CN111131424A - Service quality prediction method based on combination of EMD and multivariate LSTM - Google Patents

Service quality prediction method based on combination of EMD and multivariate LSTM Download PDF

Info

Publication number
CN111131424A
CN111131424A CN201911308215.9A CN201911308215A CN111131424A CN 111131424 A CN111131424 A CN 111131424A CN 201911308215 A CN201911308215 A CN 201911308215A CN 111131424 A CN111131424 A CN 111131424A
Authority
CN
China
Prior art keywords
data
service quality
service
time sequence
lstm
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201911308215.9A
Other languages
Chinese (zh)
Other versions
CN111131424B (en
Inventor
李兵
陈秀清
王健
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan University WHU
Original Assignee
Wuhan University WHU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Wuhan University WHU filed Critical Wuhan University WHU
Priority to CN201911308215.9A priority Critical patent/CN111131424B/en
Publication of CN111131424A publication Critical patent/CN111131424A/en
Application granted granted Critical
Publication of CN111131424B publication Critical patent/CN111131424B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00—Network arrangements or protocols for supporting network services or applications
    • H04L67/01—Protocols
    • H04L67/02—Protocols based on web technology, e.g. hypertext transfer protocol [HTTP]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/21—Design, administration or maintenance of databases
    • G06F16/215—Improving data quality; Data cleansing, e.g. de-duplication, removing invalid entries or correcting typographical errors
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/245—Query processing
    • G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
    • G06F16/2474—Sequence data queries, e.g. querying versioned data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/10—Pre-processing; Data cleansing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00—Network arrangements or protocols for supporting network services or applications
    • H04L67/50—Network services
    • H04L67/60—Scheduling or organising the servicing of application requests, e.g. requests for application data transmissions using the analysis and optimisation of the required network resources
    • H04L67/61—Scheduling or organising the servicing of application requests, e.g. requests for application data transmissions using the analysis and optimisation of the required network resources taking into account QoS or priority requirements

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computing Systems (AREA)
  • Databases & Information Systems (AREA)
  • Signal Processing (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Biophysics (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Biomedical Technology (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Medical Informatics (AREA)
  • Quality & Reliability (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Fuzzy Systems (AREA)
  • Probability & Statistics with Applications (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a service quality prediction method based on the combination of EMD and multivariable LSTM, which comprises the following steps of 1: performing data cleaning on the Web service historical call records, and detecting missing values and abnormal values of Web service quality in a data set; 2: filling the missing value and the abnormal value by using a data-based filling algorithm, and constructing a complete and effective service quality time sequence; 3: carrying out data transformation on the time sequence; 4: the service quality time sequence is decomposed into a plurality of eigenmode functions and residual wave parts by using an EMD method, then a multivariate time sequence is constructed, and a multivariate LSTM model is established for service quality prediction. The invention has the beneficial effects that: 1) unknown service quality can be accurately predicted according to the historical calling record of the Web service, and the method has good practicability. 2) By predicting possible SLA violations, the user is helped to select a service that provides the best quality of service with a higher probability of meeting the SLA constraints.

