Performance index prediction methods, devices, electronic equipment and storage media

Abstract

This application relates to the field of artificial intelligence technology, and provides a performance indicator prediction method, apparatus, electronic device, and storage medium. The method includes: constructing a feature matrix with uniform granularity from multi-source network data; extracting features from the feature matrix to obtain time-series features, non-time-series features, and time-series correlated features, and integrating these features into a fusion feature; classifying the fusion feature at multiple scales to obtain classification data corresponding to different scales; and mapping and predicting the classification data to obtain performance indicators for a wireless load balancing scenario. This application provides a performance indicator prediction method for a wireless load balancing scenario, effectively solving the problem that base station configuration parameters and network performance indicators cannot be formalized in complex wireless network environments. It also overcomes the problems of long development cycles for new service scenarios caused by the time-consuming nature of traditional simulation platforms, lack of real-time guidance from the actual network, and limitations of expert experience.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to methods, apparatus, electronic devices and storage media for predicting performance indicators. Background Technology

[0002] Wireless load balancing is a crucial function for ensuring efficient utilization of network resources, used to avoid overload in a single cell and balance the load across cells. In complex wireless network environments, to guarantee effective load balancing, it is often necessary to predict network performance indicators under wireless load balancing scenarios. However, in existing networks, there are often issues with the relationship between the configuration parameters of various base station functions and network performance indicators, which cannot be formalized, increasing the difficulty of predicting network performance indicators.

[0003] To address the aforementioned issues that cannot be formalized, existing communication systems typically rely on traditional simulation platforms and expert guidance from network optimization personnel to predict performance metrics based on parameter modifications. For example, load balancing parameters are set using preliminary simulation results, and new design schemes are evaluated on traditional simulation platforms in conjunction with existing architectures. This attempts to predict the impact of parameter adjustments on network performance and guide network optimization operations.

[0004] However, traditional simulation platforms suffer from drawbacks such as long simulation times and detachment from real-world network conditions. Furthermore, new scenarios often require re-evaluation due to a lack of adjustable parameters, resulting in lengthy development cycles. Expert experience also has limitations and struggles to meet the requirements of complex real-world network scenarios. Therefore, to improve the quality and efficiency of network optimization decisions, providing a performance indicator prediction method for wireless load balancing scenarios has become a pressing issue in this field. Summary of the Invention

[0005] This application provides a performance indicator prediction method, apparatus, electronic device, and storage medium to address the technical problem of how to provide a performance indicator prediction method for wireless load balancing scenarios.

[0006] In a first aspect, embodiments of this application provide a performance index prediction method, including: Construct a feature matrix with uniform granularity from multi-source network data; Feature extraction is performed on the feature matrix to obtain time-series features, non-time-series features, and time-series related features, and the time-series features, non-time-series features, and time-series related features are integrated into a fusion feature; The fused features are classified at multiple scales to obtain classification data corresponding to different scales; The classification data is mapped and predicted to obtain performance indicators for the wireless load balancing scenario.

[0007] In one embodiment, the multi-source network data includes cell-level parameter data, performance index data, and parameter modification record data; Constructing a feature matrix with uniform granularity from multi-source network data includes: The cell-level parameter data of each cell are concatenated to obtain the first matrix; The first matrix is ​​copied and expanded according to the time dimension of the performance index data to obtain the second matrix; Based on the parameters, modify the recorded data to construct a sparse matrix; The values ​​of the same time and the same parameter in the sparse matrix are replaced with the corresponding positions in the second matrix to obtain the feature matrix with uniform granularity.

[0008] In one embodiment, feature extraction is performed on the feature matrix to obtain the time-series features, including: The feature matrix is ​​extracted using a long short-term memory network, and the output of the last time step is used as the temporal feature.

[0009] In one embodiment, feature extraction is performed on the feature matrix to obtain the non-temporal features, including: The configuration parameter features that affect inter-cell equilibrium are selected from the feature matrix as the non-time-series features.

[0010] In one embodiment, feature extraction is performed on the feature matrix to obtain the time-related features, including: The correlation coefficients of each index in the feature matrix at different time granularities are calculated by level transformation, and the correlation coefficients are mapped to the time-series correlation features.

[0011] In one embodiment, the correlation coefficients of each index in the feature matrix at different time granularities are calculated through a rank transformation, and the correlation coefficients are mapped to the time-series correlation features, including: The different indicators in the feature matrix are subjected to rank transformation based on the rank correlation coefficient method to obtain rank variables; Calculate the linear correlation coefficients between the different ordinal variables; The linear correlation coefficient at each time granularity is mapped to the sample points to obtain the time-series correlation features.

[0012] In one embodiment, the fused features are classified at multiple scales to obtain classification data corresponding to different scales, including: The fused features are input into the encoder, and the attention weights of the fused features are calculated through a self-attention mechanism. The value matrix is ​​then weighted and summed according to the attention weights to obtain the attention output features. The attention output features are sequentially processed by global average pooling and normalized exponential function to obtain the probability distribution of the fused features belonging to different preset categories; The fused features are divided according to the probability distribution to obtain the classification data corresponding to different scales.

[0013] In one embodiment, mapping and predicting the classified data to obtain performance metrics for a wireless load balancing scenario includes: The classification data at different scales are respectively input into the corresponding multi-layer fully connected neural network; Abstract features are extracted by nonlinearly mapping the classified data through the hidden layers in the multi-layer fully connected neural network. The performance index is obtained by performing linear numerical extrapolation on the abstract features through the output layer of the multi-layer fully connected neural network.

[0014] In one embodiment, nonlinear mapping of the classification data through hidden layers in the multilayer fully connected neural network to extract abstract features includes: Preliminary features are obtained by performing feature mapping on the classification data using a first-level neural network in the hidden layer based on a first non-linear activation function. The abstract features are obtained by further processing the preliminary features using a second-level neural network in the hidden layer based on a second nonlinear activation function.

[0015] In one embodiment, the multilayer fully connected neural network is trained through the following steps: Construct a long short-term memory network model for processing historical sample labels and an initial multilayer fully connected neural network model; The first loss value generated by the predicted value of the multilayer fully connected neural network model is calculated based on the historical sample labels, and the second loss value generated by the predicted value of the long short-term memory network model is calculated. The first loss value and the second loss value are weighted and summed to obtain the comprehensive loss function; The network parameters of the multilayer fully connected neural network model are updated based on the comprehensive loss function.

[0016] In one embodiment, updating the network parameters of the multilayer fully connected neural network model according to the comprehensive loss function includes: The parameter gradients of the multilayer fully connected neural network model are calculated based on the comprehensive loss function. The network parameters of the multilayer fully connected neural network model are updated backward using the parameter gradient by means of the adaptive learning rate optimization algorithm, with the goal of minimizing the mean absolute percentage error between the predicted value of the multilayer fully connected neural network model and the historical sample label.

