A parking lot scheduling optimization method based on dynamic high-order graph network
Through dynamic high-order graph network processing of real-time parking lot data, optimizing resource scheduling between multiple commercial parking lots, solving the problem of insufficient management and coordination in the existing technology, and improving the operational efficiency and emergency response capabilities of parking lots.
Patent Information
- Application Number
- CN202410751130.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-12
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2044-06-12
AI Technical Summary
The overall coordination and management of the prior art between multiple adjacent commercial parking lots has insufficient results, resulting in waste of resources, unstable operational efficiency and insufficient emergency response capabilities, affecting traffic flow and safety guarantees.
Adopting the parking lot scheduling optimization method based on dynamic high-order graph network, by building an efficient resource scheduling system, using real-time data feature tensors for one-dimensional convolution processing, sliding window gated time convolution and high-order differential spatial convolution, Generate optimized data to enable resource prediction and scheduling decisions.
It realizes optimal management of traffic flow between multiple commercial parking lots, reduces resource waste, improves parking experience and emergency response speed, and ensures the stable operation and safety of parking lots.
Smart Images

Figure CN118313522B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of parking lot scheduling optimization, and in particular to a parking lot scheduling optimization method based on a dynamic high-order graph network. Background Art
[0002] In the urban environment, commercial parking lots, as key facilities, play a vital role in maintaining the smoothness of urban traffic. With the acceleration of urbanization and the continuous growth of car ownership, commercial parking lots are facing increasing pressure from traffic flow. Especially for multiple commercial parking lots adjacent to each other, effective resource scheduling and management is an increasingly prominent challenge.
[0003] At present, most parking lot management technologies focus on the management and scheduling of a single parking lot, but pay insufficient attention to the overall coordination of multiple parking lots. These technologies often ignore the important impact of time and space factors on parking lot management. This limitation may lead to resource waste, unstable parking lot operation efficiency, and even affect emergency response and safety when predicting future demand, responding to peak periods or abnormal situations.
[0004] Therefore, developing a new technology to comprehensively handle demand forecasting, resource scheduling, daily operations, emergency response, and safety configuration issues for multiple adjacent commercial parking lots is crucial to improving the efficiency and effectiveness of commercial parking lot management. Summary of the invention
[0005] In view of the shortcomings of the prior art, the present invention provides a parking lot scheduling optimization method based on a dynamic high-order graph network, which ensures the stable operation of the parking lot and improves the efficiency of emergency support and security configuration by constructing an efficient parking lot resource scheduling system.
[0006] To achieve the above object, the present invention provides a parking lot scheduling optimization method based on a dynamic high-order graph network, comprising:
[0007] Using the parking lot scheduling real-time data, a parking lot scheduling real-time data feature tensor is obtained;
[0008] Obtaining parking lot scheduling optimization data based on the parking lot scheduling real-time data feature tensor based on a dynamic high-order graph network;
[0009] The parking lot scheduling optimization data is used to obtain a parking lot scheduling optimization result.
[0010] Preferably, the method of obtaining a parking lot scheduling real-time data feature tensor using the parking lot scheduling real-time data includes:
[0011] The number of data types for obtaining real-time data on parking lot scheduling;
[0012] The calculation formula for establishing the parking lot scheduling real-time data feature tensor using the number of data types is as follows:
[0013] ,
[0014] Among them, X0 is the feature tensor of parking lot scheduling real-time data, P is the time length of parking lot scheduling real-time data, N is the number of parking lots, C is the number of node features of parking lot scheduling real-time data, and R is a real number set, which is a high-dimensional array composed of real numbers.
