Bill sequence prediction method and system based on composite neural network

Abstract

The invention provides a bill sequence prediction method and system based on a composite neural network in the technical field of bill management, and the method comprises the steps: S1, obtaining a large number of historical electronic bills, sequentially carrying out the data cleaning, structural aggregation and feature extraction of all historical electronic bills, and obtaining a plurality of historical electronic bills; marking the historical electronic bills based on the extracted features, segmenting the historical electronic bills through a sliding window to construct sample pairs, and constructing a data set based on the sample pairs; s2, constructing a bill sequence prediction model based on the input layer, the local feature extraction layer, the global feature extraction layer, the feature enhancement layer and the output layer, and setting a loss function; s3, training the bill sequence prediction model through the data set and the loss function, and deploying the trained bill sequence prediction model; and S4, performing bill sequence prediction through the deployed bill sequence prediction model. The method has the advantage that the accuracy and robustness of bill sequence prediction are greatly improved.

Description

Technical Field

[0001] This invention relates to the field of bill management technology, and in particular to a bill sequence prediction method and system based on a composite neural network. Background Technology

[0002] With the rapid development of e-government, digital currency, and digital finance, electronic invoices (such as e-invoices) are gradually replacing traditional paper invoices, becoming the core voucher in the financial management systems of universities, research institutions, hospitals, and many other organizations. In these organizations, key financial processes, including invoice reimbursement, budget control, and compliance auditing, all heavily rely on the efficient and stable operation of electronic invoice systems.

[0003] Currently, for predicting future trends in the quantity or amount of electronic invoices (to support financial cash flow management, audit resource allocation, etc.), relevant entities mainly employ empirical methods based on historical data, rule-based methods, or traditional statistical models (e.g., moving average method, linear regression method). However, these methods have significant shortcomings in the specific scenario of electronic invoice prediction:

[0004] 1. Difficulty in effectively modeling periodic patterns: Electronic invoice data often exhibits complex periodic fluctuations, such as the significant "peak season at the start of the semester" and "low season during winter and summer vacations" in university settings. Traditional linear models struggle to accurately capture these inherent patterns. 2. Delayed response to unforeseen events: Unforeseen events frequently occur in reality (such as the launch of temporary projects, the holding of large-scale events, and adjustments to fiscal policies), causing drastic changes in the volume or amount of invoices. Models based on historical statistics are often slow to react to such sudden changes, with predictions lagging behind actual changes. 3. Lack of deep nonlinear feature extraction capabilities: Invoice data may contain complex nonlinear correlation features; for example, large-scale purchases by a specific department in a particular month may exhibit specific data patterns. Traditional methods generally rely on simple linear or shallow nonlinear transformations, making it difficult to effectively fit such deep, nonlinear relationships.

[0005] In recent years, deep learning technology has been gradually introduced into the field of time series forecasting. Among them, Long Short-Term Memory (LSTM) networks excel at capturing long-term dependencies in sequences, Convolutional Neural Networks (CNNs) effectively extract local trend features, and attention-based models (such as Transformers) demonstrate advantages in multivariate modeling. However, directly applying these models to electronic invoice sequence forecasting still faces the following challenges:

[0006] 1. Limited Model Structure and Insufficient Feature Fusion: Existing solutions typically employ a single neural network structure (such as using LSTM, CNN, or Transformer alone), lacking an effective mechanism for efficient collaborative modeling of different temporal features (such as long-term trends, periodic patterns, local details, and external event influences). 2. Complex Pattern Coupling Leads to Large Errors: The inherent periodicity and frequent mutations in electronic invoice data are intertwined, making it difficult for a single model structure to accurately characterize both characteristics simultaneously, easily leading to the accumulation and amplification of prediction errors. 3. Limited Nonlinear Representation Capability: When handling complex nonlinear mappings, existing technologies generally rely on traditional fully connected networks for the final prediction output, which have limited function approximation capabilities and struggle to fully explore the deep nonlinear correlations within the data.

[0007] Therefore, how to provide a bill sequence prediction method and system based on composite neural networks to improve the accuracy and robustness of bill sequence prediction has become an urgent technical problem to be solved. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a method and system for predicting bill sequences based on a composite neural network, so as to improve the accuracy and robustness of bill sequence prediction.

[0009] In a first aspect, the present invention provides a method for predicting ticket sequences based on a composite neural network, comprising the following steps:

[0010] Step S1: Obtain a large number of historical electronic tickets, perform data cleaning, structured aggregation and feature group extraction on each of the historical electronic tickets in sequence, label each of the historical electronic tickets based on the extracted feature groups, and construct sample pairs by segmenting each of the historical electronic tickets through a sliding window, and construct a dataset based on each of the sample pairs.

[0011] Step S2: Construct a ticket sequence prediction model based on the input layer, local feature extraction layer, global feature extraction layer, feature enhancement layer and output layer, and set the loss function of the ticket sequence prediction model;

[0012] Step S3: Train the ticket sequence prediction model using the dataset and loss function, and deploy the trained ticket sequence prediction model.

[0013] Step S4: Predict the bill sequence using the deployed bill sequence prediction model.

[0014] Furthermore, step S1 specifically includes:

[0015] A large number of historical electronic invoices are obtained from a preset data source using an ETL tool. The historical electronic invoices carry at least the invoice date, invoice amount, project name, project code, invoice unit, invoice type, region, project amount, and project quantity.

[0016] The historical electronic invoices are cleaned by at least the following steps: standardizing the invoice date format, handling missing values, handling outliers, handling duplicate invoices, handling empty invoices, and handling abnormal invoices. Using the combination of the invoicing unit and invoice type as the granularity, the cleaned historical electronic invoices are aggregated at preset time intervals to construct a time-series sample table. The time interval is daily, weekly, or monthly. Feature groups, including at least periodic features, categorical variable features, and statistical features, are extracted from the time-series sample table.

[0017] After feature normalization and encoding of the extracted feature group, each historical electronic ticket is labeled based on the feature group, and each historical electronic ticket is segmented by a sliding window based on preset sliding parameters to construct sample pairs including input data and prediction data. A dataset is constructed based on each sample pair; the sliding parameters include the length of the input data, the length of the prediction data, and the sliding step size.

[0018] Furthermore, in step S2, the input layer is constructed based on a time series input module, a feature input module, an embedding vector input module, and a concatenation module;

[0019] The time series input module is used to extract time series data from the input electronic invoices; the feature input module is used to extract periodic features from the input electronic invoices; the embedding vector input module is used to extract dense vectors of categorical variables from the input electronic invoices; the categorical variables include project name, invoice type, region, and issuing unit; the concatenation module is used to concatenate the time series data, periodic features, and dense vectors into input invoice data;

[0020] The local feature extraction layer is used to extract temporal local pattern features from the input ticket data through a one-dimensional convolutional neural network;

[0021] The global feature extraction layer is used to extract long-term dependency features from the input ticket data through a long short-term memory network.

[0022] The feature enhancement layer is used to enhance and fuse temporal local pattern features and long-term dependency features through a Kolmogorov network to obtain enhanced features.

[0023] The output layer is used to output the bill sequence prediction result with a preset prediction time based on the enhanced features; the bill sequence prediction result carries the total amount of future invoices, the number of future bills, the amount of future projects, and a high-risk volatility score;

[0024] The formula for the loss function is:

[0025] L total =α*L1+β*L2+γ*L3;

[0026] Among them, L total L1, L2, and L3 represent the loss value of the loss function; L1, L2, and L3 represent the loss of invoice amount, loss of invoice number, and loss of average amount, respectively, and all use the mean square error function; α, β, and γ all represent weighting coefficients, and α+β+γ=1.

[0027] Furthermore, step S3 specifically includes:

[0028] The dataset is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The ticket sequence prediction model is trained using the training set until the loss value of the loss function is less than a preset loss threshold.

[0029] The mean squared error, mean absolute error, and prediction accuracy are calculated using the validation set to validate the trained ticket sequence prediction model. If the validation fails, the training set is expanded and training continues; if the validation succeeds, then:

[0030] The confidence level is calculated using the test set to test the validated invoice sequence prediction model. If the test fails, the training set is expanded and training continues. If the test passes, training ends and the validated invoice sequence prediction model is deployed using containerization technology.

[0031] Furthermore, step S4 specifically includes:

[0032] The system acquires real-time electronic tickets, preprocesses them using a streaming engine, and then inputs them into a deployed ticket sequence prediction model. The ticket sequence prediction model performs inference using hardware acceleration technology and outputs real-time ticket sequence prediction results for ticket sequence prediction.

[0033] Secondly, the present invention provides a bill sequence prediction system based on a composite neural network, comprising the following modules:

[0034] The dataset construction module is used to acquire a large number of historical electronic tickets, perform data cleaning, structured aggregation and feature group extraction on each of the historical electronic tickets in sequence, label each of the historical electronic tickets based on the extracted feature groups, and construct sample pairs by segmenting each of the historical electronic tickets through a sliding window, and construct a dataset based on each of the sample pairs.

[0035] The bill sequence prediction model construction module is used to construct a bill sequence prediction model based on an input layer, a local feature extraction layer, a global feature extraction layer, a feature enhancement layer, and an output layer, and to set the loss function of the bill sequence prediction model.

[0036] The bill sequence prediction model training module is used to train the bill sequence prediction model using the dataset and loss function, and to deploy the trained bill sequence prediction model.

[0037] The bill sequence prediction module is used to predict bill sequences using the deployed bill sequence prediction model.

[0038] Furthermore, the dataset construction module is specifically used for:

[0039] A large number of historical electronic invoices are obtained from a preset data source using an ETL tool. The historical electronic invoices carry at least the invoice date, invoice amount, project name, project code, invoice unit, invoice type, region, project amount, and project quantity.

[0040] The historical electronic invoices are cleaned by at least the following steps: standardizing the invoice date format, handling missing values, handling outliers, handling duplicate invoices, handling empty invoices, and handling abnormal invoices. Using the combination of the invoicing unit and invoice type as the granularity, the cleaned historical electronic invoices are aggregated at preset time intervals to construct a time-series sample table. The time interval is daily, weekly, or monthly. Feature groups, including at least periodic features, categorical variable features, and statistical features, are extracted from the time-series sample table.

[0041] After feature normalization and encoding of the extracted feature group, each historical electronic ticket is labeled based on the feature group, and each historical electronic ticket is segmented by a sliding window based on preset sliding parameters to construct sample pairs including input data and prediction data. A dataset is constructed based on each sample pair; the sliding parameters include the length of the input data, the length of the prediction data, and the sliding step size.

[0042] Furthermore, in the bill sequence prediction model construction module, the input layer is constructed based on a time series input module, a feature input module, an embedding vector input module, and a splicing module;

[0043] The time series input module is used to extract time series data from the input electronic invoices; the feature input module is used to extract periodic features from the input electronic invoices; the embedding vector input module is used to extract dense vectors of categorical variables from the input electronic invoices; the categorical variables include project name, invoice type, region, and issuing unit; the concatenation module is used to concatenate the time series data, periodic features, and dense vectors into input invoice data;

[0044] The local feature extraction layer is used to extract temporal local pattern features from the input ticket data through a one-dimensional convolutional neural network;

[0045] The global feature extraction layer is used to extract long-term dependency features from the input ticket data through a long short-term memory network.

[0046] The feature enhancement layer is used to enhance and fuse temporal local pattern features and long-term dependency features through a Kolmogorov network to obtain enhanced features.

[0047] The output layer is used to output the bill sequence prediction result with a preset prediction time based on the enhanced features; the bill sequence prediction result carries the total amount of future invoices, the number of future bills, the amount of future projects, and a high-risk volatility score;

[0048] The formula for the loss function is:

[0049] L total =α*L1+β*L2+γ*L3;

[0050] Among them, L total L1, L2, and L3 represent the loss value of the loss function; L1, L2, and L3 represent the loss of invoice amount, loss of invoice number, and loss of average amount, respectively, and all use the mean square error function; α, β, and γ all represent weighting coefficients, and α+β+γ=1.

[0051] Furthermore, the ticket sequence prediction model training module is specifically used for:

[0052] The dataset is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The ticket sequence prediction model is trained using the training set until the loss value of the loss function is less than a preset loss threshold.

[0053] The mean squared error, mean absolute error, and prediction accuracy are calculated using the validation set to validate the trained ticket sequence prediction model. If the validation fails, the training set is expanded and training continues; if the validation succeeds, then:

[0054] The confidence level is calculated using the test set to test the validated invoice sequence prediction model. If the test fails, the training set is expanded and training continues. If the test passes, training ends and the validated invoice sequence prediction model is deployed using containerization technology.

[0055] Furthermore, the ticket sequence prediction module is specifically used for:

[0056] The system acquires real-time electronic tickets, preprocesses them using a streaming engine, and then inputs them into a deployed ticket sequence prediction model. The ticket sequence prediction model performs inference using hardware acceleration technology and outputs real-time ticket sequence prediction results for ticket sequence prediction.

[0057] The advantages of this invention are:

[0058] 1. By acquiring a large number of historical electronic tickets, data cleaning, structured aggregation, and feature group extraction are performed on each historical electronic ticket sequentially. Based on the extracted feature groups, each historical electronic ticket is labeled, and sample pairs are constructed by segmenting each historical electronic ticket using a sliding window. A dataset is then built based on these sample pairs. Next, a ticket sequence prediction model is constructed based on an input layer, a local feature extraction layer, a global feature extraction layer, a feature enhancement layer, and an output layer. A loss function is set for the ticket sequence prediction model, and it is trained using the dataset and the loss function. The trained ticket sequence prediction model is then deployed, and finally, ticket sequence prediction is performed using the deployed model. In other words, the time series data and periodic features of electronic tickets are explicitly extracted through the input layer. Dense vectors, combined with a one-dimensional convolutional neural network (CNN) in the local feature extraction layer to capture the temporal local pattern features (local details and sudden fluctuation patterns) of bill data, and a long short-term memory network (LSTM) in the global feature extraction layer to model long-term dependent features, and then a Kolmogorov network (KAN) in the feature enhancement layer to perform deep nonlinear fusion and enhancement of the above multi-source heterogeneous features, thereby synergistically solving the core problems of traditional methods that are difficult to characterize complex periodic patterns, have a lag in response to sudden events, and have insufficient nonlinear expression capabilities. Finally, it achieves accurate and stable prediction of multi-dimensional indicators such as bill amount (future total bill amount, future project amount), bill quantity (future bill quantity), and high-risk fluctuation score, thereby greatly improving the accuracy and robustness of bill sequence prediction.

[0059] 2. Improved prediction accuracy and stability: Compared with traditional single time series models (such as ARIMA, pure LSTM, etc.), by integrating CNN to extract local temporal pattern features, LSTM to model long-term dependent features, and KAN to construct nonlinear knowledge interpolation (feature enhancement), the prediction error in high-fluctuation periods (such as the start of the school season and fiscal settlement period) is effectively reduced, and the overall prediction accuracy is significantly improved.

[0060] 3. Enhance the ability to perceive cyclical changes in non-tax business: By introducing cyclical auxiliary features such as "month", "week", and "whether it is the start of school season / holiday", it can adaptively capture annual business patterns; for example, universities concentrate on issuing invoices in September each year, and medical institutions experience peak periods before and after holidays, thereby better monitoring and forecasting seasonal financial data.

[0061] 4. Enable parallel output of multi-objective business forecasts: It can simultaneously output multiple key indicators (invoice amount, number of invoices, proportion of invoice types, average project amount, etc.), supporting multi-angle analysis and early warning, effectively supporting budget preparation, revenue forecasting, performance evaluation and other work scenarios of universities, hospitals and other units.

[0062] 5. Possesses cross-unit and cross-regional migration and generalization capabilities: Utilizing the model's modeling capabilities in the "Invoicing Unit" and "Regional Division" fields, it can achieve generalized predictions for different units (such as multiple universities or multiple hospitals) and different regions (provinces, cities, districts), possessing good promotion and deployment value.

[0063] 6. Supports rolling updates and dynamic predictions to adapt to actual business needs: The model adopts a sliding window data generation method and a rolling prediction mechanism, which allows the model to continuously receive new data for iterative training and updates. It supports the combination of daily predictions and medium- and long-term budget forecasts, improving the practicality and robustness of the model's online services.

[0064] 7. Enhance the automation level of intelligent management of non-tax invoices: By constructing an invoice sequence prediction model, the dynamic trend of non-tax revenue expenditure can be grasped in advance, realizing "pre-emptive monitoring" and "risk prediction" of invoice issuance behavior, and improving the information and intelligent management capabilities of electronic invoice business. Attached Figure Description

[0065] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0066] Figure 1 This is a flowchart of a bill sequence prediction method based on a composite neural network according to the present invention.

[0067] Figure 2 This is a schematic diagram of the structure of a bill sequence prediction system based on a composite neural network according to the present invention. Figure 3This is a schematic diagram of the one-dimensional convolutional neural network of the present invention. Figure 4 This is a schematic diagram of the Long Short-Term Memory (LSTM) network of the present invention. Figure 5 This is a schematic diagram of the Kolmogorov network of the present invention. Detailed Implementation

[0068] The technical solution in this application embodiment has the following general idea: the time series data, periodic features, and dense vectors of electronic invoices are explicitly extracted by the input layer. A one-dimensional convolutional neural network (CNN) is used to capture the temporal local pattern features of the invoice data. A long short-term memory network (LSTM) is used to model long-term dependent features. Then, a Kolmogorov network (KAN) is used to perform deep nonlinear fusion and enhancement of the above-mentioned multi-source heterogeneous features. This collaboratively solves the core problems of traditional methods, such as difficulty in characterizing complex periodic patterns, lag in response to sudden events, and insufficient nonlinear expression ability. Finally, it achieves accurate and stable prediction of multi-dimensional indicators such as invoice amount, invoice quantity, and high-risk fluctuation score, thereby improving the accuracy and robustness of invoice sequence prediction.

[0069] Please refer to Figures 1 to 5 As shown, a preferred embodiment of the present invention, a bill sequence prediction method based on a composite neural network, includes the following steps:

[0070] Step S1: Obtain a large number of historical electronic tickets, perform data cleaning, structured aggregation and feature group extraction on each of the historical electronic tickets in sequence, label each of the historical electronic tickets based on the extracted feature groups, and construct sample pairs by segmenting each of the historical electronic tickets through a sliding window, and construct a dataset based on each of the sample pairs.

[0071] Step S2: Construct a ticket sequence prediction model based on the input layer, local feature extraction layer, global feature extraction layer, feature enhancement layer and output layer, and set the loss function of the ticket sequence prediction model;

[0072] Step S3: Train the ticket sequence prediction model using the dataset and loss function, and deploy the trained ticket sequence prediction model.

[0073] Step S4: Predict the bill sequence using the deployed bill sequence prediction model.

[0074] This invention proposes a periodic electronic invoice prediction method integrating CNN-LSTM-KAN networks, primarily targeting time series data prediction scenarios with periodic and nonlinear characteristics. It is particularly suitable for predicting the amount or quantity of electronic invoices in settings such as universities and research institutions. By deeply fusing Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), and Kolmogorov Networks (KAN), the prediction accuracy and the model's adaptability to periodic fluctuations are effectively improved.

[0075] Step S1 specifically involves:

[0076] A large number of historical electronic invoices are obtained from a preset data source using an ETL tool. These historical electronic invoices carry at least the invoice date, invoice amount, project name, project code, invoicing unit, invoice type, region, project amount, and project quantity. These historical electronic invoices are non-tax revenue electronic invoices.

[0077] The data fields included in the historical electronic tickets are described below:

[0078]

[0079] The historical electronic invoices undergo data cleaning, including at least the following steps: unifying the invoice date format (e.g., converting to a standard date format: YYYY-MM-DD), handling missing values ​​(filling missing fields with the mapping relationship between project code and project name, and then deleting missing key fields (such as invoice amount and project code)), handling outliers (performing anomaly detection on amount fields (such as negative values), and removing or marking suspicious data based on the 3 sigma principle), handling duplicate invoices (merging duplicate records of the same unit and the same invoice number), handling empty invoices, and handling abnormal invoices (deleting illegal invoices with a project quantity of 0 but an amount of 0). Using the combination of the invoice unit and invoice type as the granularity, the cleaned historical electronic invoices are aggregated at preset time intervals to construct a time-series sample table; the time interval is daily, weekly, or monthly. Feature groups, including at least periodic features, categorical variable features, and statistical features, are extracted from the time-series sample table.

[0080] An example of a time series sample table is as follows:

[0081]

[0082] Where: Total Amount = Sum of aggregated invoice amounts; Total Number of Items = Cumulative sum of the Item Quantity field; Average Single Item Amount = Sum of Item Amounts ÷ Item Quantity; Total Number of Invoices = Aggregate Row Count.

[0083] To capture potential cyclical patterns in non-tax revenue (such as peak seasons for the start of the school year, winter and summer vacations, and medical insurance settlement periods), we extract cyclical features, as exemplified below:

[0084]

[0085] The process for processing categorical variable features is as follows:

[0086] Project Name / Code: The project name is converted into an Embedding representation based on the Qwen3-Embedding-0.6B model.

[0087] Invoicing Unit: The invoicing methods differ greatly among different types of units, and the patterns of large units within the same category are even more unique. Therefore:

[0088] Categorize by type of organization (e.g., "universities", "hospitals", "government departments");

[0089] Large organizations (such as top-tier hospitals and key universities) should retain separate codes.

[0090] Ticket types and regions: Each type is coded independently, or "province + city + district (county)" is used as an enumeration type code.

[0091] Statistical features are auxiliary features constructed by combining business common sense, such as: average amount per invoice = total amount / total number of invoices; average amount per item = item amount / number of items; daily unit income density (total unit income / number of items); proportion of high-frequency items (such as the proportion of "registration fee" and "tuition fee" in the total amount).

[0092] After feature normalization and encoding of the extracted feature groups (logarithmic transformation + MinMax normalization is used for monetary features; frequency encoding (low-frequency merging) or embedding is used for categorical variables; periodic features (such as months) are encoded in sin / cos form to capture the period), each historical electronic ticket is labeled based on the feature groups, and each historical electronic ticket is segmented by a sliding window based on preset sliding parameters to construct sample pairs including input data and predicted data, and a dataset is constructed based on each sample pair; the sliding parameters include the length of input data (such as 60 days), the length of predicted data (such as the next 7 days or 30 days), and the sliding step size (such as 1 day).

[0093] An example of how each sample pair is constructed is as follows:

[0094] Input X: T-day bill amount sequence + time period features + categorical variable features;

[0095] Output Y: Predicts the total invoice amount, number of invoices, number of items, or average amount per item for the next P days.

[0096] The sample construction is illustrated below:

[0097] Sample No. 1:

[0098] Time period X: September 1, 2023 to October 30, 2023 (60 days in total)

[0099] X-feature:

[0100]

[0101]

[0102] Target Y: Total invoice amount (7-day rolling) from November 1, 2023 to November 7, 2023

[0103] date Invoice amount 2023 / 11 / 1 7000 2023 / 11 / 2 8000 2023 / 11 / 3 6000 2023 / 11 / 4 0 2023 / 11 / 5 0 2023 / 11 / 6 5000 2023 / 11 / 7 9000

[0104] Sample No. 2:

[0105] Time period X: 2024 / 03 / 01~2024 / 04 / 30

[0106] X-feature:

[0107]

[0108]

[0109] Target Y: Total invoices issued in May 2024

[0110] In step S2, the input layer is constructed based on a time series input module, a feature input module, an embedded vector input module, and a concatenation module.

[0111] The time series input module is used to extract time series data (such as invoice amount series, invoice quantity series, and average price series) from the input electronic invoices; the feature input module is used to extract periodic features (such as month, whether it is the start of the school season, and whether it is winter or summer vacation) from the input electronic invoices; the embedding vector input module is used to extract dense vectors (learnable) of categorical variables from the input electronic invoices; the categorical variables include project name, invoice type, region, and invoicing unit; the concatenation module is used to concatenate the time series data, periodic features, and dense vectors into input invoice data;

[0112]

[0113] Where T represents the time step (e.g., 60 days); D represents the feature dimension of a single step; Represents a real number.

[0114] The local feature extraction layer is used to extract temporal local pattern features (such as high-frequency fluctuations and abrupt change nodes) from the input ticket data through a one-dimensional convolutional neural network.

[0115] The core idea is to identify typical cycles (such as monthly or weekly) by sliding several sliding convolution kernels along the time dimension. Its advantage is that it is suitable for dimensionality reduction of high-dimensional time series and can discover patterns in time "segments", such as the peak fluctuations on the monthly settlement days of a hospital.

[0116] Y_cnn=Conv1D(X_input, kernel_size=3~7, stride=1)

[0117] The global feature extraction layer is used to extract long-term dependent features from the input invoice data through a long short-term memory network (capturing trend evolution across time, especially suitable for predicting trends across months and quarters, such as predicting fluctuations in invoice volume before and after summer and winter vacations).

[0118] Y_lstm=LSTM(units=64~128)(Y_cnn)

[0119] LSTM effectively handles the following characteristics: strong sequence correlation (such as invoice volume related to three consecutive months); strong periodicity (such as price increases every September during the school season); and node-dependent nonlinearity (such as policy adjustments or medical insurance reimbursement windows).

[0120] The feature enhancement layer is used to enhance and fuse temporal local pattern features and long-term dependency features through a Kolmogorov network (knowledge enhancement network) to obtain enhanced features;

[0121] The core idea of ​​Kolmogorov networks is to learn nonlinear mappings over time using explicit piecewise functions (such as ReLU combinations); and to guide the model to be more interpretable and generalizable in scenarios where "structural priors" such as units, items, and seasons exist.

[0122] Kolmogorov networks are particularly well-suited for tasks such as: structural shifts in school non-tax revenue, such as "peak season surges" and "zero returns during summer and winter breaks"; modeling heterogeneous behaviors of "specific units + specific invoices" (e.g., significant differences between universities and hospitals); and improving the model's ability to fit boundary and peak nodes. In this invention, Kolmogorov networks are specifically used to: strengthen the nonlinear relationship between time, units, and items; make more accurate fits to "breakpoints" and "structural changes" in long-term trends; and reduce the smoothing trend error of LSTM in long-term predictions.

[0123] Y_kan = KAN_Layer()(Y_lstm)

[0124] The output layer is used to output the bill sequence prediction result with a preset prediction time based on the enhanced features; the bill sequence prediction result carries the total amount of future invoices, the number of future bills, the amount of future projects, and a high-risk volatility score;

[0125] The output layer supports a multi-output, multi-task modeling structure. Each prediction head can share underlying features or model independently. In specific implementation, the following prediction objectives can be set as needed:

[0126] Output name type illustrate Total amount of future invoices Regression value Forecast the amount of bills for the next 7 / 30 / 365 days Future bill quantity Regression value Total number of bills predicted by the unit Future project amount Regression value The unit predicts the total amount of a certain project. High-risk volatility score Fraction Is there a risk of abnormal surge / decline?

[0127] The formula for the loss function is:

[0128] L total =α*L1+β*L2+γ*L3;

[0129] Among them, L total The loss function represents the loss value; L1, L2, and L3 represent the loss on invoice amount, loss on number of invoices, and loss on average amount, respectively, and all use the mean squared error function; α, β, and γ represent weighting coefficients, and α + β + γ = 1. In practice, L1 regularization restrictions can be added to enhance interpretability.

[0130] Step S3 specifically involves:

[0131] The dataset is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The ticket sequence prediction model is trained using the training set until the loss value of the loss function is less than a preset loss threshold.

[0132] The mean squared error, mean absolute error, and prediction accuracy are calculated using the validation set to validate the trained ticket sequence prediction model. If the validation fails, the training set is expanded and training continues; if the validation succeeds, then:

[0133] The confidence level is calculated using the test set to test the validated invoice sequence prediction model. If the test fails, the training set is expanded and training continues. If the test passes, training ends and the validated invoice sequence prediction model is deployed using containerization technology.

[0134] Step S4 specifically involves:

[0135] The system acquires real-time electronic tickets, preprocesses them using a streaming engine, and then inputs them into a deployed ticket sequence prediction model. The ticket sequence prediction model performs inference using hardware acceleration technology and outputs real-time ticket sequence prediction results for ticket sequence prediction.

[0136] A preferred embodiment of the bill sequence prediction system based on a composite neural network of the present invention includes the following modules:

[0137] The dataset construction module is used to acquire a large number of historical electronic tickets, perform data cleaning, structured aggregation and feature group extraction on each of the historical electronic tickets in sequence, label each of the historical electronic tickets based on the extracted feature groups, and construct sample pairs by segmenting each of the historical electronic tickets through a sliding window, and construct a dataset based on each of the sample pairs.

[0138] The bill sequence prediction model construction module is used to construct a bill sequence prediction model based on an input layer, a local feature extraction layer, a global feature extraction layer, a feature enhancement layer, and an output layer, and to set the loss function of the bill sequence prediction model.

[0139] The bill sequence prediction model training module is used to train the bill sequence prediction model using the dataset and loss function, and to deploy the trained bill sequence prediction model.

[0140] The bill sequence prediction module is used to predict bill sequences using the deployed bill sequence prediction model.

[0141] This invention proposes a periodic electronic invoice prediction method integrating CNN-LSTM-KAN networks, primarily targeting time series data prediction scenarios with periodic and nonlinear characteristics. It is particularly suitable for predicting the amount or quantity of electronic invoices in settings such as universities and research institutions. By deeply fusing Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), and Kolmogorov Networks (KAN), the prediction accuracy and the model's adaptability to periodic fluctuations are effectively improved.

[0142] The dataset construction module is specifically used for:

[0143] A large number of historical electronic invoices are obtained from a preset data source using an ETL tool. These historical electronic invoices carry at least the invoice date, invoice amount, project name, project code, invoicing unit, invoice type, region, project amount, and project quantity. These historical electronic invoices are non-tax revenue electronic invoices.

[0144] The data fields included in the historical electronic tickets are described below:

[0145]

[0146] The historical electronic invoices undergo data cleaning, including at least the following steps: unifying the invoice date format (e.g., converting to a standard date format: YYYY-MM-DD), handling missing values ​​(filling missing fields with the mapping relationship between project code and project name, and then deleting missing key fields (such as invoice amount and project code)), handling outliers (performing anomaly detection on amount fields (such as negative values), and removing or marking suspicious data based on the 3 sigma principle), handling duplicate invoices (merging duplicate records of the same unit and the same invoice number), handling empty invoices, and handling abnormal invoices (deleting illegal invoices with a project quantity of 0 but an amount of 0). Using the combination of the invoice unit and invoice type as the granularity, the cleaned historical electronic invoices are aggregated at preset time intervals to construct a time-series sample table; the time interval is daily, weekly, or monthly. Feature groups, including at least periodic features, categorical variable features, and statistical features, are extracted from the time-series sample table.

[0147] An example of a time series sample table is as follows:

[0148]

[0149] Where: Total Amount = Sum of aggregated invoice amounts; Total Number of Items = Cumulative sum of the Item Quantity field; Average Single Item Amount = Sum of Item Amounts ÷ Item Quantity; Total Number of Invoices = Aggregate Row Count.

[0150] To capture potential cyclical patterns in non-tax revenue (such as peak seasons for the start of the school year, winter and summer vacations, and medical insurance settlement periods), we extract cyclical features, as exemplified below:

[0151]

[0152] The process for processing categorical variable features is as follows:

[0153] Project Name / Code: The project name is converted into an Embedding representation based on the Qwen3-Embedding-0.6B model.

[0154] Invoicing Unit: The invoicing methods differ greatly among different types of units, and the patterns of large units within the same category are even more unique. Therefore:

[0155] Categorize by type of organization (e.g., "universities", "hospitals", "government departments");

[0156] Large organizations (such as top-tier hospitals and key universities) should retain separate codes.

[0157] Ticket types and regions: Each type is coded independently, or "province + city + district (county)" is used as an enumeration type code.

[0158] Statistical features are auxiliary features constructed by combining business common sense, such as: average amount per invoice = total amount / total number of invoices; average amount per item = item amount / number of items; daily unit income density (total unit income / number of items); proportion of high-frequency items (such as the proportion of "registration fee" and "tuition fee" in the total amount).

[0159] After feature normalization and encoding of the extracted feature groups (logarithmic transformation + MinMax normalization is used for monetary features; frequency encoding (low-frequency merging) or embedding is used for categorical variables; periodic features (such as months) are encoded in sin / cos form to capture the period), each historical electronic ticket is labeled based on the feature groups, and each historical electronic ticket is segmented by a sliding window based on preset sliding parameters to construct sample pairs including input data and predicted data, and a dataset is constructed based on each sample pair; the sliding parameters include the length of input data (such as 60 days), the length of predicted data (such as the next 7 days or 30 days), and the sliding step size (such as 1 day).

[0160] An example of how each sample pair is constructed is as follows:

[0161] Input X: T-day bill amount sequence + time period features + categorical variable features;

[0162] Output Y: Predicts the total invoice amount, number of invoices, number of items, or average amount per item for the next P days.

[0163] The sample construction is illustrated below:

[0164] Sample No. 1:

[0165] Time period X: September 1, 2023 to October 30, 2023 (60 days in total)

[0166] X-feature:

[0167]

[0168]

[0169] Target Y: Total invoice amount (7-day rolling) from November 1, 2023 to November 7, 2023

[0170] date Invoice amount 2023 / 11 / 1 7000 2023 / 11 / 2 8000 2023 / 11 / 3 6000 2023 / 11 / 4 0 2023 / 11 / 5 0 2023 / 11 / 6 5000 2023 / 11 / 7 9000

[0171] Sample No. 2:

[0172] X Time Period: 2024 / 03 / 01~2024 / 04 / 30 X Characteristics:

[0173]

[0174]

[0175] Target Y: Total invoices issued in May 2024

[0176] In the bill sequence prediction model construction module, the input layer is constructed based on a time series input module, a feature input module, an embedding vector input module, and a splicing module.

[0177] The time series input module is used to extract time series data (such as invoice amount series, invoice quantity series, and average price series) from the input electronic invoices; the feature input module is used to extract periodic features (such as month, whether it is the start of the school season, and whether it is winter or summer vacation) from the input electronic invoices; the embedding vector input module is used to extract dense vectors (learnable) of categorical variables from the input electronic invoices; the categorical variables include project name, invoice type, region, and invoicing unit; the concatenation module is used to concatenate the time series data, periodic features, and dense vectors into input invoice data;

[0178]

[0179] Where T represents the time step (e.g., 60 days); D represents the feature dimension of a single step; Represents a real number.

[0180] The local feature extraction layer is used to extract temporal local pattern features (such as high-frequency fluctuations and abrupt change nodes) from the input ticket data through a one-dimensional convolutional neural network.

[0181] The core idea is to identify typical cycles (such as monthly or weekly) by sliding several sliding convolution kernels along the time dimension. Its advantage is that it is suitable for dimensionality reduction of high-dimensional time series and can discover patterns in time "segments", such as the peak fluctuations on the monthly settlement days of a hospital.

[0182] Y_cnn=Conv1D(X_input, kernel_size=3~7, stride=1)

[0183] The global feature extraction layer is used to extract long-term dependent features from the input invoice data through a long short-term memory network (capturing trend evolution across time, especially suitable for predicting trends across months and quarters, such as predicting fluctuations in invoice volume before and after summer and winter vacations).

[0184] Y_lstm=LSTM(units=64~128)(Y_cnn)

[0185] LSTM effectively handles the following characteristics: strong sequence correlation (such as invoice volume related to three consecutive months); strong periodicity (such as price increases every September during the school season); and node-dependent nonlinearity (such as policy adjustments or medical insurance reimbursement windows).

[0186] The feature enhancement layer is used to enhance and fuse temporal local pattern features and long-term dependency features through a Kolmogorov network (knowledge enhancement network) to obtain enhanced features;

[0187] The core idea of ​​Kolmogorov networks is to learn nonlinear mappings over time using explicit piecewise functions (such as ReLU combinations); and to guide the model to be more interpretable and generalizable in scenarios where "structural priors" such as units, items, and seasons exist.

[0188] Kolmogorov networks are particularly well-suited for tasks such as: structural shifts in school non-tax revenue, such as "peak season surges" and "zero returns during summer and winter breaks"; modeling heterogeneous behaviors of "specific units + specific invoices" (e.g., significant differences between universities and hospitals); and improving the model's ability to fit boundary and peak nodes. In this invention, Kolmogorov networks are specifically used to: strengthen the nonlinear relationship between time, units, and items; make more accurate fits to "breakpoints" and "structural changes" in long-term trends; and reduce the smoothing trend error of LSTM in long-term predictions.

[0189] Y_kan = KAN_Layer()(Y_lstm)

[0190] The output layer is used to output the bill sequence prediction result with a preset prediction time based on the enhanced features; the bill sequence prediction result carries the total amount of future invoices, the number of future bills, the amount of future projects, and a high-risk volatility score;

[0191] The output layer supports a multi-output, multi-task modeling structure. Each prediction head can share underlying features or model independently. In specific implementation, the following prediction objectives can be set as needed:

[0192] Output name type illustrate Total amount of future invoices Regression value Forecast the amount of bills for the next 7 / 30 / 365 days Future bill quantity Regression value Total number of bills predicted by the unit Future project amount Regression value The unit predicts the total amount of a certain project. High-risk volatility score Fraction Is there a risk of abnormal surge / decline?

[0193] The formula for the loss function is:

[0194] L total =α*L1+β*L2+γ*L3;

[0195] Among them, L total The loss function represents the loss value; L1, L2, and L3 represent the loss on invoice amount, loss on number of invoices, and loss on average amount, respectively, and all use the mean squared error function; α, β, and γ represent weighting coefficients, and α + β + γ = 1. In practice, L1 regularization restrictions can be added to enhance interpretability.

[0196] The ticket sequence prediction model training module is specifically used for:

[0197] The dataset is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The ticket sequence prediction model is trained using the training set until the loss value of the loss function is less than a preset loss threshold.

[0198] The mean squared error, mean absolute error, and prediction accuracy are calculated using the validation set to validate the trained ticket sequence prediction model. If the validation fails, the training set is expanded and training continues; if the validation succeeds, then:

[0199] The confidence level is calculated using the test set to test the validated invoice sequence prediction model. If the test fails, the training set is expanded and training continues. If the test passes, training ends and the validated invoice sequence prediction model is deployed using containerization technology.

[0200] The ticket sequence prediction module is specifically used for:

[0201] The system acquires real-time electronic tickets, preprocesses them using a streaming engine, and then inputs them into a deployed ticket sequence prediction model. The ticket sequence prediction model performs inference using hardware acceleration technology and outputs real-time ticket sequence prediction results for ticket sequence prediction.

[0202] Experimental verification:

[0203] Using non-tax revenue electronic invoice data (including fields such as invoice amount, number of invoices, and project amount) from a university over the past three years, this invention is compared with traditional ARIMA models, single LSTM models, and single CNN models. Mean squared error (MSE), mean absolute error (MAE), and prediction accuracy are used as evaluation metrics.

[0204] Experimental results:

[0205] Prediction accuracy: When predicting the total invoice amount for the next 7 days, 30 days, and 1 year, the MSE of this invention is 0.02, 0.05, and 0.10, respectively, which is significantly lower than that of the traditional ARIMA model (MSE of 0.08, 0.15, and 0.30), the single LSTM model (MSE of 0.04, 0.09, and 0.20), and the single CNN model (MSE of 0.06, 0.12, and 0.25, respectively).

[0206] Perception of periodic changes: During the school season and winter and summer vacations, the prediction accuracy of this invention is 90% and 85% respectively, while the prediction accuracy of traditional models is only 70% and 65%.

[0207] Multi-objective prediction capability: When predicting the number of bills, average project amount, and proportion of each type of bill, the MAE of this invention is 0.03, 0.02, and 0.01, respectively, which is significantly better than other models.

[0208] In summary, the advantages of this invention are:

[0209] 1. By acquiring a large number of historical electronic tickets, data cleaning, structured aggregation, and feature group extraction are performed on each historical electronic ticket sequentially. Based on the extracted feature groups, each historical electronic ticket is labeled, and sample pairs are constructed by segmenting each historical electronic ticket using a sliding window. A dataset is then built based on these sample pairs. Next, a ticket sequence prediction model is constructed based on an input layer, a local feature extraction layer, a global feature extraction layer, a feature enhancement layer, and an output layer. A loss function is set for the ticket sequence prediction model, and it is trained using the dataset and the loss function. The trained ticket sequence prediction model is then deployed, and finally, ticket sequence prediction is performed using the deployed model. In other words, the time series data and periodic features of electronic tickets are explicitly extracted through the input layer. Dense vectors, combined with a one-dimensional convolutional neural network (CNN) in the local feature extraction layer to capture the temporal local pattern features (local details and sudden fluctuation patterns) of bill data, and a long short-term memory network (LSTM) in the global feature extraction layer to model long-term dependent features, and then a Kolmogorov network (KAN) in the feature enhancement layer to perform deep nonlinear fusion and enhancement of the above multi-source heterogeneous features, thereby synergistically solving the core problems of traditional methods that are difficult to characterize complex periodic patterns, have a lag in response to sudden events, and have insufficient nonlinear expression capabilities. Finally, it achieves accurate and stable prediction of multi-dimensional indicators such as bill amount (future total bill amount, future project amount), bill quantity (future bill quantity), and high-risk fluctuation score, thereby greatly improving the accuracy and robustness of bill sequence prediction.

[0210] 2. Improved prediction accuracy and stability: Compared with traditional single time series models (such as ARIMA, pure LSTM, etc.), by integrating CNN to extract local temporal pattern features, LSTM to model long-term dependent features, and KAN to construct nonlinear knowledge interpolation (feature enhancement), the prediction error in high-fluctuation periods (such as the start of the school season and fiscal settlement period) is effectively reduced, and the overall prediction accuracy is significantly improved.

[0211] 3. Enhance the ability to perceive cyclical changes in non-tax business: By introducing cyclical auxiliary features such as "month", "week", and "whether it is the start of school season / holiday", it can adaptively capture annual business patterns; for example, universities concentrate on issuing invoices in September each year, and medical institutions experience peak periods before and after holidays, thereby better monitoring and forecasting seasonal financial data.

[0212] 4. Enable parallel output of multi-objective business forecasts: It can simultaneously output multiple key indicators (invoice amount, number of invoices, proportion of invoice types, average project amount, etc.), supporting multi-angle analysis and early warning, effectively supporting budget preparation, revenue forecasting, performance evaluation and other work scenarios of universities, hospitals and other units.

[0213] 5. Possesses cross-unit and cross-regional migration and generalization capabilities: Utilizing the model's modeling capabilities in the "Invoicing Unit" and "Regional Division" fields, it can achieve generalized predictions for different units (such as multiple universities or multiple hospitals) and different regions (provinces, cities, districts), possessing good promotion and deployment value.

[0214] 6. Supports rolling updates and dynamic predictions to adapt to actual business needs: The model adopts a sliding window data generation method and a rolling prediction mechanism, which allows the model to continuously receive new data for iterative training and updates. It supports the combination of daily predictions and medium- and long-term budget forecasts, improving the practicality and robustness of the model's online services.

[0215] 7. Enhance the automation level of intelligent management of non-tax invoices: By constructing an invoice sequence prediction model, the dynamic trend of non-tax revenue expenditure can be grasped in advance, realizing "pre-emptive monitoring" and "risk prediction" of invoice issuance behavior, and improving the information and intelligent management capabilities of electronic invoice business.

[0216] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting ticket sequences based on a composite neural network, characterized in that: Includes the following steps: Step S1: Obtain a large number of historical electronic tickets, perform data cleaning, structured aggregation and feature group extraction on each of the historical electronic tickets in sequence, label each of the historical electronic tickets based on the extracted feature groups, and construct sample pairs by segmenting each of the historical electronic tickets through a sliding window, and construct a dataset based on each of the sample pairs. Step S2: Construct a ticket sequence prediction model based on the input layer, local feature extraction layer, global feature extraction layer, feature enhancement layer and output layer, and set the loss function of the ticket sequence prediction model; Step S3: Train the ticket sequence prediction model using the dataset and loss function, and deploy the trained ticket sequence prediction model. Step S4: Predict the bill sequence using the deployed bill sequence prediction model.

2. The method for predicting ticket sequences based on a composite neural network as described in claim 1, characterized in that: Step S1 specifically involves: A large number of historical electronic invoices are obtained from a preset data source using an ETL tool. The historical electronic invoices carry at least the invoice date, invoice amount, project name, project code, invoice unit, invoice type, region, project amount, and project quantity. The historical electronic invoices shall be cleaned by at least the following methods: standardizing the invoice date format, handling missing values, handling outliers, handling duplicate invoices, handling empty invoices, and handling abnormal invoices. Using the combination of the invoicing unit and the type of invoice as the granularity, the historical electronic invoices after data cleaning are aggregated at preset time intervals to construct a time-series sample table; the time interval is daily, weekly, or monthly; feature groups including at least periodic features, categorical variable features, and statistical features are extracted from the time-series sample table; After feature normalization and encoding of the extracted feature group, each historical electronic ticket is labeled based on the feature group, and each historical electronic ticket is segmented by a sliding window based on preset sliding parameters to construct sample pairs including input data and prediction data. A dataset is constructed based on each sample pair; the sliding parameters include the length of the input data, the length of the prediction data, and the sliding step size.

3. The method for predicting ticket sequences based on a composite neural network as described in claim 1, characterized in that: In step S2, the input layer is constructed based on a time series input module, a feature input module, an embedded vector input module, and a concatenation module. The time series input module is used to extract time series data from the input electronic invoices; the feature input module is used to extract periodic features from the input electronic invoices; the embedding vector input module is used to extract dense vectors of categorical variables from the input electronic invoices; the categorical variables include project name, invoice type, region, and issuing unit; the concatenation module is used to concatenate the time series data, periodic features, and dense vectors into input invoice data; The local feature extraction layer is used to extract temporal local pattern features from the input ticket data through a one-dimensional convolutional neural network; The global feature extraction layer is used to extract long-term dependency features from the input ticket data through a long short-term memory network. The feature enhancement layer is used to enhance and fuse temporal local pattern features and long-term dependency features through a Kolmogorov network to obtain enhanced features. The output layer is used to output the prediction result of the bill sequence with a preset prediction duration based on the enhanced features; The bill sequence prediction results carry the total amount of future bills issued, the number of future bills, the amount of future projects, and a high-risk volatility score; The formula for the loss function is: L total =α*L1+β*L2+γ*L3; Among them, L total L1, L2, and L3 represent the loss value of the loss function; L1, L2, and L3 represent the loss of invoice amount, loss of invoice number, and loss of average amount, respectively, and all use the mean square error function; α, β, and γ all represent weighting coefficients, and α+β+γ=1.

4. The method for predicting ticket sequences based on a composite neural network as described in claim 1, characterized in that: Step S3 specifically involves: The dataset is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The ticket sequence prediction model is trained using the training set until the loss value of the loss function is less than a preset loss threshold. The mean squared error, mean absolute error, and prediction accuracy are calculated using the validation set to validate the trained ticket sequence prediction model. If the validation fails, the training set is expanded and training continues; if the validation succeeds, then: The confidence level is calculated using the test set to test the validated invoice sequence prediction model. If the test fails, the training set is expanded and training continues. If the test passes, training ends and the validated invoice sequence prediction model is deployed using containerization technology.

5. The method for predicting ticket sequences based on a composite neural network as described in claim 1, characterized in that: Step S4 specifically involves: The system acquires real-time electronic tickets, preprocesses them using a streaming engine, and then inputs them into a deployed ticket sequence prediction model. The ticket sequence prediction model performs inference using hardware acceleration technology and outputs real-time ticket sequence prediction results for ticket sequence prediction.

6. A ticket sequence prediction system based on a composite neural network, characterized in that: Includes the following modules: The dataset construction module is used to acquire a large number of historical electronic tickets, perform data cleaning, structured aggregation and feature group extraction on each of the historical electronic tickets in sequence, label each of the historical electronic tickets based on the extracted feature groups, and construct sample pairs by segmenting each of the historical electronic tickets through a sliding window, and construct a dataset based on each of the sample pairs. The bill sequence prediction model construction module is used to construct a bill sequence prediction model based on an input layer, a local feature extraction layer, a global feature extraction layer, a feature enhancement layer, and an output layer, and to set the loss function of the bill sequence prediction model. The bill sequence prediction model training module is used to train the bill sequence prediction model using the dataset and loss function, and to deploy the trained bill sequence prediction model. The bill sequence prediction module is used to predict bill sequences using the deployed bill sequence prediction model.

7. The bill sequence prediction system based on a composite neural network as described in claim 6, characterized in that: The dataset construction module is specifically used for: A large number of historical electronic invoices are obtained from a preset data source using an ETL tool. The historical electronic invoices carry at least the invoice date, invoice amount, project name, project code, invoice unit, invoice type, region, project amount, and project quantity. The historical electronic invoices shall be cleaned by at least the following methods: standardizing the invoice date format, handling missing values, handling outliers, handling duplicate invoices, handling empty invoices, and handling abnormal invoices. Using the combination of the invoicing unit and the type of invoice as the granularity, the historical electronic invoices after data cleaning are aggregated at preset time intervals to construct a time-series sample table; the time interval is daily, weekly, or monthly; feature groups including at least periodic features, categorical variable features, and statistical features are extracted from the time-series sample table; After feature normalization and encoding of the extracted feature group, each historical electronic ticket is labeled based on the feature group, and each historical electronic ticket is segmented by a sliding window based on preset sliding parameters to construct sample pairs including input data and prediction data. A dataset is constructed based on each sample pair; the sliding parameters include the length of the input data, the length of the prediction data, and the sliding step size.

8. The bill sequence prediction system based on a composite neural network as described in claim 6, characterized in that: In the bill sequence prediction model construction module, the input layer is constructed based on a time series input module, a feature input module, an embedding vector input module, and a splicing module. The time series input module is used to extract time series data from the input electronic invoices; the feature input module is used to extract periodic features from the input electronic invoices; the embedding vector input module is used to extract dense vectors of categorical variables from the input electronic invoices; the categorical variables include project name, invoice type, region, and issuing unit; the concatenation module is used to concatenate the time series data, periodic features, and dense vectors into input invoice data; The local feature extraction layer is used to extract temporal local pattern features from the input ticket data through a one-dimensional convolutional neural network; The global feature extraction layer is used to extract long-term dependency features from the input ticket data through a long short-term memory network. The feature enhancement layer is used to enhance and fuse temporal local pattern features and long-term dependency features through a Kolmogorov network to obtain enhanced features. The output layer is used to output the prediction result of the bill sequence with a preset prediction duration based on the enhanced features; The bill sequence prediction results carry the total amount of future bills issued, the number of future bills, the amount of future projects, and a high-risk volatility score; The formula for the loss function is: L total =α*L1+β*L2+γ*L3; Among them, L total L1, L2, and L3 represent the loss value of the loss function; L1, L2, and L3 represent the loss of invoice amount, loss of invoice number, and loss of average amount, respectively, and all use the mean square error function; α, β, and γ all represent weighting coefficients, and α+β+γ=1.

9. The bill sequence prediction system based on a composite neural network as described in claim 6, characterized in that: The ticket sequence prediction model training module is specifically used for: The dataset is divided into a training set, a validation set, and a test set in an 8:1:1 ratio. The ticket sequence prediction model is trained using the training set until the loss value of the loss function is less than a preset loss threshold. The mean squared error, mean absolute error, and prediction accuracy are calculated using the validation set to validate the trained ticket sequence prediction model. If the validation fails, the training set is expanded and training continues; if the validation succeeds, then: The confidence level is calculated using the test set to test the validated invoice sequence prediction model. If the test fails, the training set is expanded and training continues. If the test passes, training ends and the validated invoice sequence prediction model is deployed using containerization technology.

10. The bill sequence prediction system based on a composite neural network as described in claim 6, characterized in that: The ticket sequence prediction module is specifically used for: The system acquires real-time electronic tickets, preprocesses them using a streaming engine, and then inputs them into a deployed ticket sequence prediction model. The ticket sequence prediction model performs inference using hardware acceleration technology and outputs real-time ticket sequence prediction results for ticket sequence prediction.

Images