Method, computer readable medium for communication load prediction

Abstract

The present disclosure provides a method for communication load prediction, comprising: obtaining information of a plurality of load parameters; the information of each load parameter comprises a parameter value distribution of the load parameter; determining each load parameter as a strong regularity load parameter or a weak regularity load parameter according to the regularity of the load parameter; in the case that the load parameter is a strong regularity load parameter, determining a historical load shift model of the strong regularity load parameter; the historical load shift model is used to determine the parameter value distribution of the load parameter in a future predetermined time period according to the parameter value distribution of the load parameter in a historical predetermined time period; in the case that the load parameter is a weak regularity load parameter, training a machine learning model of the weak regularity load parameter; the machine learning model is used to predict the parameter value distribution of the load parameter in a future time period after a historical time period according to the parameter value distribution of the load parameter in the historical time period.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of communication and artificial intelligence, in particular to a communication load prediction method and computer readable medium. BACKGROUND

[0002] There are a large number of time-related indicators (such as load parameters) in communication (especially wireless communication), which dynamically change with the change of users and services. In many cases, the algorithm and strategy (such as energy saving strategy) need to be determined according to the data of the load parameters.

[0003] In some related technologies, the load parameters are pre-configured or detected in real time: if pre-configured, different cells, different scenes and different time periods are configured with the same strategy, which is homogeneous and has poor scene applicability; and if detected in real time, the algorithm and strategy can be adjusted only after the load parameters have changed, which has a lag and affects user experience. SUMMARY

[0004] The present disclosure provides a communication load prediction method and computer readable medium.

[0005] In a first aspect, the embodiments of the present disclosure provide a communication load prediction method, which includes:

[0006] obtaining information of a plurality of load parameters; the information of each load parameter includes a parameter value distribution of the load parameter;

[0007] determining each load parameter as a strong regularity load parameter or a weak regularity load parameter according to the regularity of the load parameter; the regularity of each load parameter represents the similarity of the parameter value distribution of the load parameter in two predetermined time periods;

[0008] in the case that the load parameter is a strong regularity load parameter, determining a historical load shift model of the strong regularity load parameter; the historical load shift model is used to determine the parameter value distribution of the load parameter in a future predetermined time period according to the parameter value distribution of the load parameter in a historical predetermined time period;

[0009] in the case that the load parameter is a weak regularity load parameter, training a machine learning model of the weak regularity load parameter; the machine learning model is used to predict the parameter value distribution of the load parameter in a future time period after a historical time period according to the parameter value distribution of the load parameter in the historical time period.

[0010] In a second aspect, the embodiments of the present disclosure provide a computer readable medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement any of the communication load prediction methods of the embodiments of the present disclosure.

[0011] In the embodiments of the present disclosure, the load parameter is predicted by the model, so that the load parameter can be known in advance and the algorithm and strategy (such as the energy-saving strategy) can be determined in advance to avoid hysteresis. Meanwhile, in the embodiments of the present disclosure, the corresponding prediction model is set for different cells and different parameters, so that accurate prediction can be realized according to the real situation of the load parameter of each cell to avoid homogeneity. In addition, in the embodiments of the present disclosure, different prediction models are set according to the strength of the regularity of the load parameter. For the load parameter with strong regularity, a simple historical load translation model is directly used, so that the amount of calculation required for model training and load parameter prediction is greatly reduced, and the realization is easy. BRIEF DESCRIPTION OF DRAWINGS

[0012] In the drawings of the embodiments of the present disclosure:

[0013] Figure 1 A flowchart of a method for communication load prediction provided by an embodiment of the present disclosure is shown in the figure;

[0014] Figure 2 A flowchart of a method for communication load prediction provided by another embodiment of the present disclosure is shown in the figure;

[0015] Figure 3 A process schematic diagram of a method for communication load prediction provided by an embodiment of the present disclosure is shown in the figure;

[0016] Figure 4 A logic schematic diagram of regularity judgment in a method for communication load prediction provided by an embodiment of the present disclosure is shown in the figure;

[0017] Figure 5 A schematic diagram of a sliding window in a method for communication load prediction provided by an embodiment of the present disclosure is shown in the figure;

[0018] Figure 6 A structure schematic diagram of LSTM in a method for communication load prediction provided by an embodiment of the present disclosure is shown in the figure;

[0019] Figure 7 A component block diagram of a computer readable medium provided by an embodiment of the present disclosure is shown in the figure. DETAILED DESCRIPTION

[0020] In order for those skilled in the art to better understand the technical solutions of the present disclosure, the method for communication load prediction and the computer readable medium provided by the embodiments of the present disclosure are described in detail below with reference to the drawings.

[0021] The embodiments shown will be described in more detail in the following with reference to the drawings, but the embodiments shown can be embodied in different forms and the present disclosure should not be interpreted as being limited to the embodiments set forth below. On the contrary, the purpose of providing these embodiments is to make the present disclosure thorough and complete and to enable those skilled in the art to fully understand the scope of the present disclosure.

[0022] The accompanying drawings are used to provide a further understanding of embodiments of the present disclosure and constitute a part of the specification, and together with the detailed embodiments, serve to explain the present disclosure, and do not constitute a limitation of the present disclosure. The above and other features and advantages will become more apparent from the detailed embodiments described below taken in conjunction with the accompanying drawings.