[0017] Secondly, embodiments of this application provide a performance indicator prediction device for a wireless load balancing scenario, comprising: The matrix construction module is used to construct a feature matrix with uniform granularity from multi-source network data. The feature extraction module is used to extract features from the feature matrix to obtain time-series features, non-time-series features, and time-related features, and to integrate the time-series features, non-time-series features, and time-related features into a fusion feature. The feature classification module is used to perform multi-scale classification on the fused features to obtain classification data corresponding to different scales; The indicator prediction module is used to map and predict the classified data to obtain the performance indicators of the wireless load balancing scenario.

[0018] Thirdly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the performance index prediction method described in the first aspect.

[0019] Fourthly, embodiments of this application provide a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the performance index prediction method described in the first aspect.

[0020] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the performance indicator prediction method as described above.

[0021] This application provides a performance index prediction method for wireless load balancing scenarios. By using a data-driven feature extraction and classification mapping prediction method, it effectively solves the problem that the relationship between base station configuration parameters and network performance indicators cannot be formalized in complex wireless network environments. At the same time, through automated feature fusion and prediction, it overcomes the problems of long development cycles for new business scenarios caused by the long simulation time, lack of real-time guidance from the actual network, and limitations of expert experience in traditional simulation platforms. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in this 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 some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the performance index prediction method provided in the embodiments of this application.

[0024] Figure 2 This is a schematic diagram illustrating the principle of the performance index prediction method provided in the embodiments of this application.

[0025] Figure 3 This is a schematic diagram illustrating the test results of the performance index prediction method provided in the embodiments of this application. Figure 4 This is a schematic diagram of the configuration parameters provided in the embodiments of this application.

[0026] Figure 5 This is a schematic diagram of the performance index prediction device for a wireless load balancing scenario provided in this application embodiment.

[0027] Figure 6 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0029] Wireless load balancing is a key function for ensuring efficient use of network resources. It primarily prevents overload of individual cells while balancing the load across cells within the network to improve overall service quality. Load balancing mechanisms monitor cell resource usage, such as the number of users and bandwidth utilization, to determine when and how to reallocate User Equipment (UE) to less busy cells. This involves functions such as cell selection and reselection, and handover control, ensuring that UEs migrate to less loaded cells at the appropriate time, thereby improving network performance. 3GPP (3rd Generation Partnership Project) defines load balancing strategies for connected and idle states. In connected state, optimized handover decisions guide UEs to migrate to more suitable cells; while in idle state, cell reselection parameters are adjusted to guide UEs to select less loaded cells for access. Furthermore, technologies such as load sharing and interference coordination work together to achieve optimal allocation of network resources.

[0030] Related technology 1 provides a method for optimizing multi-standard intelligent configurable wireless networks. First, it identifies network optimization needs and objectives based on the user's own network. This is then analyzed to determine the network optimization standard and establish a simple model of the wireless network, preparing optimization schemes and configuring optimization parameters. Next, it optimizes the network from five different perspectives (antenna, power, site location, frequency, and load balancing) using different optimization algorithms and cost functions (such as capacity, coverage, and network quality). The optimization algorithms include three heuristic algorithms (simulated annealing, particle swarm optimization, and ant colony optimization) and a traditional greedy algorithm, improving network performance to an ideal level. Finally, it summarizes the wireless network optimization results and provides optimization schemes for user reference and practical guidance. This method is intelligently configurable, can meet reasonable user needs, and can also optimize the coexistence of 2G and 3G networks, offering high flexibility and providing a good reference for improving the performance of current networks.

[0031] Related technology two provides a method for optimizing wireless communication network parameters based on causal effect estimation using knowledge graphs. Starting from measurement data of wireless communication network parameters, it constructs a corresponding local knowledge graph based on the endogenous factors of the wireless communication network protocol. It separates the research object, intervention variables, covariates, and observation results from the measurement data, constructs randomized experiments using the observation data, and uses counterfactual reasoning to evaluate the impact of intervention variables on network performance indicators. Extending this method to the entire knowledge graph, it can obtain the impact of each adjustable parameter on each network performance indicator, thus providing directions for adjusting wireless communication network performance and offering optimization options based on the existing network. The system includes a knowledge graph construction module, an environment interaction module, and a causal reasoning module. This method, with limited computational resources and manual costs, assists in optimizing wireless communication network parameters starting from the global optimum of the wireless communication network.

[0032] Related technology 3 provides an artificial intelligence-based wireless communication network scheduling method and system, including: wireless communication network data collection, network scheduling, network parameter prediction, inter-regional collaborative scheduling, and performance monitoring. This method belongs to the field of network scheduling technology, specifically referring to an artificial intelligence-based wireless communication network scheduling method and system. It clarifies that computing resource scheduling is based on user preferences, with different preferences employing different scheduling strategies, and is carried out in two stages. A comprehensive score is calculated to reflect the importance of nodes in resource allocation and scheduling. A roulette wheel selection algorithm is used to select computing resource nodes to balance load and meet user preferences. A dynamic Bayesian network is constructed for parameter estimation. Historical data is used to estimate the dynamic Bayesian network parameters, predict future network performance trends, enhance foresight, improve network stability and reliability, and enhance the ability to respond to emergencies.

[0033] The aforementioned technologies have the following drawbacks: (1) In complex wireless network environments, there are problems that cannot be formalized among the configuration parameters of various functions of the base station and network performance indicators; (2) The adjustment range of parameters such as the upper limit of the number of connected load balancers and the offset of the number of uplink users in load balancing is limited. The ease of switching load balancers depends on the simulation results, algorithms and the experience of on-site network optimization personnel. There is a certain deviation from the actual application scenario. Furthermore, unreasonable parameter adjustments may cause network failures. Therefore, this method results in the adjustment results of various user numbers and offset thresholds failing to meet the requirements of complex actual scenarios. (3) In communication systems, performance indicators based on parameter modification rely on the guidance of experts such as simulation platforms and network optimization personnel. Traditional simulation platforms have the following drawbacks: long simulation time; implementation methods detached from the actual network; and inability to provide real-time guidance to on-site personnel for parameter modification prediction. Even with an existing simulation platform, when adding a simulation function for a new business scenario, it is often found that there are no relevant adjustable parameters (because the number of adjustable parameters in the actual network is huge, it is impossible to include them all in the design when building the simulation platform). At this time, it is necessary to re-evaluate the new design scheme in conjunction with the existing platform architecture, which results in a long development cycle.

[0034] The following is combined with Figures 1 to 6 This application describes a performance index prediction method and apparatus according to embodiments of the present application.

[0035] Figure 1 This is a flowchart illustrating the performance index prediction method provided in the embodiments of this application, such as... Figure 1 As shown, the method includes, but is not limited to, steps S1, S2, S3 and S4.