[0015] Furthermore, obtaining parking lot scheduling optimization data based on the dynamic high-order graph network according to the parking lot scheduling real-time data feature tensor includes:
[0016] The calculation formula for the initial unified feature convolution of the parking lot scheduling real-time data feature tensor obtained by performing one-dimensional convolution processing on the parking lot scheduling real-time data feature tensor is as follows:
[0017] ,
[0018] Among them, X is the initial unified feature convolution, C is the number of node features of parking lot scheduling real-time data, N is the number of parking lots, T is the time length of parking lot scheduling real-time data, and R is a real number set, which is a high-dimensional array composed of real numbers;
[0019] The initial unified feature convolution is used to establish sliding window gated temporal convolution and high-order differential spatial convolution of parking lot scheduling real-time data based on a dynamic high-order graph network.
[0020] Furthermore, the initial unified feature convolution is used to establish sliding window gated time convolution and high-order differential spatial convolution of parking lot scheduling real-time data based on a dynamic high-order graph network, including:
[0021] The initial unified feature convolution is used to perform segmentation and splicing processing to obtain a sliding window gated time convolution of the parking lot scheduling real-time data;
[0022] The initial unified feature convolution is used to perform multi-level processing to obtain high-order differential spatial convolution of parking lot scheduling real-time data.
[0023] Furthermore, the sliding window gated time convolution of the parking lot scheduling real-time data obtained by segmenting and splicing the initial unified feature convolution includes:
[0024] Using the initial unified feature convolution, segmentation processing is performed according to the time dimension of the time length of the real-time data of the parking lot scheduling to obtain the initial unified feature convolution fragment;
[0025] The calculation formula of the two-dimensional convolution of the parking lot scheduling real-time data obtained by splicing the initial unified feature convolution fragments based on the two-dimensional convolution is as follows:
[0026] ,
[0027] ,
[0028] Among them, SWC is the sliding window convolution operator, X is the initial unified feature convolution, conv2D is the two-dimensional convolution, S is the number of initial unified feature convolution fragments, is the operation of splicing S fragments along the new dimension, t is the length of each initial unified feature convolution fragment, which is obtained by dividing the time length T of the parking lot scheduling real-time data by the number S of the initial unified feature convolution fragments and rounding up. is the sth segment of the initial unified feature convolution X on the channel dimension c and spatial dimension n, and the time range is from (s-1)t to st, ∀ To perform the aforementioned operation for each value in the set, C is the number of node features of the real-time data of parking lot scheduling, and N is the number of parking lots;
[0029] The calculation formula for obtaining the gated unit cycle output result of the parking lot scheduling real-time data by using the two-dimensional convolution of the parking lot scheduling real-time data based on the gated cycle unit structure is as follows:
[0030] ,
[0031] Among them, X g The gated recurrent unit output result of the parking lot scheduling real-time data, GRU is the gated recurrent unit, g is the tanh function, the output range is [-1, 1], σ is the sigmoid function, the output range is [0, 1];
[0032] The calculation formula for obtaining the output tensor of the parking lot scheduling real-time data based on the time hole convolution using the gate control unit cyclic output result of the parking lot scheduling real-time data is as follows:
[0033] ,
[0034] Among them, X c is the output tensor of real-time parking scheduling data, conv1D is a one-dimensional convolution, and f r is the temporal hole convolution reshaping, kernel is the dilated causal convolution kernel parameter, is the reorganization length of the sequence, R is a real number set, which is a high-dimensional array composed of real numbers;
[0035] The two-dimensional convolution of the parking lot scheduling real-time data, the gate control unit cyclic output result and the output tensor are used as the sliding window gated time convolution of the parking lot scheduling real-time data.
