Hospital computing power capacity prediction method based on strategy optimization

Abstract

The present application relates to the technical field of hospital computing power prediction, and discloses a hospital computing power capacity prediction method based on strategy optimization. The method collects hospital business request logs, image archiving call records and computing power node resource occupation sequences, and constructs a three-dimensional original tensor; according to the dynamic window division of the treatment wave boundary, the business semantics and computing power load representation are fused through cross attention; the deformable convolution and the gated recurrent unit stacking network are adopted to output the future multi-period computing power capacity prediction sequence; the prediction sequence is mapped to the expansion and contraction instructions according to the computing power node topology, and the window sliding step is dynamically adjusted according to the prediction residual, so as to realize the self-optimization of the prediction strategy, improve the forward-looking of hospital computing power scheduling and the resource utilization rate.

Description

Technical Field

[0001] This invention relates to the field of hospital computing power prediction technology, and specifically to a method for predicting hospital computing power capacity based on strategy optimization. Background Technology

[0002] As hospitals become increasingly information-based, the resource consumption of backend computing nodes is growing daily due to the operation of various business systems such as outpatient registration, emergency reception, inpatient registration, examination appointments, surgical scheduling, and medical image archiving and communication systems. Traditional computing capacity management relies heavily on the experience and judgment of operations and maintenance personnel or simple static threshold rules, which are insufficient to cope with the drastic fluctuations in hospital business traffic caused by factors such as patient flow and holidays.

[0003] Existing prediction methods typically only consider time-series data in a single dimension, ignoring the coupling relationship between different types of business requests and the load of computing nodes, and lack the ability to adaptively identify the boundaries of business waves, resulting in prediction results lagging behind actual load changes, which in turn leads to excessive reservation or insufficient supply of computing resources.

[0004] Furthermore, prediction models with fixed window lengths cannot dynamically adjust the granularity of analysis based on prediction errors, causing the model to gradually deviate from its optimal performance over long-term operation. Therefore, there is an urgent need for a hospital computing capacity prediction method that can deeply integrate multi-source heterogeneous data, adaptively identify business waves, and dynamically optimize prediction strategies. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the purpose of this invention is to provide a strategy-optimized method for predicting hospital computing capacity, thereby solving the problem of low accuracy in existing hospital computing capacity prediction methods.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides a hospital computing capacity prediction method based on strategy optimization, the method comprising: Step S1: Collect business request logs, image archiving call records, and computing node resource usage sequences from the hospital information system. Align them by timestamp and perform nearest neighbor interpolation on missing values ​​to construct a three-dimensional original tensor. The dimensions are time step, business type identifier, and computing node identifier, respectively. Step S2: Dynamically divide the original tensor into windows based on the boundary of the wave of visits. Encode the business semantic embedding along the business type identifier dimension and the computing load embedding along the computing node identifier dimension for each window. Merge the two types of embeddings into a single-window coupling representation matrix through a cross-attention mechanism. Stack the coupling representation matrices of all windows in chronological order to form a multi-window coupling representation sequence. Step S3: Input the multi-window coupled representation sequence into the prediction network consisting of deformable convolutional layers and gated recurrent units. The deformable convolutional layers learn offsets on the feature plane to extract local patterns, and the gated recurrent units capture long-term dependencies between windows, outputting a future multi-period computing capacity prediction sequence. Step S4: Map the predicted sequence to a scaling up / down instruction sequence according to the computing node topology. Update the sliding step size of the dynamic window in step S2 based on the residual between the predicted sequence and the actual value collected in the next round. Return to step S1 to enter the next round of prediction.

[0007] Preferably, in one possible implementation of the first aspect, the service request log in step S1 extracts the arrival time and number of requests for outpatient registration, emergency admission, inpatient registration, examination appointment, and surgery arrangement. The image archiving retrieval record extracts the retrieval start time, the length of the retrieved image sequence, and the target storage node. The resource usage sequence of computing nodes extracts the CPU utilization rate, memory utilization rate, graphics processor utilization rate, and input / output throughput.

[0008] Preferably, in one possible implementation of the first aspect, the three-dimensional primitive tensor construction process includes: Align the request arrival time and request volume of the business request log, the call start time and call image sequence length of the image archive call record with the target storage node, and the CPU utilization, memory utilization, graphics processor utilization, and input / output throughput of the computing power node resource occupancy sequence to a unified time grid by timestamp. When a missing value appears at any position after alignment, nearest neighbor non-missing values ​​are taken along the time step dimension for nearest neighbor interpolation to fill the missing value. After filling, the components are stacked in dimensional order to form a three-dimensional primitive tensor. The first dimension of the three-dimensional primitive tensor is the time step, the second dimension is the business type identifier, and the third dimension is the computing power node identifier. The tensor element values ​​are the load of the corresponding business type on the corresponding computing power node at the corresponding timestamp.