[0023] The present disclosure can be described with reference to plan views and / or cross-sectional views by virtue of the idealized schematic drawings of the present disclosure. Thus, the example illustrations can be modified according to manufacturing techniques and / or tolerances.

[0024] In the case of no conflict, each embodiment of the present disclosure and each feature in the embodiments can be combined with each other.

[0025] The terms used in the present disclosure are only used to describe specific embodiments, and are not intended to limit the present disclosure. As used in the present disclosure, the term "and / or" includes any and all combinations of one or more of the associated listed items. As used in the present disclosure, the singular forms "a" and "the" are also intended to include the plural forms, unless the context clearly indicates otherwise. As used in the present disclosure, the terms "comprise", "made of" designate the presence of the stated features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0026] Unless otherwise defined, all terms used in the present disclosure, including technical and scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art. It will also be understood that terms such as those defined in a generally used dictionary should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and the present disclosure, and should not be interpreted in an idealized or overly formal sense, unless the present disclosure expressly so defines.

[0027] The present disclosure is not limited to the embodiments shown in the drawings, but includes modifications of the configuration formed based on the manufacturing process. Therefore, the regions exemplified in the drawings have a schematic property, and the shape of the regions shown in the drawings exemplifies the specific shape of the region of the element, but is not intended to be restrictive.

[0028] In a first aspect, embodiments of the present disclosure provide a method for communication load prediction.

[0029] The method of the embodiments of the present disclosure is used to predict the load of a communication network element, specifically to determine the prediction model of the load parameter in a certain range (such as a network element, a cell, a cell lease, etc.), and send the prediction model to the corresponding network element, so that the network element uses the prediction model for prediction, and makes decisions on algorithms, strategies (such as energy saving strategies) according to the predicted load parameters.

[0030] In the embodiments of the present disclosure, the network element refers to a device or module in a communication network that needs to make some decisions according to load parameters, which can be in the form of network management, network element, base station, etc.

[0031] In the embodiments of the present disclosure, the load parameter refers to a parameter that will affect the load of the network element, which includes but is not limited to the number of RRC (Radio Resource Control) connections of an NR (New Radio) carrier, the usage rate of uplink PRB (Physical Resource Block) of the NR carrier, the usage rate of downlink PRB of the NR carrier, the usage rate of uplink PRB of a cell group, the usage rate of downlink PRB of a cell group, the number of uplink PRB of an LTE (Long Term Evolution) DSS (Dynamic Spectrum Sharing) cell group, the number of downlink PRB of the LTE DSS cell group, the number of RRC connections of the LTE DSS cell group, the number of uplink PRB of an LTE cell, the number of downlink PRB of the LTE cell, the number of RRC of the LTE cell, etc.

[0032] Reference Figure 1 The method for predicting communication load in the embodiments of the present disclosure at least includes but is not limited to the following steps:

[0033] S101, obtaining information of multiple load parameters.

[0034] The information of each load parameter includes the parameter value distribution of the load parameter.

[0035] The information of the load parameter is collected by the network management, network element, base station image, data service module, etc. corresponding to different communication scenarios, which is the parameter value of the load parameter at each time within a certain historical time period (of course, it is the actual parameter value of the historical time), that is, the parameter value distribution of the load parameter.

[0036] It should be understood that the specific data format (such as sampling rate), data volume, etc. of the information of the load parameter obtained above can be set as needed. For example, the parameter value of the load parameter per minute within a month before the sampling time can be obtained, that is, the granularity CollectDataStep can be 15 minutes, and the sampling time CollectDataTime can be 30 days.

[0037] It should be understood that the information of the load parameter can include other information in addition to the parameter value distribution of the load parameter, such as label information of the time corresponding to each parameter value, such as the day of the week, whether it is a special holiday, whether a major event occurs, etc. Through these additional information, the embodiments of the present disclosure can consider more factors when predicting the load parameter, so as to achieve more accurate and comprehensive prediction.

[0038] It should be understood that the obtained information of the load parameter can also be subjected to data cleaning, data completion, etc. which will not be described in detail here.

[0039] S102, determining, according to the regularity of the load parameter, that each load parameter is a strong regularity load parameter or a weak regularity load parameter.

[0040] The regularity of each load parameter represents the similarity of the parameter value distribution of the load parameter in two predetermined time periods.

[0041] For each load parameter, the parameter value distribution in different predetermined time periods is generally not exactly the same, but may have some similarity, and this similarity reflects whether the parameter value distribution of the load parameter is "regular"; if the parameter value distribution of the load parameter in two predetermined time periods is very similar, it means that the regularity of the load parameter is strong, and it is called "strong regularity load parameter"; if the parameter value distribution of the load parameter in two predetermined time periods is less similar, it means that the change of the load parameter is irregular, and it is called "weak regularity load parameter".

[0042] The selection of the above two predetermined time periods is various, but from the perspective of accuracy, the two adjacent predetermined time periods closest to the sampling time can be used, for example, the previous week and the previous two weeks of the sampling time are used as the two predetermined time periods.

[0043] S1031, in the case that the load parameter is a strong regularity load parameter, determining a historical load shift model of the strong regularity load parameter.

[0044] The historical load shift model is used to determine the parameter value distribution of the load parameter in a future predetermined time period according to the parameter value distribution of the load parameter in a historical predetermined time period.

[0045] S1032, in the case that the load parameter is a weak regularity load parameter, training a machine learning model of the weak regularity load parameter.

