Urban ventilation environment simulation prediction method and system based on different planning targets

Abstract

The invention discloses an urban ventilation environment simulation and prediction method and system based on different planning objectives, and particularly relates to the field of urban ventilation environment simulation and prediction, and the method comprises the steps: obtaining ventilation characteristic data, land utilization data and regulation information data for many years as data sources; unifying the data sources to the same coordinate system and spatial resolution, and performing rasterization and numerical normalization processing; constructing a land utilization prediction model based on the PLUS based on the land utilization data and the regulation information data; and rasterizing the driving factors of the land utilization change. Regulation information is introduced in a land utilization prediction process, a dynamic land utilization conversion matrix is constructed, and a space-time diagram neural network and an entropy weight method are combined for calculation, so that joint modeling and quantitative evaluation of an urban land utilization pattern and a ventilation environment under a multi-level management and control condition are realized; the problem of systematic deviation between a simulation result and actual execution in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of urban ventilation environment simulation and prediction technology, and more specifically, to urban ventilation environment simulation and prediction methods and systems based on different planning objectives. Background Technology

[0002] In current urban ventilation simulation studies, land use conversion matrices are typically used to predict future land patterns by defining which types of land can be converted into each other and which must remain unchanged. This method is relatively clear in theory and can meet the prediction needs in many conventional scenarios. However, when this method is applied to complex scenarios where urban renewal and new town expansion occur simultaneously, problems gradually emerge.

[0003] In real cities, many plots of land are not simply classified as land use types; they are often subject to multiple restrictions. For example, there may be underground utility tunnels or subway tunnels, or there may be protected red lines for historical districts on the surface, or there may be high-voltage power corridors and ecological green spaces that need to be strictly preserved. These restrictions do not appear directly in regular land use data and are difficult to describe with simple permission or prohibition. As a result, when the model simulates, it will calculate an ideal land use pattern according to a predetermined matrix, which may seem to optimize the ventilation environment. However, when planners implement it, they find that many plots that the model judges to be convertible are actually not suitable for development or renewal.

[0004] This discrepancy not only leads to a serious disconnect between model results and real-world planning, but may also bring greater risks; areas that most need to maintain ventilation may be easily replaced in the simulation map by other plots that do not meet the requirements, ultimately rendering the planning meaningless.

[0005] This leads to the conclusion that the existing land use conversion matrix lacks the ability to express the multi-level control conditions in cities, and cannot accurately reflect them, resulting in a systematic deviation between the simulation results and the actual implementation. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for simulating and predicting urban ventilation environment based on different planning objectives. By introducing regulatory information into the land use prediction process, constructing a dynamic land use transformation matrix, and combining it with a spatiotemporal graph neural network and entropy weight method for calculation, the method achieves joint modeling and quantitative evaluation of urban land use patterns and ventilation environment under multi-level control conditions, thereby solving the problem of systematic deviation between simulation results and actual implementation proposed in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for simulating and predicting urban ventilation environment based on different planning objectives, comprising:

[0008] S1. Obtain multi-year ventilation characteristic data, land use data, and regulatory information data as data sources; unify the data sources to the same coordinate system and spatial resolution and perform rasterization and numerical normalization processing;

[0009] S2. Construct a PLUS-based land use prediction model based on the land use data and regulatory information data; rasterize the driving factors of land use change; set conversion restriction markers for the raster to be converted according to the regulatory information data, generate a land use conversion matrix, and complete model training and verification.

[0010] S3. Set total land use constraints based on different planning objectives, input the total land use constraints and the land use transformation matrix into the land use prediction model, and generate the land use spatial pattern for the target year.

[0011] S4. Construct a spatiotemporal graph neural network prediction model based on multi-year ventilation characteristic data and land use spatial pattern of the target year; use grid cells as graph nodes and construct adjacency relationships based on geographical adjacency; learn spatial dependence through graph convolution operation and learn temporal dependence through recurrent cells; complete model training and output the prediction results of urban ventilation environment characteristics of the target year.

[0012] S5. The predicted results of the urban ventilation environment characteristics are weighted using the entropy weight method. Based on the spatial heterogeneity statistical weight coefficient of each ventilation environment characteristic, the comprehensive ventilation environment index is calculated and the spatial distribution results of ventilation environment changes under different planning objectives are output.

[0013] In a preferred embodiment, the ventilation characteristic data includes surface temperature gradient, surface roughness, and forest canopy density; the land use data includes construction land, arable land, forest, grassland, water bodies, and unused land; and the regulatory information data includes the distribution of underground municipal facilities, legal protection boundaries, and ventilation corridor control boundaries.

[0014] The driving factors of land use change include elevation, slope, distance from highways, distance from main roads, distance from secondary roads, population density, GDP per capita, and building density.

[0015] In a preferred embodiment, in S1, multi-year ventilation characteristic data, land use data, and regulatory information data are unified to the same coordinate system and the same spatial resolution.

[0016] Spatial rasterization is performed on the unified data to transform each data source into a numerical matrix in the form of fixed raster cells, forming a raster dataset.

[0017] Linear normalization is performed on the rasterized dataset: the original value of the index of each data type in the raster cell is used as input, the upper limit and lower limit of the index of that type in all cells are set, and the normalization function is used to calculate the processed value, wherein the normalization function is: the processed value is equal to the original value of the raster cell minus the lower limit value, and then divided by the difference between the upper limit value and the lower limit value.

[0018] The normalized numerical matrix maintains the original positive and negative directions of change and outputs a unified standardized data input set.

[0019] In a preferred embodiment, in S2, based on historical land use data and rasterized data of driving factors, conversion samples of different land use types are statistically analyzed, and the probability values ​​of conversion of various types of rasterized data to the target type are calculated to form a development potential map.