[0036] Step S1: Construct a feature matrix with uniform granularity from the multi-source network data.

[0037] Multi-source network data refers to a collection of data from wireless network communication systems that has different sources, dimensions, and time spans, combined with... Figure 2 Specifically, this can include base station cell-level engineering parameters (engineering parameters), measurement report (MR) data, default parameter configuration data, minute-level parameter modification records, and performance measurement (PM) data with a minimum granularity of 15 minutes. Among these, 3GPP defines load balancing strategies for connected and idle states. These data reflect the network's resource usage in different states, such as the number of users and bandwidth utilization.

[0038] Unified granularity means aligning the data with inconsistent time or spatial dimensions to the same time resolution (e.g., unifying to 15-minute or hourly granularity) and spatial resolution (e.g., unifying to base station cell level) on the time and spatial axes.

[0039] A feature matrix with uniform granularity can be constructed through operations such as concatenation, forward padding, interpolation completion, downsampling aggregation, or feature duplication. For example, the collected minute-level parameter modification records can be converted into a format aligned with 15-minute performance metrics through forward padding or aggregation, and then combined into a structured feature matrix.

[0040] Step S1 can eliminate the data heterogeneity caused by the complex sources and different collection frequencies of various function configuration parameters and performance indicators in the wireless network environment, and transform the originally unprocessable scattered data into a standardized format that is readable and usable by the algorithm.

[0041] Step S2: Extract features from the feature matrix to obtain time-series features, non-time-series features, and time-series related features, and integrate these features into a fusion feature.

[0042] Time-series characteristics are the dynamic trends or patterns of data over time, such as the periodic fluctuations in traffic. Non-time-series characteristics are static configurations or fixed threshold attributes unrelated to the passage of time, such as user number thresholds. Time-related characteristics are the degree of interdependence or correlation between different data variables over time, such as the positive correlation between business volume and user number.

[0043] Temporal features can be extracted using Long Short-Term Memory (LSTM) networks, non-temporal features can be extracted using feature filtering, and temporally relevant features can be extracted using correlation coefficient calculation. These three features can then be fused using methods such as column-wise vector concatenation, feature map overlay, multimodal feature cross-interaction, or matrix addition.

[0044] Step S2 can extract key elements affecting network load balancing from complex multi-dimensional data and form a fusion representation that includes a global perspective of time, space and correlation, laying the foundation for subsequent accurate classification and prediction.

[0045] Step S3: Perform multi-scale classification on the fused features to obtain classification data corresponding to different scales.

[0046] Multi-scale classification categorizes input data based on multiple different dimensions or levels (i.e., scales, such as the size of a cell's users, the time period's busy / off-peak status, signal coverage strength, etc.). Classified data consists of subsets of data that have been categorized and assigned to specific category labels.

[0047] A classification model based on self-attention mechanism, such as the Transformer multi-scale classifier, can be introduced. By extracting contextual information from the fused features through forward computation, samples with similar network behavior and load status can be classified into the same category (e.g., into 4 different clusters) through clustering algorithms (such as K-means), threshold partitioning (such as segmentation by idle and busy periods), or Softmax probability allocation to obtain classification data corresponding to different scales.

[0048] For example, the data for the whole day can be divided into four subsets of different scales, such as "high traffic during the morning peak" and "low traffic during the night off-peak".

[0049] Step S3 can refine the complex large-scale scene of wireless networks into multiple more targeted sub-scenes, enabling subsequent models to perform differentiated processing on features at different scales, thereby improving the model's generalization ability when facing diverse and complex real-world scenarios.

[0050] Step S4: Map and predict the classified data to obtain the performance indicators of the wireless load balancing scenario.

[0051] Mapping prediction establishes a mathematical relationship between input data features and output target variables. Performance metrics are specific quantitative values ​​used to measure the operational quality and load balancing effectiveness of wireless networks.

[0052] Combination Figure 2 The forward propagation structure of the network model on the right and Figure 3 Performance metrics can specifically include rate metrics, traffic volume metrics, load metrics, and quality metrics, such as: downlink average data rate (AvgDataRateDl), uplink average data rate (AvgDataRateUl), downlink PDCP (Packet Data Convergence Protocol) layer transmission traffic (pdcp_upoctdl), uplink PDCP layer transmission traffic (pdcp_upoctul), downlink PRB (Physical Resource Block) utilization (PrbAssnRateDl), and uplink physical resource block utilization (PrbAssnRateUl), etc.

[0053] Pre-trained deep learning regression models, such as Multilayer Perceptrons (MLPs) or Residual Networks, can be used to perform nonlinear calculations and forward propagation on the classification data for each category, including matrix multiplication and activation function activation, to directly predict future performance metrics. For example, inputting the classification data for "high traffic volume during morning rush hour" in a residential community into the corresponding MLP model can output the predicted PRB utilization rate for the community over the next 15 minutes.

[0054] Step S4 can realize the transformation from abstract features to specific business guidance values, providing intuitive data support for network optimization.

[0055] As can be seen from the above, the embodiments of this application provide a performance index prediction method for wireless load balancing scenarios. By using a data-driven feature extraction and classification mapping prediction method, it effectively solves the problem that the base station configuration parameters and network performance indicators cannot be formalized in complex wireless network environments. At the same time, through automated feature fusion and prediction, it overcomes the problems of long development cycles for new business scenarios caused by the long time required by traditional simulation platforms, the lack of real-time guidance from the actual network, and the limitations of expert experience.

[0056] In one embodiment, multi-source network data may include cell-level parameter data, performance index data, and parameter modification record data; Step S1 may further include: The cell-level parameter data of each cell are concatenated to obtain the first matrix; The first matrix is ​​copied and expanded according to the time dimension of the performance index data to obtain the second matrix; Construct a sparse matrix by modifying the recorded data based on the parameters; The values ​​of the same parameters at the same time in the sparse matrix are replaced with the corresponding positions in the second matrix to obtain a feature matrix with uniform granularity.

[0057] Cell-level parameter data are statistically analyzed at the base station cell level, and may include engineering parameters (such as antenna height and azimuth), MR data (measurement report data, such as reference signal received power), and default parameter configuration data. Performance metrics data are PM data (Performance Measurement), such as throughput and packet loss rate. Parameter modification records are historical operation records of network optimization personnel adjusting parameters in the live network, typically at the minute-level granularity.

[0058] The cell-level parameter data of each cell are concatenated to obtain the first matrix. Engineering parameters, MR data, and default parameter configurations can be aligned and combined at the cell level according to column dimensions through vector or matrix concatenation and horizontal merging. Assuming the number of cells is m1 and the selected feature dimension is n1 (e.g., combining features such as antenna height and azimuth), the first matrix with dimension m1×n1 can be directly constructed by concatenation.

