Parking pile available transfer bed adjustment method and system based on demand prediction

By optimizing the feature fusion of dense networks through the contribution prediction module and the reference configuration of the preceding layer, the problem of redundant information in dense networks is solved, the accuracy of demand prediction for transfer beds and the reliability of scheduling adjustments are improved, and the rational allocation of resources is realized.

CN123070633APending Publication Date: 2026-11-10NINGBO QIFA MEDICAL TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611544964.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-08
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Dense networks employ a fixed connection pattern during feature forward propagation, leading to the introduction of redundant information, making it difficult to filter relevant features, and affecting the accuracy of demand forecasting for transfer beds and the matching of scheduling adjustments.

Method used

The feature contribution matrix is ​​calculated by the contribution prediction module, and the input features are processed by weighting. The reference configuration of the preceding layer is generated according to the learning state of the previous network layer, thereby optimizing the hierarchical information transmission and feature fusion of the dense network.

Benefits of technology

This improved the accuracy of demand forecasting for transfer beds and the reliability of scheduling adjustments, enabling more rational resource allocation and enhancing overall resource utilization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123070633A_ABST
    Figure CN123070633A_ABST
Patent Text Reader

Abstract

The application provides a parking pile available transfer bed adjustment method and system based on demand prediction, including collecting historical transfer bed use data, real-time demand data and spatial position information of each parking pile, extracting a time sequence feature set of historical data, splicing the time sequence features, real-time demand and position information to obtain an initial feature vector, calculating a feature contribution degree matrix through a contribution degree prediction module, weighting the initial feature vector to obtain a weighted input feature vector, sending the vector into a dense network, generating a previous layer reference configuration according to the learning state of the output features of the previous network layer from the second network layer, selecting the corresponding previous layer output for feature transformation and fusion, outputting a prediction result through layer-by-layer forward propagation, obtaining a transfer bed demand prediction value of each parking pile, calculating a difference value in combination with the current transfer bed retention, and outputting a transfer bed increase / decrease adjustment instruction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application belongs to the field of adjustment, and in particular relates to a method and system for adjusting the availability of a transfer bed for parking piles based on demand forecasting. Background Technology

[0002] As medical institutions expand and their operational processes become increasingly complex, transport beds, as crucial equipment for patient and material transport, are typically centrally located at various designated docking stations for immediate deployment. To ensure timely access to transport beds across different areas, management systems usually collect historical usage data, real-time demand data, and spatial location information for each docking station. This data is used to predict future demand and adjust the number of transport beds at each station accordingly. Deep learning technology is employed in the demand forecasting phase due to its ability to handle multidimensional and complex data. For example, dense network models are used to perform forward propagation analysis on feature vectors containing temporal and spatial information to determine the transport bed demand at each docking station over a future period, thereby assisting in the generation of scheduling and adjustment instructions.

[0003] However, dense networks typically employ a fixed-connection pattern during feature forward propagation, where the current layer receives output features from each preceding layer. When processing spatiotemporal operational data of in-hospital transport beds, this fixed-level connection method easily introduces redundant information, making it difficult to dynamically select more relevant preceding layer features as input sources based on the current network layer's learning state. This limits the efficiency of feature transformation and fusion, affecting the accuracy of transport bed demand prediction. Prediction bias further leads to discrepancies between the transport bed increase / decrease adjustment instructions generated based on predicted values ​​and current inventory levels and the actual scheduling needs. Therefore, a new data processing approach is needed to more reasonably assess the impact of multi-source input features and optimize the hierarchical information transmission and feature fusion mechanisms within dense networks, thereby improving the reliability of transport bed demand prediction and scheduling adjustments. Summary of the Invention

[0004] To improve the accuracy of predicting future demand for parking bollards within hospitals, this invention provides a method and system for adjusting available parking bollards based on demand prediction.

[0005] The first aspect of this invention provides a method for adjusting the availability of transfer beds for parking piles based on demand forecasting, comprising the following steps: The system acquires historical transfer bed usage data, real-time demand data, and spatial location information for each parking pile; extracts a time-series feature set based on the historical transfer bed usage data; concatenates the time-series feature set, the real-time demand data, and the spatial location information to obtain an initial feature vector; inputs the initial feature vector into the contribution prediction module to calculate the expected impact weight to obtain a feature contribution matrix; and multiplies the initial feature vector by the feature contribution matrix to obtain a weighted input feature vector. The weighted input feature vector is input into a dense network for forward propagation. Starting from the second network layer of the dense network, a predecessor layer reference configuration is generated based on the learning state of the output features of the previous network layer. A predecessor layer is selected as the input source according to the predecessor layer reference configuration. The output features of the selected predecessor layer are transformed and structured and fused to generate the input features of the current layer. The current layer output features are processed and propagated sequentially to the end of the network to obtain the network output result. Based on the network output results, the non-negative integer demand forecast value of each parking pile is obtained; the difference between the demand forecast value and the current inventory of each parking pile is calculated to generate a transfer bed increase / decrease adjustment instruction.

[0006] Furthermore, the step of extracting a time-series feature set based on the historical transfer bed usage data includes: The historical transfer bed usage data of each parking pile is arranged in ascending order by timestamp to form a time series set; Establish a short-cycle time window with a set duration of 15 minutes, a medium-cycle time window with a set duration of 1 hour, and a long-cycle time window with a set duration of 24 hours. Using the current moment as the cutoff endpoint, slide the short-cycle time window, the medium-cycle time window, and the long-cycle time window in reverse along the time axis on the time series set, respectively, and extract all historical event records within the three window periods; The number of borrowed and returned transfer beds generated in the records of each window is counted and spliced ​​together to form an independent data array. The data arrays generated by the three time windows of short period, medium period and long period are concatenated into a whole vector in dimensional order, and the whole vector is used as the time series feature set.