[0036] Furthermore, the high-order differential spatial convolution of parking lot scheduling real-time data obtained by multi-stage processing using the initial unified feature convolution includes:
[0037] The calculation formula for obtaining the high-order features of the initial unified feature convolution according to the initial unified feature convolution is as follows:
[0038] ,
[0039]
[0040] in, is the value of the h-order difference sequence at time step t, and are the values of the time series at time step t and t-1 after h-1 order differences, and the difference between the two is the value of the h-th order difference sequence at time step t. is a high-dimensional tensor that connects all difference orders along the time dimension, For the connection operation, A high-dimensional tensor that connects all difference orders along the time dimension Dimensions;
[0041] The conversion and merging result of the initial unified feature convolution obtained by using the high-order features of the initial unified feature convolution is calculated as follows:
[0042] ,
[0043] ,
[0044] ,
[0045] in, is the dimension conversion value of the high-order information tensor, is the conversion function, and are the global node embeddings representing the left and right sides respectively, is the spatial node embedding, for and The same dimension, and is the conversion and merging result of the initial unified feature convolution, It is two linear function layers and one activation function layer. is the parameter;
[0046] The calculation formula for obtaining the adjacency matrix of the initial unified feature convolution using the conversion and merging results of the initial unified feature convolution is as follows:
[0047] ,
[0048] Among them, A d is the adjacency matrix, for Activation function;
[0049] The calculation formula for obtaining the diffusion graph convolution of the initial unified feature convolution using the adjacency matrix of the initial unified feature convolution is as follows:
[0050] ,
[0051] Among them, Z is the diffusion graph convolution, and As parameters, is the order of the graph convolutional network, is the activation function;
[0052] The high-order features of the initial unified feature convolution, the conversion merging results and the adjacency matrix are used as high-order differential spatial convolution of the real-time data of parking lot scheduling.
[0053] Furthermore, obtaining a parking lot scheduling optimization result using the parking lot scheduling optimization data includes:
[0054] The calculation formula for establishing the parking lot scheduling optimization prediction result using the parking lot scheduling optimization data is as follows:
[0055] ,
[0056] in, is the output value predicted by the model, Z is the diffusion graph convolution, and The convolution kernel size is of operator, L is the number of layers in the network, and each layer Generate an output ;
[0057] The calculation formula for establishing the parking lot scheduling optimization loss function based on the parking lot scheduling optimization prediction result and the corresponding parking lot scheduling real-time data is as follows:
[0058] ,
[0059] Among them, Y is the real-time data of parking lot scheduling, Optimize the loss function for parking lot scheduling, is the threshold parameter;
[0060] The parking lot scheduling optimization loss function is used as the parking lot scheduling optimization result.
[0061] Compared with the closest prior art, the present invention has the following beneficial effects:
[0062] 1. Optimize resource utilization: Through real-time demand forecasting and resource scheduling, optimal management of traffic flow between multiple commercial parking lots can be achieved, reducing resource waste and improving parking experience;
[0063] 2. Ensure service quality: by predicting future demand and providing decision support, we can maintain the stable operation of the parking lot, improve service quality, and meet the needs of car owners;
[0064] 3. Improve safety and emergency response capabilities: By optimizing the scheduling process, the parking lot can effectively manage and schedule traffic during peak hours or abnormal situations, improve emergency response speed, and ensure the safety of the parking lot. BRIEF DESCRIPTION OF THE DRAWINGS
[0065] Figure 1 It is a flow chart of a parking lot scheduling optimization method based on a dynamic high-order graph network provided by the present invention. DETAILED DESCRIPTION
[0066] The specific implementation modes of the present invention will be further described in detail below in conjunction with the accompanying drawings.
[0067] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Example 1
[0068] The present invention provides a parking lot scheduling optimization method based on a dynamic high-order graph network, such as Figure 1 As shown, including:
[0069] S1. Using the parking lot scheduling real-time data to obtain the parking lot scheduling real-time data feature tensor;
[0070] S2. Obtaining parking lot scheduling optimization data based on the dynamic high-order graph network according to the parking lot scheduling real-time data feature tensor;
[0071] S3. Obtaining a parking lot scheduling optimization result using the parking lot scheduling optimization data.
[0072] S1 specifically includes:
[0073] S1-1. Obtain the number of data types of real-time parking lot scheduling data;
[0074] S1-2. The calculation formula for establishing the parking lot scheduling real-time data feature tensor using the number of data types is as follows:
[0075] ,
[0076] Among them, X0 is the feature tensor of parking lot scheduling real-time data, P is the time length of parking lot scheduling real-time data, N is the number of parking lots, C is the number of node features of parking lot scheduling real-time data, and R is a real number set, which is a high-dimensional array composed of real numbers.
