynamic splitting operation on the development intensity prediction value, and assign the prediction value with confidence above the threshold α to the direct flow channel, and assign the prediction value with confidence below the threshold α to the correction flow channel, forming two development intensity output streams with different processing paths; Based on the spatial dimension diversion weight matrix generated by S3.3, a dynamic threshold comparison and channel allocation algorithm (parameters: confidence threshold α, diversion weight matrix, prediction development intensity matrix) is used to determine the diversion channel for the predicted value of each spatial unit. Furthermore, by using a matrix element-wise comparison method (parameters: prediction unit confidence score, threshold α), the predicted values ​​are logically divided, and a Boolean mask matrix is ​​generated to identify the locations of high-confidence units and low-confidence units. Furthermore, based on the Boolean mask matrix, matrix element selection operation (parameters: prediction development strength matrix, through-flow mask, correction flow mask) is used to extract the through-flow of high-confidence prediction values ​​and the correction flow of low-confidence prediction values, forming two independent prediction matrices; Furthermore, the proportional adjustment process is performed using the splitting weight matrix (parameters: direct flow matrix, correction flow matrix, corresponding splitting weight) to achieve numerical weighting correction of the two prediction matrices, so as to ensure the weight consistency of the subsequent fusion layer; Furthermore, a data channel identifier implantation method (parameters: direct flow matrix, correction flow matrix, channel identifier) ​​is adopted to append channel metadata to the splitting results, providing structured support for the subsequent buffer correction module to accurately identify low-confidence prediction data; By using dynamic threshold comparison and matrix splitting algorithm, the standardized confidence matrix and the predicted development strength matrix are transformed into two types of data structures: direct flow and correction flow, so as to achieve the accurate allocation effect of predicted data on different processing paths. For example, in a regional prediction scenario, the dynamic confidence threshold α is set to 0.72, the standardized confidence matrix ranges from [0.35, 0.96], and the predicted development intensity matrix values ​​range from [0, 1] (normalized intensity values). Using an element-wise matrix comparison method, 48 spatial cells in the generated direct flow mask matrix satisfy a confidence level ≥ 0.72, and their corresponding prediction intensities are directly included in the direct flow matrix. The remaining 22 spatial cells with prediction intensities below the threshold are assigned to the correction flow matrix. The splitting weight matrix takes values ​​in the high-confidence region (0.85, 0.95) and the low-confidence region (0.45, 0.65). After proportional adjustment, the high-confidence direct flow matrix values ​​are increased to an average of 1.08 times the original value, while the low-confidence correction flow matrix values ​​are reduced to an average of 0.92 times the original value. The additional channel identifiers "TC" and "CC" are embedded in the metadata tags of the two matrices, respectively, and form structured data packets in the output stage. The total data volume of the pass-through stream data packets is approximately the set of predicted development intensities for high-confidence regions, while the total data volume of the correction stream data packets is approximately the set of predicted development intensities for low-confidence regions. This splitting result enables accurate screening of low-confidence prediction data and direct transmission of high-confidence data when the subsequent buffer correction module is loaded, significantly improving the coupling stability of prediction and evaluation. S3.5: Perform channel identification marking processing on the output of the split-through stream and the correction stream, and generate development intensity prediction data packets with channel attribute tags to provide a structured output with traceable split-through path for the subsequent buffer correction module and matching degree fusion layer.