[0009] Preferably, in one possible implementation of the first aspect, the dynamic window segmentation in step S2 includes: The boundary of the outpatient wave is detected along the time step dimension of the three-dimensional original tensor. The boundary of the outpatient wave is determined by the zero-crossing position of the daily and weekly periodic differences of the outpatient registration request volume, emergency room admission request volume, inpatient registration request volume, examination appointment request volume, and surgical arrangement request volume. Two adjacent patient visit wave boundaries are defined as a wave interval, and the start and end points of the wave interval are taken as the window boundaries. The three-dimensional original tensor within each wave interval is truncated along the time step dimension to obtain several dynamic windows whose length varies with the wave interval. The time steps within the windows are continuous, and the windows are connected end to end in chronological order.

[0010] Preferably, in one possible implementation of the first aspect, the fusion process of the single-window coupled representation matrix includes: For each dynamic window, the business semantic encoder is input along the business type identifier dimension. The business semantic encoder consists of a position encoding layer and a linear projection layer, and outputs a sequence of business semantic embedding vectors. The computing load encoder is input along the dimension of the computing node identifier. The computing load encoder consists of a position encoding layer and a linear projection layer, and outputs a computing load embedding vector sequence. Using business semantic embedding as the query and computing load embedding as the key and value, cross-attention calculation is performed to obtain a business-aware computing power representation. Using computing load embedding as the query and business semantic embedding as the key and value, cross-attention calculation is performed to obtain computing load-aware business representation; The two representations are concatenated according to their feature dimensions and then passed through a feedforward network and layer normalization output to form the coupled representation matrix of the current window.

[0011] Preferably, in one possible implementation of the first aspect, the stacking process of the multi-window coupled characterization sequences includes: The coupling representation matrix obtained by fusing each dynamic window through cross attention is arranged in the order of the wave interval to which the window belongs on the original time axis. When arranging, layer normalization is performed on the coupling representation matrix of adjacent windows to ensure consistent numerical scale; The normalized window coupling representation matrices are stacked along the window number dimension, and a learnable positional encoding is added to the front of the stacked tensor to preserve the temporal relationship between windows, resulting in a multi-window coupling representation sequence. The first dimension of the multi-window coupling representation sequence is the window number, the second dimension is the business type identifier, and the third dimension is the computing power node identifier, which are fused feature channels.

[0012] Preferably, in one possible implementation of the first aspect, the structure of the prediction network in step S3 includes: The prediction network consists of deformable convolutional layers, batch normalization layers, nonlinear activation layers, gated recurrent unit layers, and fully connected output layers stacked from bottom to top. Deformable convolutional layers receive multi-window coupled representation sequences as input and predict offsets at each spatial location of the feature plane. The offsets act on the sampling grid so that the convolutional kernels adaptively fit the load distribution shape. Batch normalization layers normalize the channel dimension of deformable convolution outputs; The nonlinear activation layer uses modified linear units; The gated loop unit layer consists of two layers of gated loop units stacked together, with dropout set between layers to capture long-term dependencies in the window sequence dimension; The fully connected output layer maps the hidden state of the gated cyclic unit to a future multi-period computing capacity prediction sequence.

[0013] Preferably, in one possible implementation of the first aspect, the specific process by which the deformable convolutional layer learns offsets on the feature plane to extract local patterns includes: For each spatial location of the multi-window coupled representation sequence, the offset prediction convolutional layer outputs the two-dimensional offset of that location on the feature plane. The first component of the offset corresponds to the shift of the business type identifier dimension, and the second component corresponds to the shift of the computing power node identifier dimension. The subpixel coordinates are obtained by adding the integer grid coordinates to the offset, and bilinear interpolation is performed on the feature values ​​at the subpixel coordinates to obtain the deformed features. The deformed features are fed into a standard convolution kernel for convolution operation, and the output is a feature map weighted by local patterns. The specific process of the gated recurrent unit capturing long-term dependencies between windows includes: unfolding the feature map output by the deformable convolutional layer into a temporal input along the window number dimension, calculating the current hidden state sequentially through the reset gate, update gate, and candidate activation values, and passing the hidden state of the last window to the fully connected output layer.

