Intelligent park management method based on digital twinning

Abstract

The application discloses a kind of based on digital twinning wisdom park management method, belong to wisdom park management field, including: for the physical entity in wisdom park, the multi-source heterogeneous dataset of physical entity is collected;Constitute the wisdom park digital twinning model based on space-time constraint;Constitute the space-time synchronization and two-way mapping mechanism of physical-wisdom park digital twinning model based on adaptive Kalman filtering, realize the real-time correction and two-way linkage of wisdom park digital twinning model state;Extract the time series data and spatial data corresponding to each physical entity, output the global fusion features of each physical entity;The topological graph structure of wisdom park is constructed, and the situation score of each management dimension of wisdom park is output through fully connected layer, and the management situation of wisdom park is evaluated.The application realizes the full-dimensional digital mapping of wisdom park physical entity from geometric form, physical mechanism, behavior law to management rule by constructing four-layer coupled digital twinning body model.

Description

Technical Field

[0001] This invention relates to the field of smart park management, and specifically to a smart park management method based on digital twins. Background Technology

[0002] With the rapid development of the digital economy, smart parks have become the core carrier of urban digital transformation. Combining smart park management with digital twin technology can effectively reduce the management difficulty of smart parks, improve management efficiency, and achieve management visualization.

[0003] In existing technologies, digital twin models are often one-dimensional, remaining at the level of three-dimensional visualization and only achieving static mapping of geometric shapes. They lack multi-level coupling of physical mechanisms, behavioral patterns, and management rules, resulting in low matching between the model and physical entities and an inability to simulate the dynamic operation of the park. The spatiotemporal synchronization accuracy between physical entities and virtual twins is insufficient. Traditional synchronization methods only achieve simple data forwarding without considering the impact of transmission delay, measurement noise, and model errors. In dynamic scenarios, the synchronization delay is high and the state deviation is large, making it impossible to achieve accurate bidirectional mapping and linkage control. The ability to fuse multi-source heterogeneous data in the park is weak, failing to fully explore spatiotemporal and cross-modal correlation features, resulting in severe loss of key features and low reliability of fusion results. Finally, it is impossible to achieve multi-dimensional, full-scenario comprehensive situational assessment. Summary of the Invention

[0004] In view of the above-mentioned shortcomings in the prior art, the present invention provides a smart park management method based on digital twins, which solves the defects of low accuracy of digital twin models, poor spatiotemporal synchronization, and insufficient fusion of multi-source data in the prior art.

[0005] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A smart park management method based on digital twins is provided, which includes: Step S1: For physical entities within the smart park, collect full-dimensional data of the physical entities and perform adaptive preprocessing to construct a multi-source heterogeneous dataset containing static basic data, dynamic sensing time-series data, and external environment data; Step S2: Construct a smart park digital twin model based on spatiotemporal constraints, including a four-layer coupled model structure of geometric model, physical model, behavioral model and rule model, and realize the unified spatiotemporal reference linkage of the four-layer coupled model structure through spatiotemporal coupling operators; Step S3: Construct a spatiotemporal synchronization and bidirectional mapping mechanism for the physical-smart park digital twin model based on adaptive Kalman filtering. The mapping relationship between physical entities and the smart park digital twin model is described through a state space model. The noise covariance matrix is ​​adaptively updated based on the innovation sequence to realize real-time correction and bidirectional linkage of the state of the smart park digital twin model. Step S4: Extract the temporal and spatial data corresponding to each physical entity from the multi-source heterogeneous dataset, and use a convolutional neural network to extract the original temporal features from the temporal data and the original spatial features from the spatial data respectively; then use a multi-head attention mechanism to fuse the original temporal features and the original spatial features to output the global fused features of each physical entity. Step S5: Construct the topology of the smart park, take the global fusion features of each physical entity as input, capture the spatial correlation features and temporal evolution features between physical entities through Chebyshev spectral graph convolution and gated causal convolution, and then output the status scores of each management dimension of the smart park through a fully connected layer to evaluate the management status of the smart park.

[0006] Furthermore, step S1, which involves collecting full-dimensional data of the physical entity and performing adaptive preprocessing, specifically includes: Time alignment of dynamically sensed time-series data is performed by resampling the data at different sampling frequencies using cubic spline interpolation, and a unified sampling time step is achieved. ; ; in, This is time-aligned time-series data. i Types of time series data t For time, , l For interpolation interval index, This represents the start time of the interpolation interval. The end time of the interpolation interval. These are the cubic spline interpolation coefficients.