[0007] Further, the step of inputting the initial feature vector into the contribution prediction module to calculate the expected influence weights and obtain the feature contribution matrix includes: Set the internal network node parameter matrix for the contribution prediction module; The initial feature vector is multiplied by the network node parameter matrix to obtain the intermediate hidden layer feature vector; The dimension of the intermediate hidden layer feature vector is restored to the same data length as the initial feature vector by a fully connected mapping to obtain the mapped feature vector; A non-linear activation function is used to perform numerical mapping calculations on the mapped feature vector, converting each result value into a small value between 0 and 1, representing the influence weight value of each input feature; All influence weight values ​​are assembled into a one-dimensional weight array with the same data length according to the data arrangement order of the original initial feature vector. A diagonal matrix with each weight value as the diagonal element is constructed based on the one-dimensional weight array. The diagonal matrix is ​​the feature contribution matrix.

[0008] Further, the step of multiplying the initial feature vector with the feature contribution matrix to obtain the weighted input feature vector includes: Construct an output storage vector with all initial values ​​of zero and a size consistent with the dimension of the initial feature vector; In the traversal operation step, specific single numerical features of the current ranking of the initial feature vector are extracted sequentially, and the first numerical weights corresponding to the diagonal positions in the feature contribution matrix are extracted simultaneously. Perform an arithmetic multiplication operation on the currently extracted single numerical feature and the first numerical weight corresponding to the ranking of that feature to obtain the product value; The obtained product value is filled into the rank position corresponding to the output storage vector in its original position; The extraction and multiplication steps are repeated until all values ​​have been processed, and the output storage vector filled in this way is used as the weighted input feature vector.

[0009] Further, the step of generating the pre-sequence layer reference configuration based on the learning state of the output features of the previous network layer includes: The activation value variance and feature information entropy are calculated based on the output feature data of the previous network layer to obtain a statistical summary. The statistical summary is then concatenated with the layer number of the current processing node to form the learning state index vector of the current network layer. Set up a controller with multiple output branches, where each branch corresponds to a specific single layer among all the preceding layers; The learning state index vector is input into the classification function of the controller to map and calculate the reference connection probability score pointing to each specific single layer; Compare the reference connection probability scores of all branches with the preset reference threshold values. Retain the sequence number of the preceding layer corresponding to all probability scores greater than the preset reference threshold. If all probability scores are not greater than the preset reference threshold, retain the sequence number of the preceding layer corresponding to the largest probability score. Combine the sequence numbers to form the preceding layer reference configuration.

[0010] Furthermore, the step of performing feature transformation and structured fusion on the selected output features of the preceding layer to generate the input features of the current layer includes: Based on the available connection target number in the preceding layer reference configuration, retrieve the first original multidimensional tensor set output by the specified preceding network layer node; A linear transformation algorithm is applied to each tensor in the first original multidimensional tensor set to reduce scale differences, and the data length of each tensor is aligned by interpolation resampling to obtain a standardized tensor set. The tensors in the standardized tensor group are concatenated along the channel dimension into a multi-channel fused tensor. The multi-channel fusion tensor is rearranged in an ordered manner according to the channel dimension and length dimension to form a one-dimensional convolutional input tensor that preserves the positional relationship of the channel dimension and length.

[0011] Further, the step of generating a transfer bed increase / decrease adjustment instruction by calculating the difference between the predicted demand value of the transfer bed and the current transfer bed inventory of each parking pile includes: Read the non-negative integer-based demand forecast values ​​for each parking pile in the future scheduling cycle; Obtain the current inventory of transfer beds returned by sensor scans at each corresponding physical site; The difference in the number of transfer beds is obtained by subtracting the current number of transfer beds from the predicted demand value. If the quantity difference result is determined to be positive, then an action instruction is generated to transfer the same number of transfer beds as the difference value from the central warehouse or the available transfer bed parking piles to the corresponding parking pile position. If the difference value is determined to be negative, an action instruction is generated to retrieve and transport the same number of transfer beds as the absolute value of the difference value from the corresponding parking piles to the central warehouse. The action command information is combined and encapsulated with the corresponding parking pile identification number to generate a complete set of transfer bed increase / decrease adjustment command results, which is then sent to the scheduling system for dispatch and use.

[0012] A second aspect of the present invention provides a parking pile availability transfer bed adjustment system based on demand forecasting, comprising the following modules: The calculation module is used to acquire historical transfer bed usage data, real-time demand data, and spatial location information for each parking pile; extract a time-series feature set based on the historical transfer bed usage data; concatenate the time-series feature set, the real-time demand data, and the spatial location information to obtain an initial feature vector; input the initial feature vector to the contribution prediction module to calculate the expected impact weight to obtain a feature contribution matrix; and multiply the initial feature vector by the feature contribution matrix to obtain a weighted input feature vector. The generation module is used to input the weighted input feature vector into the dense network for forward propagation. Starting from the second network layer of the dense network, it generates a predecessor layer reference configuration based on the learning state of the output features of the previous network layer. Based on the predecessor layer reference configuration, a predecessor layer is selected as the input source. The output features of the selected predecessor layer are transformed and fused with structure to generate the input features of the current layer. The current layer output features are processed and propagated sequentially to the end of the network to obtain the network output result. The adjustment module is used to obtain the non-negative integer demand forecast value of each parking pile based on the network output results; and to calculate the difference between the demand forecast value of the transfer bed and the current transfer bed inventory of each parking pile to generate a transfer bed increase or decrease adjustment instruction.

