Communication network resource optimization method and system based on artificial intelligence

Abstract

The present invention provides a communication network resource optimization method and system based on artificial intelligence, which relates to the field of communication optimization technology. By collecting historical network data, using time series decomposition and LSTM algorithm to extract features, combining the attention mechanism to calculate the weights of performance indicators, adopting a multi-scale prediction mechanism to predict network traffic, and dynamically calibrating based on historical errors, the method can accurately predict the changing trend of network traffic, improve prediction accuracy, realize the rational allocation and optimal configuration of network resources, and effectively improve network operation efficiency.

Description

Technical Field

[0001] The present invention relates to the field of communication optimization technology, and in particular to a communication network resource optimization method and system based on artificial intelligence. Background Art

[0002] With the rapid development of 5G networks and the deepening of digital transformation, the types of services and data traffic carried by communication networks have experienced explosive growth. Efficient network resource management and optimization are of great significance to ensuring network service quality and improving user experience.

[0003] Traditional communication network resource optimization methods rely primarily on manual experience and fixed rules, making them difficult to adapt to complex and changing network environments and dynamically changing business needs. Currently, statistical analysis-based methods are commonly used to predict network traffic and optimize resources, using simple time series models or regression analysis to model and analyze historical data.

[0004] Existing network resource optimization technologies still focus on predictions on a single time scale, making it difficult to simultaneously account for the changing characteristics of network traffic over different time periods. This leads to suboptimal prediction accuracy, inadequate consideration of the correlation between network device performance indicators and traffic changes, neglect of the impact of network status on traffic prediction, and a lack of effective prediction result calibration mechanisms, making it impossible to dynamically adjust and optimize the prediction model based on actual network operation conditions.

[0005] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the Invention

[0006] The embodiments of the present invention provide a communication network resource optimization method and system based on artificial intelligence, which can at least solve some of the problems existing in the prior art.

[0007] A first aspect of an embodiment of the present invention provides a method for optimizing communication network resources based on artificial intelligence, comprising:

[0008] Collect historical network traffic data, network equipment performance data, and user service demand data from the communication network. Decompose the historical network traffic data into trend items, period items, and residual items using a time series decomposition algorithm. Calculate bandwidth utilization, processor utilization, memory utilization, and transmission delay from the network equipment performance data.

[0009] The trend term, the period term, and the residual term are combined into an input vector, the input vector is added to the long short-term memory network, the input vector is gated by a forget gate, an input gate, and an output gate, the unit state is updated to obtain a hidden state, and the hidden state is linearly transformed to obtain a temporal code;

[0010] Bandwidth occupancy, processor occupancy, memory occupancy, and transmission delay are combined into a performance indicator vector. An attention mechanism is introduced to calculate the performance indicator vector to obtain the importance weight. The temporal coding and the importance weight are then fused to obtain the fused feature.

[0011] A multi-scale prediction mechanism is used to process the fused features to predict the short-term, medium-term, and long-term network traffic values ​​respectively. Adaptive weights are set based on the prediction errors at each time scale, and the prediction results at multiple time scales are weighted and combined to output the initial network traffic prediction value.

[0012] A correction model is constructed based on the statistical characteristics of historical prediction errors. The initial network traffic prediction value is dynamically calibrated using the exponential smoothing method. The prediction value is corrected in combination with the mutation detection results of network performance indicators to obtain the network traffic prediction result.

[0013] In an optional embodiment,

[0014] Collect historical network traffic data, network equipment performance data, and user service demand data from the communication network. Decompose the historical network traffic data into trend items, period items, and residual items using a time series decomposition algorithm. Calculate the bandwidth utilization, processor utilization, memory utilization, and transmission delay in the network equipment performance data, including:

[0015] Collect historical network traffic data in the communication network within a preset sampling period, the historical network traffic data being the number of traffic bytes on the network link per unit time, and network device performance data reflecting the working status of the network link, as well as user service demand data;

[0016] Inputting the historical network traffic data into a time series decomposition algorithm to extract the fluctuation pattern of the historical network traffic data over time, thereby obtaining a trend term reflecting the long-term trend, a period term reflecting the periodic change, and a residual term reflecting the random fluctuation;

[0017] Statistical calculations are performed on the network device performance data, and sliding averages are performed on the network device performance data according to a preset time window to obtain bandwidth occupancy, processor occupancy, memory occupancy, and transmission delay.

[0018] In an optional embodiment,

[0019] The trend term, the period term, and the residual term are combined into an input vector, the input vector is added to the long short-term memory network, the input vector is gated by the forget gate, the input gate, and the output gate, the unit state is updated to obtain the hidden state, and the hidden state is linearly transformed to obtain the temporal encoding, including:

[0020] Sequentially concatenate the trend term, the period term, and the residual term in the time dimension to form an input vector, wherein the input vector contains historical values ​​at different time steps;

[0021] Based on the input vector and pre-acquired historical state information, randomly forgetting the historical information through a forget gate, calculating the current update information and generating a candidate unit state through an input gate, combining the retained portion of the historical information with the current update information to update the unit state, and performing a gating operation on the unit state through an output gate to obtain a hidden state;

[0022] Performing a linear transformation on the hidden state to obtain a temporal code.

[0023] In an optional embodiment,

[0024] Bandwidth occupancy, processor occupancy, memory occupancy, and transmission delay are combined into a performance indicator vector. The attention mechanism is introduced to calculate the performance indicator vector to obtain the importance weight. The time series coding and the importance weight are fused to obtain the fused features including:

[0025] Collecting performance parameters of the network device at multiple consecutive time points, the performance parameters including bandwidth occupancy, processor occupancy, memory occupancy, and transmission delay, and forming a performance indicator vector from the performance parameters collected at each time point;

[0026] Based on the attention mechanism, the performance indicator vector is input into the query matrix, key matrix and value matrix to obtain the query vector, key vector and value vector. The correlation score between the performance indicators is obtained based on the dot product operation of the query vector and the key vector. The correlation score is normalized by softmax to obtain the importance weight. The importance weight represents the influence of each performance indicator on the network performance.

[0027] The temporal coding and the importance weight are weighted and superimposed to obtain a fusion feature.

[0028] In an optional embodiment,

[0029] A multi-scale prediction mechanism is used to process the fusion features to predict the short-term, medium-term and long-term network traffic values ​​respectively. Adaptive weights are set based on the prediction errors at each time scale. The prediction results of multiple time scales are weighted and combined to output the initial network traffic prediction values, including:

[0030] A two-layer long short-term memory network structure is used to construct prediction processing units for three time scales: short-term, medium-term, and long-term. The two-layer long short-term memory network structure processes the fused features and outputs the initial network traffic prediction value at each time scale.

[0031] Continuously collect network traffic data at each time scale within a preset sampling time window, calculate the rate of change of the network traffic data at each time scale, and calculate the temporal correlation between adjacent time scales and the degree of fluctuation of each time scale based on the rate of change; set a corresponding temperature coefficient for each time scale based on the degree of fluctuation, where the greater the degree of fluctuation, the smaller the temperature coefficient;

[0032] The historical samples are divided into a plurality of continuous training sequence segments according to a preset time interval. The difference between the actual value of the network traffic and the initial network traffic prediction value at the corresponding time scale is calculated in each training sequence segment. The difference is cumulatively summed and the average value is calculated to generate the prediction error at each time scale. The prediction error is multiplied by the temperature coefficient of the corresponding time scale, and the inverse is taken and substituted into the exponential function for operation to generate an exponential mapping value. The ratio of the exponential mapping value to the sum of the exponential mapping values ​​of the three time scales is calculated to determine the initial adaptive weight corresponding to each time scale.

[0033] The regular characteristics of temporal correlation between time scales are extracted from the graph structure network. Constraint relationships between prediction values ​​of adjacent time scales are established based on the regular characteristics. The initial network traffic prediction value is revised based on the constraint relationship and the federated learning framework to generate a revised prediction value that considers temporal correlation. The initial adaptive weight is adjusted based on the revised prediction value to obtain a revised weight that considers temporal correlation. The revised weight corresponding to the time scale with a larger degree of fluctuation is relatively increased.

[0034] Perform a weighted sum operation on the corrected prediction value and the corresponding corrected weight to output the final network traffic prediction value.

[0035] In an optional embodiment,

[0036] The regular characteristics of temporal correlation between time scales are extracted from the graph structure network. Constraint relationships between prediction values ​​at adjacent time scales are established based on the regular characteristics. The initial network traffic prediction value is modified based on the constraint relationship and the federated learning framework to generate a modified prediction value that takes into account temporal correlation. The method includes:

[0037] A dynamic heterogeneous graph network structure is constructed, connecting network traffic nodes at different time scales through dynamic edges. Node attributes include traffic values, statistical characteristics, and change trends, and edge attributes are initialized to the basic temporal correlation between adjacent time scales. A spatiotemporal attention mechanism is used to dynamically update node and edge attributes, outputting the updated node state vector and edge correlation strength.

[0038] Based on the updated node state vectors and edge correlation strengths, a temporal correlation extraction model is constructed. This model captures the changing characteristics of network traffic at different time scales through a sliding time window. Furthermore, the model combines counterfactual reasoning methods to identify the dominant factors of temporal correlation and extract the regular characteristics of temporal correlation changes over time.

[0039] A conditional time series pattern graph is constructed based on the extracted regular features, and the time series association patterns under different conditions are mapped to nodes in the graph. Node connections are established based on the evolutionary relationship between the patterns. The conditional time series pattern graph is used to characterize the constraint relationship between the prediction values ​​of adjacent time scales. A distributed federated learning framework is used to integrate the constraint relationship information of multiple network regions. The calculation parameters of the time series correlation degree are adjusted in real time based on the local prediction deviation. The global prediction performance index is used to continuously optimize the extraction method of regular features, and an adaptive probabilistic graph model is constructed to dynamically associate the update frequency of the model parameters with the fluctuation degree of network traffic.