[0020] Perform neighborhood effect calculation on candidate grids, count the number of target types within the neighborhood window and combine it with preset neighborhood weights to form a neighborhood effect coefficient;

[0021] The development potential map is combined with the neighborhood effect coefficient, and conversion restriction markers are set for restricted rasters based on regulatory information data to generate a dynamic land use conversion matrix.

[0022] Under the condition of satisfying the total land use constraint, the dynamic land use transformation matrix is ​​input into the cellular automaton iterative process to solve the land use spatial pattern of the target year.

[0023] The land use spatial pattern of the target year is compared with the corresponding historical land use data, and consistency indicators are statistically analyzed to verify the land use prediction model.

[0024] In a preferred embodiment, S3 includes:

[0025] S3-1. The land use prediction model includes defining the state set of grid cells, neighborhood window rules, conversion probability calculation methods and iterative update mechanisms, and taking the development potential map, neighborhood effect coefficient and dynamic land use conversion matrix as input variables to form the prediction model framework.

[0026] S3-2. Based on different planning objectives, set upper and lower limits for the area of ​​construction land, cultivated land, forest, grassland, water bodies and unused land for the target year as constraints, and store the constraints as total land use control parameters.

[0027] S3-3. During the model iteration process, the comprehensive conversion probability of each candidate grid is retrieved one by one. The total land use control parameters are combined with the dynamic land use conversion matrix. Land use types that have reached the area limit are stopped from being newly converted, while land use types that have not reached the area limit are updated first.

[0028] S3-4. After all land use types meet the total constraints and complete iterative convergence, output the land use spatial pattern of the target year controlled by the constraints, and use it as the input for subsequent ventilation environment characteristic prediction.

[0029] In a preferred embodiment, S4 includes:

[0030] S4-1. Align the multi-year ventilation characteristic data with the land use spatial pattern of the target year into a time-ordered sequence input set, with each grid cell corresponding to a graph node. Encode the land use type of the target year as a node attribute and then concatenate it with the historical ventilation environment characteristics in the node dimension to form a node feature sequence for joint spatial and temporal modeling.

[0031] S4-2. Construct an adjacency matrix based on the geographic adjacency relationship of grid cells, determine the connection relationship between nodes using the eight-neighbor rule, write self-loop connections on the main diagonal of the adjacency matrix to maintain the transmission of node features, and perform normalization processing on the adjacency matrix to obtain a normalized adjacency matrix.

[0032] S4-3. Perform graph convolution calculation on the normalized adjacency matrix and node feature sequence. For all nodes at each time step, aggregate the features of neighboring nodes according to the adjacency relationship and complete the linear and nonlinear mappings. Solve the node spatial dependency representation of the corresponding time step and use the node spatial dependency representation as the input representation for time modeling.

[0033] In a preferred embodiment, S4 further includes:

[0034] S4-4. Input the spatial dependency representation of nodes arranged in chronological order into the recurrent unit. Construct a sliding time window sample in the recurrent unit according to a fixed time step. Use the preceding time step in the window as input and the ventilation environment features of the adjacent subsequent time steps as supervision signals. Minimize the mean square error loss to update the model parameters until the training process converges and the spatiotemporal joint modeling is completed.

[0035] S4-5. After training, the multi-year ventilation characteristic data and the land use spatial pattern of the target year are used as inference inputs. Graph convolution calculation and cyclic unit calculation are performed in sequence, and the prediction results of urban ventilation environment characteristics of the target year are output for subsequent weighted calculation of ventilation environment index.

[0036] In a preferred embodiment, S5 includes:

[0037] S5-1. Extract each type of data from the prediction results of the ventilation characteristic data of the target year one by one, and perform linear normalization operation within all grid cells, using the minimum value as the lower limit and the maximum value as the upper limit, to obtain a unified normalized feature matrix.

[0038] S5-2. In the normalized feature matrix, for the j-th type of urban ventilation environment feature, the normalized value of each grid cell i is divided by the sum of the normalized values ​​of the feature in all grid cells to obtain the probability value p(i,j), and the probability distribution matrix is ​​formed by the probability values ​​of all grid cells.

[0039] S5-3. Based on the probability distribution matrix, calculate the information entropy value E(j) of each type of urban ventilation environment characteristic, and obtain the information utility value d(j) by subtracting the information entropy value from 1, forming an information utility value matrix.

[0040] S5-4. Normalize all d(j) in the information utility value matrix to obtain the set of weight coefficients w(j) for various urban ventilation environment characteristics. Then multiply the set of weight coefficients with the normalized feature matrix cell by cell and accumulate them to form a comprehensive ventilation environment index matrix. The comprehensive ventilation environment index matrix is ​​used as the final output of the spatial distribution results of urban ventilation environment changes under different planning objectives.

[0041] A city ventilation environment simulation and prediction system based on different planning objectives includes a data preprocessing module, a land use modeling module, a pattern generation module, a ventilation prediction module, and an index evaluation module.

[0042] The data preprocessing module is used to acquire ventilation characteristic data, land use data, and regulatory information data from multiple years as data sources; it unifies the data sources to the same coordinate system and spatial resolution and performs rasterization and numerical normalization processing.

[0043] The land use modeling module constructs a PLUS-based land use prediction model based on the land use data and regulatory information data; it rasterizes the driving factors of land use change; it sets conversion constraint markers on the raster to be converted according to the regulatory information data, generates a land use conversion matrix, and completes model training and validation.

[0044] The pattern generation module sets total land use constraints based on different planning objectives, inputs the total land use constraints and the land use transformation matrix into the land use prediction model, and generates the land use spatial pattern for the target year.