[0013] Preferably, the step of extracting a time-series feature set based on the historical transfer bed usage data includes: The historical transfer bed usage data of each parking pile is arranged in ascending order by timestamp to form a time series set; Establish a short-cycle time window with a set duration of 15 minutes, a medium-cycle time window with a set duration of 1 hour, and a long-cycle time window with a set duration of 24 hours. Using the current moment as the cutoff endpoint, slide the short-cycle time window, the medium-cycle time window, and the long-cycle time window in reverse along the time axis on the time series set, respectively, and extract all historical event records within the three window periods; The number of borrowed and returned transfer beds generated in the records of each window is counted and spliced ​​together to form an independent data array. The data arrays generated by the three time windows of short period, medium period and long period are concatenated into a whole vector in dimensional order, and the whole vector is used as the time series feature set.

[0014] Preferably, the step of inputting the initial feature vector into the contribution prediction module to calculate the expected influence weights and obtain the feature contribution matrix includes: Set the internal network node parameter matrix for the contribution prediction module; The initial feature vector is multiplied by the network node parameter matrix to obtain the intermediate hidden layer feature vector; The dimension of the intermediate hidden layer feature vector is restored to the same data length as the initial feature vector by a fully connected mapping to obtain the mapped feature vector; A non-linear activation function is used to perform numerical mapping calculations on the mapped feature vector, converting each result value into a small value between 0 and 1, representing the influence weight value of each input feature; All influence weight values ​​are assembled into a one-dimensional weight array with the same data length according to the data arrangement order of the original initial feature vector. A diagonal matrix with each weight value as the diagonal element is constructed based on the one-dimensional weight array. The diagonal matrix is ​​the feature contribution matrix.

[0015] Preferably, the step of multiplying the initial feature vector with the feature contribution matrix to obtain the weighted input feature vector includes: Construct an output storage vector with all initial values ​​of zero and a size consistent with the dimension of the initial feature vector; In the traversal operation step, specific single numerical features of the current ranking of the initial feature vector are extracted sequentially, and the first numerical weights corresponding to the diagonal positions in the feature contribution matrix are extracted simultaneously. Perform an arithmetic multiplication operation on the currently extracted single numerical feature and the first numerical weight corresponding to the ranking of that feature to obtain the product value; The obtained product value is filled into the rank position corresponding to the output storage vector in its original position; The extraction and multiplication steps are repeated until all values ​​have been processed, and the output storage vector filled in this way is used as the weighted input feature vector.

[0016] Preferably, generating the pre-sequence layer reference configuration based on the learning state of the output features of the previous network layer includes: The activation value variance and feature information entropy are calculated based on the output feature data of the previous network layer to obtain a statistical summary. The statistical summary is then concatenated with the layer number of the current processing node to form the learning state index vector of the current network layer. Set up a controller with multiple output branches, where each branch corresponds to a specific single layer among all the preceding layers; The learning state index vector is input into the classification function of the controller to map and calculate the reference connection probability score pointing to each specific single layer; Compare the reference connection probability scores of all branches with the preset reference threshold values. Retain the sequence number of the preceding layer corresponding to all probability scores greater than the preset reference threshold. If all probability scores are not greater than the preset reference threshold, retain the sequence number of the preceding layer corresponding to the largest probability score. Combine the sequence numbers to form the preceding layer reference configuration.

[0017] Preferably, the step of performing feature transformation and structured fusion on the selected output features of the preceding layer to generate the input features of the current layer includes: Based on the available connection target number in the preceding layer reference configuration, retrieve the first original multidimensional tensor set output by the specified preceding network layer node; A linear transformation algorithm is applied to each tensor in the first original multidimensional tensor set to reduce scale differences, and the data length of each tensor is aligned by interpolation resampling to obtain a standardized tensor set. The tensors in the standardized tensor group are concatenated along the channel dimension into a multi-channel fused tensor. The multi-channel fusion tensor is rearranged in an ordered manner according to the channel dimension and length dimension to form a one-dimensional convolutional input tensor that preserves the positional relationship of the channel dimension and length.

[0018] Preferably, the step of generating a transfer bed increase / decrease adjustment instruction by calculating the difference between the predicted demand value of the transfer bed and the current transfer bed inventory of each parking post includes: Read the non-negative integer-based demand forecast values ​​for each parking pile in the future scheduling cycle; Obtain the current inventory of transfer beds returned by sensor scans at each corresponding physical site; The difference in the number of transfer beds is obtained by subtracting the current number of transfer beds from the predicted demand value. If the quantity difference result is determined to be positive, then an action instruction is generated to transfer the same number of transfer beds as the difference value from the central warehouse or the available transfer bed parking piles to the corresponding parking pile position. If the difference value is determined to be negative, an action instruction is generated to retrieve and transport the same number of transfer beds as the absolute value of the difference value from the corresponding parking piles to the central warehouse. The action command information is combined and encapsulated with the corresponding parking pile identification number to generate a complete set of transfer bed increase / decrease adjustment command results, which is then sent to the scheduling system for dispatch and use.

[0019] This invention integrates historical usage data, real-time demand data, and spatial location information, and uses a contribution prediction module to calculate a feature weight matrix. This weighted processing of initial features highlights key influencing factors, reduces the interference of redundant information in the prediction process, and improves the quality of input feature representation. In the network processing stage, a preceding layer reference configuration is generated based on the learning state of the output features of the previous network layer. The input sources for the current layer are selectively determined, and the selected preceding layer output features are transformed and fused. This optimizes the connection methods between network layers, reduces invalid feature transmission and computational overhead, and improves the model's feature extraction flexibility and learning efficiency. Based on the above processing results, the demand for transfer beds for each parking pile can be predicted more reliably, and corresponding increase / decrease adjustment instructions can be generated based on the current inventory, achieving reasonable allocation and scheduling management of transfer beds and improving overall resource utilization. Attached Figure Description

[0020] Figure 1 A flowchart of a method for adjusting the availability of transfer beds for parking piles based on demand forecasting; Figure 2 This is a schematic diagram of the weight distribution of the feature contribution matrix; Figure 3 A schematic diagram showing the distribution of the number of instructions for increasing or decreasing the number of transfer beds. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] Reference Figure 1 This invention proposes a method for adjusting the availability of transfer beds for parking piles based on demand forecasting, comprising the following steps: S1, acquire historical transfer bed usage data, real-time demand data, and spatial location information for each parking pile; extract a time-series feature set based on the historical transfer bed usage data; concatenate the time-series feature set, the real-time demand data, and the spatial location information to obtain an initial feature vector; input the initial feature vector into the contribution prediction module to calculate the expected impact weight to obtain a feature contribution matrix; multiply the initial feature vector by the feature contribution matrix to obtain a weighted input feature vector.