[0040] According to the node state distribution and transition probability in the adaptive probabilistic graph model, an energy function is constructed in combination with the constraint relationship provided by the conditional timing pattern graph. The constraint relationship strength, conditional timing law and probability distribution are used as components of the energy function. The weight coefficient of each component is adaptively adjusted with the change of prediction accuracy. The initial network traffic prediction value is corrected by iteratively optimizing the energy function until all timing constraints are met, and the corrected prediction value considering timing correlation is output.

[0041] In an optional embodiment,

[0042] Based on the statistical characteristics of historical prediction errors, a correction model is constructed. The initial network traffic forecast value is dynamically calibrated using the exponential smoothing method. The forecast value is corrected based on the mutation detection results of network performance indicators. The network traffic forecast results include:

[0043] Obtaining a historical prediction error sequence, the historical prediction error sequence being obtained by subtracting a predicted network traffic value from an actual network traffic value, calculating an error mean and an error standard deviation based on the historical prediction error sequence, constructing an autocorrelation function of the error sequence, and calculating an exponentially weighted moving variance, wherein the exponentially weighted moving variance is determined by a weighted sum of a squared difference between the error and the mean at a current moment and an exponentially weighted moving variance at a previous moment;

[0044] A correction model is constructed based on the error mean, the error standard deviation, the autocorrelation function, and the exponentially weighted moving variance, wherein the correction model includes an error distribution correction term, a time series correlation correction term, and a fluctuation correction term, wherein the error distribution correction term, the time series correlation correction term, and the fluctuation correction term respectively correspond to different adaptive weight coefficients, and the adaptive weight coefficients are updated by a gradient descent method;

[0045] Based on the exponential smoothing method, a historical information storage matrix is ​​constructed, and the initial network traffic forecast value is dynamically calibrated using an adaptive smoothing factor and an attention mechanism to obtain a first revised forecast value. The adaptive smoothing factor is determined by the weighted sum of the smoothing factor at the previous moment and the absolute value of the historical forecast error at the current moment, and the weighting coefficient is dynamically adjusted by the change in the exponentially weighted moving variance.

[0046] Constructing a multidimensional network performance indicator vector, the multidimensional network performance indicator vector including values ​​of multiple network performance indicators at a current moment, performing a cumulative sum operation based on time series differences of the multidimensional network performance indicator vector to obtain a mutation detection value, and comparing the mutation detection value with the difference between the mutation detection value at a previous moment and a detection threshold;

[0047] When the mutation detection value is greater than the mutation determination threshold, the relative change rate of each network performance indicator in the multidimensional network performance indicator vector is calculated, the relative change rate of each network performance indicator is multiplied by the corresponding weight coefficient and the sum is obtained to obtain the mutation degree, and the product of the first corrected prediction value and the mutation degree is output as the network traffic prediction result; when the mutation detection value is not greater than the mutation determination threshold, the first corrected prediction value is output as the network traffic prediction result.

[0048] In an optional embodiment,

[0049] According to the exponential smoothing method, a historical information storage matrix is ​​constructed, and the initial network traffic forecast value is dynamically calibrated through the adaptive smoothing factor and attention mechanism to obtain the first revised forecast value, which includes:

[0050] The network traffic time series is divided into multiple time scale layers according to the time scale, and an independent adaptive smoothing factor is constructed for each time scale layer.

[0051] Calculating the volatility according to the changing trend of the network traffic time series, dynamically adjusting the size of the historical data window based on the volatility, wherein the size of the historical data window is determined by multiplying the basic window size by the volatility, and obtaining smoothing results of each time scale layer within the historical data window;

[0052] The smoothed result is used to construct a feature matrix according to the time scale layer, and attention weights are calculated based on the feature matrix and the query vector at the current moment. The attention weights represent the importance of different time scale features to the current prediction, and the features of different time scales are adaptively fused using the attention weights.

[0053] Constructing a historical information storage matrix, the historical information storage matrix including historical network traffic values, the prediction error, and the adaptive smoothing factor, employing an attention mechanism to extract historical information related to a current state from the historical information storage matrix, and determining a dynamic learning rate based on a change in the prediction error, wherein the dynamic learning rate decays as the change in the prediction error increases;

[0054] A residual connection structure is introduced into the initial network traffic prediction value, and the correction function output of each time scale and the adaptive smoothing factor are weighted and summed according to the attention weight to obtain a residual term, and the residual term is added to the initial network traffic prediction value to obtain a first corrected prediction value.

[0055] A second aspect of an embodiment of the present invention provides an artificial intelligence-based communication network resource optimization system, comprising:

[0056] The first unit is used to collect historical network traffic data, network equipment performance data, and user service demand data in the communication network, decompose the historical network traffic data into trend items, period items, and residual items using a time series decomposition algorithm, and calculate bandwidth utilization, processor utilization, memory utilization, and transmission delay in the network equipment performance data;

[0057] The second unit is configured to combine the trend term, the period term, and the residual term into an input vector, add the input vector to the long short-term memory network, perform a gating operation on the input vector through a forget gate, an input gate, and an output gate, update the unit state to obtain a hidden state, and perform a linear transformation on the hidden state to obtain a temporal code;

[0058] The third unit is used to combine bandwidth occupancy, processor occupancy, memory occupancy, and transmission delay into a performance indicator vector. An attention mechanism is introduced to calculate the importance weight of the performance indicator vector. The time series coding and the importance weight are then fused to obtain a fused feature.

[0059] The fourth unit is used to process the fused features using a multi-scale prediction mechanism to predict the short-term, medium-term, and long-term network traffic values ​​respectively. Adaptive weights are set based on the prediction errors at each time scale, and the prediction results at multiple time scales are weighted and combined to output the initial network traffic prediction value.

[0060] The fifth unit is used to build a correction model based on the statistical characteristics of historical prediction errors, dynamically calibrate the initial network traffic prediction value using the exponential smoothing method, and correct the prediction value in combination with the mutation detection results of the network performance indicators to obtain the network traffic prediction result.

[0061] In the present invention, historical network traffic data is processed by combining time series decomposition and long short-term memory networks, which fully considers the temporal characteristics and long-term dependencies of the data, improves the accuracy and reliability of network traffic prediction, and provides a more accurate decision-making basis for the optimal configuration of network resources. The attention mechanism is introduced to weight the importance of network performance indicators and integrate them with the temporal characteristics, thereby realizing comprehensive perception and feature extraction of network status, enhancing the model's adaptability to changes in the network environment, and making the prediction results more consistent with the actual network operation conditions. A multi-scale prediction mechanism and dynamic calibration method are adopted, and the prediction results of different time scales are combined through adaptive weights, and corrected in combination with the statistical characteristics of historical errors, which effectively reduces the prediction deviation and improves the prediction performance of the model in different time periods, providing more accurate prediction support for the intelligent scheduling and optimization of network resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a flow chart of a method for optimizing communication network resources based on artificial intelligence according to an embodiment of the present invention;

[0063] Figure 2 This is a comparison chart of prediction accuracy under different load conditions corresponding to the communication network resource optimization method based on artificial intelligence according to an embodiment of the present invention;

[0064] Figure 3 This is a graph of prediction error data under different fluctuation scenarios corresponding to the communication network resource optimization method based on artificial intelligence in an embodiment of the present invention. DETAILED DESCRIPTION

[0065] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0066] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0067] Figure 1 FIG. 1 is a flow chart of a method for optimizing communication network resources based on artificial intelligence according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0068] Collect historical network traffic data, network equipment performance data, and user service demand data from the communication network. Decompose the historical network traffic data into trend items, period items, and residual items using a time series decomposition algorithm. Calculate bandwidth utilization, processor utilization, memory utilization, and transmission delay from the network equipment performance data.

[0069] The trend term, the period term, and the residual term are combined into an input vector, the input vector is added to the long short-term memory network, the input vector is gated by a forget gate, an input gate, and an output gate, the unit state is updated to obtain a hidden state, and the hidden state is linearly transformed to obtain a temporal code;

[0070] Bandwidth occupancy, processor occupancy, memory occupancy, and transmission delay are combined into a performance indicator vector. An attention mechanism is introduced to calculate the performance indicator vector to obtain the importance weight. The temporal coding and the importance weight are then fused to obtain the fused feature.

[0071] A multi-scale prediction mechanism is used to process the fused features to predict the short-term, medium-term, and long-term network traffic values ​​respectively. Adaptive weights are set based on the prediction errors at each time scale, and the prediction results at multiple time scales are weighted and combined to output the initial network traffic prediction value.

[0072] A correction model is constructed based on the statistical characteristics of historical prediction errors. The initial network traffic prediction value is dynamically calibrated using the exponential smoothing method. The prediction value is corrected in combination with the mutation detection results of network performance indicators to obtain the network traffic prediction result.

[0073] In an optional embodiment,

[0074] Collect historical network traffic data, network equipment performance data, and user service demand data from the communication network. Decompose the historical network traffic data into trend items, period items, and residual items using a time series decomposition algorithm. Calculate the bandwidth utilization, processor utilization, memory utilization, and transmission delay in the network equipment performance data, including:

[0075] Collect historical network traffic data in the communication network within a preset sampling period, the historical network traffic data being the number of traffic bytes on the network link per unit time, and network device performance data reflecting the working status of the network link, as well as user service demand data;

[0076] Inputting the historical network traffic data into a time series decomposition algorithm to extract the fluctuation pattern of the historical network traffic data over time, thereby obtaining a trend term reflecting the long-term trend, a period term reflecting the periodic change, and a residual term reflecting the random fluctuation;

[0077] Statistical calculations are performed on the network device performance data, and sliding averages are performed on the network device performance data according to a preset time window to obtain bandwidth occupancy, processor occupancy, memory occupancy, and transmission delay.

[0078] Deploy traffic collection devices on the network, setting a sampling period of 5 minutes. The collector records the number of bytes passing through the network link at each sampling point. For example, for a 10Gbps network link, the collector records actual traffic data on the link every 5 minutes, including information such as the number of bytes and packets transmitted in both upstream and downstream directions. This raw traffic data is stored in a database in real time, forming a continuous time series.