[0045] The ventilation prediction module constructs a spatiotemporal graph neural network prediction model based on multi-year ventilation characteristic data and land use spatial pattern of the target year; it uses grid cells as graph nodes and constructs adjacency relationships based on geographical proximity, learns spatial dependence through graph convolution operations and learns temporal dependence through recurrent cells, completes model training, and outputs the prediction results of urban ventilation environment characteristics for the target year.

[0046] The index evaluation module is used to perform weighted calculations on the predicted results of the urban ventilation environment characteristics using the entropy weight method. Based on the spatial heterogeneity statistical weight coefficients of each ventilation environment characteristic, it calculates the comprehensive ventilation environment index and outputs the spatial distribution results of ventilation environment changes under different planning objectives.

[0047] The technical effects and advantages of this invention are as follows:

[0048] 1. In existing technologies, land use conversion matrices lack the ability to express multi-level regulatory conditions such as underground facilities, legally protected areas, and ventilation corridors, leading to discrepancies between simulation results and actual implementation. This solution introduces regulatory information into the conversion matrix and sets conversion restriction markers, enabling the model to simultaneously consider planning red lines and real non-convertible areas, thus solving the key problem of the disconnect between simulation results and actual implementation.

[0049] 2. This scheme introduces the calculation logic of development potential map and neighborhood effect in the land use prediction process, combines the probability distribution of driving factors with spatial adjacency characteristics, thereby reflecting the evolution trend of land types and improving the rationality and reliability of the land use pattern generation in the target year.

[0050] 3. This scheme controls the dynamic pattern evolution process by superimposing land use total constraints in the cellular automata iteration, stopping the transformation of land types that have reached the upper limit and prioritizing the update of land types that have not reached the lower limit, thereby ensuring that the predicted pattern strictly conforms to the area constraints under different planning objectives.

[0051] 4. This solution introduces a spatiotemporal graph neural network in the ventilation environment prediction stage. It uses graph convolution to capture spatial dependencies and combines it with recurrent units to extract temporal dependency features. This allows the prediction results to reflect both spatial pattern changes and time series dynamics, improving the accuracy and stability of ventilation feature prediction.

[0052] 5. In the comprehensive evaluation of the prediction results, this scheme uses the entropy weight method for weighted calculation. By statistically analyzing the spatial heterogeneity of various ventilation environment characteristics, weight coefficients are formed, and a comprehensive ventilation environment index is constructed based on this. This enables quantitative evaluation and intuitive comparison of ventilation effect changes under different planning objectives. Attached Figure Description

[0053] Figure 1 This is a flowchart of the method steps of the present invention. Detailed Implementation

[0054] 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.

[0055] Refer to the instruction manual appendix Figure 1 An embodiment of the present invention provides a method for simulating and predicting urban ventilation environment based on different planning objectives, comprising:

[0056] S1. Obtain multi-year ventilation characteristic data, land use data, and regulatory information data as data sources; unify the data sources to the same coordinate system and spatial resolution and perform rasterization and numerical normalization processing;

[0057] S2. Construct a PLUS-based land use prediction model based on the land use data and regulatory information data; rasterize the driving factors of land use change; set conversion restriction markers for the raster to be converted according to the regulatory information data, generate a land use conversion matrix, and complete model training and verification.

[0058] S3. Set total land use constraints based on different planning objectives, input the total land use constraints and the land use transformation matrix into the land use prediction model, and generate the land use spatial pattern for the target year.

[0059] S4. Construct a spatiotemporal graph neural network prediction model based on multi-year ventilation characteristic data and land use spatial pattern of the target year; use grid cells as graph nodes and construct adjacency relationships based on geographical adjacency; learn spatial dependence through graph convolution operation and learn temporal dependence through recurrent cells; complete model training and output the prediction results of urban ventilation environment characteristics of the target year.

[0060] S5. The predicted results of the urban ventilation environment characteristics are weighted using the entropy weight method. Based on the spatial heterogeneity statistical weight coefficient of each ventilation environment characteristic, the comprehensive ventilation environment index is calculated and the spatial distribution results of ventilation environment changes under different planning objectives are output.

[0061] The ventilation characteristic data includes surface temperature gradient, surface roughness, and forest canopy density; the land use data includes construction land, cultivated land, forest, grassland, water bodies, and unused land; the regulatory information data includes the distribution of underground municipal facilities, legal protection boundaries, and ventilation corridor control boundaries.

[0062] The driving factors of land use change include elevation, slope, distance from highways, distance from main roads, distance from secondary roads, population density, GDP per capita, and building density.

[0063] In S1, multi-year ventilation characteristic data, land use data, and regulatory information data are unified to the same coordinate system and the same spatial resolution.

[0064] Spatial rasterization is performed on the unified data to transform each data source into a numerical matrix in the form of fixed raster cells, forming a raster dataset.

[0065] Linear normalization is performed on the rasterized dataset: the original value of the index of each data type in the raster cell is used as input, the upper limit and lower limit of the index of that type in all cells are set, and the normalization function is used to calculate the processed value, wherein the normalization function is: the processed value is equal to the original value of the raster cell minus the lower limit value, and then divided by the difference between the upper limit value and the lower limit value.

[0066] The normalized numerical matrix maintains the original positive and negative directions of change and outputs a unified standardized data input set.

[0067] In S2, based on historical land use data and rasterized data of driving factors, conversion samples of different land use types are statistically analyzed, and the probability values ​​of the conversion of various rasterized data to the target type are calculated to form a development potential map. The process of rasterizing driving factors is to divide the factors, which originally existed in the form of vectors or statistical tables, into regular grids according to the unified spatial resolution and coordinate system of the study area. Then, the value of the factor is calculated or interpolated in each grid cell, and all cells are filled in sequence to finally form a numerical matrix consistent with other data. In other words, the factors are first aligned to a unified spatial pattern, and then the factor value of each geographical location is mapped to the value of a raster cell, so that these factors can be used in the same calculation framework as land use data and ventilation characteristic data.