[0014] Preferably, in one possible implementation of the first aspect, the output process of the computing power capacity prediction sequence includes: The fully connected output layer receives the hidden state of the last window of the gated loop unit layer, and outputs the computing power capacity prediction vector for the next T time steps after linear transformation, where T is the preset prediction step size. The dimension of the computing power capacity prediction vector at each time step is consistent with the number of computing power nodes. Each component in the vector corresponds to the four types of resource occupancy prediction values ​​of the corresponding computing power node at that time step. The four types of resource occupancy prediction values ​​are, in order, the CPU occupancy prediction value, the memory occupancy prediction value, the GPU occupancy prediction value, and the input / output throughput prediction value. The computing capacity prediction vectors of T time steps are concatenated in chronological order to form a future multi-period computing capacity prediction sequence.

[0015] Preferably, in one possible implementation of the first aspect, the topology mapping process of the expansion / contraction instruction sequence in step S4 includes: Read the physical topology map of the hospital's computing power nodes. Each node in the topology map is labeled with the type of service it carries and its communication adjacency relationship. The node load ratio is obtained by dividing the predicted value of the four types of resource occupancy for each node at each time step in the computing power capacity prediction sequence with the corresponding node's capacity limit. The instruction category of the node in the corresponding time step is determined based on the node load ratio and the preset three threshold intervals. The three threshold intervals are shrinkage interval, maintenance interval, and expansion interval, and the instruction categories are shrinkage, maintenance, and expansion, respectively. Arrange the instruction categories of each node according to the node order in the topology graph to form an expansion and contraction instruction sequence that corresponds one-to-one with the time step of the prediction sequence. Calculate the sum of the absolute values ​​of the pointwise residuals between the predicted sequence and the actual resource occupancy sequence collected in the next round. Adjust the sliding step size of the dynamic window in step S2 in reverse order according to the magnitude of the sum of the absolute values ​​of the residuals. When the sum of the absolute values ​​of the residuals increases, decrease the sliding step size. When the sum of the absolute values ​​of the residuals decreases, increase the sliding step size.

[0016] The beneficial effects of this invention are as follows: This invention achieves unified representation and alignment of multi-source heterogeneous data by collecting business request logs, image archiving call records and computing node resource occupancy sequences from the hospital information system, and constructing a three-dimensional original tensor using nearest neighbor interpolation.

[0017] Dynamic window segmentation is performed based on the boundaries of the consultation waves, and business semantics and computing load features are integrated by combining cross-attention mechanism, which effectively captures the deep coupling relationship between business type and computing power node and improves the pertinence of feature expression.

[0018] The prediction network employs a combination of deformable convolutional layers and gated recurrent units. The deformable convolutional layers can adaptively learn the local offset patterns of the load distribution, while the gated recurrent units are adept at capturing long-term temporal dependencies between windows. The combination of the two significantly improves the accuracy of future multi-period computing capacity prediction.

[0019] Furthermore, by mapping the prediction sequence to a scaling up / down instruction sequence and dynamically adjusting the sliding step size based on the prediction residual, adaptive optimization of the prediction strategy is achieved, enabling the model to continuously track load change trends and avoiding performance degradation caused by fixed parameters.

[0020] Overall, this invention significantly improves the foresight and flexibility of hospital computing resource scheduling, reduces operation and maintenance costs, and ensures the efficient and stable operation of critical businesses. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart of a hospital computing capacity prediction method based on strategy optimization is provided for this application. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1: As Figure 1 As shown, this invention provides a hospital computing capacity prediction method based on strategy optimization, including: Step S1: Collect business request logs, image archiving call records, and computing node resource usage sequences from the hospital information system. Align them by timestamp and perform nearest neighbor interpolation on missing values ​​to construct a three-dimensional original tensor. The dimensions are time step, business type identifier, and computing node identifier, respectively.

[0025] In this embodiment, business request logs are collected from the hospital information system database. These logs include the arrival time and volume of requests for five types of services: outpatient registration, emergency room admission, inpatient registration, examination appointment, and surgery scheduling. Image archiving call records are collected from the medical image archiving and communication system database. These records include the start time of each call, the length of the called image sequence, and the identifier of the target storage node for this call. Resource usage sequences for computing nodes are collected from the server cluster monitoring platform. These sequences include four indicators for each computing node: CPU utilization, memory utilization, GPU utilization, and I / O throughput.

[0026] All timestamp fields in the collected business request logs, image archiving call records, and computing node resource usage sequences are uniformly aligned to a fixed time grid. This time grid is divided into 60-second intervals, with each time step in the grid corresponding to a 60-second time interval. During the alignment process, if a data item is missing within a certain time step, the nearest non-missing value is searched forward along the time step dimension, and then the nearest non-missing value is searched backward. The arithmetic mean of these two non-missing values ​​is taken as the filler value for the missing position.

