Hybrid prediction method based on integrated model, storage medium and electronic equipment

Abstract

The invention relates to the technical field of hybrid prediction, in particular to a hybrid prediction method based on an integrated model, a storage medium and electronic equipment. Comprising the following steps: inputting a full-dimensional input feature vector into an integrated model to generate a target predicted value; the integrated model comprises a plurality of sub-prediction models and a weight distribution model; the weight distribution model is used for generating an output weight coefficient corresponding to each sub-prediction model according to the full-dimensional input feature vector; the target prediction value is generated by weighted summation of the sub-output values of all the sub-prediction models and the corresponding output weight coefficients. According to the invention, different sub-prediction models can better utilize the relationship between different feature values and the final target prediction result to carry out higher-precision prediction. And then sub output values of each sub prediction model are summarized in a weighted summation mode to generate a target prediction value with more credibility, so that the accuracy of a prediction result in a complex prediction task is improved.

Description

Technical Field

[0001] The present invention relates to the field of hybrid prediction technology, and in particular to a hybrid prediction method, storage medium and electronic equipment based on an integrated model. Background Art

[0002] With the development of deep learning technology, the predictive capabilities of neural networks continue to improve, especially with the support of big data and high-performance computing, and their application prospects are even broader. This technology excels in handling complex nonlinear relationships, pattern recognition, classification, regression, and other problems. In recent years, it has been widely used in fields such as weather forecasting, financial market analysis, medical diagnosis, and natural language processing.

[0003] However, different neural network models often exhibit varying performance when handling different types of tasks. Furthermore, for prediction tasks in more complex scenarios (such as weather forecasting, disaster warning, and GDP forecasting), the feature dimensions that influence the final prediction result are very broad and complex, and the relationships between different types of feature values and the final prediction result vary, such as linear and nonlinear relationships. Conventional prediction models often have low accuracy when used to complete prediction tasks in these scenarios. Summary of the Invention

[0004] In order to solve one of the above technical problems, the present invention adopts the following technical solution:

[0005] According to one aspect of the present invention, a hybrid prediction method based on an integrated model is provided, the method comprising the following steps:

[0006] Obtaining a full-dimensional input feature vector; the full-dimensional input feature vector includes comparative feature elements of the current feature values of some dimensions and the feature values of the same period in history;

[0007] The full-dimensional input feature vector is fed into the integrated model to generate the target prediction value;

[0008] The integrated model includes multiple sub-prediction models and weight distribution models;

[0009] The weight allocation model is used to generate the output weight coefficient corresponding to each sub-prediction model based on the full-dimensional input feature vector; the target prediction value is generated by the weighted sum of the sub-output values of all sub-prediction models and the corresponding output weight coefficients;

[0010] Features of different dimensions in the full-dimensional input feature vector contribute differently to the prediction results of different sub-prediction models; different sub-prediction models correspond to different sub-input feature vectors, and the sub-input feature vector is a subset of the full-dimensional input feature vector. The feature dimensions in the sub-input feature vector corresponding to each sub-prediction model are all feature dimensions whose contribution is greater than the contribution threshold;

[0011] The accuracy of the prediction results generated by each sub-prediction model using the corresponding sub-input feature vector is greater than the accuracy of the prediction results generated by the remaining other sub-prediction models using the same sub-input feature vector.

[0012] Furthermore, the method further comprises:

[0013] Obtain the importance score of each feature dimension in the full-dimensional input feature vector for each sub-prediction model;

[0014] The feature dimension corresponding to the importance score greater than the importance threshold is used as the feature dimension in the sub-input feature vector corresponding to the corresponding sub-prediction model.

[0015] Furthermore, the method further comprises:

[0016] Based on the model-based recursive feature elimination method, the sub-input feature vector corresponding to each sub-prediction model is obtained.

[0017] Furthermore, the multiple sub-prediction models include linear models, tree models and deep learning models.

