Method, device, medium and program product for shared base station load energy efficiency prediction

Abstract

Embodiments of the present application provide a shared base station load energy efficiency prediction method, device, medium and program product. A multi-dimensional feature vector sequence corresponding to a preset historical period of a shared base station is obtained. The multi-dimensional feature vector sequence is processed by a long short-term memory network to obtain a long-term feature of the shared base station. The long-term feature is processed by a time attention mechanism to obtain a time-enhanced feature. The long-term feature is processed by a feature attention mechanism to obtain a dimension-enhanced feature. The time-enhanced feature and the dimension-enhanced feature are weighted and fused to obtain an initial fused feature. The initial fused feature and the long-term feature are connected by a residual connection to obtain a fused feature. The fused feature is mapped according to a preset mapping manner of a preset load energy efficiency indicator to obtain a predicted load energy efficiency indicator at a preset future time point. According to the method provided by the embodiments of the present application, the accuracy of the predicted load energy efficiency indicator is improved.

Description

Technical Field

[0001] This application belongs to the field of data processing technology, and in particular relates to a method, device, medium and program product for predicting the load energy efficiency of a shared base station. Background Technology

[0002] In existing technologies, with the evolution from 5G to 6G, the characteristics of high-frequency bands necessitate a significant increase in base station density, leading to a surge in operator load. Against this backdrop, the co-construction and sharing of communication base stations has become an industry consensus, thereby saving social resources and accelerating network coverage through unified planning and construction.

[0003] In existing technologies, macro-historical data can be processed by human experience to determine static indicators, and the load energy efficiency of shared base stations can be predicted by statistical models or machine learning models based on historical data.

[0004] However, the sharing model of shared base stations complicates the structure of revenue and load. Traditional static financial assessment methods or calculations using a single model make it difficult to capture the dynamic fluctuations of multi-operator business flows and energy consumption loads, and also make it impossible to accurately predict the long-term impact of technology upgrades or market strategy adjustments. Summary of the Invention

[0005] This application provides a method, apparatus, medium, and program product for predicting the load energy efficiency of a shared base station, which can improve the accuracy of the predicted load energy efficiency indicators of a shared base station.

[0006] In a first aspect, embodiments of this application provide a method for predicting the load energy efficiency of a shared base station, the method comprising: Obtain the multi-dimensional feature vector sequence corresponding to the preset historical time period of the shared base station. The feature dimensions of the multi-dimensional feature vector sequence include time dimension, performance dimension, resource dimension and load dimension. By processing multi-dimensional feature vector sequences through a long short-term memory network, the long-term features of the shared base station are obtained. By processing long-term features through a time attention mechanism, time-enhanced features are obtained. Long-term features are processed through a feature attention mechanism to obtain dimension-enhanced features; The temporal enhancement features and dimensionality enhancement features are weighted and fused to obtain the initial fused features; Residual connections are performed on the initial fused features and long-term features to obtain the fused features; The fusion features are mapped according to the preset mapping method of the preset load energy efficiency index to obtain the predicted load energy efficiency index for the preset future time point.

[0007] Secondly, embodiments of this application provide a shared base station load energy efficiency prediction device, the device comprising: The acquisition module is used to acquire a multi-dimensional feature vector sequence corresponding to a preset historical time period of the shared base station. The feature dimensions of the multi-dimensional feature vector sequence include time dimension, performance dimension, resource dimension and load dimension. The processing module is used to process multi-dimensional feature vector sequences through a long short-term memory network to obtain the long-term features of the shared base station; The temporal enhancement module is used to process long-term features through a temporal attention mechanism to obtain temporally enhanced features; The dimension enhancement module is used to process long-term features through a feature attention mechanism to obtain dimension-enhanced features. The fusion module is used to perform weighted fusion of temporal enhancement features and dimensionality enhancement features to obtain initial fused features; The residual module is used to perform residual connections on the initial fused features and long-term features to obtain fused features; The mapping module is used to map the fused features according to the preset mapping method of the preset load energy efficiency index to obtain the predicted load energy efficiency index at the preset future time point.

[0008] Thirdly, embodiments of this application provide a terminal device, the device including: a processor, and a memory storing computer program instructions; the processor reads and executes the computer program instructions to implement the shared base station load energy efficiency prediction method as described in the first aspect.

[0009] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the shared base station load energy efficiency prediction method as described in the first aspect.

[0010] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed, implements the shared base station load energy efficiency prediction method as described in the first aspect.

[0011] The shared base station load energy efficiency prediction method, device, medium, and program product of this application embodiment obtain a multi-dimensional feature vector sequence corresponding to a preset historical time period of the shared base station. The feature dimensions of the multi-dimensional feature vector sequence include time dimension, performance dimension, resource dimension, and load dimension. This enables the learning of multivariate nonlinear coupling relationships through multi-dimensional feature vectors, making the prediction closer to the real-world scenario. By processing the multi-dimensional feature vectors through a long short-term memory network, the long-term characteristics of the shared base station are obtained, which can accurately predict the long-term trend and periodic pattern of load energy efficiency. By processing the long-term characteristics through a time attention mechanism, time-enhanced features are obtained. By processing the long-term characteristics through a feature attention mechanism, dimension-enhanced features are obtained. This solves the technical problem of difficulty in simultaneously capturing short-term patterns such as fluctuations in the time dimension of base station load and long-term dependencies such as energy efficiency payback period. The time-enhanced features and dimension-enhanced features are weighted and fused to obtain initial fused features. The fused features are mapped according to a preset mapping method of preset load energy efficiency indicators to obtain the predicted load energy efficiency indicators for future time points. This solves the problem of difficulty in co-modeling short-term fluctuations and long-term trends in the load energy consumption prediction of shared base stations, adapts to changing scenarios, and improves the accuracy of predicted load energy efficiency indicators. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating a shared base station load energy efficiency prediction method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a shared base station load energy efficiency prediction device provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. Detailed Implementation

[0014] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0015] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0016] It should be noted that the acquisition, storage, use, and processing of data in this application embodiment all comply with the relevant provisions of laws and regulations.

[0017] It should be noted that in the embodiments of this application, certain software, components, models and other existing solutions in the industry may be mentioned. These should be regarded as exemplary and are only intended to illustrate the feasibility of implementing the technical solution of this application. However, it does not mean that the applicant has used or necessarily used the solution.