[0077] S2 specifically includes:
[0078] S2-1. The calculation formula for the initial unified feature convolution of the parking lot scheduling real-time data feature tensor obtained by performing one-dimensional convolution processing on the parking lot scheduling real-time data feature tensor is as follows:
[0079] ,
[0080] Among them, X is the initial unified feature convolution, C is the number of node features of parking lot scheduling real-time data, N is the number of parking lots, T is the time length of parking lot scheduling real-time data, and R is a real number set, which is a high-dimensional array composed of real numbers;
[0081] S2-2. Using the initial unified feature convolution, a sliding window gated temporal convolution and a high-order differential spatial convolution are respectively established for parking lot scheduling real-time data based on a dynamic high-order graph network.
[0082] In this embodiment, a parking lot scheduling optimization method based on a dynamic high-order graph network is provided, wherein the spatial relationship definition , node set : middle, express spatial nodes, relationship edge sets : Represents the relationship between nodes, adjacency matrix : , describes the strength of edge connections, feature tensor , data representation: in spatial relationship In the node The features of are expressed in tensor form, namely , : historical time series length, : Number of spatial nodes, : Number of node features (such as vehicle information, parking lot information, real-time traffic and weather), short-term spatiotemporal prediction, definition: use historical data Predicting future feature data , : Input time window length, : Output time window length, usually less than , prediction method: learning mapping function from arrive ,Right now ,in is the model parameter, the goal is to reduce the prediction data With real future data The difference between them improves accuracy and advantages: Compared with long-term prediction, short-term prediction is more real-time and efficient, and is suitable for scenarios that require quick response.
[0083] In this embodiment, a parking lot scheduling optimization method based on a dynamic high-order graph network, the initial convolution establishment process: feature tensor After one-dimensional convolution, Dimensions are mapped to Dimensional space. Purpose: Increase the number of parameters and enhance the model's ability to abstract and understand data. Dimensional reordering: Perform dimensional reordering on the feature tensors of different data sources to unify them into , simplifying subsequent calculations and processing.
[0084] S2-2 specifically includes:
[0085] S2-2-1. Use the initial unified feature convolution to perform segmentation and splicing processing to obtain a sliding window gated time convolution of the parking lot scheduling real-time data;
[0086] S2-2-2. Use the initial unified feature convolution to perform multi-level processing to obtain high-order differential spatial convolution of parking lot scheduling real-time data.
[0087] In this embodiment, a parking lot scheduling optimization method based on a dynamic high-order graph network is provided, wherein the multi-level processing includes: high-order feature extraction: extracting high-order features from spatiotemporal data; feature conversion and merging: converting and merging the extracted features; adjacency matrix generation: creating a dynamic adjacency matrix based on the converted features; high-order differential spatial convolution: applying the high-order differential spatial convolution generated by the adjacency matrix processing. Through this process, the dynamic spatial relationship in spatiotemporal data is effectively captured, and the model's ability to process spatial variation information is improved.
[0088] S2-2-1 specifically includes:
[0089] S2-2-1-1, using the initial unified feature convolution to perform segmentation processing according to the time dimension of the time length of the real-time data of the parking lot scheduling to obtain an initial unified feature convolution segment;
[0090] S2-2-1-2, the calculation formula of the two-dimensional convolution of the parking lot scheduling real-time data obtained by splicing the initial unified feature convolution fragments based on the two-dimensional convolution is as follows:
[0091] ,
[0092] ,
[0093] Among them, SWC is the sliding window convolution operator, X is the initial unified feature convolution, conv2D is the two-dimensional convolution, S is the number of initial unified feature convolution fragments, is the operation of splicing S fragments along the new dimension, t is the length of each initial unified feature convolution fragment, which is obtained by dividing the time length T of the parking lot scheduling real-time data by the number S of the initial unified feature convolution fragments and rounding up, is the sth segment of the initial unified feature convolution X on the channel dimension c and spatial dimension n, and the time range is from (s-1)t to st, ∀ To perform the aforementioned operation on each value in the set, C is the number of node features of the parking lot scheduling real-time data, and N is the number of parking lots.