[0068] The probability values ​​of various types of raster data being transformed into the target type include comparing samples that have actually been transformed in historical periods with samples that have not been transformed, using driving factor values ​​as input variables, training the model using classification or regression methods (including logistic regression or random forest), and then inputting the driving factor values ​​into each raster cell, with the model outputting the probability value of that cell being transformed into the target type.

[0069] The development potential map is formed by filling the probability values ​​of each grid cell obtained from training into the corresponding spatial position one by one, and finally generating a probability distribution map with spatial grid as the unit and numerical values ​​representing the possibility of conversion.

[0070] The neighborhood effect is calculated for the candidate grid, and the number of target types within the neighborhood window is counted and combined with the preset neighborhood weight to form the neighborhood effect coefficient. The neighborhood effect calculation method includes: within the neighborhood window centered on the candidate grid, the number of grids belonging to the target type is counted, and the number of grids is multiplied by the preset neighborhood weight parameter. The result is the neighborhood effect coefficient of the candidate grid.

[0071] The development potential map is combined with the neighborhood effect coefficient, and conversion restriction markers are set for restricted grids based on regulatory information data to generate a dynamic land use conversion matrix. The generation process of the dynamic land use conversion matrix can be divided into two steps: First, the conversion probability value of each grid cell in the development potential map is multiplied by the neighborhood effect coefficient corresponding to that grid to obtain a comprehensive conversion probability matrix; Second, regulatory information data is introduced on the basis of the comprehensive conversion probability matrix, conversion restriction markers are set for the constrained grid cells, and their corresponding entries are replaced with prohibited conversion or restricted conversion status, thereby generating the dynamic land use conversion matrix.

[0072] Under the condition of satisfying the total land use constraint, the dynamic land use transformation matrix is ​​input into the cellular automaton iterative process to solve the land use spatial pattern of the target year.

[0073] The land use spatial pattern of the target year is compared with the corresponding historical land use data, consistency indicators are calculated, and the land use prediction model is validated.

[0074] It should be noted that the conditions for the total land use constraint mainly include the upper and lower limits of the area of ​​various types of land use in the target year, that is, the target total amount requirements set for construction land, cultivated land, forest, grassland, water body and unused land respectively; during the simulation process, it is necessary to ensure that the cumulative area of ​​each type of land use meets the preset upper and lower limits, so that the generated land use pattern is consistent with the planning target.

[0075] The cellular automaton includes a grid state set, a neighborhood window definition, a transformation rule set, a time step advancement mechanism, and a termination condition. Its iterative process is as follows: using the latest land use grid as the initial state, it reads the comprehensive transformation probability matrix and the dynamic land use transformation matrix. For each candidate grid, it searches for the permitted state between the current type and each candidate target type, calculating the corresponding comprehensive transformation probability score. Grids with transformation restriction markers are included in the non-transformable set and kept in their original state. Within the convertible set, a list to be updated is generated in descending order of the comprehensive transformation probability score. Combined with the total land use constraint, the area of ​​each type is statistically analyzed and verified in real time. Types reaching the upper limit are stopped from new transformations, while types not reaching the lower limit are prioritized. The state of this round is written in batches, advancing one time step. The above retrieval, statistics, verification, and updating are repeated until each type meets the total land use constraint and no new grids are added to the list to be updated. The land use spatial pattern for the target year is then solved and output.

[0076] S3 includes:

[0077] S3-1. The land use prediction model includes defining the state set of grid cells, neighborhood window rules, conversion probability calculation methods, and iterative update mechanisms. It uses the development potential map, neighborhood effect coefficient, and dynamic land use conversion matrix as input variables to form an executable prediction model framework.

[0078] The state set of the grid cell is defined as the combination of the land use type and its conversion restriction mark values ​​for each grid cell in the study area at the current time step. The type values ​​are consistent with the aforementioned land use data, and the conversion restriction mark comes from the regulatory information data and distinguishes between prohibited conversion status and restricted conversion status.

[0079] The neighborhood window rule calculates the number of target types in the neighborhood by setting a fixed size and shape of the neighborhood window for each grid cell and giving a neighborhood weight. Cells outside the study area at the boundary are not included in the statistics. The statistical result is multiplied by the neighborhood weight to form the neighborhood effect coefficient of the grid cell.

[0080] The conversion probability calculation method takes the conversion probability corresponding to the grid cell in the development potential map as the base, multiplies the probability by the neighborhood effect coefficient to obtain the initial value of the comprehensive conversion probability, and then assigns zero value to the items in the prohibited conversion state according to the dynamic land use conversion matrix, and adjusts the items in the restricted conversion state by the restriction coefficient to form the comprehensive conversion probability matrix.

[0081] The iterative update mechanism sorts all candidate rasters in descending order according to the comprehensive transformation probability matrix at each time step, and performs real-time statistics on the cumulative area of ​​each type in combination with the total land use constraint. It stops adding new transformations for types that have reached the upper limit, and prioritizes the selection of types that have not reached the lower limit. It uses a batch update method to write the current state at once and advance the time step. When each type meets the total land use constraint and no new rasters are added to the list to be updated, the iteration terminates and the land use spatial pattern of the target year is generated.