[0079] Collect performance data from network devices. Access network devices using the SNMP protocol to obtain operational status parameters. The collector sends SNMP queries to the devices every five minutes, reading performance indicators such as CPU usage, memory usage, and port status. This performance data reflects the current operating status and load of the network link.

[0080] Collect user service demand data, including bandwidth application records, service types, and service quality requirements submitted by users. This data is used to understand the source and changing patterns of network load.

[0081] The collected historical network traffic data is processed using a time series decomposition algorithm. First, the continuous traffic data is arranged in chronological order, and the seasonal decomposition method is used to decompose the data into three components: the trend term reflects the long-term trend of traffic, such as whether it is rising or falling overall; the cycle term reflects the periodic change pattern of traffic, such as daily peaks and troughs, and weekly weekday-weekend differences; the residual term reflects random short-term fluctuations.

[0082] For network device performance data, a sliding time window is used for statistical calculations. A fixed-size time window (e.g., 30 minutes) is set and the average value of each performance indicator is calculated within the window. The window slides over time, continuously updating the calculation results.

[0083] For example, a 10Gbps backbone link is monitored. Traffic collection equipment records data every five minutes: 3.2Gbps at 8:00, 3.5Gbps at 8:05, and 3.8Gbps at 8:10. Time series analysis of 24 hours of collected data reveals a trend of 100Mbps per hour. The periodicity indicates peak traffic of 5Gbps between 9:00-11:00 and 14:00-16:00 each day, with a trough of 1Gbps. The residual indicates random fluctuations of ±500Mbps.

[0084] Performance statistics show that within a 30-minute sliding window, the link's average bandwidth utilization was 35%, the average processor utilization was 45%, the average memory utilization was 60%, and the average transmission latency was 15 milliseconds. User service demand data shows that the link primarily carries data center synchronization services (2 Gbps) and video transmission services (3 Gbps).

[0085] In this embodiment, by performing time series decomposition on historical traffic data, the changing patterns of network traffic are accurately grasped, providing an important basis for network planning and optimization, improving the efficiency of network resource utilization, and using a sliding time window method to perform statistical analysis on performance indicators. The operating status of network equipment is fully grasped, performance bottlenecks are discovered in a timely manner, and stable network operation is ensured. Combined with user business demand data, on-demand allocation and dynamic adjustment of network resources are achieved, thereby improving user experience quality and reducing network operation and maintenance costs.

[0086] In an optional embodiment,

[0087] The trend term, the period term, and the residual term are combined into an input vector, the input vector is added to the long short-term memory network, the input vector is gated by the forget gate, the input gate, and the output gate, the unit state is updated to obtain the hidden state, and the hidden state is linearly transformed to obtain the temporal encoding, including:

[0088] Sequentially concatenate the trend term, the period term, and the residual term in the time dimension to form an input vector, wherein the input vector contains historical values ​​at different time steps;

[0089] Based on the input vector and pre-acquired historical state information, randomly forgetting the historical information through a forget gate, calculating the current update information and generating a candidate unit state through an input gate, combining the retained portion of the historical information with the current update information to update the unit state, and performing a gating operation on the unit state through an output gate to obtain a hidden state;

[0090] Performing a linear transformation on the hidden state to obtain a temporal code.

[0091] Obtain trend terms, cycle terms, and residual terms. Taking the average daily temperature data for a region over a year as an example, time series decomposition can yield a trend term reflecting long-term trends, a cycle term reflecting seasonal variations, and a residual term reflecting random fluctuations.

[0092] The obtained trend, cycle, and residual terms are organized into an input vector in chronological order. A 30-day window is used, and each time step contains the trend value, cycle value, and residual value for that day. For example, the input vector for day 1 contains a trend value of 20 degrees, a cycle value of 2 degrees, and a residual value of 0.5 degrees. The input vector for day 2 contains a trend value of 19.8 degrees, a cycle value of 1.8 degrees, and a residual value of -0.3 degrees. This continues for 30 time steps.

[0093] The input vector is processed using a long short-term memory network. The network consists of three gating units: a forget gate, an input gate, and an output gate. The forget gate calculates the proportion of historical information to be forgotten based on the current input vector and the previous hidden state. For temperature prediction, for example, if the current temperature differs significantly from the historical temperature, the forget gate will reduce the degree of retention of historical information.

[0094] The input gate calculates the information that needs to be updated at the current moment. It combines the current input vector and the historical hidden state to generate a candidate cell state. The input gate determines how much new information to write to the memory cell. For example, if a sudden temperature change is detected, the input gate will increase the proportion of new information written.

[0095] The unit state is updated by combining historical information filtered by the forget gate with new information updated by the input gate. Based on the updated unit state, the output gate controls the degree of information output to obtain the current hidden state. Finally, a linear transformation is performed on the hidden state to obtain an encoding vector containing temporal features.

[0096] For example, at a certain moment, the trend term value is 3.5 Gbps, the period term value is 1.2 Gbps, and the residual term value is -0.3 Gbps. The data of the most recent 6 time steps are concatenated to form an input vector of length 18 (6 time steps × 3 components).

[0097] The input vector enters the LSTM network. The previous cell state is [0.8, 0.6, 0.4], and the hidden state is [0.7, 0.5, 0.3]. The forget gate calculates the forgetting coefficient [0.4, 0.3, 0.5], which represents the proportion of historical information retained. The input gate calculates the update information [0.6, 0.5, 0.4] and generates the candidate cell state [0.9, 0.7, 0.5]. After combining the updates, the new cell state is [0.85, 0.65, 0.45]. The output gate performs a gating operation and the hidden state is [0.75, 0.55, 0.35]. The hidden state is mapped to a two-dimensional temporal code [0.65, 0.45] through a linear transformation.

[0098] In this embodiment, multiple characteristic components of data are extracted through time series decomposition, which can more comprehensively capture the changing patterns of data and improve the accuracy of time series feature extraction. The gating mechanism of the long short-term memory network can adaptively adjust the degree of retention of historical information and the degree of updating of new information, effectively handle long-term dependencies, combine trend terms, periodic terms and residual terms to model, and perform nonlinear feature extraction through neural networks, which can learn richer time series patterns and enhance the expressive power of the model.

[0099] In an optional embodiment,

[0100] Bandwidth occupancy, processor occupancy, memory occupancy, and transmission delay are combined into a performance indicator vector. The attention mechanism is introduced to calculate the performance indicator vector to obtain the importance weight. The time series coding and the importance weight are fused to obtain the fused features including:

[0101] Collecting performance parameters of the network device at multiple consecutive time points, the performance parameters including bandwidth occupancy, processor occupancy, memory occupancy, and transmission delay, and forming a performance indicator vector from the performance parameters collected at each time point;

[0102] Based on the attention mechanism, the performance indicator vector is input into the query matrix, key matrix and value matrix to obtain the query vector, key vector and value vector. The correlation score between the performance indicators is obtained based on the dot product operation of the query vector and the key vector. The correlation score is normalized by softmax to obtain the importance weight. The importance weight represents the influence of each performance indicator on the network performance.

[0103] The temporal coding and the importance weight are weighted and superimposed to obtain a fusion feature.

[0104] Four performance parameters of network devices are collected at consecutive time points: bandwidth utilization reflects link usage, processor utilization indicates device computing load, memory utilization reflects resource consumption, and transmission latency indicates packet transmission quality. The four parameters collected at each time point are arranged in order to form a performance indicator vector, with the sampled data at each time point forming a set of elements in the vector.

[0105] An attention mechanism is used to process performance metric vectors. These performance metric vectors are fed into three transformation matrices: a query matrix for extracting the target features to be analyzed, a key matrix for extracting reference features, and a value matrix for extracting actual eigenvalues. These transformations yield query, key, and value vectors. The dot product between the query and key vectors is calculated to generate a correlation score, which represents the degree of association between different performance metrics. The correlation scores are then normalized using softmax, mapped to a range between 0 and 1, to generate importance weights. These importance weights reflect the impact of each performance metric on overall network performance.

[0106] The resulting time series code is then weighted and superimposed with the importance weights. The time series code captures the temporal characteristics of network traffic, while the importance weights reflect the impact of performance metrics. The resulting fusion features reflect both temporal variation patterns and performance influencing factors.

[0107] For example, four performance parameters are collected at a certain moment: bandwidth utilization 35%, processor utilization 45%, memory utilization 60%, and transmission delay 15ms. Data from six consecutive time points is sampled to form a 24-dimensional performance indicator vector (6 time points × 4 parameters).

[0108] The performance metric vector is fed into the attention mechanism and transformed through three matrices to yield an 8-dimensional query vector [0.4, 0.5, 0.3, 0.6, 0.4, 0.2, 0.5, 0.3], an 8-dimensional key vector [0.3, 0.6, 0.4, 0.5, 0.3, 0.4, 0.6, 0.2], and an 8-dimensional value vector [0.5, 0.4, 0.6, 0.3, 0.5, 0.3, 0.4, 0.6]. The dot product of the query and key vectors yields a relevance score of [0.8, 0.6, 0.7, 0.5]. Softmax normalization yields importance weights of [0.35, 0.20, 0.30, 0.15], indicating that bandwidth utilization and memory utilization have a significant impact on network performance.

[0109] The obtained two-dimensional temporal coding [0.65, 0.45] is weightedly superimposed with the four-dimensional importance weight [0.35, 0.20, 0.30, 0.15] to obtain the fusion feature [0.55, 0.40] that reflects the temporal characteristics and performance impact.

[0110] In this embodiment, the importance weights of performance indicators are calculated through the attention mechanism, and the key indicators with a greater impact on network performance are accurately identified, thereby improving the accuracy of network performance evaluation. Temporal coding is used to retain the characteristic information of performance indicators changing over time, so that the fused features can reflect the dynamic changes in network performance and enhance the temporal expression ability of the features. The importance weights are fused with the temporal coding to obtain fused features with stronger expression ability, which can simultaneously characterize the importance and temporal change characteristics of the performance indicators, and provide more effective feature support for subsequent network performance prediction and fault diagnosis.

