Service prediction method, device and equipment based on LNM large numerical model

Abstract

The invention relates to the technical field of data processing, and discloses a service prediction method, device and equipment based on an LNM large numerical model, and the method comprises the steps: carrying out the preprocessing and multi-scale decomposition of internal historical service data and external macroscopic, mesoscopic and microscopic associated data, and obtaining long-term trend feature data and short-term fluctuation feature data; inputting a deep service association network for feature processing to obtain first comprehensive feature data; obtaining second comprehensive feature data through feature importance score and feature dependency relationship analysis; inputting an LNM large numerical model fused with multi-scale gating prediction to carry out service prediction, and obtaining a service prediction result; according to the method, distributed learning is carried out, iterative optimization is carried out through the local evaluation module and the local participation module, a target business prediction model is obtained, deep fusion of multi-dimensional features is realized, and the capture capability of an LNM large numerical model on data features is improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a business prediction method, device and equipment based on a LNM large numerical model. Background Art

[0002] With the rapid growth of business data and the continuous changes in the market environment, traditional business forecasting methods are unable to meet the needs of enterprises for high-precision forecasting. Existing forecasting methods mainly rely on a single statistical model or a simple machine learning algorithm, which cannot effectively capture the long-term trend characteristics and short-term fluctuation characteristics in the data, resulting in low stability and accuracy of the forecast results.

[0003] At present, most prediction models lack the ability to deeply analyze the correlation of multi-dimensional business data, cannot fully explore the intrinsic connections between different business segments, and fail to effectively deal with problems such as noise and missing values ​​in the data. Especially when processing large-scale data, the model's computational efficiency and resource consumption issues are more prominent, which restricts the application of prediction models in actual business scenarios. In addition, existing prediction methods generally have the problem of insufficient model generalization ability and are difficult to adapt to the characteristics of different business scenarios. Due to the lack of effective distributed learning mechanisms and dynamic optimization strategies, the model cannot be adaptively adjusted according to real-time prediction results, which affects the continuous optimization and performance improvement of the prediction model. Summary of the invention

[0004] The present invention provides a business prediction method, device and equipment based on the LNM large numerical model, which are used to achieve deep fusion of multi-dimensional features and improve the LNM large numerical model's ability to capture data features.

[0005] In a first aspect, the present invention provides a service prediction method based on a LNM large numerical model, the service prediction method based on the LNM large numerical model comprising: Preprocess and multi-scale decomposition of internal historical business data and external macro, meso and micro related data to obtain long-term trend characteristic data and short-term fluctuation characteristic data; Inputting the long-term trend feature data and the short-term fluctuation feature data into a deep business association network, performing feature processing through a plate association layer, a high-order feature extraction layer, and an information fusion layer to obtain first comprehensive feature data; Performing feature dimension reduction and optimization processing on the first comprehensive feature data, and obtaining second comprehensive feature data through feature importance scoring and feature dependency analysis; The second comprehensive feature data is input into the LNM large numerical model integrated with multi-scale gating prediction, and business prediction is performed through the long-term trend prediction unit and the short-term fluctuation prediction unit to obtain a business prediction result; According to the business prediction results, the LNM large numerical model integrating multi-scale gating prediction is subjected to distributed learning, and iterative optimization is performed through a local criticism module and a local participation module to obtain a target business prediction model.

[0006] In a second aspect, the present invention provides a service prediction device based on a LNM large numerical model, the service prediction device based on a LNM large numerical model comprising: The preprocessing module is used to preprocess and multi-scale decompose the internal historical business data and external macro, meso and micro related data to obtain long-term trend characteristic data and short-term fluctuation characteristic data; A feature processing module, used for inputting the long-term trend feature data and the short-term fluctuation feature data into a deep business association network, performing feature processing through a plate association layer, a high-order feature extraction layer and an information fusion layer, and obtaining first comprehensive feature data; An analysis module, used to perform feature dimension reduction and optimization processing on the first comprehensive feature data, and obtain second comprehensive feature data through feature importance scoring and feature dependency analysis; A business prediction module, used for inputting the second comprehensive feature data into the LNM large numerical model integrating multi-scale gating prediction, performing business prediction through a long-term trend prediction unit and a short-term fluctuation prediction unit, and obtaining a business prediction result; The iterative optimization module is used to perform distributed learning on the LNM large numerical model integrating multi-scale gating prediction according to the business prediction results, and iteratively optimize through the local criticism module and the local participation module to obtain the target business prediction model.

[0007] The third aspect of the present invention provides a business prediction device based on the LNM large numerical model, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the business prediction device based on the LNM large numerical model executes the above-mentioned business prediction method based on the LNM large numerical model.

[0008] A fourth aspect of the present invention provides a computer-readable storage medium, in which instructions are stored. When the computer-readable storage medium is run on a computer, the computer executes the above-mentioned business forecasting method based on the LNM large numerical model.

[0009] In the technical solution provided by the present invention, the long-term trend and short-term fluctuation feature extraction of internal historical business data is carried out through a multi-scale decomposition mechanism, and the deep fusion of multi-dimensional features is realized by combining a deep business association network, thereby improving the prediction model's ability to capture data features; a three-level association calculation unit and a high-order feature extraction layer are adopted to realize the effective association of features of different time dimensions and space dimensions, thereby enhancing the model's ability to understand complex business scenarios; a feature importance scoring and feature dependency analysis mechanism are introduced, and redundant features are reduced through dimensionality reduction optimization processing, thereby reducing the computational complexity of the model; an LNM large numerical model integrating multi-scale gated prediction is designed, and the long-term trend prediction unit and the short-term fluctuation prediction unit are combined to enhance the model's time series feature learning ability; a distributed learning architecture is adopted, and continuous optimization and dynamic adjustment of the model are realized through iterative optimization of a local criticism module and a local participation module; a probability calibration mechanism is introduced, and the prediction results are calibrated through a multi-layer density estimation network, thereby improving the reliability of the prediction results; a complete model verification mechanism is designed, and the convergence and stability of the model optimization are guaranteed by setting specific optimization stop conditions.

[0010] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0011] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 A schematic diagram of an embodiment of a business prediction method based on a LNM large numerical model in an embodiment of the present invention; Figure 2 A schematic diagram of an embodiment of a service prediction device based on a LNM large numerical model in an embodiment of the present invention; Figure 3 It is a schematic diagram of an embodiment of a business prediction device based on the LNM large numerical model in an embodiment of the present invention. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0015] To facilitate understanding of this embodiment, a business prediction method based on the LNM large numerical model disclosed in an embodiment of the present invention is first introduced in detail. Figure 1 As shown, the method comprises the following steps: 101. Preprocess and multi-scale decomposition of internal historical business data and external macro, meso and micro related data to obtain long-term trend characteristic data and short-term fluctuation characteristic data; It is understandable that the execution subject of the present invention may be a service prediction device based on the LNM large numerical model, or may be a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0016] Specifically, outlier detection processing is performed on internal historical business data and external macro, meso and micro related data to remove anomalies in the data. By setting the normal value range of business data, data that deviates from the range and significantly does not conform to the actual logic of the business is filtered out to obtain the first processed data that is initially cleaned. The missing values ​​in the first processed data are supplemented. The supplementary values ​​are determined by establishing an association matrix between the time dimension and the feature dimension to obtain the second processed data. The second processed data is segmented according to the time dimension. By setting the size of the time window, the continuous time series data is divided into multiple time series segments, so as to control the time characteristics of the data in a more fine-grained manner. The size of the window is determined according to the characteristics of the business and the forecasting requirements, so as to take into account sufficient historical information and the real-time nature of the data. Data smoothing processing is performed on multiple time series segments to extract stable features in the data. The data smoothing processing adopts the mean filtering method. By calculating the mean of the data within a certain time window, the stable part of the sequence is extracted to obtain the first decomposed data. Smoothing processing helps to reduce high-frequency random fluctuations in the data and make the long-term trend of the data more significant. For the first decomposed data, a low-frequency filter is used to extract the long-term trend. The low-frequency filter can filter out high-frequency fluctuation noise and retain the low-frequency components in the data. These low-frequency components reflect the long-term change trend of the business data. After this process, the long-term trend component is obtained. This component represents the part of the data that does not change frequently over time and is an important basis for reflecting the long-term development law of the business. At the same time, the fluctuation characteristics are extracted for the original multiple time series fragments. In order to extract the short-term fluctuation characteristics in the data, the data is filtered through a high-frequency filter, and the high-frequency components are extracted to obtain the short-term fluctuation component. The short-term fluctuation component reflects the short-term change law in the business data. The short-term fluctuation is caused by instantaneous market changes, seasonal fluctuations or other external environmental factors. The long-term trend component is feature extracted and statistically analyzed to obtain the final long-term trend feature data and short-term fluctuation feature data. The long-term trend feature data is the feature extraction result of the long-term trend component. By performing feature statistics on the trend component, a series of indicators describing long-term changes, such as mean, growth rate, etc., are obtained. The short-term fluctuation feature data is the feature extraction result of the short-term fluctuation component. By statistically analyzing the amplitude, frequency and other characteristics of short-term fluctuations, it can better reflect the short-term change characteristics of the business.

