Urban space resource optimal allocation method based on urban information model

Abstract

The application provides a kind of urban space resource optimization configuration method based on city information model, it is related to space resource configuration field, comprising: step S1, data acquisition and pretreatment, step S2, dynamic association feature extraction, step S3, space-time coordination supply-demand matching, step S4, dynamic demand prediction, step S5, facility layout optimization.The method is based on city information model, realizes the iterative optimization of infrastructure and other urban space resources in the whole life cycle of planning, construction and management, effectively avoids the problems such as demand reduction and public resource redundancy caused by uncontrollable factors of urban development during the construction process or after the completion of infrastructure, effectively improves the utilization rate of urban space and reduces the planning and construction cost of infrastructure.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of space resource configuration, and particularly relates to a city space resource optimization configuration method based on a city information model. BACKGROUND

[0002] City information modeling (CIM) is a three-dimensional spatial information model taking digital technology as the core and integrating whole-element and whole-life-cycle information of a city, and is a digital foundation for smart city construction and fine city management; the CIM constructs a "digital city" that is synchronized with a city by fusing geographic spatial information, building information modeling (BIM), Internet of Things (IoT), government data and other multi-dimensional information, and provides intelligent support for city planning, construction, management and service in an integrated and systematic manner to manage city operation.

[0003] In the process of city construction, infrastructure planning is the basis for supporting sustainable economic and social development, livelihood needs (green, livable, etc.) and ecological security of a region or city, and is a systematic and forward-looking arrangement for the layout, scale, construction sequence and operation mechanism of various basic engineering facilities (for example, transportation, energy, water conservancy, communication, environmental protection, security, medical treatment, etc.), so as to realize orderly construction of a region or city, and directly affects the development quality, operation efficiency and life quality of residents of a region or city. However, due to the super-advance nature of infrastructure planning, factors such as sudden changes in social and economic patterns (for example, unexpected population flow, drastic adjustment of industrial structure, etc.), accelerated technology iteration (leading to "uncompleted and backward" infrastructure), sudden conditions of natural environment and ecological constraints (for example, sudden earthquakes, floods, landslides, etc. during construction, leading to project suspension or line change), uncontrollable risks of construction and supply chain, etc. cause the space-time mismatch of "demand chain" and "supply chain", leading to the risk of "excess" or "backwardness" of super-advance planned infrastructure, and the built infrastructure loses its original planning function, resulting in problems such as resource waste and increased construction cost. SUMMARY

[0004] In view of the problems in the prior art, the present application aims to provide a city space resource optimization configuration method based on a city information model, which realizes iterative optimization of city space resources such as infrastructure in the whole life cycle of planning, construction and management based on the city information model, effectively avoids problems such as reduced infrastructure demand and redundant public resources caused by uncontrollable factors of city development after planning and during construction, and improves the utilization rate of city space and reduces the planning and construction cost of infrastructure.

[0005] The object of the present application is achieved by the following technical solutions:

[0006] A city space resource optimization configuration method based on a city information model, comprising:

[0007] Step S1, data acquisition and preprocessing: acquiring static data and dynamic data in the city based on a city information model (CIM), and performing spatiotemporal dynamic alignment preprocessing on the data to obtain multi-element heterogeneous data;

[0008] Step S2, dynamic correlation feature extraction: converting the multi-element heterogeneous data into spatiotemporal features to provide robust and semantically rich inputs for subsequent supply-demand matching, dynamic demand prediction, and layout optimization;

[0009] Step S3, spatiotemporal collaborative supply-demand matching: using a spatiotemporal potential field to depict the supply-demand relationship between facilities and planning areas; that is, first, assigning weights and time decay factors to facilities according to facility types to obtain the supply potential of facilities of the corresponding type, then using regional demand factors and spatiotemporal feature codes in step S2 to obtain the demand potential of the corresponding facilities, and finally obtaining the supply-demand relationship through the supply potential and the demand potential;

[0010] Step S4, dynamic demand prediction: obtaining facility demand prediction for future time points to provide accurate and causally interpretable prediction results for layout optimization;

[0011] Step S5, facility layout optimization: using proximal policy optimization to realize dynamic optimization of facility layout.

[0012] Based on further optimization of the above scheme, the static data includes infrastructure vectors, planning maps, etc.; the dynamic data includes human flow, traffic, POI (Point of Interest) activity (a core indicator that measures the dynamic performance of a map interest point in a specific scenario, which is usually evaluated by combining multidimensional data such as human flow, user interaction, and time-space features), historical demand, social media heat, festival calendar, traffic congestion index, etc.

[0013] Based on further optimization of the above scheme, the "spatiotemporal dynamic alignment preprocessing" in step S1 includes temporal dynamic alignment and spatial adaptive alignment;

[0014] Temporal dynamic alignment: for dynamic data, the timestamps of all dynamic data are made consistent to eliminate the effects of data collection lag, including:

[0015] For example, the traffic congestion index T co (t) As the core of temporal dynamic alignment, the dynamic lag function is obtained:

[0016]

[0017] In the formula: Indicates k The hysteresis of the data at time t ; Indicates k The historical average hysteresis of the data; ;

[0018] The timestamps of all dynamic data are corrected by using a dynamic hysteresis function:

[0019]

[0020] In the formula: Indicates the minimum time granularity; t k Indicates k The original timestamp of the data;

[0021] Spatial adaptive alignment: taking the seats of infrastructure vectors (such as hospitals, shopping malls, etc.) as data source positions s j , and dividing the grid cells by the land use properties (such as commercial areas, residential areas, etc.) of the planning map g i ; for the aligned dynamic data, the decay coefficient of the data type j is allocated to obtain the spatial weight:

[0022]

[0023] In the formula: IOU Indicates the spatial intersection ratio;

[0024] All data sources (including static data and dynamic data) are weighted by spatial weight to obtain spatiotemporally consistent grid-level fusion data, i.e., multi-element heterogeneous data G i (t) :

[0025]

[0026] In the formula: Data j (t) Indicates the original observation data of the data of the j type at time t .