[0111] In an optional embodiment,

[0112] A multi-scale prediction mechanism is used to process the fusion features to predict the short-term, medium-term and long-term network traffic values ​​respectively. Adaptive weights are set based on the prediction errors at each time scale. The prediction results of multiple time scales are weighted and combined to output the initial network traffic prediction values, including:

[0113] A two-layer long short-term memory network structure is used to construct prediction processing units for three time scales: short-term, medium-term, and long-term. The two-layer long short-term memory network structure processes the fused features and outputs the initial network traffic prediction value at each time scale.

[0114] Continuously collect network traffic data at each time scale within a preset sampling time window, calculate the rate of change of the network traffic data at each time scale, and calculate the temporal correlation between adjacent time scales and the degree of fluctuation of each time scale based on the rate of change; set a corresponding temperature coefficient for each time scale based on the degree of fluctuation, where the greater the degree of fluctuation, the smaller the temperature coefficient;

[0115] The historical samples are divided into a plurality of continuous training sequence segments according to a preset time interval. The difference between the actual value of the network traffic and the initial network traffic prediction value at the corresponding time scale is calculated in each training sequence segment. The difference is cumulatively summed and the average value is calculated to generate the prediction error at each time scale. The prediction error is multiplied by the temperature coefficient of the corresponding time scale, and the inverse is taken and substituted into the exponential function for operation to generate an exponential mapping value. The ratio of the exponential mapping value to the sum of the exponential mapping values ​​of the three time scales is calculated to determine the initial adaptive weight corresponding to each time scale.

[0116] The regular characteristics of temporal correlation between time scales are extracted from the graph structure network. Constraint relationships between prediction values ​​of adjacent time scales are established based on the regular characteristics. The initial network traffic prediction value is revised based on the constraint relationship and the federated learning framework to generate a revised prediction value that considers temporal correlation. The initial adaptive weight is adjusted based on the revised prediction value to obtain a revised weight that considers temporal correlation. The revised weight corresponding to the time scale with a larger degree of fluctuation is relatively increased.

[0117] Perform a weighted sum operation on the corrected prediction value and the corresponding corrected weight to output the final network traffic prediction value.

[0118] A two-layer LSTM network structure is constructed, and each time-scale processing unit contains two LSTM layers. The number of hidden units in the first layer of LSTM is set to 128, which is used to extract the temporal pattern in the input features; the number of hidden units in the second layer of LSTM is set to 64, focusing on sequence prediction tasks. The short-term prediction unit processes the data of the last 24 hours, the medium-term prediction unit processes the data of the last 7 days, and the long-term prediction unit processes the data of the last 4 weeks. The input of each processing unit contains the temporal encoding and performance weight information in the fused features. The LSTM network updates the unit state and hidden state at each time step through forward propagation, and the final output layer uses a fully connected layer to map the hidden state to the predicted value;

[0119] Within a 30-minute sampling window, flow data is collected every 5 minutes. The short-term scale calculates the rate of change of adjacent sampling points: (current flow - flow at the previous moment) / sampling interval; the medium-term scale calculates the rate of change of adjacent hours; the long-term scale calculates the rate of change of adjacent days. The correlation between adjacent time scales is calculated using the Pearson correlation coefficient: the data series of the two time scales are standardized and their covariance is calculated, and then divided by the product of the standard deviation. The degree of fluctuation is obtained by calculating the standard deviation of the flow series in the sliding window, and the window size is 2 hours for short-term, 2 days for medium-term, and 2 weeks for long-term. The temperature coefficient is set using the exponential decay formula: exp(-degree of fluctuation);

[0120] The 30-day historical samples are divided into fixed intervals. The short-term prediction uses a 1-hour interval to generate 720 training sequence segments; the medium-term prediction uses a 6-hour interval to generate 120 training sequence segments; the long-term prediction uses a 24-hour interval to generate 30 training sequence segments. In each sequence segment, the prediction error is accumulated: the difference between the predicted value and the true value at each time point is summed, and then divided by the length of the sequence segment to obtain the average error. The average error is multiplied by the temperature coefficient and negative, and nonlinear mapping is performed through the exp function to obtain the exponential mapping value. The exponential mapping values ​​of the three time scales are added to obtain the normalized denominator, and the respective exponential mapping values ​​are divided by the denominator to obtain the initial adaptive weight;

[0121] Construct a three-node graph network, where the nodes represent the predicted values ​​of three time scales, and the edge weights are the temporal correlation. Each node contains the current predicted value and the historical predicted sequence features. The dynamic correlation features between nodes are extracted through the graph attention layer to generate an edge feature matrix. A constraint equation is established based on the edge feature matrix: the difference in predicted values ​​of adjacent nodes should be inversely proportional to their correlation. In the federated learning framework, the prediction models of each time scale act as federated members, sharing constraint information but not directly exchanging data. By iteratively optimizing the predicted values ​​that violate the constraints, a revised predicted value is generated. The calculation of the revised weight takes into account the original weight and the degree of volatility: revised weight = original weight * (1 + normalized degree of volatility);

[0122] A weighted summation operation is performed on the revised forecast values ​​and their revised weights for the three time scales. The weights reflect the credibility of the forecast results at each time scale, while the forecast values ​​reflect the revised results after accounting for temporal correlations. This weighted summation balances the forecast deviations at different time scales, resulting in a more accurate final forecast value.

[0123] For example, the collected traffic data shows that the standard deviation of the short-term scale is 0.8 Gbps, the medium-term is 0.5 Gbps, and the long-term is 0.3 Gbps. Based on this, the temperature coefficients are set as follows: short-term 0.4, medium-term 0.6, and long-term 0.8.

[0124] The two-layer LSTM outputs initial predictions for three time scales: 4.2 Gbps in the short term, 4.0 Gbps in the medium term, and 3.8 Gbps in the long term. The actual traffic value is 4.1 Gbps, and the calculated prediction errors are: 0.1 Gbps in the short term, -0.1 Gbps in the medium term, and -0.3 Gbps in the long term.

[0125] Multiply the prediction error by the temperature coefficient, negate it, and substitute it into the exponential function to calculate: short-term exp(-0.04) = 0.96, medium-term exp(0.06) = 1.06, and long-term exp(0.24) = 1.27. Normalized, the initial adaptive weights are: short-term 0.29, medium-term 0.32, and long-term 0.39.

[0126] The graph network extracted a short-term-medium-term correlation of 0.7 and a medium-term-long-term correlation of 0.6. Based on the constraints, the revised forecasts are: 4.15 Gbps in the short term, 4.05 Gbps in the medium term, and 3.9 Gbps in the long term. The adjusted weights are: 0.35 in the short term, 0.33 in the medium term, and 0.32 in the long term.

[0127] The weighted summation yields the traffic prediction value: 4.15×0.35+4.05×0.33+3.9×0.32=4.04 Gbps.

[0128] In this embodiment, a multi-scale prediction mechanism is used to fully utilize the characteristics of network traffic data at different time scales, thereby improving the prediction accuracy and robustness. An adaptive weight distribution scheme is adopted to dynamically adjust the weights according to the prediction error and data fluctuation degree of each time scale, making the prediction results more accurate and reliable. A graph structure network and a federated learning framework are introduced to effectively extract time series correlation features and realize distributed collaborative optimization, thereby enhancing the generalization ability and practicality of the model.

[0129] In an optional embodiment,

[0130] The regular characteristics of temporal correlation between time scales are extracted from the graph structure network. Constraint relationships between prediction values ​​at adjacent time scales are established based on the regular characteristics. The initial network traffic prediction value is modified based on the constraint relationship and the federated learning framework to generate a modified prediction value that takes into account temporal correlation. The method includes:

[0131] A dynamic heterogeneous graph network structure is constructed, connecting network traffic nodes at different time scales through dynamic edges. Node attributes include traffic values, statistical characteristics, and change trends, and edge attributes are initialized to the basic temporal correlation between adjacent time scales. A spatiotemporal attention mechanism is used to dynamically update node and edge attributes, outputting the updated node state vector and edge correlation strength.

[0132] Based on the updated node state vectors and edge correlation strengths, a temporal correlation extraction model is constructed. This model captures the changing characteristics of network traffic at different time scales through a sliding time window. Furthermore, the model combines counterfactual reasoning methods to identify the dominant factors of temporal correlation and extract the regular characteristics of temporal correlation changes over time.

[0133] A conditional time series pattern graph is constructed based on the extracted regular features, and the time series association patterns under different conditions are mapped to nodes in the graph. Node connections are established based on the evolutionary relationship between the patterns. The conditional time series pattern graph is used to characterize the constraint relationship between the prediction values ​​of adjacent time scales. A distributed federated learning framework is used to integrate the constraint relationship information of multiple network regions. The calculation parameters of the time series correlation degree are adjusted in real time based on the local prediction deviation. The global prediction performance index is used to continuously optimize the extraction method of regular features, and an adaptive probabilistic graph model is constructed to dynamically associate the update frequency of the model parameters with the fluctuation degree of network traffic.

[0134] According to the node state distribution and transition probability in the adaptive probabilistic graph model, an energy function is constructed in combination with the constraint relationship provided by the conditional timing pattern graph. The constraint relationship strength, conditional timing law and probability distribution are used as components of the energy function. The weight coefficient of each component is adaptively adjusted with the change of prediction accuracy. The initial network traffic prediction value is corrected by iteratively optimizing the energy function until all timing constraints are met, and the corrected prediction value considering timing correlation is output.