[0017] 102. Input the long-term trend feature data and the short-term fluctuation feature data into the deep business association network, perform feature processing through the plate association layer, the high-order feature extraction layer and the information fusion layer, and obtain the first comprehensive feature data; Specifically, the long-term trend feature data and the short-term fluctuation feature data are input into the three-level association calculation unit of the plate association layer respectively. The plate association layer captures various associations between features through self-attention, cross-attention and spatial attention. In the first-level association unit, the self-attention mechanism is used to calculate the feature correlation of the same time dimension. Through this mechanism, the model can automatically pay attention to the mutually influential or highly correlated features at the same time point, and fully mine the correlation information within the time dimension. The second-level association unit adopts the cross-attention mechanism to calculate the feature correlation between different time dimensions. In this way, the model captures the correlation characteristics of business data changing over time, especially the influence and interaction between different time periods. In the third-level association unit, the spatial correlation between plates is calculated through the spatial attention mechanism, the interaction between features is mined, and the plate association matrix is ​​generated. The matrix records the comprehensive association information between features in multiple dimensions such as time and space. The plate association matrix is ​​subjected to feature decomposition processing, and the main feature components are extracted by setting the singular value decomposition operation of the three decomposition dimensions, and the feature importance threshold is set for each dimension to perform feature screening to obtain the plate feature vector. The singular value decomposition process can effectively reduce the complexity of the data while retaining the main amount of information. Through threshold screening, it ensures that only features that are important for business prediction are retained, reducing redundant information and improving the prediction efficiency and accuracy of the model. The plate feature vector is input into the first feature extraction unit of the high-order feature extraction layer. The first feature extraction unit contains a four-layer feedforward network structure. Each layer uses the GELU activation function and batch normalization layer. A residual connection structure is added between adjacent layers to effectively alleviate the gradient vanishing problem while maintaining effective information transmission. The GELU activation function can perform nonlinear mapping on the input data and capture the nonlinear characteristics of the data, while the batch normalization layer helps to speed up the convergence of the model and improve the generalization performance of the model. The residual connection ensures the effective transmission of feature information between different layers to obtain the first layer of feature data. The first layer of feature data is input into the second feature extraction unit of the high-order feature extraction layer. The second feature extraction unit contains three sets of parallel feature extraction channels, each of which consists of two convolutional layers and one pooling layer. Through parallel design, multi-path extraction is performed on different features to capture more diverse feature representations. At the same time, the feature extraction channels are fused by feature splicing, so that the information extracted by each channel can be integrated to obtain the second layer of feature data. The second layer of feature data is input into the third feature extraction unit of the high-order feature extraction layer. The third feature extraction unit sets three spatial pyramid pooling modules with different receptive field sizes. These modules can extract spatial features of different scales from the input data, and integrate these pooling results by weighted averaging to obtain the third layer of feature data. The first layer of feature data, the second layer of feature data, and the third layer of feature data are input into the feature aggregation unit of the information fusion layer.The feature aggregation unit includes a channel attention module and a spatial attention module. The channel attention module adopts a squeeze-and-excitation structure. By compressing and activating each feature channel, the model performs weighted processing according to the importance of different channels. The spatial attention module adopts a non-local attention structure. By calculating the correlation between global features, the model can pay attention to the long-distance dependencies in the data and obtain fused feature data. The fused feature data is processed by grouping belief propagation, so that feature information can be transmitted and shared between different groups, improving the model's understanding of data features, and obtaining optimized feature data. The optimized feature data is input into the feature reconstruction unit. The feature reconstruction unit includes an encoder and a decoder structure. The encoder includes three downsampling layers. Each layer uses a convolution operation with a step size of 2 to gradually reduce the spatial dimension of the data and extract high-level features from the data. The decoder includes three upsampling layers. Each layer uses a deconvolution operation to reconstruct low-dimensional features back to the original high-dimensional representation, and adds jump connections between the encoder and the decoder so that features can be maintained and transmitted between different stages, ensuring that the reconstructed data can retain the original information and obtain the first comprehensive feature data.

[0018] 103. Performing feature dimension reduction and optimization processing on the first comprehensive feature data, and obtaining second comprehensive feature data through feature importance scoring and feature dependency analysis; Specifically, the first comprehensive feature data is subjected to feature correlation analysis, and the feature correlation matrix is ​​obtained by calculating the correlation coefficient between the features. The matrix can quantify the degree of mutual dependence and similarity between each pair of features. In the feature correlation matrix, the features with high correlation are marked. These features have strong redundancy and are merged by clustering method. The hierarchical clustering method is adopted to classify the features with high correlation into the same category to generate feature clustering data. Hierarchical clustering can gradually merge features from bottom to top to form a tree-like hierarchical structure, which helps the model understand the hierarchical association between features. The feature clustering data is scored for feature importance, and the importance of each feature is measured by calculating the variance contribution rate, information gain and feature weight of the feature. The variance contribution rate is used to measure the degree of change of the feature in the data set, the information gain is used to evaluate the contribution of each feature to the target variable, and the feature weight is combined with multiple information to comprehensively calculate the influence of the feature on the model prediction. The feature importance data is obtained by calculation. The feature importance data is sorted according to the score, and the features are screened according to the set importance threshold of 0.6 to obtain important feature data. Construct a dependency graph for important feature data to represent the dependency between features. A dependency graph is a graph structure used to describe the mutual influence and degree of dependency between features. In the dependency graph, nodes represent important features, while edges represent the dependency between features. Through the graph structure, the synergy between features can be analyzed to more accurately understand the feature pattern of the data. Group the feature dependency data according to the dependency strength. By calculating the correlation strength between groups and within groups, the relationship between features can be refined to obtain feature grouping data. Perform feature mapping on the feature grouping data, and map the features to a low-dimensional space through linear combination to obtain reduced-dimensional feature data. The linear combination method can retain the main information of the original features, while reducing the number of features and reducing the multicollinearity between features, thereby reducing the computational overhead and complexity of the model. Regularize the reduced-dimensional feature data. Adjust the numerical range of the feature data to a relatively consistent interval, standardize the data to a mean of 0 and a variance of 1, or normalize the data to the interval [0,1]. Regularization processing can eliminate the dimensional differences between different features and avoid unnecessary influence of certain features with large values ​​on the model, thereby improving the convergence speed and prediction accuracy of the model.

[0019] 104. Input the second comprehensive feature data into the LNM large numerical model integrating multi-scale gating prediction, perform business prediction through the long-term trend prediction unit and the short-term fluctuation prediction unit, and obtain a business prediction result; Specifically, the second comprehensive feature data is input into the LNM large numerical modeling layer, which consists of three layers of large numerical analysis networks. The first layer contains five fully connected layers, each of which uses the ReLU activation function to introduce nonlinear mapping, so that the model can effectively learn complex data features. A Dropout layer is set after each fully connected layer to randomly ignore a part of neurons during the training process to prevent overfitting and enhance the generalization ability of the model. The second layer contains three groups of linear regression layers, which capture the linear relationship between features through linear transformation. In the third layer, the model contains two layers of nonlinear transformation layers, which mine more complex relationships between features through nonlinear mapping to obtain numerical analysis features. The long-term trend prediction unit enhances the long-term trend component of the numerical analysis features. The long-term trend prediction unit contains four gated recurrent network (GRU) layers, each of which contains a data selection gate, a forget gate, and an output gate, each of which uses different activation functions. The data selection gate uses the sigmoid activation function to determine whether the input data is selected or not, the forget gate uses the tanh activation function to determine the data retention of the current time step, and the output gate generates the output of each time step. Through these gating mechanisms, the model can effectively capture the time dependency and long-term trend information in the data and obtain long-term trend data. The long-term trend data is processed by multi-scale decomposition to capture the data features at different scales. The multi-scale decomposition contains three scale decomposition units, each of which contains two convolutional layers and one pooling layer. The convolution kernel sizes of the convolutional layers are 3×3 and 5×5, respectively, which are designed to capture local features and a larger range of contextual information in the data. The convolution operation can effectively extract the change pattern of the data in different time windows, while the pooling layer uses the maximum pooling operation to reduce the dimension of the data while retaining the most significant features. Through multi-scale decomposition processing, multi-scale feature data containing information of different scales is obtained. The multi-scale feature data is input into the short-term fluctuation prediction unit for processing. The short-term fluctuation prediction unit consists of six bidirectional gating units, each of which contains a forward channel and a reverse channel. The bidirectional structure captures the dynamic changes in the time series in both the forward and backward directions. Each channel contains three convolutional layers and a batch normalization layer. The convolutional layers extract the short-term variation features of the time series, while the batch normalization layer is used to standardize each batch of data, accelerate the training of the model and improve stability. Through this step, the short-term fluctuation prediction unit generates short-term fluctuation data, which effectively reflects the frequent fluctuations in the business data. The long-term trend data and the short-term fluctuation data are gated and fused. The gated fusion controls the fusion ratio of features of different scales through the weight gating matrix. The size of the weight gating matrix is ​​N×N, where N is the number of feature dimensions. The role of the weight gating matrix is ​​to dynamically adjust the weight of each feature by learning the contribution of different features in the overall prediction, so as to achieve effective integration of information of different scales and obtain fused prediction data.The fused prediction data is input into the LNM prediction layer for the final business prediction. The LNM prediction layer contains four numerical prediction units, each of which includes three linear transformation layers and two nonlinear transformation layers. The output dimensions of the linear transformation layer are set to 256, 128, and 64, respectively, to gradually reduce the dimension of the data so that the model can learn the main feature representation more focusedly. The role of the nonlinear transformation layer is to capture complex feature relationships through nonlinear mapping, enhance the expression ability of the model, and generate initial prediction results. The initial prediction results are probability calibrated. The probability calibration contains three layers of density estimation networks, each of which contains four fully connected layers and a probability distribution layer. The number of neurons in the fully connected layer is 128, 64, 32, and 16, respectively. By reducing the number of neurons layer by layer, the features are gradually refined to improve the prediction accuracy of the model. The probability distribution layer is used to model the probability distribution of the prediction results to ensure that the output prediction results conform to the actual distribution of the data and obtain calibrated prediction data. The calibrated prediction data is threshold constrained, and the initial prediction results are corrected by setting upper and lower thresholds and confidence intervals. The upper and lower thresholds are the 95% and 5% quantiles of historical data, respectively. By setting them, the prediction results are limited to the reasonable range of historical data to avoid extreme prediction values. At the same time, the confidence interval is set to 90% to ensure that the prediction results are reasonable and stable at a higher confidence level, and obtain business prediction results.