[0018] Furthermore, each sub-prediction model includes the following structure:

[0019] Seasonal model, time series model, feature concatenation module, first multilayer perceptron, Random Forest, XGBoost, and second multilayer perceptron;

[0020] The output ends of the seasonal model and the time series model are both connected to the input end of the feature splicing module; the feature splicing module is used to perform channel splicing on the feature vectors extracted by the preset feature extraction layer in the seasonal model and the feature vectors extracted by the preset feature extraction layer in the time series model to generate a first fused feature vector;

[0021] The output end of the feature concatenation module is connected to the input end of the first multilayer perceptron; the first multilayer perceptron is used to assign a corresponding adjustment weight to each feature dimension in the first fused feature vector to generate a second fused feature vector;

[0022] The output of the first multilayer perceptron is connected to the input of Random Forest and XGBoost respectively;

[0023] The output ends of Random Forest and XGBoost are both connected to the input ends of the second multi-layer perceptron; the second multi-layer perceptron is used to perform weighted summation on the outputs of Random Forest and XGBoost to generate the output results corresponding to the sub-prediction model.

[0024] Furthermore, the seasonal model includes a seasonal autoregressive integrated moving average model; and the time series model includes an autoregressive integrated moving average model.

[0025] Furthermore, the method further comprises:

[0026] Based on the multiple target prediction values and the corresponding target true values corresponding to the preset historical period at the current monitoring moment, the model abnormality P corresponding to the current monitoring moment is obtained. P satisfies the following conditions:

[0027]

[0028] Among them, x i Y is the target forecast value of the preset historical period; x i Z is x i Y The corresponding target actual value; n is the total number of target forecast values obtained in the preset historical period;

[0029] If P>Y1, an alarm message of the output weight coefficient is generated; Y1 is the first accuracy threshold.

[0030] Furthermore, after generating the warning information of the output weight coefficient, the method further includes:

[0031] Obtain the sub-output values of each sub-prediction model corresponding to multiple target prediction values corresponding to the preset historical period at the current monitoring moment;

[0032] According to the sub-output value of each sub-prediction model and the corresponding target true value, the sub-model abnormality of each sub-prediction model is generated; among them, the sub-model abnormality Accuracy of the mth sub-prediction model m The following conditions are met:

[0033]

[0034] Zx i mY is x i Y The corresponding sub-output value of the m-th sub-prediction model;

[0035] According to the abnormality of each sub-model, the output weight coefficient corresponding to each sub-model is adjusted to generate the updated output weight coefficient corresponding to each sub-model; wherein, the updated output weight coefficient K corresponding to the mth sub-model is m G The following conditions are met:

[0036]

[0037] Among them, w is the total number of sub-prediction models, S 2 is the second accuracy threshold, S 2 Greater than the minimum value of the sub-model abnormality and less than the maximum value of the sub-model abnormality; K m Y is the output weight coefficient before updating corresponding to the m-th sub-prediction model.

[0038] According to a second aspect of the present invention, a non-transitory computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned hybrid prediction method based on the integrated model.

[0039] According to a third aspect of the present invention, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned hybrid prediction method based on an integrated model when executing the computer program.

[0040] The present invention has at least one of the following beneficial effects:

[0041] In the present invention, multiple sub-prediction models and weight distribution models are set in the integrated model. The weight distribution model can generate the output weight coefficient corresponding to each sub-prediction model based on the full-dimensional input feature vector. Then each sub-prediction model will obtain the corresponding sub-input feature vector from the full-dimensional input feature vector based on its own advantages, and then generate a more accurate prediction result. Since different sub-prediction models can process data from different angles or based on different assumptions, and the working mechanisms of different sub-prediction models are different, the relationship between different eigenvalues and the final target prediction results (such as linear relationships and nonlinear relationships, etc.) can be better utilized to make higher-precision predictions. Therefore, they can select different requirements or preferred input features from the full-dimensional input feature vector for prediction. Then, the sub-output values of each sub-prediction model are summarized by weighted summation to generate a target prediction value with more credibility, thereby improving the accuracy of the prediction results in complex prediction tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0043] Figure 1 A flowchart of a hybrid prediction method based on an integrated model provided in an embodiment of the present invention.