[0135] A dynamic heterogeneous graph network structure is constructed, organizing network traffic data at different time scales into a graph network. The network traffic data at each time point is constructed as a node in the graph. Each node contains the actual traffic value at that time point, statistical characteristics, and trend information. Statistical characteristics are calculated using a sliding time window, including statistics such as the mean, standard deviation, skewness, and kurtosis of the data distribution within the window range. Trend information is obtained by analyzing traffic changes at adjacent time points and is used to characterize the changing state of traffic. For nodes adjacent in time, the system establishes dynamic connecting edges, and the initial attributes of the edges are determined by calculating the correlation coefficient of the traffic data at adjacent time points.

[0136] To dynamically update node and edge attributes, a spatiotemporal attention mechanism is designed. This calculates the degree of association between each node and its surrounding nodes in the spatial dimension, and uses a multi-head attention mechanism to focus on different types of spatial association features. In the temporal dimension, a temporal attention layer is introduced to capture temporal position information, enabling the model to understand data dependencies in the temporal dimension.

[0137] Based on the updated node status and edge association information, a temporal correlation extraction model is constructed. A dynamically adjusted sliding window mechanism is used to analyze the trend, periodicity, and burst characteristics of network traffic within each window. Counterfactual reasoning is used to identify key factors influencing temporal correlation. A systematic controlled experiment is conducted to observe the degree of change in temporal correlation by varying individual factors, thereby assessing the importance of each factor.

[0138] Based on the extracted regular features, a conditional time series pattern graph is constructed. Conditional dimensions, including network load level and time period, are defined, and corresponding pattern nodes are created for each condition combination. By analyzing the evolution of patterns in historical data, connections between pattern nodes are established. A distributed federated learning framework is used to integrate constraint relationship information from multiple network regions. Each region maintains a local model and regularly exchanges parameters with a central node, enabling knowledge sharing while protecting data privacy.

[0139] Constructing an adaptive probabilistic graphical model. First, a multi-dimensional state space is established, encompassing the traffic level dimension (dividing traffic into multiple load levels), the trend dimension (indicating rising, falling, or stable traffic), and the temporal attribute dimension (distinguishing between weekdays, weekends, and holidays). The model's core components include a state transition module, an observation mapping module, and an adaptive adjustment module.

[0140] The state transition module is responsible for learning and updating the transition patterns between states. It establishes initial state transition relationships by analyzing the changing patterns of state sequences in historical data. As new data is collected, the system dynamically updates these transition relationships, taking into account the importance of both new and old data to ensure the model can adapt to changes in the network environment.

[0141] The observation mapping module establishes a correspondence between states and actual flow values, statistically analyzes the distribution of possible flow values ​​under each state, and forms an observation probability map. The mapping relationship is continuously optimized as new data accumulates, making the prediction results more accurate.

[0142] The adaptive regulation module is responsible for dynamic adjustments at three levels: first, the adjustment of parameter update frequency. The system dynamically changes the update frequency of model parameters according to the degree of traffic fluctuation; second, the adjustment of state transition probability. The transition probability is updated through the weighted fusion of new and old data; third, the optimization of the state space structure. The granularity of state division is regularly evaluated and adjusted to merge low-usage states or subdivide high-frequency states.

[0143] An energy function is constructed that integrates multiple pieces of information, encompassing constraint strength, conditional temporal patterns, and probability distribution. Weight coefficients are adaptively adjusted based on changes in prediction accuracy. The initial prediction is modified by iteratively optimizing the energy function using a gradient descent algorithm until all temporal constraints are met. A revised prediction that takes temporal correlation into account is then output. During the optimization process, prediction accuracy metrics are continuously monitored, and model parameters are adjusted based on performance feedback.

[0144] For example, at a certain moment, the initial attributes of the three types of nodes are: short-term node [4.2 Gbps, (4.0, 0.3, 4.5), (0.1, 0.8)] represents the current traffic of 4.2 Gbps, mean 4.0, variance 0.3, peak 4.5, growth rate 0.1, and cycle intensity 0.8; medium-term node [4.0 Gbps, (3.8, 0.2, 4.2), (0.05, 0.6)]; long-term node [3.8 Gbps, (3.6, 0.1, 4.0), (0.02, 0.4)]. The initial edge correlation is 0.7 for short-term to medium-term and 0.6 for medium-term to long-term. After being updated by the attention mechanism, the node state vector becomes: short-term [4.15, 4.0, 0.12], medium-term [4.05, 3.9, 0.06], long-term [3.85, 3.7, 0.03], and the edge association strength is updated to 0.75 and 0.65.

[0145] Within a 6-hour sliding window, the short-term traffic amplitude was calculated to be 1.2 Gbps, with a frequency of 0.2 times / hour and a peak at 10:00 AM; the medium-term amplitude was 0.8 Gbps, with a frequency of 0.1 times / hour and a peak at 11:00 AM; and the long-term amplitude was 0.5 Gbps, with a frequency of 0.05 times / hour and a peak at 12:00 PM. Counterfactual analysis showed that a short-term traffic increase of 1 Gbps resulted in a medium-term change of 0.6 Gbps, while a medium-term increase of 1 Gbps resulted in a long-term change of 0.4 Gbps. Feature importance scores: amplitude 0.8, frequency 0.6, and phase 0.4.

[0146] Conditional time series modes include: peak mode (trigger condition: short-term traffic flow >4.5 Gbps, correlation strength 0.85, lasting 2 hours), stable mode (3.5-4.5 Gbps, strength 0.7, lasting 6 hours), and trough mode (<3.5 Gbps, strength 0.6, lasting 4 hours). Mode transition probabilities: peak to stable 0.8, stable to trough 0.6, trough to stable 0.7. When the current traffic standard deviation is 0.8 Gbps, the model updates parameters every 10 minutes.

[0147] During the energy function optimization process, the initial weight coefficients were all around 0.33. The current constraint value was -0.65, the probability of the regularity term was 0.8, and the density of the probability term was 0.7. After 50 iterations of optimization, the revised predictions were: 4.12 Gbps in the short term, 4.08 Gbps in the medium term, and 3.95 Gbps in the long term. The weights were updated to: 0.35 for the constraint term, 0.35 for the regularity term, and 0.30 for the probability term. The final energy value dropped to -0.2, meeting the convergence criteria.

[0148] In this embodiment, by introducing a dynamic heterogeneous graph network structure and a spatiotemporal attention mechanism, the modeling capability of multi-scale temporal correlations is significantly improved. The use of counterfactual reasoning and conditional temporal pattern graphs enhances the understanding of the changing patterns of network traffic. Based on the design of a federated learning framework and an adaptive probabilistic model, the model's ability to perceive and respond to changes in network status is improved. Through the optimization mechanism of a unified energy function, precise correction and adaptive adjustment of prediction results are achieved.

[0149] In existing technologies, network traffic forecasting typically uses a single-time-scale prediction model or simply weights and combines prediction results from multiple time scales. This ignores the complex temporal correlations between different time scales, making it difficult to accurately capture the dynamic evolution characteristics of network traffic at multiple time scales. Existing technologies often model temporal correlations based on static correlation coefficients, which cannot adapt to changes in correlation strength caused by network traffic fluctuations. This results in a significant decrease in the accuracy of prediction results during periods of significant fluctuations. Furthermore, existing methods often use a fixed weighting system to fuse prediction results, lacking the ability to adapt to changes in network status.

[0150] This embodiment constructs a heterogeneous graph structure containing nodes at multiple time scales, combines the spatiotemporal attention mechanism to realize the dynamic update of node attributes and edge correlation strength, effectively improves the modeling ability of network traffic time series characteristics, adopts the counterfactual reasoning method to analyze the dominant factors of time series correlation, and extracts correlation rules, so that the prediction model can better understand the mutual influence mechanism between different time scales, constructs a conditional time series pattern graph, and uses a probabilistic graph model to characterize the constraint relationship between the prediction values ​​of adjacent time scales. The update frequency of the model parameters is dynamically associated with the degree of traffic fluctuation, ensuring the ability to respond quickly in periods of severe fluctuations. In summary, while maintaining the prediction accuracy, this embodiment significantly improves the model's adaptability to severe fluctuations in network traffic, enhances the temporal consistency of the prediction results, better meets the requirements for traffic prediction accuracy and stability in actual network environments, and provides a more reliable decision-making basis for network resource scheduling and performance optimization.

[0151] Figure 2 This is a comparison chart of prediction accuracy under different load conditions corresponding to the communication network resource optimization method based on artificial intelligence according to an embodiment of the present invention. Figure 2The figure shows a comparison of the prediction accuracy of three different prediction methods under different network load conditions. As the network load changes from low to high, the prediction accuracy of all methods shows a downward trend, but this technical solution maintains the highest accuracy under various load conditions. Under low load (30%) conditions, the accuracy of traditional neural networks is 82.3%, the graph neural network (GNN) model reaches 86.8%, and this technical solution is as high as 91.2%, which is 4.4 percentage points higher than the graph neural network model and 8.9 percentage points higher than the traditional neural network. Under medium load (60%) conditions, the accuracy of the three methods is 77.4% (traditional neural network), 83.2% (graph neural network model) and 88.7% (this technical solution). Under high load (80%) conditions, the accuracy further decreases to 71.2% (traditional neural network), 79.6% (graph neural network model) and 86.3% (this technical solution). The most challenging condition was peak load (95%), where the accuracy of traditional neural networks was only 65.1% and that of graph neural network models was 74.1%, while our proposed solution maintained a high level of 82.9%. Notably, our proposed solution experienced minimal performance degradation as load increased, with the accuracy dropping by only 8.3 percentage points from low to peak load, compared to 17.2 and 12.7 percentage points for traditional neural network and graph neural network models, respectively. This demonstrates our proposed solution's greater stability and robustness under high load conditions.