[0020] like Figure 3 As shown, step S4 involves designing a lightweight adversarial reconstruction buffer correction module. This module receives the low-confidence correction stream data from S3, constructs a training set using the historical development trajectories of similar plots within the region, reconstructs local features of the low-confidence prediction using a generative adversarial network (GAN) architecture, and outputs the corrected development intensity estimate and its residual compensation coefficient. Specifically, this includes: S4.1: Based on the historical development trajectory data of similar plots in the region, a training set of low-confidence prediction samples is constructed for training the Generative Adversarial Network (GAN) model, thereby providing the buffer correction module with sample input that has spatial consistency and temporal continuity. The input conditions are the spatial unit development intensity prediction data after being diverted by the dynamic confidence gating unit in step S3 and identified as low confidence correction stream, as well as the historical land development trajectory data of these spatial units that have geographical proximity and functional attribute similarity. A spatial similarity clustering algorithm (parameters: spatial location coordinates, land use attribute code, plot ratio range) is used to divide the target spatial unit into similar groups with historical land parcels, and obtain the land parcel index set within each group; Furthermore, by using a time series pattern matching algorithm (parameters: time window length m, delay alignment threshold δ), the phase alignment of historical land development trajectories in the time dimension is achieved, and a time-aligned trajectory data matrix is ​​generated; Furthermore, a normalization processing algorithm (parameter: minimum-maximum normalization interval [0,1]) is adopted to achieve scale unification of development intensity values ​​in the trajectory data matrix and obtain unitized spatiotemporal trajectory samples; Furthermore, by using data smoothing and denoising algorithms (parameters: second-order Savitzky-Golay filter window width w, order p), short-term fluctuations of spatiotemporal trajectory samples are suppressed, and a stable spatiotemporal trajectory feature matrix is ​​obtained. A sample labeling and stratified sampling algorithm (parameters: confidence level label, sample ratio r) is used to achieve a balanced distribution of samples with different confidence levels in the training set and generate a low-confidence prediction training set data package containing input features and confidence labels. By constructing the training set as described above, the low-confidence prediction and historical trajectory of similar plots from the previous step are transformed into GAN training inputs with dual constraints of spatial consistency and temporal continuity, thereby achieving the sample base construction technology effect of the buffer correction module. For example, in a low-confidence prediction spatial unit set of a city, the coordinate precision of each unit is set to 0.001 degrees. A spatial similarity clustering algorithm divides the data into 8 similar groups, each containing no fewer than 50 historical plot indices. The window length m of the time series pattern matching algorithm is set to 36 months, and the delay alignment threshold δ is set to 3 months, resulting in a 50×36 dimension trajectory matrix after alignment. The normalization processing interval is [0,1], where the original building area data ranges from [2000,10000] square meters before normalization, and after normalization, it is all mapped to the [0,1] interval. Short-term fluctuation suppression uses a second-order Savitzky-Golay filter with a window width w set to 5 and an order p set to 2. The filtered trajectory curve is smooth but maintains the long-term trend. The stratified sampling ratio r is set to 0.6. Training sets are extracted from samples with high and low confidence levels respectively, resulting in a training data package containing 400 trajectory samples, each with a confidence label. After being used by the subsequent GAN generator, this data package can significantly improve the spatial rationality and temporal trend consistency when outputting the corrected candidate map. S4.2: Design a generator model based on a generative adversarial network architecture. Input the low-confidence development intensity prediction data output from S3, and use a convolutional neural network to extract local development pattern features to generate a development intensity correction candidate map with regional consistency, so as to improve the spatial rationality of the prediction results. S4.3: Construct a discriminator model and conduct adversarial training on the development intensity correction candidate map generated in S4.2 and real historical development data. By minimizing the discriminator's recognition accuracy of the generated samples, optimize the spatial feature distribution of the generator output to enhance the matching degree between the generated results and actual development behavior. S4.4: Based on the residual between the development intensity correction candidate map generated in S4.2 and the original low-confidence prediction, calculate the residual compensation coefficient to quantify the uncertainty level of the low-confidence prediction and provide an error correction basis for the subsequent matching degree fusion layer; Based on the development intensity correction candidate map generated by S4.2 and the original low-confidence prediction data, a point-by-point difference calculation method is used (parameter: spatial resolution). Time resolution This enables the quantitative measurement of the deviation from the predicted value; Furthermore, by using the mean square error calculation method (parameter: total number of samples N), the mean square value of the residuals at each time step for the same spatial unit is estimated, and the residual energy index is obtained. Furthermore, by using the residual standard deviation calculation method (parameter: total number of samples N), the degree of dispersion of the residuals is quantified, and a prediction stability index is generated; Furthermore, the residual compensation coefficient is achieved through a weighted fusion method of normalized residual energy and normalized residual standard deviation (the weighting coefficient β is determined by model sensitivity analysis). The generation of [the data] yields a unified index for quantifying the level of uncertainty in low-confidence predictions;