[0082] S3-2. Based on different planning objectives, set upper and lower limits for the area of ​​construction land, cultivated land, forests, grasslands, water bodies, and unused land in the target year as constraints, and store the constraints as total land use control parameters. In practical applications, different planning objectives include the land use direction and control conditions preset in the process of urban development. For example, the baseline scenario represents the state of maintaining the historical development trend unchanged, the rapid expansion scenario represents the state of construction land increasing significantly in the target year, the cultivated land protection scenario represents the state of strictly maintaining the lower limit of cultivated land area in the target year, and the ecological protection scenario represents the state of ensuring that forests and grasslands remain stable or increase in the target year.

[0083] In other words, different planning objectives set different upper and lower limits on the area of ​​various types of land use in the target year, thereby forming several land use total control schemes with differentiated constraints.

[0084] S3-3. During the model iteration process, the comprehensive conversion probability of each candidate grid is retrieved one by one. The total land use control parameters are combined with the dynamic land use conversion matrix. The conversion of land use types that have reached the area limit is stopped, and the land use types that have not reached the area limit are updated first, so as to ensure that the total constraint is continuously satisfied during the iteration process.

[0085] S3-4. After all land use types meet the total constraints and complete the iterative convergence, output the land use spatial pattern of the target year controlled by the constraints, and use it as the input for subsequent ventilation environment characteristic prediction.

[0086] Additionally, it should be noted in S3-3 that during the model iteration process, the comprehensive conversion probability of each candidate raster cell is retrieved one by one. First, this comprehensive conversion probability is combined with the constraints in the dynamic land use conversion matrix to determine whether the candidate raster cell is in a state where conversion is allowed or prohibited. Then, for raster cells in a state where conversion is allowed, the area is statistically analyzed according to the total land use control parameters. For land use types that have reached the area limit, the new conversion operation is stopped directly, and their current state remains unchanged in this iteration. For land use types that have not reached the area limit, the update operation is performed first, and the raster cells with the highest comprehensive conversion probability values ​​are written into the new state set first. Finally, after completing the retrieval, constraint determination, and area constraint statistics of all candidate raster cells, the raster state is updated uniformly and the next iteration begins, until all land use types meet the total control requirements.

[0087] In S3-4, it should be noted that during the iteration process, when the cumulative area of ​​all land use types meets the total land use constraint, and no new grid cells are generated in consecutive iterations to enter the set to be updated, the iteration process should be considered to have reached a convergence state. At this time, the land use type status of each grid cell in the final iteration result is uniformly written into the output set to form the land use spatial pattern under the target year conditions. This output result has been constrained by both the dynamic land use transformation matrix and the total land use control parameter, so its spatial distribution meets both regulatory requirements and planning objectives. Finally, the land use spatial pattern of the target year controlled by the constraints is used as the input data source to provide the subsequent urban ventilation environment characteristic prediction model to achieve the connection with the ventilation simulation stage.

[0088] S4 includes:

[0089] S4-1. Align the multi-year ventilation characteristic data with the land use spatial pattern of the target year into a time-ordered sequence input set, with each grid cell corresponding to a graph node. Encode the land use type of the target year as a node attribute and then concatenate it with the historical ventilation environment characteristics in the node dimension to form a node feature sequence for joint spatial and temporal modeling.

[0090] S4-2. Construct an adjacency matrix based on the geographic adjacency relationship of grid cells, determine the connection relationship between nodes using the eight-neighbor rule, write self-loop connections on the main diagonal of the adjacency matrix to maintain the transmission of node features, and perform normalization processing on the adjacency matrix to obtain a normalized adjacency matrix.

[0091] Geographic adjacency refers to the adjacency relationship between each grid cell and its directly contacting surrounding grid cells in a gridded space, including definitions using four-neighborhood (top, bottom, left, right) or eight-neighborhood (plus four diagonal directions);

[0092] The process of constructing an adjacency matrix based on the geographic adjacency relationship of grid cells is to first assign a unique node number to each grid cell, and then check whether the node is adjacent to other nodes in the matrix. If they are adjacent, the value is assigned to 1 in the corresponding row and column position; otherwise, the value is assigned to 0, thus forming a complete adjacency matrix.

[0093] S4-3. Perform graph convolution calculation on the normalized adjacency matrix and node feature sequence. For all nodes at each time step, aggregate the features of the neighboring nodes according to the adjacency relationship and complete the linear and nonlinear mappings. Solve the node spatial dependency representation of the corresponding time step and use the node spatial dependency representation as the input representation for time modeling.

[0094] In S4-3, at each time step, the feature vector of each node at the current time step is first weighted and summed with the feature vectors of all its neighboring nodes using the normalized adjacency matrix as the weight, thereby realizing the aggregation of neighborhood information. Then, the aggregation result is subjected to a linear mapping (mapping the aggregated features to a new feature space through matrix multiplication) and a nonlinear mapping (nonlinear transformation of the linear output through activation functions such as ReLU or Sigmoid), so that the model can capture both the linear and nonlinear relationships between nodes and their neighborhoods.

[0095] After completing the above aggregation, linear mapping, and nonlinear mapping, each node will obtain an updated feature representation at this time step. This representation already contains the node's own feature information and spatial relationship information from neighboring nodes. This updated node feature vector is the node spatial dependency representation at this time step, representing the node's dependency result in the spatial topology, and providing input for subsequent time series modeling.

[0096] S4 also includes:

[0097] S4-4. Input the spatial dependency representation of nodes arranged in chronological order into the recurrent unit. Construct a sliding time window sample in the recurrent unit according to a fixed time step. Use the preceding time step in the window as input and the ventilation environment features of the adjacent subsequent time steps as supervision signals. Minimize the mean square error loss to update the model parameters until the training process converges and the spatiotemporal joint modeling is completed.

