Enterprise fund income and expenditure prediction method based on AI large model and multi-modal interaction

Abstract

The application discloses an enterprise fund income and expenditure prediction method based on an AI large model and multi-modal interaction, and particularly relates to the field of enterprise asset prediction, and is used for solving the problem that the existing fund prediction method relies on a single data source and lacks causal explainability. The method collects text, image and time sequence multi-modal data related to enterprise fund income and expenditure, constructs a hierarchical causal diagram containing an upper layer and a lower layer, aligns features in a unified semantic space by using a multi-modal Transformer to generate modal feature vectors, dynamically adjusts the weight of a modal node in the hierarchical causal diagram based on reinforcement learning, extracts short-term and long-term dependencies by combining a time sequence causal convolution and a memory network, generates a time sequence enhanced multi-modal representation, and infers the fund income and expenditure change trajectory under different external event scenarios by using counterfactual reasoning and a Bayesian network model, so that long-period and scenario-based prediction of the enterprise fund operation state is realized, and the prediction accuracy and explainability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of enterprise asset prediction, more specifically, the present application relates to an enterprise fund income and expenditure prediction method based on AI large model and multi-modal interaction. BACKGROUND

[0002] In the current enterprise fund income and expenditure prediction scenario, the operating environment faced by enterprises presents highly dynamic and multi-factor coupling characteristics. The inflow and outflow of funds no longer depend only on internal operating data, but are also affected by multi-dimensional factors such as policy adjustment, industry fluctuation, supply chain resilience and market psychological expectation. Traditional financial prediction methods mainly rely on single-dimensional historical transaction data or static regression models, which are difficult to capture the implicit correlation between cross-time and cross-modal information. For example, when external factors such as macro-policy changes, regional market demand fluctuations or key infrastructure renovations occur, the cash flow structure, working capital allocation and debt level of the enterprise often show a nonlinear response. However, traditional linear models or rule-based statistical methods cannot timely perceive this potential causal driving relationship, resulting in lagging or even inaccurate prediction results.

[0003] In addition, existing multi-modal prediction models still remain at the level of modal fusion and fail to build a causal structure to explain the logical relationship between external events and enterprise fund dynamics. They also fail to establish a counterfactual reasoning mechanism based on time dependence to analyze long-term trends under different scenarios.

[0004] Therefore, under complex economic environment, there is an urgent need for an enterprise fund income and expenditure prediction method that can integrate the capabilities of AI large models, combine multi-modal feature alignment and hierarchical causal reasoning mechanisms to achieve long-period, scenario-based and dynamic prediction of enterprise fund operation status. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide an enterprise fund income and expenditure prediction method based on AI large model and multi-modal interaction to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] The enterprise fund income and expenditure prediction method based on AI large model and multi-modal interaction comprises the following steps:

[0008] S1, collecting multi-modal data related to enterprise fund income and expenditure, the multi-modal data including text data, image data and time series data;

[0009] S2, constructing a hierarchical causal graph based on the multi-modal data, the hierarchical causal graph including an upper causal graph and a lower causal graph;

[0010] S3, inputting the multi-modal data into a multi-modal Transformer architecture, performing cross-modal feature alignment in a unified semantic space, and outputting corresponding modal feature vectors;

[0011] S4, dynamically adjusting global weights of each modality in the hierarchical causal graph based on a change pattern of the multi-modal feature vectors;

[0012] S5, processing the multi-modal feature vectors using a time series causal convolution and memory network to capture long-period time series dependencies, and outputting time series enhanced multi-modal representations;

[0013] S6, performing counterfactual reasoning based on the hierarchical causal graph with adjusted weights and the time series enhanced multi-modal representations to generate long-period prediction data of enterprise fund collection and disbursement.

[0014] In a preferred embodiment, in S1, multi-modal data related to enterprise fund collection and disbursement is collected, and the multi-modal data includes text data, image data, and time series data, specifically including:

[0015] Data extraction is performed from a structured database of enterprise information to extract text data containing policy and regulation texts and industry analysis reports;

[0016] Image data is obtained from a geographic information platform, which is image records reflecting resource distribution status and infrastructure layout changes;

[0017] Time series data is collected from financial transaction records and market data sources, including fund flow rate indicators and market fluctuation sequences, and the time series data is sampled through a sliding window to generate continuous time series segments;

[0018] Data cleaning and data format standardization processing is performed on the extracted text data, image data, and time series data, and different sources of data are converted into a unified time series format, and pre-processed multi-modal data is output.