[0027] Based on the further optimization of the above scheme, the dynamic neighborhood spatiotemporal autoencoder introduced in the step S2 converts the multi-element heterogeneous data into spatiotemporal features, including an encoder and a decoder:

[0028] Encoder for capturing dynamic neighborhood associations:

[0029] Real-time mutual information I(i,j,t) Screening moment t With grid i Strongly associated neighborhood N(i,t) :

[0030]

[0031] In the formula: Indicates the moment t The maximum mutual information of the whole area below;

[0032] Real-time mutual information is the KL divergence between the joint distribution of the feature sequence of the grid i And j In the local time window and the independent distribution, that is:

[0033]

[0034]

[0035] In the formula: S And D Respectively represent the category label after feature discretization, which is a bridge from continuous features to discrete probability, used to solve the calculation difficulty of mutual information on continuous data; p ij (S,D) Indicates the joint probability of grid i In category S , grid j In category D ; n ij (S,D) Indicates the actual observation count of feature pair (S,D) , that is, the number of samples of grid i Data S And grid j Belongs to D ; p i (S) , p j (D) Respectively represent the marginal probability of grid i In category S , the marginal probability of grid j In category D ; n i (S) Indicates the marginal observation count of feature S , that is, grid i In category Sthe number of samples, n j (D) representing features D edge observation count, i.e., grid j individually in a class D the number of samples; N representing the total amount of samples;

[0036] based on dynamic neighborhood N(i,t) spatiotemporal feature encoding is completed:

[0037]

[0038] wherein: z i,t representing the feature vector output by the encoder; Encoder representing the encoder function; representing the dynamic neighborhood N(i,t) all grids j fusion data set at time t, used to capture the grid i real-time interaction with strong grids; representing the grid i fusion data sequence of the past L time steps, used to capture the historical time trend of grid i; L representing the length of the time window, i.e., the number of historical time steps, used to control the amount of historical data included by the encoder;

[0039] decoder, used to verify the validity of spatiotemporal features and strengthen neighborhood association:

[0040] Based on the reconstruction error, dynamic neighborhood constraints are added to ensure that spatiotemporal features contain temporary associations, and the loss function is:

[0041]

[0042] wherein: Decoder representing the decoder function; is the neighborhood constraint weight.

[0043] Based on the further optimization of the above scheme, in step S2, during the process of converting multi-element heterogeneous data into spatiotemporal features, the Dirichlet prior + pseudo count method is also used to solve the problem of misjudgment of sparse data (such as sparse data at night, remote areas, etc.).

[0044] First, by adding pseudo count , the probability estimation is smoothed, and joint probability smoothing and marginal probability smoothing are obtained respectively:

[0045] Joint probability smoothing:

[0046] Edge probability smoothing:

[0047] wherein: K denotes the number of feature categories;

[0048] Then, mutual information calculation is performed:

[0049] .

[0050] Based on further optimization of the above scheme, the step S3 of "adopting a space-time potential field to depict the supply-demand relationship between the facility and the planning land" is specifically:

[0051] First, according to the facility type, a facility weight is given w k and a time decay factor T s (k) , to obtain the supply potential of the corresponding type of facility:

[0052]

[0053] wherein: C k denotes the scale of the k category data (i.e., the facility); denotes the real-time utilization rate of the k category data (i.e., the facility) at the t moment; x denotes the spatial position variable, i.e., the target position; x k denotes the actual position of the k category data (i.e., the facility); t denotes the time variable, i.e., the current moment; t k denotes the service peak period of the k category data (i.e., the facility); denotes the spatial decay coefficient;

[0054] The demand potential of the corresponding facility is obtained through a regional demand factor B(x) (the regional demand factor is defined according to the regional nature of the corresponding facility):

[0055]

[0056] wherein: w q denotes the demand weight vector of the corresponding facility;

[0057] The grid i is obtained t at the moment M(i,t) , and the supply-demand matching degree is

[0058]

[0059] wherein: denotes the total quantity penalty coefficient; t reach (i,k) denotes the grid i spatial-temporal reachable time to the facility k ; T 0 denotes the reachable time threshold.

[0060] Based on further optimization of the above scheme, the step S4 dynamic demand prediction is specifically:

[0061] Introduce causal gate in the LSTM (Long Short-Term Memory) model, so as to distinguish the causal influence of exogenous variables such as policy release, emergency, social media heat, and festival day;

[0062] First, according to the specific type, the causal variable is decomposed into endogenous variable X end (t) and exogenous variable X exo (t) The endogenous variable includes historical demand, flow, and other time series dependent variables, and the exogenous variable includes policy release, emergency, social media heat, and festival day, etc.

[0063] Then, introduce the causal gate in the LSTM model to correct the cell state to include the intervention effect:

[0064] Input gate:

[0065] wherein: denotes the Sigmiod activation function; W i denotes the input weight, connecting the current time input x t to the input gate weight matrix; x t denotes the current time input; U i denotes the hidden state weight, connecting the previous time hidden state h t-1 to the input gate weight matrix; h t-1 denotes the previous time hidden state; V i denotes the cell state weight, connecting the previous time cell state c t-1weight matrix of input gate; c t-1 denotes the cell state at the last time step (LSTM's core carrier to store history information); b i denotes the input gate bias;

[0066] forget gate:

[0067] wherein: W f , U f , V f denote the weight matrix of input, hidden state, cell state to forget gate, respectively; b f denotes the forget gate bias;

[0068] candidate state:

[0069] wherein: tanh() denotes the Tanh activation function; W c , U c denote the weight matrix of input, hidden state to candidate state, respectively; b c denotes the candidate state bias;

[0070] causally-modified cell state:

[0071] wherein: denotes the element-wise multiplication; denotes the causal strength; denotes the cell state change caused by intervention;

[0072] output gate:

[0073] wherein: W o , U o , V o denote the weight matrix of input, hidden state, current cell state to output gate, respectively; b o denotes the output gate bias;

[0074] output state:

[0075] Causal strength is obtained by MLP (Multilayer Perceptron), which is composed of an input layer, two hidden layers and an output layer, and is fully connected between layers, and the activation function adopts ReLU:

[0076]

[0077] The PC (Peter-Clark) algorithm is used to construct the causal graph of exogenous variables to obtain the changes in cell state caused by intervention :

[0078]

[0079] In the formula: do() represents the intervention operator; represents the causal coefficient obtained by data mining, which is used to measure the marginal influence of variables k on demand;

[0080] Finally, the demand prediction at the future time is output:

[0081]

[0082] In the formula: W h represents the weight matrix input to the output layer; b h represents the bias of the output layer; represents the predicted fluctuation coefficient.

[0083] Based on the further optimization of the above scheme, the "facility layout optimization" in step S5 is specifically:

[0084] First, define the state space St and action space At :

[0085]

[0086] In the formula: G i (t) represents the dynamically fused multi-element heterogeneous data reflecting the real-time supply and demand relationship, which is output by step S1;{ JS k (t) represents the construction progress, which constrains the "time sequence feasibility" of the decision, and is obtained according to the actual construction progress; represents the demand prediction at the future time , which is output by step S4; t build represents the construction time sequence; Crepresents the facility scale; C max represents the resource upper limit (i.e. the upper limit of the infrastructure budget of the corresponding area); A O represents the actionable space, i.e. the set of all facility layout decisions that meet the resource constraints;

[0087] Then, the reward function is set to distinguish the importance of the facility area and realize dynamic matching:

[0088]

[0089] In the formula: B(x) i represents the area demand factor (used to measure the priority of resource matching); respectively represent the corresponding weight coefficients;

[0090] represents the average matching degree when the facility is built, which is obtained by the supply-demand matching degree in step S3 M(i, t) obtained, i.e. the supply-demand matching degree when built is predicted by step S3 M(i,t+t build ) Then, the average matching degree is obtained according to the area demand factor for all grids i

[0091]

[0092] represents the redundancy (i.e. the space-time proportion when the facility utilization rate is lower than the threshold);

[0093] Finally, the strategy network outputs the layout decision probability, and then iteratively updates the network parameters to realize layout optimization (PPO, Proximal Policy Optimization):

[0094]

[0095] In the formula: represents the parameters before the n th update; represents the learning rate (obtained through experimental data); E represents the expectation of the experience data distribution of "state-action-advantage"; Y t represents the advantage function, which guides the direction of policy update; clip() represents the clipping operation; represents the clipping coefficient (hyperparameter). ​

[0096] Based on the further optimization of the above scheme, the redundancy acquisition method is: first, taking the position coordinates of the facility as the center, a circular area with a radius of R is generated as the service range corresponding to the facility; then, the demand of the first grid at the construction time is predicted by introducing the causal gate in the LSTM model in step S4; the spatial intersection analysis of the geometric range of the grid and the service range of the facility is carried out, and the intersection area ratio is obtained; a preset intersection threshold is set, if the intersection area is greater than the intersection threshold, the demand is summarized; then, the predicted utilization rate is obtained. k r k i g i k IOU(g i ,k)

[0097]

[0098]

[0099]

[0100] t yy t yc

[0101]

[0102] The technical effects of the scheme are as follows:

[0103] ​​​​​​​​​​​​​This invention transforms diverse and heterogeneous data (static and dynamic data) into high-quality spatiotemporal features through dynamic correlation feature extraction, effectively addressing issues such as lag, decay, and sparsity in the original data. This provides robust and semantically rich input for subsequent supply-demand matching and layout optimization. Subsequently, through spatiotemporally coordinated supply-demand matching, it not only accurately characterizes supply-demand relationships and improves resource allocation efficiency but also effectively corrects matching degrees and avoids resource mismatch, thereby supporting dynamic layout optimization, quantifying the value of spatial resources, and forming efficient and accurate long-term planning. Dynamic demand forecasting not only accurately captures demand drivers and improves forecast accuracy but also quantifies forecast uncertainties, making layout decision risks controllable. This provides accurate, robust, and interpretable demand forecasts for subsequent layout optimization. Finally, through facility layout optimization, it not only ensures regional fairness and optimizes resource utilization but also effectively responds to urban emergencies and long-term trend changes, thereby enhancing the dynamic responsiveness of the entire model. This invention transforms resource allocation from "experience-driven" to "data-model-driven" through a combination of steps: data acquisition and preprocessing, dynamic correlation feature extraction, spatiotemporal collaborative supply and demand matching, dynamic demand forecasting, and facility layout optimization. This enables precise spatiotemporal optimization, type-adaptive and dynamically flexible optimization of urban spatial resources, achieving iterative optimization of urban spatial resources throughout their entire lifecycle of planning, construction, and management. It avoids problems such as reduced urban demand for facilities due to time lags and unforeseen factors after planning and during construction, leading to resource waste and planning mismatch. This improves the spatiotemporal matching accuracy of resource allocation with urban operational needs and reduces the redundancy of public resources. Attached Figure Description

[0104] Figure 1 This is a structural block diagram of the urban spatial resource collaborative optimization allocation method in an embodiment of the present invention. Detailed Implementation

[0105] The technical solutions in the embodiments of the present invention will be clearly and completely described below. In the following description, specific details such as specific system structures and technologies are presented for illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present invention.