[0098] S4-5. After training, the multi-year ventilation characteristic data and the land use spatial pattern of the target year are used as inference inputs. Graph convolution calculation and cyclic unit calculation are performed in sequence. The predicted results of urban ventilation environment characteristics of the target year are output for subsequent weighted calculation of ventilation environment index.

[0099] In S4-4, it should be noted that: the spatial dependency representations of nodes arranged in chronological order are sequentially input into the recurrent unit. Within the recurrent unit, sliding time window samples are constructed with a fixed time step. For each time window, the preceding time step within the window is used as the input feature, and the ventilation environment features of the immediately following time step are used as the supervision signal. The predicted value is calculated through forward propagation and compared with the true value. During the comparison process, the mean squared error loss is calculated, and this loss value is used to perform gradient updates on the model parameters during backpropagation. The above process of sample construction, prediction, loss calculation, and parameter update is repeated until the mean squared error loss stabilizes within the convergence interval, ultimately completing the spatiotemporal joint modeling and obtaining a trained recurrent unit model that can be used for prediction.

[0100] In S4-5, it should be noted that after the spatiotemporal graph neural network has been trained and reached convergence, the ventilation environment characteristic data of multiple years and the land use spatial pattern of the target year should be used as input data for the inference stage. First, graph convolution is used to aggregate neighborhood features for each grid node under the action of the normalized adjacency matrix and perform linear and nonlinear mappings to generate node representations containing spatial dependency information. Then, these node representations are input into the recurrent unit in chronological order. The recurrent unit interprets the time series features in turn and propagates forward in combination with historical states to output the node feature prediction value corresponding to the target year. After the prediction values ​​of all nodes are generated, they are written into the output set to form a complete prediction result of the urban ventilation environment characteristics for the target year. This prediction result serves as the input basis for the next stage of entropy weighted calculation to solve the comprehensive ventilation environment index and support the evaluation of ventilation environment under different planning objectives.

[0101] S5 includes:

[0102] S5-1. Extract each type of data from the prediction results of the ventilation characteristic data of the target year one by one, and perform linear normalization operation within all grid cells, using the minimum value as the lower limit and the maximum value as the upper limit, to obtain a unified normalized feature matrix.

[0103] S5-2. In the normalized feature matrix, for the j-th type of urban ventilation environment feature, the normalized value of each grid cell i is divided by the sum of the normalized values ​​of the feature in all grid cells to obtain the probability value p(i,j), and the probability distribution matrix is ​​formed by the probability values ​​of all grid cells.

[0104] S5-3. Based on the probability distribution matrix, calculate the information entropy value E(j) of each type of urban ventilation environment feature, and obtain the information utility value d(j) by subtracting the information entropy value from 1 to form an information utility value matrix. The information utility value matrix represents the spatial heterogeneity of various types of urban ventilation environment features.

[0105] S5-4. Normalize all d(j) in the information utility value matrix to obtain the set of weight coefficients w(j) for various urban ventilation environment characteristics. Then multiply the set of weight coefficients with the normalized feature matrix cell by cell and accumulate them to form a comprehensive ventilation environment index matrix. The comprehensive ventilation environment index matrix is ​​used as the final output of the spatial distribution results of urban ventilation environment changes under different planning objectives.

[0106] A city ventilation environment simulation and prediction system based on different planning objectives includes a data preprocessing module, a land use modeling module, a pattern generation module, a ventilation prediction module, and an index evaluation module.

[0107] The data preprocessing module is used to acquire ventilation characteristic data, land use data, and regulatory information data from multiple years as data sources; it unifies the data sources to the same coordinate system and spatial resolution and performs rasterization and numerical normalization processing.

[0108] The land use modeling module constructs a PLUS-based land use prediction model based on the land use data and regulatory information data; it rasterizes the driving factors of land use change; it sets conversion constraint markers on the raster to be converted according to the regulatory information data, generates a land use conversion matrix, and completes model training and validation.

[0109] The pattern generation module sets total land use constraints based on different planning objectives, inputs the total land use constraints and the land use transformation matrix into the land use prediction model, and generates the land use spatial pattern for the target year.

[0110] The ventilation prediction module constructs a spatiotemporal graph neural network prediction model based on multi-year ventilation characteristic data and land use spatial pattern of the target year; it uses grid cells as graph nodes and constructs adjacency relationships based on geographical proximity, learns spatial dependence through graph convolution operations and learns temporal dependence through recurrent cells, completes model training, and outputs the prediction results of urban ventilation environment characteristics for the target year.

[0111] The index evaluation module is used to perform weighted calculations on the predicted results of the urban ventilation environment characteristics using the entropy weight method. Based on the spatial heterogeneity statistical weight coefficients of each ventilation environment characteristic, it calculates the comprehensive ventilation environment index and outputs the spatial distribution results of ventilation environment changes under different planning objectives.

[0112] Working principle: First, in step S1, urban ventilation environment characteristic data, land use data and regulatory information data from multiple years are acquired. After being unified to the same coordinate system and spatial resolution, spatial rasterization is performed to convert each data source into a numerical matrix of fixed raster units. Then, linear normalization is performed according to the upper and lower limits of each indicator to keep the positive and negative directions consistent, forming a standardized data input set.

[0113] Subsequently, in step S2, based on the rasterized data of historical land use and driving factors, the samples are statistically transformed. Classification or regression methods are trained to solve the probability value of each raster transforming into the target type and fill it into the spatial location to form a development potential map. At the same time, the number of target types in the neighborhood window centered on the candidate raster is counted and multiplied by the preset neighborhood weight to obtain the neighborhood effect coefficient. The development potential map and the neighborhood effect coefficient are multiplied to generate a comprehensive transformation probability matrix. Then, based on the regulatory information data, transformation restriction marks are set for restricted rasteres, and prohibited or restricted transformations are written into the initial matrix to form a dynamic land use transformation matrix.