[0019] In a preferred embodiment, in S2, a hierarchical causal graph is constructed based on the multi-modal data, and the hierarchical causal graph includes an upper causal graph and a lower causal graph, specifically including:

[0020] The pre-processed multi-modal data is analyzed to identify causal relationships between enterprise fund collection and direct factors, and a hierarchical causal graph is generated based on conditional independence testing and causal edge direction inference, and the hierarchical causal graph structure is divided into upper and lower layers;

[0021] The hierarchical causal graph uses an online learning algorithm to incrementally update the graph structure, and when new multi-modal data is input, the weights and directions of the causal edges are recalculated.

[0022] In a preferred embodiment, the hierarchical causal graph structure is divided into upper and lower layers, specifically:

[0023] The upper layer of the causal graph represents the causal dependency relationship between the enterprise fund flow nodes and the direct factor nodes, including policy adjustment factors and industry demand change factors, through directed edges.

[0024] The lower layer of the causal graph represents the causal influence of the multi-modal driving factor nodes on the direct factor nodes through directed edges, where the multi-modal driving factor nodes are the modal nodes corresponding to the multi-modal data.

[0025] The weights of the directed edges in the upper and lower layers of the causal graph are calculated by maximum likelihood estimation.

[0026] In a preferred embodiment, in S3, the multi-modal data is input into a multi-modal Transformer architecture to perform cross-modal feature alignment in a unified semantic space and output corresponding modal feature vectors, specifically including:

[0027] The preprocessed text data, image data, and time series data are input into the input channels of the multi-modal Transformer model, and the model sets independent modal encoding units and shared multi-head attention layers.

[0028] The modal encoding units perform feature mapping on different modal data and output corresponding high-dimensional feature vector sequences.

[0029] The multi-head attention layer calculates the semantic correlation between each modal feature vector sequence in the unified semantic space as the correlation strength, and generates a cross-modal weight distribution based on the correlation strength.

[0030] The representation of each modal feature vector is adjusted through the cross-modal weight distribution, and the text feature vector, image feature vector, and time series feature vector are fused using a weighted sum operation to align different modal data semantically.

[0031] In a preferred embodiment, in S4, the global weights of each modal in the hierarchical causal graph are dynamically adjusted based on the change pattern of the multi-modal feature vector, specifically including:

[0032] The current environmental state is identified based on the change pattern of the multi-modal feature vector, including the policy release period, industry transformation period, and market fluctuation period, and the change pattern is identified through the distribution offset of the modal feature vector and the time series fluctuation.

[0033] The importance score of each modal in the hierarchical causal graph is calculated using a reinforcement learning algorithm, and the importance score calculates the contribution of the modal to the causal relationship through a reward function, which is set based on the principle of minimizing prediction error.

[0034] The global weight distribution of the modal nodes in the upper and lower causal graphs is adjusted in real time according to the importance score.

[0035] In a preferred embodiment, the S5, the time series causal convolution and memory network are applied to process the multi-modal feature vector, capture long-period time series dependence, and output a time series enhanced multi-modal representation, specifically comprising:

[0036] The time series causal convolution is applied to process the multi-modal feature vector, the convolution kernel size is adaptively adjusted according to the time series period, and the short-term local mode of the local change in the continuous time window is outputted;

[0037] The memory network is used to store the historical multi-modal feature vector sequence, and the long-period dependence is extracted by retrieving the historical feature vector similar to the current multi-modal feature vector;

[0038] The output of the causal convolution and the output of the memory network are integrated, the short-term local mode and the global dependence information are fused using a fully connected layer, and the time series enhanced multi-modal representation is generated.