[0106] Example 1:

[0107] A method for optimizing the allocation of urban spatial resources based on urban information models includes:

[0108] Step S1, data acquisition and preprocessing: based on the city information model (CIM), static data and dynamic data in the city are obtained, the static data includes infrastructure vector, planning map, etc. (which can be obtained by surveying and mapping and three-dimensional modeling, for example: using oblique photography, laser radar, etc. to collect urban geographic space data, generate high-precision three-dimensional model; at the same time, cooperate with natural resources, planning, housing, etc. Department, auxiliary satellite image data, obtain administrative division, land use planning, traffic network, etc. Data); dynamic data includes people flow, traffic, POI activity (a core index for measuring the dynamic performance of map interest points in a specific scene, which usually combines people flow, user interaction, time and space characteristics, etc. Multidimensional data comprehensive evaluation), historical demand (can be collected by deploying flow sensors, cameras on roads, public transportation sites; can be obtained from population census, traffic management system, etc. Government information system to export historical demand, POI activity, etc. Data), social media heat, festival calendar, traffic congestion index (social media heat information and festival calendar information can be obtained by web crawler technology; traffic congestion index can be obtained by deploying geomagnetic sensors, radar detectors, etc. In key road sections through real-time traffic data from traffic management departments) and other data.

[0109] And the data is preprocessed by spatiotemporal dynamic alignment, and multi-element heterogeneous data is obtained, which includes time dynamic alignment and space adaptive alignment.

[0110] Time dynamic alignment: for dynamic data, make the timestamps of all dynamic data consistent, eliminate the influence of data collection lag, including:

[0111] Traffic congestion index T co (t) As the core of time dynamic alignment, the dynamic lag function is obtained:

[0112]

[0113] In the formula: Indicates k The lag of similar data at time t ; Indicates kThe historical average lag of the data (obtained by statistically averaging historical data, for example, the historical average lag of traffic data, 1000 historical traffic accidents can be collected, the actual time of the historical traffic accident or the actual time of the congestion start is obtained as the event time, the time corresponding to the historical traffic accident uploaded by the sensor or camera instrument is obtained as the collection time, the difference between the collection time and the event time is the lag of the corresponding traffic accident, and the average value of the lag of 1000 historical traffic accidents is the historical average lag, which can realize the collection chain delay correction, thereby accurately quantifying the historical lag characteristics of each type of data and providing key parameters for spatio-temporal causal alignment); (0 represents smooth, and 1 represents severe congestion);

[0114] The timestamps of all dynamic data are corrected by using a dynamic lag function:

[0115]

[0116] In the formula: represents the minimum time granularity (generally 4-6 min); t k represents k the original timestamp of the data;

[0117] For example, for each type of data (such as people flow, traffic, POI activity, historical demand, social media heat, festival calendar, and traffic congestion index), the delay on the collection chain is obtained by using a dynamic lag function (for example, people flow data is collected in real time, POI activity is reported in real time, historical demand has no collection delay, and social media heat is crawled in real time, i.e. , , k is any one of people flow, POI activity, historical demand, and social media heat; for traffic flow, the influence of historical average lag on traffic congestion index is obtained as lag influence, i.e. traffic flow reflects after traffic congestion, so the corresponding is positive, and the timestamp is obtained according to the corresponding historical average lag; for the festival calendar, the influence of historical average lag on traffic congestion index is obtained as advance influence, for example, the influence of the National Day holiday on traffic congestion index is reflected 3 days in advance, so the corresponding is negative, and the corresponding timestamp is obtained by statistically averaging the historical average lag of the three days before the National Day holiday).

[0118] Spatial adaptive alignment: the seats of infrastructure vectors (such as hospitals and shopping malls) are taken as data source positions s jAt the same time, grid units are divided according to the land use nature (such as commercial area, residential area, etc.) on the planning map. g i For aligned dynamic data, categorized by data type j Assign its attenuation coefficient (The attenuation coefficient is obtained based on the actual dynamic data type and empirical data; for example, if the impact of pedestrian flow data on short distances is significant, then...) ; POI data has a stable long-distance impact, then ), obtain spatial weights:

[0119]

[0120] In the formula: IOU Indicates the intersection-union ratio of spaces;

[0121] All data sources (including static and dynamic data) are spatially weighted to obtain spatiotemporally consistent grid-level fused data, i.e., multi-dimensional heterogeneous data. G i (t) :

[0122]

[0123] In the formula: Data j (t) Indicates the first j Class data at time t The original observation data.

[0124] Step S2, Dynamic Relationship Feature Extraction: Transform diverse heterogeneous data into spatiotemporal features, providing robust and semantically rich input for subsequent supply and demand matching, dynamic demand prediction, and layout optimization;

[0125] Including encoders and decoders:

[0126] Encoder, used to capture dynamic neighborhood associations:

[0127] Based on real-time mutual information I(i,j,t) Filtering time t With grid i Strongly related neighborhoods N(i,t) :

[0128]

[0129] In the formula: Indicates time t Maximize the mutual information across the entire region (ensuring that only strongly correlated grids are included, such as large venues and their surrounding transportation hubs).

[0130] Real-time mutual information is grid i andj The joint distribution of the feature sequence within a local time window and the KL divergence of the independent distribution, i.e.,

[0131]

[0132]

[0133] wherein: S and D respectively represent the class label after feature discretization, and is a bridge from continuous features to discrete probability, used to solve the calculation difficulty of mutual information on continuous data; p ij (S,D) represents the joint probability of the grid i in the category S , the grid j in the category D ; n ij (S,D) represents the actual observation count of the feature pair (S,D) , i.e., the number of samples of which the grid i data S and the grid j belongs to D ; p i (S) , p j (D) respectively represent the marginal probability of the grid i in the category S alone, and the marginal probability of the grid j in the category D ; n i (S) represents the marginal observation count of the feature S , i.e., the number of samples of which the grid i is in the category S alone, n j (D) represents the marginal observation count of the feature D , i.e., the number of samples of which the grid j is in the category D alone; N represents the total number of samples;