[0152] In an optional embodiment,

[0153] Based on the statistical characteristics of historical prediction errors, a correction model is constructed. The initial network traffic forecast value is dynamically calibrated using the exponential smoothing method. The forecast value is corrected based on the mutation detection results of network performance indicators. The network traffic forecast results include:

[0154] Obtaining a historical prediction error sequence, the historical prediction error sequence being obtained by subtracting a predicted network traffic value from an actual network traffic value, calculating an error mean and an error standard deviation based on the historical prediction error sequence, constructing an autocorrelation function of the error sequence, and calculating an exponentially weighted moving variance, wherein the exponentially weighted moving variance is determined by a weighted sum of a squared difference between the error and the mean at a current moment and an exponentially weighted moving variance at a previous moment;

[0155] A correction model is constructed based on the error mean, the error standard deviation, the autocorrelation function, and the exponentially weighted moving variance, wherein the correction model includes an error distribution correction term, a time series correlation correction term, and a fluctuation correction term, wherein the error distribution correction term, the time series correlation correction term, and the fluctuation correction term respectively correspond to different adaptive weight coefficients, and the adaptive weight coefficients are updated by a gradient descent method;

[0156] Based on the exponential smoothing method, a historical information storage matrix is ​​constructed, and the initial network traffic forecast value is dynamically calibrated using an adaptive smoothing factor and an attention mechanism to obtain a first revised forecast value. The adaptive smoothing factor is determined by the weighted sum of the smoothing factor at the previous moment and the absolute value of the historical forecast error at the current moment, and the weighting coefficient is dynamically adjusted by the change in the exponentially weighted moving variance.

[0157] Constructing a multidimensional network performance indicator vector, the multidimensional network performance indicator vector including values ​​of multiple network performance indicators at a current moment, performing a cumulative sum operation based on time series differences of the multidimensional network performance indicator vector to obtain a mutation detection value, and comparing the mutation detection value with the difference between the mutation detection value at a previous moment and a detection threshold;

[0158] When the mutation detection value is greater than the mutation determination threshold, the relative change rate of each network performance indicator in the multidimensional network performance indicator vector is calculated, the relative change rate of each network performance indicator is multiplied by the corresponding weight coefficient and the sum is obtained to obtain the mutation degree, and the product of the first corrected prediction value and the mutation degree is output as the network traffic prediction result; when the mutation detection value is not greater than the mutation determination threshold, the first corrected prediction value is output as the network traffic prediction result.

[0159] The historical forecast error sequence is obtained by calculating the difference between actual and predicted network traffic values. This error sequence is statistically analyzed, with the mean error calculated as a measure of system forecast deviation and the standard deviation of the error calculated to characterize forecast volatility. An autocorrelation function of the error sequence is constructed to analyze the correlation between error values ​​at different time intervals, capturing the temporal dependence of the forecast error. An exponentially weighted moving variance is calculated, combining the squared difference between the current error and the mean with the exponentially weighted moving variance from the previous moment. The weighting coefficient decays over time to highlight the impact of recent error fluctuations.

[0160] A three-term correction model is constructed based on the statistical characteristics of the error. The error distribution correction term uses the error mean and standard deviation to characterize the distribution characteristics of the forecast error; the time series correlation correction term models the temporal dependence of the error sequence based on the autocorrelation function; and the fluctuation correction term reflects the dynamic fluctuation characteristics of the forecast error through the exponentially weighted moving variance. Initial weight coefficients are assigned to the three correction terms, and the weight coefficients are dynamically updated based on the degree of improvement in forecast performance using the gradient descent method, so that the correction effect can be adaptively adjusted as the forecast error characteristics change.

[0161] A historical information storage matrix is ​​constructed to store past forecast data, and exponential smoothing is applied to calibrate the initial forecast. An adaptive smoothing factor is determined by a weighted combination of the previous smoothing factor and the absolute value of the current historical forecast error. The weighting coefficient is dynamically adjusted based on the exponentially weighted moving variance, increasing the weight of the current error as error fluctuations intensify. An attention mechanism is used to calculate the impact of historical data on the current forecast, and the initial forecast is dynamically calibrated to obtain the first revised forecast.

[0162] Construct a multidimensional indicator vector containing the current values ​​of multiple network performance indicators. Calculate the time-series differences of the indicator vector and perform cumulative summation to obtain a mutation detection value. This detection value is compared with the difference between the previous detection value and a preset threshold to determine whether a mutation has occurred. When the detection value exceeds the mutation threshold, calculate the relative rate of change of each performance indicator. Multiply the rate of change by the corresponding weight coefficient and sum the results to obtain the degree of mutation. This degree of mutation is used to adjust the first revised prediction value, and the final prediction result is output. If the detection value does not exceed the threshold, the first revised prediction value is directly output as the final prediction result.

[0163] For example, the historical forecast error series calculated for the last 24 hours showed a mean error of 0.2 Gbps and a standard deviation of 0.5 Gbps. The autocorrelation function showed correlation coefficients of 0.6 for one-hour intervals, 0.4 for two-hour intervals, and 0.2 for three-hour intervals. The squared difference between the error and the mean at the current moment was 0.09, the exponentially weighted moving variance at the previous moment was 0.16, and the weighting coefficient was 0.3, resulting in a calculated current exponentially weighted moving variance of 0.14.

[0164] The initial weights of the three correction models are all 0.33. After gradient descent, the weights are: 0.35 for the error distribution correction, 0.4 for the time series correlation correction, and 0.25 for the volatility correction. The historical information storage matrix stores the forecast data for the last 12 hours. The previous smoothing factor was 0.7, the current absolute error value was 0.3, and the change in exponentially weighted moving variance was -0.02. Based on this, the current smoothing factor is determined to be 0.65.

[0165] The attention mechanism calculates the influence weight distribution of historical data, calibrating the initial prediction of 4.5 Gbps to obtain a first revised prediction of 4.3 Gbps. The multidimensional performance indicator vector includes: bandwidth utilization of 75%, processor utilization of 60%, memory utilization of 50%, and transmission latency of 25 ms. The calculated mutation detection value is 85, which exceeds the mutation threshold of 80. The relative change rates of each indicator are 0.2, 0.15, 0.1, and 0.25, respectively, with corresponding weights of 0.3, 0.2, 0.2, and 0.3, resulting in a calculated mutation degree of 1.2. Finally, the first revised prediction of 4.3 Gbps is multiplied by the mutation degree of 1.2 to output a prediction of 5.16 Gbps.

[0166] In this embodiment, by comprehensively analyzing the historical prediction error sequence, comprehensively considering the distribution characteristics, time series correlation and dynamic fluctuation characteristics of the errors, the exponential smoothing method and the attention mechanism are used to calibrate the prediction value. By dynamically updating the adaptive smoothing factor, the influence of historical information can be adjusted in real time according to the fluctuation of the prediction error, thereby improving the response speed of the model to changes in network status and enhancing the timeliness of the prediction results. The mutation degree is calculated based on the relative change rate of each performance indicator, and the prediction results are adjusted in a targeted manner. This improves the prediction accuracy of the model during periods of drastic changes in network status, realizes accurate correction of network traffic prediction results, and provides an effective solution for improving the accuracy and reliability of network traffic prediction.

[0167] In an optional embodiment,

[0168] According to the exponential smoothing method, a historical information storage matrix is ​​constructed, and the initial network traffic forecast value is dynamically calibrated through the adaptive smoothing factor and attention mechanism to obtain the first revised forecast value, which includes:

[0169] The network traffic time series is divided into multiple time scale layers according to the time scale, and an independent adaptive smoothing factor is constructed for each time scale layer.

[0170] Calculating the volatility according to the changing trend of the network traffic time series, dynamically adjusting the size of the historical data window based on the volatility, wherein the size of the historical data window is determined by multiplying the basic window size by the volatility, and obtaining smoothing results of each time scale layer within the historical data window;

[0171] The smoothed result is used to construct a feature matrix according to the time scale layer, and attention weights are calculated based on the feature matrix and the query vector at the current moment. The attention weights represent the importance of different time scale features to the current prediction, and the features of different time scales are adaptively fused using the attention weights.

[0172] Constructing a historical information storage matrix, the historical information storage matrix including historical network traffic values, the prediction error, and the adaptive smoothing factor, employing an attention mechanism to extract historical information related to a current state from the historical information storage matrix, and determining a dynamic learning rate based on a change in the prediction error, wherein the dynamic learning rate decays as the change in the prediction error increases;

[0173] A residual connection structure is introduced into the initial network traffic prediction value, and the correction function output of each time scale and the adaptive smoothing factor are weighted and summed according to the attention weight to obtain a residual term, and the residual term is added to the initial network traffic prediction value to obtain a first corrected prediction value.

[0174] Network traffic time series are divided into three time scales: hourly, daily, and weekly. Each time scale layer independently configures an adaptive smoothing factor to smooth the historical data at that scale. The initial value of the smoothing factor is set based on the span of the time scale. Larger spans result in smaller initial smoothing factors, reflecting the degree of dependence of different time scales on historical data.

[0175] The volatility of network traffic time series is calculated as the ratio of the change in traffic values ​​between adjacent time points to the mean. Based on the calculated volatility, the size of the historical data window is dynamically adjusted, with the window size being equal to the product of the pre-set base window size and the current volatility. When traffic fluctuations are large, the window size is increased to capture more historical information; when fluctuations are small, the window size is decreased to emphasize the influence of recent data. Within the specified window range, the historical data is smoothed using adaptive smoothing factors at each time scale layer, yielding smoothed results at different time scales.

[0176] The smoothed results at each time scale are organized into a feature matrix, with rows representing different time scales and columns representing time steps. A query vector is constructed for the current moment, containing the current traffic value and recent trend information. Attention weights are calculated based on the query vector and the feature matrix. Using a scaled dot-product attention mechanism, the similarity between the query vector and each time scale feature in the feature matrix is ​​calculated and normalized using a softmax function to obtain the attention weights. The weights reflect the importance of features at different time scales to the current prediction.

[0177] A historical information storage matrix is ​​constructed to store historical traffic values, prediction errors, and adaptive smoothing factors. An attention mechanism is used to calculate the similarity between the current state and the historical state, extracting relevant historical information. A dynamic learning rate is determined based on the change in prediction error. When the change in prediction error increases, the learning rate is reduced to minimize the correction and improve prediction stability.