[0039] In a preferred embodiment, the S6, the counterfactual reasoning is performed based on the weight-adjusted hierarchical causal graph and the time series enhanced multi-modal representation, and the long-period prediction data of the enterprise fund flow is generated, specifically comprising:

[0040] The hierarchical causal graph structure and the global weight of the modal nodes are obtained, the direct factor nodes affected by external events or the multi-modal driving factor nodes are identified, and the corresponding counterfactual scenario is constructed;

[0041] The intervention operation is applied to the identified nodes, the node value is set to the counterfactual condition, the intervention influence is propagated to the enterprise fund flow nodes layer by layer on the directed edge structure of the causal graph according to the historical dependence information provided by the time series enhanced multi-modal representation, and the change amount probability distribution of the fund flow nodes is inferred using the Bayesian network model.

[0042] The change amount probability distribution of the fund flow nodes is inferred using the Bayesian network model, and the fund flow prediction trajectory under multiple external event scenarios is outputted.

[0043] The technical effects and advantages of the enterprise fund flow prediction method based on the AI large model and multi-modal interaction of the present application are as follows:

[0044] The feature alignment of text, image and time series data by multi-modal Transformer can map the policy, industry, market and transaction signals originally scattered in different systems to a unified semantic space, significantly improving the integrity and semantic consistency of information fusion. The structured modeling based on hierarchical causal graph can explicitly express the causal dependence between enterprise fund collection and payment and various external direct factors and multi-modal driving factors, thereby realizing the explainable prediction logic. The reinforcement learning driven modal weight dynamic adjustment mechanism enables the model to automatically focus on key information sources during policy fluctuations or market transformation, improving the adaptability of the model to environmental changes. By combining time series causal convolution and memory network, the model can capture both short-term trading fluctuations and long-term periodic dependencies, effectively solving the problem of long-term dependency decay in traditional methods.

[0045] The model can simulate the potential impact of policy, market or industry events on enterprise fund collection and payment under different external situations, output long-term prediction trajectories with uncertainty measurement, and realize the logical closed loop from prediction results to prediction reasons, providing enterprises with more explainable and decision-making value intelligent financial analysis capabilities. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 The figure is a schematic diagram of the enterprise fund collection and payment prediction method based on AI large model and multi-modal interaction of the present application. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0048] Embodiment 1, Figure 1 The enterprise fund collection and payment prediction method based on AI large model and multi-modal interaction of the present application is given, which includes the following steps:

[0049] S1, collecting multi-modal data related to enterprise fund collection and payment, the multi-modal data including text data, image data and time series data;

[0050] S2, constructing a hierarchical causal graph based on the multi-modal data, the hierarchical causal graph including an upper causal graph and a lower causal graph;

[0051] S3, inputting the multi-modal data into a multi-modal Transformer architecture, performing cross-modal feature alignment in a unified semantic space and outputting corresponding modal feature vectors;

[0052] S4, dynamically adjusting the global weights of each modality in the hierarchical causal graph based on the change pattern of the multi-modal feature vector;

[0053] S5, processing the multi-modal feature vector by applying a time series causal convolution and memory network to capture long-period time series dependencies, and outputting a time series enhanced multi-modal representation;

[0054] S6, performing counterfactual reasoning based on the weight-adjusted hierarchical causal graph and the time series enhanced multi-modal representation to generate long-period prediction data of enterprise fund collection and payment.

[0055] In S1, multi-modal data related to enterprise fund collection and payment is collected, including text data, image data and time series data, specifically including:

[0056] When performing data extraction from the structured database of enterprise information, first determine the text types related to fund collection and payment, including policies and regulations, industry analysis reports, financial announcements and enterprise internal operation documents. The record fields in the database include file title, publication time, source agency and content. In the extraction stage, a text segmentation parsing process is used to structure the four types of regions of each document, including title, text paragraph, table description and appendix. After parsing, a set of text segments with hierarchical indexes is generated, each segment is attached with a timestamp and a data source label. Then the extracted content is semantically regularized to remove invalid information such as format symbols, headers and footers, and to retain key phrases and sentences. For example, for policy texts, retain core expression fields such as "fund supervision", "interest rate adjustment" and "tax incentives"; for industry reports, retain key phrases such as "revenue changes", "capacity utilization" and "market forecasts". After structured parsing, the texts arranged in chronological order form text feature indicators, each text feature indicator corresponds to a complete semantic unit.