[0007] Furthermore, the expression for the digital twin model of the smart park is: ; in, u For the physical entity's identifier, U The number of physical entities, For time t The digital twin model of the smart park These are respectively geometric model, physical model, behavioral model, and rule model. It is a spatiotemporal coupling operator. For the first u A physical entity in time t Geometric model, physical model, behavioral model, rule model, For coupling weights.

[0008] Furthermore, geometric model , VFor the set of vertices, E Let be the set of edges. F For a set of faces, S For a collection of entities, A collection of spatiotemporal attributes; Physical Model , P A set of physical properties C For a set of physical coupling constraints, For state space, For physical mechanism equations, For time-dimensional attributes; Behavioral Model Based on finite state machine construction, For a finite set of states, For triggering event sets, This is the state transition function. This is the initial state. For the set of terminating states, It is a set of time-series constraints; rule model , For a set of rules, L For a set of logical constraints, G For the target constraint set, It is a set of spatiotemporal constraints.

[0009] Further, step S3 includes: Step S31: Construct the true state vector of the physical entity based on the multi-source heterogeneous dataset, and construct the state space model of the physical entity-virtual mapping based on the true state vector, including the state prediction equation and the measurement update equation: ; in, For time t The true state vector of a physical entity For time t The prior state prediction value of the digital twin model of the smart park. Here is the state transition matrix. To control the input matrix, To control the input vector, This is the process noise vector. , For process noise covariance, For the measurement value vector, For the measurement matrix, To measure the noise vector, , To measure the noise covariance; Step S32: Execute the adaptive Kalman filtering process to perform prior covariance prediction, Kalman gain calculation, posterior state update, and posterior covariance update, so as to realize the real-time correction of the state of the digital twin model of the smart park. Prior covariance prediction: ; Kalman gain calculation: ; Post-hoc state update: ; Posterior covariance update: ; in, Let be the prior state covariance matrix. Let be the posterior state covariance matrix. Here is the Kalman gain matrix. It is the identity matrix; Step S33: Perform adaptive noise update based on the innovation sequence and adjust the process noise covariance in real time. Covariance of measurement noise To adapt to the dynamically changing noise environment of the smart park; Measurement noise covariance update: ; Process noise covariance update: ; in, The actual covariance matrix of the new information; Based on the measurement value vector and prior state prediction value Calculate the new information sequence Then based on length L The actual covariance matrix of the innovation sequence within the sliding window is obtained. ; ; in, a The time within the sliding window, For time a New information sequence ; Step S34: Achieve a positive mapping between the physical entity and the digital twin model of the smart park through posterior state updates, by controlling the input vector. Control commands from the digital twin model of the smart park are sent to the physical entity, enabling reverse control from virtual to physical.

[0010] Further, step S4 includes: Step S41: Extract the temporal and spatial data corresponding to each physical entity from the multi-source heterogeneous dataset, and use a one-dimensional convolutional neural network (1D-CNN) to extract the original temporal features of the temporal data. Two-dimensional convolutional neural networks (2D-CNN) are used to extract the original spatial features of spatial data. ; Step S42: Process the original spatial features using the spatial attention module. Perform weighted analysis and output the weighted spatial features. ; ; in, Position in space Attention weights This indicates element-wise multiplication. It is the Sigmoid activation function. For convolution operations, For average pooling, For max pooling; Step S43: Process the original temporal features using the temporal attention module. Weighting is applied to highlight the characteristics of key time points, and the weighted time series features are output. ; ; in, For time step k Attention weights This is the weight matrix. For bias vectors, The hyperbolic tangent activation function is used. For classification activation function; Step S44: Weighted spatial features Weighted time series features Each feature is mapped to the same latent feature space through a linear transformation, and the transformed spatial features are output. and time characteristics ; in, These are the projection matrices of linear transformations of spatial features and temporal features, respectively. These are the projection matrices and bias vectors of the linear transformations of spatial and temporal features, respectively. Step S45: Use time series features as a query Spatial features as keys K Sum V Calculate the single-head spatiotemporal cross attention weights; ; in, For key K vector dimension, This is a single-head spatiotemporal cross-attention operation; Step S46: Calculate the multi-head parallel cross-attention weights based on the single-head spatiotemporal cross-attention weights, and output the global fusion features of each physical entity. ; ; in, H For the number of attention heads, h Number the attention head. For the first h Head-spatial-temporal cross attention weights For multi-head parallel cross-attention computation, To output the concatenation function, For the fusion matrix, These are the query weight matrix, key weight matrix, and value weight matrix, respectively. This is the layer normalization function.