[0044] Figure 2 A schematic diagram of the architecture of the sub-prediction model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0046] As a possible embodiment of the present invention, Figure 1 As shown, a hybrid prediction method based on an integrated model is provided, which includes the following steps:

[0047] S100: Obtain the full-dimensional input feature vector. The full-dimensional input feature vector includes comparative feature elements of the current feature values of some dimensions and the feature values of the same period in history. In addition, the full-dimensional input feature vector of this embodiment also includes the sub-input feature vectors required by each sub-prediction model. The feature dimensions corresponding to the comparative feature elements can be part or all of the feature dimensions in the sub-input feature vector. The same period in history can specifically be a historical period that is relatively long from the current time, such as the previous year or previous years; it can also be a historical period that is relatively short from the current time, such as the previous month or previous months, the previous quarter or previous quarters, and can be flexibly determined based on the dependency law of the change cycle of the corresponding feature dimension.

[0048] For example, in a weather forecasting task, the sub-input feature vector includes the average humidity, total precipitation, precipitation frequency, average temperature, maximum temperature, minimum temperature, average sea level pressure, average ground pressure, average surface wind speed, average high-altitude wind speed, and average soil moisture in the current month. The corresponding feature dimensions can be compared with the feature values of the same month last year and / or the previous month to form feature values corresponding to the comparative feature elements. The specific feature values can be obtained by calculating the difference and / or similarity and / or year-on-year value and / or month-on-month value between the two values.

[0049] For example, in the GDP forecasting task, the sub-input feature vector includes the current quarter's GDP total, growth rate, total retail sales of consumer goods, per capita consumer expenditure, completed fixed asset investment, new credit scale, total import and export volume, trade surplus / deficit, industrial added value growth rate, manufacturing PMI, service industry production index, GDP growth rate of major trading partners, global commodity prices, interest rate level (such as the central bank's benchmark interest rate), money supply, and stock market capitalization / GDP ratio. The corresponding feature dimensions can be compared with the feature values of the same quarter and / or the previous quarter last year (that is, the difference between the year-on-year value or the quarter-on-quarter value and the year-on-year or cycle average is obtained) to form the feature values corresponding to the comparative feature elements.

[0050] S200: Input the full-dimensional input feature vector into the integrated model to generate a target prediction value.

[0051] The integrated model includes multiple sub-prediction models and weight distribution models.

[0052] The weight allocation model is used to generate the output weight coefficients corresponding to each sub-prediction model based on the full-dimensional input feature vector. The target prediction value is generated by the weighted sum of the sub-output values of all sub-prediction models and the corresponding output weight coefficients.

[0053] Specifically, the weight allocation model can select all feature dimensions in the full-dimensional input feature vector, or it can only use the feature dimensions corresponding to the comparative feature elements. Since the comparative feature elements can roughly reflect the similarity between the current state and the historical state, this can also provide certain guidance for the model, thereby more accurately generating the output weight coefficient.

[0054] Specifically, in this embodiment, the weight distribution model can be an MLP (Multilayer Perceptron), which can be designed to form corresponding training feature samples based on historical data during training. The corresponding labels can be formed according to the deviation between the prediction results output by each sub-prediction model and the actual target results to form corresponding weight labels.

[0055] For example, for the three sub-prediction models, for training feature sample A (A is in the same form as the full-dimensional input feature vector), the precipitation prediction value generated by sub-prediction model a is 200ml, the precipitation prediction value generated by sub-prediction model b is 240ml, and the precipitation prediction value generated by sub-prediction model c is 150ml, while the final actual precipitation is 206ml. The deviation value corresponding to sub-prediction model a can be calculated as follows:

[0056] 200-206 / |206-200|+|206-240|+|206-150|=-0.061.

[0057] Similarly, the deviation value of sub-prediction model b is +0.367, and the deviation value of sub-prediction model c is -0.571. In the above deviation, the positive sign indicates that the predicted value is larger than the true value, and the negative sign indicates that the predicted value is smaller than the true value. The specific data value indicates the degree of deviation. It can be seen that the deviation of sub-prediction model a, sub-prediction model b and sub-prediction model c gradually increases. When setting the corresponding weight label, its weight label value can be adjusted according to the adaptive setting of the deviation. The specific weight label value can be determined by those skilled in the art according to the actual scenario.