[0094] In this embodiment, a parking lot scheduling optimization method based on a dynamic high-order graph network is provided. The two-dimensional convolution of the parking lot scheduling real-time data includes dividing the data into S segments along the time dimension on each channel and node. The length of each segment t is obtained by dividing the original time length T by the number of segments S and rounding up. The time segments are concatenated along a new dimension to form a new feature tensor. The new feature tensor is subjected to a two-dimensional convolution in the segment dimension and the time dimension (using 3 3 convolution kernels). This method can effectively extract the time dimension features and related information between positions in time series data, and is suitable for processing time series signals such as video and audio.
[0095] S2-2-1-3, the calculation formula for obtaining the gated unit cycle output result of the parking lot scheduling real-time data based on the gated cycle unit structure by using the two-dimensional convolution of the parking lot scheduling real-time data is as follows:
[0096] ,
[0097] Among them, X gThis is the gated recurrent unit output result of the real-time parking lot scheduling data. GRU is a gated recurrent unit. g is a tanh function with an output range of [-1, 1]. σ is a sigmoid function with an output range of [0, 1].
[0098] The tanh function is used to process the results of SWC to generate new candidate features and map the input to the interval [-1,1]. The sigmoid function is used to calculate and update the gate output, mapping the input to the interval [0,1] for simulating logic gates, and the output is regarded as the probability of an event. The new candidate features and the updated gate output are combined through element-by-element multiplication operations to control the retention and discarding of information. Through this mechanism, GRU can effectively balance long-term and short-term dependencies in time series data, thereby more accurately capturing the dynamic characteristics of time series.
[0099] S2-2-1-4. The calculation formula for obtaining the output tensor of the parking lot scheduling real-time data based on the time hole convolution by using the gate control unit cyclic output result of the parking lot scheduling real-time data is as follows:
[0100] ,
[0101] Among them, X c is the output tensor of real-time parking scheduling data, conv1D is a one-dimensional convolution, and f r is the temporal hole convolution reshaping, kernel is the dilated causal convolution kernel parameter, is the reorganization length of the sequence, R is a real number set, which is a high-dimensional array composed of real numbers.
[0102] The original time series is divided into S time segments, each of which is t long, and all time segments are re-integrated into a continuous time series in order to handle the time dependency of the entire series while maintaining time continuity. Represents the length of the recombined sequence. The shape is The tensor of Reshape , to match the input format of the high-order differential spatial convolution, apply a one-dimensional convolution using a dilated causal convolution kernel (size 1 3, the hole size is 2), and the processed feature tensor is obtained , the shape is adjusted to Through time-dwelling convolution, the receiving field of view of time convolution can be effectively expanded while maintaining low computational complexity, thereby enhancing the global perception ability of time series data.
[0103] S2-2-1-5. Utilize the two-dimensional convolution of the parking lot scheduling real-time data, the gate control unit cyclic output result and the output tensor as the sliding window gated time convolution of the parking lot scheduling real-time data.
[0104] S2-2-2 specifically includes:
[0105] S2-2-2-1. The calculation formula for obtaining the high-order features of the initial unified feature convolution according to the initial unified feature convolution is as follows:
[0106] ,
[0107] ,
[0108] in, is the value of the h-order difference sequence at time step t, and are the values of the time series at time step t and t-1 after h-1 order differences, and the difference between the two is the value of the h-th order difference sequence at time step t. is a high-dimensional tensor that connects all difference orders along the time dimension, For the connection operation, A high-dimensional tensor that connects all difference orders along the time dimension Dimension.