[0027] After imputing missing values, the data is organized into a three-dimensional primitive tensor. The first dimension of the three-dimensional primitive tensor is the time step number, the second dimension is the business type identifier, and the third dimension is the computing node identifier. The index is located within the three-dimensional primitive tensor. element value at Indicates the first Within the first time step The first type of business The formula for calculating the total load generated on each computing node is as follows:

[0028] in For the first Within the first time step The first type of business Request volume per computing node For the first Within the first time step The first type of business Total length of image call sequence triggered on each computing node For the first Within the first time step Comprehensive resource utilization metrics for each computing node. The average of the normalized utilization rates of the four resources is obtained as follows:

[0029] In the formula , , , They represent the first Within the first time step CPU utilization, memory utilization, GPU utilization, and I / O throughput of each computing node; , , , These are the rated maximum values ​​for the corresponding indicators. Weighting coefficients. , , Based on historical data regression, in this embodiment, we take... , , The final shape of the original three-dimensional tensor is as follows: ,in The total number of time steps. For the number of business types, This represents the number of computing nodes.

[0030] Step S2: Dynamically divide the original tensor into windows based on the boundary of the treatment wave. Encode the business semantic embedding along the business type identifier dimension and the computing load embedding along the computing node identifier dimension for each window. Merge the two types of embeddings into a single-window coupling representation matrix through a cross-attention mechanism. Stack the coupling representation matrices of all windows in chronological order to form a multi-window coupling representation sequence.

[0031] In this embodiment, visitor wave boundary detection is performed along the time step dimension of the three-dimensional original tensor. The visitor wave boundary is determined by the zero-crossing positions of the daily and weekly periodic differences of outpatient registration requests, emergency room visits, inpatient registration requests, examination appointment requests, and surgical scheduling requests. Let the first... The first time step The request volume for this type of service is The daily periodic difference is defined as The periodic difference is defined as The time steps are in minutes, with 1440 representing the number of time steps in a day and 10080 representing the number of time steps in a week.

[0032] When any kind of business exists that makes and When the time step changes from positive to negative or from negative to positive, that time step Marked as patient visit wave boundaries. Using two adjacent patient visit wave boundaries... and For a wave interval, take and As the window boundary, the three-dimensional original tensor within each wave interval is truncated along the time step dimension to obtain the dynamic window. Window length is The different windows are arranged sequentially from beginning to end according to time.

[0033] For each dynamic window The business semantic encoder is input along the business type identifier dimension. The business semantic encoder consists of a position encoding layer and a linear projection layer. The position encoding layer uses sine and cosine functions to encode the position at each time step within the window, and the position encoding dimension is set to... For the first [item] in the window time step Location encoding vector The peacekeeping The dimensions are as follows:

[0034] in The location code is added to the service load features at each time step within the window along the feature dimension, resulting in a dimension of... The intermediate features are obtained. The linear projection layer uses a fully connected layer to map the 128-dimensional features to 256 dimensions, with the activation function being the linear identity mapping. The output is a sequence of business semantic embedding vectors. , where 5 represents the number of business types.

[0035] The computational load encoder is input along the dimension of the computational node identifier. The structure of the computational load encoder is the same as that of the business semantic encoder, with the position encoding dimension also set to 128. The linear projection layer maps the 128 dimensions to 256 dimensions. The output is a sequence of computational load embedding vectors. ,in This represents the number of computing nodes.

[0036] The cross-attention module performs calculations using business semantic embeddings as queries and computing load embeddings as keys and values. The input dimensions of the cross-attention module are as follows: Attention count Dimensions of each head Query matrix Key matrix Value matrix ,in This is a learnable weight matrix. The attention weights are calculated as follows:

[0037] Multi-head attention concatenates the outputs of all heads and then projects them linearly to obtain a business-aware computing power representation. Symmetrically, using the computing load embedding as the query and the business semantic embedding as the key and value, the same multi-head cross-attention computation is performed to obtain the computing load-aware business representation. .

[0038] The two representations are concatenated along the feature dimension to obtain The concatenated features are input into a feedforward network, which consists of two fully connected layers: the first layer maps 512 dimensions to 1024 dimensions, and the activation function is... The second layer maps the 1024 dimensions back to 512 dimensions without an activation function. The output of the feedforward network undergoes layer normalization, which independently calculates the mean for each sample's feature dimension. With variance The normalization formula is:

[0039] in , These are learnable parameters. The coupling representation matrix of the current window is obtained. .