[0011] Further, step S5 includes: Step S51: Obtain physical entities as nodes from the smart park digital twin model to form a node set. Obtain the set of edges between physical entities from the geometric model. And construct an adjacency matrix based on whether physical entities are associated. ; Step S52: Construct the topology of the smart park Adjacency matrix Convert to normalized Laplace matrix ; ; in, D Let be the degree matrix of the nodes. It is an adjacency matrix with self-loops; Step S53: Globally fuse the features of each physical entity The value is assigned to each node in each topological graph structure as the initial feature input to the 0th layer of the spatiotemporal graph convolution. ; Step S54: Apply Chebyshev spectral convolution from the input... Capturing spatial relationships between physical entities ; ; in, for J Chebyshev polynomials Number the order of the Chebyshev polynomial. J Let the order be the Chebyshev polynomial. The kernel weights for Chebyshev spectral graph convolutions; Step S55: Use gated causal convolution to process the input... Capture temporal evolution patterns and output temporal evolution features. ; ; in, This represents the causal convolution operation. The weights of the convolution kernel for gated causal convolution. The kernel bias for gated causal convolution; Step S56: Transform the temporal evolution features Spatial correlation features By splicing the data, we obtain the spatiotemporal joint features. ; ; Step S57: Combine spatiotemporal features Input the fully connected layer, and output the status scores of each management dimension of the smart park through the fully connected layer. Evaluate the management status of the smart park based on the status score thresholds of different management dimensions. ; in, This represents the fully connected layer operation. For the situational scoring vector of the management dimension, The status scores are based on the dimensions of security, energy consumption, operation and maintenance, and service management.

[0012] The beneficial effects of this invention are as follows: By constructing a four-layer coupled digital twin model, this invention achieves full-dimensional digital mapping of the physical entities of smart parks, from geometric shape, physical mechanism, behavioral patterns to management rules; through the bidirectional mapping mechanism of adaptive Kalman filtering, it realizes real-time and accurate linkage between the physical park and the virtual twin, providing real-time assurance for closed-loop management and control; by improving the spatiotemporal attention fusion network, it fully explores the spatiotemporal correlation and cross-modal correlation features of multi-source heterogeneous data in smart parks, providing highly reliable feature support for situational awareness and decision-making; and through the spatiotemporal graph convolutional network, it realizes multi-dimensional comprehensive assessment of the situation of smart parks. Attached Figure Description

[0013] Figure 1 This is a flowchart of a smart park management method based on digital twins. Detailed Implementation

[0014] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0015] like Figure 1 As shown, a smart park management method based on digital twins includes: Step S1: For physical entities within the smart park, collect full-dimensional data of the physical entities and perform adaptive preprocessing to construct a multi-source heterogeneous dataset containing static basic data, dynamic sensing time-series data, and external environment data.

[0016] The full-dimensional data includes static basic data, dynamic sensing time-series data, and external environment data. In this embodiment, static basic data includes park BIM model, GIS geographic information, building structure parameters, equipment ledger, pipeline layout, etc.; dynamic sensing time-series data includes real-time data collected from IoT sensors, video surveillance, access control gates, parking lots, elevators, fire-fighting equipment, etc.; and external environment data includes meteorological, traffic, and other data.

[0017] Time alignment of dynamically sensed time-series data is performed by resampling the data at different sampling frequencies using cubic spline interpolation, and a unified sampling time step is achieved. ; ; in, This is time-aligned time-series data. i Types of time series data t For time, , l For interpolation interval index, This represents the start time of the interpolation interval. The end time of the interpolation interval. These are the cubic spline interpolation coefficients; The seven-parameter coordinate transformation method is used to transform the BIM local coordinate system, GIS geodetic coordinate system, and sensor local coordinate system corresponding to multi-source heterogeneous data into the geocentric rectangular coordinate system of the smart park. ; in, The coordinates are in the geocentric rectangular coordinate system. The original coordinates of the multi-source heterogeneous data. Let be a rotation matrix. It is a translation matrix.