[0059] The first matrix is ​​copied and expanded according to the time dimension of the performance index data to obtain the second matrix. The second matrix adds time dimension information to the first matrix. Assuming that the time dimension of a certain cell 'a' in the performance index data is m2 (for example, the performance index has m2 15-minute time point spans), then the spliced ​​data of cell 'a' in the first matrix is ​​directly copied m2 times. This process is repeated for all cells, ultimately generating a second matrix Z1 with m×n dimensional features (here it is assumed that the expanded overall matrix has such a dimensional mapping of m×n).

[0060] A sparse matrix is ​​constructed based on parameter modification records. A sparse matrix is ​​a matrix where most elements are zero or invalid default values, with non-zero valid values ​​only at specific positions. The minute-level time points and modification values ​​of actual parameter modifications can be extracted to construct the sparse matrix. Assuming that after filling with parameter modification records, there are p smallest granularity (number of time points) within a 15-minute time range, and s load balancing parameters have been modified (the number of parameters filtered for the current scenario), then a sparse matrix Z2 with dimension p×s is constructed. For example, if the load balancing user limit is adjusted only in the third 15-minute period, the matrix will only have values ​​in the corresponding row and column, with the rest being empty.

[0061] The values ​​of the same time and the same parameters in the sparse matrix are replaced with the corresponding positions in the second matrix to obtain a feature matrix with uniform granularity. Replacement refers to using the non-zero valid values ​​in the sparse matrix to cover the historical old values ​​at the corresponding positions in the second matrix. For each cell, the values ​​of the sparse matrix Z2 at the corresponding time point and the corresponding parameters can replace the data of the same time and the same parameters in the second matrix Z1, thus finally obtaining a feature matrix Z with a dimension of M×N.

[0062] This application embodiment obtains a first matrix by concatenating cell-level parameter data, then expands it by copying and expanding it according to the time dimension of performance indicators to obtain a second matrix. A sparse matrix is ​​constructed using parameter modification records, and corresponding replacements of non-zero elements are performed. This method perfectly reconstructs, aligns, and integrates multi-source heterogeneous data with different spatial granularities (cell-level) and temporal granularities (minute-level modification records versus larger granularities such as 15-minute performance indicators) into a feature matrix of uniform granularity. This approach accurately preserves the time nodes and numerical mutation characteristics of parameter modifications, avoiding information loss caused by crude aggregation, and providing a high-quality and uniform-dimensional data source for subsequent deep feature extraction.

[0063] In one embodiment, feature extraction from the feature matrix to obtain time-series features may further include: Long Short-Term Memory (LSTM) networks are used to extract time-step features from the feature matrix, and the output of the last time step is used as the temporal feature.

[0064] Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network used to process time series data. Time-step feature extraction refers to the network processing the data at each time point sequentially according to the time series. The final output of each time step is the hidden layer feature state vector, representing the result of processing the data sequences at all defined time steps (time points).

[0065] Assume the input feature matrix Z contains M samples, each sample has N features, and each sample is divided into T time steps, with F features at each time step. First, the two-dimensional feature matrix is ​​reshaped into a three-dimensional tensor form that meets the input requirements of an LSTM through a matrix transformation (Reshape). The corresponding reshaping formula is: → ; Where R represents the real number set space mapping, M represents the number of samples, N represents the total number of features per sample, T represents the number of time steps, and F represents the number of features per dimension per time step.

[0066] Subsequently, the reshaped tensor is fed into the LSTM layer as input. The LSTM layer is configured to return only the output of the last time step as the extraction result, and its output dimension satisfies the following formula definition: Lstm output shape=M u; Where u is the number of hidden units set in the LSTM layer.

[0067] Finally, M A u-dimensional vector is used as a temporal feature.

[0068] This application embodiment extracts time-step features from the feature matrix using a Long Short-Term Memory (LSTM) network, and only extracts the output of the last time step as the temporal feature. This effectively utilizes the memory characteristics of LSTM to extract and condense deep historical evolution patterns in time series data. This approach eliminates redundant intermediate outputs from historical moments, ensuring that the model can remember the core information of past load change trends, greatly enhancing the final feature's ability to represent and capture the dynamic changes in the network state.

[0069] In one embodiment, feature extraction is performed on the feature matrix to obtain non-temporal features, including: Configuration parameter features that influence inter-cell equilibrium are selected from the feature matrix as non-time-series features.

[0070] The configuration parameters affecting inter-cell load balancing fall under the aforementioned non-time-series features. Specifically, these refer to scalar values ​​such as threshold values ​​that are relatively static and unchanging in network configuration settings and directly control or influence load handover decisions and traffic balancing between adjacent cells. Parameters affecting inter-cell load balancing can be filtered from the overall feature matrix and selected as the non-time-series features for this scenario.

[0071] Combination Figure 4 These non-time-series characteristics include, but are not limited to, parameters such as: user number threshold, upper limit of connected load balancing users (ConnMlbUeNumThld, corresponding to the user number limit in RRC_CONNECTED state), uplink user number difference threshold (UlUeNumDiffThld, bias controlling switching), and load balancing algorithm switch (MlbAlgoSwitch).

[0072] Assuming the total number of samples is M and the number of selected features is v, then the dimension of the non-temporal feature data output from the segmentation is M. v.

[0073] This application embodiment selects configuration parameter features that affect inter-cell equilibrium from the feature matrix as non-time-series features. This allows the static configuration physical rules that are strongly correlated with the load balancing mechanism and determine the ease of switching to be directly introduced into the model input. This provides clear prior constraints and boundary reference parameters for the prediction model, avoids the blindness brought by pure data-driven approaches, and improves the rationality and physical comprehensiveness of the final fusion feature combination.

[0074] In one embodiment, feature extraction is performed on the feature matrix to obtain time-related features, including: The correlation coefficients of each indicator in the feature matrix at different time granularities are calculated by level transformation, and the correlation coefficients are mapped to time-series correlation features.

[0075] Rank transformation is a nonparametric statistical method that, independent of the specific numerical distribution of the data, replaces the original continuous numerical data with their rank in a sorting order. Correlation coefficients, such as Pearson or Spearman coefficients, are used to measure the strength of a linear or nonlinear association between two or more random variables.

[0076] For each specific time granularity, such as 15 minutes or 1 hour, time series of multiple indicators can be extracted, such as user count sequences, uplink and downlink traffic volume sequences, and feature combination sequences of uplink and downlink traffic volume and rate. Correlation coefficients based on level transformation are calculated between these indicator sequences. Subsequently, the coefficient values ​​between these obtained indicator pairs are used to construct a correlation matrix, and this matrix is ​​mapped and assigned as the time-series correlation features of the sample.