[0020] The initial prediction results are input to the first fully connected layer of the first density estimation network for feature mapping. The number of neurons in this layer is 128, and the rectified linear unit (ReLU) is used as the activation function to enable the model to perform nonlinear transformation on the data and capture more complex features. The features of this layer are batch normalized to ensure that the mean of the feature data is 0 and the variance is 1, which improves the stability and convergence speed of the training and obtains the first layer of density feature data. The first layer of density feature data is input to the second, third and fourth fully connected layers of the first layer of density estimation network in turn. The number of neurons in these fully connected layers is set to 64, 32 and 16 respectively. Each fully connected layer uses the hyperbolic tangent function (tanh) as the activation function. The tanh function has the characteristic of mapping the input data to the [-1,1] interval, so that the model compresses and expands the feature data and captures more nonlinear features. A jump connection structure is added between adjacent fully connected layers so that feature information can be effectively transmitted between different layers, reducing the gradient vanishing problem and improving the learning efficiency of the model. The probability density is calculated through the probability distribution layer to obtain the probability distribution data of the first layer. The Gaussian mixture model is calculated for the probability distribution data of the first layer. By setting the number of components of the mixed Gaussian distribution to 5, the expectation maximization algorithm is used to iteratively optimize the mixture weight parameters and Gaussian distribution parameters. The number of optimization iterations is set to 100, and the log-likelihood function value is calculated in each iteration to ensure that the model maximizes the possibility of the data in the process of gradual convergence, and the first layer of calibration data is obtained, which contains a detailed probability description of the input data characteristics. The first layer of calibration data is input into the four fully connected layers of the second layer density estimation network. The number of neurons in the four fully connected layers is set to 128, 64, 32 and 16 respectively. The activation function uses the rectified linear unit ReLU, and a random dropout layer with an inactivation probability of 0.3 is set between each adjacent layer. Through the Dropout process, the model is effectively prevented from overfitting and the generalization ability of the model is enhanced. After the feature extraction of each fully connected layer, the data enters the probability distribution layer to calculate the probability density and obtain the probability distribution data of the second layer. For the probability distribution data of the second layer, the cumulative distribution function (CDF) calculation and inverse transformation sampling are performed to map the probability distribution to the uniform distribution interval of [0.05, 0.95], and the number of sampling points is set to 1000. The probability calibration is performed by the quantile method to ensure that the model prediction results have good probability consistency and obtain the second layer of calibration data. Through the calculation of the cumulative distribution function, the model can map the complex probability distribution to a standardized uniform distribution interval, making the prediction results more balanced and eliminating the influence of extreme probabilities.The second layer calibration data is input into the four fully connected layers of the third layer density estimation network. The number of neurons in the four fully connected layers is also set to 128, 64, 32 and 16 respectively. Each fully connected layer uses the rectified linear unit ReLU as the activation function to ensure the full extraction of nonlinear features. A softmax normalization operation is added after the last fully connected layer. Softmax can convert feature data into probability distribution so that the sum of the probabilities of each feature is 1, which is convenient for measuring the credibility of each prediction. The probability distribution data of the third layer is obtained through the calculation of the probability distribution layer. For the probability distribution data of the third layer, kernel density estimation is performed. In the kernel density estimation, the Gaussian kernel function is used as the kernel function. The bandwidth parameter of the kernel function is set to 0.1. By weighted averaging the probability density of each data point, the smoothed probability density distribution is calculated to obtain a smoother third layer calibration data. Kernel density estimation eliminates the influence of discrete points on the overall probability distribution by smoothing the data, and obtains a more continuous probability distribution that conforms to business rules. The first-layer calibration data, the second-layer calibration data, and the third-layer calibration data are weighted and fused according to weight coefficients 0.3, 0.3, and 0.4, and the weighted average of the three-layer calibration data is calculated as the final calibration prediction data.

[0021] 105. According to the business prediction results, the LNM large numerical model integrating multi-scale gating prediction is distributedly learned, and iterative optimization is performed through the local criticism module and the local participation module to obtain the target business prediction model.

[0022] Specifically, the business prediction results are input into the local criticism module for evaluation. The local criticism module consists of a four-layer evaluation network structure. The first layer contains 256 neurons, the second layer contains 128 neurons, the third layer contains 64 neurons, and the fourth layer contains 32 neurons. Each layer uses the LeakyReLU activation function. The LeakyReLU activation function can retain a small gradient in the negative area, which helps to alleviate the gradient vanishing problem and ensures that the parameters are effectively adjusted during the back propagation process. Through the processing of the fully connected layer and the activation function, the local evaluation data obtained can reflect the performance and deviation of the prediction results at the local level. The value function is calculated for the local evaluation data, and the long-term impact of the prediction results is evaluated by setting the temporal difference algorithm with a discount factor of 0.9. The discount factor is used to weigh the relationship between current rewards and future rewards. The temporal difference algorithm calculates the cumulative return value by comparing the predicted results with the actual results to obtain the strategy value data. The strategy value data is input into the local participation module for strategy optimization. The local participation module consists of three parallel policy networks, each of which contains five layers of feedforward networks, with the number of neurons set to 512, 256, 128, 64, and 32 respectively. Each feedforward network uses the tanh activation function, which can compress the input data into the interval [-1,1] and capture the nonlinear characteristics of the data. Through multi-layer processing of the input data, the local participation module obtains the policy parameter data describing the current policy of the model, which is used to guide the prediction and decision-making of the model in different situations. The policy parameter data is optimized proximally. The policy gradient is constrained by using a loss function with a clipping parameter of 0.2 to prevent excessive deviation from the original policy during the policy update process and ensure the stability of the model. The entropy regularization coefficient is set to 0.01. Entropy regularization helps to increase the randomness of the policy and prevent the model from falling into the local optimal solution. Through the optimization process, the optimized policy data is obtained. The optimized policy data is assigned tasks according to the distributed architecture. In the distributed architecture, 8 parallel observer nodes are set, and each observer node contains an independent local criticism module and local participation module. Through this design, each observer node can independently evaluate local data and optimize strategies to speed up the learning speed of the overall model. The data processed by each node is parallel, which can make full use of computing resources, greatly improve the training efficiency of the model, and obtain distributed learning data. Parameter synchronization is performed on the distributed learning data. By setting an asynchronous parameter server with a synchronization period of 100 steps, the parameters of each observer node are summarized and updated to obtain global parameter data. The asynchronous parameter server can update some parameters without waiting for all nodes, reduce synchronization overhead, improve training efficiency, and ensure timely update of model parameters, making the distributed learning process more robust. The global parameter data is input into the LNM large numerical model that integrates multi-scale gating predictions for parameter update.The stochastic gradient descent algorithm is used in the parameter update process, and the learning rate is set to 0.001 to ensure that the model parameters can converge smoothly during the update process. By gradually updating the model weights, the model can continuously adjust its response to the input data to better adapt to complex business scenarios. The number of iterations is set to 1000. Through multiple iterations, the model is continuously corrected and optimized in each round to obtain the optimized model data. The performance of the optimized model data is evaluated and verified. The performance of the model is evaluated by calculating the root mean square value and mean absolute percentage error of the prediction error. These two indicators can quantify the accuracy and error level of the model prediction. The optimization stop condition is set to stop the optimization when the performance index changes less than 0.001 for 5 consecutive iterations to prevent the model from falling into invalid iterations and wasting computing resources. The stop condition can ensure that the model stops training in time when it reaches a stable state and obtains the final target business prediction model.

[0023] In the embodiment of the present invention, a multi-scale decomposition mechanism is used to extract the features of long-term trends and short-term fluctuations of internal historical business data, and a deep business association network is combined to realize the deep fusion of multi-dimensional features, thereby improving the prediction model's ability to capture data features; a three-level association calculation unit and a high-order feature extraction layer are adopted to realize the effective association of features of different time dimensions and spatial dimensions, thereby enhancing the model's ability to understand complex business scenarios; a feature importance scoring and feature dependency analysis mechanism are introduced, and redundant features are reduced through dimensionality reduction optimization processing, thereby reducing the computational complexity of the model; a LNM large numerical model integrating multi-scale gated prediction is designed, and a long-term trend prediction unit and a short-term fluctuation prediction unit are combined to enhance the model's time series feature learning ability; a distributed learning architecture is adopted, and continuous optimization and dynamic adjustment of the model are realized through iterative optimization of local criticism modules and local participation modules; a probability calibration mechanism is introduced, and the prediction results are calibrated through a multi-layer density estimation network, thereby improving the reliability of the prediction results; a complete model verification mechanism is designed, and the convergence and stability of the model optimization are guaranteed by setting specific optimization stop conditions.

[0024] In a specific embodiment, the process of executing step 101 may specifically include the following steps: Perform outlier detection processing on internal historical business data and external macro, meso and micro related data, and screen the data by setting a normal value range to obtain first processed data; The missing values ​​in the first processed data are supplemented, and the supplementary values ​​are determined by establishing a correlation matrix between the time dimension and the feature dimension to obtain the second processed data; The second processed data is segmented according to the time dimension, and the data is divided by setting the time window size to obtain multiple time series segments; Performing data smoothing processing on multiple time series segments, extracting stable features through a mean filtering method to obtain first decomposition data, extracting long-term trends from the first decomposition data, and filtering the data through a low-frequency filter to obtain long-term trend components; Extract fluctuation characteristics of multiple time series segments, filter data through high-frequency filters, and obtain short-term fluctuation components; The long-term trend component is subjected to feature extraction and statistical analysis to obtain long-term trend feature data, and the short-term fluctuation component is subjected to feature extraction and statistical analysis to obtain short-term fluctuation feature data.

[0025] Specifically, perform outlier detection on internal historical business data and external macro, meso and micro related data. Outliers are caused by various reasons such as data entry errors and sensor failures, and are filtered by setting a normal value range. Set an outlier detection rule, such as defining a threshold range: ; in, represents the mean of the data, represents the standard deviation of the data, is the range of normal values. If the data point is not within this range, it is considered an outlier and the first processed data is obtained after screening. The missing values ​​in the first processed data are supplemented. The existence of missing values ​​is caused by incomplete data transmission, equipment collection errors, etc. The supplementary values ​​are determined by establishing the association matrix between the time dimension and the feature dimension. The matrix completion method is used to estimate the missing values ​​by using the continuity of the time dimension and the correlation between products. If the data at a specific time point is missing, it is estimated by using the average value of the data at the previous and next time points, or supplemented by the linear interpolation method to obtain the second processed data. The second processed data is segmented according to the time dimension. The data is divided into multiple time series segments by setting the time window size. The choice of time window depends on the specific business scenario. For example, if the seasonal sales data is analyzed, the window is set to one quarter to capture the trend and fluctuation within the quarter. Assume that there is a time series containing daily sales data for a whole year, and set every three months as a time window to divide four time series segments. Dividing the entire data set into smaller segments allows each segment to be analyzed independently for its characteristics, especially when there is a seasonal effect, different segments show different characteristics. Perform data smoothing on multiple time series segments to extract stable features from the data. Mean filtering is a smoothing method that eliminates high-frequency noise in the data by calculating the average value of the data in each time window. For example, for a time series with a length of Time series data , the smoothed data is obtained by mean filtering ,in: in, Indicates the size of the filter window, is the value after smoothing, and t represents the current time step in the time series. Mean filtering can reduce short-term random fluctuations and better reflect the long-term trend of the data, and obtain the first decomposition data. The data is filtered by a low-frequency filter, and the long-term trend of the first decomposition data is extracted. The low-frequency filter can eliminate the high-frequency part of the data and only retain the long-term trend with slow changes. After low-frequency filtering, the long-term trend component of the data is obtained. At the same time, for multiple time series fragments, the fluctuation characteristics are extracted. The fluctuation characteristics reflect the short-term changes in the data and are obtained by filtering the data with a high-frequency filter. The high-frequency filter can extract the frequently changing parts of the data as the short-term fluctuation component. The long-term trend component and the short-term fluctuation component are respectively feature extracted and statistically analyzed to obtain the final long-term trend feature data and short-term fluctuation feature data. For the long-term trend component, its mean, slope and other features are calculated to describe the overall market change trend. For example, if the slope of the long-term trend is positive, it indicates that the market is gradually growing; if the slope is negative, it indicates that the market is shrinking. For the short-term fluctuation component, its amplitude and frequency and other features are calculated to describe the fluctuation characteristics in the data. For example, by calculating the average amplitude of fluctuations, we can understand the severity of short-term market fluctuations.