[0018] Step S2: Construct a smart park digital twin model based on spatiotemporal constraints, including a four-layer coupled model structure of geometric model, physical model, behavioral model and rule model, and realize the unified spatiotemporal reference linkage of the four-layer coupled model structure through spatiotemporal coupling operators.

[0019] The expression for the digital twin model of the smart park is: ; in, u For the physical entity's identifier, U The number of physical entities, For time t The digital twin model of the smart park These are respectively geometric model, physical model, behavioral model, and rule model. It is a spatiotemporal coupling operator. For the first u A physical entity in time t Geometric model, physical model, behavioral model, rule model, For coupling weights; Geometric model , V For the set of vertices, E Let be the set of edges. F For a set of faces, S For a collection of entities, It is a set of spatiotemporal attributes; a three-dimensional geometric model of all elements of the smart park is constructed based on the boundary representation method to achieve millimeter-level spatial morphology mapping; Physical Model , P A set of physical properties C For a set of physical coupling constraints, For state space, For physical mechanism equations, The time dimension is an attribute; the physical model assigns physical attributes and operating mechanism equations to each geometric physical entity, simulating the real physical characteristics of the physical entity, such as energy consumption, thermodynamics, fluid mechanics, circuits and other mechanisms. Behavioral Model Based on finite state machine construction, For a finite set of states, For triggering event sets, This is the state transition function. This is the initial state. For the set of terminating states, It is a set of temporal constraints; a dynamic behavior model of physical entities is constructed based on finite state machines to describe the state transitions and behavioral responses of physical entities under different triggering conditions; rule model , For a set of rules, L For a set of logical constraints, G For the target constraint set, It is a set of spatiotemporal constraints; it constructs a smart park full-dimensional management rule system, including rules for compliance, security, operation and maintenance, emergency response, energy consumption, etc., to realize the digital mapping of management rules.

[0020] Digital twin models are the core carriers of smart park management. This invention breaks through the limitations of traditional static visualization models and constructs a four-layer coupled digital twin model of geometry, physics, behavior and rules. Through spatiotemporal coupling operators, the four-layer model is linked under a unified spatiotemporal benchmark, realizing the digital mapping of the park's physical entities in all dimensions and throughout their entire life cycle.

[0021] Step S3: Construct a spatiotemporal synchronization and bidirectional mapping mechanism for the physical-smart park digital twin model based on adaptive Kalman filtering. The mapping relationship between physical entities and the smart park digital twin model is described through a state space model, and the noise covariance matrix is ​​adaptively updated based on the innovation sequence to realize real-time correction and bidirectional linkage of the state of the smart park digital twin model.

[0022] Step S3 specifically includes: Step S31: Construct the true state vector of the physical entity based on the multi-source heterogeneous dataset, and construct the state space model of the physical entity-virtual mapping based on the true state vector, including the state prediction equation and the measurement update equation: ; in, For time t The true state vector of a physical entity For time t The prior state prediction value of the digital twin model of the smart park. Here is the state transition matrix. To control the input matrix, To control the input vector, This is the process noise vector. , For process noise covariance, For the measurement value vector, For the measurement matrix, To measure the noise vector, , To measure the noise covariance; Step S32: Execute the adaptive Kalman filtering process to perform prior covariance prediction, Kalman gain calculation, posterior state update, and posterior covariance update, so as to realize the real-time correction of the state of the digital twin model of the smart park. Prior covariance prediction: ; Kalman gain calculation: ; Post-hoc state update: ; Posterior covariance update: ; in, Let be the prior state covariance matrix. Let be the posterior state covariance matrix. Here is the Kalman gain matrix. It is the identity matrix; Step S33: Perform adaptive noise update based on the innovation sequence and adjust the process noise covariance in real time. Covariance of measurement noise To adapt to the dynamically changing noise environment of the smart park; Measurement noise covariance update: ; Process noise covariance update: ; in, The actual covariance matrix of the new information; Based on the measurement value vector and prior state prediction value Calculate the new information sequence Then based on length L The actual covariance matrix of the innovation sequence within the sliding window is obtained. ; ; in, a The time within the sliding window, For time a New information sequence ; New sequence This is used to characterize the error between the actual measured values ​​of physical entities and the predicted values ​​of the smart park digital twin model, reflecting the accuracy of the smart park digital twin model's predictions, whether there are abrupt changes in the physical system, and the level of noise. It is expressed through the actual covariance matrix of the innovation. For a recent period of time (length) L The covariance of the true, actual, and online estimated innovation sequence is obtained by statistical averaging within a sliding window; this covariance is used to adaptively correct the noise covariance during the process. Covariance of measurement noise .