[0077] This application's embodiments calculate the correlation coefficients of each indicator at different time granularities through rank transformation and map them to time-series correlation features. This can effectively overcome noise interference caused by fluctuations (outliers) in individual extreme indicator data and robustly uncover the inherent correlation and coupling characteristics between different network indicators at different time granularity levels, supplementing the overall fusion features with a key multi-dimensional indicator interaction global perspective.

[0078] In one embodiment, the correlation coefficients of each indicator in the feature matrix at different time granularities are calculated through a rank transformation, and the correlation coefficients are mapped to time-series correlation features, including: The rank-related coefficient method is used to perform rank transformation on different indicators in the feature matrix to obtain rank variables. Calculate the linear correlation coefficients between variables of different orders; By mapping the linear correlation coefficient at each time granularity to the sample points, we obtain the time-series correlation features.

[0079] The rank correlation coefficient method is used to transform different indicators in the feature matrix into rank variables. Rank correlation coefficient methods can be non-parametric methods such as Spearman's rank correlation coefficient. The different indicators are two specific data feature sequences whose correlation needs to be analyzed, denoted as variable sequence X and variable sequence Y. Rank variables are the variable values ​​transformed into new sequences with rankings, represented by ranks RX and RY, respectively.

[0080] The numerical sorting algorithm can be used to perform a ranking transformation and replacement process on all cell data variables X and Y at each time point, thereby converting absolute values ​​into relative ranks to obtain RX and RY. For example, the traffic sequence [100, 50, 200] becomes [2, 1, 3] after ranking transformation.

[0081] Calculate the linear correlation coefficient between different ordinal variables. Specifically, the linear correlation coefficient can be the Pearson correlation coefficient. The linear correlation coefficient between the previously generated ordinal series RX and RY can be calculated using the formula for Pearson correlation analysis (i.e., the product of covariance and standard deviation).

[0082] Mapping the linear correlation coefficient at each time granularity to the sample points yields the time-series correlation features. Assuming the correlation calculation targets w combinations, such as the combination of user count and uplink / downlink traffic volume, or the combination of uplink / downlink traffic volume and speed, etc., if the data has a minimum granularity of 15 minutes, the calculation will result in a dimension of (24...) 4) The correlation matrix of w (24 hours a day, 4 x 15-minute intervals per hour, totaling 96 time intervals) is used. The w-dimensional matrix feature map corresponding to each time interval in this matrix is ​​matched to the corresponding M sample points, and the final output feature dimension is expanded to M. w, that is, the time-related features were obtained.

[0083] This application's embodiments obtain rank variables by performing rank transformation on different indicators based on the rank correlation coefficient method, then calculate the linear correlation coefficient between different rank variables, and finally map them onto sample points. This design cleverly combines the advantages of both nonparametric (rank transformation resists extreme outliers) and parametric (linear correlation is easy to calculate) statistical methods. It can extremely robustly and accurately capture and quantify the convergent and stable coupling strength between multivariate time series, generating high-order correlation features that are highly sensitive to model predictions and extremely valuable.

[0084] The data dimension after integrating the constructed time-series features, non-time-series features, and time-related features is M. (u+v+w)=M n, where M represents the number of samples and n represents the number of features per sample. Subsequent data preprocessing includes data encoding and data standardization. Data encoding can use LabelEncode to convert the data into integer codes. Due to the different feature scales, data standardization can use z-standardization to make the data mean zero and the standard deviation one, thereby improving model convergence speed and performance.

[0085] In one embodiment, the fused features are classified at multiple scales to obtain classification data corresponding to different scales, including: The fused features are input into the encoder, and the attention weights of the fused features are calculated through a self-attention mechanism. The value matrix is ​​then weighted and summed according to the attention weights to obtain the attention output features. The attention output features are processed sequentially by global average pooling and normalized exponential function to obtain the probability distribution of the fused features belonging to different preset categories; The fused features are divided according to the probability distribution to obtain classification data corresponding to different scales.

[0086] The fused features are input into the encoder, and attention weights are calculated using a self-attention mechanism. The value matrix is ​​then weighted and summed based on these attention weights to obtain the attention output features. The encoder can be a stacked structure similar to the encoder in a Transformer architecture model. The self-attention mechanism is an algorithm that calculates the importance of the correlation between features at different positions in a sequence, specifically including multi-head self-attention mechanisms. The attention weights are probability weight score matrices that measure the magnitude of this feature correlation. The value matrix is ​​the feature matrix used to store the actual representation information after the input features are linearly transformed by the weight matrix. The attention output features are the new high-dimensional representation of the features obtained after attention-weighted extraction.

[0087] For specific feature combinations in the fused features used as input for multi-scale classification, five features can be extracted from n features: maximum number of users in the cell (rrc_max), average number of users in the cell (rrc_mean), busy time indicator (busy_time), weekday indicator (weekday), and reference signal received power. These features are then input into the encoder stack. For each input sample, a fully connected network is used to compute the query, key, and value matrices, where the query and key matrices are both m in dimension. d model d model The hidden dimension for the model is typically set to a large value, such as 512 or 1024, to preserve feature capacity. Then, attention weights are calculated using the dot product of the query matrix and the key matrix, combined with Softmax. This weight matrix also has a dimension of m. dmodel; then, based on these calculated weights, tensor weighted summation is performed on the value matrix to obtain the preliminary attention output features.

[0088] The attention output features are sequentially processed by global average pooling and exponential function normalization to obtain the probability distribution of the fused features belonging to different preset categories. The preset categories can be multiple predefined categories derived from multi-scale segmentation of the base station features, for example, four-scale segmentation. The attention features, after auxiliary stabilization training processes such as residual connections, layer normalization, and feedforward neural networks, are input into the output layer network. First, global average pooling is performed sequentially to compress the feature output dimension to 1. d model Then, after exponential normalization using the Softmax function layer, the output dimension becomes m. A probability tensor of size 4, where each row of the tensor represents the probability distribution of each input sample belonging to one of the four categories.

[0089] The fused features are divided according to probability distribution to obtain classification data corresponding to different scales. The model divides the entire feature set containing a total of M×n features into four independent classification data subsets corresponding to four preset categories, denoted as data=[M1...]. n, M2 n, M3 n, M4 [n], where M1+M2+M3+M4=M.

[0090] This application's embodiments utilize an encoder and a self-attention mechanism to calculate attention weights for the fused features and perform weighted summation. Global average pooling and normalized exponential function processing are then applied sequentially to obtain a probability distribution, which is subsequently used for data partitioning. This multi-scale classification mechanism based on the Transformer architecture can adaptively perform intelligent weighting and contextual pattern recognition on the extremely complex interdependencies within the fused features, thereby outputting a highly accurate classification probability distribution and performing data partitioning. This advanced feature extraction and clustering mechanism ensures high rationality, flexibility, and robustness in multimodal partitioning of live network data, enabling subsequent models to exhibit stronger adaptability when dealing with data mutations and complex live network scenarios.