[0021] in, This is the normalized value of the residual energy. This represents the normalized value of the residual standard deviation. These are the weighting coefficients; By using the above residual compensation coefficient calculation method, the difference between the corrected candidate map and the original low-confidence prediction value in the previous step is transformed into an error correction basis that can be used for the matching degree fusion layer, thereby realizing the explicit quantification of the uncertainty of low-confidence prediction. For example, in a new urban area planning scenario, the input spatial resolution... Set to 100 meters, time resolution Set to a one-year period, the total sample size N is 500 spatial units multiplied by 5 prediction years. For each spatial unit, the difference between the revised candidate value and the original low-confidence prediction value is calculated year by year, yielding an average residual ē of 0.12, a residual energy E of 0.018, and a residual standard deviation σ of 0.095. After normalization... It is 0.36. The value is 0.28. Sensitivity analysis determined the weighting coefficient β to be 0.65. Substituting this into the formula, the compensation coefficient C = 0.65 × 0.36 + (1 (0.65)×0.28=0.336. This coefficient is used in the subsequent matching degree fusion process to correct the predictive impact of low confidence areas, significantly improving the stability and robustness of facility load pressure early warning; S4.5: The corrected development intensity generated in S4.2 is fused with the residual compensation coefficient calculated in S4.4 to output a standardized corrected development intensity estimate, which is then passed to the matching degree fusion layer to support the assessment of facility carrying capacity pressure under the dynamic confidence weighted merging mechanism.

[0022] Step S5: Establish an improved fuzzy comprehensive evaluation matching degree fusion layer, weighted and merged the outputs of S3 direct flow and S4 correction flow according to confidence weights, construct a supply-demand matching degree feature vector by combining the municipal facility capacity calculation results, calculate the facility carrying capacity pressure index through fuzzy membership function, and generate a graded early warning signal. Specifically, this includes: S5.1: The confidence level of the development intensity direct flow output by S3 and the correction flow output by S4 is normalized. A weighted fusion coefficient matrix is ​​constructed based on the confidence score of each spatial unit to realize the dynamic weighted merging of the two development intensity estimates and output the fused development intensity feature map. S5.2: Based on the water supply, power supply, drainage and traffic carrying capacity data provided by the municipal facilities capacity calculation module, construct a standardized municipal facilities capacity feature vector, and perform spatial alignment and attribute mapping with the development intensity feature map output by S5.1 to form a regional supply and demand matching feature vector set; The input conditions are the regional water supply, power supply, drainage and traffic carrying capacity data provided by the municipal facilities capacity calculation module, and the fused development intensity feature map output by S5.1 is received as another set of inputs for spatial matching; A standardized mapping method (parameters: facility type coding table, capacity unit conversion factor) is adopted to achieve unit normalization processing of water supply, power supply, drainage and traffic carrying capacity data, forming an initial capacity vector set with the same dimensions; Furthermore, by using a vector normalization algorithm (parameters: maximum and minimum capacity values), a linear mapping of the capacity values ​​of each facility in the [0,1] interval is achieved, and a standardized municipal facility capacity feature vector is obtained, ensuring that different facility types are comparable in the matching degree calculation; Furthermore, a spatial coordinate alignment algorithm (parameters: GIS geographic reference coordinate system, spatial unit division rules) is adopted to realize the spatial location mapping between the standardized capacity feature vector and the development intensity feature map, and to generate a correspondence matrix of spatial unit numbers; Furthermore, by using the attribute mapping function (parameters: facility type attribute code, spatial unit index set), the attribute binding of capacity characteristics and development intensity characteristics of each spatial unit is realized, and spatial unit-level supply and demand matching degree input entries are generated. The supply-demand ratio for each spatial unit is calculated using the supply-demand matching degree combination function and the following formula:

[0023] in, To standardize the values ​​of the facility capacity feature vector in the corresponding spatial unit, This represents the development intensity characteristic value of this spatial unit; Through the above calculation process, the results of the previous step are transformed into a feature vector set of regional supply and demand matching degree, so as to achieve precise integration of facility capacity and development intensity in the spatial dimension. For example, in the analysis scenario of a new urban area, the water supply capacity is 50,000 cubic meters per day, which is converted to 0.82 after unit normalization; the power supply capacity is 800 megawatts per day, which is normalized to 0.78; the drainage capacity is 40,000 cubic meters per day, which is normalized to 0.79; and the traffic capacity is 100,000 person-times per day, which is normalized to 0.85. In the development intensity characteristic map, the development intensity characteristic value of a spatial unit numbered A-12 is 0.64. After GIS coordinate alignment and attribute mapping binding, the supply-demand ratio of this unit is calculated: The result, 1.28125, indicates that the water supply capacity is significantly more than the predicted development intensity of this unit. Within the same spatial unit, the transportation capacity supply-demand ratio is... The result was 1.328125, indicating sufficient traffic capacity. These supply-demand matching ratios are then fed into the subsequent fuzzy comprehensive evaluation module, which can significantly improve the accuracy of the load-bearing pressure index calculation for different facility types and the reliability of regional hierarchical early warning. S5.3: An improved fuzzy comprehensive evaluation model is adopted. The supply and demand matching degree feature vector set constructed in S5.2 is fuzzified. A multi-level fuzzy membership function is set based on the facility type. The fuzzy membership matrix of the bearing pressure of each facility type is calculated, and the fuzzy evaluation set of the facility bearing status is output. For the supply and demand matching feature vector set constructed by S5.2, an improved fuzzy comprehensive evaluation model (parameters: facility type classification rules, membership function parameter set) is adopted to realize the fuzzification of the carrying pressure of different facility types; Furthermore, by using a facility type identification algorithm (parameters: coded label sets for water supply, power supply, drainage, and transportation), each dimension in the supply-demand matching feature vector is mapped to a specific facility type, thereby realizing the type mapping of the feature vector and providing type constraint data for subsequent fuzzification processing; Furthermore, a multi-level fuzzy membership function set (parameters: level division threshold, function shape coefficient) is used to perform membership degree calculation on the matching degree features of each facility type, and a membership value matrix divided by facility type is obtained; Furthermore, through the membership function The calculation formula derives the membership degree of the matching degree index for each facility type:

[0024] in, For the characteristic value of supply and demand matching degree, The membership function center value, This represents the maximum possible deviation in the matching degree of this facility type; Furthermore, based on the membership results calculated by the above formula, a fuzzy membership matrix of bearing pressure arranged by facility type is constructed, and the fuzzy evaluation vector set of bearing status of each spatial unit under different facility types is stored in a matrix structure. By using a fuzzy comprehensive evaluation model, the membership matrix of the previous step is transformed into a fuzzy evaluation set of facility carrying capacity status, thereby realizing the fuzzy technical indicators required for subsequent classification and early warning. For example, in the analysis scenario of a new urban area, the regional supply and demand matching feature vector output by S5.2 includes four facility types: water supply feature value 0.78, power supply feature value 0.65, drainage feature value 0.82, and transportation feature value 0.55. The facility type identification algorithm maps these four values ​​to the corresponding facility labels. For water supply facilities, a fuzzy membership function center value is set. =0.8, maximum deviation =0.3, substituting into the formula to calculate the membership degree is =0.933. For power supply facilities, set... =0.7、 =0.4, membership degree is calculated as follows =0.875. The membership values ​​for drainage and transportation facilities were calculated sequentially, and the results were stored in matrix form: the first row shows the membership degree for water supply (0.933), the second row for power supply (0.875), the third row for drainage (0.933), and the fourth row for transportation (0.875). This matrix was used as the fuzzy evaluation set input to the subsequent classification and determination model. Verification showed that it significantly improved the accuracy and foresight of the classification of facility load-bearing pressure states in different spatial units. S5.4: Based on the principle of maximum membership in the fuzzy comprehensive evaluation results, and combined with the preset facility load-bearing pressure level classification standard, the municipal facility load-bearing status of each spatial unit is classified and judged, and a load-bearing pressure level label matrix is ​​generated as the qualitative evaluation result of the regional facility matching degree. S5.5: Jointly analyze the bearing pressure level label matrix output from S5.4 with the confidence score of spatial unit development intensity prediction, dynamically correct the warning level of low confidence areas based on the confidence weighted grading rules, generate the final facility bearing pressure grading warning signal map, and output it to the visualization display module.