[0040] The coupling representation matrices of all dynamic windows are arranged in chronological order according to the wave intervals on the original time axis. After arrangement, layer normalization is performed on the coupling representation matrices of adjacent windows, with the same normalization parameters as described above, ensuring that the numerical scales of each window are consistent after normalization. The normalized coupling representation matrices of each window are then stacked along the window index dimension. Due to the length of each window... If they are not equal, take the maximum window length. The coupling representation matrix for each window is padded with zeros at the end of the time step dimension to the length. The eigenvalues ​​of the zero-padding portion are all set to zero. The shape of the stacked tensor is... ,in The total number of windows. Learnable positional encodings are appended to the front of the stacked tensors. The positional encoding is randomly initialized using a normal distribution with a mean of 0 and a standard deviation of 0.02, and updated via backpropagation during training. During appending, P is expanded to the same dimension as the stacked tensor and then summed element-wise. The final result is a multi-window coupled representation sequence. The first dimension is the window number, the second dimension is the time step number within the window, the third dimension is the business type identifier, the fourth dimension is the computing power node identifier, and the fifth dimension is the fusion feature channel.

[0041] Step S3: Input the multi-window coupled representation sequence into the prediction network consisting of deformable convolutional layers and gated recurrent units. The deformable convolutional layers learn offsets on the feature plane to extract local patterns, and the gated recurrent units capture long-term dependencies between windows, outputting a prediction sequence of computing capacity for future multi-period periods.

[0042] In this embodiment, the prediction network is composed of deformable convolutional layers, batch normalization layers, nonlinear activation layers, gated recurrent unit layers, and fully connected output layers stacked from bottom to top. The deformable convolutional layers receive a multi-window coupled representation sequence as input, and the shape of this multi-window coupled representation sequence is... ,in The total number of windows, The maximum window length, For the number of business types, 512 represents the number of computing nodes and the fusion feature channel dimension. The kernel size of the deformable convolutional layer is set to... The input channel count is 512, the output channel count is 512, the step size is 1, and the padding method is... .

[0043] At each spatial location on the feature plane, the deformable convolutional layer outputs a two-dimensional offset through an independent offset prediction convolutional layer with a kernel size of [size missing]. The number of input channels is 512 and the number of output channels is 18, which means that each spatial location corresponds to a two-dimensional offset of 9 sampling points.

[0044] The first component of the offset corresponds to the shift in the business type identifier dimension, and the second component corresponds to the shift in the computing node identifier dimension. For the multi-window coupled representation sequence, the coordinates are... The spatial location, in which Corresponding business type identifier dimension index, The offset prediction convolutional layer outputs the offset of 9 sampling points at the corresponding computing node identifier dimension index. , The value range is from 0 to 8.

[0045] The sub-pixel coordinates are obtained by adding the integer grid coordinates to the offset. The formula for calculating the sub-pixel coordinates is as follows: ,in The original integer grid coordinates are used. Bilinear interpolation is performed on the feature values ​​at sub-pixel coordinates to obtain the deformed features. The formula for bilinear interpolation is: ,in Integer grid coordinates It is a bilinear interpolation kernel function. .

[0046] The deformed features are sent to the standard. The convolution kernel performs a convolution operation, and the output feature map of the convolution operation is... ,in These are the kernel weights. The feature map output by the deformable convolutional layer has the following shape: The batch normalization layer performs channel-dimensional normalization on the deformable convolution output. The calculation formula for the batch normalization layer is as follows: ,in This represents the mean value of the channel dimension within the batch. The variance of the channel dimension within the batch. Values , and For learnable parameters, the initial value is... Set to 1, Set to 0.

[0047] The nonlinear activation layer uses a modified linear unit. The activation function is The gated loop unit layer consists of two stacked gated loop units. The first gated loop unit has an input dimension of 512 and a hidden layer dimension of 512, as does the second gated loop unit. A space is set between the two gated loop unit layers. , The ratio is set to 0.2.

[0048] The specific process by which the gated recurrent unit captures long-term dependencies between windows is as follows: the feature map output by the deformable convolutional layer is unfolded along the window index dimension into a temporal input, and for each window index... The shape of the input feature map is Flatten the feature map along the last three dimensions into a one-dimensional vector. The dimensions of the flattened vector are... The flattened vector is used as a gated loop unit at time step. Input .