Description

Service quality prediction method based on combination of EMD and multivariate LSTM
Technical Field
The invention relates to the technical field of service calculation, in particular to a service quality prediction method based on the combination of EMD and multivariate LSTM.
Background
Web services are loosely coupled software systems that interact through a network to support interoperating machines, which provides a standardized solution for Service-oriented architecture (SOA). The number of Web services is increasing every year, and there are a large number of services providing similar functions, for example, according to programable Web (http:// www.programmableweb.com/, PWeb for short) statistics, the number of webapis capable of providing search services is as high as 3546 by 2019, 10 and 11. In the face of such a multifunctional identical or similar Web service, we need to select the most appropriate service according to the service quality to better meet the user requirements.
In recent years, researchers have proposed many service quality prediction methods for predicting unknown QoS values of Web services based on historical call records of Web services, and there are three main methods: a method based on similarity measure, a method based on time series prediction technique and a method based on deep learning model.
The inventor of the present application finds that the method of the prior art has at least the following technical problems in the process of implementing the present invention:
the existing QoS prediction models have the defects that the time dynamics of QoS cannot be captured by a similarity-based method, and QoS information from a plurality of data sources (such as users) is needed; the method based on the time series prediction technology cannot ensure accurate prediction in the face of data with high volatility and time-varying characteristics. Neural network models (ANN, MLP, RNN) have the ability to handle non-linear complex relationships, but can only be used to handle short-term data dependencies. Inaccurate predictions may lead to key problems such as incorrect selection of Web services or unnecessary management decisions, which further affect subsequent development and cause unnecessary losses.
Therefore, the method in the prior art has the technical problem that the prediction result is not accurate enough.
Disclosure of Invention
In view of the above, the present invention provides a service quality prediction method based on the combination of EMD and multivariate LSTM, so as to solve or at least partially solve the technical problem of the prior art that the prediction result is not accurate enough.
In order to solve the above technical problem, the present invention provides a service quality prediction method based on the combination of EMD and multivariate LSTM, comprising:
step S1: carrying out data cleaning on the Web service historical calling record, and detecting an abnormal value and a missing value;
step S2: performing data completion on the detected abnormal values and the detected missing values by adopting a preset data-based filling algorithm to construct a complete and effective service quality time sequence;
step S3: carrying out data transformation on the complete and effective service quality time sequence;
step S4: decomposing the service quality time sequence subjected to data transformation based on an EMD method to obtain a multivariate time sequence, and obtaining training data according to the multivariate time sequence;
step S5: training a pre-constructed LSTM model by using training data to obtain a trained multivariate LSTM model;
step S6: and performing quality prediction on the Web service call record to be predicted by using the trained multivariate LSTM model, and selecting the target Web service according to the quality prediction result.
In one embodiment, step S1 specifically includes:
step S1.1: identifying service abnormality in the Web service history calling record, screening out a preliminary abnormal value and marking the preliminary abnormal value as a missing value;
step S1.2: and detecting abnormal values in the Web service history calling record by using the box line graph, and taking observed values which are larger than or smaller than the upper and lower boundaries set by the box line graph in the detection result as target abnormal values and identifying the target abnormal values.
In one embodiment, step S2 specifically includes:
step S2.1: extracting a change trend on a date level and a change trend within one day from normal data subjected to data cleaning, wherein the rest is residual partial data, and the service quality data subjected to data cleaning comprises normal data and abnormal data;
step S2.2: selecting basic time characteristics of different time granularities and historical lag values of service quality data from normal data as characteristics, using residual error part data as a label, training an Xgboost model, and predicting residual error parts of abnormal values and missing values by using the trained Xgboost model to obtain a residual error predicted value;
step S2.3: and adding the variation trend of the extracted normal data on the date level, the variation trend within one day and the residual prediction value of the abnormal data, and filling the abnormal value and the missing value to obtain a data filling result.
In one embodiment, step S3 specifically includes:
step S3.1: carrying out Box _ Cox transformation on the complete and effective service quality time sequence;
step S3.2: and carrying out logarithmic transformation on the service quality time series subjected to the Box _ Cox transformation.
In one embodiment, step S4 specifically includes:
step S4.1: decomposing the service quality time sequence with irregular frequency into a plurality of frequency single IMF components and a residual part Res by using an EMD method;
step S4.2: carrying out scale transformation on the decomposed data, and converting the numerical values of all components of the service quality time sequence into the range of an activation function of the LSTM model;
step S4.3: and constructing a 3D tensor according to the multivariable time sequence subjected to scale transformation, wherein the form of the 3D tensor is (samples, times, features), the samples represent independent observation results from the domain and are data rows, the times represent that current values are related to data continuously input in the previous time step and are history lag values, and the features represent characteristics observed at each time point and are a plurality of IMF components and residual parts Res.
In one embodiment, the EMD method is used to decompose the time series into a set of eigenmode functions and a residual waveform in step S4.1, which results in the following equation:
Figure BDA0002323751630000031
wherein X (t) represents the original time series, IMFn(t) denotes the nth IMF component, and Res (t) denotes the residual wave portion.
In one embodiment, the previously constructed LSTM model includes two long-term and short-term memory networks and a fully connected layer, and step S5 specifically includes:
step S5.1: inputting the multivariate time sequence into a pre-constructed LSTM model, and learning high-level time sequence representation through a two-layer long-time and short-time memory network;
step S5.2: and transmitting the last step of the output sequence to a full-connection layer to realize the prediction of the next time step and obtain the trained multivariate LSTM model.
In one embodiment, Box _ Cox transformation is performed on the time series in step S3.1, using the following formula:
Figure BDA0002323751630000032
performing a transformation, wherein λ represents an introduced parameter, Y represents a time series to be transformed, Y(λ)Representing the transformed result.
In one embodiment, the calculation procedure of the long and short term memory network in step S6 is as follows:
step S6.1: selecting information discarded from the cell state by forgetting gate and determining that the previous long-term memory c is to be usedt-1To what extent c is savedtThe method comprises the following steps:
Figure BDA0002323751630000041
step S6.2: new information saved into the long-term memory state is determined by the input gate:
Figure BDA0002323751630000042
step S6.3: calculating a candidate value for the long-term state based on determining new information to be saved into the long-term memory state:
Figure BDA0002323751630000043
step S6.4: updating the long-term state with the candidate values:
Figure BDA0002323751630000044
step S6.5: determine the current LSTM output value through the output gate:
Figure BDA0002323751630000045
wherein the hidden layer state htThe calculation formula of (a) is as follows:
Figure BDA0002323751630000046
sigma represents a sigmoid function, ⊙ represents a bitwise multiplication operation, and the calculation formula of the sigmoid function and the tanh function is as follows:
Figure BDA0002323751630000047
Figure BDA0002323751630000048
m represents the dimension of the input variable accepted by the LSTM, u represents that the LSTM possesses u neurons,
Figure BDA0002323751630000049
Figure BDA00023237516300000410
and
Figure BDA00023237516300000411
respectively showing the states of a forgetting gate, an input gate and a control gate of the LSTM at the time t, all being u-dimensional vectors,
Figure BDA00023237516300000412
and
Figure BDA00023237516300000413
the output and the cell state of the LSTM at time t are u-dimensional vectors, V is a weight matrix in a hidden state, is a u × u matrix, W is an input weight matrix, is a u × m matrix, and b is an offset term.
In one embodiment, after obtaining the prediction, the method further comprises: and judging whether the prediction result meets the SLA constraint of the service level agreement, if not, determining that the SLA violates the rules, and if so, not violating the rules.
One or more technical solutions in the embodiments of the present application have at least one or more of the following technical effects:
the invention discloses a service quality prediction method based on the combination of EMD and multivariable LSTM, which comprises the steps of firstly, carrying out data cleaning on Web service history calling records, and detecting abnormal values and missing values; then, a preset data-based filling algorithm is adopted to perform data completion on the detected abnormal values and the detected missing values, and a complete and effective service quality time sequence is constructed; then, carrying out data transformation on the complete and effective service quality time sequence; then decomposing the service quality time sequence subjected to data transformation based on an EMD method to obtain a multivariate time sequence, and obtaining training data according to the multivariate time sequence; training a pre-constructed LSTM model by using training data to obtain a trained multivariate LSTM model; and finally, performing quality prediction on the Web service call record to be predicted by using the trained multivariate LSTM model, and selecting the target Web service according to a quality prediction result.
By the method, the service quality time sequence can be decomposed into a plurality of IMF components and residual wave parts by an Empirical Mode Decomposition (EMD) method, then the multivariate time sequence is constructed and used for training the preset LSTM model, so that the trained multivariate LSTM model is obtained, further the service quality prediction is carried out, the unknown service quality can be accurately predicted according to the historical calling record of the Web service, the accuracy of the prediction result is improved, the WEB service with good service quality can be selected based on the prediction result, and the method has good practicability.
Further, by predicting possible Service Level Agreement (SLA) violations, users are helped to select services that provide the best quality of service and have a higher probability of meeting SLA constraints.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.
FIG. 1 is a flow chart of the method of the present invention for predicting quality of service based on the combination of EMD and multivariate LSTM;
FIG. 2 is a general framework diagram of the method of the present invention based on the combination of EMD and multivariate LSTM for predicting QoS;
FIG. 3 is a response time box line diagram of the Web service "QuoteOfTheday" in an embodiment of the present invention;
fig. 4 is a schematic diagram of a process of extracting search _ Date _ tend for response time data of a Web service "Amazon" in an embodiment of the present invention, where dots represent original data distribution, and lines represent general trends of extracted QoS data;
fig. 5 is a schematic diagram of a process of extracting data _ home _ tree for response time data of a Web service "Amazon" in the embodiment of the present invention, where a circle indicates a data distribution after extracting the seaclean _ Date _ tree, and a curve indicates a Trend of the extracted QoS data within one day;
fig. 6 is a diagram illustrating a result of applying a data-based padding algorithm to perform missing value and outlier padding on QoS data of a Web service "Amazon" in the embodiment of the present invention, where dots indicate a result of the algorithm padding;
fig. 7 is a diagram showing a result of EMD decomposition of QoS data of the Web service "Amazon" in the embodiment of the present invention;
FIG. 8 is a schematic diagram of a multivariate input pattern constructed in an embodiment of the invention;
FIG. 9 is a block diagram of a multivariate LSTM model constructed in an embodiment of the invention;
FIG. 10 is a diagram of the actual logical architecture inside the multivariate LSTM neuron of FIG. 9 at time t;
figure 11 is a graphical illustration of historical QoS data versus prediction results,
fig. 12 is a diagram showing the result of applying the service quality prediction method based on EMD and multivariate LSTM to the QoS data of the Web service "Amazon" in the embodiment of the present invention, where the star line is the true value and the dotted line is the prediction result.
Detailed Description
The inventor of the application finds out through a great deal of research and practice that:
quality of service (QoS), which describes the non-functional characteristics of a Web service, including response time, throughput, service price, etc., becomes a determining factor in picking out the best service from among services providing similar functions. Generally, a service provider declares and exposes QoS data as part of a service description so that some static QoS data, such as service prices, etc., can be obtained therefrom or some dynamic QoS data, such as response time, throughput, etc., can be obtained through service caller sharing. The dynamic QoS data has a time-related characteristic during service operation and is not in a stable state, so that prediction is performed according to past observed values or QoS data shared by service callers, so that the requirements of dynamic service combinations in a Web service recommendation system and a service-oriented architecture on the QoS data are met, and the dynamic QoS data prediction method is one of the key problems to be solved in the field of service calculation.
Based on the consideration, the invention provides a service quality prediction method based on the combination of EMD and multivariable LSTM, which can accurately predict unknown service quality according to the historical calling record of Web service and provide basis for Web service recommendation and selection in a dynamic environment, thereby recommending optimal service for users and having good practicability.
In order to solve the above technical problems, the main concept of the present invention is as follows:
1: performing data cleaning on the Web service historical call records, and detecting missing values and abnormal values of Web service quality in a data set; 2: filling the missing value and the abnormal value by using a data-based filling algorithm, and constructing a complete and effective service quality time sequence; 3: performing data transformation on the time sequence to enable the transformed data to be more consistent with the assumption of a prediction model on data distribution; 4: and decomposing the service quality time sequence into a plurality of eigen-mode functions and residual wave parts by using an Empirical Mode Decomposition (EMD) method, then constructing a multivariate time sequence, and establishing a multivariate LSTM model for service quality prediction.
Furthermore, the invention also provides a Service Level Agreement (SLA) violation prediction mechanism based on the combination of EMD and multivariate LSTM, which can accurately predict future QoS values through a multivariate LSTM prediction model based on EMD and predict possible SLA violation behaviors, thereby helping a user select a service which can provide the best QoS and has higher probability to meet SLA constraints.
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The embodiment provides a service quality prediction method based on combination of EMD and multivariate LSTM, please refer to fig. 1, the method includes:
step S1: and (4) performing data cleaning on the Web service historical call record, and detecting abnormal values and missing values.
Specifically, the history of Web service invocation, i.e., the history of Web services, is used to predict the quality of service.
Step S2: and performing data completion on the detected abnormal values and the detected missing values by adopting a preset data-based filling algorithm to construct a complete and effective service quality time sequence.
Specifically, the abnormal value and the missing value of the data summary can be detected through step S1, and the abnormal value and the missing value are completed in this step in order to obtain complete quality of service data.
Step S3: and carrying out data transformation on the complete and effective service quality time sequence.
Specifically, the purpose of data transformation is to make the transformed data more consistent with the assumption of data distribution by the prediction model, and the existing data transformation method can be adopted.
Step S4: and decomposing the service quality time sequence subjected to data transformation based on an EMD method to obtain a multivariate time sequence, and obtaining training data according to the multivariate time sequence.
In particular, the EMD method, Empirical Mode Decomposition (EMD), is a new method for processing non-stationary signals, and the essence of this method is to identify all vibration modes (Intrinsic vibration modes) contained in the signals by means of a characteristic time scale. In this process, the feature time scale and the definition of the IMF are both empirical and approximate. Compared with other signal processing methods, the EMD method is intuitive, indirect, a posteriori, adaptive, with the characteristic time scale used for its decomposition being derived from the original signal. After obtaining the multivariate time sequence, carrying out scale transformation on the multivariate time sequence to ensure that the multivariate time sequence conforms to the activation function of the LSTM model.
Step S5: and training the pre-constructed LSTM model by using the training data to obtain the trained multivariate LSTM model.
Specifically, after high quality training data is obtained, a previously constructed LSTM model may be trained using the training data.
Step S6: and performing quality prediction on the Web service call record to be predicted by using the trained multivariate LSTM model, and selecting the target Web service according to the quality prediction result.
Specifically, the step is to specifically apply the trained multivariate LSTM model, predict the service quality by using the trained model, and finally select the target Web service with the quality meeting the requirements according to the prediction result.
In one embodiment, step S1 specifically includes:
step S1.1: identifying service abnormality in the Web service history calling record, screening out a preliminary abnormal value and marking the preliminary abnormal value as a missing value;