[0046] The machine learning model is used to predict the parameter value distribution of the load parameter in a future time period after a historical time period according to the parameter value distribution of the load parameter in the historical time period.

[0047] The type of the regularity of each load parameter is determined, and different subsequent processing is specifically made.

[0048] For the above strong regularity load parameter, a historical load shift model is determined to be used; since the regularity of the strong regularity load parameter is strong, the historical load shift model can directly determine the parameter value distribution in a future predetermined time period (such as the next week of the prediction time point) according to the parameter value distribution in a historical predetermined time period (such as the previous week of the prediction time point), to realize the prediction.

[0049] As for the weakly regular load parameters above, a machine learning model is determined to be used, and data in the information of the load parameters is used as training samples to train the machine learning model. The machine learning model is used to predict the future parameter value distribution according to some historical parameter value distribution.

[0050] The historical time and the predicted future time on which the machine learning model is based can be different in length. For example, the machine learning model can predict the parameter value distribution of the next day according to the parameter value distribution of the previous week.

[0051] It should be understood that the specific way of training the machine learning model is various, but the overall idea is to let the machine learning model predict the parameter value distribution at a later time (also belonging to the historical time) according to the historical parameter value distribution, and then calculate the loss of the current prediction of the machine learning model according to the actual parameter value distribution at the later time, and adjust the parameters of the machine learning model according to the loss until the machine learning model can achieve better prediction.

[0052] It should be understood that the above steps S1031 and S1032 are two processing methods for different situations, so their description and label order do not represent their execution order.

[0053] It should be understood that when there are multiple load parameters, the prediction model of each load parameter is determined according to its own regularity, so all load parameters can use the historical load translation model, all load parameters can use the machine learning model, or part of the load parameters can use the historical load translation model, and part of the load parameters can use the machine learning model.

[0054] In the embodiments of the present disclosure, the load parameters are predicted by the model, so the load parameters can be known in advance and the algorithm and strategy decision can be made in advance to avoid hysteresis. At the same time, in the embodiments of the present disclosure, the corresponding prediction model is set for different cells and different parameters, so that accurate prediction can be achieved according to the real situation of each cell load parameter, and homogeneity can be avoided. In addition, in the embodiments of the present disclosure, different prediction models are set according to the strength of the load parameter regularity. For the load parameter with strong regularity, a simple historical load translation model is directly used, so the required amount of calculation for model training and load parameter prediction is greatly reduced, and it is easy to implement.

[0055] In some embodiments, the predetermined time period is seven days.

[0056] As one way of the embodiments of the present disclosure, the above predetermined time period can be seven days (one week). This is because, in addition to the regular change of the parameter value distribution of the load parameter at each time of the day, there is also a certain change rule on each day of the week, so the predetermined time period of one week can well include such changes.

[0057] It should be understood that the predetermined time period is not limited to this, and it can also be one day, one month, etc.

[0058] In some embodiments, the historical load translation model is used to take the parameter value distribution of the load parameter in a nearest historical predetermined time period as the parameter value distribution of the load parameter in a nearest future predetermined time period.

[0059] As a way of the embodiments of the present disclosure, the historical load translation model corresponding to the strongly regular load parameter can directly take the parameter value distribution of the nearest previous predetermined time period as the (translated) parameter value distribution of the nearest predicted future predetermined time period.

[0060] For example, the parameter value distribution of the previous week of the prediction time point can be directly taken as the parameter value distribution of the next week of the prediction time point.

[0061] It should be understood that the specific way of the historical load translation model is not limited to this. For example, the historical load translation model can also take the distribution of the average parameter value of two corresponding times in the previous two weeks as the parameter value of the corresponding time in the next week.

[0062] In some embodiments, the machine learning model includes any one of a long short-term memory neural network model LSTM, a one-dimensional convolutional neural network model, and a Transformer model.

[0063] As a way of the embodiments of the present disclosure, the machine learning model corresponding to the weakly regular load parameter can be a long short-term memory neural network model (LSTM, Long-Short Term Memory).

[0064] Alternatively, the machine learning model can also be a one-dimensional convolutional neural network model, a Transformer model, and other deep learning models.

[0065] In some embodiments, with reference to Figure 2 After determining the historical load translation model of the strongly regular load parameter (S1031), the method further includes:

[0066] S1041, sending the historical load translation model of the load parameter to the network element corresponding to the load parameter.

[0067] After training the machine learning model of the weakly regular load parameter (S1032), the method further includes:

[0068] S1042, sending the machine learning model of the load parameter to the network element corresponding to the load parameter.

[0069] As a manner of the embodiments of the present disclosure, after obtaining the prediction model, the prediction model (the historical load shift model or the machine learning model) can also be sent to the corresponding network element, that is, the prediction model for a certain load parameter can be sent to the network element generating the load parameter, so that the network element can predict the load parameter according to the received prediction model and guide the network element to perform algorithm, strategy planning and the like.

[0070] It should be understood that if the prediction model is not sent to the network element, but the prediction is directly performed by the execution subject (such as a network management, a server and the like) of the present method, it is also feasible.

[0071] In some embodiments, referring to Figure 2 , the obtaining of the information of the plurality of load parameters (S101) comprises:

[0072] S1011, obtaining the information of the plurality of load parameters, and normalizing the parameter value distribution of each load parameter.

[0073] S1012, sending the normalized parameter of each load parameter to the network element corresponding to the load parameter.