[0091] In one embodiment, the classification data is mapped and predicted to obtain performance metrics for a wireless load balancing scenario, including: Different scales of classification data are input into the corresponding multi-layer fully connected neural networks; Abstract features are extracted by non-linear mapping of categorical data through hidden layers in a multi-layer fully connected neural network. The performance index is obtained by performing linear numerical extrapolation on the abstract features through the output layer of a multi-layer fully connected neural network.

[0092] Different scales of classification data are input into corresponding multilayer fully connected neural networks (MLPs). All artificial nodes between layers of a MLP are fully connected via weight coefficients. Each of the multiple sets of classification data (such as the four sets M1, M2, etc. mentioned above) pre-divided according to scale in the previous stage can be used as input data for the MLP, and then input into four pre-constructed, independent, and parallel MLP models, each corresponding to one of the four categories.

[0093] Abstract features are extracted by nonlinearly mapping categorical data through hidden layers in multi-layer fully connected neural networks. Hidden layers are fully connected matrix computation layers located between the input and output layers of the neural network. Nonlinear mapping utilizes nonlinear activation functions to disrupt the linear combination properties of data, performing high-dimensional folding and spatial transformation. Abstract features are high-order semantic vector expressions without intuitive physical meaning, formed after multiple layers of mathematical matrix transformations. Data can undergo multiple weight matrix multiplications and forward computations of nonlinear activation functions within the hidden layers of each MLP to extract deep, intrinsic relationships.

[0094] Performance metrics are obtained by performing linear numerical extrapolation on abstract features through the output layer of a multi-layer fully connected neural network. The output layer is the last layer of the neural network computation graph, and the number of its neurons is determined by the number of performance features to be predicted. Linear numerical extrapolation refers to forgoing the use of non-linear activation functions and instead directly outputting continuous predicted numerical results through weighted summation of linear matrices. In the output layer of an MLP, the final weight matrix can be directly combined with a linear activation function to calculate the output regression prediction results, thereby obtaining the predicted performance metrics, such as... Figure 3 The line chart displays continuous metrics such as average uplink and downlink data rates (AvgDataRateDl / Ul), transmission traffic (pdcp_upoctdl / ul), and physical resource block utilization (PrbAssnRateDl / Ul). During actual inference, the system matches the sample classification criteria to the corresponding class and calls the corresponding MLP model to deduce the output metrics.

[0095] This application's embodiments achieve customized regression predictions for different network data scales by feeding classified data at different scales into their corresponding independent multi-layer fully connected neural networks. Leveraging the powerful nonlinear mapping of the hidden layers and the linear numerical regression deduction of the output layers, it effectively approximates the extremely complex unknown nonlinear function mapping boundary between massive, complex wireless parameters and final performance indicators. This significantly reduces prediction bias caused by mutual interference between multi-scale features, thereby greatly improving the accuracy of indicator prediction results in various scenarios.

[0096] In one embodiment, abstract features are extracted by non-linearly mapping categorical data through hidden layers in a multi-layer fully connected neural network, including: Preliminary features are obtained by feature mapping of the classification data using the first-level neural network in the hidden layer based on the first non-linear activation function. Abstract features are obtained by further processing the preliminary features using a second-level neural network in the hidden layer based on a second nonlinear activation function.

[0097] The first-level neural network is the first fully connected parametric layer in the hidden layer stack structure, and can specifically contain 1024 computational neurons. The first non-linear activation function is the Rectified Linear Unit (ReLU) activation function. Feature mapping involves performing tensor multiplication, weighted summation, and biasing on the input classification data, followed by filtering through the activation layer. The preliminary features are the intermediate results of the low-dimensional hidden vectors output from the first layer's forward computation.

[0098] The second-level neural network is the second fully connected parametric layer in the hidden layer stack, immediately following the first layer. To form the bottleneck structure, it can contain 512 computational neurons. The second nonlinear activation function can also be the ReLU activation function.

[0099] In this embodiment, preliminary features are obtained through a first-level neural network and a first nonlinear activation function in the hidden layer. Then, abstract features are obtained through further processing using a second-level neural network and a second nonlinear activation function. This carefully designed two-layer cascaded processing structure (such as a shrinking dimension from 1024 to 512) can extract deeper, more representative, and more abstract intrinsic high-order features from classification data in an extremely efficient and progressive manner. This combination of deep nonlinear feature mapping significantly enhances the nonlinear expression and generalization extraction capabilities of the entire multilayer fully connected neural network in dealing with complex data patterns.

[0100] In one embodiment, a multilayer fully connected neural network is trained through the following steps: Construct a long short-term memory network model for processing historical sample labels and an initial multilayer fully connected neural network model; The first loss value is calculated based on the predictions of the multilayer fully connected neural network model using historical sample labels, and the second loss value is calculated based on the predictions of the long short-term memory network model. The first loss value and the second loss value are weighted and summed to obtain the comprehensive loss function; The network parameters of the multilayer fully connected neural network model are updated based on the comprehensive loss function.

[0101] Historical sample labels are real and actual performance metric observations recorded in the historical training dataset, such as actual throughput, which serve as absolute guidance for supervised learning.

[0102] The predicted value is the estimated result of various indicators calculated and extrapolated by the model's forward propagation in the current round. The loss value is used to measure the deviation between the current predicted value and the true historical sample label. For each training sample i, its true historical sample label is y. i The actual predicted value obtained by the model forward calculation for this sample is The mean squared error loss function (MSE) is defined as follows: ; Following the mathematical formula, the difference and squaring operations are performed respectively, and the calculated loss value of MLP is recorded as L1. i (i.e., the first loss value), and the loss value of the LSTM model is calculated in parallel and recorded as L2. i (i.e., the second loss value).

[0103] Assuming the weights of the MLP (first weight) are pre-set as weight1 and the weights of the LSTM (second weight) are pre-set as weight2, then the comprehensive loss function L is: ; Can be used Figure 2 The backpropagation shown updates the internal network parameter values ​​such as weights and biases in the MLP model. Specifically, it calculates the gradient tensor of the comprehensive loss function with respect to the parameters of each layer, and then multiplies the gradient by a preset learning rate before applying it to the iterative optimization and update of the network parameters.

[0104] This application's embodiments calculate the first predicted loss value generated by a multi-layer fully connected neural network and the second predicted loss value generated by a long short-term memory network model based on unified historical sample labels. These two values ​​are then weighted and summed to construct a comprehensive loss function that uniformly guides network parameter updates. This joint loss correction training method, under limited computational resources, enables the model to simultaneously learn the time-series evolution patterns (temporal dependencies) and complex nonlinear static mapping relationships of wireless data. It achieves global-perspective optimization of wireless communication network prediction, effectively avoiding the problems of local optima and prediction instability inherent in single network architectures, and realizing a more comprehensive and stable fusion of model parameter optimization.