step S1.2: and detecting abnormal values in the Web service history calling record by using the box line graph, and taking observed values which are larger than or smaller than the upper and lower boundaries set by the box line graph in the detection result as target abnormal values and identifying the target abnormal values.
Specifically, the Web service history call record may have a service abnormality and data of a missing part of time period, and the preliminary abnormal value and the missing value are screened and identified by identifying whether the service is abnormal or not.
The box plot is a statistical graph used for displaying a group of data dispersion situation data, which has no special requirement on data distribution, and on the basis of S1.1, observed values which are larger than or smaller than the upper and lower bounds set by the box plot can be regarded as abnormal values and identified.
In a specific implementation process, the box plot is used to detect abnormal values, and the following formula is adopted:
{outlier}＝{X>Q3+1.5Q}∪{X<Q1-1.5Q}
outlier screening was performed, where Q1, Q3 represent the 25 th percentile (lower quartile), the 75 th percentile (upper quartile), and Q is the interquartile distance, i.e., the distance between Q3 and Q1, respectively, after all data were arranged from small to large.
In one embodiment, step S2 specifically includes:
step S2.1: extracting a change trend on a date level and a change trend within one day from normal data subjected to data cleaning, wherein the rest is residual partial data, and the service quality data subjected to data cleaning comprises normal data and abnormal data;
step S2.2: selecting basic time characteristics of different time granularities and historical lag values of service quality data from normal data as characteristics, using residual error part data as a label, training an Xgboost model, and predicting residual error parts of abnormal values and missing values by using the trained Xgboost model to obtain a residual error predicted value;
step S2.3: and adding the variation trend of the extracted normal data on the date level, the variation trend within one day and the residual prediction value of the abnormal data, and filling the abnormal value and the missing value to obtain a data filling result.
In particular, in step S2.1, trend descriptions are performed at different levels according to the temporal granularity of the normal data. First, looking at Seasonal changes of service quality data from the aspect of large Trend, fitting the data by using linear regression or curve regression according to the distribution of the data, extracting the Trend on the Date level of the data, recording the Trend as Seasonal _ Date _ Trend, and then utilizing Res of the residual part after the Seasonal _ Date _ Trend is extractedt(Rest1＝Originalt-Seasonal_Date_Trendt) To extract a more detailed Trend, similarly, linear regression or curve regression may be used to fit the Trend, and the Trend of the change of the quality of service data within one day is extracted and is denoted as Daily _ home _ tend, and the remaining data after the seasal _ Date _ tend and Daily _ home _ tend are extracted are:
Rest2＝Originalt-Seasonal_Date_Trendt-Daily_Hour_Trendt。
abnormal data are abnormal values and missing values.
Since the Residual part fluctuates around 0 after the search _ Date _ tend and the daisy _ Hour _ tend are extracted through Trend description, in the embodiment, the Xgboost model is used to train the Residual, basic time characteristics such as minute, Hour, weekday, month and the like and historical lag values step _1 and step _2.. step _ n of the service quality data are selected from normal data, the Residual of the normal data is used as a label, the Xgboost model is trained, and then the trained model is used to predict the Residual of the abnormal data and is marked as Residual _ Prediction (similarly, different basic time characteristics and historical lag values are extracted from the abnormal data and input into the trained Xgboost model, and the Residual Prediction value of the abnormal data can be obtained); for abnormal data (namely each missing value and each abnormal value), the Trend parts Seasonal _ Date _ Trend and Daily _ Hour _ Trend of the extracted normal data are added with the residual prediction value of the abnormal data to obtain a filling result.
In one embodiment, step S3 specifically includes:
step S3.1: carrying out Box _ Cox transformation on the complete and effective service quality time sequence;
step S3.2: and carrying out logarithmic transformation on the service quality time series subjected to the Box _ Cox transformation.
Specifically, Box _ Cox transformation is carried out on a time sequence, so that the correlation between an unobservable error and a prediction variable can be reduced to a certain extent; and the time sequence is further subjected to logarithmic transformation, so that the vibration amplitude of the data can be reduced, and the linear rule of the data is more obvious.
In a specific implementation, Box _ Cox transform is performed on the time sequence, using the following formula:
Figure BDA0002323751630000101
the data transformation mode is based on the data without any prior information, and the whole process is more objective and accurate compared with other transformation modes.
In one embodiment, step S4 specifically includes:
step S4.1: decomposing the service quality time sequence with irregular frequency into a plurality of frequency single IMF components and a residual part Res by using an EMD method;
step S4.2: carrying out scale transformation on the decomposed data, and converting the numerical values of all components of the service quality time sequence into the range of an activation function of the LSTM model;
step S4.3: and constructing a 3D tensor according to the multivariable time sequence subjected to scale transformation, wherein the form of the 3D tensor is (samples, times, features), the samples represent independent observation results from the domain and are data rows, the times represent that current values are related to data continuously input in the previous time step and are history lag values, and the features represent characteristics observed at each time point and are a plurality of IMF components and residual parts Res.
Specifically, IMF is an Intrinsic Mode Function (IMF), and Res is a Residual Waveform (Res). The EMD method for decomposing the quality of service time series is as follows: firstly, determining local extreme points of an original time sequence X (t); then, connecting maximum value points through cubic spline interpolation to form an upper envelope line, connecting minimum value points to form a lower envelope line, and calculating the average value m (t) of the upper envelope line and the lower envelope line; then, subtracting m (t) from X (t) to obtain candidate IMF; next, it is determined whether one of the shutdown criteria is satisfied: (1) x (t) is zero or not, (2) h (t) is zero or not, (1) the difference between the local extreme point and the zero crossing point in the whole time range is not more than 1, secondly, the average envelope curve is zero at any time), the maximum iteration number is reached, (3) the maximum iteration number is reached, (1) any item of (3) is satisfied, otherwise, if h (t) is not satisfied, the step (i) is repeated by taking h (t) as an initial signal (replacing X (t)), namely, the step (i) of determining the local extreme point of the original time sequence X (t), connecting the maximum point by cubic spline interpolation to form an upper envelope curve, connecting the minimum point to form a lower envelope curve, and calculating the average value m (t) of the upper and lower envelope curves, (m (t) is subtracted from X (t) to obtain a candidate IMF, the step of judging whether one of the shutdown criteria is satisfied or not, and if h (t) is satisfied, determining h (t) as a new IMF, subtracting h (t) from X (t) to obtain a residual signal r (t), using r (t) as a new initial signal (replacing X (t)), if the extreme point of r (t) is greater than 2, repeating the above steps (same as above), otherwise, using r (t) as a residual waveform to complete decomposition.
In step S4.2, the range of the activation function is [ -1,1], and in step 4.3, a plurality of IMFs and Res extracted by the EMD method are multivariate time series and further converted into a supervised learning problem.
In a specific implementation, taking into account the sensitivity of the LSTM model, the data is scaled, and the input is controlled to be in the range of [ -1,1] using the following equation:
Figure BDA0002323751630000111
after transformation, the data is within the range of the activation function of the LSTM model.
Wherein, x' is data after scale transformation, x is data to be processed, and x isminAnd xmaxRespectively, the minimum value and the maximum value of the data to be processed.
In one embodiment, the EMD method is used to decompose the time series into a set of eigenmode functions and a residual waveform in step S4.1, which results in the following equation:
Figure BDA0002323751630000112
wherein X (t) represents the original time series, IMFn(t) denotes the nth IMF component, and Res (t) denotes the residual wave portion.
In one embodiment, the previously constructed LSTM model includes two long-term and short-term memory networks and a fully connected layer, and step S5 specifically includes:
step S5.1: inputting the multivariate time sequence into a pre-constructed LSTM model, and learning high-level time sequence representation through a two-layer long-time and short-time memory network;
step S5.2: and transmitting the last step of the output sequence to a full-connection layer to realize the prediction of the next time step and obtain the trained multivariate LSTM model.
In one embodiment, the calculation procedure of the long and short term memory network in step S6 is as follows:
step S6.1: selecting information discarded from the cell state by forgetting gate and determining that the previous long-term memory c is to be usedt-1To what extent c is savedtThe method comprises the following steps:
Figure BDA0002323751630000121
step S6.2: new information saved into the long-term memory state is determined by the input gate:
Figure BDA0002323751630000122
step S6.3: calculating a candidate value for the long-term state based on determining new information to be saved into the long-term memory state:
Figure BDA0002323751630000123
step S6.4: updating the long-term state with the candidate values:
Figure BDA0002323751630000124
step S6.5: determine the current LSTM output value through the output gate:
Figure BDA0002323751630000125
wherein the hidden layer state htThe calculation formula of (a) is as follows:
Figure BDA0002323751630000126
sigma represents a sigmoid function, ⊙ represents a bitwise multiplication operation, and the calculation formula of the sigmoid function and the tanh function is as follows:
Figure BDA0002323751630000127
Figure BDA0002323751630000128
m represents the dimension of the input variable accepted by the LSTM, u represents that the LSTM possesses u neurons,
Figure BDA0002323751630000129
Figure BDA00023237516300001210
and
Figure BDA00023237516300001211
respectively showing the states of a forgetting gate, an input gate and a control gate of the LSTM at the time t, all being u-dimensional vectors,
Figure BDA00023237516300001212
and
Figure BDA00023237516300001213
the output and the cell state of the LSTM at time t are u-dimensional vectors, V is a weight matrix in a hidden state, is a u × u matrix, W is an input weight matrix, is a u × m matrix, and b is an offset term.
In one embodiment, after obtaining the prediction, the method further comprises: and judging whether the prediction result meets the SLA constraint of the service level agreement, if not, determining that the SLA violates the rules, and if so, not violating the rules.
The following is a specific embodiment of performing service quality prediction by applying the method of the present invention, taking QoS prediction based on the combination of EMD and multivariate LSTM performed on ten QoS data sets of actual Web services collected by rice cavalo in empirical research as an embodiment, and describing in detail the implementation process of the present invention with reference to the accompanying drawings, wherein fig. 2 is a general framework schematic diagram of the service quality prediction method based on the combination of EMD and multivariate LSTM in the present invention.
The QoS dataset for ten real Web services was derived from a empirical study by Bice Cavallo, calling 10 real services every hour for up to 4 months, and then recording 1) the time of calling the services, 2) the SOAP output message, 3) the observed response time, 4) the throughput, i.e., the size of the SOAP message divided by the response time, 5) whether the requested service is abnormal or not.
First, step S101 (corresponding to step S1 above) is executed to read the original QoS data from the data source, determine whether the QoS data at the current time point is missing according to whether the requested service is abnormal, and complete the detection and marking of the missing value.
Outliers in the QoS data are then detected using the box plots, and outliers greater than or less than the upper and lower bounds set by the box plots are identified and flagged. Fig. 3 illustrates a process of detecting abnormal values of response time of the service "QuoteOfTheDay" using a box plot, in which there are many sharp outliers, meaning that there are particularly large abnormal values in the original data.
Then, step S102 (corresponding to step S2 above) is executed, a data-based padding algorithm is applied to fill in missing values and abnormal values, and a complete and effective qos time series is constructed;
firstly, trend description is carried out on different levels according to the time granularity of data, and the trend description is divided into two steps: the method comprises the steps that firstly, Seasonal changes of QoS data are extracted from a Date level, the data are rolled up to the Date level in a time dimension, then linear regression is used for fitting the data according to the distribution of the data, the general Trend of the data on the Date level is extracted and is marked as Seasonal _ Date _ Trend, the process of extracting the Seasonal change Trend of the QoS data of Web service Amazon is shown in figure 4, the origin of the process represents the distribution of real data, and a line is the Trend obtained through linear regression fitting; and secondly, extracting a more detailed Trend by using the residual error, fitting the Trend by using curve regression, and extracting a variation Trend of the QoS data within one day, wherein the variation Trend is marked as Daily _ Hour _ Trend, and a process of extracting the Daily _ Hour _ Trend from the QoS data of the Web service Amazon is shown in FIG. 5, wherein a circle point represents an average value per Hour, and a curve is a Trend obtained by curve regression fitting.
Then, the Xgboost model is used for training the QoS data to extract residual parts after Seasonal _ Date _ Trend and Daily _ Hour _ Trend through Trend description, the residual parts basically fluctuate around 0, four basic time characteristics of Hour, weekday, Date and month and 12 historical lag values step _1 and step _2.
And finally, adding the result predicted by the Xgboost model and the extracted trend part to obtain a filling result of the missing value and the abnormal value. Fig. 6 shows a result of applying a data-based padding algorithm to QoS data of Web service "Amazon" to fill missing values and abnormal values, where dots represent padding results of the algorithm, lines represent original data, and results of the algorithm more conform to fluctuation characteristics of an original time series.
Next, step 103 is executed (corresponding to step S3 above), and the time series of the quality of service of the Web service is subjected to data transformation so that the transformed data more conforms to the assumption of the model on the data distribution. Firstly, the Box _ Cox transformation is carried out on the QoS data to reduce the correlation between the unobservable error and the prediction variable to a certain extent, then, the logarithmic transformation is carried out on the QoS data to lead the sequence to be stable,
finally, step S104 (corresponding to the above steps S4-S6) is executed to decompose the QoS time sequence into a plurality of eigenmode functions and a residual waveform by using an EMD method, and then, a multivariate time sequence is constructed and a multivariate LSTM is trained to predict the service quality;
first, an EMD method is applied to an original QoS time series to decompose the time series into a plurality of IMFs and a residual wave, and fig. 7 shows a result of EMD decomposition of QoS data of a Web service "Amazon".
Then, taking into account the sensitivity of the LSTM model, the input data is scaled to transform the values of all components between [ -1,1 ].
Then, a multivariate time sequence is constructed by using a plurality of IMFs and Res extracted by EMD, the multivariate time sequence is converted into a supervision problem, a 3D tensor (samples, time ranges and features) is constructed, the samples are the number of samples, the time ranges represent the number of data learning data distribution rules of how many data are seen backwards, 12 historical lag values are selected, and the features observed at each time point are a plurality of IMF components and Res parts. Fig. 8 shows a schematic diagram of a multivariate input pattern.
And finally, constructing a multivariate LSTM model to solve the multivariate time sequence prediction problem, wherein the model receives a mixed sequence consisting of IMFs and Res obtained by EMD as multivariate input, and outputs a QoS predicted value of the next time step. FIG. 9 shows a block diagram of a multivariate LSTM model that stacks two layers of LSTMs to learn high-level timing characterization, the first layer returning the full output sequence h1,h2,h3…htThe second layer only returns the last step h of the output sequencetAnd abandoning a time sequence dimension to convert the input sequence into a single vector and transmitting the single vector to a full-connection layer which has only one neuron and is provided with a linear activation function so as to realize the prediction of the next time step. FIG. 10 shows the actual logical architecture inside a multivariate LSTM neuron at time t.
FIG. 11 is a diagram showing the relationship between historical QoS data and prediction results, first, using a historical QoS time series x1,x2,x3…xtThe result of EMD decomposition forms a multivariate sequence, and then the next time step y is predicted through a multivariate LSTM modelt+1. Fig. 12 is a result diagram of service quality prediction based on the combination of EMD and multivariate LSTM for QoS data of Web service "Amazon", where the star line is a true value and the dotted line is a prediction result, and the model can predict the service quality more accurately. Table 1 shows the statistics of the results of service quality prediction based on the combination of EMD and multivariate LSTM for ten actual Web services' QoS data. After the prediction of unknown service quality is completed, SLA violation behaviors which may occur are further predicted, and table 2 is a result of performing EMD-based multivariate LSTM service quality SLA violation behavior prediction on QoS data sets of ten actual Web services, and records the percentage of correct prediction of SLA violation by EMD-based multivariate LSTM prediction models for each service and statistics of average performance.
TABLE 1 statistics of forecasts across QoS datasets for ten actual Web services
Figure BDA0002323751630000151
TABLE 2 SLA violation prediction statistics on QoS datasets for ten actual Web services
Figure BDA0002323751630000152
Figure BDA0002323751630000161
The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments or alternatives may be employed by those skilled in the art without departing from the spirit or ambit of the invention as defined in the appended claims.