[0109] In this embodiment, a parking lot scheduling optimization method based on a dynamic high-order graph network is provided, wherein h is the differential order, and its corresponding range is , represents the number of times the difference is performed. For example, h = 0 means no difference is performed, and the difference sequence at this time is the original time series. h = 1 means one difference is performed, h = 2 means two differences are performed, and so on. t is the time step, and its corresponding range is ,A time series can be viewed as a series of data points ordered by time, and each data point corresponds to a time step.
[0110] In addition to the original sequence, all orders of differences are connected together to obtain a high-dimensional tensor that connects all difference orders along the time dimension , It means that the difference tensors of each order are connected along the last dimension (here, the time dimension). In the sum of the lengths of the time dimension of the differences of each order, the length corresponding to the first order is , the length of the time series will decrease by 1 after each difference. The length corresponding to the first order is 1, so from the first order to the The sum of the lengths of the differential time series of the order is calculated according to the arithmetic sequence summation formula as follows: High-order feature acquisition extracts key information of time series through differential operations, effectively eliminates trend and seasonal effects, and provides a basis for in-depth analysis.
[0111] S2-2-2-2. The conversion and merging result of the initial unified feature convolution is obtained by using the high-order features of the initial unified feature convolution to calculate the following formula:
[0112] ,
[0113] ,
[0114] ,
[0115] in, is the dimension conversion value of the high-order information tensor, is the conversion function, and are the global node embeddings representing the left and right sides respectively, is the spatial node embedding, for and The same dimension, and is the conversion and merging result of the initial unified feature convolution, It is two linear function layers and one activation function layer. as a parameter.
[0116] S2-2-2-3. The calculation formula for obtaining the adjacency matrix of the initial unified feature convolution using the conversion and merging results of the initial unified feature convolution is as follows:
[0117] ,
[0118] Among them, A d is the adjacency matrix, for Activation function.
[0119] S2-2-2-4. The calculation formula for obtaining the diffusion graph convolution of the initial unified feature convolution using the adjacency matrix of the initial unified feature convolution is as follows:
[0120] ,
[0121] Among them, Z is the diffusion graph convolution, and As parameters, is the order of the graph convolutional network, is the activation function, specifically using the ReLU activation function, defined as , which can simply and effectively alleviate the problem of gradient disappearance and speed up the training of the model.
[0122] The high-order features of the initial unified feature convolution, the conversion merging results and the adjacency matrix are used as high-order differential spatial convolution of the real-time data of parking lot scheduling.
[0123] In this embodiment, a parking lot scheduling optimization method based on a dynamic high-order graph network is provided, wherein the conversion and merging process sequentially includes a linear transformation layer, a node embedding and a linear transformation layer (Linear).
[0124] In this embodiment, a parking lot scheduling optimization method based on a dynamic high-order graph network is provided. Dynamic and global.
[0125] S3 specifically includes:
[0126] 3-1. The calculation formula for establishing the parking lot scheduling optimization prediction result using the parking lot scheduling optimization data is as follows:
[0127] ,
[0128] in, is the output value predicted by the model, Z is the diffusion graph convolution, and The convolution kernel size is of operator, L is the number of layers in the network, and each layer Generate an output ;
[0129] The calculation formula for establishing the parking lot scheduling optimization loss function based on the parking lot scheduling optimization prediction result and the corresponding parking lot scheduling real-time data is as follows:
[0130] ,
[0131] Among them, Y is the real-time data of parking lot scheduling, The loss function for optimizing parking lot scheduling represents the Huber loss function. is a threshold parameter used to distinguish between "small errors" and "large errors". When the absolute value of the prediction deviation is greater than The loss function becomes a linear function, which helps limit the problem of excessive gradients caused by outliers.
[0132] The parking lot scheduling optimization loss function is used as the parking lot scheduling optimization result.