[0178] A residual connection structure is introduced to correct the initial prediction value. First, the correction function output is calculated based on the characteristics of each time scale. The corrected output is multiplied by the corresponding adaptive smoothing factor, and then weighted summed with the attention weight to obtain the residual term. Finally, the residual term is added to the initial prediction value to obtain a corrected prediction value that takes into account the characteristics of multiple time scales.

[0179] For example, the network traffic time series is divided into three scales: hourly, daily, and weekly. The initial smoothing factors are set to 0.8, 0.6, and 0.4, respectively. The current volatility is calculated to be 1.5, and the base window size is set to 24 hours, resulting in an actual window size of 36 hours. Smoothing the three time scales within the 36-hour window yields smoothed results of [4.2, 4.0, 3.8, ...] Gbps at the hourly level, [4.1, 3.9, 3.7, ...] Gbps at the daily level, and [4.0, 3.8, 3.6, ...] Gbps at the weekly level.

[0180] Construct a feature matrix of size 3×36, representing the smoothed values ​​of the three time scales over 36 time steps. The current query vector is [4.5, 0.2, 0.1], indicating a current traffic flow of 4.5 Gbps, a recent growth rate of 0.2, and an acceleration of 0.1. Calculate the attention weights: 0.5 for the hourly level, 0.3 for the daily level, and 0.2 for the weekly level.

[0181] The historical information storage matrix records the last 72 hours of data, including traffic values, prediction errors, and smoothing factors at each point in time. The current prediction error change is 0.3 Gbps, which results in a dynamic learning rate of 0.6. Historical information with a high degree of similarity to the current state is extracted for reference.

[0182] The initial prediction is 4.8 Gbps, and the correction function outputs for each time scale are -0.3, -0.2, and -0.1, respectively. The correction outputs are multiplied by the corresponding smoothing factors: -0.24, -0.12, and -0.04, and then weighted summed with the attention weight to obtain a residual term of -0.16. Finally, the residual term is added to the initial prediction to obtain a corrected prediction of 4.64 Gbps.

[0183] In this embodiment, through multi-layer division of time series and configuration of independent smoothing factors, differentiated processing of features at different time scales is achieved. The historical data window dynamically adjusted based on volatility enables the data sampling range to be adaptively scaled according to traffic changes, improving the pertinence and effectiveness of historical data utilization. An attention mechanism is used to adaptively fuse features at multiple time scales. By calculating the correlation between features at different scales and the current state, dynamic evaluation and selective fusion of feature importance are achieved. The introduction of a historical information storage matrix and a dynamic learning rate mechanism enables the correction process to fully utilize historical experience and flexibly adjust the correction strength according to changes in prediction error.

[0184] Existing network traffic prediction and correction methods typically use a fixed time window and a uniform smoothing factor to process historical data. This method is unable to adapt to the dynamic changes in network traffic at different time scales. Furthermore, they often simply superimpose or average features at different time scales, ignoring the differences in the importance of features at each time scale to the current prediction. Furthermore, the correction parameters in existing technologies are usually fixed and lack the ability to adaptively adjust to changes in network status, resulting in poor correction results during periods of drastic traffic fluctuations.

[0185] This embodiment enables the correction process to better adapt to the changing characteristics of network traffic through dynamic fusion of multi-scale features and adaptive parameter adjustment. The introduced dynamic window mechanism and residual connection structure effectively improve the responsiveness to abnormal fluctuations and the correction stability. While ensuring the correction effect, it significantly enhances the adaptability and robustness of the correction process, can better meet the requirements for traffic prediction accuracy in actual network environments, and provide more reliable technical support for the efficient scheduling and optimized management of network resources.

[0186] Figure 3 This is a graph of prediction error data under different fluctuation scenarios corresponding to the communication network resource optimization method based on artificial intelligence in an embodiment of the present invention, such as Figure 3 The figure shows a comparison of the mean absolute percentage errors (MAPEs) of various prediction methods under different volatility scenarios. The data clearly demonstrates that our proposed method achieves the lowest prediction error across all volatility scenarios. In low-volatility scenarios (volatility < 0.5), our proposed method achieves a MAPE of 3.42%, a 41.7% reduction from the 5.87% of the traditional exponential smoothing method and a 30.5% reduction from the 4.92% of the SARIMA model. As traffic volatility increases, our proposed method's advantages become even more pronounced. In high-volatility scenarios (1.5 ≤ volatility < 3.0), our proposed method achieves a MAPE of 8.76%, a 31.9% reduction from the 12.87% of the GRU network and a 23.3% reduction from the 11.42% of the TCN model. In extreme volatility scenarios (volatility ≥ 3.0), our proposed method achieves a MAPE of 17.52%, a 46.6% reduction from the 32.78% of the exponential smoothing method and a 24.1% reduction from the 23.08% of the closest TCN model. The average mean absolute percentage error (MAPE) of this solution is 8.64%, a 22.0% reduction compared to the best-performing TCN model (11.08%). This demonstrates the effectiveness of the adaptive smoothing factor and dynamic window adjustment mechanism in this solution in addressing diverse volatility scenarios, particularly in high and extreme volatility scenarios, where it offers significant predictive stability advantages.

[0187] The AI-based communication network resource optimization system includes:

[0188] The first unit is used to collect historical network traffic data, network equipment performance data, and user service demand data in the communication network, decompose the historical network traffic data into trend items, period items, and residual items using a time series decomposition algorithm, and calculate bandwidth utilization, processor utilization, memory utilization, and transmission delay in the network equipment performance data;

[0189] The second unit is configured to combine the trend term, the period term, and the residual term into an input vector, add the input vector to the long short-term memory network, perform a gating operation on the input vector through a forget gate, an input gate, and an output gate, update the unit state to obtain a hidden state, and perform a linear transformation on the hidden state to obtain a temporal code;

[0190] The third unit is used to combine bandwidth occupancy, processor occupancy, memory occupancy, and transmission delay into a performance indicator vector. An attention mechanism is introduced to calculate the importance weight of the performance indicator vector. The time series coding and the importance weight are then fused to obtain a fused feature.

[0191] The fourth unit is used to process the fused features using a multi-scale prediction mechanism to predict the short-term, medium-term, and long-term network traffic values ​​respectively. Adaptive weights are set based on the prediction errors at each time scale, and the prediction results at multiple time scales are weighted and combined to output the initial network traffic prediction value.

[0192] The fifth unit is used to build a correction model based on the statistical characteristics of historical prediction errors, dynamically calibrate the initial network traffic prediction value using the exponential smoothing method, and correct the prediction value in combination with the mutation detection results of the network performance indicators to obtain the network traffic prediction result.

[0193] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0194] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A communication network resource optimization method based on artificial intelligence, characterized in that: include: Collect historical network traffic data, network equipment performance data, and user service demand data from the communication network. Decompose the historical network traffic data into trend items, period items, and residual items using a time series decomposition algorithm. Calculate bandwidth utilization, processor utilization, memory utilization, and transmission delay from the network equipment performance data. The trend term, the period term, and the residual term are combined into an input vector, the input vector is added to the long short-term memory network, the input vector is gated by a forget gate, an input gate, and an output gate, the unit state is updated to obtain a hidden state, and the hidden state is linearly transformed to obtain a temporal code; Bandwidth occupancy, processor occupancy, memory occupancy, and transmission delay are combined into a performance indicator vector. An attention mechanism is introduced to calculate the performance indicator vector to obtain the importance weight. The temporal coding and the importance weight are then fused to obtain the fused feature. A multi-scale prediction mechanism is used to process the fused features to predict the short-term, medium-term, and long-term network traffic values ​​respectively. Adaptive weights are set based on the prediction errors at each time scale, and the prediction results at multiple time scales are weighted and combined to output the initial network traffic prediction value. A correction model is constructed based on the statistical characteristics of historical prediction errors. The initial network traffic prediction value is dynamically calibrated using the exponential smoothing method. The prediction value is corrected in combination with the mutation detection results of network performance indicators to obtain the network traffic prediction result.

2. The method according to claim 1, characterized in that Collect historical network traffic data, network equipment performance data, and user service demand data from the communication network. Decompose the historical network traffic data into trend items, period items, and residual items using a time series decomposition algorithm. Calculate the bandwidth utilization, processor utilization, memory utilization, and transmission delay in the network equipment performance data, including: Collect historical network traffic data in the communication network within a preset sampling period, the historical network traffic data being the number of traffic bytes on the network link per unit time, and network device performance data reflecting the working status of the network link, as well as user service demand data; Inputting the historical network traffic data into a time series decomposition algorithm to extract the fluctuation pattern of the historical network traffic data over time, thereby obtaining a trend term reflecting the long-term trend, a period term reflecting the periodic change, and a residual term reflecting the random fluctuation; Statistical calculations are performed on the network device performance data, and sliding averages are performed on the network device performance data according to a preset time window to obtain bandwidth occupancy, processor occupancy, memory occupancy, and transmission delay.

3. The method according to claim 1, characterized in that The trend term, the period term, and the residual term are combined into an input vector, the input vector is added to the long short-term memory network, the input vector is gated by the forget gate, the input gate, and the output gate, the unit state is updated to obtain the hidden state, and the hidden state is linearly transformed to obtain the temporal encoding, including: Sequentially concatenate the trend term, the period term, and the residual term in the time dimension to form an input vector, wherein the input vector contains historical values ​​at different time steps; Based on the input vector and pre-acquired historical state information, randomly forgetting the historical information through a forget gate, calculating the current update information and generating a candidate unit state through an input gate, combining the retained portion of the historical information with the current update information to update the unit state, and performing a gating operation on the unit state through an output gate to obtain a hidden state; Performing a linear transformation on the hidden state to obtain a temporal code.