[0018] In existing technologies, because shared base stations carry equipment from multiple operators, their revenue sources and load composition are complex. This makes it difficult for financial evaluation methods and statistical models based on static slicing and simple allocation to dynamically capture the temporal fluctuations of multi-operator business flows, accurately measure energy consumption load, or predict the long-term impact of technological evolution or market strategy adjustments on load energy efficiency. The core issue is that load energy efficiency is regarded as a static indicator derived from a few macro-historical data through linear or simple nonlinear relationships, which requires a lot of human experience intervention.

[0019] However, existing technologies using single time-series data or simple models do not fully consider multi-dimensional nonlinear dynamic characteristics such as regional population density and seasonal call volume, and cannot characterize complex multivariate relationships, resulting in limited prediction accuracy. Traditional time-series algorithms and machine learning algorithms have limited memory capacity, making it difficult to capture long-term dependencies in load energy efficiency and accurately reflect the long-term trends and periodic patterns of long-period, non-stationary time-series data. The load data of shared base stations has many outliers and is non-stationary, making it easy to lead to fitting or underfitting when calculated using existing technologies. It is also highly sensitive to outliers, resulting in large prediction deviations and poor stability. Furthermore, relying on manual feature engineering is inefficient and highly subjective, and the algorithm cannot automatically extract deep time-series features during the process, making it difficult to adapt to changes in data distribution and new influencing factors.

[0020] To address the problems in the prior art, embodiments of this application provide a method, apparatus, medium, and program product for predicting the load energy efficiency of a shared base station.

[0021] The shared base station load energy efficiency prediction method provided in the embodiments of this application will be introduced first below.

[0022] Figure 1 A flowchart illustrating a shared base station load energy efficiency prediction method according to an embodiment of this application is shown. Figure 1 As shown, the method may include the following steps: S101: Obtain the multi-dimensional feature vector sequence corresponding to the preset historical time period of the shared base station.

[0023] The feature dimensions of the multi-dimensional feature vector sequence include time dimension, performance dimension, resource dimension and load dimension.

[0024] Specifically, multi-dimensional data corresponding to a preset historical period of the shared base station is obtained. The preset historical period can be the past 10 consecutive months to cover the regular operation cycle, seasonal business fluctuation cycle and statutory holiday business change cycle of the shared base station.

[0025] By using the base station's unique identifier and timestamp as the primary key, multi-dimensional data can be aligned and associated to obtain a multi-dimensional feature vector for each time step, thereby obtaining a multi-dimensional feature vector sequence corresponding to a preset historical time period.

[0026] The multi-dimensional feature vector sequence includes time, performance, resource, and load dimensions. The time dimension is used to characterize the time attributes of multi-dimensional data collection, including but not limited to sub-features such as timestamps, weekday attributes, holiday identifiers, and large-scale social event identifiers. The performance dimension is used to characterize the network operation performance status of the shared base station, including but not limited to sub-features such as base station access success rate, network transmission rate, service interruption duration, and 5G service carrying ratio. The resource dimension is used to characterize the infrastructure resource occupation and sharing status of the shared base station, including but not limited to sub-features such as tower reuse coefficient, equipment room space occupancy rate, number of transmission port reuses, and power resource sharing ratio. The load dimension is used to characterize the service carrying load level of the shared base station, including but not limited to sub-features such as peak concurrent users, average monthly data traffic, base station CPU utilization rate, and radio frequency unit transmit power ratio.

[0027] Furthermore, multi-dimensional data can be processed by minimax normalization to obtain multi-dimensional feature vectors with uniform scale features, thereby eliminating numerical biases caused by features of different dimensions in model processing.

[0028] The embodiments of this application can obtain multi-dimensional feature vector sequences, which can reduce prediction variance and improve the prediction accuracy of load energy efficiency indicators.

[0029] S102: The long-term features of the shared base station are obtained by processing the multi-dimensional feature vector sequence through a long short-term memory network.

[0030] Specifically, the multi-dimensional feature vector sequence is input into a long short-term memory network to process the multi-dimensional feature vector sequence and obtain the long-term features of the shared base station.

[0031] Among them, the Long Short-Term Memory network also includes a gating mechanism, that is, through the coordinated regulation of input gate, forget gate, output gate and cell state, selective memory and past are performed on time-series information, reducing the gradient vanishing and gradient explosion when traditional recurrent neural networks process long time-series data, continuously retaining long-term correlation information in time-series data, and obtaining the long-term characteristics of shared base stations.

[0032] The embodiments of this application process multi-dimensional feature vector sequences through a long short-term memory network to obtain long-term features, which can avoid the gradient vanishing problem and improve the accuracy of predicting load energy efficiency indicators.

[0033] S103: Long-term features are processed through a time attention mechanism to obtain time-enhanced features.

[0034] Specifically, long-term features are input into the temporal attention mechanism, which then traverses each time step of the long-term features and determines their corresponding temporal attention scores. The temporal attention scores are then normalized to obtain the temporal attention weights for each time step.

[0035] We perform weighted fusion of the long-term features of each time step and the corresponding temporal attention weights to obtain the temporal augmentation features in the time dimension.

[0036] Among them, the temporal attention mechanism is used to assign adaptive weights to time steps of temporal features, thereby strengthening the time step features.

[0037] In this embodiment of the application, since the performance data of the shared base station at different time steps have different degrees of influence on the predicted load energy efficiency index, the time attention mechanism is used to adaptively allocate weights, which reduces the noise interference of irrelevant time steps and amplifies the information of key time steps, thereby improving the accuracy of the predicted load energy efficiency index.

[0038] S104: Long-term features are processed through a feature attention mechanism to obtain dimension-enhanced features.

[0039] Specifically, long-term features are synchronously input into the feature attention mechanism so that the feature attention mechanism performs global average pooling of the long-term features in the time dimension. Then, the dimension attention weights of each feature dimension are calculated through a fully connected layer. The long-term features and dimension attention weights are then considered element by element to complete the feature dimension enhancement process and obtain dimension-enhanced features.

[0040] Among them, the feature attention mechanism is used to assign adaptive weights to feature dimensions, thereby strengthening the core feature dimension information.

[0041] The embodiments of this application adaptively allocate weights based on the different contributions of each feature dimension in the multi-dimensional feature vector sequence to resource utilization efficiency and energy consumption parameters, thereby suppressing redundant or interfering features, improving generalization ability and robustness against outliers, and enhancing the effectiveness of feature representation.

[0042] S105: Perform weighted fusion of temporal enhancement features and dimensionality enhancement features to obtain the initial fused features.