[0023] Step S34: Achieve a positive mapping between the physical entity and the digital twin model of the smart park through posterior state updates, by controlling the input vector. Control commands from the digital twin model of the smart park are sent to the physical entity, enabling reverse control from virtual to physical.

[0024] Spatiotemporal synchronization is the core of achieving bidirectional interaction between digital twin models and physical entities. This invention describes the mapping relationship between the two through a state-space model. Addressing the shortcomings of traditional Kalman filtering with its fixed noise covariance, it adaptively updates the noise matrix based on the innovation sequence, achieving high-precision real-time synchronization and bidirectional mapping in dynamic environments. This solves the problems of spatiotemporal asynchrony, low mapping accuracy, and weak bidirectional interaction capability of digital twin models in existing technologies.

[0025] Step S4: Extract the temporal and spatial data corresponding to each physical entity from the multi-source heterogeneous dataset, and use a convolutional neural network to extract the original temporal features from the temporal data and the original spatial features from the spatial data respectively; then use a multi-head attention mechanism to fuse the original temporal features and the original spatial features to output the global fused features of each physical entity.

[0026] The multi-source heterogeneous data of smart parks have strong heterogeneity and strong spatiotemporal correlation. Traditional fusion methods cannot fully explore the cross-modal and cross-spatiotemporal correlation features. This invention constructs a spatiotemporal attention fusion network, which highlights key features through spatial and temporal attention modules and achieves cross-modal fusion through multi-head cross attention, resulting in highly discriminative global fusion features.

[0027] Step S4 specifically includes: Step S41: Extract the temporal and spatial data corresponding to each physical entity from the multi-source heterogeneous dataset, and use a one-dimensional convolutional neural network (1D-CNN) to extract the original temporal features of the temporal data. Two-dimensional convolutional neural networks (2D-CNN) are used to extract the original spatial features of spatial data. ; Step S42: Process the original spatial features using the spatial attention module. Weighting is applied to highlight the characteristics of key areas in the smart park (security priority areas, high-energy-consumption areas, and densely populated areas), and the weighted spatial characteristics are output. ; ; in, Position in space Attention weights This indicates element-wise multiplication. It is the Sigmoid activation function. For convolution operations, For average pooling, For max pooling; Step S43: Process the original temporal features using the temporal attention module. Weighting is applied to highlight the characteristics of key time nodes (peak hours, equipment start-up and shutdown, emergency events), capture long-distance temporal dependencies, and output the weighted temporal features. ; ; in, For time step k Attention weights This is the weight matrix. For bias vectors, The hyperbolic tangent activation function is used. For classification activation function; Step S44: Weighted spatial features Weighted time series features Each feature is mapped to the same latent feature space through a linear transformation, and the transformed spatial features are output. and time characteristics ; in, These are the projection matrices of linear transformations of spatial features and temporal features, respectively. These are the projection matrices and bias vectors of the linear transformations of spatial and temporal features, respectively. Output spatial features and time characteristics The dimensions are the same, eliminating the dimensional differences between modalities and preparing for cross-attention.

[0028] Step S45: Use time series features as a query Spatial features as keys K Sum V Calculate the single-head spatiotemporal cross attention weights; ; in, For key K vector dimension, This is a single-head spatiotemporal cross-attention operation; At each time step k It dynamically focuses on the most relevant regions in space, enabling joint modeling of time dependence and spatial structure.

[0029] Step S46: Calculate the multi-head parallel cross-attention weights based on the single-head spatiotemporal cross-attention weights, and output the global fusion features of each physical entity. ; ; in, H For the number of attention heads, h Number the attention head. For the first h Head-spatial-temporal cross attention weights For multi-head parallel cross-attention computation, To output the concatenation function, For the fusion matrix, These are the query weight matrix, key weight matrix, and value weight matrix, respectively. This is the layer normalization function.