[0074] As a manner of the embodiments of the present disclosure, the prediction model can process the "normalized" data, so it is necessary to normalize the parameter value of the load parameter first; thus, the data generated by the prediction model is also normalized, but the actual predicted value of the load parameter is required by the network element, so the normalized parameter is also sent to the network element corresponding to the load parameter, so that the network element can perform corresponding data normalization and restoration.

[0075] The specific manner of normalization is various, for example, it can be maximum-minimum normalization.

[0076] In some embodiments, referring to Figure 2 , after sending the historical load shift model of the load parameter to the network element corresponding to the load parameter (S1041) or sending the machine learning model of the load parameter to the network element corresponding to the load parameter (S1042), the method further comprises:

[0077] S105, receiving the prediction effect information of the network element.

[0078] The prediction effect information represents the accuracy of the parameter value distribution of the load parameter predicted by the network element according to the historical load shift model or the machine learning model.

[0079] S106, when the prediction effect information is lower than a predetermined standard, returning to the step of obtaining the information of the plurality of load parameters.

[0080] As a manner of the embodiments of the present disclosure, after the prediction model is determined, the prediction effect (i.e. whether the parameter value distribution of the predicted load parameter conforms to the actual parameter value distribution) can be analyzed (e.g. periodically analyzed), and when the prediction effect becomes poor due to aging, degradation or changes of the prediction model, the information of the load parameter is collected again, the prediction model is obtained again (i.e. the prediction model is updated) and is reissued to the network element. Of course, if the prediction effect does not become poor, the current prediction model can be continuously used and the prediction effect is analyzed.

[0081] In some embodiments, according to the regularity of the load parameter, determining each load parameter as a strong regularity load parameter or a weak regularity load parameter comprises:

[0082] The load parameter whose waveform trend similarity and numerical value similarity conform to the corresponding threshold range respectively is determined as a strong regularity load parameter, and other load parameters are determined as weak regularity load parameters; wherein the waveform trend similarity of each load parameter represents the similarity of the change trend of the parameter value distribution of the load parameter in two predetermined time periods, and the numerical value similarity of each load parameter represents the similarity of the magnitude of the parameter value of the load parameter in two predetermined time periods.

[0083] As a manner of the embodiments of the present disclosure, the regularity of the load parameter can include two aspects of waveform trend similarity and numerical value similarity.

[0084] The waveform trend similarity refers to that the change law of the parameter value distribution in two predetermined time periods is similar, for example, in two weeks, if the parameter value of a certain load parameter is higher on Monday and Thursday and lower on Saturday, it is considered that the change law of the two is similar, and the waveform trend similarity is high.

[0085] And the numerical value similarity refers to that the "absolute value" of the load parameter in two predetermined time periods cannot differ too much.

[0086] In the embodiments of the present disclosure, the corresponding threshold range (e.g. greater than 0.9 and less than 20%) can be set for the waveform trend similarity and the numerical value similarity respectively, and only when the waveform trend similarity and the numerical value similarity of a load parameter both conform to the corresponding threshold range, it is considered that the load parameter is a strong regularity load parameter.

[0087] It should be understood that the specific manner of regularity judgment according to the waveform trend similarity and the numerical value similarity is various. For example, the waveform trend similarity can be calculated first and compared with the corresponding threshold range, and only when it conforms to the threshold range, the numerical value similarity is calculated and compared with the corresponding threshold range; and if the waveform trend similarity does not conform to the threshold range, it is directly determined as a weak regularity load parameter, and the numerical value similarity does not need to be calculated.

[0088] In some embodiments, the waveform trend similarity is calculated by the following formula:

[0089]

[0090] The numerical similarity is calculated by the following formula:

[0091]

[0092] wherein self_cov represents the waveform trend similarity, σ represents the operation of Pearson Correlation Coefficient, train_y_front represents the parameter value distribution of the load parameter in the first predetermined time period, train_y_behind represents the parameter value distribution of the load parameter in the second predetermined time period, relative_mean_diff represents the numerical similarity, mean represents the operation of averaging, and max represents the operation of taking the maximum value.

[0093] As a manner of the embodiments of the present disclosure, the waveform trend similarity can be calculated according to the Pearson Correlation Coefficient by the above formula. Compared with the similarity represented by relative Euclidean distance, cosine value, etc., the above manner eliminates the influence of different numerical dimensions by standardized calculation.

[0094] The numerical similarity can be calculated according to the above formula, i.e., the difference between the average values of the load parameter in the two predetermined time periods, the proportion of the relative larger average value, obviously, the smaller the value of the proportion, the higher the numerical similarity.

[0095] It should be understood that the specific algorithms for calculating the waveform trend similarity and the numerical similarity are not limited thereto.

[0096] Next, referring to Figure 3 , a specific method of communication load prediction of the embodiments of the present disclosure is exemplarily described, which can include:

[0097] Stage one: data collection and data preprocessing

[0098] The data collection is to collect the information of the required load parameter from the network management, network element, base station image or data service module, etc. After the data collection is completed, the data preprocessing such as data cleaning and missing data completion is performed, so as to be used for the subsequent steps.

[0099] And the data collection and data preprocessing can specifically include:

[0100] A1.1, data collection: collect the load parameter information of multiple time lengths CollectDataTime (unit: day) from network management, network element or base station image module with granularity CollectDataStep (unit: minute), that is, collect training data TrainingData (parameter value of load parameter).

[0101] For example, 30 days of data can be collected with 15 minutes as the granularity, so there are 96 data per day.