[0026] In a specific embodiment, the process of executing step 102 may specifically include the following steps: The long-term trend feature data and the short-term fluctuation feature data are respectively input into the three-level correlation calculation unit in the plate correlation layer, wherein the first-level correlation unit uses the self-attention mechanism to calculate the feature correlation of the same time dimension, the second-level correlation unit uses the cross-attention mechanism to calculate the feature correlation of different time dimensions, and the third-level correlation unit uses the spatial attention mechanism to calculate the spatial correlation between plates, and obtain the plate correlation matrix; The plate correlation matrix is ​​subjected to characteristic decomposition, and the main characteristic components are extracted by setting the singular value decomposition operation of three decomposition dimensions, and the characteristic importance threshold is set for each dimension for feature screening to obtain the plate characteristic vector; The plate feature vector is input into the first feature extraction unit of the high-order feature extraction layer. The first feature extraction unit contains a four-layer feedforward network structure. Each layer uses a GELU activation function and a batch normalization layer, and a residual connection structure is added between adjacent layers to obtain the first layer of feature data. The first layer feature data is input into the second feature extraction unit of the high-order feature extraction layer. The second feature extraction unit includes three sets of parallel feature extraction channels. Each feature extraction channel includes two convolutional layers and one pooling layer. The feature extraction channels are fused by feature splicing to obtain the second layer feature data. The second layer feature data is input into the third feature extraction unit of the high-order feature extraction layer. The third feature extraction unit is provided with three spatial pyramid pooling modules with different receptive field sizes. The modules are integrated by weighted average to obtain the third layer feature data. Input the first layer feature data, the second layer feature data and the third layer feature data into the feature aggregation unit of the information fusion layer. The feature aggregation unit includes a channel attention module and a spatial attention module. The channel attention module adopts a squeeze-and-excitation structure, and the spatial attention module adopts a non-local attention structure to obtain fused feature data. The fused feature data is processed by group belief propagation to obtain optimized feature data, and the optimized feature data is input into a feature reconstruction unit. The feature reconstruction unit includes an encoder and a decoder structure. The encoder includes three downsampling layers, each layer uses a convolution operation with a step size of 2, and the decoder includes three upsampling layers, each layer uses a deconvolution operation, and a jump connection is added between the encoder and the decoder to obtain the first comprehensive feature data.

[0027] Specifically, the long-term trend feature data and short-term fluctuation feature data are input into the three-level correlation calculation unit in the plate correlation layer. The first-level correlation unit uses the self-attention mechanism to calculate the feature correlation of the same time dimension and measure the relationship between each feature at the same time point. The self-attention mechanism obtains the correlation weight between features by calculating the dot product between features. Let the feature matrix be , then the calculation of self-attention is expressed as: ; in, Represent the feature matrix The query matrix, key matrix and value matrix obtained by different linear transformations, is the dimension of the feature vector, which is used for normalization. Through the self-attention mechanism, the interaction between features at the same time point is effectively captured to obtain the first-level correlation features. The second-level correlation unit adopts the cross-attention mechanism to calculate the feature correlation of different time dimensions. The goal of the cross-attention mechanism is to measure the feature correlation between different time points, which helps to capture the temporal dependencies in time series data. For example, changes in business data are affected by the situation in the past few days or weeks, and the feature relationships at different time points can provide important predictive information. Through the cross-attention mechanism, more refined correlation data on the time dimension is obtained. The third-level correlation unit uses the spatial attention mechanism to calculate the spatial correlation between sectors. "Sectors" are different feature sets or feature categories that represent different parts or modules of the business, such as sales, marketing, supply chain, etc. By calculating the spatial correlation between sectors, the sector correlation matrix is ​​obtained. , which is used to represent the overall dependency between different characteristic plates. Perform eigendecomposition. By setting the singular value decomposition operation of three decomposition dimensions to extract the main eigencomponents, the singular value decomposition can convert the correlation matrix Expressed as the product of three matrices: ; in, and are the left singular matrix and the right singular matrix respectively, It is a diagonal matrix, and the diagonal elements are singular values. Singular values ​​can represent the importance of features. By setting the feature importance threshold for each dimension, the most representative features are screened out to obtain the plate feature vector . The plate feature vector is input to the first feature extraction unit of the high-order feature extraction layer. The first feature extraction unit consists of a four-layer feedforward network structure, each layer uses the GELU activation function and the batch normalization layer, and a residual connection structure is added between adjacent layers. The GELU activation function is a nonlinear activation function that enhances the expressiveness of the model by approximating the probability distribution of the input, while batch normalization is used to stabilize the training process and prevent gradient explosion or disappearance. The residual connection allows the input to flow freely between layers, avoiding the degradation problem common in deep networks. Through these processes, the first layer of feature data is obtained. The first layer of feature data is input to the second feature extraction unit of the high-order feature extraction layer. The second feature extraction unit contains three sets of parallel feature extraction channels, each of which contains two convolutional layers and one pooling layer. The function of the convolutional layer is to extract local features in the data, while the pooling layer is used to downsample the features to reduce the dimension of the data and retain the most important information. The design of the three sets of parallel channels captures the diversity of features, and the features extracted by each channel are fused by feature splicing to obtain the second layer of feature data. The second-layer feature data is input into the third feature extraction unit of the high-order feature extraction layer. The third feature extraction unit contains three spatial pyramid pooling modules with different receptive field sizes. The receptive field represents the "field of view" range of the convolution kernel on the input data. Setting receptive fields of different sizes captures information of different scales. The output of the module is integrated by weighted average to obtain the third-layer feature data. The first-layer feature data, the second-layer feature data, and the third-layer feature data are input into the feature aggregation unit of the information fusion layer. The feature aggregation unit contains a channel attention module and a spatial attention module. The channel attention module adopts a squeeze-and-excitation structure. By weighting each feature channel, it automatically selects important features to increase the robustness of the model. The spatial attention module adopts a non-local attention structure. By calculating the correlation between global features, the model can capture long-distance dependencies and obtain fused feature data. The fused feature data is processed by grouping belief propagation to optimize the feature data. Belief propagation is used to transfer information between features to strengthen the dependency between features and obtain optimized feature data. The optimized feature data is input into the feature reconstruction unit, which consists of an encoder and a decoder structure. The encoder contains three downsampling layers, each of which reduces the dimension of features through a convolution operation with a step size of 2 to extract high-level features in the data, while the decoder contains three upsampling layers, which restore the features through deconvolution operations. In order to ensure that the information can be effectively retained during the encoding and decoding process, a jump connection is set between the encoder and the decoder to directly pass the features in the encoder to the corresponding layer of the decoder, and finally obtain the first comprehensive feature data.

[0028] In a specific embodiment, the process of executing step 103 may specifically include the following steps: Performing feature correlation analysis on the first comprehensive feature data, obtaining a feature correlation matrix by calculating correlation coefficients between features, performing clustering processing on the features marked in the feature correlation matrix, classifying related features by a hierarchical clustering method, and obtaining feature clustering data; The importance of feature clustering data is scored, and feature importance data is obtained by calculating the variance contribution rate, information gain and feature weight of the feature; Sort the feature importance data by score, and filter the features by setting the importance threshold of 0.6 to obtain the important feature data; Construct a dependency graph for important feature data, express the dependency between features through the graph structure, obtain feature dependency data, group the feature dependency data according to the dependency strength, and obtain feature grouping data by calculating the correlation strength between and within groups; Perform feature mapping on the feature grouping data, map the features to a low-dimensional space through a linear combination method to obtain reduced-dimensional feature data, and perform regularization processing on the reduced-dimensional feature data to obtain second comprehensive feature data.