[0105] In one embodiment, updating the network parameters of a multilayer fully connected neural network model based on a comprehensive loss function includes: Calculate the parameter gradients of the multilayer fully connected neural network model based on the comprehensive loss function; By using an adaptive learning rate optimization algorithm, the network parameters of the multilayer fully connected neural network model are updated backward using parameter gradients, with the goal of minimizing the mean absolute percentage error between the predicted values ​​of the multilayer fully connected neural network model and the historical sample labels.

[0106] The parameter gradient is a set of tensors consisting of the partial derivatives of the scalar of the comprehensive loss function with respect to each learnable weight parameter in the MLP, indicating the direction of the fastest error descent. The adaptive learning rate optimization algorithm can specifically be the Adam optimizer (Adaptive Moment Estimation), which dynamically and adaptively adjusts and allocates different learning rates for each parameter based on historical first- and second-moment estimates of the parameter gradient.

[0107] By calling adaptive algorithm engines such as the Adam optimizer, the parameter gradient tensor obtained from the backpropagation in the previous step can be read. Combined with the exponential moving average information of historical gradients, the step size of each iteration can be dynamically and adaptively updated. The learning step size is multiplied by the gradient and then subtracted from the current MLP model parameter matrix to achieve backpropagation. After each or multiple iterations, the MAPE error index is used to evaluate and check whether convergence has been achieved.

[0108] This application's embodiments combine the gradients of multi-layer fully connected network parameters calculated using the comprehensive loss function, and employ an adaptive learning rate optimization algorithm to directly minimize the mean absolute percentage error (MAPE), thereby updating the network parameters in reverse using the parameter gradients. This optimization mechanism allows the model to adaptively and dynamically adjust the optimization step size based on the gradient change characteristics of different parameters during reverse training. This not only significantly improves the convergence speed of the complex parameter space of the network but also significantly enhances the stability of the training process, thereby ensuring that the final model has extremely low relative error when outputting predictive performance metrics, providing highly reliable theoretical support for modifying parameters in the external field.

[0109] As described above, this application's embodiments construct a simulation framework driven by both data and knowledge. Based on real-world load balancing scenarios, this framework systematically adjusts and sets parameters to collect network performance data and conducts in-depth analysis of the collected data. It explores the intrinsic correlation between parameter configuration and network performance, thereby extracting key factors for optimizing network strategies. This not only helps deepen the understanding of network dynamics but also provides accurate and efficient decision support for on-site network optimization engineers, significantly improving the quality and efficiency of network optimization work.

[0110] This application's embodiments integrate the time-series characteristics of existing network performance indicators, the correlation features between different performance indicators, and multi-scale information under load balancing environments. By employing nonlinear activation functions and a multi-level model design, it achieves the ability to efficiently learn nonlinear relationships from complex datasets. This method significantly enhances the model's generalization ability and prediction accuracy.

[0111] This application's embodiments utilize offline generated models, effectively reducing the time consumed by large-scale traversal and repeated trials required during parameter exploration, thereby avoiding the problem of system performance degradation due to frequent parameter adjustments and improving the overall system stability and efficiency.

[0112] Compared with existing technologies, the embodiments of this application comprehensively evaluate the impact of existing network data on network performance indicators and cost functions by effectively dividing the data into multiple modalities. This comprehensive approach greatly enhances the model's generalization ability, enabling it to more robustly cope with various complex scenarios during network optimization, especially demonstrating stronger adaptability and reliability when dealing with unexpected situations.

[0113] Unlike traditional methods that rely solely on static data analysis, this application innovatively introduces a combination of temporal and non-temporal features to deeply explore the patterns of wireless network parameter measurement data changes over time and their relationship with network performance. This method not only better captures the changing trends of network status but also, under limited computing resources, optimizes the wireless communication network from a global perspective through loss correction and other means, improving the model's prediction accuracy and stability.

[0114] This application's embodiments achieve efficient prediction of future performance trends in wireless communication networks by introducing the correlation coefficients of various performance indicators at the finest granular time level as input variables. Compared to traditional prediction methods that rely solely on historical data, this application's embodiments not only enable more refined predictions but also demonstrate significant performance gains in tests on large-scale datasets. Particularly when handling large-scale datasets, it effectively improves prediction efficiency and accuracy, thus providing a more solid data foundation for network optimization.

[0115] The embodiments of this application can predict and output relevant performance indicators for multiple time periods under various effective parameter configuration combinations in load balancing scenarios for on-site network optimization engineers. It can quickly compare the positive / negative gain impact of different parameter configurations on the indicators, providing accurate and efficient decision support, and greatly improving the quality and efficiency of network optimization work.

[0116] The embodiments of this application can achieve efficient multi-region learning and prediction, and are capable of efficiently learning nonlinear relationships from complex datasets. This method significantly enhances the model's generalization ability and prediction accuracy.

[0117] This application embodiment uses a method of multi-dimensional data segmentation and comprehensive consideration of parameter influence to more accurately assess the impact of different factors on network performance, thereby optimizing network configuration and improving service quality.

[0118] The application of time-series and non-time-series feature fusion optimization technology in this application helps operators more accurately predict changes in network demand, rationally plan resource allocation, and avoid resource waste or shortages. Furthermore, through in-depth analysis of the time-varying patterns of wireless network parameter measurement data, potential problems can be effectively identified and resolved, reducing failure rates and further lowering maintenance costs.

[0119] The embodiments of this application can, to a certain extent, achieve the goal of internalizing external field indicators. By utilizing the twins and simulations of the network simulation environment, and then outputting data and parameters that have theoretical support for the existing network based on the model trained by AI, the high-performance prediction efficiency and global analysis capabilities are improved, providing strong technical support for operators. This technology also makes it possible to explore new service models.

[0120] This application's embodiments were tested and verified using data from multiple provincial companies. The error was reduced from 170% to approximately 30%, and the uplink and downlink PRB utilization rate was reduced to 10%. The test results are as follows. Figure 3 As shown.

[0121] By using an offline-generated model, this embodiment effectively reduces the time and resource consumption during online parameter exploration, avoiding the negative impact of frequent parameter adjustments on system performance. This not only improves the overall operating efficiency of the product in commercial applications but also enhances its stability, while reducing the risk of service interruptions or performance degradation due to improper parameters, which is particularly important for network environments that require continuous high-quality service.

[0122] By integrating time series characteristics, correlations between performance metrics, and multi-scale information, and employing nonlinear activation functions and multi-level model design, the embodiments of this application significantly enhance the model's ability to handle complex datasets, particularly in learning nonlinear relationships. This means the model can more accurately predict the changing trends of network performance, providing solid technical support for the automated management and optimization of products.