[0043] Specifically, learnable fusion parameters that can be autonomously optimized are obtained, and time-enhanced features and dimensionality-enhanced features are weighted and fused to obtain initial fusion features. The learnable fusion parameters are scalar parameters that can be adaptively optimized through the backpropagation algorithm and are used to match time-enhanced features and dimensionality-enhanced features that conform to the actual data of the shared base station.

[0044] The initial fusion features include key information from both the time dimension and the feature dimension, and are synergistically integrated through a weighted fusion process.

[0045] The embodiments of this application reduce the information redundancy caused by simple splicing or addition by weighted fusion of time-enhanced features and dimensional-enhanced features, and can improve the adaptability of the initial fusion features by performing weighted fusion on time-enhanced features and dimensional-enhanced features by ..., and can improve the adaptability of the initial fusion features by weighted fusion parameters according to the actual situation of the shared base station.

[0046] S106: Perform residual connections on the initial fused features and long-term features to obtain fused features.

[0047] Specifically, the initial fused features and the long-term features output by the long short-term memory network are residually connected to obtain fused features, reducing the loss of feature information that occurs during the processing of temporal attention and dimensional attention mechanisms.

[0048] Among them, residual connection represents a feature connection method that adds the initial fused features and long-term features element by element, while retaining the original feature information.

[0049] The formula for residual connection can be expressed as: , This represents the fusion feature, with a shape of (T, 32). Indicates long-term characteristics, This represents the initial fusion features. The two are added element by element, which can preserve the original information while enhancing gradient flow.

[0050] in, It needs to be expanded to be similar to H2 In the same dimension, the residual path can preserve the long-term trends extracted by the long short-term memory network (such as the increase in 5G penetration rate and the cumulative annual load energy efficiency), while the attention path provides short-term focus driven by events (such as holiday fluctuations). The combination of the two enables the model to optimize both long-term and short-term indicators at the same time.

[0051] The embodiments of this application compensate for the information lost in the initial fused features through residual connections. The residual structure can alleviate the gradient vanishing phenomenon, improve model stability, and enhance robustness.

[0052] S107: Map the fusion features according to the preset mapping method of the preset load energy efficiency index to obtain the predicted load energy efficiency index for the preset future time point.

[0053] Specifically, according to the mapping structure of the fully connected layer, a preset mapping method for the preset load energy efficiency index is obtained. The fusion features are then fused according to the preset mapping method to obtain the predicted load energy efficiency index for the preset future time point.

[0054] Among them, the preset future time point can be four consecutive natural months, and the predicted load energy efficiency index can quantitatively characterize the resource utilization efficiency or energy consumption parameters of the shared base station in the future period.

[0055] Furthermore, a target report can be generated based on the predicted load energy efficiency index in a preset report format, and the target report can be sent to the target client.

[0056] This application embodiment maps fused features into predicted load energy efficiency indicators through a preset mapping method, outputs quantified resource utilization efficiency and energy consumption parameters, and accurately predicts the future load energy efficiency of shared base stations.

[0057] This application embodiment captures the resource sharing and load coupling relationship unique to shared base stations through multi-dimensional feature vector sequences, solves the long-term dependency modeling problem through long short-term memory networks, achieves adaptive focusing through time attention mechanism and feature attention mechanism, and maps fused features according to a preset mapping method to obtain predicted load energy efficiency index, thereby enabling the prediction of load energy efficiency index in various scenarios and improving the accuracy of predicted load energy efficiency index.

[0058] In some embodiments, a long short-term memory network is used to process multi-dimensional feature vector sequences to obtain long-term features of the shared base station, including: By using the first layer of the Long Short-Term Memory network and the gating update mechanism in the Long Short-Term Memory network, the temporal features of the multi-dimensional feature vector sequence are extracted to obtain the short-term memory features of the shared base station. By using the second layer of the Long Short-Term Memory (LSTM) network and the gating update mechanism in the LSM network, the temporal dependency features of the short-term memory features are extracted to obtain the long-term features of the shared base station.

[0059] Specifically, the temporal features of the multi-dimensional feature vector sequence are extracted through the first layer of the long short-term memory network and the gating update mechanism to obtain short-term memory features. The first layer of the long short-term memory network has 64 units, which is used to quickly respond to monthly changes.

[0060] The gating update mechanism is formed by combining input gate, forget gate and output gate. It traverses the multi-dimensional feature vector sequence at each time step to complete the temporal correlation analysis, and filters and retains the feature content in the sequence that reflects short-cycle changes such as monthly business fluctuations, temporary equipment debugging and sudden traffic surges.

[0061] The input gate processing can be represented as follows: , Indicates the input gate output. This represents the input gate weight matrix. This indicates the hidden state at the previous time step. This represents the sequence of multi-dimensional feature vectors input at the current time step. This represents the bias term of the input gate. σ It can be the Sigmoid activation function.

[0062] The process of forgetting gating can be represented as follows: , Indicates the output of the forget gate. This represents the forget gate weight matrix. This represents the bias term for forgetting gating.

[0063] The process of output gate regulation of the information flow from the memory cell to the current output can be represented as: , Indicates the output of the output gate. This represents the output gate weight matrix. This represents the bias term used to adjust the information flow from the memory cell to the current output by the output gate.

[0064] The process of calculating candidate cell states can be represented as follows: , Indicates the state of candidate cells. Represents the weight matrix. represents the bias term, and tanh represents the hyperbolic tangent activation function.

[0065] The process of cell state renewal can be represented as follows: , This indicates the cell state at the current time step. The symbol represents the cell state at the previous time step, and ⊙ represents element-wise multiplication.

[0066] The process of handling hidden state output can be represented as follows: , The output represents the hidden state at the current time step.

[0067] In addition, by using the second layer of the long short-term memory network and the gating update mechanism in the long short-term memory network, the hidden state sequence (i.e. short-term memory feature) output by the first layer of the long short-term memory network is processed, and its temporal dependency feature is extracted to obtain the long-term feature of the shared base station.

[0068] The second layer of the long short-term memory network has 32 units, fewer than the first layer, in order to compress and abstract the short-term memory features of the input, discard short-term noise, and retain long-term dependencies, which can be cross-year time dependencies.

[0069] The gating update mechanism is formed by combining input gate, forget gate, and output gate. It traverses each time step in the short-term memory features to complete the temporal correlation analysis, deeply mines the cross-period correlation information in the short-term memory features, filters out short-period noise data, locks the long-term correlation patterns across months and quarters in the shared base station operation data, and extracts the temporal dependency features of the short-term memory features.