[0058] Therefore, according to the training data set set above, the MLP can more accurately generate the output weight coefficient corresponding to each sub-prediction model based on the full-dimensional input feature vector.

[0059] Features of different dimensions in the full-dimensional input feature vector contribute differently to the prediction results of different sub-prediction models. Different sub-prediction models correspond to different sub-input feature vectors. The sub-input feature vectors are subsets of the full-dimensional input feature vector. The feature dimensions in the sub-input feature vector corresponding to each sub-prediction model are all feature dimensions whose contribution is greater than the contribution threshold.

[0060] The accuracy of the prediction results generated by each sub-prediction model using the corresponding sub-input feature vector is greater than the accuracy of the prediction results generated by the remaining other sub-prediction models using the same sub-input feature vector.

[0061] The sub-prediction models provided in the present invention may perform predictions based on different physical processes, initial condition processing methods, parameterization schemes, or spatial resolutions. Therefore, features that are more suitable for their own use can be selected from the full-dimensional input feature vector as sub-input feature vectors to achieve higher prediction accuracy of the sub-prediction model itself.

[0062] Specifically, multiple sub-prediction models can be set up in the following two ways:

[0063] First, the sub-prediction model is set to multiple prediction models with different architecture types. For example, multiple sub-prediction models may include linear models, tree models, and deep learning models.

[0064] Taking the meteorological forecasting task as an example, multiple sub-prediction models can include at least two of the following: numerical weather forecast models (such as the European Centre for Medium-Range Weather Forecasts (ECMWF) model), regional weather forecast models (such as WRF (Weather Research and Forecasting Model)), linear models (such as ARIMA (AutoRegressive Integrated Moving Average)), tree models (such as Random Forests), and deep learning models (such as GraphCast models). All of the above models are existing models that can be used for weather forecasting. Of course, the above sub-models can also adopt a combination of other existing models. The various sub-prediction models can be integrated by calling the corresponding interface, and then according to the corresponding feature dimensions that need to be input, the feature values of the corresponding feature dimensions are adaptively added to the full-dimensional input feature vector.

[0065] Second, the architecture of each sub-prediction model is the same, and through different training data sets, each sub-prediction model can have different prediction accuracy performance for different types of input features.

[0066] like Figure 2 As shown, each sub-prediction model can include the following structure:

[0067] Seasonal model, time series model, feature concatenation module, first multilayer perceptron, Random Forest, XGBoost, and second multilayer perceptron. Specifically, the seasonal model can be a Seasonal Autoregressive Integrated Moving Average (SARIMA) model. The time series model can be an Autoregressive Integrated Moving Average (ARIMA) model. Correspondingly, for the GDP forecasting task, multiple different sub-forecasting models from existing technologies can also be combined for use.

[0068] The outputs of the seasonal model and the time series model are connected to the input of the feature concatenation module. The feature concatenation module is used to perform channel concatenation on the feature vectors extracted by the preset feature extraction layer in the seasonal model and the feature vectors extracted by the preset feature extraction layer in the time series model to generate a first fused feature vector.

[0069] The output end of the feature concatenation module is connected to the input end of the first multilayer perceptron. The first multilayer perceptron is used to assign a corresponding adjustment weight to each feature dimension in the first fused feature vector to generate a second fused feature vector.

[0070] The output of the first multilayer perceptron is connected to the input of Random Forest and XGBoost respectively.

[0071] The outputs of Random Forest and XGBoost are connected to the input of the second multilayer perceptron. The second multilayer perceptron is used to perform a weighted summation of the outputs of Random Forest and XGBoost to generate the output corresponding to the sub-prediction model.

[0072] In this embodiment, seasonal models and time series models are used to extract features from the input data and then splice them. ARIMA can handle non-stationary time series: it converts non-stationary series into stationary series by differencing, which is suitable for series with trends or random walks. It is more suitable for predicting data that has no obvious seasonality but has trends (such as long-term temperature change trends). And ARIMA can model linear relationships: it can predict by linearly combining historical values and error terms, which is suitable for scenarios dominated by linear relationships.