[0114] In step S3, the upper and lower limits of the area constrained by the total land use are used as the total land use control parameters. They are input into the cellular automaton along with the dynamic land use transformation matrix. Candidate rasters are sorted according to the comprehensive transformation probability. In each time step, rasters that are prohibited from transformation are first excluded. Then, new transformations are stopped for types that have reached the upper limit. Types that have not reached the lower limit are selected first. The batch is written into the current state and the time step is advanced until each type meets the total land use constraint and no new rasters to be updated are generated. The land use spatial pattern of the target year is output and the consistency index is calculated using historical data to complete the verification.

[0115] Next, in step S4, grid cells are used as graph nodes. The ventilation environment features of multiple years are aligned with the land use spatial pattern of the target year to form a node feature sequence. An adjacency matrix is ​​constructed based on four or eight neighbors and written into a self-loop before normalization. At each time step, the normalized adjacency matrix is ​​used to weight and aggregate the node and neighborhood features and complete linear and nonlinear mapping to obtain the node spatial dependency representation. Then, the node spatial dependency representation arranged in time order is input into the recurrent unit. A sliding time window with a fixed time step is used, with the previous time step as input and the ventilation environment features of the adjacent subsequent time steps as supervision signals. The mean square error loss is minimized until convergence. After training, graph convolution calculation and recurrent unit calculation are performed in sequence to output the prediction result of the urban ventilation environment features of the target year.

[0116] Finally, in step S5, the prediction result is linearly normalized again according to the upper and lower limits of each ventilation environment feature in the whole area to form a normalized feature matrix. The probability distribution matrix of each feature in each grid cell is calculated and the information entropy and information utility value are statistically analyzed. The normalized information utility value is used to obtain the weight coefficient set. The weight coefficient set is multiplied by the normalized feature matrix cell by cell and accumulated to form a comprehensive ventilation environment index matrix. The spatial distribution result of urban ventilation environment changes under different planning objectives is output.

[0117] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for simulating and predicting urban ventilation environment based on different planning objectives, characterized in that, include: S1. Obtain multi-year ventilation characteristic data, land use data, and regulatory information data as data sources; Unify the data sources to the same coordinate system and spatial resolution, and perform rasterization and numerical normalization processing; S2. Construct a PLUS-based land use prediction model based on the land use data and regulatory information data; rasterize the driving factors of land use change; set conversion restriction markers for the raster to be converted according to the regulatory information data, generate a land use conversion matrix, and complete model training and verification. S3. Set total land use constraints based on different planning objectives, input the total land use constraints and the land use transformation matrix into the land use prediction model, and generate the land use spatial pattern for the target year. S4. Construct a spatiotemporal graph neural network prediction model based on multi-year ventilation characteristic data and land use spatial pattern of the target year; use grid cells as graph nodes and construct adjacency relationships based on geographical adjacency; learn spatial dependence through graph convolution operation and learn temporal dependence through recurrent cells; complete model training and output the prediction results of urban ventilation environment characteristics of the target year. S5. The predicted results of the urban ventilation environment characteristics are weighted using the entropy weight method. Based on the spatial heterogeneity statistical weight coefficient of each ventilation environment characteristic, the comprehensive ventilation environment index is calculated and the spatial distribution results of ventilation environment changes under different planning objectives are output.

2. The urban ventilation environment simulation and prediction method based on different planning objectives according to claim 1, characterized in that: The ventilation characteristic data includes surface temperature gradient, surface roughness, and forest canopy density; the land use data includes construction land, cultivated land, forest, grassland, water bodies, and unused land; the regulatory information data includes the distribution of underground municipal facilities, legal protection boundaries, and ventilation corridor control boundaries. The driving factors of land use change include elevation, slope, distance from highways, distance from main roads, distance from secondary roads, population density, GDP per capita, and building density.

3. The urban ventilation environment simulation and prediction method based on different planning objectives according to claim 2, characterized in that: In S1, multi-year ventilation characteristic data, land use data, and regulatory information data are unified to the same coordinate system and the same spatial resolution. Spatial rasterization is performed on the unified data to transform each data source into a numerical matrix in the form of fixed raster cells, forming a raster dataset. Linear normalization is performed on the rasterized dataset: the original value of the index of each data type in the raster cell is used as input, the upper limit and lower limit of the index of that type in all cells are set, and the normalization function is used to calculate the processed value, wherein the normalization function is: the processed value is equal to the original value of the raster cell minus the lower limit value, and then divided by the difference between the upper limit value and the lower limit value. The normalized numerical matrix maintains the original positive and negative directions of change and outputs a unified standardized data input set.

4. The urban ventilation environment simulation and prediction method based on different planning objectives according to claim 3, characterized in that: In S2, based on historical land use data and rasterized data of driving factors, conversion samples of different land use types are statistically analyzed, the probability values ​​of conversion of various rasterized data to the target type are calculated, and a development potential map is formed. Perform neighborhood effect calculation on candidate grids, count the number of target types within the neighborhood window and combine it with preset neighborhood weights to form a neighborhood effect coefficient; The development potential map is combined with the neighborhood effect coefficient, and conversion restriction markers are set for restricted rasters based on regulatory information data to generate a dynamic land use conversion matrix. Under the condition of satisfying the total land use constraint, the dynamic land use transformation matrix is ​​input into the cellular automaton iterative process to solve the land use spatial pattern of the target year. The land use spatial pattern of the target year is compared with the corresponding historical land use data, and consistency indicators are statistically analyzed to verify the land use prediction model.