[0134] The spatio-temporal feature encoding is completed based on the dynamic neighborhood: N(i,t)

[0135]

[0136] wherein: z ​i,t represents the feature vector output by the encoder; Encoder represents the encoder function (usually a neural network such as LSTM, CNN, Transformer, etc.); represents the dynamic neighborhood N(i,t) all grids in j fusion data set at time t, used to capture the grid i real-time interaction with strongly related grids; represents the grid i in the past L fusion data sequence of time steps, used to capture the historical time trend of grid i; L represents the time window length, i.e. the number of historical time steps, used to control the amount of historical data included by the encoder;

[0137] decoder, used to verify the validity of spatio-temporal features and strengthen neighborhood association:

[0138] Add dynamic neighborhood constraints based on reconstruction error to ensure that spatio-temporal features contain temporary associations, and the loss function is:

[0139]

[0140] In the formula: Decoder represents the decoder function; is the neighborhood constraint weight (usually 0.1-0.3).

[0141] In the process of converting multi-element heterogeneous data into spatio-temporal features, Dirichlet prior + pseudo count method is also used to solve the problem of misjudgment of sparse data (such as sparse data at night, remote areas, etc.);

[0142] First, by adding pseudo count (generally 0.1), the probability estimate is smoothed, and joint probability smoothing and marginal probability smoothing are obtained respectively:

[0143] Joint probability smoothing:

[0144] Marginal probability smoothing:

[0145] In the formula: K represents the number of feature categories;

[0146] Then, mutual information is calculated:

[0147] ;

[0148] J(S,D) By introducing pseudo count, the problem of misjudgment of sparse data is solved, and the purpose is to modify real-time mutual information.

[0149] Step S3, Spatiotemporal Coordinated Supply and Demand Matching: The spatiotemporal potential energy field is used to characterize the supply and demand relationship between facilities and the planned area. First, facilities are assigned weights and time decay factors based on their type to obtain the supply potential energy for the corresponding type of facility. Then, the regional demand factor and the spatiotemporal feature encoding from Step S2 are used to obtain the demand potential energy for the corresponding facility. Finally, the supply and demand relationship is obtained through the supply and demand potential energy. Specifically:

[0150] First, assign weights to facilities based on their type. w k and time decay factor T s (k) (For example: hospitals have a higher priority than shopping malls, and hospitals require 24 / 7 response and have slow time decay, while shopping malls have concentrated peak hours and fast time decay; therefore, the facility weight of hospitals can be [missing information].) w k1 =1.5, the time decay factor can be: T s (k 1 ) =24h, while the weight of the mall's facilities can be 24h. w k2 =1, the time decay factor can be: T s (k 2 ) =6h), obtain the supply potential energy of the corresponding type of facility:

[0151]

[0152] In the formula: C k express k The scale of the data type (i.e., facilities); express k Class data (i.e., facilities) in t Real-time utilization rate at any given moment (obtained by the actual amount of resources used and the total amount of available resources). x This represents the spatial location variable, i.e., the target location; x k express k The actual location of the data (i.e., the facility); t This represents a time variable, i.e., the current moment. t k express k Peak service hours for data-related (i.e., facilities); Indicates the spatial attenuation coefficient (obtained from experimental data);

[0153] Through regional demand factors B(x)obtain the demand potential of the corresponding facility (the regional demand factor is defined according to the regionality of the corresponding facility, for example, the facility demand in the central city is higher than that in the suburb, and the regional demand factor of the central city is B(x 1 )=1.2 , and the regional demand factor of the suburb is B (x 2 )=0.8 ):

[0154]

[0155] In the formula, w q , represents the demand weight vector of the corresponding facility (combined with the multi-source characteristics of urban space-time data and the corresponding facility type, the historical demand data of the corresponding facility is obtained through statistics);

[0156] obtain the grid i The supply-demand matching degree at time t : M(i,t)

[0157]

[0158] In the formula, , represents the total amount of punishment coefficient (usually 0.1); t reach (i,k) , represents the space-time reachable time of the grid i to the facility k (available through real-time traffic data); T 0 represents the reachable time threshold.

[0159] Step S4, dynamic demand prediction: obtaining the facility demand prediction at the future time, providing accurate and causally interpretable prediction results for layout optimization; specifically:

[0160] Introducing causal gates in the LSTM (Long Short-Term Memory, Long Short-Term Memory Network) model, thereby distinguishing the causal effects of exogenous variables such as policy release, sudden events, social media heat, and festival days (for example: demand mutation caused by new district planning);

[0161] First, according to the specific type, the causal variables are divided into endogenous variables X end (t) and exogenous variables X exo (t) The endogenous variables include time series dependent variables such as historical demand and people flow, and the exogenous variables include causal intervention variables such as policy release, sudden events, social media heat, and festival days;

[0162] ​Then, the causal gate is introduced in the LSTM model to revise the cell state to incorporate the intervention effect:

[0163] Input gate:

[0164] where: denotes the Sigmiod activation function (output range [0, 1]); W i denotes the input weight, connecting the current time input x t to the input gate (obtained by training parameters); x t denotes the current time input; U i denotes the hidden state weight, connecting the previous time hidden state h t-1 to the input gate (obtained by training parameters); h t-1 denotes the previous time hidden state; V i denotes the cell state weight, connecting the previous time cell state c t-1 to the input gate (obtained by training parameters); c t-1 denotes the previous time cell state (the core carrier of LSTM to store historical information); b i denotes the input gate bias (obtained by training parameters);

[0165] Forget gate:

[0166] where: W f , U f , V f denote the input, hidden state, cell state weight matrix to the forget gate, respectively; b f denotes the forget gate bias;

[0167] Candidate state:

[0168] where: tanh() denotes the Tanh activation function (output range [-1, 1]); W c , U c denote the input, hidden state weight matrix to the candidate state, respectively; b cdenotes the candidate state bias;