[0073] Compared to ARIMA, SARIMA can handle seasonal data: SARIMA adds a seasonal parameter to ARIMA, specifically designed to capture cyclical fluctuations. Therefore, by simultaneously extracting features from the input data using both ARIMA and SARIMA, we can better capture both long-term trends and short-term patterns in the data. The extracted features are then concatenated along the channel dimension. This concatenated feature vector contains richer long- and short-term semantic information, thereby improving the predictive accuracy of subsequent models. The first multi-layer sensor then adjusts the corresponding weights to generate a second fused feature vector, which is then fed into Random Forest and XGBoost to generate their respective outputs.

[0074] Random Forest is better at processing high-dimensional, unstructured data, while XGBoost is more efficient in gradient optimization of structured data. Combining the two, you can simultaneously utilize the robustness of Random Forest and the gradient optimization ability of XGBoost to make up for their respective shortcomings. In this embodiment, by utilizing the algorithm differences and complementary advantages of the two models (Random Forest and XGBoost), the limitations of the individual models are reduced through fusion, and the accuracy and robustness of the prediction are ultimately improved. This model fusion strategy is particularly effective in high-dimensional, complex, and noisy data, and is more suitable for scenarios with complex feature data such as weather forecasts and GDP forecasts.

[0075] Each sub-prediction model in this embodiment has the same architecture. To achieve high predictive power in different situations, it is necessary to rely on differentiated design of the training data. For example, in weather forecasting, there are different variations: years or quarters with stable weather changes and years or quarters with large weather changes. Similarly, in GDP forecasting, there are different variations: years and quarters with stable GDP changes, years and quarters with rapid development of new industries and large fluctuations in overall GDP, and years and quarters with balanced industrial development but large GDP fluctuations.

[0076] Based on this, when training multiple sub-prediction models, historical data can be clustered into training data for different situations according to the changing and fluctuating characteristics of each historical period, and then different sub-prediction models can be trained separately, so that different models can have high adaptability to different changing situations, that is, a certain sub-prediction model has higher prediction accuracy for situations with certain changing characteristics.

[0077] In this embodiment, when determining each feature dimension in the sub-input feature vector of the sub-prediction model, the determination can be performed according to the following method:

[0078] S101: Obtain the importance score of each feature dimension in the full-dimensional input feature vector for each sub-prediction model.

[0079] S102: The feature dimension corresponding to the importance score greater than the importance threshold is used as the feature dimension in the sub-input feature vector corresponding to the corresponding sub-prediction model.

[0080] Typically, different models may have different sensitivities to the same feature (e.g., linear models are sensitive to linear features, while tree models are sensitive to nonlinear features). Therefore, it is necessary to determine which feature dimensions in the full-dimensional input feature vector are more important to which sub-prediction model, and then determine the model's sensitivity to specific features or feature combinations, that is, the contribution of the features to the model's prediction results. Specifically, this can be quantified through feature importance, such as the model's built-in feature importance indicators (e.g., XGBoost's gain, Random Forest's importance), to evaluate feature importance.

[0081] In addition, the method of S103 may also be used to determine each feature dimension in the sub-input feature vector of the sub-prediction model.

[0082] S103: Based on the model-based recursive feature elimination method, obtain the sub-input feature vector corresponding to each sub-prediction model.

[0083] As another possible embodiment of the present invention, the method further includes:

[0084] S300: Based on multiple target prediction values and corresponding target true values corresponding to the preset historical period at the current monitoring moment, obtain the model abnormality P corresponding to the current monitoring moment, where P satisfies the following conditions:

[0085]

[0086] Among them, x i Y is the target forecast value of the i-th time period in the preset historical period. i Z is x i Y The corresponding target true value. n is the total number of target forecast values obtained in the preset historical period.

[0087] S400: If P>Y1, then generate an alarm message for outputting the weight coefficient, where Y1 is a first accuracy threshold.