[0102] The load parameters include, but are not limited to, the number of RRC connections of the NR carrier, the uplink PRB usage rate of the NR carrier, the downlink PRB usage rate of the NR carrier, the uplink PRB usage rate of the cell group, the downlink PRB usage rate of the cell group, the uplink PRB usage number of the LTE DSS cell group, the downlink PRB usage number of the LTE DSS cell group, the RRC connection number of the LTE DSS cell group, the uplink PRB usage number of the LTE cell, the downlink PRB usage number of the LTE cell, the RRC number of the LTE cell, etc.

[0103] A1.2, data time axis completion: during data collection, some data may be missing within a few granularities, so the time axis needs to be filled accordingly. The filled data is empty.

[0104] A1.3, data time axis deduplication: during data collection, some data may be repeatedly collected within a few granularities, so the repeatedly collected data needs to be deduplicated. The deduplication rule can be to retain the first data in the data set and delete the subsequent repeated data.

[0105] A1.4, data completion: if the missing data is at the head of the data set, the empty data at the head can be filled with the first non-empty data from the head; if the missing data is at the tail of the data set, the empty data at the tail can be filled with the first non-empty data from the tail; if the missing data is in the middle of the data set, the first non-empty data can be found forward and backward respectively, and linear interpolation filling (or mean filling or other ways) is performed.

[0106] Phase two: regular threshold judgment

[0107] Referring to Figure 4 , the regular threshold judgment distinguishes different regularities of the time series load parameter data and establishes a prediction model respectively, so as to adaptively implement targeted solutions for different types of load parameter data, thereby accelerating the operation effect and improving the prediction accuracy on the premise of reducing the overall computing resource investment.

[0108] The strength of the load parameter regularity is respectively embodied in: (1) the parameter value distribution of the load parameter is similar in the waveform change trend of the adjacent two time series (a predetermined time period, in "weeks" as a unit), that is, the waveform consistency; (2) the parameter value of the load parameter is basically stable in size in the adjacent two weeks, that is, the numerical consistency.

[0109] Therefore, we use the two-week correlation coefficient matrix mean and the relative mean difference to judge the waveform trend similarity and the numerical similarity respectively, and compare the two scores with the respective regularity judgment threshold (threshold range) to make a comprehensive judgment on the strength of the regularity.

[0110] A2.1, normalize the data set: obtain the maximum value MaxData of the parameter value of each load parameter in TrainingData, and perform maximum and minimum normalization (or other normalization methods) on the parameter value of the load parameter, that is, calculate TrainingData / MaxData; then send the MaxData index (normalization parameter) to the corresponding network element, so that the network element can perform reverse normalization (restoration) processing on the prediction result data generated by the prediction model.

[0111] A2.2, calculate the correlation coefficient matrix mean: the waveform trend similarity (correlation coefficient) reflects the similarity of the waveform between the data distribution of the last week (train_y_behind) and the data distribution of the week before last (train_y_front) in the training data, that is, to what extent the new data after repeats the change pattern of the existing data before:

[0112] The calculation can be based on the Pearson Correlation Coefficient. Compared with the similarity represented by Euclidean distance and cosine, it eliminates the influence of different numerical dimensions through standardized calculation.

[0113] Of course, if other alternative vector distance or similarity calculation schemes are used, it is also feasible.

[0114] A2.3, calculate the relative mean difference: after it is judged that the correlation coefficient matrix mean is large (such as greater than 0.9), that is, it is confirmed that the waveforms of the two weeks are highly similar, the data can be additionally calculated and judged by the relative mean difference:

[0115]

[0116] That is, under the premise that the waveform regularity has been confirmed to be similar, the parameter value of the load parameter is re-determined to ensure that the overall difference is not too large (e.g., less than 20%).

[0117] A2.4, Regularity Threshold Judgment: According to different application scenarios, the regularity judgment threshold parameters ThresholdRegularity and ThresholdMeanDiff can be configured to determine the division basis of strong regularity and non-strong regularity (weak regularity) load parameters. ThresholdRegularity is used to preliminarily divide the regularity, and the correlation coefficient matrix mean greater than or equal to ThresholdRegularity is the strong regularity load parameter preliminarily judged, and further judgment is made on whether the relative mean difference is less than the threshold ThresholdMeanDiff. If both meet the requirements, LSTM is used; if one of them does not meet the requirements, the historical load shifting model is used.

[0118] Phase Three: Model Training

[0119] Model training is based on time series statistics of historical data to build a prediction model for each load parameter (periodicity, volatility, trend, flow size, etc.), i.e., historical load shifting model LoadShifting and prediction model LoadForcasting. Model training specifically includes:

[0120] A3.1, for strong regularity load parameters, determine to use historical load shifting model LoadShifting. As can be seen, when judging the load parameter of the (t+1)th week, if it has been proved that the load data of the tth week and the (t-1)th week have strong similarity, we can a priori believe that the data of the (t+1)th week will also continue such similarity, that is, the data of the (t+1)th week should be similar to the data of the tth week and the (t-1)th week. Therefore, LoadShifting model is used, and is sent to the network element.

[0121] A3.2, for weak regularity load parameters, determine to train the corresponding prediction model LoadForcasting, i.e., learn the periodic fluctuation regularity, increasing and decreasing trend, etc. of the data time series through the prediction model, and deeply combine the internal characteristics to make model prediction.

[0122] A3.2.1, data construction.