[0029] Specifically, the first comprehensive feature data is subjected to feature correlation analysis. By calculating the correlation coefficients between the features, a feature correlation matrix is ​​obtained. Each element in the correlation matrix represents the linear relationship between two features. The Pearson correlation coefficient is used to represent the correlation between the features, and the formula of the Pearson correlation coefficient is: ; in, Representation characteristics and Features The correlation coefficient between It is a feature and Features The covariance of and Characteristics and By calculating the correlation coefficients between all feature pairs, a feature correlation matrix is ​​constructed. , where each element Represents the degree of correlation between features. Cluster the highly correlated features marked in the feature correlation matrix. In order to reduce the redundancy between features, hierarchical clustering is used to group the features. Hierarchical clustering finds the hierarchical structure between features by gradually merging features from bottom to top. For example, if In the data, the correlation coefficient between some features exceeds a certain set threshold (such as 0.8), and these features are classified into the same group. In this way, highly correlated features are effectively classified into a feature cluster, and feature clustering data is obtained. . Importance scoring of feature clustering data. By calculating the variance contribution rate, information gain and feature weight of the feature, the feature importance data is obtained. The variance contribution rate indicates the degree of variation of the feature in the data. The feature with larger variance tends to have stronger descriptive ability for the data. The formula for calculating the variance contribution rate is: ; in, Representation characteristics The variance of Features No. Sample values, It is a feature The mean of is the number of samples. Information gain is used to measure the contribution of a feature to reducing the uncertainty of the target variable. The larger the information gain, the more important the feature. The feature weight is calculated by combining the variance and information gain of the feature, thereby obtaining a comprehensive importance score for each feature. The feature importance data is sorted according to the score, and the features are filtered by setting the importance threshold to 0.6. Only features with scores higher than the threshold are retained to obtain important feature data S. This screening method can effectively remove features that contribute less to business forecasts, reduce the complexity of the model, and improve the efficiency and accuracy of forecasts. Construct a dependency graph for important feature data. The dependency graph represents the dependency relationship between features through a graph structure, where nodes represent features and edges represent the dependency relationship between features. In order to construct the dependency graph, the dependency strength between features is calculated, and features with strong dependencies are connected to obtain feature dependency data. The feature dependency data is grouped according to the dependency strength to obtain feature grouping data G. By calculating the correlation strength between and within groups, the features are divided into different groups. The features within each group have strong similarities, and the features between groups are as independent as possible. For feature grouping data Perform feature mapping to reduce the dimension of features and retain the most important information. Feature mapping maps high-dimensional features to low-dimensional space through linear combination to reduce the complexity of data. For example, suppose there is a feature vector , the reduced eigenvector is obtained by linear combination ; in, is the linear mapping matrix, is the original feature vector, is the eigenvector after dimensionality reduction. By adjusting the mapping matrix , control the representation ability of the features after dimensionality reduction. After dimensionality reduction is completed, regularize the reduced feature data. Adjust the numerical range of the feature data to a relatively consistent interval, standardize the data to a mean of 0 and a variance of 1, or normalize the data to the interval [0,1]. The regularization operation can prevent the model from being affected by some feature values ​​that are too large or too small during the training process, ensure the convergence and stability of the model, and finally obtain the second comprehensive feature data. .

[0030] In a specific embodiment, the process of executing step 104 may specifically include the following steps: The second comprehensive feature data is input into the LNM large numerical modeling layer. The LNM large numerical modeling layer includes three layers of large numerical analysis networks. The first layer includes five fully connected layers. Each fully connected layer uses a ReLU activation function and a Dropout layer. The second layer includes three linear regression layers. The third layer includes two nonlinear transformation layers to obtain numerical analysis features. The long-term trend component of the numerical analysis features is enhanced. The long-term trend prediction unit contains 4 gated recurrent network layers. Each network layer contains a data selection gate, a forget gate, and an output gate. The data selection gate uses a sigmoid activation function, and the forget gate uses a tanh activation function to obtain long-term trend data. The long-term trend data is decomposed at multiple scales. The multi-scale decomposition includes three scale decomposition units. Each decomposition unit includes two convolutional layers and one pooling layer. The convolution kernel sizes of the convolutional layers are 3×3 and 5×5 respectively. The pooling layer uses the maximum pooling operation to obtain multi-scale feature data. The multi-scale feature data is input into the short-term fluctuation prediction unit, which contains 6 bidirectional gating units, each of which contains a forward channel and a reverse channel, and each channel contains 3 convolution layers and a batch normalization layer to obtain short-term fluctuation data; Gated fusion is performed on long-term trend data and short-term fluctuation data. Gated fusion controls the fusion ratio of features of different scales through a weighted gating matrix. The size of the weighted gating matrix is ​​N×N, where N is the number of feature dimensions, to obtain fused prediction data. The fused prediction data is input into the LNM prediction layer. The LNM prediction layer contains 4 numerical prediction units. Each prediction unit contains 3 linear transformation layers and 2 nonlinear transformation layers. The output dimensions of the linear transformation layer are 256, 128 and 64 respectively. The initial prediction results are obtained. The initial prediction results are probability calibrated. The probability calibration includes a 3-layer density estimation network. Each layer includes 4 fully connected layers and a probability distribution layer. The number of neurons in the fully connected layers is 128, 64, 32, and 16, respectively, to obtain the calibrated prediction data. The calibration forecast data is subjected to threshold constraint processing, and the initial forecast results are corrected by setting upper and lower thresholds and confidence intervals. The upper and lower thresholds are the 95% quantile and 5% quantile of the historical data, respectively, and the confidence interval is set to 90%, to obtain the business forecast results.

[0031] Specifically, the second comprehensive feature data is input into the LNM large numerical modeling layer. The LNM large numerical modeling layer contains a three-layer large numerical analysis network structure. The first layer consists of five fully connected layers. Each fully connected layer uses the ReLU activation function to introduce nonlinear mapping, so that the model can capture complex feature relationships. A Dropout layer is set after each fully connected layer to randomly ignore some neurons during the training process to prevent overfitting and improve the generalization ability of the model. For example, if the input feature is , after linear transformation of the fully connected layer and ReLU activation, the intermediate features are obtained , which can be expressed as: ; in, Indicates The weight matrix of the fully connected layer, is the bias term, is the intermediate feature output by this layer. After being processed by the five fully connected layers of the first layer, a rich feature representation is obtained. The second layer contains three sets of linear regression layers, which are used to capture the linear relationship between features, so as to retain the basic linear pattern between features during the modeling process. The third layer contains two nonlinear transformation layers. By performing nonlinear transformation on the feature data, the model can learn more complex relationships and obtain numerical analysis features. The long-term trend component of the numerical analysis features is enhanced. The long-term trend prediction unit contains four gated recurrent network (GRU) layers. Each GRU layer controls the transmission and memory of the data flow through a data selection gate, a forget gate, and an output gate. The data selection gate uses the sigmoid activation function, which determines whether the input data is selected or not, while the forget gate uses the tanh activation function to determine the degree of retention and forgetting of the data at the current time step. For example, the output of the data selection gate is expressed as: ; in, is the output of the data selection gate, is the hidden state of the previous time step, is the input feature of the current time step, and is the weight matrix, is the sigmoid activation function. Through the gating structure, GRU can effectively capture the time dependency of the data and obtain long-term trend data. The long-term trend data is decomposed at multiple scales. The multi-scale decomposition contains three scale decomposition units, each of which contains two convolution layers and one pooling layer. The convolution kernel sizes of the convolution layer are 3×3 and 5×5 respectively. Through convolution kernels of different sizes, local features of the data and feature patterns in a larger range are extracted. The pooling layer uses the maximum pooling operation to downsample the features to retain the most significant information. The goal of multi-scale decomposition is to extract features of different scales in long-term trend data and generate multi-scale feature data. For example, the convolution operation is expressed as: ; in, is the first feature after the convolution operation elements, is the weight of the convolution kernel, To input feature data, is the size of the convolution kernel, is the bias term. Through multi-scale decomposition, the model can understand the feature information of data at different scales. The multi-scale feature data is input into the short-term fluctuation prediction unit, which contains six bidirectional gating units, each of which contains a forward channel and a backward channel. Each channel contains three convolutional layers and a batch normalization layer. The convolutional layer is used to extract the short-term fluctuation features in the data, while the batch normalization layer is used to standardize the data, thereby accelerating the convergence speed of the model and improving the stability. Through the bidirectional structure of the forward and backward channels, the model simultaneously captures the dynamic changes in the forward and backward directions of the time series to obtain short-term fluctuation data. The long-term trend data and the short-term fluctuation data are gated and fused. During the fusion process, the weight gating matrix is ​​used to control the fusion ratio of features of different scales. The size of the weight gating matrix is ​​N×N, where N is the number of feature dimensions, which is used to represent the weight of each feature in the fusion process. Through gated fusion, the advantages of long-term trends and short-term fluctuations are effectively combined to obtain fused prediction data. The fused prediction data is input into the LNM prediction layer for prediction processing. The LNM prediction layer contains four numerical prediction units, each of which includes three linear transformation layers and two nonlinear transformation layers. The output dimensions of the linear transformation layer are set to 256, 128, and 64, respectively, to gradually reduce the dimension of the data so that the model can focus on the most important features. The role of the nonlinear transformation layer is to introduce nonlinear mapping, enhance the model's ability to express complex relationships, and obtain the initial prediction results. The initial prediction results are probability calibrated to ensure the rationality and credibility of the model output. Probability calibration is achieved through a three-layer density estimation network, each of which contains four fully connected layers and a probability distribution layer. The number of neurons in the fully connected layer is set to 128, 64, 32, and 16, respectively, to gradually extract features and adjust the feature space. Finally, the probability density of the output is estimated through the probability distribution layer to obtain the calibrated prediction data. The purpose of calibration is to make the prediction results of the model more consistent with the actual probability distribution and reduce the impact of extreme predictions. Threshold constraint processing is performed on the calibrated prediction data, and the initial prediction results are corrected by setting upper and lower thresholds and confidence intervals. The upper and lower thresholds are the 95% and 5% quantiles of historical data, respectively, and the confidence interval is set to 90%. The threshold constraint prevents the prediction result from exceeding the reasonable range, and finally obtains the business prediction result.