Claims (10)

1. A service quality prediction method based on the combination of EMD and multivariate LSTM is characterized by comprising the following steps:
step S1: carrying out data cleaning on the Web service historical calling record, and detecting an abnormal value and a missing value;
step S2: performing data completion on the detected abnormal values and the detected missing values by adopting a preset data-based filling algorithm to construct a complete and effective service quality time sequence;
step S3: carrying out data transformation on the complete and effective service quality time sequence;
step S4: decomposing the service quality time sequence subjected to data transformation based on an EMD method to obtain a multivariate time sequence, and obtaining training data according to the multivariate time sequence;
step S5: training a pre-constructed LSTM model by using training data to obtain a trained multivariate LSTM model;
step S6: and performing quality prediction on the Web service call record to be predicted by using the trained multivariate LSTM model, and selecting the target Web service according to the quality prediction result.
2. The method according to claim 1, wherein step S1 specifically comprises:
step S1.1: identifying service abnormality in the Web service history calling record, screening out a preliminary abnormal value and marking the preliminary abnormal value as a missing value;
step S1.2: and detecting abnormal values in the Web service history calling record by using the box line graph, and taking observed values which are larger than or smaller than the upper and lower boundaries set by the box line graph in the detection result as target abnormal values and identifying the target abnormal values.
3. The method according to claim 2, wherein step S2 specifically comprises:
step S2.1: extracting a change trend on a date level and a change trend within one day from normal data subjected to data cleaning, and the balance being residual partial data, wherein the service quality data subjected to data cleaning comprises normal data and abnormal data;
step S2.2: selecting basic time characteristics of different time granularities and historical lag values of service quality data from normal data as characteristics, using residual error part data as a label, training an Xgboost model, and predicting residual error parts of abnormal values and missing values by using the trained Xgboost model to obtain a residual error predicted value;
step S2.3: and adding the variation trend of the extracted normal data on the date level, the variation trend within one day and the residual prediction value of the abnormal data, and filling the abnormal value and the missing value to obtain a data filling result.
4. The method according to claim 1, wherein step S3 specifically comprises:
step S3.1: carrying out Box _ Cox transformation on the complete and effective service quality time sequence;
step S3.2: and carrying out logarithmic transformation on the service quality time series subjected to the Box _ Cox transformation.
5. The method according to claim 1, wherein step S4 specifically comprises:
step S4.1: decomposing the service quality time sequence with irregular frequency into a plurality of frequency single IMF components and a residual part Res by using an EMD method;
step S4.2: carrying out scale transformation on the decomposed data, and converting numerical values of all components of the service quality time sequence into the range of an activation function of a multivariate LSTM model;
step S4.3: and constructing a 3D tensor according to the multivariable time sequence subjected to scale transformation, wherein the form of the 3D tensor is (samples, times, features), the samples represent independent observation results from the domain and are data rows, the times represent that current values are related to data continuously input in the previous time step and are history lag values, and the features represent characteristics observed at each time point and are a plurality of IMF components and residual parts Res.
6. The method of claim 5, wherein the EMD method is used to decompose the time series into a set of eigenmode functions and a residual waveform in step S4.1, resulting in the following equation:
Figure FDA0002323751620000021
wherein X (t) represents the original time series, IMFn(t) denotes the nth IMF component, and Res (t) denotes the residual wave portion.
7. The method of claim 1, wherein the pre-constructed LSTM model includes two layers of long-and-short memory networks and a fully-connected layer, and step S5 specifically includes:
step S5.1: inputting the multivariate time sequence into a pre-constructed LSTM model, and learning high-level time sequence representation through a two-layer long-time and short-time memory network;
step S5.2: and transmitting the last step of the output sequence to a full-connection layer to realize the prediction of the next time step and obtain the trained multivariate LSTM model.
8. The method according to claim 4, characterized in that in step S3.1 the Box _ Cox transform is performed on the time sequence, using the following formula:
Figure FDA0002323751620000022
performing a transformation, wherein λ represents an introduced parameter, Y represents a time series to be transformed, Y(λ)Representing the transformed result.
9. The method as claimed in claim 7, wherein the calculation process of the long and short term memory network in step S6 is:
step S6.1: selecting information discarded from the cell state by forgetting gate and determining that the previous long-term memory c is to be usedt-1To what extent c is savedtThe method comprises the following steps:
Figure FDA0002323751620000031
step S6.2: new information saved into the long-term memory state is determined by the input gate:
Figure FDA0002323751620000032
step S6.3: calculating a candidate value for the long-term state based on determining new information to be saved into the long-term memory state:
Figure FDA0002323751620000033
step S6.4: updating the long-term state with the candidate values:
Figure FDA0002323751620000034
step S6.5: determine the current LSTM output value through the output gate:
Figure FDA0002323751620000035
the calculation formula of the hidden layer state ht is as follows:
Figure FDA0002323751620000036
sigma represents a sigmoid function, ⊙ represents a bitwise multiplication operation, and the calculation formula of the sigmoid function and the tanh function is as follows:
Figure FDA0002323751620000037
Figure FDA0002323751620000038
m represents the dimension of the input variable accepted by the LSTM, u represents that the LSTM possesses u neurons,
Figure FDA0002323751620000039
Figure FDA00023237516200000310
and
Figure FDA00023237516200000311
respectively showing the states of a forgetting gate, an input gate and a control gate of the LSTM at the time t, all being u-dimensional vectors,
Figure FDA00023237516200000312
and
Figure FDA00023237516200000313
the output and the cell state of the LSTM at time t are u-dimensional vectors, V is a weight matrix in a hidden state, is a u × u matrix, W is an input weight matrix, is a u × m matrix, and b is an offset term.
10. The method of claim 1, wherein after obtaining the prediction, the method further comprises: and judging whether the prediction result meets the SLA constraint of the service level agreement, if not, determining that the SLA violates the rules, and if so, not violating the rules.
CN201911308215.9A 2019-12-18 2019-12-18 Service quality prediction method based on combination of EMD and multivariate LSTM Active CN111131424B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911308215.9A CN111131424B (en) 2019-12-18 2019-12-18 Service quality prediction method based on combination of EMD and multivariate LSTM