[0023] By using IoT RFID readers installed at each parking station and integrating them with the hospital's information management system, historical transfer bed usage data is collected, including timestamped records of bed borrowing and returning events within a set number of days. Simultaneously, the number of current borrowing requests is obtained via a mobile application, providing real-time demand data, and the three-dimensional coordinates of each parking station are acquired via indoor positioning base stations, serving as spatial location information. Subsequently, the historical transfer bed usage data is segmented using a sliding window approach, and the number of borrowed and returned transfer beds within each time window is statistically analyzed. The statistical results are then combined to form a time-series feature set.

[0024] The time-series feature set, real-time demand data, and spatial location information are subjected to maximum and minimum value normalization processing, mapping each data point to a range of 0 to 1. These data are then concatenated along the feature dimension in a preset order to obtain a one-dimensional initial feature vector. This initial feature vector is input into a contribution prediction module composed of a multilayer perceptron network. Through fully connected mapping and activation function processing, the expected influence weights corresponding to each element in the initial feature vector are calculated. These weights are then arranged in their original index order to form a weight array of the same length as the initial feature vector. Next, an N×N diagonal matrix with each weight as a diagonal element is constructed based on this weight array, serving as the feature contribution matrix. Subsequently, element-wise multiplication is performed between the weight array and the initial feature vector, or a diagonal matrix multiplication is performed between the initial feature vector and the feature contribution matrix to obtain a weighted input feature vector. This process relatively preserves key features and reduces the influence of irrelevant features on the subsequent prediction process. The weight distribution of the feature contribution matrix is ​​illustrated below. Figure 2 As shown.

[0025] As an optional implementation, the step of extracting a time-series feature set based on the historical transfer bed usage data includes: The historical transfer bed usage data of each parking pile is arranged in ascending order by timestamp to form a time series set; Establish a short-cycle time window with a set duration of 15 minutes, a medium-cycle time window with a set duration of 1 hour, and a long-cycle time window with a set duration of 24 hours. Using the current moment as the cutoff endpoint, slide the short-cycle time window, the medium-cycle time window, and the long-cycle time window in reverse along the time axis on the time series set, respectively, and extract all historical event records within the three window periods; The number of borrowed and returned transfer beds generated in the records of each window is counted and spliced ​​together to form an independent data array. The data arrays generated by the three time windows of short period, medium period and long period are concatenated into a whole vector in dimensional order, and the whole vector is used as the time series feature set.

[0026] At the present moment The historical transfer bed usage data is represented as a set of sequences containing timestamps and borrowing / returning event types. If the original data is stored as aggregated quantities for time periods, these aggregated quantities are first bound to the corresponding timestamps, serving as equivalent timestamped quantity records for window statistics. Based on this, data from three windows are extracted in reverse along the time axis to derive the short-cycle window. Medium-term window and long-term windows The data records are used to perform event counting within each window, and the number of borrowed and returned devices within a short period is statistically analyzed to form a short-period data array. The number of units lent out and returned within a medium cycle forms a medium-cycle data array. The number of devices lent out and returned over a long period forms a long-term data array. Following the feature extraction order of short-period, medium-period, and long-period, the array concatenation operation is performed, sequentially connecting the beginning and end of each independent array to generate a single overall vector. The numerical distribution is used to represent the flow trend at different time granularities.

[0027] As an optional implementation, the step of inputting the initial feature vector into the contribution prediction module to calculate the expected influence weights and obtain the feature contribution matrix includes: Set the internal network node parameter matrix for the contribution prediction module; The initial feature vector is multiplied by the network node parameter matrix to obtain the intermediate hidden layer feature vector; The dimension of the intermediate hidden layer feature vector is restored to the same data length as the initial feature vector by a fully connected mapping to obtain the mapped feature vector; A non-linear activation function is used to perform numerical mapping calculations on the mapped feature vector, converting each result value into a small value between 0 and 1, representing the influence weight value of each input feature; All influence weight values ​​are assembled into a one-dimensional weight array with the same data length according to the data arrangement order of the original initial feature vector. A diagonal matrix with each weight value as the diagonal element is constructed based on the one-dimensional weight array. The diagonal matrix is ​​the feature contribution matrix.

[0028] This contribution prediction module includes a neural network model. The network structure comprises an input layer, an intermediate hidden layer, and an output layer. The input layer receives the input feature data. The intermediate hidden layer contains multiple neurons using a linear activation function. The output layer contains multiple neurons corresponding to the feature dimension and uses a non-linear activation function to map feature values. The input to this network model is an initial feature vector, and the output is a one-dimensional weight array with the same length as the initial feature vector. A feature contribution matrix is ​​constructed based on this one-dimensional weight array. Assuming the concatenated initial feature vector... The first network node parameter matrix of the contribution prediction module is defined as having a dimension of 1×N. and bias vector First network node parameter matrix The dimension is N × the number of hidden layer nodes, and the bias vector is... The dimension is 1 × the number of hidden layer nodes, and the matrix multiplication and addition operations are used. Find the eigenvector H of the intermediate hidden layer.