[0088] The current monitoring moment in the above steps corresponds to a preset historical time period, which is often a historical time period adjacent to the current monitoring moment, such as a historical time period of the previous week or the previous month. Thus, the prediction performance of the model in the historical time period closer to the current moment can be used to more accurately and in real time reflect the prediction accuracy performance of the model in the recent period, so as to more promptly discover the anomaly of the model or the problem of decreased prediction accuracy, and to make certain adjustments to the model in a timely manner. This embodiment reflects the degree of deviation (i.e., abnormality) between the predicted value and the true value by comparing the variance between the multiple target predicted values and the corresponding target true values of the preset historical time period, thereby reflecting the prediction accuracy of the model. Specifically, the model abnormality is negatively correlated with the prediction accuracy of the model, so when P>Y1, it can be determined that the model is abnormal.

[0089] Specifically, in this embodiment, when P>Y1, an alarm message of the output weight coefficient is generated, thereby reminding the corresponding staff to promptly pay attention to whether the current output weight coefficient is appropriate and whether it needs to be updated or adjusted.

[0090] At the same time, the present invention also provides a method for updating or adjusting the current output weight coefficient, and the specific steps are as follows:

[0091] Furthermore, after generating the warning information of the output weight coefficient, the method further includes:

[0092] S500: Obtain the sub-output value of each sub-prediction model corresponding to multiple target prediction values corresponding to a preset historical period at the current monitoring moment.

[0093] S600: Generate the sub-model abnormality of each sub-prediction model according to the sub-output value of each sub-prediction model and the corresponding target true value. m The following conditions are met:

[0094]

[0095] Zx i mY is x i Y The corresponding sub-output value of the m-th sub-prediction model.

[0096] S700: According to the sub-model abnormality of each sub-prediction model, the output weight coefficient corresponding to each sub-prediction model is adjusted to generate an updated output weight coefficient corresponding to each sub-prediction model. Among them, the updated output weight coefficient K corresponding to the m-th sub-prediction model is m G The following conditions are met:

[0097]

[0098] Among them, w is the total number of sub-prediction models, S 2 is the second accuracy threshold, S 2 Greater than the minimum value of the sub-model abnormality and less than the maximum value of the sub-model abnormality. K m Y is the output weight coefficient before updating corresponding to the m-th sub-prediction model.

[0099] In this embodiment, the variance between the sub-output value of each sub-prediction model and the corresponding target true value is calculated to indicate the degree of abnormality of each sub-prediction model's prediction. Specifically, the larger the variance value, the greater the deviation between the sub-prediction model's prediction result and the true value, and the higher the degree of abnormality.

[0100] Based on this, when the output weight coefficient before updating is adjusted by the sub-model abnormality, S 2 -Accuracy m Calculate S 2 The size and difference between the abnormality of the sub-model can reflect whether the output weight coefficient of the corresponding sub-prediction model should be lowered or increased, as well as the adjustment range. 2 -Accuracy m The positive or negative value of the value is used to determine whether the current output weight coefficient is adjusted downward or upward. 2 -Accuracy mThe value of is used to determine the magnitude of the adjustment of the current output weight coefficient. For example, the sub-model abnormality of sub-prediction model a is 4, the sub-model abnormality of sub-prediction model b is 9, S 2 If is 5, then the corresponding sub-prediction model a Sub-prediction model b corresponds to From the above calculation results, we can see that the sub-model abnormality of sub-prediction model a is less than S 2 , which means that the prediction accuracy of sub-prediction model a is relatively high, and the results of this model should be trusted more, so the current output weight coefficient corresponding to sub-prediction model a will be adjusted upward, and the adjusted output weight coefficient will be 1.2 times the current output weight coefficient. Similarly, the sub-model abnormality of sub-prediction model b is greater than S 2 , which means that the prediction accuracy of sub-prediction model b is relatively low, and the results of this model should be trusted less, so the current output weight coefficient corresponding to sub-prediction model b will be adjusted downward, and the adjusted output weight coefficient will change to 0.2 times the current output weight coefficient.

[0101] Therefore, adaptive adjustment of the current output weight coefficient through the above method can better maintain the prediction accuracy of the model.