[0070] The input gate processing can be represented as follows: The superscript 2 indicates the second layer of the Long Short-Term Memory network. This represents the hidden state output of the first layer at the current time step, which serves as the input to the second layer. Indicates the input gate output. This represents the input gate weight matrix. This indicates the hidden state at the previous time step. This represents the sequence of multi-dimensional feature vectors input at the current time step. This represents the bias term of the input gate. σ It can be the Sigmoid activation function.

[0071] The process of using the forget gate can be represented as follows: , This represents the weight matrix of the second-layer forget gate. This represents the bias term for the forget gate.

[0072] The processing procedure of the output gate can be represented as follows: , This indicates the output of the output gate. This represents the output gate weight matrix of the second-layer Long Short-Term Memory network. This represents the bias term used to adjust the information flow from the memory cell to the current output by the output gate.

[0073] The process of processing candidate cell states can be represented as follows: , This is the candidate cell state weight matrix for the second-layer long short-term memory network. Indicates the state of candidate cells. The term represents the bias term for the candidate cell state, and tanh represents the hyperbolic tangent activation function.

[0074] The process of cell state renewal can be represented as follows: , This represents the cell state at the current time step in the second-layer long short-term memory network.

[0075] The process of handling hidden state output can be represented as follows: Output the hidden state at the current time step for the second-layer Long Short-Term Memory network.

[0076] This application embodiment uses a progressive extraction method with a dual-layer long short-term memory network and a gating update mechanism. First, short-cycle time-series features are extracted to generate short-term memory features. Then, long-cycle time-series dependent features are mined based on the short-term memory features to obtain long-term features. This can separate short-cycle fluctuations from long-cycle patterns, so that the long-term features accurately reflect the long-cycle energy efficiency attributes of the shared base station.

[0077] In some embodiments, long-term features are processed through a time attention mechanism to obtain time-enhanced features, including: Based on the learnable attention parameters and the long-term feature vectors corresponding to each time step in the long-term features, assign corresponding initial time attention weights to each time step. Normalize the initial temporal attention weights to obtain the temporal attention weights for each time step; We obtain time-enhanced features by weighting based on temporal attention weights and long-term features.

[0078] Specifically, the long-term features consist of long-term feature vectors arranged sequentially by time step. The long-term features are input into the time attention mechanism, which then calls the learnable attention parameters that are autonomously optimized during the model training phase. Based on the learnable attention parameters, the mechanism iterates through the long-term feature vectors corresponding to each time step within the long-term features and determines the corresponding initial time attention weights for each time step.

[0079] Alternatively, the initial temporal attention weights can be processed using the exponential normalization method, mapping the values ​​of the initial temporal attention weights at each time step to the 0-1 interval, and the sum of all mapped initial temporal attention weight values ​​is 1, thus obtaining the temporal attention weights corresponding to each time step.

[0080] Furthermore, the temporal attention weights are matched with the long-term feature vectors of the corresponding time steps in the long-term features. The long-term feature vectors are then weighted using the temporal attention weights as coefficients. The weighting process can be element-wise multiplication, thereby amplifying the feature representation of time steps with high temporal attention weights and compressing the feature proportion of time steps with low temporal attention weights, resulting in temporally enhanced features.

[0081] Furthermore, key time points and corresponding initial temporal attention weights are determined based on the patterns encoded by time steps in long-term features. The initial temporal attention weights can be expressed as: , Indicates time step t The initial attention weights, This represents the learnable attention parameters. Represents the weight matrix. Indicates the bias term. This indicates that the second layer of the Long Short-Term Memory network is in time step t The hidden state output.

[0082] ; in, This represents the time attention weights obtained after normalization at time step t, where T represents the preset historical period, exp represents the exponential function, and the sum of the weights for all time steps is 1.

[0083] ; The time-enhanced feature is represented by a shape (T, 32), where the hidden state at each time step is multiplied by its corresponding time attention weight.

[0084] In this embodiment, the initial temporal attention weight is calculated using learnable attention parameters. After normalization, the temporal attention weight is obtained and then weighted for long-term features. Based on the time step information in the long-term features, redundant noise in the temporal dimension is filtered out, so that the time-enhanced features carry the time information that affects the load energy efficiency of the shared base station.

[0085] In some embodiments, long-term features are processed through a feature attention mechanism to obtain dimension-enhanced features, including: The pooling vector is obtained by performing global average pooling on the time step set based on long-term features; By processing the pooling vector through a fully connected network and activation function, the dimensional attention weights of each feature dimension corresponding to the long-term features are obtained. The feature vectors and dimensional attention weights of each time step in the long-term features are multiplied according to the feature dimension to obtain the dimension-enhanced features.

[0086] Specifically, for long-term features, global average pooling is performed on the set of time steps consisting of all time steps. The average value of the corresponding dimension of all time steps is calculated for each feature dimension, and the feature information of the time series dimension is compressed into a single vector, namely the pooling vector. The pooling vector aggregates the global statistical features of long-term features in the time dimension, which can filter out local fluctuation data of a single time step.

[0087] In addition, the pooling vector is input into the fully connected network so that the fully connected network performs a linear transformation on the pooling vector. The transformed pooling features are then non-linearly mapped through the activation function to obtain the dimensional attention weights of each feature dimension corresponding to the long-term features. The activation function constrains the values ​​of the transformed pooling features within a preset range, and the dimensional attention weights are used to reflect the degree of influence of each feature dimension on the shared base station load energy efficiency index.

[0088] Furthermore, according to the correspondence of feature dimensions, the feature vector of each time step in the long-term features is multiplied element-wise with the dimensional attention weight along the feature dimension to obtain the dimension-enhanced features. The dimensional attention weight can be adapted to the feature vector of all time steps in a broadcast manner, which improves the feature expression intensity of the feature dimension of the high-dimensional attention weight and reduces the information proportion of the feature dimension of the low-dimensional attention weight.

[0089] Furthermore, the process of determining the pooling vector can be expressed as: ; in, This represents the pooling vector in the time dimension, with a shape of (32,), used to generate the query vector for feature attention.

[0090] The process of determining dimensional attention weights can be expressed as: , This represents the attention weights for the dimension, with a shape of (32,). , This represents the weight matrix of the fully connected layer. , σ represents the bias term, ReLU represents the linear rectified activation function, and σ represents the sigmoid activation function, which is used to restrict the dimensional attention weights to between 0 and 1.