[0049] The formula for calculating the reset gate of the gated loop unit is as follows: ,in To reset the gate weight matrix, To reset the gate bias vector, for Activation function This sets the previous window to a hidden state. The update door calculation formula is as follows: ,in To update the gate weight matrix, To update the gate bias vector, the candidate activation value is calculated using the following formula: ,in The candidate activation value weight matrix, This is the candidate activation value bias vector. This represents element-wise multiplication. The formula for calculating the current hidden state is: .

[0050] After processing through two layers of gated recurrent units, the hidden state of the last window is passed to the fully connected output layer. The hidden state of the last window has a dimension of 512. The fully connected output layer receives the hidden state of the last window from the gated recurrent unit layer, performs a linear transformation, and outputs the future... The computing capacity prediction vector at each time step. In this embodiment, the preset prediction step size is used. The value is 12. The dimension of the weight matrix of the fully connected output layer is... The bias vector dimension is The formula for calculating the linear transformation is: ,in This is the weight matrix of the fully connected layer. This is the bias vector for the fully connected layer. This sets the last window to a hidden state.

[0051] The output dimension of a fully connected output layer is Reshape the output to the shape of The tensor. The dimension of the computing power capacity prediction vector at each time step and the number of computing power nodes. Consistent, each component in the vector corresponds to one of the four types of resource utilization prediction values ​​for the corresponding computing node at that time step. These four types of resource utilization prediction values ​​are, in order: CPU utilization prediction value, memory utilization prediction value, GPU utilization prediction value, and I / O throughput prediction value. The computing capacity prediction vectors from T time steps are concatenated in chronological order to form a future multi-period computing capacity prediction sequence. The shape of this prediction sequence is as follows: The training process of the prediction network adopts... The optimizer has an initial learning rate of 0.001, and the learning rate is increased every 10 training cycles. The learning rate decayed to 0.9 times its original value.

[0052] The loss function uses mean squared error. The calculation formula is: ,in This represents the actual resource usage value. To predict resource usage, Total number of samples. Batch size. Set to 32, number of training rounds Set it to 200.

[0053] Step S4: Map the predicted sequence to a scaling up / down instruction sequence according to the computing node topology. Update the sliding step size of the dynamic window in step S2 based on the residual between the predicted sequence and the actual value collected in the next round. Return to step S1 to enter the next round of prediction.

[0054] In this embodiment, the physical topology map of the computing power nodes is first read from the hospital data center. This topology map is stored in the form of an adjacency list, where each node records its unique identifier, a list of the service types it carries (including outpatient registration, emergency reception, inpatient registration, examination appointment, and surgery scheduling), and a list of communication adjacency nodes. For the computing power capacity prediction sequence for the next T time steps, the shape is... ,in To predict the step size, 4 represents the total number of computing nodes, corresponding to the predicted CPU utilization, memory utilization, graphics processor utilization, and input / output throughput.

[0055] For each time step and each node , will node Dividing the predicted resource occupancy values ​​of the four types by their corresponding capacity limits yields four dimensionless ratios: CPU load ratio Memory load ratio Graphics processor load ratio Input / output load ratio .

[0056] in , , , For nodes The rated capacity limit is predetermined by the hardware configuration. The four ratios are weighted and summed with weights of 0.3, 0.3, 0.2, and 0.2 to obtain the node's capacity. At time step Overall node load ratio Three preset threshold ranges are defined: the shrinkage range is... The maintenance interval is The expansion range is .like Then determine the node At time step The instruction category is shrinking; if If, then it is determined to be maintained; if If so, it is determined to be an expansion.

[0057] Arrange the instruction categories of all nodes within the same time step in ascending order of node identifiers in the topology graph, forming a list of length [length missing]. The instruction vector has elements where 0 represents shrinking, 1 represents maintaining, and 2 represents expanding. Concatenating the instruction vectors from the 12 time steps in chronological order yields the shrinking / expanding instruction sequence, with the shape [formula missing]. .

[0058] After generating the instruction sequence, wait for the next data collection cycle to arrive and obtain the actual resource usage sequence for the time period corresponding to the predicted sequence. The shape is also [missing information]. Calculate the sum of the absolute values ​​of the pointwise residuals between the predicted and actual sequences: ,in Iterate through the four resource types.

[0059] Record the sum of the absolute values ​​of the residuals from the previous prediction. (Initial value set to 0). If Then multiply the sliding step size of the dynamic window in step S2 by 0.9; if If the value is equal to 1, the sliding step size is multiplied by 1.1; otherwise, it remains unchanged. The sliding step size has a lower limit of 1 and an upper limit of 10, and the boundary value is used when the value exceeds the range. The updated sliding step size is used in the dynamic window segmentation process of step S2 in the next round of prediction, and then returns to step S1 to start a new round of data collection and prediction.