[0123] A3.2.1.1, data sample construction: sample data is obtained by sliding window method.

[0124] Specifically, the load assessment period (15-minute granularity) for a cell can be any time point that is an integer multiple of 15 minutes from the absolute time (00:00) (0, 15, 30, 45), with a window length of (LoadPeriod*96+96) and a sliding window step size of 96 (i.e., one day). Thus, referring to... Figure 5 21 training samples can be obtained by constructing a 28-day data sliding window.

[0125] Each sample needs to be constructed according to the input format requirements of LSTM. The constructed sample data consists of two parts: feature data and label data.

[0126] Feature data (X): LoadPeriod * 96 consecutive normalized load parameter data, where LoadPeriod is 7 by default, which is one week of continuous data.

[0127] Tag data (Y): 96 consecutive normalized load parameter data that immediately follow the feature data X, i.e., continuous data one day after the feature data.

[0128] A3.2.1.2 Training and Test Set Partitioning: During training, to evaluate the accuracy of the model, the data needs to be divided into a training set D_norm_train and a test set D_norm_test. The training set is used to train the model, and the test set is used to verify the training effect of the model. The ratio of the training set is alpha, and the ratio of the test set is (1-alpha).

[0129] A3.2.1.3 Parameter Normalization: Typically, the network management system and network elements use the same MaxData parameter. That is, when the network element side determines that the maximum value TestMaxData in the test data is less than or equal to the maximum value parameter MaxData set in the training data on the network management side, then the set MaxData is used as the maximum value parameter. However, when the network element side determines that the maximum value TestMaxData in the test data is greater than the maximum value parameter MaxData set in the training data on the network management side, the reasons may be: (1) sudden abnormal value; (2) the overall load increases (i.e., the actual generated data is greater than the maximum value of the previously collected data). Therefore, we can first distinguish between the two possible situations and then take appropriate measures accordingly.

[0130] (1) TestMaxData>MaxData is because of a small number of burst values / outliers:

[0131] Identification: When the mean of TestData (MeanTestData) and the mean of TrainData (MeanTrainData) are not significantly different, meaning that the proportion of increasing values ​​in TestData is small, it indicates a small number of sudden values. For example:

[0132] |MeanTestData-MeanTrainData| / MeanTrainData<20%;

[0133] Processing: For a small number of burst values, the MaxData parameter is not modified, and only abnormal values greater than 1 after normalization are processed, such as replacing TestMaxData with the original MaxData, that is, replacing values greater than 1 after normalization with 1.

[0134] (2) TestMaxData>MaxData is because the overall load increases:

[0135] Recognition: The mean value MeanTestData of TestData and the mean value MeanTrainData of TrainData differ greatly, that is, the proportion of large values in TestData is large, indicating that the overall load is large, for example:

[0136] |MeanTestData-MeanTrainData| / MeanTrainData≧20%;

[0137] Processing: Since the overall load is large, the MaxData value is modified to the maximum value of the new data, that is, MaxData=TestMaxData.

[0138] A3.2.2, model structure selection: refer to Figure 6 , the basic structure of LSTM includes an input layer, an output layer, and several hidden layers.

[0139] A3.2.2.1, parameter setting.

[0140] The parameter setting can use the following examples:

[0141] UnitsNum (number of hidden layer neurons): number of single-layer network neurons, default 16 is selected;

[0142] LoadPeriod (data period): 7 days of data are selected;

[0143] ActFunc (activation function): value range enum(‘tanh’,‘relu’,‘elu’,‘linear’,‘sigmoid’), default ‘elu’ is selected;

[0144] DropRatio (deactivation ratio): to prevent overfitting, the value range is (0-1), and the default is 0;

[0145] According to the above configuration, the model size is about 50KB.

[0146] BatchSize (Batch Size): The amount of data for each training batch LoadPeriodBatchSize, the default is 16;

[0147] Epoch (training rounds): The number of training batches LoadPeriodEpoch, the default is 100 times;

[0148] Optimizer (optimizer): Determine the convergence method of model gradient descent, the value range is enum ('adam', 'adagrad', 'rmsprop','momentem'), the default is adam;

[0149] The model is initialized with the determined mean and variance parameters and the random seed is set, which facilitates the reproduction of different scenarios.

[0150] A3.2.3, model effect judgment.

[0151] A3.2.3.1, judgment process: by setting the error threshold LoadPredictMError, the validity of the model prediction result is verified.

[0152] A3.2.3.1.1, model training: calculate the error Error_train of the model LoadPredictModel on the training set D_norm_train, and judge by the following logic:

[0153] IF Error_train<LoadPredictModel:

[0154] Output model LoadPredictModel;

[0155] Else:

[0156] Do not output the model LoadPredictModel.

[0157] A3.2.3.1.2, model verification: calculate the error Error_test of the model LoadPredictModel on the test set D_norm_test, and judge by the following logic:

[0158] If Error_test<LoadPredictModel:

[0159] Output the model LoadPredictModel to the online inference engine (network element side);

[0160] Else:

[0161] LoadPredictModel is not exported to the online inference engine (network element side).

[0162] A3.2.3.2, error calculation.

[0163] The error is set as the relative error between the predicted value load_predict and the true value load_true, both load_predict and load_true are time series containing N 15-minute granularity data of future periods, denoted as:

[0164] load_predict = [load_predict_1,..., load_predict_i,..., load_predict_N];

[0165] Note: If there is a negative value in the prediction result load_predict, it needs to be replaced with 0.