[0169] causal correction cell state:

[0170] where: denotes the element-wise multiplication; denotes the causal strength; denotes the intervention-induced cell state change;

[0171] output gate:

[0172] where: W o , U o , V o denote the input, hidden state, current cell state to output gate weight matrices, respectively; b o denotes the forget gate bias;

[0173] output state:

[0174] The causal strength is obtained by an MLP (Multilayer Perceptron) composed of an input layer, two hidden layers and an output layer, fully connected between layers, with ReLU as the activation function:

[0175]

[0176] The causal graph of exogenous variables is constructed using the PC (Peter-Clark) algorithm, and the intervention-induced cell state change is obtained :

[0177]

[0178] where: do() denotes the intervention operator; denotes the causal coefficient obtained by data mining, which is used to measure the marginal influence of variables k on demand (obtained by actual data);

[0179] Finally, the demand prediction at the future time is output:

[0180]

[0181] where: W h denotes the weight matrix from input to output layer; b h denotes the bias of the output layer; This represents the predicted volatility coefficient (obtained from previously measured experimental data).

[0182] Step S5, Facility Layout Optimization: Dynamic optimization of facility layout is achieved using near-end strategy optimization, specifically as follows:

[0183] First, define the state space. St With action space At :

[0184]

[0185] In the formula: { G i (t)} represents dynamically fused multi-source heterogeneous data, reflecting real-time supply and demand relationships, and is output by step S1; { JS k (t) The} indicates the construction progress and constrains the "temporal feasibility" of decision-making, which is obtained based on the actual construction progress; Indicating a future moment The demand forecast is output by step S4; t build Indicates the construction sequence; C Indicates the scale of the facility; C max This indicates the upper limit of resources (i.e., the upper limit of the infrastructure budget for the corresponding region). A O It represents the available operational space, which is the set of all facility layout decisions that meet resource constraints; x Indicates the location of the facility;

[0186] Then, a reward function is set to differentiate the importance of facility areas and achieve dynamic matching:

[0187]

[0188] In the formula: B(x) i This indicates the regional demand factor (used to measure the priority of resource allocation, such as higher priority for central urban areas than suburbs). These represent the corresponding weighting coefficients (obtained from experimental data);

[0189] This indicates the average matching degree when the facilities are completed, expressed as the supply-demand matching degree in step S3. M(i, t) The supply and demand matching degree at the time of completion is obtained by using step S3 to predict. M(i,t+t build ) Then for all grids iThe average matching degree is obtained according to the regional demand factor:

[0190]

[0191] Redundancy (i.e., the spatial and temporal proportion of facility utilization below a threshold) is obtained by first using facilities... k Generate a radius centered on the position coordinates. r k The circular area is designated as the service area corresponding to the facility; subsequently, a causal gate prediction is introduced into the LSTM model in step S4 to predict the first... i The requirements of each grid at the time of construction ; grid g i Geometric range and facilities k Spatial intersection analysis of service areas to obtain the percentage of intersection area IOU(g i ,k) A preset intersection threshold (e.g., 0.5) is set. If the intersection area is greater than the threshold, the requirements are summarized.

[0192]

[0193] Then, the predicted utilization rate is obtained:

[0194]

[0195] In the formula: t yy Indicates the actual operating time of the facility; t yc This indicates the predicted application duration during the facility planning process;

[0196] A preset utilization threshold (set based on facility type and experimental data statistics, e.g., 0.8 for schools, 0.7 for hospitals, and 0.5 for shopping malls) is used to statistically predict the utilization rate. Redundancy is obtained by measuring the proportion of time and space that is less than the utilization threshold.

[0197] Finally, through policy networks Output the placement decision probability, and then iteratively update the network parameters to achieve placement optimization (PPO, Proximal Policy Optimization):

[0198]

[0199] In the formula: Indicates the first n Parameters before the last update; Indicates the learning rate (obtained from experimental data);E denotes taking expectation over the empirical data distribution of "state-action-advantage" (computed by averaging over a batch sampled from the training data); Y t denotes the advantage function, which guides the direction of policy update; clip() denotes the clipping operation; denotes the clipping coefficient (hyperparameter, usually 0.1~0.3);

[0200]

[0201] wherein: Q(St,At) is the action value function (is the cumulative reward expectation of state St for action At ); V(St) is the state value function (is the cumulative reward expectation of state St ); the advantage function guides the direction of policy update: positive advantage encourages the action, and negative advantage inhibits the action.

[0202] Embodiment 2:

[0203] As another preferred embodiment of the technical scheme of the present application, on the basis of the scheme of the above-mentioned embodiment 1, iterative optimization is further performed, that is, through dynamic error feedback, self-adaptive parameter update and phased trigger layout, dynamic adaptation to the city development stage is realized.

[0204] Dynamic error feedback:

[0205] Through the city information model (CIM, City Information Modeling), the actual operation data of the built facilities are collected D real (s) , the error between the actual operation data and the predicted value is obtained:

[0206]

[0207] wherein: denotes the decay coefficient (generally between (0, 1), which is set according to the actual situation);

[0208] Self-adaptive FTRL (Follow The Regularized Leader, follow the regularized leader) update:

[0209]

[0210] wherein: N denotes the cumulative data volume (generally N = t , that is, to the timet the number of samples) ; represents dynamic L1 regularization, the less data (e.g. new area), the stronger regularization; represents dynamic L2 regularization, the more data (i.e. mature urban area), the weaker regularization;

[0211]

[0212] In the formula: , represents the initial regularization coefficient;

[0213] Phased trigger release layout:

[0214] First, principal component analysis (PCA) is used to fuse multi-dimensional indicators to construct a city comprehensive development index; multi-dimensional indicators include population growth rate, GDP growth rate, infrastructure investment rate, and employment rate; each indicator is normalized to [0, 1] and reduced by PCA to obtain its corresponding principal component PC i (t) ; the city comprehensive development index is obtained through the principal component of each indicator ZH(t) :