[0029] Call the second network node parameter matrix and bias vector Perform a fully connected mapping, second network node parameter matrix The dimension is the number of hidden layer nodes × N, and the bias vector is... The dimension is 1×N, and the formula is used to... The hidden features are restored to an initial 1×N dimension, outputting a mapped feature vector Y. A non-linear activation function is independently applied to each scalar element of vector Y. The influence weights are derived by compressing and mapping eigenvalues ​​of any real number range to a probability interval of 0 to 1. These weights are then arranged in their original indices to obtain a 1×N one-dimensional weight array, which is then used to construct an N×N diagonal matrix as the feature contribution matrix. The one-dimensional weight array is used in conjunction with... Perform element-wise multiplication, or Used in conjunction with the feature contribution matrix Perform diagonal matrix multiplication; when using matrix notation... It is equivalently represented as an N×N diagonal matrix, where the diagonal elements are the influence weights in the one-dimensional array, used to indicate the information value of each input feature for the task of predicting the demand for transport beds.

[0030] As an optional implementation, multiplying the initial feature vector by the feature contribution matrix to obtain the weighted input feature vector includes: Construct an output storage vector with all initial values ​​of zero and a size consistent with the dimension of the initial feature vector; In the traversal operation step, specific single numerical features of the current ranking of the initial feature vector are extracted sequentially, and the first numerical weights corresponding to the diagonal positions in the feature contribution matrix are extracted simultaneously. Perform an arithmetic multiplication operation on the currently extracted single numerical feature and the first numerical weight corresponding to the ranking of that feature to obtain the product value; The obtained product value is filled into the rank position corresponding to the output storage vector in its original position; The extraction and multiplication steps are repeated until all values ​​have been processed, and the output storage vector filled in this way is used as the weighted input feature vector.

[0031] When the initial feature vector is input and the corresponding feature contribution matrix At that time, construct an output storage vector with all initial values ​​of zero and a size consistent with the dimension of the initial feature vector. Perform traversal operations to extract the initial feature vectors sequentially. Corresponding bit values ​​and feature contribution matrix For the weights at the corresponding diagonal positions, perform multiplication to obtain the product, and store the product in the output storage vector. The corresponding bits are then used. This process is repeated element by element, performing calculations to ensure that each original feature value is scaled and adjusted according to its own information contribution. Once all calculations are complete, the generated vector fully preserves the feature scale relationship and relative weight distribution. This vector is then used as a weighted input feature vector and passed to the next stage of the network structure.

[0032] Before applying the model, a training dataset is constructed based on historical samples. Each training sample includes historical transfer bed usage data, real-time demand data, and spatial location information. The training label is a one-dimensional label vector composed of the actual transfer bed demand quantity of each parking pile in the future scheduling cycle. During training, the contribution prediction module, dense network, and controller used to generate the preceding layer reference configuration are trained as a joint model. The mean squared error or mean absolute error between the predicted demand quantity vector and the actual demand quantity vector is used as the loss function. The trainable parameters in the contribution prediction module, controller, and dense network are updated through backpropagation. During the training phase, the reference connection probability score output by the controller is used as a continuous weight in the weighted fusion of preceding layer features, and the controller parameters are updated through backpropagation. After training, the model parameters are fixed for the real-time prediction phase. In the real-time prediction phase, the reference connection probability score is compared with a preset reference threshold to generate discrete preceding layer reference configurations. The preset reference threshold can be set to 0.5.

[0033] S2, the weighted input feature vector is input into the dense network for forward propagation. Starting from the second network layer of the dense network, a predecessor layer reference configuration is generated based on the learning state of the output features of the previous network layer. The predecessor layer is selected as the input source according to the predecessor layer reference configuration. The output features of the selected predecessor layer are transformed and structured to generate the input features of the current layer. The current layer output features are processed and propagated to the end of the network to obtain the network output result.

[0034] A dense network containing multiple dense blocks and transition layers is constructed. Before the weighted input feature vector is input into the dense network, the one-dimensional weighted input feature vector is first converted into an initial feature tensor through an input mapping layer. The input mapping layer includes fully connected mapping and shape rearrangement processing to convert the 1×N-dimensional weighted input feature vector into an initial multi-dimensional feature tensor with channel and length dimensions. This initial multi-dimensional feature tensor is then input into the first network layer of the dense network to initiate the forward propagation process. During the layer-by-layer propagation process, starting from the second network layer, the statistical summary of the output feature data of the previous network layer and the current layer index are used together as the learning state index. The statistical summary includes the activation value variance and feature information entropy of the output features of the previous network layer. The learning state index is input to the configuration to generate a sub-network, and the selection results of each preceding layer are output using a gating mechanism. The selection results are used as the preceding layer reference configuration. According to the preceding layer reference configuration, the unselected preceding layers are excluded, and the selected preceding layers are retained as the input source of the current layer.

[0035] Linear projection is applied to the output features of each selected preceding layer to unify the feature dimension and complete the feature transformation. The transformed features from each layer are then concatenated along the channel dimension to achieve structured fusion, forming the actual input features of the current layer. Upon receiving these input features, the current layer sequentially performs batch normalization, linear rectified activation, and one-dimensional convolutional feature extraction to generate the current layer's output features. This state evaluation, configuration generation, and feature processing process is repeated until the forward propagation reaches the global average pooling layer and fully connected layer at the end of the dense network. The output is a one-dimensional regression vector representing the future demand for transfer beds for each parking pile, which serves as the network output.