[0123] The performance index prediction device for wireless load balancing scenarios provided in the embodiments of this application is described below. The performance index prediction device for wireless load balancing scenarios described below can be referred to in correspondence with the performance index prediction method described above.

[0124] like Figure 5 As shown in the figure, the performance indicator prediction device for wireless load balancing scenarios provided in this application embodiment includes: The matrix construction module is used to construct a feature matrix with uniform granularity from multi-source network data. The feature extraction module is used to extract features from the feature matrix, obtaining time-series features, non-time-series features, and time-series related features, and integrating these features into a fusion feature. The feature classification module is used to perform multi-scale classification on the fused features to obtain classification data corresponding to different scales; The indicator prediction module is used to map and predict categorized data to obtain performance indicators for wireless load balancing scenarios.

[0125] Figure 6 A schematic diagram of the physical structure of an electronic device is provided. This electronic device may include a processor, a communications interface, memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions from the memory to execute a performance metric prediction method.

[0126] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0127] On the other hand, embodiments of this application also provide a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the performance index prediction methods provided by the above methods.

[0128] In another aspect, embodiments of this application also provide a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the performance index prediction methods provided by the methods described above.

[0129] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0130] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting performance indicators, characterized in that, include: Construct a feature matrix with uniform granularity from multi-source network data; Feature extraction is performed on the feature matrix to obtain time-series features, non-time-series features, and time-series related features, and the time-series features, non-time-series features, and time-series related features are integrated into a fusion feature; The fused features are classified at multiple scales to obtain classification data corresponding to different scales; The classification data is mapped and predicted to obtain performance indicators for the wireless load balancing scenario.

2. The performance index prediction method according to claim 1, characterized in that, The multi-source network data includes cell-level parameter data, performance index data, and parameter modification record data; Constructing a feature matrix with uniform granularity from multi-source network data includes: The cell-level parameter data of each cell are concatenated to obtain the first matrix; The first matrix is ​​copied and expanded according to the time dimension of the performance index data to obtain the second matrix; Based on the parameters, modify the recorded data to construct a sparse matrix; The values ​​of the same time and the same parameter in the sparse matrix are replaced with the corresponding positions in the second matrix to obtain the feature matrix with uniform granularity.

3. The performance index prediction method according to claim 1, characterized in that, Feature extraction is performed on the feature matrix to obtain the time-series features, including: The feature matrix is ​​extracted using a long short-term memory network, and the output of the last time step is used as the temporal feature.

4. The performance index prediction method according to claim 1, characterized in that, Feature extraction is performed on the feature matrix to obtain the non-temporal features, including: The configuration parameter features that affect inter-cell equilibrium are selected from the feature matrix as the non-time-series features.

5. The performance index prediction method according to claim 1, characterized in that, Feature extraction is performed on the feature matrix to obtain the time-series related features, including: The correlation coefficients of each index in the feature matrix at different time granularities are calculated by level transformation, and the correlation coefficients are mapped to the time-series correlation features.

6. The performance index prediction method according to claim 5, characterized in that, The correlation coefficients of each index in the feature matrix at different time granularities are calculated through rank transformation, and the correlation coefficients are mapped to the time-series correlation features, including: The different indicators in the feature matrix are subjected to rank transformation based on the rank correlation coefficient method to obtain rank variables; Calculate the linear correlation coefficients between the different ordinal variables; The linear correlation coefficient at each time granularity is mapped to the sample points to obtain the time-series correlation features.

7. The performance index prediction method according to claim 1, characterized in that, The fused features are then subjected to multi-scale classification to obtain classification data corresponding to different scales, including: The fused features are input into the encoder, and the attention weights of the fused features are calculated through a self-attention mechanism. The value matrix is ​​then weighted and summed according to the attention weights to obtain the attention output features. The attention output features are sequentially processed by global average pooling and normalized exponential function to obtain the probability distribution of the fused features belonging to different preset categories; The fused features are divided according to the probability distribution to obtain the classification data corresponding to different scales.

8. The performance index prediction method according to claim 1, characterized in that, The classification data is mapped and predicted to obtain performance indicators for the wireless load balancing scenario, including: The classification data at different scales are respectively input into the corresponding multi-layer fully connected neural network; Abstract features are extracted by nonlinearly mapping the classified data through the hidden layers in the multi-layer fully connected neural network. The performance index is obtained by performing linear numerical extrapolation on the abstract features through the output layer of the multi-layer fully connected neural network.

9. The performance index prediction method according to claim 8, characterized in that, Abstract features are extracted by nonlinearly mapping the classification data through the hidden layers of the multi-layer fully connected neural network, including: Preliminary features are obtained by performing feature mapping on the classification data using a first-level neural network in the hidden layer based on a first non-linear activation function. The abstract features are obtained by further processing the preliminary features using a second-level neural network in the hidden layer based on a second nonlinear activation function.

10. The performance index prediction method according to claim 8, characterized in that, The multilayer fully connected neural network is trained through the following steps: Construct a long short-term memory network model for processing historical sample labels and an initial multilayer fully connected neural network model; The first loss value generated by the predicted value of the multilayer fully connected neural network model is calculated based on the historical sample labels, and the second loss value generated by the predicted value of the long short-term memory network model is calculated. The first loss value and the second loss value are weighted and summed to obtain the comprehensive loss function; The network parameters of the multilayer fully connected neural network model are updated based on the comprehensive loss function.

11. The performance index prediction method according to claim 10, characterized in that, Updating the network parameters of the multilayer fully connected neural network model according to the comprehensive loss function includes: The parameter gradients of the multilayer fully connected neural network model are calculated based on the comprehensive loss function. The network parameters of the multilayer fully connected neural network model are updated backward using the parameter gradient by means of the adaptive learning rate optimization algorithm, with the goal of minimizing the mean absolute percentage error between the predicted value of the multilayer fully connected neural network model and the historical sample label.

12. A performance index prediction device for a wireless load balancing scenario, characterized in that, include: The matrix construction module is used to construct a feature matrix with uniform granularity from multi-source network data. The feature extraction module is used to extract features from the feature matrix to obtain time-series features, non-time-series features, and time-related features, and to integrate the time-series features, non-time-series features, and time-related features into a fusion feature. The feature classification module is used to perform multi-scale classification on the fused features to obtain classification data corresponding to different scales; The indicator prediction module is used to map and predict the classified data to obtain the performance indicators of the wireless load balancing scenario.

13. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the performance metric prediction method as described in any one of claims 1 to 11.

14. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the performance metric prediction method as described in any one of claims 1 to 11.

15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the performance metric prediction method as described in any one of claims 1 to 11.

Images