[0166] load_true = [load_true_1,..., load_true_i,..., load_true_N];

[0167] Then we define the sample relative average error (MAE / sample mean MeanTrain) as:

[0168]

[0169] The error threshold LoadPredictMError is configured by the network management, and the closer the value is to 1, the more relaxed the standard for model adoption. The threshold can be set to 0.4.

[0170] A3.2.4, model export.

[0171] After model training, it is exported as an HDF5 format file: model.save('XXX.h5')

[0172] Model naming rules: The model package issued by the network management side adopts zip package format, and the model package naming rule is: network element ID + time.zip. The model package contains multiple model directories, for example:

[0173]

[0174]

[0175] A3.2.4.1, model_name1 model name format: algorithm name_business object id; Where the algorithm name, business object id is agreed by the network management intelligent application and the network element intelligent application.

[0176] A3.2.4.1.1, for history load shift models, for example, there are four in total, corresponding to four model names as follows:

[0177] - NR carrier uses RB number model, model of NR carrier id = 1: PRBloadshift_carrierid_1;

[0178] - NR carrier RRC user number model, model of NR carrier id = 1: RRCloadshift_carrierid_1;

[0179] - NR cell group uses RB number model, model of NR cell NCGI = {PLMN: 46001, gNBid: 1, cellid: 1}: PRBloadshift_PLMN_460-01_gNBid_1_cellid_1;

[0180] - NR cell group RRC user number model, model of NR cell NCGI = {PLMN: 46001, gNBid: 1, cellid: 1}: RRCloadshift_PLMN_460-01_gNBid_1_cellid_1.

[0181] A3.2.4.1.2, for LSTM, for example, there are four in total, corresponding to four model names as follows:

[0182] - NR carrier uses RB number model, model of NR carrier id = 1: PRBloadforecast_carrierid_1;

[0183] - NR carrier RRC user number model, model of NR carrier id = 1: RRCloadforecast_carrierid_1;

[0184] - NR cell group uses RB number model, model of NR cell NCGI = {PLMN: 46001, gNBid: 1, cellid: 1}: PRBloadforecast_PLMN_460-01_gNBid_1_cellid_1;

[0185] - NR cell group RRC user number model, model of NR cell NCGI = {PLMN: 46001, gNBid: 1, cellid: 1}: RRCloadforecast_PLMN_460-01_gNBid_1_cellid_1.

[0186] A3.2.4.2, version1: must be a number, 32-bit integer, can be automatically generated by network management intelligent application, manually specified or automatically generated by LSE.

[0187] A3.2.4.3, trainpara.json training parameter information content must contain: model name, build time, version number, other related parameters are customized according to different algorithms.

[0188] For LSTM, the following parameters also need to be included:

[0189] -NR carrier RB number normalization denominator;

[0190] -NR carrier RRC user number normalization denominator;

[0191] -NR cell group RB number normalization denominator;

[0192] -NR cell group RRC user number normalization denominator.

[0193] For historical load translation model, no additional parameters are required.

[0194] Stage four: load parameter prediction

[0195] The prediction model obtained by training is issued to the network element, and the load condition is predicted according to the prediction model to guide the energy saving and shutdown work. Since the load condition in the cell may change greatly over time, the model needs to be updated at a certain period, and the updated model is reissued to the network element.

[0196] A4.1, start prediction.

[0197] Determine whether the conditions for load parameter prediction are met: (1) the load prediction switch is on; (2) the RSE model has been loaded. If all the above conditions are met, start the load prediction timer (tentatively 00:00:00), after the timer expires, call the load prediction service interface of the middle station, initiate load prediction, and the load prediction component of the middle station extracts the load data of the NR carrier and the NR cell group for prediction. It should be understood that the prediction model of the user number load of the NR carrier and the NR cell group and the prediction model of the PRB load are different, and need to be extracted and predicted respectively.

[0198] A4.2, when predicting the load parameters of date X, the prediction input data needs to extract the load parameter data of (X-8)~(X-1) days. The output data is the load parameter data of 96 granularities of the whole day of X date.

[0199] A4.3 After receiving the prediction model from UME, RSE decompresses the file to obtain the prediction model file model.tflite and the algorithm configuration file trainpara.json, and performs the following processing:

[0200] RSE automatically loads algorithm models. Even if the model file contains an empty model, it still needs to be loaded, which is equivalent to deleting the model that RSE is currently using. It saves the algorithm configuration file, which is used to organize the prediction input data and set the input parameters for the prediction algorithm. The file uses JSON format, and a fixed template for the algorithm description is defined for each algorithm. It queries RSE for prediction model capabilities; if RSE has already loaded the prediction model, RSE returns success; otherwise, it returns failure.

[0201] A4.4 Prediction model loaded successfully.

[0202] RSE broadcasts model information. After receiving the broadcast, the middleware prediction component determines whether the energy-saving carrier and cell support load prediction and which algorithm model to use to perform the prediction based on the model type and the corresponding model object primary key ID carried in the broadcast, and saves the corresponding information. After completion, the middleware prediction component broadcasts a load prediction capability notification to the application. After receiving the notification, the application determines whether it can initiate load prediction based on the broadcast content.

[0203] Secondly, referring to Figure 7 This disclosure provides a computer-readable medium having a computer program stored thereon, which, when executed by a processor, implements any of the communication load prediction methods of this disclosure.