[0091] The process of dimensionality enhancement features can be represented as follows: , The output of the feature attention is represented by the shape (T, 32), where ⊙ represents element-wise multiplication. This is a broadcast mechanism, meaning that β is the same for each time step.

[0092] In this embodiment, a pooling vector is obtained through global average pooling. Then, a dimension attention weight is generated by a fully connected network and an activation function. The long-term features are multiplied dimension by dimension to filter and strengthen the feature dimensions related to load energy efficiency in the long-term features and remove redundant information in the feature dimensions.

[0093] In some embodiments, the temporal enhancement features and dimensionality enhancement features are weighted and fused to obtain initial fused features, including: The temporal enhancement features are subjected to global average pooling in the temporal dimension to obtain the first temporal pooling vector. The dimension-enhanced features are subjected to global flat pooling in the time dimension to obtain the second time pooling vector. The first time-pooling vector and the second time-pooling vector are concatenated to obtain the concatenated vector; The concatenated vectors are processed using a fully connected network and activation functions to obtain the fusion parameters; Based on the fusion parameters, determine the temporal fusion weights corresponding to the temporal enhancement features and the dimensional fusion weights corresponding to the dimensional enhancement features; The time fusion weight, time enhancement feature, dimension fusion weight, and dimension enhancement feature are weighted and fused to obtain the initial fused feature.

[0094] Specifically, the time-enhanced features are subjected to global average pooling across the time dimension formed by all their time steps. The feature mean of each time step is calculated for each feature dimension, and the feature information of the time sequence dimension is compressed into a first time pooling vector of a preset dimension.

[0095] In addition, global average pooling is performed on the dimension-enhancing features in the time dimension to obtain the second time pooling vector.

[0096] Furthermore, the first and second time pooling vectors are concatenated according to the order of their vector dimensions to obtain a continuous single vector, namely the concatenated vector. The concatenated vector can represent the global pooling information of the time-enhanced features and the dimension-enhanced features in the time dimension.

[0097] Furthermore, the concatenated vector is input into a fully connected network, which performs a linear transformation on the concatenated vector. The transformed data undergoes a non-linear mapping via an activation function, and the scalar value output by the mapping is the fusion parameter. The fusion parameter enables the model to be autonomously optimized and updated during the model training phase through the backpropagation algorithm.

[0098] In addition, based on the fusion parameters, the temporal fusion weight corresponding to the temporal enhancement feature and the dimensional fusion weight corresponding to the dimensional enhancement feature are determined. The sum of the temporal fusion weight and the dimensional fusion weight is 1. The temporal fusion weight and the temporal enhancement feature are weighted element by element, and the dimensional fusion weight and the dimensional enhancement feature are weighted element by element. The two sets of weighted results are added element by element to obtain the initial fusion feature.

[0099] The calculation process of the initial fusion features can be expressed as follows: ,in, Let represent the initial fusion features, with shape (T, 32), and λ represent the learnable fusion parameters (scalar, ranging from 0 to 1).

[0100] The process of determining λ can be expressed as: Where λ is the adaptive generation, MeanPool represents global average pooling in the time dimension, resulting in a vector of shape (32,), and [;] represents vector concatenation. Represents the weight matrix. As a bias term, σ restricts λ to the range of 0 to 1.

[0101] Learnable fusion parameters During training, automatic optimization determines whether the time dimension or the feature dimension is more important in the current scene. As a result, the key time steps identified by the time attention are amplified and then enter the pooling calculation of the feature attention, making the feature attention pay more attention to the feature dimensions corresponding to these time steps. During backpropagation, the feature attention also affects the gradient of the time attention, forming a two-way information flow.

[0102] This application embodiment processes temporal enhancement features and dimensional enhancement features separately, divides corresponding fusion weights according to fusion parameters and performs weighted fusion, which can adaptively balance the information ratio of temporal enhancement features and dimensional enhancement features, generate initial fusion features that integrate the two types of attention enhancement information, provide suitable intermediate feature data for residual connections, and improve the accuracy of predicting load energy efficiency indicators.

[0103] In some embodiments, the fused features are mapped according to a preset mapping method for preset load energy efficiency indicators to obtain predicted load energy efficiency indicators for preset future time points, including: The fused features are nonlinearly transformed by a fully connected layer to obtain intermediate mapped features; According to the preset mapping method of the preset load energy efficiency index, the intermediate mapping features are mapped into multi-dimensional prediction vectors to obtain the predicted load energy efficiency index at the preset future time point. The dimension of the prediction vector is the same as the number of preset load energy efficiency indices. The calculation process for the preset mapping method is expressed as follows: ; in, Represents a multi-dimensional prediction vector. Represents the adaptive weight matrix. Indicates intermediate mapping features, This represents the bias vector.

[0104] Specifically, after obtaining the fusion characteristics The fused feature vector at each time step A nonlinear transformation is performed, that is, the input is fed into the first fully connected layer, which maps the 32 dimensions to 16 dimensions, and the intermediate mapped features are output through the ReLU activation function. The process can be represented as follows: ,in, Let represent a weight matrix of shape (32, 16). The 16-dimensional bias is used to enhance the abstract representation of features, compressing the high-dimensional information extracted by Long Short-Term Memory Networks, Temporal Attention Mechanisms, and Dimensional Attention Mechanisms into a more compact representation space.

[0105] In addition, according to the preset mapping method of the preset load energy efficiency index, the 16-dimensional intermediate mapping features are... Mapped to multidimensional prediction vector The mapping process can be completed through a second fully connected layer, and the process can be represented as follows: ,in This represents the adaptive weight matrix, with the following shape: , This indicates the number of preset load energy efficiency indicators. express dimensional bias vector, This represents the predicted load energy efficiency index at the t-th preset future time point.

[0106] in, and It automatically updates through backpropagation during the training process, thereby learning the optimal mapping relationship from intermediate features to various energy efficiency indicators.

[0107] For example, The value can be 4, corresponding to the monthly total energy efficiency, monthly load saving rate, network resource utilization trend, and load recovery period of the shared base station. The monthly total energy efficiency represents the total energy efficiency generated by the shared base station in that month; the load recovery period represents the remaining number of months required for the cumulative energy efficiency to cover the initial load up to that month; the network resource utilization rate represents the average utilization rate of the shared base station resources in that month; and the monthly load saving rate can be calculated using the following formula: ( (Actual load in the current month / self-built load of a single operator) × 100% (Single operator's self-built load / actual load for the month) × 100%.