[0036] As an optional implementation, the step of generating the pre-sequence layer reference configuration based on the learning state of the output features of the previous network layer includes: The activation value variance and feature information entropy are calculated based on the output feature data of the previous network layer to obtain a statistical summary. The statistical summary is then concatenated with the layer number of the current processing node to form the learning state index vector of the current network layer. Set up a controller with multiple output branches, where each branch corresponds to a specific single layer among all the preceding layers; The learning state index vector is input into the classification function of the controller to map and calculate the reference connection probability score pointing to each specific single layer; Compare the reference connection probability scores of all branches with the preset reference threshold values. Retain the sequence number of the preceding layer corresponding to all probability scores greater than the preset reference threshold. If all probability scores are not greater than the preset reference threshold, retain the sequence number of the preceding layer corresponding to the largest probability score. Combine the sequence numbers to form the preceding layer reference configuration.

[0037] The controller with multiple output branches comprises a neural network model. The network structure includes an input layer and a fully connected classification layer with multiple output branches. The input to this network model is the learning state index vector of the current network layer, and the output is the reference connection probability score pointing to each specific single layer. When the number of preceding layers corresponding to different current network layers varies, the controller only activates the output branches corresponding to existing preceding layers according to the layer number of the current network layer. Output branches not corresponding to existing preceding layers are masked and do not participate in the reference connection probability score calculation. During computation, feature data from the output of the previous network layer is extracted, and the activation value variance and feature information entropy are calculated from the feature data as a statistical summary. The current layer number is converted into a configurable feature tensor and appended to the end of the statistical summary to form the learning state index vector. It is used to simultaneously represent the feature representation distribution and the network depth environment.

[0038] index vector The network structure input to the controller is parsed using a multi-label classification function to calculate the connection probability of each potential direct path. These probability scores are then compared to a preset reference threshold, which can be set to 0.5, to select the preceding layer numbers that meet the criteria. To ensure that each network layer gradient has at least one backhaul path, the preceding layer number corresponding to the highest probability score is retained when all branch probability scores are not greater than the preset reference threshold. The selected number encoding set is combined to form a preceding layer reference configuration, used to guide the establishment of subsequent skip connections.

[0039] As an optional implementation, the step of performing feature transformation and structured fusion on the selected preceding layer output features to generate the current layer input features includes: Based on the available connection target number in the preceding layer reference configuration, retrieve the first original multidimensional tensor set output by the specified preceding network layer node; A linear transformation algorithm is applied to each tensor in the first original multidimensional tensor set to reduce scale differences, and the data length of each tensor is aligned by interpolation resampling to obtain a standardized tensor set. The tensors in the standardized tensor group are concatenated along the channel dimension into a multi-channel fused tensor. The multi-channel fusion tensor is rearranged in an ordered manner according to the channel dimension and length dimension to form a one-dimensional convolutional input tensor that preserves the positional relationship of the channel dimension and length.

[0040] After parsing the preceding layer reference configuration to obtain the set of available connection target indices, the output tensors of the selected preceding network layer nodes are retrieved to form the first original multidimensional tensor set. Due to size differences, a linear transformation operation is used on the selected output tensors to unify the number of feature channels, and one-dimensional linear interpolation resampling is used to make the length of each tensor consistent, resulting in a standardized tensor group; then, it is concatenated along the channel dimension to obtain a multi-channel fused tensor. When the current network layer uses one-dimensional convolution for feature extraction, the multi-channel fused tensor is organized according to the channel index and length index into a one-dimensional convolution input tensor that retains the channel dimension and length dimension. This input tensor retains the correspondence between each channel and length position and serves as the standard input feature of the one-dimensional convolutional layer.

[0041] S3. Based on the network output results, obtain the non-negative integer demand forecast value of each parking pile for the transfer bed; calculate the difference between the transfer bed demand forecast value and the current transfer bed inventory of each parking pile to generate a transfer bed increase / decrease adjustment instruction.

[0042] The one-dimensional regression numerical vector output from the end of the dense network is denormalized and then converted into discrete integer values ​​using a rounding-up algorithm. If the converted value is less than zero, it is corrected to zero. This integer value is the predicted demand for transfer beds at each parking post within a specified future time period. The number of transfer beds currently parked within the range of each parking post is counted in real time using gravity sensors at the bottom of the parking posts or an electronic fence system, which serves as the current transfer bed inventory. The demand difference is calculated by subtracting the corresponding current transfer bed inventory from the predicted demand value. When the demand difference is greater than zero, it indicates a shortage of transfer beds for that parking post, and a transfer bed increase instruction is generated, containing the target parking post number and the number of transfer beds to be transferred in. When the demand difference is less than zero, it indicates redundancy of transfer beds for that parking post, and the absolute value of the difference is taken to generate a transfer bed decrease instruction, containing the source parking post number and the number of transfer beds to be transferred out. When the demand difference is equal to zero, no action instruction is generated for that parking post. The generated increase / decrease adjustment instructions, containing specific quantity and location information, are sent to the scheduling execution terminal to execute the actual allocation task. The distribution of the number of transfer bed increase / decrease adjustment instructions is shown in the diagram below. Figure 3 As shown.

[0043] As an optional implementation, the step of generating a transfer bed increase / decrease adjustment instruction by calculating the difference between the predicted demand value of the transfer bed and the current transfer bed inventory of each parking post includes: Read the non-negative integer-based demand forecast values ​​for each parking pile in the future scheduling cycle; Obtain the current inventory of transfer beds returned by sensor scans at each corresponding physical site; The difference in the number of transfer beds is obtained by subtracting the current number of transfer beds from the predicted demand value. If the quantity difference result is determined to be positive, then an action instruction is generated to transfer the same number of transfer beds as the difference value from the central warehouse or the available transfer bed parking piles to the corresponding parking pile position. If the difference value is determined to be negative, an action instruction is generated to retrieve and transport the same number of transfer beds as the absolute value of the difference value from the corresponding parking piles to the central warehouse. The action command information is combined and encapsulated with the corresponding parking pile identification number to generate a complete set of transfer bed increase / decrease adjustment command results, which is then sent to the scheduling system for dispatch and use.