[0133] In this embodiment, a parking lot scheduling optimization method based on a dynamic high-order graph network is implemented, and the actual application process is as follows:
[0134] Step 1: Data preparation, including parking lot capacity, number of vehicles entering and leaving the parking lot every day, surrounding pedestrian flow, road conditions, weather, surrounding activities, etc.; forming feature tensors , where C represents the number of features, N represents the number of parking lots, and T represents the length of time.
[0135] Step 2: Initial convolution to unify the feature tensor size.
[0136] Step 3: Sliding window gated temporal convolution, using sliding window convolution to process time series data and integrate local and global information; gated structure: controls information flow, optimizes the number of parameters and computational complexity; dilated convolution: merges data dimensions and expands the range of perception.
[0137] Step 4: High-order differential spatial convolution, dynamic adjacency matrix: calculate the dynamic adjacency matrix; diffuse graph convolution: learn spatiotemporal patterns; residual structure: enhance learning ability and generalization performance.
[0138] Step 5: Final convolution: Perform final convolution on the processed features to achieve feature fusion.
[0139] Step 6: Calculate the loss and predict the data With real future data Compare and calculate the prediction error.
[0140] Step 7: Model optimization, use optimization algorithms such as gradient descent to minimize the loss function and optimize the parameters.
[0141] Step 8: Model deployment. After model training is completed, it is used for new data prediction and scheduling decisions to predict future parking lot demand and adjust vehicle scheduling accordingly.
[0142] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0143] The present invention is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0144] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0145] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.
Claims
1. A parking lot scheduling optimization method based on dynamic high-order graph network, characterized in that: include: S1. Using the parking lot scheduling real-time data to obtain the parking lot scheduling real-time data feature tensor; S1-1. Obtain the number of data types of real-time parking lot scheduling data; S1-2. The calculation formula for establishing the parking lot scheduling real-time data feature tensor using the number of data types is as follows: , Among them, X0 is the feature tensor of parking lot scheduling real-time data, P is the time length of parking lot scheduling real-time data, N is the number of parking lots, C is the number of node features of parking lot scheduling real-time data, and R is a real number set, which is a high-dimensional array composed of real numbers; S2. Obtaining parking lot scheduling optimization data based on the dynamic high-order graph network according to the parking lot scheduling real-time data feature tensor; S2-1. The calculation formula for the initial unified feature convolution of the parking lot scheduling real-time data feature tensor obtained by performing one-dimensional convolution processing on the parking lot scheduling real-time data feature tensor is as follows: , Among them, X is the initial unified feature convolution, C is the number of node features of parking lot scheduling real-time data, N is the number of parking lots, T is the time length of parking lot scheduling real-time data, and R is a real number set, which is a high-dimensional array composed of real numbers; S2-2, using the initial unified feature convolution to establish sliding window gated temporal convolution and high-order differential spatial convolution of parking lot scheduling real-time data based on a dynamic high-order graph network; S2-2-1. Use the initial unified feature convolution to perform segmentation and splicing processing to obtain a sliding window gated time convolution of the parking lot scheduling real-time data; S2-2-1-1, using the initial unified feature convolution to perform segmentation processing according to the time dimension of the time length of the real-time data of the parking lot scheduling to obtain an initial unified feature convolution segment; S2-2-1-2, the calculation formula of the two-dimensional convolution of the parking lot scheduling real-time data obtained by splicing the initial unified feature convolution fragments based on the two-dimensional convolution is as follows: , , Among them, SWC is the sliding window convolution operator, X is the initial unified feature convolution, conv2D is the two-dimensional convolution, S is the number of initial unified feature convolution fragments, is the operation of splicing S fragments along the new dimension, t is the length of each initial unified feature convolution fragment, which is obtained by dividing the time length T of the parking lot scheduling real-time data by the number S of the initial unified feature convolution fragments and rounding up, is the sth segment of the initial unified feature convolution X on the channel dimension c and spatial dimension n, and the time range is from (s-1)t to st, To