[0204] Among them, the processor is a device with data processing capabilities, including but not limited to the central processing unit (CPU); the memory is a device with data storage capabilities, including but not limited to random access memory (RAM, more specifically SDRAM, DDR, etc.), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), and flash memory (FLASH); the I / O interface (read-write interface) is connected between the processor and the memory, enabling information exchange between the memory and the processor, including but not limited to the data bus (Bus).

[0205] Those skilled in the art will understand that all or some of the steps, systems, and devices disclosed above, as functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0206] In hardware implementations, the division between functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be executed by several physical components working together.

[0207] Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit (CPU), a digital signal processor, or a microprocessor, or hardware, or a combination of software and / or hardware. Such software can be distributed on computer readable media, which can comprise computer storage media (or non-transitory media), and communication media (or transitory media). Computer storage media, as used herein, includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer storage media includes, but is not limited to, random access memory (RAM), such as SDRAM, double data rate SDRAM (DDR SDRAM), and / or the like, read only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and that can be accessed by a computer. Further, it should be appreciated by those skilled in the art that computer storage media generally includes any media that can be used for storing and / or transferring data accessible by a computer, including but not limited to, carrier waves, modulated data signals, and / or the like.

[0208] The present disclosure has disclosed example embodiments, and although the specific terms are employed, they are used in a generic sense only and should not be construed to be limited to the specific embodiments described herein. In some instances, it will be readily apparent to those skilled in the art that the certain features, characteristics and / or elements described in connection with a particular embodiment can be used independently, or in combination with respect to other embodiments, and that various modifications can be made without departing from the scope of the disclosure as set forth in the appended claims. Thus, the skilled artisan will understand that various changes can be made in the form, details, and / or arrangements of the exemplary embodiments without departing from the scope of the disclosure set forth in the appended claims.

Claims

1. A method of communication load prediction, comprising: obtaining information of a plurality of load parameters; the information of each of the load parameters comprises a parameter value distribution of the load parameter; determining each of the load parameters as a strong regularity load parameter or a weak regularity load parameter according to a regularity of the load parameter; the regularity of each of the load parameters represents the similarity of the parameter value distribution of the load parameter in two predetermined time periods; in the case of a strong regularity load parameter, determining a historical load shift model of the strong regularity load parameter; the historical load shift model is used to determine the parameter value distribution of the load parameter in a future predetermined time period according to the parameter value distribution of the load parameter in a historical predetermined time period; in the case of a weak regularity load parameter, training a machine learning model of the weak regularity load parameter; the machine learning model is used to predict the parameter value distribution of the load parameter in a future time period after a historical time period according to the parameter value distribution of the load parameter in the historical time period. 2.According to claim 1, wherein, the historical load shift model is used to take the parameter value distribution of the load parameter in a nearest historical predetermined time period as the parameter value distribution of the load parameter in a nearest future predetermined time period.

3. The method of claim 1, wherein, the machine learning model comprises: any one of a long short-term memory neural network model LSTM, a one-dimensional convolutional neural network model, and a transformation template Transformer model. 4.According to claim 1, wherein, after the determination of the historical load shift model of the strong regularity load parameter, further comprising: sending the historical load shift model of the load parameter to a network element corresponding to the load parameter; after the training of the machine learning model of the weak regularity load parameter, further comprising: sending the machine learning model of the load parameter to a network element corresponding to the load parameter.

5. The method of claim 4, wherein, the obtaining of the information of a plurality of load parameters comprises: obtaining information of a plurality of load parameters, and normalizing the parameter value distribution of each of the load parameters; sending the normalized parameter of each of the load parameters to a network element corresponding to the load parameter.

6. The method of claim 4, wherein, after the sending of the historical load shift model of the load parameter to a network element corresponding to the load parameter or the sending of the machine learning model of the load parameter to a network element corresponding to the load parameter, further comprising: receiving prediction effect information of the network element; the prediction effect information represents the accuracy of the parameter value distribution of the load parameter predicted by the network element according to the historical load shift model or the machine learning model; when the prediction effect information is lower than a predetermined standard, returning to the step of obtaining the information of a plurality of load parameters.

7. The method of claim 1, wherein, the determination of each of the load parameters as a strong regularity load parameter or a weak regularity load parameter according to the regularity of the load parameter comprises: The load parameters with waveform trend similarity and numerical similarity meeting corresponding threshold ranges are determined as strong regularity load parameters, and other load parameters are determined as weak regularity load parameters; wherein, the waveform trend similarity of each load parameter represents the similarity of the change trend of the parameter value distribution of the load parameter in two predetermined time periods, and the numerical similarity of each load parameter represents the similarity of the parameter value size of the load parameter in two predetermined time periods.

8. The method of claim 7, wherein, The waveform trend similarity is calculated by the following formula: The numerical similarity is calculated by the following formula: wherein, self_cov represents the waveform trend similarity, σ represents the operation of Pearson correlation coefficient, train_y_front represents the parameter value distribution of the load parameter in the first predetermined time period, train_y_behind represents the parameter value distribution of the load parameter in the second predetermined time period, relative_mean_diff represents the numerical similarity, mean represents the operation of taking average value, and max represents the operation of taking maximum value.

9. The method of claim 1, wherein, The predetermined time period is seven days.

10. A computer readable medium having stored thereon a computer program, the computer program being executed by a processor to implement the method of communication load prediction of any one of claims 1 to 9.

Images