[0108] Furthermore, based on the shared base station implementation logic, service constraint regular expressions can be embedded, the expression of which is: ; in, Indicates business constraint loss. t Iterate through the four time steps of the predicted load energy efficiency index. k Traverse the 4 business constraints, Indicates the first k The function of the degree of violation of the constraints. Ensure that only positive values ​​(i.e., violations of constraints) are penalized.

[0109] By using business constraint regularization terms, the predicted load energy efficiency indicators are made to meet the business logic 100%, reducing the occurrence of predicted load energy efficiency indicators that violate common sense.

[0110] Furthermore, scene feature vectors can be extracted from multi-dimensional feature vector sequences, and the relevant weights can be dynamically adjusted according to the predicted scene.

[0111] The scene feature vector can be determined based on the average value of multi-dimensional feature data over time, using the following formula: Where s represents the extracted scene features (25 dimensions). Indicates time step t The multi-dimensional feature vector sequence (25 dimensions) is used to determine the average value within a preset historical period.

[0112] S includes scenario information such as holiday intensity (i.e., whether it is a holiday average), call load (i.e., peak concurrent user average), business evolution (i.e., 5G penetration average), sharing degree (i.e., spectrum sharing average), and load level (electricity cost average).

[0113] A small fully connected neural network is trained for each predicted load energy efficiency index, and an importance score is output according to scenario s, which can be expressed as: ,m=1,2,3,4 ,in, This represents the unnormalized importance score of the m-th predicted load energy efficiency index in scenario s. This represents a small fully connected neural network (the architecture and the main model are trained together).

[0114] The process of adjusting the relevant weights can be represented as: ,m=1,2,3,4 ; in, This represents the normalized weight of the m-th predicted load energy efficiency index in scenario s, which can be calculated using the Softmax function. The sum of the weights of all indices is 1.

[0115] The above-mentioned small fully connected neural network automatically learns the following: When the intensity of holidays is high in s Improve (energy efficiency fluctuates greatly during holidays, overall energy efficiency is important); When 5G penetration is high in s Increase (new technology load, payback period is important); When the call load in S suddenly increases Improve (sudden surges in traffic can cause load spikes, making cost savings important); When the base station load rate in s remains high Improvement is needed (capacity expansion is required; utilization rate is important).

[0116] In this embodiment, the fused features are nonlinearly transformed by a fully connected layer to obtain intermediate mapping features. Then, the intermediate mapping features are mapped into a multi-dimensional prediction vector with the same number of indicators by a linear formula of the preset mapping method in Gnu. The fused features are converted into preset load energy efficiency indicators to obtain stable quantitative prediction results of shared base station load energy efficiency at preset future time points. Moreover, the preset load energy consumption indicators are closer to the actual situation.

[0117] In some embodiments, the training process of a long short-term memory network includes: Obtain the training sample vector sequence and the training real labels corresponding to the training sample data, wherein the length of the training sample vector sequence is the length of a preset historical period; The time average of the training sample vector sequence in the time dimension is used to determine the scene feature vector; The scene feature vector is mapped to the training dynamic weights corresponding to the predicted load energy efficiency index, and the training prediction index is determined based on the training dynamic weights. Determine the mean squared error of training between each training prediction metric and the corresponding training true label, and obtain the total prediction loss. Based on the total prediction loss, the parameters of the Long Short-Term Memory (LSTM) network, the temporal attention mechanism, and the feature attention mechanism are updated through backpropagation until the training completion condition is met, thus obtaining the LSM network.

[0118] Specifically, training sample vector sequences and corresponding training real labels are extracted from the historical data of shared base stations. The length of the training sample vector sequence is consistent with the length of the preset historical time period, the feature dimension of the training sample vector sequence is fully aligned with the input feature dimension of the actual prediction stage, and the training real label is the predicted load energy efficiency index at the preset future time point.

[0119] In addition, the mean of the training sample vector sequence is calculated dimension by dimension in the time dimension, and the time-extended vector data is compressed into a single vector of a preset dimension, which is the scene feature vector. The scene feature vector is used to represent the global statistical attributes of the actual scene of the base station corresponding to the training sample.

[0120] In addition, the scene feature vector is converted into training dynamic weights that correspond one-to-one with the predicted load energy efficiency index. The original prediction results are then weighted and adjusted according to the training dynamic weights to obtain training prediction indexes that are suitable for the current scene.

[0121] In addition, the training prediction index and its corresponding training true label are traversed, and the training mean square error between the two is calculated for each index. The training mean square errors of each index are summed to obtain the total prediction loss of the model in the current iteration round.

[0122] Furthermore, using the total prediction loss as the basis for gradient calculation, the gradient value is derived in reverse along the forward propagation path, and all learnable parameters within the Long Short-Term Memory network, the temporal attention mechanism, and the feature attention mechanism are updated synchronously through the backpropagation algorithm.

[0123] Repeat the above iterative process until the model converges or the total prediction loss drops to a preset threshold, which satisfies the training completion condition and yields the Long Short-Term Memory network.

[0124] The total prediction loss can be determined based on the total loss function, which includes the mean squared error. The process of determining the mean squared error can be expressed as follows: ,m=1,2,3,4 ; in, This represents the mean square error (MSE) of the m-th load energy efficiency index. This represents the training true label for the m-th load energy efficiency metric at time step t. This represents the corresponding training prediction metric, and the length of the summation window is the length of a preset future time period.

[0125] The total loss function can be expressed as: ,in, This represents the weighted total predicted loss. This indicates scene-adaptive dynamic weights. This represents the predicted loss for each load energy efficiency index.

[0126] The objective function can be expressed as: Where L represents the objective function, Indicates the weighted predicted loss. This represents the regularization term for business constraints. The coefficient 0.1 is a hyperparameter that controls the penalty strength of the constraint.

[0127] This application embodiment extracts scene feature vectors and maps them to dynamic weights, enabling the model to allocate different optimization priorities to different types of input samples, thereby improving the prediction accuracy in different scenarios. By extracting scene features through time averaging, the generalization ability is improved.

[0128] Figure 2 This is a schematic diagram of a shared base station load energy efficiency prediction device provided in an embodiment of this application. Figure 2 As shown, the shared base station load energy efficiency prediction device may include an acquisition module 210, a processing module 220, a time enhancement module 230, a dimension enhancement module 240, a fusion module 250, a residual module 260, and a mapping module 270.