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911308215.9A CN111131424B (en) 2019-12-18 2019-12-18 Service quality prediction method based on combination of EMD and multivariate LSTM

Publications (2)

Publication Number Publication Date
CN111131424A true CN111131424A (en) 2020-05-08
CN111131424B CN111131424B (en) 2020-12-18

Family

ID=70499604

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911308215.9A Active CN111131424B (en) 2019-12-18 2019-12-18 Service quality prediction method based on combination of EMD and multivariate LSTM

Country Status (1)

Country Link
CN (1) CN111131424B (en)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111914873A (en) * 2020-06-05 2020-11-10 华南理工大学 Two-stage cloud server unsupervised anomaly prediction method
CN112162860A (en) * 2020-09-25 2021-01-01 中国民航大学 CPU load trend prediction method based on IF-EMD-LSTM
CN112231307A (en) * 2020-09-18 2021-01-15 苏州冷王网络科技有限公司 Cold chain monitoring temperature and humidity data fitting method based on LSTM
CN112532429A (en) * 2020-11-11 2021-03-19 北京工业大学 Multivariable QoS prediction method based on position information
CN113177672A (en) * 2021-05-28 2021-07-27 阜新博远科技有限公司 PM2.5 prediction method based on long-time and short-time memory network
CN113688771A (en) * 2021-09-02 2021-11-23 重庆大学 LNG storage tank acceleration response data completion method and device
CN114037551A (en) * 2021-11-15 2022-02-11 中国水产科学研究院渔业机械仪器研究所 Pond culture pH value missing data interpolation method
CN114826988A (en) * 2021-01-29 2022-07-29 中国电信股份有限公司 Method and device for anomaly detection and parameter filling of time sequence data
CN117807380A (en) * 2024-01-02 2024-04-02 中国科学院西北生态环境资源研究院 Time sequence data complement method and device, storage medium and electronic equipment