5. The urban ventilation environment simulation and prediction method based on different planning objectives according to claim 4, characterized in that: S3 includes: S3-1. The land use prediction model includes defining the state set of grid cells, neighborhood window rules, conversion probability calculation methods and iterative update mechanisms, and taking the development potential map, neighborhood effect coefficient and dynamic land use conversion matrix as input variables to form the prediction model framework. S3-2. Based on different planning objectives, set upper and lower limits for the area of ​​construction land, cultivated land, forest, grassland, water bodies and unused land for the target year as constraints, and store the constraints as total land use control parameters. S3-3. During the model iteration process, the comprehensive conversion probability of each candidate grid is retrieved one by one. The total land use control parameters are combined with the dynamic land use conversion matrix. Land use types that have reached the area limit are stopped from being newly converted, while land use types that have not reached the area limit are updated first. S3-4. After all land use types meet the total constraints and complete iterative convergence, output the land use spatial pattern of the target year controlled by the constraints, and use it as the input for subsequent ventilation environment characteristic prediction.

6. The urban ventilation environment simulation and prediction method based on different planning objectives according to claim 5, characterized in that: S4 includes: S4-1. Align the multi-year ventilation characteristic data with the land use spatial pattern of the target year into a time-ordered sequence input set, with each grid cell corresponding to a graph node. Encode the land use type of the target year as a node attribute and then concatenate it with the historical ventilation environment characteristics in the node dimension to form a node feature sequence for joint spatial and temporal modeling. S4-2. Construct an adjacency matrix based on the geographic adjacency relationship of grid cells, determine the connection relationship between nodes using the eight-neighbor rule, write self-loop connections on the main diagonal of the adjacency matrix to maintain the transmission of node features, and perform normalization processing on the adjacency matrix to obtain a normalized adjacency matrix. S4-3. Perform graph convolution calculation on the normalized adjacency matrix and node feature sequence. For all nodes at each time step, aggregate the features of neighboring nodes according to the adjacency relationship and complete the linear and nonlinear mappings. Solve the node spatial dependency representation of the corresponding time step and use the node spatial dependency representation as the input representation for time modeling.

7. The urban ventilation environment simulation and prediction method based on different planning objectives according to claim 6, characterized in that: S4 also includes: S4-4. Input the spatial dependency representation of nodes arranged in chronological order into the recurrent unit. Construct a sliding time window sample in the recurrent unit according to a fixed time step. Use the preceding time step in the window as input and the ventilation environment features of the adjacent subsequent time steps as supervision signals. Minimize the mean square error loss to update the model parameters until the training process converges and the spatiotemporal joint modeling is completed. S4-5. After training, the multi-year ventilation characteristic data and the land use spatial pattern of the target year are used as inference inputs. Graph convolution calculation and cyclic unit calculation are performed in sequence, and the prediction results of urban ventilation environment characteristics of the target year are output for subsequent weighted calculation of ventilation environment index.

8. The urban ventilation environment simulation and prediction method based on different planning objectives according to claim 7, characterized in that: S5 includes: S5-1. Extract each type of data from the prediction results of the ventilation characteristic data of the target year one by one, and perform linear normalization operation within all grid cells, using the minimum value as the lower limit and the maximum value as the upper limit, to obtain a unified normalized feature matrix. S5-2. In the normalized feature matrix, for the j-th type of urban ventilation environment feature, the normalized value of each grid cell i is divided by the sum of the normalized values ​​of the feature in all grid cells to obtain the probability value p(i,j), and the probability distribution matrix is ​​formed by the probability values ​​of all grid cells. S5-3. Based on the probability distribution matrix, calculate the information entropy value E(j) of each type of urban ventilation environment characteristic, and obtain the information utility value d(j) by subtracting the information entropy value from 1, forming an information utility value matrix. S5-4. Normalize all d(j) in the information utility value matrix to obtain the set of weight coefficients w(j) for various urban ventilation environment characteristics. Then multiply the set of weight coefficients with the normalized feature matrix cell by cell and accumulate them to form a comprehensive ventilation environment index matrix. The comprehensive ventilation environment index matrix is ​​used as the final output of the spatial distribution results of urban ventilation environment changes under different planning objectives.

9. A city ventilation environment simulation and prediction system based on different planning objectives, comprising a data preprocessing module, a land use modeling module, a pattern generation module, a ventilation prediction module, and an index evaluation module, characterized in that: The data preprocessing module is used to acquire ventilation characteristic data, land use data, and regulatory information data from multiple years as data sources; it unifies the data sources to the same coordinate system and spatial resolution and performs rasterization and numerical normalization processing. The land use modeling module constructs a PLUS-based land use prediction model based on the land use data and regulatory information data; and rasterizes the driving factors of land use change. Based on the regulatory information data, set conversion restriction markers for the raster to be converted, generate a land use conversion matrix, and complete model training and validation. The pattern generation module sets total land use constraints based on different planning objectives, inputs the total land use constraints and the land use transformation matrix into the land use prediction model, and generates the land use spatial pattern for the target year. The ventilation prediction module constructs a spatiotemporal graph neural network prediction model based on multi-year ventilation characteristic data and land use spatial pattern of the target year; it uses grid cells as graph nodes and constructs adjacency relationships based on geographical proximity, learns spatial dependence through graph convolution operations and learns temporal dependence through recurrent cells, completes model training, and outputs the prediction results of urban ventilation environment characteristics for the target year. The index evaluation module is used to perform weighted calculations on the predicted results of the urban ventilation environment characteristics using the entropy weight method. Based on the spatial heterogeneity statistical weight coefficients of each ventilation environment characteristic, it calculates the comprehensive ventilation environment index and outputs the spatial distribution results of ventilation environment changes under different planning objectives.

Images