[0215]

[0216] In the formula: represents the principal component weight coefficient of the i-th indicator (obtained through experimental data) ; i

[0217] Then, the mismatch entropy threshold is obtained through the city comprehensive development index H 0 (t) :

[0218]

[0219] In the formula: H 00 represents the stable development threshold (obtained through experimental data), i.e. the static benchmark of mismatch tolerance;

[0220] The actual mismatch entropy of the grid is obtained H(t) :

[0221]

[0222] In the formula: N w represents the total number of city grids; represents the time decay factor (0 t - t i The greater, the lower the weight of historical mismatch, and the current mismatch is given priority.​

[0223] When the actual mismatch entropy over K consecutive time periods H(t) Greater than the mismatch entropy threshold H 0 (t) time, Triggering layout re-optimization:

[0224]

[0225] In the formula: This indicates that the strategy network in step S5 is adopted; S(t+K) Indicates the updated status; A(t+K) Indicates an update action; This indicates the updated parameters.

[0226] Iterative optimization closed-loop logic:

[0227] Data Acquisition and Error Calculation: Real-time acquisition of required data D real (s) Calculate dynamic error e(t) To mitigate the impact of outdated data;

[0228] Model parameter update: based on e(t) and data volume N Adaptive FTRL adjusts parameters to balance underfitting / overfitting;

[0229] Triggering condition judgment: Based on the city's comprehensive development index ZH(t) Adjust the mismatch entropy threshold H 0 (t) ,continuous K If the limit is exceeded once, a new optimization will be triggered;

[0230] Layout optimization execution: based on the new state S(t+K) and new parameters Generate a new layout scheme A(t+K) ;

[0231] Iterative loop: After the new layout is implemented, repeat the above steps to achieve a closed-loop evolution of "data → model → decision → data".

[0232] Example 3:

[0233] A system for optimizing urban infrastructure configuration includes a data acquisition and preprocessing module, a dynamic correlation feature extraction module, a spatiotemporal collaborative supply and demand matching module, a dynamic demand prediction module, a facility layout optimization module, and an iterative optimization module. Each module adopts the corresponding method in Embodiment 1 and Embodiment 2, respectively.

Claims

1. A method for optimizing allocation of urban space resources based on a city information model, characterized in that: Comprise: Step S1, data acquisition and preprocessing: based on the city information model to obtain static data and dynamic data in the city, and to preprocess the data for spatiotemporal dynamic alignment, obtain multi-element heterogeneous data; Step S2, dynamic correlation feature extraction: convert multi-element heterogeneous data into spatiotemporal features, provide robust and semantically rich inputs for subsequent supply and demand matching, dynamic demand prediction, and layout optimization; Step S3, spatiotemporal collaborative supply and demand matching: use spatiotemporal potential field to depict the supply and demand relationship between facilities and planning areas, i.e., first assign weights and time decay factors to facilities according to facility type, obtain supply potential for corresponding type facilities, then use regional demand factor and spatiotemporal feature coding in step S2 to obtain demand potential for corresponding facilities, and finally obtain supply and demand relationship through supply potential and demand potential; "Use spatiotemporal potential field to depict the supply and demand relationship between facilities and planning areas" specifically refers to: First, a facility is given a weight according to the facility type w k and a time decay factor T s (k) , the supply potential of the corresponding type of facility is obtained: In the formula: C k represents k the size of the class data; represents k the real-time utilization of the class data at the moment; t represents x a spatial position variable, i.e. a target position; x k represents k the actual position of the class data; t represents a time variable, i.e. a current moment; t k represents k the service peak period of the class data; represents a spatial attenuation coefficient; By regional demand factor B(x) Obtain demand potential for corresponding facility: In the formula: w q denotes the demand weight vector of the corresponding facility; z i,t denotes the feature vector output by the encoder in S2 that converts the multi-variate heterogeneous data into spatio-temporal features; Obtaining a grid i At time t Supply-demand matching degree at time M(i,t) : In the formulae: represents the total amount of the penalty coefficient; t reach (i,k) represents the grid i to the facility k spatiotemporal reachable time; T 0 represents the threshold of reachable time; Step S4, dynamic demand prediction: obtain facility demand prediction for future time, provide accurate and causally interpretable prediction results for layout optimization; Step S5, facility layout optimization: use proximal policy optimization to realize dynamic optimization of facility layout.

2. The urban space resource optimization configuration method based on the urban information model according to claim 1, characterized in that: The static data includes infrastructure vector and planning map; the dynamic data includes human flow, traffic, POI activity, historical demand, social media heat, festival calendar, and traffic congestion index.

3. The method of claim 2, wherein the method is based on a city information model. The "preprocessing of spatiotemporal dynamic alignment of data" in step S1 includes time dynamic alignment and spatial adaptive alignment; Time dynamic alignment: for dynamic data, make the timestamps of all dynamic data consistent, eliminate the influence of data collection lag, including: With the traffic congestion index T co (t) As the core of time dynamic alignment, the dynamic lag function is obtained: In the formula: represents k the hysteresis of the data of the same kind at time t ; represents k the historical average hysteresis of the data of the same kind; ; Use dynamic lag function to correct the timestamps of all dynamic data: In the formula: represents the minimum time granularity; t k represents k the original timestamp of the class data; Spatial adaptive alignment: taking the seats of infrastructure vector as data source location s j , meanwhile, by planning the land use property to divide the grid cell g i ; for the aligned dynamic data, according to the data type j Assign its attenuation coefficient , get the spatial weight: In the formulae: IOU represents the spatial overlap ratio; All data sources are weighted by spatial weight to obtain spatio-temporally consistent grid-level fusion data, i.e. multi-element heterogeneous data G i (t) : In the formula: Data j (t) represents the first j class data at time t of the original observation data.