4. The method according to claim 1, wherein Bandwidth occupancy, processor occupancy, memory occupancy, and transmission delay are combined into a performance indicator vector. The attention mechanism is introduced to calculate the performance indicator vector to obtain the importance weight. The time series coding and the importance weight are fused to obtain the fused features including: Collecting performance parameters of the network device at multiple consecutive time points, the performance parameters including bandwidth occupancy, processor occupancy, memory occupancy, and transmission delay, and forming a performance indicator vector from the performance parameters collected at each time point; Based on the attention mechanism, the performance indicator vector is input into the query matrix, key matrix and value matrix to obtain the query vector, key vector and value vector. The correlation score between the performance indicators is obtained based on the dot product operation of the query vector and the key vector. The correlation score is normalized by softmax to obtain the importance weight. The importance weight represents the influence of each performance indicator on the network performance. The temporal coding and the importance weight are weighted and superimposed to obtain a fusion feature.

5. The method according to claim 1, wherein A multi-scale prediction mechanism is used to process the fusion features to predict the short-term, medium-term and long-term network traffic values ​​respectively. Adaptive weights are set based on the prediction errors at each time scale. The prediction results of multiple time scales are weighted and combined to output the initial network traffic prediction values, including: A two-layer long short-term memory network structure is used to construct prediction processing units for three time scales: short-term, medium-term, and long-term. The two-layer long short-term memory network structure processes the fused features and outputs the initial network traffic prediction value at each time scale. Continuously collect network traffic data at each time scale within a preset sampling time window, calculate the rate of change of the network traffic data at each time scale, and calculate the temporal correlation between adjacent time scales and the degree of fluctuation of each time scale based on the rate of change; set a corresponding temperature coefficient for each time scale based on the degree of fluctuation, where the greater the degree of fluctuation, the smaller the temperature coefficient; The historical samples are divided into a plurality of continuous training sequence segments according to a preset time interval. The difference between the actual value of the network traffic and the initial network traffic prediction value at the corresponding time scale is calculated in each training sequence segment. The difference is cumulatively summed and the average value is calculated to generate the prediction error at each time scale. The prediction error is multiplied by the temperature coefficient of the corresponding time scale, and the inverse is taken and substituted into the exponential function for operation to generate an exponential mapping value. The ratio of the exponential mapping value to the sum of the exponential mapping values ​​of the three time scales is calculated to determine the initial adaptive weight corresponding to each time scale. The regular characteristics of temporal correlation between time scales are extracted from the graph structure network. Constraint relationships between prediction values ​​of adjacent time scales are established based on the regular characteristics. The initial network traffic prediction value is revised based on the constraint relationship and the federated learning framework to generate a revised prediction value that considers temporal correlation. The initial adaptive weight is adjusted based on the revised prediction value to obtain a revised weight that considers temporal correlation. The revised weight corresponding to the time scale with a larger degree of fluctuation is relatively increased. Perform a weighted sum operation on the corrected prediction value and the corresponding corrected weight to output the final network traffic prediction value.

6. The method according to claim 5, characterized in that The regular characteristics of temporal correlation between time scales are extracted from the graph structure network. Constraint relationships between prediction values ​​at adjacent time scales are established based on the regular characteristics. The initial network traffic prediction value is modified based on the constraint relationship and the federated learning framework to generate a modified prediction value that takes into account temporal correlation. The method includes: A dynamic heterogeneous graph network structure is constructed, connecting network traffic nodes at different time scales through dynamic edges. Node attributes include traffic values, statistical characteristics, and change trends, and edge attributes are initialized to the basic temporal correlation between adjacent time scales. A spatiotemporal attention mechanism is used to dynamically update node and edge attributes, outputting the updated node state vector and edge correlation strength. Based on the updated node state vectors and edge correlation strengths, a temporal correlation extraction model is constructed. This model captures the changing characteristics of network traffic at different time scales through a sliding time window. Furthermore, the model combines counterfactual reasoning methods to identify the dominant factors of temporal correlation and extract the regular characteristics of temporal correlation changes over time. A conditional time series pattern graph is constructed based on the extracted regular features, and the time series association patterns under different conditions are mapped to nodes in the graph. Node connections are established based on the evolutionary relationship between the patterns. The conditional time series pattern graph is used to characterize the constraint relationship between the prediction values ​​of adjacent time scales. A distributed federated learning framework is used to integrate the constraint relationship information of multiple network regions. The calculation parameters of the time series correlation degree are adjusted in real time based on the local prediction deviation. The global prediction performance index is used to continuously optimize the extraction method of regular features, and an adaptive probabilistic graph model is constructed to dynamically associate the update frequency of the model parameters with the fluctuation degree of network traffic. According to the node state distribution and transition probability in the adaptive probabilistic graph model, an energy function is constructed in combination with the constraint relationship provided by the conditional timing pattern graph. The constraint relationship strength, conditional timing law and probability distribution are used as components of the energy function. The weight coefficient of each component is adaptively adjusted with the change of prediction accuracy. The initial network traffic prediction value is corrected by iteratively optimizing the energy function until all timing constraints are met, and the corrected prediction value considering timing correlation is output.

7. The method according to claim 1, characterized in that Based on the statistical characteristics of historical prediction errors, a correction model is constructed. The initial network traffic forecast value is dynamically calibrated using the exponential smoothing method. The forecast value is corrected based on the mutation detection results of network performance indicators. The network traffic forecast results include: Obtaining a historical prediction error sequence, the historical prediction error sequence being obtained by subtracting a predicted network traffic value from an actual network traffic value, calculating an error mean and an error standard deviation based on the historical prediction error sequence, constructing an autocorrelation function of the error sequence, and calculating an exponentially weighted moving variance, wherein the exponentially weighted moving variance is determined by a weighted sum of a squared difference between the error and the mean at a current moment and an exponentially weighted moving variance at a previous moment; A correction model is constructed based on the error mean, the error standard deviation, the autocorrelation function, and the exponentially weighted moving variance, wherein the correction model includes an error distribution correction term, a time series correlation correction term, and a fluctuation correction term, wherein the error distribution correction term, the time series correlation correction term, and the fluctuation correction term respectively correspond to different adaptive weight coefficients, and the adaptive weight coefficients are updated by a gradient descent method; Based on the exponential smoothing method, a historical information storage matrix is ​​constructed, and the initial network traffic forecast value is dynamically calibrated using an adaptive smoothing factor and an attention mechanism to obtain a first revised forecast value. The adaptive smoothing factor is determined by the weighted sum of the smoothing factor at the previous moment and the absolute value of the historical forecast error at the current moment, and the weighting coefficient is dynamically adjusted by the change in the exponentially weighted moving variance. Constructing a multidimensional network performance indicator vector, the multidimensional network performance indicator vector including values ​​of multiple network performance indicators at a current moment, performing a cumulative sum operation based on time series differences of the multidimensional network performance indicator vector to obtain a mutation detection value, and comparing the mutation detection value with the difference between the mutation detection value at a previous moment and a detection threshold; When the mutation detection value is greater than the mutation determination threshold, the relative change rate of each network performance indicator in the multidimensional network performance indicator vector is calculated, the relative change rate of each network performance indicator is multiplied by the corresponding weight coefficient and the sum is obtained to obtain the mutation degree, and the product of the first corrected prediction value and the mutation degree is output as the network traffic prediction result; when the mutation detection value is not greater than the mutation determination threshold, the first corrected prediction value is output as the network traffic prediction result.

8. The method according to claim 7, characterized in that According to the exponential smoothing method, a historical information storage matrix is ​​constructed, and the initial network traffic forecast value is dynamically calibrated through the adaptive smoothing factor and attention mechanism to obtain the first revised forecast value, which includes: The network traffic time series is divided into multiple time scale layers according to the time scale, and an independent adaptive smoothing factor is constructed for each time scale layer. Calculating the volatility according to the changing trend of the network traffic time series, dynamically adjusting the size of the historical data window based on the volatility, wherein the size of the historical data window is determined by multiplying the basic window size by the volatility, and obtaining smoothing results of each time scale layer within the historical data window; The smoothed result is used to construct a feature matrix according to the time scale layer, and attention weights are calculated based on the feature matrix and the query vector at the current moment. The attention weights represent the importance of different time scale features to the current prediction, and the features of different time scales are adaptively fused using the attention weights. Constructing a historical information storage matrix, the historical information storage matrix including historical network traffic values, the prediction error, and the adaptive smoothing factor, employing an attention mechanism to extract historical information related to a current state from the historical information storage matrix, and determining a dynamic learning rate based on a change in the prediction error, wherein the dynamic learning rate decays as the change in the prediction error increases; A residual connection structure is introduced into the initial network traffic prediction value, and the correction function output of each time scale and the adaptive smoothing factor are weighted and summed according to the attention weight to obtain a residual term, and the residual term is added to the initial network traffic prediction value to obtain a first corrected prediction value.

9. A communication network resource optimization system based on artificial intelligence, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to collect historical network traffic data, network equipment performance data, and user service demand data in the communication network, decompose the historical network traffic data into trend items, period items, and residual items using a time series decomposition algorithm, and calculate bandwidth utilization, processor utilization, memory utilization, and transmission delay in the network equipment performance data; The second unit is configured to combine the trend term, the period term, and the residual term into an input vector, add the input vector to the long short-term memory network, perform a gating operation on the input vector through a forget gate, an input gate, and an output gate, update the unit state to obtain a hidden state, and perform a linear transformation on the hidden state to obtain a temporal code; The third unit is used to combine bandwidth occupancy, processor occupancy, memory occupancy, and transmission delay into a performance indicator vector. An attention mechanism is introduced to calculate the importance weight of the performance indicator vector. The time series coding and the importance weight are then fused to obtain a fused feature. The fourth unit is used to process the fused features using a multi-scale prediction mechanism to predict the short-term, medium-term, and long-term network traffic values ​​respectively. Adaptive weights are set based on the prediction errors at each time scale, and the prediction results at multiple time scales are weighted and combined to output the initial network traffic prediction value. The fifth unit is used to build a correction model based on the statistical characteristics of historical prediction errors, dynamically calibrate the initial network traffic prediction value using the exponential smoothing method, and correct the prediction value in combination with the mutation detection results of the network performance indicators to obtain the network traffic prediction result.

Images