[0129] The acquisition module 210 is used to acquire a multi-dimensional feature vector sequence corresponding to a preset historical time period of the shared base station, wherein the feature dimensions of the multi-dimensional feature vector sequence include time dimension, performance dimension, resource dimension and load dimension. The processing module 220 is used to process multi-dimensional feature vector sequences through a long short-term memory network to obtain the long-term features of the shared base station; The temporal enhancement module 230 is used to process long-term features through a temporal attention mechanism to obtain temporally enhanced features; The dimension enhancement module 240 is used to process long-term features through a feature attention mechanism to obtain dimension-enhanced features. The fusion module 250 is used to perform weighted fusion of temporal enhancement features and dimensionality enhancement features to obtain initial fused features; Residual module 260 is used to perform residual connection on the initial fused features and long-term features to obtain fused features; The mapping module 270 is used to map the fusion features according to the preset mapping method of the preset load energy efficiency index to obtain the predicted load energy efficiency index at the preset future time point.

[0130] In some embodiments, the processing module 220 processes the multi-dimensional feature vector sequence through a long short-term memory network to obtain the long-term features of the shared base station, for use in: By using the first layer of the Long Short-Term Memory network and the gating update mechanism in the Long Short-Term Memory network, the temporal features of the multi-dimensional feature vector sequence are extracted to obtain the short-term memory features of the shared base station. By using the second layer of the Long Short-Term Memory (LSTM) network and the gating update mechanism in the LSM network, the temporal dependency features of the short-term memory features are extracted to obtain the long-term features of the shared base station.

[0131] In some embodiments, the time enhancement module 230 processes long-term features through a time attention mechanism to obtain time-enhanced features, which are used for: Based on the learnable attention parameters and the long-term feature vectors corresponding to each time step in the long-term features, assign corresponding initial time attention weights to each time step. Normalize the initial temporal attention weights to obtain the temporal attention weights for each time step; We obtain time-enhanced features by weighting based on temporal attention weights and long-term features.

[0132] In some embodiments, the dimensionality enhancement module 240 processes long-term features through a feature attention mechanism to obtain dimensionality-enhanced features, which are used for: The pooling vector is obtained by performing global average pooling on the time step set based on long-term features; By processing the pooling vector through a fully connected network and activation function, the dimensional attention weights of each feature dimension corresponding to the long-term features are obtained. The feature vectors and dimensional attention weights of each time step in the long-term features are multiplied according to the feature dimension to obtain the dimension-enhanced features.

[0133] In some embodiments, the fusion module 250 performs weighted fusion of the temporal enhancement features and the dimensionality enhancement features to obtain initial fused features, which are used for: The temporal enhancement features are subjected to global average pooling in the temporal dimension to obtain the first temporal pooling vector. The dimension-enhanced features are subjected to global flat pooling in the time dimension to obtain the second time pooling vector. The first time-pooling vector and the second time-pooling vector are concatenated to obtain the concatenated vector; The concatenated vectors are processed using a fully connected network and activation functions to obtain the fusion parameters; Based on the fusion parameters, determine the temporal fusion weights corresponding to the temporal enhancement features and the dimensional fusion weights corresponding to the dimensional enhancement features; The time fusion weight, time enhancement feature, dimension fusion weight, and dimension enhancement feature are weighted and fused to obtain the initial fused feature.

[0134] In some embodiments, the mapping module 270 maps the fused features according to a preset mapping method for preset load energy efficiency indicators to obtain a predicted load energy efficiency indicator for a preset future time point, used for: The fused features are nonlinearly transformed by a fully connected layer to obtain intermediate mapped features; According to the preset mapping method of the preset load energy efficiency index, the intermediate mapping features are mapped into multi-dimensional prediction vectors to obtain the predicted load energy efficiency index at the preset future time point. The dimension of the prediction vector is the same as the number of preset load energy efficiency indices. The calculation process for the preset mapping method is expressed as follows: ; in, Represents a multi-dimensional prediction vector. Represents the adaptive weight matrix. Indicates intermediate mapping features, This represents the bias vector.

[0135] In some embodiments, the processing module 220 is used during the training process of the Long Short-Term Memory network to: Obtain the training sample vector sequence and the training real labels corresponding to the training sample data, wherein the length of the training sample vector sequence is the length of a preset historical period; The time average of the training sample vector sequence in the time dimension is used to determine the scene feature vector; The scene feature vector is mapped to the training dynamic weights corresponding to the predicted load energy efficiency index, and the training prediction index is determined based on the training dynamic weights. Determine the mean squared error of training between each training prediction metric and the corresponding training true label, and obtain the total prediction loss. Based on the total prediction loss, the parameters of the Long Short-Term Memory (LSTM) network, the temporal attention mechanism, and the feature attention mechanism are updated through backpropagation until the training completion condition is met, thus obtaining the LSM network.

[0136] Figure 3 A schematic diagram of the hardware structure of the terminal device provided in an embodiment of this application is shown.

[0137] The terminal device may include a processor 301 and a memory 302 storing computer program instructions.

[0138] Specifically, the processor 301 may include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0139] Memory 302 may include mass storage for data or instructions. For example, and not limitingly, memory 302 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. In one instance, memory 302 may include removable or non-removable (or fixed) media, or memory 302 may be non-volatile solid-state memory. Memory 302 may be internal or external to the integrated gateway disaster recovery device.

[0140] In one instance, memory 302 may be read-only memory (ROM). In one instance, the ROM may be a mask-programmed ROM, a programmable ROM (PROM), an erasable PROM (EPROM), an electrically erasable PROM (EEPROM), an electrically rewritable ROM (EAROM), or flash memory, or a combination of two or more of these.

[0141] Memory 302 may include read-only memory (ROM), random access memory (RAM), disk storage media device, optical storage media device, flash memory device, electrical, optical, or other physical / tangible memory storage device. Therefore, generally, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method according to the first aspect of this disclosure.

[0142] The processor 301 reads and executes computer program instructions stored in the memory 302 to achieve... Figure 1 The shared base station load energy efficiency prediction method in the illustrated embodiment.

[0143] In one example, the terminal device may also include a communication interface 303 and a bus 304. Wherein, for example... Figure 3 As shown, the processor 301, memory 302, and communication interface 303 are connected through bus 304 and complete communication with each other.