[0025] Step S6: Deploy an online feedback learning mechanism, periodically collect actual development progress data and initial prediction deviation values, update the time attention weight entropy calculation parameters of the TCN model in S2, and optimize the dynamic adjustment rule of the confidence threshold α in S3 based on the deviation distribution characteristics to achieve adaptive calibration of the gating strategy parameters. Specifically, this includes: S6.1: Based on the actual development progress data provided by the urban planning and management department, the data is standardized, cleaned, and timestamp aligned to obtain a dataset of actual development behavior trajectories that can be used for model feedback; S6.2: Compare the actual development behavior trajectory data obtained in S6.1 with the historical prediction output of the development behavior trend prediction sub-model in S2 on a cycle-by-cycle basis, calculate the development intensity prediction deviation value sequence of each spatial unit within the prediction cycle, and form a prediction error time series database. S6.3: Based on the prediction error time series database generated in S6.2, the standard deviation and skewness coefficient of the prediction error in each period are calculated using the sliding window statistical method, and a dynamic distribution model of prediction uncertainty is constructed to quantify the stability characteristics of the prediction model in different regions and time scales. S6.4: Based on the prediction uncertainty dynamic distribution model constructed in S6.3, optimize the calculation parameters of the temporal attention weight entropy value of the temporal convolutional network (TCN) in S2. Specifically, perform weighted decay processing on the attention weight entropy value in the high uncertainty period to improve the sensitivity of the model to historical trend changes and obtain the updated temporal attention weight parameter set. S6.5: Based on the dynamic distribution model of prediction uncertainty built in S6.3, a nonlinear regression method is used to fit the mapping function relationship between the prediction error distribution and the confidence threshold α, generate an adaptive adjustment rule for the dynamic confidence threshold α, and inject the rule into the dynamic confidence gating unit in S3 to realize the real-time calibration of the gating diversion strategy. S6.6: The updated dynamic confidence threshold α adjustment rule in S6.5 is jointly analyzed with the historical correction residual data of the buffer correction module in S4 to construct a feedback closed-loop control model, which is used to dynamically adjust the training loss function weight of the generative adversarial network (GAN) in the correction stream, thereby improving the reconstruction accuracy and stability of low-confidence predictions.

[0026] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0027] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for analyzing the matching of building complex development intensity and municipal infrastructure capacity, characterized in that, Includes the following steps: S1: Collect multi-source heterogeneous data of urban areas, perform standardization cleaning and spatiotemporal alignment preprocessing on the multi-source heterogeneous data, and generate a basic dataset for predicting development behavior. S2: Construct a development behavior trend prediction sub-model based on the temporal convolutional network architecture, input the basic dataset for development behavior prediction, extract spatial correlation features and temporal evolution laws through multi-scale dilated convolutional layers, and output the building increment probability distribution map and the corresponding spatial unit confidence score matrix. S3: Construct a dynamic confidence gating unit. Based on the confidence score matrix of the spatial unit, establish the mapping relationship between the confidence threshold and the diversion weight using a piecewise linear function to generate a dynamic allocation strategy for the development intensity direct flow and the correction flow. S4: Design an adversarial reconstruction buffer correction module to receive low-confidence correction stream data, construct a training set using the historical development trajectory of similar plots in the region, reconstruct local features of the low-confidence prediction through a generative adversarial network architecture, and output the corrected development intensity estimate and its residual compensation coefficient. S5: Establish an improved fuzzy comprehensive evaluation matching degree fusion layer, weight and merge the development intensity direct flow and the corrected development intensity estimate according to the confidence degree weight, construct the supply and demand matching degree feature vector by combining the municipal facility capacity calculation results, calculate the facility bearing pressure index through the fuzzy membership function, and generate a graded early warning signal.