[0044] The system reads the predicted demand for transfer beds for a specific parking pile within a future scheduling cycle and simultaneously obtains the current inventory of transfer beds. A subtraction operation is performed to obtain the quantity difference, i.e., the difference value. If the difference value is positive, an action instruction is generated with the action type "transfer in," the source endpoint being the central warehouse or available transfer bed surplus parking piles, and the destination endpoint being the corresponding parking pile location. If the difference value is negative, the absolute value of the difference is calculated, and an action instruction is generated with the action type "recovery out," and the destination endpoint being the central warehouse. The action instruction information is combined with the corresponding parking pile identifier to form a transfer bed increase / decrease adjustment instruction result set, which is then sent to the scheduling system for dispatch.

[0045] This invention also proposes a parking pile availability transfer bed adjustment system based on demand forecasting, comprising the following modules: The calculation module is used to acquire historical transfer bed usage data, real-time demand data, and spatial location information for each parking pile; extract a time-series feature set based on the historical transfer bed usage data; concatenate the time-series feature set, the real-time demand data, and the spatial location information to obtain an initial feature vector; input the initial feature vector to the contribution prediction module to calculate the expected impact weight to obtain a feature contribution matrix; and multiply the initial feature vector by the feature contribution matrix to obtain a weighted input feature vector. The generation module is used to input the weighted input feature vector into the dense network for forward propagation. Starting from the second network layer of the dense network, it generates a predecessor layer reference configuration based on the learning state of the output features of the previous network layer. Based on the predecessor layer reference configuration, a predecessor layer is selected as the input source. The output features of the selected predecessor layer are transformed and fused with structure to generate the input features of the current layer. The current layer output features are processed and propagated sequentially to the end of the network to obtain the network output result. The adjustment module is used to obtain the non-negative integer demand forecast value of each parking pile based on the network output results; and to calculate the difference between the demand forecast value of the transfer bed and the current transfer bed inventory of each parking pile to generate a transfer bed increase or decrease adjustment instruction.

[0046] The above description represents the preferred embodiments of the present invention. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for adjusting parking piles using a transfer bed based on demand forecasting, characterized in that, Includes the following steps: Obtain historical usage data, real-time demand data, and spatial location information of each parking pile for the transfer bed; Extract a time-series feature set based on the historical transfer bed usage data; The time-series feature set, the real-time demand data, and the spatial location information are concatenated to obtain an initial feature vector; the initial feature vector is input to the contribution prediction module to calculate the expected impact weights to obtain a feature contribution matrix; the initial feature vector is multiplied by the feature contribution matrix to obtain a weighted input feature vector; The weighted input feature vector is input into a dense network for forward propagation. Starting from the second network layer of the dense network, a predecessor layer reference configuration is generated based on the learning state of the output features of the previous network layer. A predecessor layer is selected as the input source according to the predecessor layer reference configuration. The output features of the selected predecessor layer are transformed and structured and fused to generate the input features of the current layer. The current layer output features are processed and propagated sequentially to the end of the network to obtain the network output result. Based on the network output results, the non-negative integer demand forecast values ​​for each parking pile are obtained; Based on the predicted demand for transfer beds and the current number of transfer beds at each parking post, a transfer bed increase / decrease adjustment instruction is generated by calculating the difference.

2. The method according to claim 1, characterized in that, The extraction of the time-series feature set based on the historical transfer bed usage data includes: The historical transfer bed usage data of each parking pile is arranged in ascending order by timestamp to form a time series set; Establish a short-cycle time window with a set duration of 15 minutes, a medium-cycle time window with a set duration of 1 hour, and a long-cycle time window with a set duration of 24 hours. Using the current moment as the cutoff endpoint, slide the short-cycle time window, the medium-cycle time window, and the long-cycle time window in reverse along the time axis on the time series set, respectively, and extract all historical event records within the three window periods; The number of borrowed and returned transfer beds generated in the records of each window is counted and spliced ​​together to form an independent data array. The data arrays generated by the three time windows of short period, medium period and long period are concatenated into a whole vector according to the dimensional order, and the whole vector is used as the time series feature set.

3. The method according to claim 2, characterized in that, The step of inputting the initial feature vector into the contribution prediction module to calculate the expected influence weights and obtain the feature contribution matrix includes: Set the internal network node parameter matrix for the contribution prediction module; The initial feature vector is multiplied by the network node parameter matrix to obtain the intermediate hidden layer feature vector; The dimension of the intermediate hidden layer feature vector is restored to the same data length as the initial feature vector by a fully connected mapping to obtain the mapped feature vector; A non-linear activation function is used to perform numerical mapping calculations on the mapped feature vector, converting each result value into a small value between 0 and 1, representing the influence weight value of each input feature; All influence weight values ​​are assembled into a one-dimensional weight array with the same data length according to the data arrangement order of the original initial feature vector. A diagonal matrix with each weight value as the diagonal element is constructed based on the one-dimensional weight array. The diagonal matrix is ​​the feature contribution matrix.

4. The method according to claim 1, characterized in that, The step of multiplying the initial feature vector by the feature contribution matrix to obtain the weighted input feature vector includes: Construct an output storage vector with all initial values ​​of zero and a size consistent with the dimension of the initial feature vector; In the traversal operation step, specific single numerical features of the current ranking of the initial feature vector are extracted sequentially, and the first numerical weights corresponding to the diagonal positions in the feature contribution matrix are extracted simultaneously. Perform an arithmetic multiplication operation on the currently extracted single numerical feature and the first numerical weight corresponding to the ranking of that feature to obtain the product value; The obtained product value is filled into the rank position corresponding to the output storage vector in its original position; The extraction and multiplication steps are repeated until all values ​​have been processed, and the output storage vector filled in this way is used as the weighted input feature vector.