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP3908627B2 (en) * 2002-08-21 2007-04-25 日本電信電話株式会社 Web page transfer time estimation device, Web page transfer time estimation program, and computer readable recording medium recording Web page transfer time estimation program
US9864704B2 (en) * 2014-04-03 2018-01-09 SK Hynix Inc. Memory controller communicating with host, operating method thereof, and computing system including the same
CN109131452A (en) * 2018-08-14 2019-01-04 西安理工大学 A kind of train status on-line prediction method based on long memory network in short-term
CN109257217A (en) * 2018-09-19 2019-01-22 河海大学 Web service QoS prediction technique based on secret protection under mobile peripheral surroundings
CN109919082A (en) * 2019-03-05 2019-06-21 东南大学 Modal identification method based on LSTM and EMD
CN110362772A (en) * 2019-06-11 2019-10-22 北京邮电大学 Real-time webpage method for evaluating quality and system based on deep neural network

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP3908627B2 (en) * 2002-08-21 2007-04-25 日本電信電話株式会社 Web page transfer time estimation device, Web page transfer time estimation program, and computer readable recording medium recording Web page transfer time estimation program
US9864704B2 (en) * 2014-04-03 2018-01-09 SK Hynix Inc. Memory controller communicating with host, operating method thereof, and computing system including the same
CN109131452A (en) * 2018-08-14 2019-01-04 西安理工大学 A kind of train status on-line prediction method based on long memory network in short-term
CN109257217A (en) * 2018-09-19 2019-01-22 河海大学 Web service QoS prediction technique based on secret protection under mobile peripheral surroundings
CN109919082A (en) * 2019-03-05 2019-06-21 东南大学 Modal identification method based on LSTM and EMD
CN110362772A (en) * 2019-06-11 2019-10-22 北京邮电大学 Real-time webpage method for evaluating quality and system based on deep neural network

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
ZAIFA CHEN,ET.AL: "Mechanical State Prediction Based on LSTM Neural Network", 《IEEE》 *

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111914873A (en) * 2020-06-05 2020-11-10 华南理工大学 Two-stage cloud server unsupervised anomaly prediction method
CN112231307A (en) * 2020-09-18 2021-01-15 苏州冷王网络科技有限公司 Cold chain monitoring temperature and humidity data fitting method based on LSTM
CN112162860A (en) * 2020-09-25 2021-01-01 中国民航大学 CPU load trend prediction method based on IF-EMD-LSTM
CN112532429A (en) * 2020-11-11 2021-03-19 北京工业大学 Multivariable QoS prediction method based on position information
CN112532429B (en) * 2020-11-11 2023-01-31 北京工业大学 Multivariable QoS prediction method based on position information
CN114826988A (en) * 2021-01-29 2022-07-29 中国电信股份有限公司 Method and device for anomaly detection and parameter filling of time sequence data
CN113177672A (en) * 2021-05-28 2021-07-27 阜新博远科技有限公司 PM2.5 prediction method based on long-time and short-time memory network
CN113177672B (en) * 2021-05-28 2023-12-15 阜新博远科技有限公司 PM2.5 prediction method based on long-short-time memory network
CN113688771A (en) * 2021-09-02 2021-11-23 重庆大学 LNG storage tank acceleration response data completion method and device
CN113688771B (en) * 2021-09-02 2024-04-05 重庆大学 LNG storage tank acceleration response data supplementing method and device
CN114037551A (en) * 2021-11-15 2022-02-11 中国水产科学研究院渔业机械仪器研究所 Pond culture pH value missing data interpolation method
CN117807380A (en) * 2024-01-02 2024-04-02 中国科学院西北生态环境资源研究院 Time sequence data complement method and device, storage medium and electronic equipment

Also Published As

Publication number Publication date
CN111131424B (en) 2020-12-18

Similar Documents

Publication Publication Date Title
CN111131424B (en) Service quality prediction method based on combination of EMD and multivariate LSTM
Kumar et al. A big data driven framework for demand-driven forecasting with effects of marketing-mix variables
Krishna et al. Sales-forecasting of retail stores using machine learning techniques
US20200311615A1 (en) Systems and methods for forecast alerts with programmable human-machine hybrid ensemble learning
Yerpude et al. Impact of internet of things (IoT) data on demand forecasting
US20110153419A1 (en) System and method for intelligent modeling for insurance marketing
KR20050007309A (en) Automatic neural-net model generation and maintenance
US11501107B2 (en) Key-value memory network for predicting time-series metrics of target entities
CN114971748B (en) Prediction data generation method, model training method, computer device, and storage medium
CN115965463A (en) Model training method and device, computer equipment and storage medium
Mansouri et al. The moderating role of master production scheduling method on throughput in job shop systems
Lepenioti et al. Human-augmented prescriptive analytics with interactive multi-objective reinforcement learning
CN113469482A (en) Method, apparatus, and computer-readable storage medium for data analysis
Jadli et al. A Novel LSTM-GRU-Based Hybrid Approach for Electrical Products Demand Forecasting.
Shetty et al. A comparative study of automobile sales forecasting with ARIMA, SARIMA and deep learning LSTM model
Hota et al. Prediction of customer churn in telecom industry: a machine learning perspective
Borysenko et al. Intelligent forecasting in multi-criteria decision-making.
Sarhani et al. Forecasting demand with support vector regression technique incorporating feature selection in the presence of calendar effect
CN114118570A (en) Service data prediction method and device, electronic equipment and storage medium
Vaitkus et al. Electrical spare parts demand forecasting
Hoffmann et al. Forecasting irregular demand using single hidden layer neural networks
Lee et al. Design and development of inventory knowledge discovery system
Chen et al. A fuzzy polynomial fitting and mathematical programming approach for enhancing the accuracy and precision of productivity forecasting
Krishna et al. Identifying demand forecasting using machine learning for business intelligence
Qureshi et al. A comparative analysis of traditional SARIMA and machine learning models for CPI data modelling in Pakistan

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant