Time sequence prediction method and device, equipment and medium

Abstract

The embodiment of the invention provides a time sequence prediction method and device, equipment and a medium, and the method specifically comprises the steps: decomposing historical time sequence data, so as to obtain k components corresponding to the historical time sequence data; determining an error evaluation index of the ith component corresponding to the historical time sequence data under the jth neural network; selecting the neural network with the minimum error evaluation index from the M neural networks as a target neural network corresponding to the ith component; according to the data of the ith component in the historical time sequence data, performing optimization processing on the hyper-parameter of the target neural network; determining component prediction results respectively corresponding to k components of the to-be-predicted time sequence data by using the trained target neural network; and reconstructing the component prediction results corresponding to the k components to obtain a prediction result corresponding to the to-be-predicted time sequence data. According to the embodiment of the invention, the accuracy and reliability of time sequence prediction can be improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of computer technology, and in particular to a method, apparatus, device and medium for predicting a time series. Background Art

[0002] With the continuous development of computer technology, time series prediction technology has emerged. It can predict the development trend of future things and bring great convenience to users' work and life. Time series prediction technology is based on an ordered observation data set that is closely related to the time sequence. With the help of random process theory and mathematical statistics methods, it deeply studies the statistical laws followed by the data set, and then predicts the development trend of the data, providing powerful guidance for solving practical problems. At present, time series prediction and analysis technology has been widely used in many technical fields such as stock market prediction, urban traffic management, equipment failure warning, power generation prediction, etc.

[0003] In related technologies, neural network technology is usually used to predict time series. The corresponding prediction process specifically covers the following aspects: First, collect and organize historical time series data. Then, select a neural network architecture to build a prediction model. Common neural network architectures include: RNN (Recurrent Neural Network), LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Unit). Then, the prediction model is trained with the help of historical time series data. In addition, the time series data to be predicted is input into the prediction model to obtain the prediction result of the time series data to be predicted.

[0004] However, since different types of time series data to be predicted have different characteristics, a single neural network architecture is difficult to fully adapt to each type of time series data to be predicted. This largely restricts the generalization ability of the prediction model, resulting in a decrease in the accuracy and reliability of the prediction model when facing different types of time series data to be predicted. Summary of the invention

[0005] The embodiment of the present application provides a time series prediction method, which can improve the accuracy and reliability of time series prediction.

[0006] Correspondingly, an embodiment of the present application also provides a time series prediction device, an electronic device and a machine-readable medium to ensure the implementation and application of the above method.

[0007] In order to solve the above problems, the embodiment of the present application discloses a time series prediction method, the method comprising: obtaining historical time series data;

[0008] Decompose the historical time series data to obtain k components corresponding to the historical time series data;

[0009] Integrate M neural networks in the neural network integration pool; k and M are both positive integers greater than 1;

[0010] Determine the error evaluation index of the i-th component corresponding to the historical time series data under the j-th neural network;

[0011] Select the neural network with the smallest error evaluation index from the M neural networks as the target neural network corresponding to the i-th component; i is a positive integer;

[0012] According to the data of the i-th component in the historical time series data, the hyperparameters of the target neural network are optimized;

[0013] According to the data of the i-th component in the historical time series data, the optimized target neural network is trained;

[0014] Using the trained target neural network, determine the component prediction results corresponding to the k components of the time series data to be predicted;

[0015] The component prediction results corresponding to the k components are reconstructed to obtain the prediction results corresponding to the time series data to be predicted.

[0016] In order to solve the above problems, the present application discloses a time series prediction device, which includes:

[0017] Acquisition module, used to obtain historical time series data;

[0018] A decomposition module is used to decompose the historical time series data to obtain k components corresponding to the historical time series data;

[0019] An integration module, used for integrating M neural networks in a neural network integration pool; k and M are both positive integers greater than 1;

[0020] An error determination module is used to determine the error evaluation index of the i-th component corresponding to the historical time series data under the j-th neural network;

[0021] A selection module is used to select a neural network with the smallest error evaluation index from M neural networks as a target neural network corresponding to the i-th component; i is a positive integer;

[0022] The optimization module is used to optimize the hyperparameters of the target neural network according to the data of the i-th component in the historical time series data;

[0023] A training module is used to train the optimized target neural network according to the data of the i-th component in the historical time series data;

[0024] The prediction module is used to determine the component prediction results corresponding to the k components of the time series data to be predicted by using the trained target neural network;

[0025] The reconstruction module is used to reconstruct the component prediction results corresponding to the k components to obtain the prediction results corresponding to the time series data to be predicted.

[0026] Optionally, the optimization module includes:

[0027] A verification sample acquisition module is used to obtain a verification sample from a verification set corresponding to the historical time series data; the verification sample corresponds to the first true value of the i-th component at the prediction time point;

[0028] A first prediction value determination module, used to determine, using the target neural network, a first prediction value of the i-th component corresponding to the verification sample at a preset time point and under current hyperparameters of the target neural network;

[0029] A fitness value determination module, used to determine the fitness value corresponding to the current hyperparameter according to the first true value and the first predicted value;

[0030] The end module is used to end the optimization if the fitness value meets the preset conditions;

[0031] The update module is used to update the current hyperparameters according to the optimization algorithm if the fitness value does not meet the preset conditions.

[0032] Optionally, the decomposition module includes:

[0033] A preprocessing module is used to preprocess the historical time series data; the preprocessing includes: outlier removal and missing value filling;

[0034] A historical decomposition module, used for decomposing the preprocessed historical time series data to obtain first initial values ​​of k components;

[0035] The normalization module is used to normalize the first initial values ​​of the k components to obtain first normalized values ​​of the k components.

[0036] Optionally, the decomposition module includes:

[0037] A first noise adding module is used to add N kinds of white noise to the historical time series data respectively to obtain N first constructed sequences;

[0038] A first decomposition module is used to decompose each first construction sequence using an empirical mode decomposition device, and determine a first group of residual sequences and first-order components according to a local mean of the decomposition results of the N first construction sequences;

[0039] The second noise adding module is used to add N kinds of white noise to the m-1th group of residual sequences respectively to obtain N second constructed sequences; N and m are positive integers;

[0040] The second decomposition module is used to decompose each second construction sequence using an empirical mode decomposition device, and determine the mth group of residual sequences and the mth order component according to the local mean of the decomposition results of the N second construction sequences.

[0041] Optionally, the error determination module is configured to include:

[0042] A verification sample acquisition module is used to obtain a verification sample from a verification set corresponding to the historical time series data; the verification sample corresponds to the first true value of the i-th component at the prediction time point;

[0043] A second prediction value determination module, used to determine the second prediction value of the i-th component corresponding to the verification sample at a preset time point and under the j-th neural network using the j-th neural network;

[0044] The error index determination module is used to determine the error evaluation index of the i-th component corresponding to the verification sample under the j-th neural network according to the first true value and the second predicted value.

[0045] Optionally, the reconstruction module includes:

[0046] A denormalization module, used for performing denormalization processing on the component prediction results corresponding to the k components respectively, so as to obtain the component denormalized prediction results corresponding to the k components respectively;

[0047] The fusion module is used to fuse the component denormalization prediction results corresponding to the k components to obtain the prediction results corresponding to the time series data to be predicted.

[0048] Optionally, the M neural networks include: an extreme learning machine, a radial basis function neural network and a convolutional neural network.

[0049] The embodiment of the present application further discloses an electronic device, including: a processor; and a memory, on which executable code is stored. When the executable code is executed, the processor executes the method described in the embodiment of the present application.

[0050] The embodiment of the present application further discloses a machine-readable medium on which executable code is stored. When the executable code is executed, a processor executes the method described in the embodiment of the present application.

[0051] The embodiments of the present application include the following advantages:

[0052] In the technical solution of the embodiment of the present application, the historical time series data is decomposed to obtain k components that can reflect the trend, cycle and other characteristics in the historical time series data. The above k components can help the neural network learn the different characteristics in the historical time series data more specifically, thereby improving the prediction accuracy of the time series data.

[0053] Moreover, the embodiment of the present application integrates a variety of neural networks in the neural network integration pool, and different neural network architectures have different adaptability to different types of features. On this basis, the embodiment of the present application determines the error evaluation index of each component of the historical time series data under different neural networks, and selects the neural network with the smallest error evaluation index as the target neural network corresponding to the component. Therefore, the embodiment of the present application can select the target neural network that best matches each component, thereby improving the accuracy of component prediction.

[0054] In addition, the embodiment of the present application optimizes the hyperparameters of the target neural network according to the data of specific components in the historical time series data. The optimization process can improve the generalization ability of the target neural network and reduce the risk of overfitting and underfitting by automatically searching for the optimal hyperparameter combination, thereby enhancing the stability and prediction effect of the target neural network. For example, a suitable learning rate can speed up the convergence speed of the target neural network, and a suitable number of layers and neurons can improve the expressive power of the target neural network.

[0055] Furthermore, the embodiment of the present application uses data of specific components in historical time series data to train the optimized target neural network, so that the target neural network can fully learn the patterns and laws of the specific components, and further improve the accuracy of prediction of the component.

[0056] Furthermore, the embodiment of the present application utilizes the trained target neural network to determine the component prediction results corresponding to the k components of the time series data to be predicted, and reconstructs the prediction results of each component. The above reconstruction can integrate the component prediction results obtained for different component predictions to obtain the final prediction result of the time series data to be predicted.

[0057] In summary, the embodiments of the present application effectively improve the accuracy and reliability of time series prediction through a series of steps such as data decomposition, neural network integration, hyperparameter optimization, targeted training and result reconstruction. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 This is a schematic diagram of the steps of a time series prediction method according to an embodiment of the present application;

[0059] Figure 2 This is a schematic diagram of a process for decomposing historical time series data according to an embodiment of the present application;

[0060] Figure 3 is a schematic diagram of a decomposition result of an embodiment of the present application;

[0061] Figure 4 It is a flowchart of a time series prediction method according to an embodiment of the present application;

[0062] Figure 5 It is a schematic diagram of the structure of a time series prediction device according to an embodiment of the present application;

[0063] Figure 6 It is a schematic diagram of the structure of a device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0064] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0065] The time series prediction method of the embodiment of the present application can be applied to many technical fields such as stock market prediction, urban traffic management, equipment failure warning, network security, energy industry, supply chain management, etc.

[0066] For example, in the field of network security, the time series prediction method of the embodiment of the present application can be used for network traffic prediction, anomaly detection, security event prediction, strengthening network security protection, preventing DdoS attacks (Distributed Denial of Service attack), and ensuring the security of critical infrastructure.

[0067] For example, in the energy industry, the time series prediction method of the embodiment of the present application can be used for power load forecasting, new energy power generation power forecasting, power generation monitoring and early warning, energy price forecasting, optimizing energy distribution, reducing waste, improving grid stability, and supporting green energy policies.

[0068] For example, in the field of supply chain management, the time series prediction method of the embodiment of the present application can be used for demand forecasting, inventory management, logistics optimization, and supplier selection. By accurately predicting demand, reducing inventory costs, avoiding overproduction and out-of-stock, and improving supply chain efficiency.

[0069] Alternatively, in urban traffic management, the time series prediction method of the embodiment of the present application can be used for traffic flow prediction, air quality prediction, public facility demand prediction, improving urban planning, reducing traffic congestion, improving public service efficiency, and enhancing the quality of life of residents.

[0070] It can be understood that the above-mentioned technical fields are only used as examples, and those skilled in the art can apply the prediction method of the time series of the embodiment of the present application to any technical field according to actual application requirements. For example, the embodiment of the present application can also be used for weather forecasting and natural disaster warning, providing timely warnings for the government and the public, and reducing the loss of life and property. For another example, the embodiment of the present application can also be used in the field of education and human resources, and can assist in educational decision-making, optimize human resource allocation, and improve organizational performance. For another example, the embodiment of the present application can also bring direct economic benefits to enterprises and institutions, and can also indirectly promote the overall progress and development of society by improving decision-making quality and response speed.

[0071] In related technologies, neural network technology is usually used to predict time series. The corresponding prediction process specifically covers the following aspects: First, collect and organize historical time series data. Then, select a neural network architecture to build a prediction model. Then, train the prediction model with the help of historical time series data. In addition, input the time series data to be predicted into the prediction model to obtain the prediction result of the time series data to be predicted.

[0072] However, since different types of time series data to be predicted have different characteristics, a single neural network architecture is difficult to fully adapt to each type of time series data to be predicted. This largely restricts the generalization ability of the prediction model, resulting in a decrease in the accuracy and reliability of the prediction model when facing different types of time series data to be predicted.

[0073] In view of the technical problems of low accuracy and low reliability of time series prediction in related technologies, an embodiment of the present application provides a time series prediction method, which may specifically include the following steps:

[0074] Get historical time series data;

[0075] Decompose the historical time series data to obtain k components corresponding to the historical time series data;

[0076] Integrate M neural networks in the neural network integration pool; k and M are both positive integers greater than 1;

[0077] Determine the error evaluation index of the i-th component corresponding to the historical time series data under the j-th neural network;

[0078] Select the neural network with the smallest error evaluation index from the M neural networks as the target neural network corresponding to the i-th component; i is a positive integer;

[0079] According to the data of the i-th component in the historical time series data, the hyperparameters of the target neural network are optimized;

[0080] According to the data of the i-th component in the historical time series data, the optimized target neural network is trained;

[0081] Using the trained target neural network, determine the component prediction results corresponding to the k components of the time series data to be predicted;

[0082] The component prediction results corresponding to the k components are reconstructed to obtain the prediction results corresponding to the time series data to be predicted.

[0083] First, the embodiment of the present application decomposes the historical time series data, and the obtained k components can reflect the trend, cycle and other characteristics in the historical time series data. The above k components can help the neural network learn the different characteristics in the historical time series data more specifically, thereby improving the prediction accuracy of the time series data.

[0084] Moreover, the embodiment of the present application integrates a variety of neural networks in the neural network integration pool, and different neural network architectures have different adaptability to different types of features. On this basis, the embodiment of the present application determines the error evaluation index of each component of the historical time series data under different neural networks, and selects the neural network with the smallest error evaluation index as the target neural network corresponding to the component. Therefore, the embodiment of the present application can select the target neural network that best matches each component, thereby improving the accuracy of component prediction.

[0085] In addition, the embodiment of the present application optimizes the hyperparameters of the target neural network according to the data of specific components in the historical time series data. The optimization process can improve the generalization ability of the target neural network and reduce the risk of overfitting and underfitting by automatically searching for the optimal hyperparameter combination, thereby enhancing the stability and prediction effect of the target neural network. For example, a suitable learning rate can speed up the convergence speed of the target neural network, and a suitable number of layers and neurons can improve the expressive power of the target neural network.

[0086] Furthermore, the embodiment of the present application uses data of specific components in historical time series data to train the optimized target neural network, so that the target neural network can fully learn the patterns and laws of the specific components, and further improve the accuracy of prediction of the component.

[0087] Furthermore, the embodiment of the present application utilizes the trained target neural network to determine the component prediction results corresponding to the k components of the time series data to be predicted, and reconstructs the prediction results of each component. The above reconstruction can integrate the component prediction results obtained for different component predictions to obtain the final prediction result of the time series data to be predicted.

[0088] In summary, the embodiments of the present application effectively improve the accuracy and reliability of time series prediction through a series of steps such as data decomposition, neural network integration, hyperparameter optimization, targeted training and result reconstruction.

[0089] Method Example 1

[0090] refer to Figure 1 , shows a schematic flow chart of the steps of a time series prediction method according to an embodiment of the present application, the method may specifically include the following steps:

[0091] Step 101: Obtain historical time series data;

[0092] Step 102: decompose the historical time series data to obtain k components corresponding to the historical time series data;

[0093] Step 103, integrating M neural networks in the neural network integration pool; k and M are both positive integers greater than 1;

[0094] Step 104, determining the error evaluation index of the i-th component corresponding to the historical time series data under the j-th neural network;

[0095] Step 105: Select a neural network with the smallest error evaluation index from the M neural networks as the target neural network corresponding to the i-th component; i is a positive integer;

[0096] Step 106: Optimize the hyperparameters of the target neural network according to the data of the i-th component in the historical time series data;

[0097] Step 107: training the optimized target neural network according to the data of the i-th component in the historical time series data;

[0098] Step 108: using the trained target neural network, determine the component prediction results corresponding to the k components of the time series data to be predicted;

[0099] Step 109: reconstruct the component prediction results corresponding to the k components to obtain the prediction results corresponding to the time series data to be predicted.

[0100] In the embodiments of the present application, time series prediction refers to using the acquired data to be arranged in a time sequence, analyzing its change direction and degree, and thus speculating on the levels that may be reached in several future periods. The basic idea of ​​time series prediction is to treat the time series as a sample of a random variable and use probability statistics to minimize the impact of accidental factors.

[0101] Among them, the time series is to measure a variable or a group of variables X(t) of the objective process, and obtain an ordered set of discretization with time t as the independent variable at the moment: t1<t2<…<tn. X(t1), X(t2),…, X(tn) The independent variable t can have different physical meanings, such as length, temperature or other physical quantities. The fluctuation of the time series is the result of the joint action of many factors. The effects of various factors include long-term trends, seasonal changes, cyclic changes and random changes. For convenience, the variable can be called a prediction object. For example, in the energy industry, power generation data is a common prediction object.

[0102] In step 101, at least one data source may be used to collect historical time series data. Common data sources may include: sensors, databases, etc. It is understood that the embodiments of the present application do not limit the specific data source.

[0103] In one example, assuming that the prediction object is power generation data, the corresponding data source may include: a relational database, a dedicated time series database, etc.

[0104] In an embodiment of the present application, historical time series data can be presented in the following two situations: first, it can be data whose time series length is exactly the preset time length; second, the historical time series data can include data whose time series length is the preset time length.

[0105] For example, assuming that the preset time length is set to one month, the historical time series data can be the power generation data for one consecutive month, and each time point corresponds to the power generation value of one day. For another example, if the preset time length is one day, the historical time series data can be the power generation data for one consecutive day, and each time point corresponds to the power generation data of one hour. The time series prediction of the embodiment of the present application can be used to predict the power generation value at one time point or multiple time points. It can be understood that the embodiment of the present application does not limit the specific time unit represented by the time point.

[0106] In step 102, the historical time series data is decomposed to obtain k components that can reflect the trend, cycle and other characteristics of the historical time series data. The above k components can help the neural network learn different characteristics of the historical time series data more specifically, thereby improving the prediction accuracy of the time series data.

[0107] The embodiment of the present application can use data decomposition methods such as EMD (Empirical Mode Decomposition) to decompose historical time series data.

[0108] Data decomposition methods are a method of breaking down a complex signal into a set of simpler, more interpretable components. Data decomposition methods help identify different patterns in signals, thus playing a key role in areas such as fault diagnosis, noise removal, feature extraction, and data compression.

[0109] The empirical mode decomposition method is a time-frequency signal processing technique that can decompose complex signals into a limited set of intrinsic mode functions and a residual component. Its uniqueness lies in that it is an analysis method based entirely on data, which does not require any preset basis functions or models, but directly extracts the intrinsic structure of the signal from the data. However, it is also prone to problems such as mode mixing and over-decomposition.

[0110] The embodiments of the present application do not limit the specific empirical mode decomposition method. For example, the empirical mode decomposition method specifically includes: EEMD (Ensemble Empirical Mode Decomposition), CEEMD (Complete Ensemble Empirical Mode Decomposition), CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) and ICEEMDAN (Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise, Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise), etc.

[0111] Among them, in order to optimize the residual auxiliary noise and poor reconstruction accuracy caused by the introduction of Gaussian white noise in EEMD, CEEMD adds white noise that can offset positive and negative noise to the original signal, which can improve the decomposition effect. However, due to the differences in the intrinsic mode functions, CEEMD will make it difficult to completely offset the aggregated average components.

[0112] CEEMDAN borrows the idea of EEMD for multi - order iteration. After obtaining the m - th order IMF (Intrinsic Mode Function) component, the residual sequence is updated and the m - th order IMF component of white noise is added to achieve the purpose of white noise adaption to reduce the residual noise after reconstruction. However, due to multiple iterations, CEEMDAN may lead to many unnecessary components. The ICEEMDAN decomposition method further solves the problem of pseudo - modes in CEEMDAN by adding an operator that generates the local mean of the signal. Among them, IMF is a set of functions generated during the empirical mode decomposition process, which represent different oscillation modes in the original signal. c Mod e Function) component, the residual sequence is updated and the m - th order IMF component of white noise is added to achieve the purpose of white noise adaption to reduce the residual noise after reconstruction. However, due to multiple iterations, CEEMDAN may lead to many unnecessary components. The ICEEMDAN decomposition method further solves the problem of pseudo - modes in CEEMDAN by adding an operator that generates the local mean of the signal. Among them, IMF is a set of functions generated during the empirical mode decomposition process, which represent different oscillation modes in the original signal.

[0113] In the case of adopting the ICEEMDAN decomposition method, the above - mentioned process of decomposing historical time - series data specifically includes:

[0114] Step A1: Add N kinds of white noise to the historical time - series data respectively to obtain N first constructed sequences;

[0115] Step A2: Use the empirical mode decomposition method to decompose each first constructed sequence, and determine the first group of residual sequences and the first - order components according to the local means of the decomposition results of the N first constructed sequences;

[0116] Step A3: Add N kinds of white noise to the (m - 1) - th group of residual sequences respectively to obtain N second constructed sequences; N and m are positive integers; 1 < m ≤ k; For example, when m = 2, the (m - 1) - th group of residual sequences can be the first group of residual sequences. In this case, the second group of residual sequences and the second - order components can be determined according to the first group of residual sequences.

[0117] Step A4: Use the empirical mode decomposition method to decompose each second constructed sequence, and determine the m - th group of residual sequences and the m - th order components according to the local means of the decomposition results of the N second constructed sequences. The m - th order component can be any one of the k components. For example, the m - th order component can be the m - th component among the k components.

[0118] Refer to Figure 2 , which shows the flow chart of decomposing historical time - series data in an embodiment of the present application. This decomposition process introduces two operators E k (·) and F(·), where E m (·) represents the m - th order component generated after the input signal is decomposed by EMD, and F(·) represents the local mean of the signal. The above - mentioned decomposition process specifically includes the following steps:

[0119] Step B1: Add N kinds of white noise ω to the signal x to be decomposed(n) , generate N construction sequences x (n) =x+β0E1(ω (n) ), where n = 1, 2, ..., N, β0, β1, etc. are noise weights;

[0120] The signal x to be decomposed is specifically historical time series data.

[0121] Step B2: For x (n) Perform EMD decomposition and use the local mean operator F(·) to calculate a set of residual sequences r1=F(x (n) ); According to F(x (n) )=x-E1(x), and the first-order component IMF1=xF(x (n) );

[0122] Step B3: Continue to add the second-order white noise component to r1, and use the local mean operator F(·) to calculate the second residual sequence r2=F(r1+β1E2(ω (n) )) and the second order component IMF2 = r1-r2;

[0123] Step B4: Repeatedly add the m-th order white noise component to the m-1-th order residual sequence to obtain the m-th group of residual sequence r m =F(r m-1 +β m-1 E k (ω (n) )) and the mth order component IMF m =r m-1 -r m ;

[0124] Step B5: Repeat step B4 until the decomposition is completed, and finally obtain a set of intrinsic mode components IMF1, IMF2, ..., IMF k . Figure 2 In the above example, the k-th order component IMF k =r k-1 -r k .

[0125] Reference Figure 3 , shows a schematic diagram of the decomposition result of an embodiment of the present application, wherein the decomposition result specifically includes: 11 components and 1 residual sequence (Residual), and the 11 components are: IMF1 to IMF11.

[0126] It should be noted that, before decomposing the historical time series data, the historical time series data may also be preprocessed; the preprocessing specifically includes: outlier removal and missing value filling.

[0127] Outliers may be unreasonable data points caused by measurement errors, data entry errors or special events. Removing outliers can prevent these erroneous data from misleading subsequent analysis and prediction, so that the prediction results can more accurately reflect the real time series change trends and laws.

[0128] For all positive time series data, statistical methods can be used to detect outliers. Common statistical methods include standard deviation, mean, and confidence interval. Setting a detection threshold can effectively identify outliers. For example, the detection threshold is set to three times the mean, and the points exceeding the detection threshold are replaced with the detection threshold.

[0129] For time series data with both positive and negative values, the average value may be affected by the mutual offset of positive and negative values, so it is not suitable as a criterion for detecting outliers. In this case, methods such as confidence intervals can be used. Confidence intervals can more accurately determine the range of normal data based on the distribution characteristics of the data, thereby effectively identifying outliers. By eliminating these outliers, the quality and reliability of the data can be improved, providing a more accurate basis for subsequent time series analysis and prediction.

[0130] Missing values ​​can destroy the integrity of time series data and affect the coherence of data analysis. Missing value filling can restore the integrity of the data, making the data continuous in time and facilitating subsequent analysis and modeling.

[0131] In scenarios where data fluctuations are relatively small and the overall trend is relatively stable, the average value can be used to fill the missing values. In scenarios where the data has certain local variation characteristics, adjacent values ​​can be used for filling. It is understood that the embodiments of the present application do not limit the specific missing value filling method.

[0132] The embodiment of the present application can also divide the preprocessed historical time series data into a training set, a validation set, and a test set according to a set ratio.

[0133] The training set is used to train the target neural network, allowing the target neural network to learn the patterns and rules in the historical data. By continuously adjusting the parameters of the target neural network, the target neural network can fit the features in the training data as accurately as possible. For example, when predicting power generation, the training set contains a large amount of power generation data at known time points and related influencing factors. The target neural network establishes the relationship between power generation and various factors by learning these power generation data.

[0134] The validation set is used to select the best matching target neural network for the i-th component and to optimize the hyperparameters of the target neural network. For example, when adjusting the model's hyperparameters such as complexity and learning rate, the performance of the target neural network on the validation set can be observed, and the hyperparameter combination that minimizes the validation set error can be selected.

[0135] The test set is used to evaluate the generalization ability of the target neural network after training, that is, the performance of the target neural network after training on new and unseen data. The test set is independent of the training set and the validation set, and can more objectively reflect the actual performance of the target neural network after training.

[0136] The embodiment of the present application does not limit the set ratios corresponding to the training set, validation set, and test set, respectively. For example, the common set ratios are: 60%-80% for the training set, 10%-20% for the validation set, and 10%-20% for the test set. The specific ratios can be adjusted according to the scale of the data and actual needs. If the amount of data is large, the proportion of the training set can be appropriately increased to improve the learning effect of the prediction model; if the performance requirements of the prediction model are high, the proportions of the validation set and the test set can be appropriately increased to more strictly evaluate the performance of the prediction model. The prediction model of the embodiment of the present application may specifically include: a target neural network corresponding to k components.

[0137] Since the target neural network of the embodiment of the present application is used to predict the components, the embodiment of the present application needs to decompose the historical time series data in the training set, the validation set and the test set respectively.

[0138] In addition, the embodiment of the present application may also normalize the first initial values ​​of the k components to obtain first normalized values ​​of the k components.

[0139] In the case of non-normalization, some features of the component may dominate the training process of the target neural network because of their excessively large or small values, which will cause the parameter update of the target neural network to over-focus on these features and ignore other features, resulting in deviations in the parameter update direction and slower convergence. After normalization, the different features of the component are at the same scale, and the target neural network can learn the contribution of each feature more fairly. Parameter update algorithms such as gradient descent can more stably approach the direction of the global optimal solution, reducing the oscillation between different features, thereby accelerating the convergence of the target neural network.

[0140] The process of normalizing the first initial values ​​of k components specifically includes: determining the maximum value and the minimum value of the i-th component; 1≤i≤k; and normalizing each data point in the i-th component according to the maximum value and the minimum value of the i-th component.

[0141] Assuming that the value of the data point is y, the minimum value of the component is min_i, and the maximum value is max_i, the normalized value y_normalized is: y_normalized = (y-min_i) / (max_i-min_i).

[0142] After normalization, the values ​​of all data points of the i-th component are mapped to the interval [0,1], which makes the data of this component more comparable and stable in subsequent analysis and processing, and helps to improve the convergence speed of the target neural network.

[0143] In step 103, those skilled in the art may integrate M types of neural networks in the neural network integration pool according to actual application requirements. The embodiment of the present application does not limit the specific M types of neural networks.

[0144] For example, the M types of neural networks specifically include: extreme learning machine, radial basis function neural network and convolutional neural network.

[0145] Among them, the extreme learning machine can quickly learn the mapping relationship between input data (time series data of the i-th component) and output data (component prediction results). The way the extreme learning machine randomly initializes and quickly solves the output weights enables it to process a large amount of time series data in a short period of time. For time series prediction problems, the extreme learning machine can quickly capture the potential patterns and trends in the data. At the same time, due to its strong generalization ability, it can accurately predict even new time series data to a certain extent.

[0146] The radial basis function neural network can locally approximate the data at different time points through its radial basis function, so as to better adapt to the changes in the time series. The local response characteristics enable the radial basis function neural network to respond promptly to small changes in the time series and accurately capture the dynamic characteristics of the time series.

[0147] In time series prediction, convolutional neural networks can automatically extract features from time series data, such as periodicity and trend, through convolutional layers and pooling layers. These features are critical for predicting values ​​at future time points. At the same time, the translation invariance of convolutional neural networks means that when processing time series data, it will not lose the ability to recognize data patterns over time. Moreover, by adjusting the depth and width of the network, convolutional neural networks can better adapt to time series data of different complexities and improve the accuracy of predictions.

[0148] In a specific implementation, a suitable deep learning framework, such as TensorFlow, etc., can be installed and configured in an electronic device, and a network structure corresponding to each of the M neural networks can be constructed in the deep learning framework.

[0149] In step 104, the error evaluation index of the i-th component corresponding to the historical time series data under the j-th neural network is determined.

[0150] The error evaluation index is an indicator used to measure the accuracy of the prediction results of the i-th component corresponding to the historical time series data under the j-th neural network. 1≤j≤k.

[0151] The error evaluation index reflects the degree of difference between the predicted value of the jth neural network and the true value. The smaller the value, the closer the predicted value is to the true value, and the better the prediction performance of the jth neural network on this component; conversely, the larger the value, the worse the prediction performance.

[0152] The embodiments of the present application do not limit the specific error evaluation index. For example, the error evaluation index may include: RMSE (Root Mean Square Error), or MAPE (Mean Absolute Percentage Error).

[0153] The specific calculation process of RMSE includes: first, calculating the difference between the predicted value and the actual value at each time point, that is, the error; then, squaring each error to obtain the square error; then calculating the average of the square errors of all time points; finally, taking the square root of this average square error to obtain the root mean square error.

[0154] MAPE is calculated by first calculating the absolute percentage error between the predicted value and the actual value at each time point, and then averaging the absolute percentage errors of all time points.

[0155] In a specific implementation, the process of determining the error evaluation index of the i-th component corresponding to the historical time series data under the j-th neural network specifically includes:

[0156] Step C1, obtaining a verification sample from a verification set corresponding to the historical time series data; the verification sample corresponds to the first true value of the i-th component at the prediction time point;

[0157] Step C2, using the j-th neural network, determining a second prediction value of the i-th component corresponding to the verification sample at a preset time point and under the j-th neural network;

[0158] Step C3: Determine the error evaluation index of the i-th component corresponding to the verification sample under the j-th neural network according to the first true value and the second predicted value.

[0159] The first true value refers to the actual observed value of the i-th component at a specific forecast time point in the validation set corresponding to the historical time series data. It represents what actually happened to the i-th component in reality at the forecast time point.

[0160] The second prediction value refers to the prediction value of the i-th component in the verification sample at a preset time point obtained by processing the verification sample in the verification set corresponding to the historical time series data using the j-th neural network.

[0161] By comparing the second predicted values ​​of the i-th component at the predicted time point by different neural networks with the first true value, the prediction performance of the neural network can be evaluated and the quality of the neural network can be judged.

[0162] In step 105, a neural network with the smallest error evaluation index is selected from the M neural networks as the target neural network corresponding to the i-th component. For the i-th component, its error evaluation index under the M neural networks is determined respectively, and the target neural network can be a neural network with the smallest error evaluation index corresponding to the i-th component.

[0163] Referring to Table 1, an example of a target neural network corresponding to the i-th component of an embodiment of the present application is shown. Different components may correspond to the same type or different types of neural networks. In a specific implementation, the mapping relationship between the component number and the neural network number may be recorded in a data table.

[0164] Table 1

[0165] The i-th component Type of target neural network The first component Extreme Learning Machine The second component Radial Basis Function Neural Network …… …… kth component Convolutional Neural Networks

[0166] In step 106, the hyperparameters of the target neural network are optimized based on the data of the i-th component in the historical time series data.

[0167] In the prediction model, hyperparameters are a type of parameters that cannot be directly learned from the training set. Hyperparameters play a decisive role in the structure and training process of the prediction model. Different hyperparameter combinations will significantly affect the performance of the prediction model in terms of training speed, accuracy, and generalization ability. For example, in a neural network, the learning rate, number of layers, number of neurons, number of iterations, batch size, etc. are all hyperparameters.

[0168] In the specific implementation, the process of optimizing the hyperparameters of the target neural network according to the data of the i-th component in the historical time series data specifically includes:

[0169] Step D1, obtaining a verification sample from a verification set corresponding to the historical time series data; the verification sample corresponds to the first true value of the i-th component at the prediction time point;

[0170] Step D2, using the target neural network, determining the first predicted value of the i-th component corresponding to the verification sample at a preset time point and under the current hyperparameters of the target neural network;

[0171] Step D3: determining a fitness value corresponding to a current hyperparameter according to the first true value and the first predicted value;

[0172] Step D4: If the fitness value meets the preset conditions, the optimization is terminated;

[0173] Step D5: If the fitness value does not meet the preset conditions, the current hyperparameters are updated according to the optimization algorithm, and step D2 is executed.

[0174] Steps D1 to D5 continuously adjust the hyperparameters of the target neural network to improve its prediction performance when processing the i-th component in the historical time series data. Specifically, a validation sample is obtained from the validation set, and the target neural network is used to predict the i-th component of the sample at a preset time point under different hyperparameters, and then the fitness value of the hyperparameter is determined based on the difference between the first predicted value and the first true value. If the fitness value meets the preset conditions, the optimization is stopped, otherwise the hyperparameter is updated according to the optimization algorithm and the above process is repeated.

[0175] The fitness value reflects the prediction performance of the target neural network under the current hyperparameters. The fitness value can be the aforementioned RMSE or MAPE, etc.

[0176] The optimization algorithm may include: particle swarm optimization algorithm, genetic algorithm, or simulated annealing algorithm, etc.

[0177] Among them, the particle swarm optimization algorithm optimizes the hyperparameters of the target neural network by simulating the foraging behavior of bird flocks. In this process, each particle represents a set of hyperparameter combinations. The particles fly in the search space and adjust their speed and position according to their own historical optimal position and the historical optimal position of the entire group.

[0178] The genetic algorithm draws on the idea of ​​biological evolution to optimize the hyperparameters of the target neural network. It gradually evolves a better hyperparameter combination by selecting, crossover and mutating a set of hyperparameter combinations (individuals).

[0179] The simulated annealing algorithm is based on the principle of solid annealing process in physics to optimize the hyperparameters of the target neural network. It starts from a higher initial temperature and gradually cools down. During the cooling process, it accepts poor solutions with a certain probability to avoid falling into the local optimum. In the process of adjusting the hyperparameters of the neural network using the simulated annealing algorithm, a set of initial hyperparameter combinations is first randomly selected and a higher initial temperature is set. During the iteration process, the current hyperparameter combination is perturbed to generate a new combination. According to the fitness value of the new combination and a certain probability, it is decided whether to accept the new combination. At the same time, the temperature is gradually lowered and the above process is repeated until the stopping condition is met. The stopping condition can be related to factors such as the number of iterations or the temperature threshold. In this way, a set of better hyperparameter combinations can be found in the end.

[0180] The above optimization algorithm can effectively search for a better combination of hyperparameters and improve the performance and generalization ability of the neural network. Different optimization algorithms have different characteristics and applicable scenarios, and a suitable algorithm can be selected according to specific problems. It is understood that the embodiments of the present application are not limited to specific optimization algorithms.

[0181] After the optimization is completed, the target hyperparameters of the target neural network can be obtained. The training in step 107 can be performed based on the target hyperparameters.

[0182] In step 107, the optimized target neural network is trained according to the data of the i-th component in the historical time series data.

[0183] In the embodiment of the present application, the training process of the target neural network may include: forward propagation and back propagation.

[0184] Among them, forward propagation can calculate the component prediction results in sequence from the embedding layer to the processing layer according to the parameters of the target neural network. The component prediction results are used to determine the loss information.

[0185] Backward propagation can calculate and update the parameters of the target neural network in sequence from the output layer to the input layer according to the loss information. In the process of backpropagation, the gradient information of the parameters of the target neural network can be determined, and the parameters of the target neural network can be updated using the gradient information. For example, backpropagation can calculate and store the gradient information of the parameters of the target neural network in sequence from the processing layer to the embedding layer according to the chain rule in calculus.

[0186] The embodiments of the present application do not limit the specific loss information. For example, the loss information may include: mean square error, mean absolute error, or root mean square error.

[0187] It should be noted that the embodiment of the present application can use the data of the i-th component in the training set to train the target neural network. Specifically, the training sample in the training set corresponds to the second true value of the i-th component at the prediction time point; the target neural network is used to determine the second predicted value of the i-th component corresponding to the training sample at the preset time point and the current hyperparameters of the target neural network; the loss information is determined based on the second true value and the second predicted value; and the parameters of the target neural network are updated based on the loss information. In this way, the parameters of the target neural network are continuously optimized, and finally, when the set conditions are met, a trained neural network is obtained. The set conditions may include but are not limited to: reaching a preset number of iterations, the loss information drops below a certain threshold, the network performance meets expectations, etc.

[0188] In step 108, the trained target neural network may be used to determine component prediction results corresponding to the k components of the time series data to be predicted.

[0189] The embodiment of the present application can determine the process of k components of the time series data to be predicted, specifically including: preprocessing the time series data to be predicted; the preprocessing includes: outlier removal and missing value filling; decomposing the preprocessed time series data to be predicted to obtain second initial values ​​of k components; normalizing the second initial values ​​of the k components to obtain second normalized values ​​of the k components. Since the process of determining the k components of the time series data to be predicted is similar to the process of determining the k components of the historical time series data, it will not be repeated here, and reference can be made to each other.

[0190] Reference Figure 4 , shows a schematic flow chart of a time series prediction method of an embodiment of the present application, wherein k components of historical time series data or time series data to be predicted can be input into a neural network integration pool. The present application embodiment can search in the mapping relationship between component numbers and neural network numbers according to the component numbers corresponding to the k components, so as to obtain the target neural network corresponding to each component, and use the target neural network to process the component to obtain the component prediction result corresponding to the component.

[0191] It should be noted that, in the screening process of the target neural network, the component prediction result corresponding to a component can be called a first prediction value. In the training process of the target neural network, the component prediction result corresponding to a component can be called a second prediction value. In the inference process of the target neural network, the component prediction result corresponding to a component can be called a component prediction result.

[0192] In addition, it should be noted that the embodiment of the present application optimizes the hyperparameters of the target neural network in the neural network integration pool. The optimization process can improve the generalization ability of the target neural network and reduce the risk of overfitting and underfitting by automatically searching for the optimal hyperparameter combination, thereby enhancing the stability and prediction effect of the target neural network. For example, a suitable learning rate can speed up the convergence speed of the target neural network, and a suitable number of layers and neurons can improve the expressive power of the target neural network.

[0193] In step 109, the component prediction results corresponding to the k components are reconstructed to obtain the prediction results corresponding to the time series data to be predicted.

[0194] In a specific implementation, the process of reconstructing the component prediction results corresponding to the k components includes:

[0195] Step E1, performing denormalization processing on the component prediction results corresponding to the k components respectively, so as to obtain the component denormalized prediction results corresponding to the k components respectively;

[0196] Step E2: Fusion the component denormalization prediction results corresponding to the k components to obtain the prediction results corresponding to the time series data to be predicted.

[0197] The denormalization process can restore the component prediction results to the original data range. The original data range is specifically: the data range corresponding to the minimum value min_i and the maximum value max_i of a component.

[0198] The denormalization process may specifically include: firstly multiplying the component prediction result by the length of the original data range (maximum value minus minimum value), and then adding the minimum value of the original data range to obtain the component denormalized prediction result.

[0199] The embodiment of the present application can add the component denormalized prediction results corresponding to k components at a specific prediction time point according to the timestamp of the prediction time point to obtain the prediction result corresponding to the time series data to be predicted at the specific prediction time point.

[0200] In one example, the prediction of the power generation data of each wind turbine in a wind power station is taken as an example. Here, the time unit of the prediction time point can have multiple options. For example, the time unit of the prediction time point can be a month. If it is a month, the power generation data of each wind turbine in recent months can be used to predict the power generation situation of the next month, so as to carry out long-term capacity planning and resource allocation. The time unit of the prediction time point can also be a day. When it is a day, it can predict tomorrow's power generation based on the power generation data of recent days, which helps to reasonably arrange daily power transmission and maintenance plans. The time unit of the prediction time point can also be an hour or a minute. When it is an hour or a minute as a time unit, it can be more timely based on the power generation data of the last few hours or minutes. The power generation of the next hour or minute can be accurately predicted, which is very important for real-time power dispatching and stable operation of the power grid. In other words, the embodiment of the present application can obtain the prediction result of the power generation data at the Q+1th time point based on the power generation data of the last Q (Q is a positive integer) time points. Regardless of whether the time unit is a month, a day, an hour or a minute, it can provide strong support for the efficient operation and power management of the wind power station.

[0201] The embodiment of the present application can also feed back the prediction result to the monitoring center, so that the monitoring center can determine whether to trigger the early warning mechanism based on the relationship between the prediction result and the preset threshold. If the prediction value exceeds the preset threshold, the early warning mechanism is triggered.

[0202] The above closed-loop mechanism can timely detect potential risk factors through continuous prediction and early warning. Whether in business operations, project management or social management, actions can be taken before or just after problems occur to reduce losses caused by risks.

[0203] For example, in the scenario of predicting the power generation of each wind turbine in a wind power station, potential abnormal situations can be discovered in a timely manner through a closed-loop mechanism. For example, if the predicted power generation at a certain moment exceeds the preset threshold, the early warning mechanism is triggered, and the monitoring center can quickly take measures (such as locking the wind blades, disconnecting the engine, etc.) to prevent the impact of abnormal power fluctuations on the power grid and ensure the safe and stable operation of the entire power system. This is like installing a solid safety line for the wind power station, effectively reducing the risk of accidents.

[0204] It should be noted that the embodiment of the present application can execute steps 101 to 107 according to the trigger conditions, determine the target neural networks corresponding to the k components in the historical time series data, and save the mapping relationship between the component number and the neural network number. In this way, when the time series data to be predicted is subsequently received, the mapping relationship between the component number and the neural network number can be searched according to the component numbers corresponding to the k components to obtain the target neural network corresponding to each component, and the target neural network is used to process the component to obtain the component prediction result corresponding to the component.

[0205] A person skilled in the art can determine the trigger conditions according to actual application requirements. For example, the trigger condition can be a preset time period. For example, the analysis of historical time series data and the determination of the target neural network are automatically triggered at regular intervals to ensure that the latest and most accurate target neural network can be used for prediction when facing new time series data to be predicted.

[0206] Alternatively, the trigger condition may be receiving a refresh instruction from the user. When the user believes it is necessary to update the neural network or hopes to obtain more accurate prediction results, the user may manually issue a refresh instruction to trigger the entire process.

[0207] In addition, receiving a user's start command can also be used as a trigger condition. For example, when a specific project is launched or a new data analysis requirement arises, the user issues a start command to start processing historical time series data and determining the target neural network, preparing for subsequent forecasting work.

[0208] In summary, the time series prediction method of the embodiment of the present application decomposes the historical time series data, and the obtained k components can reflect the trend, cycle and other characteristics in the historical time series data. The above k components can help the neural network learn the different characteristics in the historical time series data more specifically, thereby improving the prediction accuracy of the time series data.

[0209] Moreover, the embodiment of the present application integrates a variety of neural networks in the neural network integration pool, and different neural network architectures have different adaptability to different types of features. On this basis, the embodiment of the present application determines the error evaluation index of each component of the historical time series data under different neural networks, and selects the neural network with the smallest error evaluation index as the target neural network corresponding to the component. Therefore, the embodiment of the present application can select the target neural network that best matches each component, thereby improving the accuracy of component prediction.

[0210] In addition, the embodiment of the present application optimizes the hyperparameters of the target neural network according to the data of specific components in the historical time series data. The optimization process can improve the generalization ability of the target neural network and reduce the risk of overfitting and underfitting by automatically searching for the optimal hyperparameter combination, thereby enhancing the stability and prediction effect of the target neural network. For example, a suitable learning rate can speed up the convergence speed of the target neural network, and a suitable number of layers and neurons can improve the expressive power of the target neural network.

[0211] Furthermore, the embodiment of the present application uses data of specific components in historical time series data to train the optimized target neural network, so that the target neural network can fully learn the patterns and laws of the specific components, and further improve the accuracy of prediction of the component.

[0212] Furthermore, the embodiment of the present application utilizes the trained target neural network to determine the component prediction results corresponding to the k components of the time series data to be predicted, and reconstructs the prediction results of each component. The above reconstruction can integrate the component prediction results obtained for different component predictions to obtain the final prediction result of the time series data to be predicted.

[0213] In summary, the embodiments of the present application effectively improve the accuracy and reliability of time series prediction through a series of steps such as data decomposition, neural network integration, hyperparameter optimization, targeted training and result reconstruction.

[0214] It should be noted that, for the method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of the present application are not limited by the described order of actions, because according to the embodiments of the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present application.

[0215] Based on the above embodiment, this embodiment also provides a time series prediction device, referring to Figure 5 The time series prediction device may specifically include: an acquisition module 501, a decomposition module 502, an integration module 503, an error determination module 504, a selection module 505, an optimization module 506, a training module 507, a prediction module 508 and a reconstruction module 509.

[0216] Wherein, the acquisition module 501 is used to acquire historical time series data;

[0217] A decomposition module 502 is used to decompose the historical time series data to obtain k components corresponding to the historical time series data;

[0218] Integration module 503, used to integrate M neural networks in the neural network integration pool; k and M are both positive integers greater than 1;

[0219] An error determination module 504 is used to determine the error evaluation index of the i-th component corresponding to the historical time series data under the j-th neural network;

[0220] A selection module 505 is used to select a neural network with the smallest error evaluation index from the M neural networks as a target neural network corresponding to the i-th component; i is a positive integer;

[0221] An optimization module 506 is used to optimize the hyperparameters of the target neural network according to the data of the i-th component in the historical time series data;

[0222] A training module 507 is used to train the optimized target neural network according to the data of the i-th component in the historical time series data;

[0223] The prediction module 508 is used to determine the component prediction results corresponding to the k components of the time series data to be predicted by using the trained target neural network;

[0224] The reconstruction module 509 is used to reconstruct the component prediction results corresponding to the k components respectively, so as to obtain the prediction results corresponding to the time series data to be predicted.

[0225] Optionally, the optimization module includes:

[0226] A verification sample acquisition module is used to obtain a verification sample from a verification set corresponding to the historical time series data; the verification sample corresponds to the first true value of the i-th component at the prediction time point;

[0227] A first prediction value determination module, used to determine, using the target neural network, a first prediction value of the i-th component corresponding to the verification sample at a preset time point and under current hyperparameters of the target neural network;

[0228] A fitness value determination module, used to determine the fitness value corresponding to the current hyperparameter according to the first true value and the first predicted value;

[0229] The end module is used to end the optimization if the fitness value meets the preset conditions;

[0230] The update module is used to update the current hyperparameters according to the optimization algorithm if the fitness value does not meet the preset conditions.

[0231] Optionally, the decomposition module includes:

[0232] A preprocessing module is used to preprocess the historical time series data; the preprocessing includes: outlier removal and missing value filling;

[0233] A historical decomposition module, used for decomposing the preprocessed historical time series data to obtain first initial values ​​of k components;

[0234] The normalization module is used to normalize the first initial values ​​of the k components to obtain first normalized values ​​of the k components.

[0235] Optionally, the decomposition module includes:

[0236] A first noise adding module is used to add N kinds of white noise to the historical time series data respectively to obtain N first constructed sequences;

[0237] A first decomposition module is used to decompose each first construction sequence using an empirical mode decomposition device, and determine a first group of residual sequences and first-order components according to a local mean of the decomposition results of the N first construction sequences;

[0238] The second noise adding module is used to add N kinds of white noise to the m-1th group of residual sequences respectively to obtain N second constructed sequences; N and m are positive integers;

[0239] The second decomposition module is used to decompose each second construction sequence using an empirical mode decomposition device, and determine the mth group of residual sequences and the mth order component according to the local mean of the decomposition results of the N second construction sequences.

[0240] Optionally, the error determination module is configured to include:

[0241] A verification sample acquisition module is used to obtain a verification sample from a verification set corresponding to the historical time series data; the verification sample corresponds to the first true value of the i-th component at the prediction time point;

[0242] A second prediction value determination module, used to determine the second prediction value of the i-th component corresponding to the verification sample at a preset time point and under the j-th neural network using the j-th neural network;

[0243] The error index determination module is used to determine the error evaluation index of the i-th component corresponding to the verification sample under the j-th neural network according to the first true value and the second predicted value.

[0244] Optionally, the reconstruction module includes:

[0245] A denormalization module is used to perform denormalization processing on the component prediction results corresponding to the k components to obtain the component denormalized prediction results corresponding to the k components;

[0246] The fusion module is used to fuse the component denormalization prediction results corresponding to the k components to obtain the prediction results corresponding to the time series data to be predicted.

[0247] Optionally, the M neural networks include: an extreme learning machine, a radial basis function neural network and a convolutional neural network.

[0248] The embodiment of the present application also provides a non-volatile readable storage medium, which stores one or more modules (programs). When the one or more modules are applied to a device, the device can execute instructions (instructions) of each method step in the embodiment of the present application.

[0249] The present application embodiment provides one or more machine-readable media on which instructions are stored, and when executed by one or more processors, an electronic device executes one or more of the methods described in the above embodiments. In the present application embodiment, the electronic device includes various types of devices such as terminal devices and server terminals (clusters).

[0250] The embodiments of the present disclosure may be implemented as a device that uses any appropriate hardware, firmware, software, or any combination thereof to perform desired configurations. The device may include: electronic devices such as terminal devices and server (cluster). Figure 6 An exemplary apparatus 2100 that can be used to implement various embodiments described in this application is schematically shown.

[0251] For one embodiment, Figure 6 An exemplary apparatus 2100 is shown having one or more processors 2102, a control module (chip set) 2104 coupled to at least one of the (one or more) processors 2102, a memory 2106 coupled to the control module 2104, a non-volatile memory / storage device 2108 coupled to the control module 2104, one or more input / output devices 2110 coupled to the control module 2104, and a network interface 2112 coupled to the control module 2104.

[0252] The processor 2102 may include one or more single-core or multi-core processors, and the processor 2102 may include any combination of general-purpose processors or special-purpose processors (such as graphics processors, application processors, baseband processors, etc.). In some embodiments, the device 2100 can be used as a terminal device, a server (cluster), and other devices described in the embodiments of the present application.

[0253] In some embodiments, the apparatus 2100 may include one or more computer-readable media (e.g., memory 2106 or non-volatile memory / storage 2108) having instructions 2114 and one or more processors 2102 combined with the one or more computer-readable media and configured to execute the instructions 2114 to implement a module to perform the actions described in the present disclosure.

[0254] For one embodiment, the control module 2104 may include any suitable interface controller to provide any suitable interface to at least one of the processor(s) 2102 and / or any suitable device or component in communication with the control module 2104 .

[0255] The control module 2104 may include a memory controller module to provide an interface to the memory 2106. The memory controller module may be a hardware module, a software module, and / or a firmware module.

[0256] The memory 2106 may be used, for example, to load and store data and / or instructions 2114 for the device 2100. For one embodiment, the memory 2106 may include any suitable volatile memory, such as a suitable DRAM (Dynamic Random Access Memory). In some embodiments, the memory 2106 may include a double data rate type four synchronous dynamic random access memory.

[0257] For one embodiment, the control module 2104 may include one or more input / output controllers to provide an interface to the non-volatile memory / storage device 2108 and the input / output device(s) 2110 .

[0258] For example, non-volatile memory / storage 2108 may be used to store data and / or instructions 2114. Non-volatile memory / storage 2108 may include any suitable non-volatile memory (e.g., flash memory) and / or may include any suitable non-volatile storage device(s) (e.g., one or more hard disk drives, one or more optical disk drives, and / or one or more digital versatile optical disk drives).