5. The method according to claim 4, characterized in that, The step of generating the pre-sequence layer reference configuration based on the learning state of the output features of the previous network layer includes: The activation value variance and feature information entropy are calculated based on the output feature data of the previous network layer to obtain a statistical summary. The statistical summary is then concatenated with the layer number of the current processing node to form the learning state index vector of the current network layer. Set up a controller with multiple output branches, where each branch corresponds to a specific single layer among all the preceding layers; The learning state index vector is input into the classification function of the controller to map and calculate the reference connection probability score pointing to each specific single layer; Compare the reference connection probability scores of all branches with the preset reference threshold values. Retain the sequence number of the preceding layer corresponding to all probability scores greater than the preset reference threshold. If all probability scores are not greater than the preset reference threshold, retain the sequence number of the preceding layer corresponding to the largest probability score. Combine the sequence numbers to form the preceding layer reference configuration.

6. The method according to claim 1, characterized in that, The step of performing feature transformation and structured fusion on the selected output features of the preceding layer to generate the input features of the current layer includes: Based on the available connection target number in the preceding layer reference configuration, retrieve the first original multidimensional tensor set output by the specified preceding network layer node; A linear transformation algorithm is applied to each tensor in the first original multidimensional tensor set to reduce scale differences, and the data length of each tensor is aligned by interpolation resampling to obtain a standardized tensor set. The tensors in the standardized tensor group are concatenated along the channel dimension into a multi-channel fused tensor. The multi-channel fusion tensor is rearranged in an ordered manner according to the channel dimension and length dimension to form a one-dimensional convolutional input tensor that preserves the positional relationship of the channel dimension and length.

7. The method according to claim 1, characterized in that, The step of generating transfer bed increase / decrease adjustment instructions based on the difference between the predicted demand value of the transfer beds and the current transfer bed inventory of each parking pile includes: Read the non-negative integer-based demand forecast values ​​for each parking pile in the future scheduling cycle; Obtain the current inventory of transfer beds returned by sensor scans at each corresponding physical site; The difference in the number of transfer beds is obtained by subtracting the current number of transfer beds from the predicted demand value. If the quantity difference result is determined to be positive, then an action instruction is generated to transfer the same number of transfer beds as the difference value from the central warehouse or the available transfer bed parking piles to the corresponding parking pile position. If the difference value is determined to be negative, an action instruction is generated to retrieve and transport the same number of transfer beds as the absolute value of the difference value from the corresponding parking piles to the central warehouse. The action command information is combined and encapsulated with the corresponding parking pile identification number to generate a complete set of transfer bed increase / decrease adjustment command results, which is then sent to the scheduling system for dispatch and use.

8. A demand-forward-based parking pile adjustment system using a transfer bed, characterized in that, Includes the following modules: The calculation module is used to obtain historical transfer bed usage data, real-time demand data, and spatial location information for each parking pile; Based on the historical transfer bed usage data, a time-series feature set is extracted; the time-series feature set, the real-time demand data, and the spatial location information are concatenated to obtain an initial feature vector; the initial feature vector is input to the contribution prediction module to calculate the expected impact weight to obtain a feature contribution matrix; the initial feature vector is multiplied by the feature contribution matrix to obtain a weighted input feature vector; The generation module is used to input the weighted input feature vector into the dense network for forward propagation. Starting from the second network layer of the dense network, it generates a predecessor layer reference configuration based on the learning state of the output features of the previous network layer. Based on the predecessor layer reference configuration, a predecessor layer is selected as the input source. The output features of the selected predecessor layer are transformed and fused with structure to generate the input features of the current layer. The current layer output features are processed and propagated sequentially to the end of the network to obtain the network output result. The adjustment module is used to obtain the non-negative integer demand prediction value of each parking pile for the transfer bed based on the network output results; Based on the predicted demand for transfer beds and the current number of transfer beds at each parking post, a transfer bed increase / decrease adjustment instruction is generated by calculating the difference.

9. The system according to claim 8, characterized in that, The extraction of the time-series feature set based on the historical transfer bed usage data includes: The historical transfer bed usage data of each parking pile is arranged in ascending order by timestamp to form a time series set; Establish a short-cycle time window with a set duration of 15 minutes, a medium-cycle time window with a set duration of 1 hour, and a long-cycle time window with a set duration of 24 hours. Using the current moment as the cutoff endpoint, slide the short-cycle time window, the medium-cycle time window, and the long-cycle time window in reverse along the time axis on the time series set, respectively, and extract all historical event records within the three window periods; The number of borrowed and returned transfer beds generated in the records of each window is counted and spliced ​​together to form an independent data array. The data arrays generated by the three time windows of short period, medium period and long period are concatenated into a whole vector according to the dimensional order, and the whole vector is used as the time series feature set.

10. The system according to claim 8, characterized in that, The step of inputting the initial feature vector into the contribution prediction module to calculate the expected influence weights and obtain the feature contribution matrix includes: Set the internal network node parameter matrix for the contribution prediction module; The initial feature vector is multiplied by the network node parameter matrix to obtain the intermediate hidden layer feature vector; The dimension of the intermediate hidden layer feature vector is restored to the same data length as the initial feature vector by a fully connected mapping to obtain the mapped feature vector; A non-linear activation function is used to perform numerical mapping calculations on the mapped feature vector, converting each result value into a small value between 0 and 1, representing the influence weight value of each input feature; All influence weight values ​​are assembled into a one-dimensional weight array with the same data length according to the data arrangement order of the original initial feature vector. A diagonal matrix with each weight value as the diagonal element is constructed based on the one-dimensional weight array. The diagonal matrix is ​​the feature contribution matrix.