[0032] In a specific embodiment, the execution step performs probability calibration on the initial prediction result, the probability calibration includes a 3-layer density estimation network, each layer includes 4 fully connected layers and a probability distribution layer, the number of neurons in the fully connected layer is 128, 64, 32 and 16 respectively, and the process of obtaining the calibrated prediction data can specifically include the following steps: The initial prediction results are input into the first fully connected layer of the first density estimation network for feature mapping. The number of neurons in the first fully connected layer is 128, and the rectified linear unit ReLU is used as the activation function. The features are batch normalized with a mean of 0 and a variance of 1 to obtain the first layer of density feature data. The density feature data of the first layer is sequentially input into the second fully connected layer, the third fully connected layer and the fourth fully connected layer of the first density estimation network. The number of neurons in the second fully connected layer, the third fully connected layer and the fourth fully connected layer are set to 64, 32 and 16 respectively. Each fully connected layer uses the hyperbolic tangent function tanh as the activation function, and a skip connection structure is added between adjacent fully connected layers. The probability density is calculated through the probability distribution layer to obtain the probability distribution data of the first layer. The Gaussian mixture model is calculated for the first layer probability distribution data. The number of components of the mixed Gaussian distribution is set to 5. The expectation maximization algorithm is used to iteratively optimize the mixture weight parameters and Gaussian distribution parameters. The number of optimization iterations is set to 100 times. The log-likelihood function value is calculated for each iteration to obtain the first layer calibration data. The first layer calibration data is input into the four fully connected layers of the second layer density estimation network. The number of neurons in the four fully connected layers is set to 128, 64, 32 and 16 respectively. The activation function uses the rectified linear unit ReLU. A random dropout layer with an inactivation probability of 0.3 is set between every two adjacent layers. The probability density is calculated through the probability distribution layer to obtain the second layer probability distribution data. The second-layer probability distribution data is subjected to cumulative distribution function calculation and inverse transformation sampling, the probability distribution is mapped to a uniform distribution interval of [0.05, 0.95], the number of sampling points is set to 1000, and the probability is calibrated by the quantile method to obtain the second-layer calibration data; The second-layer calibration data is input into the four fully connected layers of the third-layer density estimation network. The number of neurons in the four fully connected layers is set to 128, 64, 32, and 16 respectively. Each fully connected layer uses the rectified linear unit ReLU as the activation function, and a softmax normalization operation is added after the last fully connected layer. The probability density is calculated through the probability distribution layer to obtain the third-layer probability distribution data; The kernel density of the third-layer probability distribution data is estimated by using a Gaussian kernel function as the kernel function. The bandwidth parameter of the kernel function is set to 0.1. The probability density of each data point is weighted averaged and the smoothed probability density distribution is calculated to obtain the third-layer calibration data. The first-layer calibration data, the second-layer calibration data, and the third-layer calibration data are weighted and fused according to weight coefficients of 0.3, 0.3, and 0.4, and the weighted average of the three-layer calibration data is calculated as the calibration prediction data.

[0033] Specifically, the initial prediction results are input into the first fully connected layer of the first density estimation network for feature mapping. The number of neurons in the first fully connected layer is set to 128 to extract more information from the initial input. This layer uses the rectified linear unit (ReLU) as the activation function, which is expressed as: ; in, is the output of the ReLU function, is the input feature value. The ReLU activation function has a simple and effective nonlinear characteristic, which enables the model to learn complex feature relationships. After the activation function, the feature data is batch normalized with a mean of 0 and a variance of 1. Batch normalization is used to standardize each feature so that it remains stable during training, which can accelerate the convergence speed and alleviate the problem of gradient disappearance. Through processing, the first layer of density feature data is obtained. The first layer of density feature data is sequentially input into the second fully connected layer, the third fully connected layer, and the fourth fully connected layer of the first layer density estimation network. The number of neurons in these layers is set to 64, 32, and 16, respectively. Each fully connected layer uses the hyperbolic tangent function (tanh) as the activation function. The tanh function maps the input to the interval [-1,1]. The formula is: ; in, is the output of the tanh activation function, is the input feature value. The tanh activation function has the characteristics of bidirectional compression and can better capture the relationship between features. A skip connection structure is added between adjacent fully connected layers to alleviate the gradient vanishing problem and make it easier to transfer feature information between layers. After passing through the fully connected layer, the data is input into the probability distribution layer to calculate the probability distribution data of the first layer. The Gaussian mixture model is calculated for the probability distribution data of the first layer. The Gaussian mixture model is a statistical model that approximates the distribution of complex data by weighting multiple Gaussian distributions. The number of components of the mixed Gaussian distribution is set to 5, and the expectation maximization algorithm is used to iteratively optimize the mixture weight parameters and Gaussian distribution parameters. The expectation maximization algorithm gradually maximizes the log-likelihood function value in each iteration by alternating between the E step and the M step. The expression of the log-likelihood function is: ; in, Indicates the given parameter The probability of data under For the The weights of the Gaussian components, is a Gaussian distribution, and are the mean and covariance matrices, respectively. The first layer calibration data is obtained through 100 iterations of optimization. The first layer calibration data is input into the second layer density estimation network, which contains four fully connected layers, and the number of neurons is set to 128, 64, 32 and 16 respectively. The activation function uses ReLU to continue to extract the nonlinear relationship of the features. A random dropout layer (Dropout) with an inactivation probability of 0.3 is set between each two adjacent layers. Through the Dropout operation, the model randomly ignores some neurons during the training process, improves the generalization ability of the model, and avoids overfitting. After the fully connected layer and Dropout processing, the feature data is input into the probability distribution layer to obtain the probability distribution data of the second layer. For the probability distribution data of the second layer, the cumulative distribution function is calculated and the inverse transformation sampling is performed. The cumulative distribution function is used to map the probability distribution to a standardized interval, map it to a uniform distribution interval of [0.05, 0.95], and set the number of sampling points to 1000 points. The probability calibration is performed through the quantile method to ensure that the prediction results meet the expected probability distribution, and the second layer calibration data is obtained. The second layer calibration data is input into the third layer density estimation network, which also contains four fully connected layers with 128, 64, 32 and 16 neurons respectively. Each fully connected layer uses ReLU as the activation function. After each layer is processed, a softmax normalization operation is added after the last fully connected layer. Softmax is used to convert the output into a probability distribution, and its formula is: ; in, For the The probability of an output, is the input feature, is the total number of output categories, Represents the activation value of the jth output unit in the softmax function. Through softmax normalization, ensure that each output value is between [0,1] and the sum of all outputs is 1, forming a valid probability distribution. Through the probability distribution layer calculation, the third layer of probability distribution data is obtained. The kernel density estimation is performed on the third layer of probability distribution data, and the Gaussian kernel function is used as the kernel function. The bandwidth parameter of the kernel function is set to 0.1. The expression of the Gaussian kernel function is: ; in, is the output of the kernel function, and is the input feature, is the bandwidth parameter. Kernel density estimation calculates the smoothed probability density distribution by weighted average of the probability density of each data point, and obtains smoother and more stable third-layer calibration data. The first-layer calibration data, the second-layer calibration data, and the third-layer calibration data are calculated according to The weighted fusion is performed with a weight coefficient of 0.4. The final calibration prediction data is obtained by calculating the weighted average of the three layers of calibration data.

[0034] In a specific embodiment, the process of executing step 105 may specifically include the following steps: The business prediction results are input into the local criticism module. The local criticism module contains a four-layer evaluation network structure. The first layer contains 256 neurons, the second layer contains 128 neurons, the third layer contains 64 neurons, and the fourth layer contains 32 neurons. Each layer uses the LeakyReLU activation function to obtain local evaluation data. The value function is calculated for the local evaluation data, and the prediction results are evaluated by the time series difference algorithm with a discount factor of 0.9. The cumulative return value of the prediction deviation is calculated to obtain the strategy value data; The strategy value data is input into the local participation module. The local participation module contains three parallel strategy networks. Each strategy network contains a 5-layer feedforward network with 512, 256, 128, 64 and 32 neurons respectively. The tanh activation function is used to obtain the strategy parameter data. Perform proximal policy optimization on the policy parameter data, constrain the policy gradient by setting the loss function with a clipping parameter of 0.2, and set the entropy regularization coefficient to 0.01 to obtain the optimized policy data; The optimization strategy data is assigned tasks according to the distributed architecture, and distributed learning data is obtained by setting 8 parallel observer nodes, each of which contains an independent local criticism module and local participation module; Perform parameter synchronization on distributed learning data, and summarize and update the parameters of each observer node by setting the asynchronous parameter server with a synchronization period of 100 steps to obtain global parameter data; The global parameter data is input into the LNM large numerical model integrating multi-scale gating prediction to update the parameters. The model weights are optimized by the stochastic gradient descent algorithm with a learning rate of 0.001 and the number of iterations is set to 1000 to obtain the optimized model data. The optimized model data was validated, and the model performance was evaluated by calculating the root mean square value and mean absolute percentage error of the prediction error. When the performance index change for five consecutive iterations was less than 0.001, the optimization was stopped and the target business prediction model was obtained.

[0035] Specifically, the business prediction results are input into the local criticism module, which consists of four layers of evaluation networks, including 256, 128, 64 and 32 neurons respectively. Each layer uses the LeakyReLU activation function, and the expression of the LeakyReLU activation function is ; in, represents the output of the activation function, is the input feature, is a constant less than 1, used to control the gradient of the negative part, usually set to 0.01. Compared with ReLU, the LeakyReLU activation function maintains a smaller gradient in the negative area, which helps to alleviate the gradient vanishing problem and enables the model to effectively train deep networks. After being processed by the four-layer network, local evaluation data is obtained, which is used to evaluate the performance of business prediction results. The value function is calculated for the local evaluation data, and the value of the prediction results is evaluated by the temporal difference algorithm. The value function is used to measure the cumulative return obtained by the model under the current strategy. The temporal difference algorithm combines the advantages of the Monte Carlo method and dynamic programming by setting the discount factor , weight the future returns to get the current value assessment. The expression for value function calculation is: ; in, Indicates in status The value of For the status Instant rewards, is a discount factor used to discount future returns. By continuously updating the value function, the cumulative return value of the prediction deviation is gradually calculated to obtain the strategy value data. The strategy value data is input into the local participation module. The local participation module contains three parallel strategy networks, each of which consists of five layers of feedforward networks with 512, 256, 128, 64 and 32 neurons respectively. Each layer uses the hyperbolic tangent function (tanh) as the activation function. The formula of the tanh function is: ; in, is the output of the tanh function, is the input feature. The bidirectional compression property of the tanh function enables it to better capture the relationship between features, allowing the policy network to generate more complex and flexible policies. Through the collaborative work of three parallel policy networks, the policy parameter data obtained effectively guides the behavior of the business prediction model. Proximal policy optimization is performed on the policy parameter data. Proximal policy optimization constrains the policy gradient by setting the loss function with a clipping parameter of 0.2, thereby preventing excessive deviation from the original policy during the policy update process and maintaining the stability of the optimization. The expression of the clipping part of the loss function is: ; in, is the clipping loss function, represents the probability ratio of the new strategy to the old strategy, is the advantage function, is the clipping parameter, set to 0.2, Represents the expected value at time step t. By clipping the loss function, it is ensured that the update of the strategy will not exceed the preset range, thereby controlling the change range of the strategy. The entropy regularization coefficient is set to 0.01 to increase the exploratory nature of the strategy, avoid the model from falling into the local optimal solution, and obtain the optimized strategy data. The optimized strategy data is assigned tasks according to the distributed architecture. Set up 8 parallel observer nodes, each of which contains an independent local criticism module and local participation module. Each node is responsible for independently evaluating and optimizing part of the data, and accelerates the training process of the entire model through parallel operations between each other. The distributed learning architecture makes full use of computing resources to greatly improve the training speed of the model, and finally obtains distributed learning data. Synchronize the parameters of the distributed learning data. By setting an asynchronous parameter server with a synchronization period of 100 steps, the parameters of each observer node are summarized and updated to obtain global parameter data. The role of the asynchronous parameter server is to asynchronously send the parameters of each node to the central server, summarize and update these parameters, and ensure that the model parameters can reflect the learning results from each node in a timely manner. The synchronization cycle of 100 steps means that the parameters of each node will be updated to the global model every 100 steps, thereby maintaining the parameter consistency and synchronization of the entire model. The global parameter data is input into the LNM large numerical model that integrates multi-scale gating predictions for parameter update. The stochastic gradient descent algorithm with a learning rate of 0.001 is used to optimize the model weights. Stochastic gradient descent finds the optimal model parameters by minimizing the loss function, and its update formula is: ; in, Indicates The parameters of the iterations, is the learning rate, set to 0.001, is the loss function About parameters The gradient of the model. By setting the number of iterations to 1000, the model gradually adjusts the parameters in multiple rounds to ensure the minimization of the loss function and obtain the optimized model data. The optimized model data is validated. In order to evaluate the performance of the model, the accuracy of the prediction is quantified by calculating the root mean square error (RMSE) and the mean absolute percentage error (MAPE). The formula for the root mean square error is: ; in, The model predicts values, For the actual values, is the total number of samples. The formula for the mean absolute percentage error is: ; These indicators can effectively reflect the prediction deviation and overall performance of the model. During the model verification process, if the performance indicator changes less than 0.001 for 5 consecutive iterations, the optimization is stopped to ensure that the model stops training when it reaches a stable state, thereby saving computing resources and preventing overfitting, and obtaining the target business prediction model.