2. The method for matching the development intensity of building complexes with the capacity of municipal supporting facilities according to claim 1, characterized in that, Following step S5, the following is also included: S6: Deploy an online feedback learning mechanism to periodically collect actual development progress data and initial prediction deviation values, update the time attention weight entropy value calculation parameters of the development behavior trend prediction sub-model, and optimize the dynamic adjustment rules of the confidence threshold based on the deviation distribution characteristics.

3. The method for matching the development intensity of building complexes with the capacity of municipal supporting facilities according to claim 1, characterized in that, Step S1 specifically includes: Based on the database and public data platform of the urban planning and management department, historical land transfer data within the urban area is obtained to form a historical record table of land parcel-level development behavior; The regional development plan text is extracted in a structured manner to generate a structured planning element table; Collect macroeconomic indicator data, perform standardization and normalization processing on the macroeconomic indicator data, and obtain a time-series input feature vector reflecting the overall development trend of the city. The historical record table of the land parcel-level development behavior, the structured planning element table, and the time-series input feature vector reflecting the overall development trend of the city are subjected to data cleaning operations to obtain cleaned land parcel data. Based on the geographic information system platform, the cleaned land parcel data is spatially aligned, and the historical land transfer data, planning text parsing results and macroeconomic indicators in the cleaned land parcel data are time-series aligned using the timestamp field to generate a multi-source heterogeneous dataset with a unified spatiotemporal benchmark.

4. The method for matching the development intensity of building complexes with the capacity of municipal supporting facilities according to claim 3, characterized in that, The data cleaning operations include missing value imputation, outlier removal, duplicate record deduplication and field format standardization, and the use of sliding window interpolation to imput missing parts in the time series data.

5. The method for matching the development intensity of building complexes with the capacity of municipal supporting facilities according to claim 1, characterized in that, Step S2 specifically includes: Based on the generated basic dataset for predicting development behavior, a sub-model architecture for predicting development behavior trends is constructed using a temporal convolutional network, and a spatiotemporal joint feature input tensor is built. Multi-scale dilated convolution operation is performed on the spatiotemporal joint feature input tensor. By setting convolution kernel stacking layers with different dilation coefficients, the multi-scale evolution law of development behavior in the time dimension and the neighborhood correlation features between spatial units are extracted to obtain a high-dimensional spatiotemporal feature embedding representation. The high-dimensional spatiotemporal feature embedding representation is subjected to weighted fusion processing using a temporal attention mechanism. An attention weight matrix is ​​calculated based on the similarity between feature vectors at each time step to obtain an attention-enhanced temporal feature representation. Based on the attention-enhanced temporal feature representation, a fully connected network is used to generate the probability distribution of building increments for each spatial unit in the future, thereby obtaining the probability distribution matrix of the future annual development intensity of each spatial unit and its corresponding temporal attention weight matrix. Based on the entropy value of the time attention weight matrix and the historical backtesting deviation data, a spatial unit confidence scoring model is constructed to calculate the confidence score of the development intensity prediction results for each spatial unit and output a development behavior prediction result matrix containing confidence information.

6. The method for matching the development intensity of building complexes with the capacity of municipal supporting facilities according to claim 1, characterized in that, Step S3 specifically includes: The spatial unit confidence score matrix output by the development behavior trend prediction sub-model is normalized to obtain a standardized confidence matrix; Based on historical prediction backtesting deviation distribution data, statistical analysis methods are used to determine the initial range of confidence threshold values, and confidence level intervals are divided in combination with the characteristics of regional development stages to generate a piecewise linear function mapping rule table. Based on the piecewise linear function mapping rule table, perform confidence-splitting weight mapping calculation on each spatial unit in the standardized confidence matrix to generate a splitting weight matrix in the spatial dimension; Based on the aforementioned splitting weight matrix, a dynamic splitting operation is performed on the predicted development intensity value to form two development intensity output streams with different processing paths; The output of the split-through stream and the corrected stream are processed by channel identification marking to generate development intensity prediction data packets with channel attribute tags.

7. The method for matching the development intensity of building complexes with the capacity of municipal supporting facilities according to claim 6, characterized in that, Step S3 further includes: the dynamic confidence gating unit relies on the confidence score of the spatial unit, sets a confidence threshold based on the historical prediction backtesting deviation and development stage characteristics, calculates the diversion weight for each spatial unit through a piecewise linear function, assigns the prediction value above the threshold to the direct flow, assigns the prediction value below the threshold to the correction flow, and adds channel attribute labels to each.

8. The method for matching the development intensity of building complexes with the capacity of municipal supporting facilities according to claim 1, characterized in that, Step S4 specifically includes: A training set of low-confidence prediction samples is constructed based on historical development trajectory data of similar plots within the region. A generator model is designed based on a generative adversarial network architecture. Low-confidence development intensity prediction data is input, and local development pattern features are extracted using a convolutional neural network to generate a development intensity correction candidate map with regional consistency. A discriminator model is constructed, and adversarial training is performed on the development intensity correction candidate map and real historical development data. By minimizing the discriminator's recognition accuracy of the generated samples, the spatial feature distribution of the generator output is optimized. Based on the residual between the development intensity correction candidate map and the original low-confidence prediction, the residual compensation coefficient is calculated; The development intensity correction candidate map is fused with the residual compensation coefficient to output a standardized corrected development intensity estimate, which is then passed to the matching degree fusion layer.

9. The method for matching the development intensity of building complexes with the capacity of municipal supporting facilities according to claim 1, characterized in that, Step S5 specifically includes: The confidence level of the output development intensity direct flow and the output correction flow are normalized. A weighted fusion coefficient matrix is ​​constructed based on the confidence score of each spatial unit, and the fused development intensity feature map is output. Based on the water supply, power supply, drainage and traffic carrying capacity data provided by the municipal facilities capacity calculation module, a standardized municipal facilities capacity feature vector is constructed, and spatial alignment and attribute mapping are performed with the development intensity feature map to form a regional supply and demand matching feature vector set. An improved fuzzy comprehensive evaluation model is adopted to perform fuzzification processing on the feature vector set of regional supply and demand matching degree. Based on the facility type, a multi-level fuzzy membership function is set to calculate the fuzzy membership matrix of the carrying pressure of each facility type and output the fuzzy evaluation set of facility carrying status. Based on the principle of maximum membership in the fuzzy comprehensive evaluation results, and combined with the preset facility load-bearing pressure level classification standard, the municipal facility load-bearing status of each spatial unit is classified and judged, and a load-bearing pressure level label matrix is ​​generated. The load-bearing pressure level label matrix is ​​jointly analyzed with the spatial unit development intensity prediction confidence score. Based on the confidence-weighted grading rules, the early warning level of low confidence areas is dynamically corrected to generate the final facility load-bearing pressure grading early warning signal map, which is then output to the visualization module.

10. The method for matching the development intensity of building complexes with the capacity of municipal supporting facilities according to claim 1, characterized in that, The online feedback learning mechanism periodically collects and cleans the actual development progress, calculates the prediction error time series, models the uncertainty distribution using the sliding window method, dynamically adjusts the attention weight entropy parameters and confidence threshold adaptive adjustment rules of the temporal convolutional network, and feeds back the correction residuals to the buffer correction module to optimize the training loss function weights.

Images