perform the aforementioned operation for each value in the set, C is the number of node features of the real-time data of parking lot scheduling, and N is the number of parking lots; S2-2-1-3, the calculation formula for obtaining the gated unit cycle output result of the parking lot scheduling real-time data based on the gated cycle unit structure by using the two-dimensional convolution of the parking lot scheduling real-time data is as follows: , Among them, X g The gated recurrent unit output result of the parking lot scheduling real-time data, GRU is the gated recurrent unit, g is the tanh function, the output range is [-1, 1], σ is the sigmoid function, the output range is [0, 1]; S2-2-1-4. The calculation formula for obtaining the output tensor of the parking lot scheduling real-time data based on the time hole convolution by using the gate control unit cyclic output result of the parking lot scheduling real-time data is as follows: , Among them, X c is the output tensor of the real-time data of parking lot scheduling, conv1D is the one-dimensional convolution, fr is the temporal hole convolution reshaping, kernel is the dilated causal convolution kernel parameter, is the reorganization length of the sequence, R is a real number set, which is a high-dimensional array composed of real numbers; Using the two-dimensional convolution of the parking lot scheduling real-time data, the gate unit cyclic output result and the output tensor as the sliding window gated time convolution of the parking lot scheduling real-time data; S2-2-2, using the initial unified feature convolution to perform multi-level processing to obtain a high-order differential spatial convolution of the parking lot scheduling real-time data; S2-2-2-1. The calculation formula for obtaining the high-order features of the initial unified feature convolution according to the initial unified feature convolution is as follows: , , in, is the value of the h-order difference sequence at time step t, and are the values of the time series at time step t and t-1 after h-1 order differences, and the difference between the two is the value of the h-th order difference sequence at time step t. is a high-dimensional tensor that connects all difference orders along the time dimension, For the connection operation, A high-dimensional tensor that connects all difference orders along the time dimension Dimensions; S2-2-2-2. The conversion and merging result of the initial unified feature convolution is obtained by using the high-order features of the initial unified feature convolution to calculate the following formula: , , , in, is the dimension conversion value of the high-order information tensor, is the conversion function, and are the global node embeddings representing the left and right sides respectively, is the spatial node embedding, for and The same dimension, and is the conversion and merging result of the initial unified feature convolution, It is two linear function layers and one activation function layer. is the parameter; S2-2-2-3. The calculation formula for obtaining the adjacency matrix of the initial unified feature convolution using the conversion and merging results of the initial unified feature convolution is as follows: , Among them, Ad is the adjacency matrix, for Activation function; S2-2-2-4. The calculation formula for obtaining the diffusion graph convolution of the initial unified feature convolution using the adjacency matrix of the initial unified feature convolution is as follows: , Among them, Z is the diffusion graph convolution, and As parameters, is the order of the graph convolutional network, is the activation function; Using the high-order features of the initial unified feature convolution, the conversion and merging results and the adjacency matrix as high-order differential spatial convolution of parking lot scheduling real-time data; S3. Obtaining a parking lot scheduling optimization result using the parking lot scheduling optimization data.
2. A parking lot scheduling optimization method based on a dynamic high-order graph network as claimed in claim 1, characterized in that: The parking lot scheduling optimization results obtained by using the parking lot scheduling optimization data include: The calculation formula for establishing the parking lot scheduling optimization prediction result using the parking lot scheduling optimization data is as follows: , in, is the output value predicted by the model, Z is the diffusion graph convolution, and The convolution kernel size is of operator, L is the number of layers in the network, and each layer Generate an output ; The calculation formula for establishing the parking lot scheduling optimization loss function based on the parking lot scheduling optimization prediction result and the corresponding parking lot scheduling real-time data is as follows: , Among them, Y is the real-time data of parking lot scheduling, Optimize the loss function for parking lot scheduling, is the threshold parameter; The parking lot scheduling optimization loss function is used as the parking lot scheduling optimization result.
Citation Information
Patent Citations
Gas operation data prediction method based on spatial high-order convolution
CN117726183A