[0030] After multi-head spatiotemporal cross-attention operation, a residual connection structure is introduced to add the linearly mapped temporal features and attention output features element by element, preserving the original temporal feature information to avoid feature degradation. Subsequently, the residual output features are normalized to zero mean and unit variance through layer normalization to eliminate computational instability caused by feature numerical differences, and finally outputs robust and numerically normalized spatiotemporal fusion features.

[0031] The cross-modal feature fusion step uses spatial attention-weighted spatial features Temporal features weighted by time attention For direct input, a linear transformation is first used to align the dimensions of bimodal features. Then, a query, key, and value matrix of spatiotemporal cross-attention is constructed. A multi-head cross-attention mechanism is used to dynamically focus on the spatial structure of the time series, achieving deep interaction and global fusion of spatial and temporal features. Finally, a unified spatiotemporal fused feature is output as the input basis for subsequent smart park situational awareness. Specifically, spatial attention and temporal attention enhance key features within a single modality, while cross-modal fusion enables correlation modeling between the two modalities. The two are cascaded to form a complete spatiotemporal feature extraction and fusion system.

[0032] Step S5: Construct the topology of the smart park, take the global fusion features of each physical entity as input, capture the spatial correlation features and temporal evolution features between physical entities through Chebyshev spectral graph convolution and gated causal convolution, and then output the status scores of each management dimension of the smart park through a fully connected layer to evaluate the management status of the smart park.

[0033] Step S5 specifically includes: Step S51: Obtain physical entities as nodes from the smart park digital twin model to form a node set. Obtain the set of edges between physical entities from the geometric model. And construct an adjacency matrix based on whether physical entities are associated. ; edge set The adjacency matrix integrates the relationships between physical entities, including adjacency, power supply, pipeline network, communication, and control. This indicates the association between two nodes; it is set to 0 if there is no association.

[0034] Step S52: Construct the topology of the smart park Adjacency matrix Convert to normalized Laplace matrix ; ; in, D Let be the degree matrix of the nodes. It is an adjacency matrix with self-loops; Step S53: Globally fuse the features of each physical entity The value is assigned to each node in each topological graph structure as the initial feature input to the 0th layer of the spatiotemporal graph convolution. ; Step S54: Apply Chebyshev spectral convolution from the input... Capturing spatial relationships between physical entities ; ; in, for J Chebyshev polynomials Number the order of the Chebyshev polynomial. J Let the order be the Chebyshev polynomial. The kernel weights for Chebyshev spectral graph convolutions; This embodiment captures the spatial relationships, dependencies, and transmission patterns between physical entities within a smart park through Chebyshev spectral graph convolution, and learns spatial relationship patterns such as energy consumption transmission, security linkage, equipment cluster status, and regional pedestrian flow heat.

[0035] Step S55: Use gated causal convolution to process the input... Capture temporal evolution patterns and output temporal evolution features. ; ; in, This represents the causal convolution operation. The weights of the convolution kernel for gated causal convolution. The kernel bias for gated causal convolution; Step S56: Transform the temporal evolution features Spatial correlation features By splicing the data, we obtain the spatiotemporal joint features. ; ; Step S57: Combine spatiotemporal features Input the fully connected layer, and output the status scores of each management dimension of the smart park through the fully connected layer. Evaluate the management status of the smart park based on the status score thresholds of different management dimensions. ; in, This represents the fully connected layer operation. For the situational scoring vector of the management dimension, The status scores are based on the dimensions of security, energy consumption, operation and maintenance, and service management.

[0036] By using a fully connected layer, the high-dimensional spatiotemporal joint features output by the spatiotemporal graph convolutional network are mapped to a low-dimensional multi-dimensional situational assessment of the smart park, thus completing the spatiotemporal joint feature mapping. Nonlinear transformation to decision indicators.

[0037] This embodiment can evaluate management in different dimensions by setting corresponding situational assessment thresholds. For example, if the situational assessment score of the energy consumption management dimension exceeds the set situational assessment threshold, it means that the energy consumption in the smart park under current management is too high and the energy consumption of each physical entity in the smart park needs to be optimized.

[0038] There are complex spatial relationships and temporal evolution patterns among entities in a smart park. Traditional convolutional networks cannot handle the topological relationships in non-Euclidean spaces. This invention constructs a spatiotemporal graph convolutional network to capture both the spatial relationships and temporal evolution patterns of entities, thereby enabling a multi-dimensional assessment of the operational status of the smart park.

Claims

1. A smart park management method based on digital twins, characterized in that, include: Step S1: For physical entities within the smart park, collect full-dimensional data of the physical entities and perform adaptive preprocessing to construct a multi-source heterogeneous dataset containing static basic data, dynamic sensing time-series data, and external environment data; Step S2: Construct a smart park digital twin model based on spatiotemporal constraints, including a four-layer coupled model structure of geometric model, physical model, behavioral model and rule model, and realize the unified spatiotemporal reference linkage of the four-layer coupled model structure through spatiotemporal coupling operators; Step S3: Construct a spatiotemporal synchronization and bidirectional mapping mechanism for the physical-smart park digital twin model based on adaptive Kalman filtering. The mapping relationship between physical entities and the smart park digital twin model is described through a state space model. The noise covariance matrix is ​​adaptively updated based on the innovation sequence to realize real-time correction and bidirectional linkage of the state of the smart park digital twin model. Step S4: Extract the temporal and spatial data corresponding to each physical entity from the multi-source heterogeneous dataset, and use a convolutional neural network to extract the original temporal features from the temporal data and the original spatial features from the spatial data respectively; Then, a multi-head attention mechanism is used to fuse the original temporal features and the original spatial features, and output the global fused features of each physical entity. Step S5: Construct the topology of the smart park, take the global fusion features of each physical entity as input, capture the spatial correlation features and temporal evolution features between physical entities through Chebyshev spectral graph convolution and gated causal convolution, and then output the status scores of each management dimension of the smart park through a fully connected layer to evaluate the management status of the smart park.

2. The smart park management method based on digital twins according to claim 1, characterized in that, The step S1, which involves collecting full-dimensional data of the physical entity and performing adaptive preprocessing, specifically includes: Time alignment of dynamically sensed time-series data is performed by resampling the data at different sampling frequencies using cubic spline interpolation, and a unified sampling time step is achieved. ; ； in, This is time-aligned time-series data. i Types of time series data t For time, , l For interpolation interval index, This represents the start time of the interpolation interval. The end time of the interpolation interval. These are the cubic spline interpolation coefficients.

3. The smart park management method based on digital twins according to claim 1, characterized in that, The expression for the digital twin model of the smart park is: ； in, u For the physical entity's identifier, U The number of physical entities, For time t The digital twin model of the smart park These are respectively geometric model, physical model, behavioral model, and rule model. It is a spatiotemporal coupling operator. For the first u A physical entity in time t Geometric model, physical model, behavioral model, rule model, For coupling weights.

4. The smart park management method based on digital twins according to claim 3, characterized in that, The geometric model , V For the set of vertices, E Let be the set of edges. F For a set of faces, S For a collection of entities, A collection of spatiotemporal attributes; The physical model , P A set of physical properties C For a set of physical coupling constraints, For state space, For physical mechanism equations, For time-dimensional attributes; Behavioral Model Based on finite state machine construction, For a finite set of states, For triggering event sets, This is the state transition function. This is the initial state. For the set of terminating states, It is a set of time-series constraints; rule model , For a set of rules, L For a set of logical constraints, G For the target constraint set, It is a set of spatiotemporal constraints.

5. The smart park management method based on digital twins according to claim 1, characterized in that, Step S3 includes: Step S31: Construct the true state vector of the physical entity based on the multi-source heterogeneous dataset, and construct the state space model of the physical entity-virtual mapping based on the true state vector, including the state prediction equation and the measurement update equation: ； in, For time t The true state vector of a physical entity For time t The prior state prediction value of the digital twin model of the smart park. Here is the state transition matrix. To control the input matrix, To control the input vector, This is the process noise vector. , For process noise covariance, For the measurement value vector, For the measurement matrix, To measure the noise vector, , To measure the noise covariance; Step S32: Execute the adaptive Kalman filtering process to perform prior covariance prediction, Kalman gain calculation, posterior state update, and posterior covariance update, so as to realize the real-time correction of the state of the digital twin model of the smart park. Prior covariance prediction: ; Kalman gain calculation: ; Post-hoc state update: ; Posterior covariance update: ; in, Let be the prior state covariance matrix. Let be the posterior state covariance matrix. Here is the Kalman gain matrix. It is the identity matrix; Step S33: Perform adaptive noise update based on the innovation sequence and adjust the process noise covariance in real time. Covariance of measurement noise To adapt to the dynamically changing noise environment of the smart park; Measurement noise covariance update: ; Process noise covariance update: ; in, The actual covariance matrix of the new information; Based on the measurement value vector and prior state prediction value Calculate the new information sequence Then based on length L The actual covariance matrix of the innovation sequence within the sliding window is obtained. ; ； in, a The time within the sliding window, For time a New information sequence ; Step S34: Achieve a positive mapping between the physical entity and the digital twin model of the smart park through posterior state updates, by controlling the input vector. Control commands from the digital twin model of the smart park are sent to the physical entity, enabling reverse control from virtual to physical.