[0102] Furthermore, although the steps of the method of the present disclosure are described in a particular order in the accompanying drawings, this does not require or imply that the steps must be performed in this particular order, or that all steps shown must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0103] Through the description of the above embodiments, it is easy for those skilled in the art to understand that the example embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solution according to the embodiments of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes several instructions to enable a computing device (which can be a personal computer, a server, a mobile terminal, or a network device, etc.) to execute the method according to the embodiments of the present disclosure.

[0104] In an exemplary embodiment of the present disclosure, an electronic device capable of implementing the above method is also provided.

[0105] Those skilled in the art will appreciate that various aspects of the present invention may be implemented as systems, methods, or program products. Therefore, various aspects of the present invention may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, microcode, etc.), or in a combination of hardware and software, which may be collectively referred to herein as "circuits," "modules," or "systems."

[0106] The electronic device according to this embodiment of the present invention is merely an example and should not limit the functions and scope of use of the embodiments of the present invention.

[0107] The electronic device is implemented as a general-purpose computing device. Components of the electronic device may include, but are not limited to, the aforementioned at least one processor, the aforementioned at least one storage, and a bus connecting different system components (including the storage and the processor).

[0108] The storage stores program codes, which can be executed by the processor, so that the processor executes the steps according to various exemplary embodiments of the present invention described in the above “Exemplary Method” section of this specification.

[0109] The memory may include readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory, and may further include read only memory (ROM).

[0110] The storage may also include a program / utility having a set (at least one) of program modules, such program modules including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0111] The bus may represent one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processor, or a local bus using any of a variety of bus architectures.

[0112] The electronic device may also communicate with one or more external devices (e.g., keyboards, pointing devices, Bluetooth devices, etc.), one or more devices that enable a user to interact with the electronic device, and / or any device that enables the electronic device to communicate with one or more other computing devices (e.g., routers, modems, etc.). Such communication may be performed via an input / output (I / O) interface. Furthermore, the electronic device may also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via a network adapter. The network adapter communicates with other modules of the electronic device via a bus. It should be understood that, although not shown in the figures, other hardware and / or software modules may be used in conjunction with the electronic device, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0113] Through the description of the above embodiments, it is easy for those skilled in the art to understand that the example embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solution according to the embodiments of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes several instructions to enable a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) to execute the method according to the embodiments of the present disclosure.

[0114] In exemplary embodiments of the present disclosure, a computer-readable storage medium is also provided, on which is stored a program product capable of implementing the methods described above. In some possible implementations, various aspects of the present invention may also be implemented in the form of a program product comprising program code that, when executed on a terminal device, causes the terminal device to execute the steps according to various exemplary embodiments of the present invention described in the "Exemplary Methods" section above.

[0115] The program product may employ any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or component, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0116] A computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, which carries readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0117] The program code embodied on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0118] The program code for performing the operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages such as Java, C++, and the like, as well as conventional procedural programming languages such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a stand-alone software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving a remote computing device, the remote computing device may be connected to the user computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0119] Furthermore, the above-described figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention and are not intended to be limiting. It is readily understood that the processes illustrated in the above-described figures do not indicate or limit the temporal order of these processes. Furthermore, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0120] It should be noted that although several modules or units of the device for action execution are mentioned in the detailed description above, this division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more modules or units described above can be concretized in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided into multiple modules or units to be concretized.

[0121] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A hybrid prediction method based on an integrated model, characterized in that: The method comprises the following steps: Obtaining a full-dimensional input feature vector; the full-dimensional input feature vector includes a comparison feature element of the current feature value and the historical feature value of the same period of some dimensions; Inputting the full-dimensional input feature vector into the integrated model to generate a target prediction value; The integrated model includes multiple sub-prediction models and a weight distribution model; The weight allocation model is used to generate an output weight coefficient corresponding to each sub-prediction model based on the full-dimensional input feature vector; the target prediction value is generated by weighted summing the sub-output values of all sub-prediction models and the corresponding output weight coefficients; The features of different dimensions in the full-dimensional input feature vector have different contributions to the prediction results of different sub-prediction models; different sub-input feature vectors correspond to different sub-prediction models, and the sub-input feature vectors are subsets of the full-dimensional input feature vector. The feature dimensions in the sub-input feature vectors corresponding to each sub-prediction model are all feature dimensions whose contribution is greater than the contribution threshold; The accuracy of the prediction results generated by each sub-prediction model using the corresponding sub-input feature vector is greater than the accuracy of the prediction results generated by the remaining other sub-prediction models using the same sub-input feature vector.