[0259] The non-volatile memory / storage device 2108 may include storage resources that are physically part of the device on which the apparatus 2100 is installed, or it may be accessible to the device without being part of the device. For example, the non-volatile memory / storage device 2108 may be accessed via (one or more) input / output devices 2110 over a network.

[0260] The input / output device(s) 2110 may provide an interface for the apparatus 2100 to communicate with any other appropriate device, and the input / output device 2110 may include a communication component, an audio component, a sensor component, etc. The network interface 2112 may provide an interface for the apparatus 2100 to communicate through one or more networks, and the apparatus 2100 may wirelessly communicate with one or more components of a wireless network according to any of one or more wireless network standards and / or protocols, for example, accessing a wireless network based on a communication standard, such as WiFi (Wireless Fidelity), 2G (2nd Generation Wireless Telephone Technology), 3G (3rd Generation Wireless Telephone Technology), 4G (4th Generation Wireless Telephone Technology), 5G (5th Generation Wireless Telephone Technology), etc., or a combination thereof for wireless communication.

[0261] For one embodiment, at least one of the processor(s) 2102 may be packaged together with the logic of one or more controllers (e.g., a memory controller module) of the control module 2104. For one embodiment, at least one of the processor(s) 2102 may be packaged together with the logic of one or more controllers of the control module 2104 to form a system-in-package. For one embodiment, at least one of the processor(s) 2102 may be integrated on the same die with the logic of one or more controllers of the control module 2104. For one embodiment, at least one of the processor(s) 2102 may be integrated on the same die with the logic of one or more controllers of the control module 2104 to form a system-on-chip.

[0262] In various embodiments, the device 2100 may be, but is not limited to, a terminal device such as a server, a desktop computing device, or a mobile computing device (e.g., a laptop computing device, a handheld computing device, a tablet computer, a netbook, etc.). In various embodiments, the device 2100 may have more or fewer components and / or different architectures. For example, in some embodiments, the device 2100 includes one or more cameras, a keyboard, a liquid crystal display screen (including a touch screen display), a non-volatile memory port, multiple antennas, a graphics chip, a dedicated integrated circuit, and a speaker.

[0263] Among them, the main control chip can be used as a processor or control module in the detection device, sensor data, location information, etc. are stored in a memory or non-volatile memory / storage device, the sensor group can be used as an input / output device, and the communication interface may include a network interface.

[0264] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0265] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0266] The present application embodiment is described with reference to the flowchart and / or block diagram of the method, terminal device (system) and computer program product according to the embodiment of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, and the combination of the process and / or box in the flowchart and / or block diagram can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmable data processing terminal device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device produce a device for implementing the function specified in one process or multiple processes in the flowchart and / or one box or multiple boxes in the block diagram.

[0267] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0268] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce computer-implemented processing, so that the instructions executed on the computer or other programmable terminal device provide steps for implementing the functions specified in one or more processes in the flowchart and / or one or more boxes in the block diagram.

[0269] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present application.

[0270] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the elements.

[0271] The above is a detailed introduction to a time series prediction method and device, an electronic device and a machine-readable medium provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea. At the same time, for general technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A time series prediction method, characterized in that: The method comprises: Get historical time series data; Decompose the historical time series data to obtain k components corresponding to the historical time series data; Integrate M neural networks in the neural network integration pool; k and M are both positive integers greater than 1; Determine the error evaluation index of the i-th component corresponding to the historical time series data under the j-th neural network; Select the neural network with the smallest error evaluation index from the M neural networks as the target neural network corresponding to the i-th component; i is a positive integer; According to the data of the i-th component in the historical time series data, the hyperparameters of the target neural network are optimized; According to the data of the i-th component in the historical time series data, the optimized target neural network is trained; Using the trained target neural network, determine the component prediction results corresponding to the k components of the time series data to be predicted; The component prediction results corresponding to the k components are reconstructed to obtain the prediction results corresponding to the time series data to be predicted.

2. The method according to claim 1, characterized in that The step of optimizing the hyperparameters of the target neural network according to the data of the i-th component in the historical time series data includes: Obtain a verification sample from a verification set corresponding to the historical time series data; the verification sample corresponds to the first true value of the i-th component at the prediction time point; Using the target neural network, determine the first predicted value of the i-th component corresponding to the validation sample at a preset time point and under the current hyperparameters of the target neural network; Determine a fitness value corresponding to a current hyperparameter according to the first true value and the first predicted value; If the fitness value meets the preset conditions, the optimization ends; If the fitness value does not meet the preset conditions, the current hyperparameters are updated according to the optimization algorithm.

3. The method according to claim 1, characterized in that The decomposition of the historical time series data includes: Preprocessing the historical time series data; the preprocessing includes: outlier removal and missing value filling; Decomposing the preprocessed historical time series data to obtain first initial values ​​of k components; The first initial values ​​of the k components are normalized to obtain first normalized values ​​of the k components.

4. The method according to claim 1, characterized in that: The decomposition of the historical time series data includes: Add N kinds of white noise to the historical time series data respectively to obtain N first constructed sequences; Using the empirical mode decomposition method, each first construction sequence is decomposed, and the first group of residual sequences and first-order components are determined according to the local mean of the decomposition results of the N first construction sequences; Add N kinds of white noise to the m-1th group of residual sequences respectively to obtain N second constructed sequences; N and m are positive integers; Each second construction sequence is decomposed by using the empirical mode decomposition method, and the m-th group of residual sequences and the m-th order component are determined according to the local mean of the decomposition results of the N second construction sequences.

5. The method according to claim 1, characterized in that Determining the error evaluation index of the i-th component corresponding to the historical time series data under the j-th neural network includes: Obtain a verification sample from a verification set corresponding to the historical time series data; the verification sample corresponds to the first true value of the i-th component at the prediction time point; Using the j-th neural network, determine the second predicted value of the i-th component corresponding to the verification sample at the preset time point and under the j-th neural network; According to the first true value and the second predicted value, an error evaluation index of the i-th component corresponding to the verification sample under the j-th neural network is determined.

6. The method according to claim 1, characterized in that The reconstructing the component prediction results corresponding to the k components respectively includes: Performing denormalization processing on the component prediction results corresponding to the k components respectively, so as to obtain the component denormalized prediction results corresponding to the k components respectively; The component denormalized prediction results corresponding to the k components are fused to obtain the prediction results corresponding to the time series data to be predicted.

7. The method according to any one of claims 1 to 6, characterized in that: The M types of neural networks include: extreme learning machine, radial basis function neural network and convolutional neural network.

8. A time series prediction device, characterized in that: The device comprises: Acquisition module, used to obtain historical time series data; A decomposition module is used to decompose the historical time series data to obtain k components corresponding to the historical time series data; An integration module, used for integrating M neural networks in a neural network integration pool; k and M are both positive integers greater than 1; An error determination module is used to determine the error evaluation index of the i-th component corresponding to the historical time series data under the j-th neural network; A selection module is used to select a neural network with the smallest error evaluation index from M neural networks as a target neural network corresponding to the i-th component; i is a positive integer; The optimization module is used to optimize the hyperparameters of the target neural network according to the data of the i-th component in the historical time series data; A training module is used to train the optimized target neural network according to the data of the i-th component in the historical time series data; The prediction module is used to determine the component prediction results corresponding to the k components of the time series data to be predicted by using the trained target neural network; The reconstruction module is used to reconstruct the component prediction results corresponding to the k components to obtain the prediction results corresponding to the time series data to be predicted.

9. An electronic device, characterized in that: include: processor; and A memory having executable codes stored thereon, which, when executed, causes the processor to execute the method according to any one of claims 1 to 7.

10. A machine-readable medium having executable codes stored thereon, which, when executed, causes a processor to perform the method according to any one of claims 1 to 7.

Images