[0036] The above describes the service prediction method based on the LNM large numerical model in the embodiment of the present invention. The following describes the service prediction device based on the LNM large numerical model in the embodiment of the present invention. Figure 2 In one embodiment of the present invention, a service prediction device based on a LNM large numerical model includes: The preprocessing module 201 is used to preprocess and multi-scale decompose the internal historical business data and external macro, meso and micro related data to obtain long-term trend feature data and short-term fluctuation feature data; The feature processing module 202 is used to input the long-term trend feature data and the short-term fluctuation feature data into the deep business association network, perform feature processing through the plate association layer, the high-order feature extraction layer and the information fusion layer, and obtain the first comprehensive feature data; The analysis module 203 is used to perform feature dimension reduction and optimization processing on the first comprehensive feature data, and obtain second comprehensive feature data through feature importance scoring and feature dependency analysis; The business prediction module 204 is used to input the second comprehensive feature data into the LNM large numerical model integrating multi-scale gating prediction, perform business prediction through the long-term trend prediction unit and the short-term fluctuation prediction unit, and obtain a business prediction result; The iterative optimization module 205 is used to perform distributed learning on the LNM large numerical model integrating multi-scale gating prediction according to the business prediction results, and perform iterative optimization through the local criticism module and the local participation module to obtain the target business prediction model.

[0037] Through the collaborative cooperation of the above components, the long-term trend and short-term fluctuation feature extraction of internal historical business data is carried out through the multi-scale decomposition mechanism, and the deep fusion of multi-dimensional features is realized by combining the deep business association network, which improves the prediction model's ability to capture data features; the three-level association calculation unit and the high-order feature extraction layer are adopted to realize the effective association of features of different time dimensions and spatial dimensions, and enhance the model's ability to understand complex business scenarios; the feature importance scoring and feature dependency analysis mechanism are introduced, and the redundant features are reduced through dimensionality reduction optimization processing, thereby reducing the computational complexity of the model; the LNM large numerical model integrating multi-scale gated prediction is designed, and the long-term trend prediction unit and the short-term fluctuation prediction unit are combined to enhance the model's time series feature learning ability; the distributed learning architecture is adopted, and the continuous optimization and dynamic adjustment of the model are realized through the iterative optimization of the local criticism module and the local participation module; the probability calibration mechanism is introduced, and the prediction results are calibrated through the multi-layer density estimation network, which improves the reliability of the prediction results; a complete model verification mechanism is designed, and the convergence and stability of the model optimization are guaranteed by setting specific optimization stop conditions.

[0038] above Figure 2 The business prediction device based on the LNM large numerical model in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The business prediction device based on the LNM large numerical model in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0039] Figure 3 1 is a schematic diagram of the structure of a business prediction device based on a LNM large numerical model provided by an embodiment of the present invention. The business prediction device 300 based on the LNM large numerical model may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more massive storage device terminals) storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 may be short-term storage or permanent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the business prediction device 300 based on the LNM large numerical model. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, and execute a series of instruction operations in the storage medium 330 on the business prediction device 300 based on the LNM large numerical model to implement the steps of the above-mentioned business prediction method based on the LNM large numerical model.

[0040] The business prediction device 300 based on the LNM large numerical model may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that Figure 3 The structure of the business prediction device based on the LNM large numerical model shown does not constitute a limitation on the business prediction device based on the LNM large numerical model provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0041] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the business forecasting method based on the LNM large numerical model.

[0042] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0043] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0044] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A business forecasting method based on LNM large numerical model, characterized in that: The method comprises: Preprocess and multi-scale decomposition of internal historical business data and external macro, meso and micro related data to obtain long-term trend characteristic data and short-term fluctuation characteristic data; Inputting the long-term trend feature data and the short-term fluctuation feature data into a deep business association network, performing feature processing through a plate association layer, a high-order feature extraction layer, and an information fusion layer to obtain first comprehensive feature data; Performing feature dimension reduction and optimization processing on the first comprehensive feature data, and obtaining second comprehensive feature data through feature importance scoring and feature dependency analysis; The second comprehensive feature data is input into the LNM large numerical model integrated with multi-scale gating prediction, and business prediction is performed through the long-term trend prediction unit and the short-term fluctuation prediction unit to obtain a business prediction result; According to the business prediction results, the LNM large numerical model integrating multi-scale gating prediction is subjected to distributed learning, and iterative optimization is performed through a local criticism module and a local participation module to obtain a target business prediction model.

2. The business forecasting method based on the LNM large numerical model according to claim 1 is characterized in that: The internal historical business data and external macro, meso and micro related data are preprocessed and decomposed at multiple scales to obtain long-term trend characteristic data and short-term fluctuation characteristic data, including: Perform outlier detection processing on internal historical business data and external macro, meso and micro related data, and screen the data by setting a normal value range to obtain first processed data; Performing supplementary processing on the missing values ​​in the first processed data, determining supplementary values ​​by establishing a correlation matrix between the time dimension and the feature dimension, and obtaining second processed data; The second processed data is segmented according to the time dimension, and the data is divided by setting the time window size to obtain multiple time series segments; Performing data smoothing processing on the multiple time series segments, extracting stable features by a mean filtering method to obtain first decomposed data, extracting long-term trends from the first decomposed data, and filtering the data by a low-frequency filter to obtain long-term trend components; Extracting fluctuation characteristics from the multiple time series segments, filtering the data through a high-frequency filter, and obtaining short-term fluctuation components; The long-term trend component is subjected to feature extraction and statistical analysis to obtain long-term trend feature data, and the short-term fluctuation component is subjected to feature extraction and statistical analysis to obtain short-term fluctuation feature data.

3. The business forecasting method based on the LNM large numerical model according to claim 2 is characterized in that: The long-term trend feature data and the short-term fluctuation feature data are input into the deep business association network, and feature processing is performed through the plate association layer, the high-order feature extraction layer and the information fusion layer to obtain the first comprehensive feature data, including: The long-term trend feature data and the short-term fluctuation feature data are respectively input into the three-level association calculation unit in the plate association layer, wherein the first-level association unit adopts a self-attention mechanism to calculate the feature correlation of the same time dimension, the second-level association unit adopts a cross-attention mechanism to calculate the feature correlation of different time dimensions, and the third-level association unit adopts a spatial attention mechanism to calculate the spatial correlation between plates, so as to obtain a plate association matrix; Performing feature decomposition processing on the plate correlation matrix, extracting main feature components by setting a singular value decomposition operation of three decomposition dimensions, and setting a feature importance threshold for each dimension to perform feature screening, thereby obtaining a plate feature vector; Input the plate feature vector into the first feature extraction unit of the high-order feature extraction layer, wherein the first feature extraction unit comprises a four-layer feedforward network structure, each layer uses a GELU activation function and a batch normalization layer, and a residual connection structure is added between adjacent layers to obtain the first layer feature data; Inputting the first layer feature data into the second feature extraction unit of the high-order feature extraction layer, the second feature extraction unit includes three sets of parallel feature extraction channels, each feature extraction channel includes two convolutional layers and one pooling layer, and the feature extraction channels are fused by feature splicing to obtain the second layer feature data; The second layer feature data is input into the third feature extraction unit of the high-order feature extraction layer, the third feature extraction unit is provided with three spatial pyramid pooling modules with different receptive field sizes, and the modules are integrated by weighted average to obtain the third layer feature data; Inputting the first layer feature data, the second layer feature data and the third layer feature data into a feature aggregation unit of an information fusion layer, wherein the feature aggregation unit comprises a channel attention module and a spatial attention module, wherein the channel attention module adopts a squeeze-and-excitation structure and the spatial attention module adopts a non-local attention structure, to obtain fused feature data; The fused feature data is processed by group belief propagation to obtain optimized feature data, and the optimized feature data is input into a feature reconstruction unit, wherein the feature reconstruction unit comprises an encoder and a decoder structure, the encoder comprises three downsampling layers, each layer uses a convolution operation with a step size of 2, the decoder comprises three upsampling layers, each layer uses a deconvolution operation, and a jump connection is added between the encoder and the decoder to obtain the first comprehensive feature data.

4. The business forecasting method based on the LNM large numerical model according to claim 3 is characterized in that: The performing of feature dimension reduction and optimization processing on the first comprehensive feature data, and obtaining the second comprehensive feature data through feature importance scoring and feature dependency analysis, includes: Performing feature correlation analysis on the first comprehensive feature data, obtaining a feature correlation matrix by calculating correlation coefficients between features, performing clustering processing on the features marked in the feature correlation matrix, classifying related features by a hierarchical clustering method, and obtaining feature clustering data; The feature clustering data is scored for importance, and feature importance data is obtained by calculating the variance contribution rate, information gain and feature weight of the feature; The feature importance data is sorted according to the scores, and features are screened by setting the importance threshold of 0.6 to obtain important feature data; Constructing a dependency graph for the important feature data, representing the dependency relationship between features through a graph structure, obtaining feature dependency data, and grouping the feature dependency data according to dependency strength, and obtaining feature grouping data by calculating the association strength between and within groups; Performing feature mapping on the feature grouping data, mapping the features to a low-dimensional space by linear combination to obtain reduced-dimensional feature data, and performing regularization processing on the reduced-dimensional feature data to obtain second comprehensive feature data.

5. The business forecasting method based on the LNM large numerical model according to claim 4 is characterized in that: The step of inputting the second comprehensive feature data into the LNM large numerical model integrating multi-scale gating prediction, performing business prediction through the long-term trend prediction unit and the short-term fluctuation prediction unit, and obtaining the business prediction result includes: Inputting the second comprehensive feature data into the LNM large numerical modeling layer, the LNM large numerical modeling layer comprises a three-layer large numerical analysis network, the first layer comprises 5 fully connected layers, each fully connected layer uses a ReLU activation function and a Dropout layer, the second layer comprises 3 linear regression layers, and the third layer comprises 2 nonlinear transformation layers to obtain numerical analysis features; The long-term trend component of the numerical analysis feature is enhanced, and the long-term trend prediction unit includes 4 gated recurrent network layers, each of which includes a data selection gate, a forget gate and an output gate. The data selection gate uses a sigmoid activation function, and the forget gate uses a tanh activation function to obtain long-term trend data; Performing multi-scale decomposition on the long-term trend data, the multi-scale decomposition includes three scale decomposition units, each decomposition unit includes two convolution layers and one pooling layer, the convolution kernel sizes of the convolution layers are 3×3 and 5×5 respectively, and the pooling layer adopts a maximum pooling operation to obtain multi-scale feature data; Inputting the multi-scale feature data into a short-term fluctuation prediction unit, wherein the short-term fluctuation prediction unit comprises 6 bidirectional gating units, each of which comprises a forward channel and a reverse channel, and each channel comprises 3 convolutional layers and a batch normalization layer, to obtain short-term fluctuation data; Performing gated fusion on the long-term trend data and the short-term fluctuation data, wherein the gated fusion controls the fusion ratio of features of different scales through a weighted gating matrix, wherein the size of the weighted gating matrix is ​​N×N, where N is the number of feature dimensions, to obtain fused prediction data; The fused prediction data is input into the LNM prediction layer, the LNM prediction layer includes 4 numerical prediction units, each prediction unit includes 3 linear transformation layers and 2 nonlinear transformation layers, the output dimensions of the linear transformation layer are 256, 128 and 64 respectively, to obtain an initial prediction result; Probability calibration is performed on the initial prediction result, wherein the probability calibration includes a 3-layer density estimation network, each layer includes 4 fully connected layers and a probability distribution layer, and the number of neurons in the fully connected layer is 128, 64, 32 and 16 respectively, to obtain calibration prediction data; The calibrated prediction data is subjected to threshold constraint processing, and the initial prediction result is corrected by setting upper and lower thresholds and confidence intervals. The upper and lower thresholds are the 95% quantile and 5% quantile of the historical data, respectively, and the confidence interval is set to 90%, to obtain the business prediction result.

6. The business forecasting method based on the LNM large numerical model according to claim 5 is characterized in that: The initial prediction result is probability calibrated, and the probability calibration includes a 3-layer density estimation network, each layer includes 4 fully connected layers and a probability distribution layer, and the number of neurons in the fully connected layer is 128, 64, 32 and 16 respectively, to obtain calibration prediction data, including: Input the initial prediction result into the first fully connected layer of the first density estimation network for feature mapping, the number of neurons in the first fully connected layer is 128, the rectified linear unit ReLU is used as the activation function, and the features are batch normalized with a mean of 0 and a variance of 1 to obtain the first layer of density feature data; The first layer of density feature data is sequentially input into the second fully connected layer, the third fully connected layer and the fourth fully connected layer of the first layer of density estimation network, the number of neurons in the second fully connected layer, the third fully connected layer and the fourth fully connected layer are set to 64, 32 and 16 respectively, each fully connected layer uses the hyperbolic tangent function tanh as the activation function, and a skip connection structure is added between adjacent fully connected layers, and the probability density is calculated through the probability distribution layer to obtain the first layer of probability distribution data; The Gaussian mixture model is calculated for the first layer of probability distribution data, the number of components of the mixed Gaussian distribution is set to 5, the expectation maximization algorithm is used to iteratively optimize the mixture weight parameters and Gaussian distribution parameters, the number of optimization iterations is set to 100 times, and the log-likelihood function value is calculated each time to obtain the first layer of calibration data; The first layer of calibration data is input into the four fully connected layers of the second layer density estimation network. The number of neurons in the four fully connected layers is set to 128, 64, 32 and 16 respectively. The activation function uses the rectified linear unit ReLU. A random dropout layer with an inactivation probability of 0.3 is set between every two adjacent layers. The probability density is calculated through the probability distribution layer to obtain the second layer of probability distribution data; The second-layer probability distribution data is subjected to cumulative distribution function calculation and inverse transformation sampling, the probability distribution is mapped to a uniform distribution interval of [0.05, 0.95], the number of sampling points is set to 1000, and the probability is calibrated by the quantile method to obtain the second-layer calibration data; The second layer calibration data is input into the four fully connected layers of the third layer density estimation network. The number of neurons in the four fully connected layers is set to 128, 64, 32 and 16 respectively. Each fully connected layer uses the rectified linear unit ReLU as the activation function, and a softmax normalization operation is added after the last fully connected layer. The probability density is calculated through the probability distribution layer to obtain the third layer probability distribution data; Performing kernel density estimation on the third-layer probability distribution data, using a Gaussian kernel function as the kernel function, setting the bandwidth parameter of the kernel function to 0.1, and calculating the smoothed probability density distribution by weighted averaging the probability density of each data point to obtain the third-layer calibration data; The first layer calibration data, the second layer calibration data and the third layer calibration data are weightedly fused according to weight coefficients of 0.3, 0.3 and 0.4, and the weighted average of the three layers of calibration data is calculated as calibration prediction data.

7. The business forecasting method based on the LNM large numerical model according to claim 1 is characterized in that: The method of performing distributed learning on the LNM large numerical model integrating multi-scale gating prediction according to the business prediction result, and iteratively optimizing the model through the local criticism module and the local participation module to obtain the target business prediction model includes: Input the business prediction result into the local criticism module, which includes a four-layer evaluation network structure, where the first layer includes 256 neurons, the second layer includes 128 neurons, the third layer includes 64 neurons, and the fourth layer includes 32 neurons. Each layer uses the LeakyReLU activation function to obtain local evaluation data; Performing value function calculation on the local evaluation data, evaluating the prediction results by using a temporal difference algorithm with a discount factor of 0.9, calculating the cumulative return value of the prediction deviation, and obtaining the strategy value data; The strategy value data is input into a local participation module, wherein the local participation module comprises three parallel strategy networks, each strategy network comprises a 5-layer feedforward network, and the number of neurons is 512, 256, 128, 64 and 32 respectively, and a tanh activation function is used to obtain strategy parameter data; Performing proximal policy optimization on the policy parameter data, constraining the policy gradient by setting a loss function with a clipping parameter of 0.2, and setting an entropy regularization coefficient of 0.01, to obtain optimized policy data; The optimization strategy data is task-assigned according to a distributed architecture, and distributed learning data is obtained by setting 8 parallel observer nodes, each of which includes an independent local criticism module and a local participation module; Perform parameter synchronization processing on the distributed learning data, and summarize and update the parameters of each observer node by setting an asynchronous parameter server with a synchronization period of 100 steps to obtain global parameter data; The global parameter data is input into the LNM large numerical model integrating multi-scale gating prediction to update the parameters, and the model weight is optimized by setting the learning rate to 0.001 and the number of iterations to 1000 to obtain the optimized model data; The optimization model data is validated, and the model performance is evaluated by calculating the root mean square value and the mean absolute percentage error of the prediction error. When the performance index change for five consecutive iterations is less than 0.001, the optimization is stopped to obtain the target business prediction model.

8. A business prediction device based on LNM large numerical model, characterized in that: The device is used to execute the business prediction method based on the LNM large numerical model according to any one of claims 1 to 7, comprising: The preprocessing module is used to preprocess and multi-scale decompose the internal historical business data and external macro, meso and micro related data to obtain long-term trend characteristic data and short-term fluctuation characteristic data; A feature processing module, used for inputting the long-term trend feature data and the short-term fluctuation feature data into a deep business association network, performing feature processing through a plate association layer, a high-order feature extraction layer and an information fusion layer, and obtaining first comprehensive feature data; An analysis module, used to perform feature dimension reduction and optimization processing on the first comprehensive feature data, and obtain second comprehensive feature data through feature importance scoring and feature dependency analysis; A business prediction module, used for inputting the second comprehensive feature data into the LNM large numerical model integrating multi-scale gating prediction, performing business prediction through a long-term trend prediction unit and a short-term fluctuation prediction unit, and obtaining a business prediction result; The iterative optimization module is used to perform distributed learning on the LNM large numerical model integrating multi-scale gating prediction according to the business prediction results, and iteratively optimize through the local criticism module and the local participation module to obtain the target business prediction model.

9. A business prediction device based on LNM large numerical model, characterized in that: The business prediction device based on the LNM large numerical model includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor calls the instruction in the memory so that the business prediction device based on the LNM large numerical model executes the business prediction method based on the LNM large numerical model as described in any one of claims 1-7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the business prediction method based on the LNM large numerical model as described in any one of claims 1 to 7 is implemented.

Images