6. The smart park management method based on digital twins according to claim 1, characterized in that, Step S4 includes: Step S41: Extract the temporal and spatial data corresponding to each physical entity from the multi-source heterogeneous dataset, and use a one-dimensional convolutional neural network (1D-CNN) to extract the original temporal features of the temporal data. Two-dimensional convolutional neural networks (2D-CNN) are used to extract the original spatial features of spatial data. ; Step S42: Process the original spatial features using the spatial attention module. Perform weighted analysis and output the weighted spatial features. ; ； in, Position in space Attention weights This indicates element-wise multiplication. It is the Sigmoid activation function. For convolution operations, For average pooling, For max pooling; Step S43: Process the original temporal features using the temporal attention module. Weighting is applied to highlight the characteristics of key time points, and the weighted time series features are output. ; ； in, For time step k Attention weights This is the weight matrix. For bias vectors, The hyperbolic tangent activation function is used. For classification activation function; Step S44: Weighted spatial features Weighted time series features Each feature is mapped to the same latent feature space through a linear transformation, and the transformed spatial features are output. and time characteristics ； in, These are the projection matrices of linear transformations of spatial features and temporal features, respectively. These are the projection matrices and bias vectors of the linear transformations of spatial and temporal features, respectively. Step S45: Use time series features as a query Spatial features as keys K Sum V Calculate the single-head spatiotemporal cross attention weights; ； in, For key K vector dimension, This is a single-head spatiotemporal cross-attention operation; Step S46: Calculate the multi-head parallel cross-attention weights based on the single-head spatiotemporal cross-attention weights, and output the global fusion features of each physical entity. ; ； in, H For the number of attention heads, h Number the attention head. For the first h Head-spatial-temporal cross attention weights For multi-head parallel cross-attention computation, To output the concatenation function, For the fusion matrix, These are the query weight matrix, key weight matrix, and value weight matrix, respectively. This is the layer normalization function.

7. The smart park management method based on digital twins according to claim 6, characterized in that, Step S5 includes: Step S51: Obtain physical entities as nodes from the smart park digital twin model to form a node set. Obtain the set of edges between physical entities from the geometric model. And construct an adjacency matrix based on whether physical entities are associated. ; Step S52: Construct the topology of the smart park Adjacency matrix Convert to normalized Laplace matrix ; ； in, D Let be the degree matrix of the nodes. It is an adjacency matrix with self-loops; Step S53: Globally fuse the features of each physical entity The value is assigned to each node in each topological graph structure as the initial feature input to the 0th layer of the spatiotemporal graph convolution. ; Step S54: Apply Chebyshev spectral convolution from the input... Capturing spatial relationships between physical entities ; ； in, for J Chebyshev polynomials Number the order of the Chebyshev polynomial. J Let the order be the Chebyshev polynomial. The kernel weights for Chebyshev spectral graph convolutions; Step S55: Use gated causal convolution to process the input... Capture temporal evolution patterns and output temporal evolution features. ; ； in, This represents the causal convolution operation. The weights of the convolution kernel for gated causal convolution. The kernel bias for gated causal convolution; Step S56: Transform the temporal evolution features Spatial correlation features By splicing the data, we obtain the spatiotemporal joint features. ; ； Step S57: Combine spatiotemporal features Input the fully connected layer, and output the status scores of each management dimension of the smart park through the fully connected layer. Evaluate the management status of the smart park based on the status score thresholds of different management dimensions. ； in, This represents the fully connected layer operation. For the situational scoring vector of the management dimension, The status scores are based on the dimensions of security, energy consumption, operation and maintenance, and service management.

Images