4. The urban space resource optimization configuration method based on the urban information model according to claim 3, characterized in that: In step S2, the dynamic neighborhood spatiotemporal autoencoder is introduced to convert multi-element heterogeneous data into spatiotemporal features, including encoder and decoder: Encoder, used to capture dynamic neighborhood correlation: Based on real-time mutual information I(i,j,t) Screening time t With grid i Strongly correlated neighborhood N(i,t) : In the formula: denotes the time instant t lower full region maximum mutual information; Real-time mutual information is the grid i and j Kullback-Leibler divergence of the joint distribution of the feature sequence within a local time window from the independent distribution, i.e.: In the formula: S and D respectively represent the category label after feature discretization, and are the bridge from continuous features to discrete probability, used to solve the calculation difficulty of mutual information for continuous data; p ij (S,D) represent the grid i is in the category S , the grid j is in the category D ; n ij (S,D) represent the actual observation count of feature pair (S,D) , that is, the number of samples in which the grid i data S and the grid j belongs to D ; p i (S) , p j (D) respectively represent the marginal probability of the grid i being in the category S alone, and the marginal probability of the grid j being in the category D ; n i (S) represent the marginal observation count of feature S , that is, the number of samples in which the grid i is in the category S alone, n j (D) represent the marginal observation count of feature D , that is, the number of samples in which the grid j is in the category D alone; N represent the total amount of samples; Based on dynamic neighborhood N(i,t) Complete spatiotemporal feature encoding: wherein: z i,t represents the feature vector output by the encoder; Encoder represents the encoder function; represents the dynamic neighborhood N(i,t) all grids in j fusion data set at time t, used to capture the grid i real-time interaction with strongly connected grids; represents the grid i in the past L fusion data sequence of past time steps, used to capture the historical temporal trend of grid i; L represents the length of the time window, i.e., the number of historical time steps, used to control the amount of historical data incorporated by the encoder; Decoder, used to verify the validity of spatiotemporal features and strengthen neighborhood correlation: Add dynamic neighborhood constraint based on reconstruction error to ensure that spatiotemporal features contain temporary correlation, and the loss function is: In the formulae: Decoder denotes a decoder function; is a neighborhood constraint weight.

5. The urban space resource optimization configuration method based on the urban information model according to claim 4, characterized in that: The dynamic demand prediction in step S4 specifically refers to: Introduce causal gate in LSTM model to distinguish the causal influence of exogenous variables such as policy release, sudden events, social media heat, and festival; First, we decompose the causal variables into endogenous variables according to specific types X end (t) and exogenous variables X exo (t) Endogenous variables include time-series dependent variables of historical demand and people flow, and exogenous variables include causal intervention variables of policy release, emergencies, social media heat, and festival days. Then, introduce causal gate in LSTM model to correct cell state to include intervention effect: Input gate: wherein: denotes a Sigmiod activation function; W i denotes input weights, connecting the current time input x t to the input gate; x t denotes the current time input; U i denotes hidden state weights, connecting the previous time hidden state h t-1 to the input gate; h t-1 denotes the previous time hidden state; V i denotes cell state weights, connecting the previous time cell state c t-1 to the input gate; c t-1 denotes the previous time cell state; b i denotes the input gate bias; forget gate: wherein: W f , U f , V f respectively represent input, hidden state, cell state to forget gate weight matrices; b f represent forget gate bias; Candidate status: wherein: tanh() denotes a Tanh activation function; W c 、 U c Wi, Wf, Wg, Wkdenote weight matrices from input, hidden state to candidate state, respectively; b c represents a candidate state bias; Causal corrected cell state: wherein: represents an element-wise product; represents a causal strength; represents a change in cell state resulting from an intervention; Output gate: wherein: W o , U o , V o respectively represent input, hidden state, current cell state to output gate weight matrices; b o represents a forget gate bias; Output state: Causal strength is obtained through MLP, which consists of an input layer, two hidden layers, and an output layer, with full connection between layers and ReLU activation function: A PC (Peter-Clark) algorithm was used to construct a causal graph of exogenous variables, obtaining the cell state changes caused by interventions : In the formula: do() represents the intervention operator; represents the causal coefficient obtained by data mining, which is used to measure the marginal influence of the variable k on demand; Finally, output the demand forecast for future time t: In the formula: W h represents a weight matrix input to the output layer; b h represents a bias of the output layer; represents a predicted fluctuation coefficient.

6. The urban space resource optimization configuration method based on the urban information model according to claim 5, characterized in that: The "facility layout optimization" in step S5 specifically refers to: First, define the state space St and action space At : In the formula: G i (t) represents the dynamic fusion of multi-element heterogeneous data, reflecting the real-time supply and demand relationship, output by step S1; JS k (t) represents the construction progress, constraining the "time sequence feasibility" of the decision, obtained according to the actual construction progress; represents the demand prediction for the future time , output by step S4; t build represents the construction time sequence; C represents the facility scale; C max represents the resource upper limit; A O represents the actionable space, that is, the set of all facility layout decisions that meet the resource constraints; Then, set the reward function to distinguish the importance of facility areas and achieve dynamic matching: In the formulae: B(x) i denotes the area demand factor; denotes the corresponding weight factor, respectively. represents the average match at build-out, obtained by the supply-demand match degree in step S3 M(i,t) obtained, i.e. the supply-demand match degree at build-out is predicted using step S3 M(i,t+t build ) Then, for all grids i the average match degree is obtained according to the regional demand factor: represents the redundancy; Finally, the layout decision probability is output by the policy network and the network parameters are iteratively updated to achieve layout optimization. wherein: denotes the parameter before the n denotes the learning rate; E denotes taking the expectation over the empirical data distribution of "state-action-advantage"; Y t denotes the advantage function, which guides the direction of policy update; clip() denotes the clipping operation; denotes the clipping coefficient.​

Images