[0144] The communication interface 303 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0145] Bus 304 includes hardware, software, or both, that couples components of an online data traffic metering device together. For example, and not as a limitation, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 304 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0146] Furthermore, in conjunction with the shared base station load energy efficiency prediction method in the above embodiments, this application embodiment can provide a computer-readable storage medium for implementation. The computer-readable storage medium stores computer program instructions; when executed by a processor, these computer program instructions implement any of the shared base station load energy efficiency prediction methods in the above embodiments.

[0147] This application also provides a computer program product, including a computer program that, when executed by a processor, implements any of the shared base station load energy efficiency prediction methods described in the above embodiments.

[0148] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0149] The functional blocks shown in the above-described block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, read-only memory (ROM), flash memory, erasable read-only memory (EROM), floppy disks, compact disc read-only memory (CD-ROM), optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0150] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0151] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0152] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A method for predicting the load energy efficiency of a shared base station, characterized in that, include: Obtain a multi-dimensional feature vector sequence corresponding to a preset historical time period of the shared base station, wherein the feature dimensions of the multi-dimensional feature vector sequence include time dimension, performance dimension, resource dimension and load dimension; The long-term features of the shared base station are obtained by processing the multi-dimensional feature vector sequence through a long short-term memory network. The long-term features are processed using a time attention mechanism to obtain time-enhanced features; The long-term features are processed using a feature attention mechanism to obtain dimension-enhanced features; The temporal enhancement features and the dimensionality enhancement features are weighted and fused to obtain the initial fused features; Residual connections are performed on the initial fused features and long-term features to obtain fused features; The fused features are mapped according to a preset mapping method for preset load energy efficiency indicators to obtain the predicted load energy efficiency indicators for preset future time points.

2. The method according to claim 1, characterized in that, The step of processing the multi-dimensional feature vector sequence through a long short-term memory network to obtain the long-term features of the shared base station includes: The temporal features of the multi-dimensional feature vector sequence are extracted through the first layer of the long short-term memory network and the gating update mechanism in the long short-term memory network to obtain the short-term memory features of the shared base station. By using the second layer of the Long Short-Term Memory (LSTM) network and the gating update mechanism, the temporal dependency features of the short-term memory features are extracted to obtain the long-term features of the shared base station.

3. The method according to claim 1, characterized in that, The process of processing the long-term features through a time attention mechanism to obtain time-enhanced features includes: Based on the learnable attention parameters and the long-term feature vectors corresponding to each time step in the long-term features, assign corresponding initial time attention weights to each time step. Normalize the initial temporal attention weights to obtain the temporal attention weights corresponding to each time step; The temporal enhancement features are obtained by weighting the temporal attention weights and the long-term features.

4. The method according to claim 1, characterized in that, The process of processing the long-term features through a feature attention mechanism to obtain dimension-enhanced features includes: A pooling vector is obtained by performing global average pooling on the time step set based on the long-term features. The pooling vector is processed by a fully connected network and an activation function to obtain the dimensional attention weights of each feature dimension corresponding to the long-term feature; The feature vectors of each time step in the long-term features and the dimensional attention weights are multiplied according to the feature dimensions to obtain the dimensional enhancement features.

5. The method according to claim 1, characterized in that, The weighted fusion of the temporal enhancement features and the dimensionality enhancement features to obtain the initial fused features includes: The time-enhanced features are subjected to global average pooling in the time dimension to obtain a first time pooling vector. The dimensional enhancement features are subjected to global flat pooling in the time dimension to obtain a second time pooling vector. The first time pooling vector and the second time pooling vector are concatenated to obtain the concatenated vector; The concatenated vector is processed using a fully connected network and an activation function to obtain the fusion parameters; Based on the fusion parameters, determine the temporal fusion weight corresponding to the temporal enhancement feature and the dimensional fusion weight corresponding to the dimensional enhancement feature; The time fusion weight, the time enhancement feature, the dimension fusion weight, and the dimension enhancement feature are weighted and fused to obtain the initial fusion feature.

6. The method according to claim 1, characterized in that, The step of mapping the fused features according to a preset mapping method for preset load energy efficiency indicators to obtain predicted load energy efficiency indicators for preset future time points includes: The fused features are nonlinearly transformed by a fully connected layer to obtain intermediate mapping features; According to the preset mapping method of the preset load energy efficiency index, the intermediate mapping features are mapped into multi-dimensional prediction vectors to obtain the predicted load energy efficiency index at a preset future time point, wherein the dimension of the prediction vector is the same as the number of preset load energy efficiency indices. The calculation process of the preset mapping method is expressed as follows: ； in, The predicted vector represents a multi-dimensional vector. Represents the adaptive weight matrix. This represents the intermediate mapping feature. This represents the bias vector.

7. The method according to claim 1, characterized in that, The training process of the Long Short-Term Memory network includes: Obtain the training sample vector sequence and the training real label corresponding to the training sample data, wherein the length of the training sample vector sequence is the length of the preset historical time period; The time average value of the training sample vector sequence in the time dimension is determined as the scene feature vector; The scene feature vector is mapped to the training dynamic weights corresponding to the predicted load energy efficiency index, and the training prediction index is determined based on the training dynamic weights. Determine the training mean square error between each of the training prediction metrics and the corresponding training real labels to obtain the total prediction loss; Based on the total prediction loss, the parameters of the Long Short-Term Memory Network, the Temporal Attention Mechanism, and the Feature Attention Mechanism are updated through backpropagation until the training completion condition is met, thus obtaining the Long Short-Term Memory Network.

8. A shared base station load energy efficiency prediction device, characterized in that, The device includes: The acquisition module is used to acquire a multi-dimensional feature vector sequence corresponding to a preset historical time period of the shared base station, wherein the feature dimensions of the multi-dimensional feature vector sequence include time dimension, performance dimension, resource dimension and load dimension. The processing module is used to process the multi-dimensional feature vector sequence through a long short-term memory network to obtain the long-term features of the shared base station; The time enhancement module is used to process the long-term features through a time attention mechanism to obtain time-enhanced features; The dimension enhancement module is used to process the long-term features through a feature attention mechanism to obtain dimension-enhanced features. The fusion module is used to perform weighted fusion of the temporal enhancement features and the dimensionality enhancement features to obtain initial fused features; The residual module is used to perform residual connections on the initial fused features and long-term features to obtain fused features; The mapping module is used to map the fused features according to a preset mapping method of preset load energy efficiency index to obtain the predicted load energy efficiency index at a preset future time point.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the shared base station load energy efficiency prediction method as described in any one of claims 1-7.

10. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the shared base station load energy efficiency prediction method as described in any one of claims 1-7.

Images