2. The method according to claim 1, characterized in that Also includes: Obtain the importance score of each feature dimension in the full-dimensional input feature vector for each sub-prediction model; The feature dimension corresponding to the importance score greater than the importance threshold is used as the feature dimension in the sub-input feature vector corresponding to the corresponding sub-prediction model.

3. The method according to claim 1, characterized in that Also includes: Based on the model-based recursive feature elimination method, the sub-input feature vector corresponding to each sub-prediction model is obtained.

4. The method according to claim 1, wherein The multiple sub-prediction models include linear models, tree models and deep learning models.

5. The method according to claim 1, wherein Each of the sub-prediction models includes the following structure: Seasonal model, time series model, feature concatenation module, first multilayer perceptron, Random Forest, XGBoost, and second multilayer perceptron; The output ends of the seasonal model and the time series model are both connected to the input end of the feature splicing module; the feature splicing module is used to perform channel splicing on the feature vectors extracted by the preset feature extraction layer in the seasonal model and the feature vectors extracted by the preset feature extraction layer in the time series model to generate a first fused feature vector; The output end of the feature splicing module is connected to the input end of the first multilayer perceptron; the first multilayer perceptron is used to assign a corresponding adjustment weight to each feature dimension in the first fused feature vector to generate a second fused feature vector; The output end of the first multilayer perceptron is connected to the input end of Random Forest and XGBoost respectively; The output ends of the Random Forest and XGBoost are both connected to the input ends of the second multilayer perceptron; the second multilayer perceptron is used to perform weighted summation on the outputs of the Random Forest and XGBoost to generate an output result corresponding to the sub-prediction model.

6. The method according to claim 5, characterized in that The seasonal model includes a seasonal autoregressive integrated moving average model; the time series model includes an autoregressive integrated moving average model.

7. The method according to claim 1, characterized in that Also includes: Based on the multiple target prediction values and the corresponding target true values corresponding to the preset historical period at the current monitoring moment, the model abnormality P corresponding to the current monitoring moment is obtained. P satisfies the following conditions: Among them, x i Y is the target forecast value of the preset historical period; x i Z is x i Y The corresponding target actual value; n is the total number of target forecast values obtained in the preset historical period; If P>Y1, an alarm message of the output weight coefficient is generated; Y1 is a first accuracy threshold.

8. The method according to claim 7, characterized in that After generating the warning information of the output weight coefficient, the method further includes: Obtain the sub-output values of each sub-prediction model corresponding to multiple target prediction values corresponding to the preset historical period at the current monitoring moment; According to the sub-output value of each sub-prediction model and the corresponding target true value, the sub-model abnormality of each sub-prediction model is generated; among them, the sub-model abnormality Accuracy of the mth sub-prediction model m The following conditions must be met: Zx i mY is x i Y The corresponding sub-output value of the m-th sub-prediction model; According to the abnormality of each sub-model, the output weight coefficient corresponding to each sub-model is adjusted to generate the updated output weight coefficient corresponding to each sub-model; wherein, the updated output weight coefficient K corresponding to the mth sub-model is m G The following conditions must be met: Among them, w is the total number of sub-prediction models, S 2 is the second accuracy threshold, S 2 Greater than the minimum value of the sub-model abnormality and less than the maximum value of the sub-model abnormality; K m Y is the output weight coefficient before updating corresponding to the m-th sub-prediction model.

9. A non-transitory computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the hybrid prediction method based on an integrated model as described in any one of claims 1 to 8 is implemented.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the processor implements the hybrid prediction method based on the integrated model as described in any one of claims 1 to 8.

Images