[0060] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A hospital computing capacity prediction method based on strategy optimization, characterized in that, The method includes: Step S1: Collect business request logs, image archiving call records, and computing node resource usage sequences from the hospital information system. Align them by timestamp and perform nearest neighbor interpolation on missing values ​​to construct a three-dimensional original tensor. The dimensions are time step, business type identifier, and computing node identifier, respectively. Step S2: Dynamically divide the original tensor into windows based on the boundary of the wave of visits. Encode the business semantic embedding along the business type identifier dimension and the computing load embedding along the computing node identifier dimension for each window. Merge the two types of embeddings into a single-window coupling representation matrix through a cross-attention mechanism. Stack the coupling representation matrices of all windows in chronological order to form a multi-window coupling representation sequence. Step S3: Input the multi-window coupled representation sequence into the prediction network consisting of deformable convolutional layers and gated recurrent units. The deformable convolutional layers learn offsets on the feature plane to extract local patterns, and the gated recurrent units capture long-term dependencies between windows, outputting a future multi-period computing capacity prediction sequence. Step S4: Map the predicted sequence to a scaling up / down instruction sequence according to the computing node topology. Update the sliding step size of the dynamic window in step S2 based on the residual between the predicted sequence and the actual value collected in the next round. Return to step S1 to enter the next round of prediction.

2. The hospital computing capacity prediction method based on strategy optimization as described in claim 1, characterized in that, In step S1, the service request logs are extracted to determine the arrival time and number of requests for outpatient registration, emergency admission, hospitalization registration, examination appointment, and surgery arrangement. The image archiving retrieval record extracts the retrieval start time, the length of the retrieved image sequence, and the target storage node. The resource usage sequence of computing nodes extracts the CPU utilization rate, memory utilization rate, graphics processor utilization rate, and input / output throughput.

3. The hospital computing capacity prediction method based on strategy optimization as described in claim 2, characterized in that, The process of constructing the three-dimensional primitive tensor includes: Align the request arrival time and request volume of the business request log, the call start time and call image sequence length of the image archive call record with the target storage node, and the CPU utilization, memory utilization, graphics processor utilization, and input / output throughput of the computing power node resource occupancy sequence to a unified time grid by timestamp. When a missing value appears at any position after alignment, nearest neighbor non-missing values ​​are taken along the time step dimension for nearest neighbor interpolation to fill the missing value. After filling, the components are stacked in dimensional order to form a three-dimensional primitive tensor. The first dimension of the three-dimensional primitive tensor is the time step, the second dimension is the business type identifier, and the third dimension is the computing power node identifier. The tensor element values ​​are the load of the corresponding business type on the corresponding computing power node at the corresponding timestamp.

4. The hospital computing capacity prediction method based on strategy optimization as described in claim 1, characterized in that, The dynamic window segmentation in step S2 includes: The boundary of the outpatient wave is detected along the time step dimension of the three-dimensional original tensor. The boundary of the outpatient wave is determined by the zero-crossing position of the daily and weekly periodic differences of the outpatient registration request volume, emergency room admission request volume, inpatient registration request volume, examination appointment request volume, and surgical arrangement request volume. Two adjacent patient visit wave boundaries are defined as a wave interval, and the start and end points of the wave interval are taken as the window boundaries. The three-dimensional original tensor within each wave interval is truncated along the time step dimension to obtain several dynamic windows whose length varies with the wave interval. The time steps within the windows are continuous, and the windows are connected end to end in chronological order.

5. The hospital computing capacity prediction method based on strategy optimization as described in claim 4, characterized in that, The fusion process of the single-window coupled representation matrix includes: For each dynamic window, the business semantic encoder is input along the business type identifier dimension. The business semantic encoder consists of a position encoding layer and a linear projection layer, and outputs a sequence of business semantic embedding vectors. The computing load encoder is input along the dimension of the computing node identifier. The computing load encoder consists of a position encoding layer and a linear projection layer, and outputs a computing load embedding vector sequence. Using business semantic embedding as the query and computing load embedding as the key and value, cross-attention calculation is performed to obtain a business-aware computing power representation. Using computing load embedding as the query and business semantic embedding as the key and value, cross-attention calculation is performed to obtain computing load-aware business representation; The two representations are concatenated according to their feature dimensions and then passed through a feedforward network and layer normalization output to form the coupled representation matrix of the current window.

6. The hospital computing capacity prediction method based on strategy optimization as described in claim 5, characterized in that, The stacking process of the multi-window coupled representation sequences includes: The coupling representation matrix obtained by fusing each dynamic window through cross attention is arranged in the order of the wave interval to which the window belongs on the original time axis. When arranging, layer normalization is performed on the coupling representation matrix of adjacent windows to ensure consistent numerical scale; The normalized window coupling representation matrices are stacked along the window number dimension, and a learnable positional encoding is added to the front of the stacked tensor to preserve the temporal relationship between windows, resulting in a multi-window coupling representation sequence. The first dimension of the multi-window coupling representation sequence is the window number, the second dimension is the business type identifier, and the third dimension is the computing power node identifier, which are fused feature channels.

7. The hospital computing capacity prediction method based on strategy optimization as described in claim 1, characterized in that, The structure of the prediction network in step S3 includes: The prediction network consists of deformable convolutional layers, batch normalization layers, nonlinear activation layers, gated recurrent unit layers, and fully connected output layers stacked from bottom to top. Deformable convolutional layers receive multi-window coupled representation sequences as input and predict offsets at each spatial location of the feature plane. The offsets act on the sampling grid so that the convolutional kernels adaptively fit the load distribution shape. Batch normalization layers normalize the channel dimension of deformable convolution outputs; The nonlinear activation layer uses modified linear units; The gated loop unit layer consists of two layers of gated loop units stacked together, with dropout set between layers to capture long-term dependencies in the window sequence dimension; The fully connected output layer maps the hidden state of the gated cyclic unit to a future multi-period computing capacity prediction sequence.

8. The hospital computing capacity prediction method based on strategy optimization as described in claim 7, characterized in that, The specific process by which the deformable convolutional layer learns offsets on the feature plane to extract local patterns includes: For each spatial location of the multi-window coupled representation sequence, the offset prediction convolutional layer outputs the two-dimensional offset of that location on the feature plane. The first component of the offset corresponds to the shift of the business type identifier dimension, and the second component corresponds to the shift of the computing power node identifier dimension. The subpixel coordinates are obtained by adding the integer grid coordinates to the offset, and bilinear interpolation is performed on the feature values ​​at the subpixel coordinates to obtain the deformed features. The deformed features are fed into a standard convolution kernel for convolution operation, and the output is a feature map weighted by local patterns. The specific process of the gated recurrent unit capturing long-term dependencies between windows includes: unfolding the feature map output by the deformable convolutional layer into a temporal input along the window number dimension, calculating the current hidden state sequentially through the reset gate, update gate, and candidate activation values, and passing the hidden state of the last window to the fully connected output layer.

9. The hospital computing capacity prediction method based on strategy optimization as described in claim 8, characterized in that, The output process of the computing power capacity prediction sequence includes: The fully connected output layer receives the hidden state of the last window of the gated loop unit layer, and outputs the computing power capacity prediction vector for the next T time steps after linear transformation, where T is the preset prediction step size. The dimension of the computing power capacity prediction vector at each time step is consistent with the number of computing power nodes. Each component in the vector corresponds to the four types of resource occupancy prediction values ​​of the corresponding computing power node at that time step. The four types of resource occupancy prediction values ​​are, in order, the CPU occupancy prediction value, the memory occupancy prediction value, the GPU occupancy prediction value, and the input / output throughput prediction value. The computing capacity prediction vectors of T time steps are concatenated in chronological order to form a future multi-period computing capacity prediction sequence.

10. The hospital computing capacity prediction method based on strategy optimization as described in claim 1, characterized in that, The topology mapping process of the expansion and contraction instruction sequence in step S4 includes: Read the physical topology map of the hospital's computing power nodes. Each node in the topology map is labeled with the type of service it carries and its communication adjacency relationship. The node load ratio is obtained by dividing the predicted value of the four types of resource occupancy for each node at each time step in the computing power capacity prediction sequence with the corresponding node's capacity limit. The instruction category of the node in the corresponding time step is determined based on the node load ratio and the preset three threshold intervals. The three threshold intervals are shrinkage interval, maintenance interval, and expansion interval, and the instruction categories are shrinkage, maintenance, and expansion, respectively. Arrange the instruction categories of each node according to the node order in the topology graph to form an expansion and contraction instruction sequence that corresponds one-to-one with the time step of the prediction sequence. Calculate the sum of the absolute values ​​of the pointwise residuals between the predicted sequence and the actual resource occupancy sequence collected in the next round. Adjust the sliding step size of the dynamic window in step S2 in reverse order according to the magnitude of the sum of the absolute values ​​of the residuals. When the sum of the absolute values ​​of the residuals increases, decrease the sliding step size. When the sum of the absolute values ​​of the residuals decreases, increase the sliding step size.

Images