[0057] When obtaining image data from the geographic information platform, take the areas related to enterprise production and resource distribution as the spatial range, and select geographic image data covering key areas such as factory sites, warehouses and transportation nodes. Image data includes satellite remote sensing images and ground monitoring images, the former is used to reflect infrastructure construction and land use changes, and the latter is used to reflect the production facility operation status and material storage density. All images are updated monthly or quarterly. After obtaining, the images are uniformly converted into a standard resolution two-dimensional matrix form, recording the brightness, color and texture feature values of each pixel. Then divide the space into regions, mark the factory boundaries, road directions and warehouse areas as fixed region masks, and generate average pixel feature vectors for each mask region. These feature vectors are combined into image time series in chronological order to reflect resource distribution status, infrastructure expansion or contraction trends, and output as time-indexed image feature sequence indicators.

[0058] When collecting time series data from financial transaction records and market data sources, the enterprise fund account flow, procurement payment records, sales collection records, and public market indicators are selected as data sources. Each record contains transaction time, amount, category, and transaction object information. The original data is sorted in chronological order, and the fund inflow rate and outflow rate are extracted according to the business nature. The time interval is set to one day. When calculating the fund flow rate, the difference between the transaction amounts of adjacent time periods is divided by the time interval to form a continuous change curve; when calculating the market fluctuation sequence, the price change rate of the closing price sequence of the main market index is calculated. To prevent extreme values, a three-day moving average is used for smoothing. Then, a fixed window length (e.g., 30 days) is used for sliding window sampling, each window slides forward one time step, generating a continuous time series segment, each segment retains the transaction rate, market fluctuation, and related macro indicators as time series feature indicators.

[0059] Text, image, and time series feature indicators are labeled as text data, image data, and time series data, respectively, and data cleaning and standardization processing is performed. Duplicate and abnormal values are checked for each type of data. Text data removes duplicate documents by hashing comparison and corrects character replacement for garbled paragraphs; image data is removed by time series consistency detection to eliminate blurred, missing frames, or offset images; time series data performs linear interpolation to complete missing time points and uses local median smoothing for abnormal jump points, and integrates into preprocessed multi-modal data.

[0060] In S2, a hierarchical causal graph is constructed based on the multi-modal data, which includes an upper layer causal graph and a lower layer causal graph.

[0061] Text, image, and time series data are uniformly indexed and managed to ensure comparability of features of different types under the same time reference. For each time period, a set of multi-modal data is established, including fund collection and payment nodes and their candidate direct impact factor set. Direct factors mainly include policy adjustments, changes in industry demand, market price fluctuations, supply cycle changes, and other representative index types in economic activities. To identify the causal relationship between enterprise fund collection and payment and direct factors, a conditional independence-based inference process is used. Specifically, for each candidate factor and fund collection and payment node, detect whether the joint distribution changes with the introduction of a third-party factor. If it still shows stable statistical dependence after controlling other variables, it is determined that there is a direct causal relationship between the factor and fund collection and payment. This process is performed one by one in the multi-modal feature space, and finally an initial causal relationship set between the fund collection and payment node and multiple direct factor nodes is formed.

[0062] After generating the initial causal set, the directionality of the directed edges is determined by a causal edge direction inference method. Specifically, according to the time sequence relationship and conditional independence results, if the characteristic change of a certain direct factor always precedes the significant fluctuation of the fund flow index, it is determined that the factor points to the fund flow node to form a cause-to-effect direction; if the fund flow change leads and the direct factor change has a responsive characteristic, it is confirmed as an effect-to-cause direction. On this basis, the upper structure of the hierarchical causal graph is constructed. The node set of the upper causal graph includes the fund flow node and various direct factor nodes, and the nodes are connected by directed edges, and the dependence direction and strength parameters are recorded. The establishment of the lower layer of the causal graph is based on the feature association analysis between multi-modal driving factors and direct factors. According to the type of multi-modal data, the node set is divided into text modal nodes, image modal nodes and time series modal nodes. For each direct factor node, the correlation distribution between it and the multi-modal nodes is calculated, and when the characteristic change of a certain modal node and the direct factor remain significantly associated in multiple time windows, a directed edge from the modal node to the direct factor node is established. The structure of the lower layer reflects the driving effect of multi-modal features on economic direct factors, and provides source support for the upper causal path.

[0063] To maintain the timeliness of the causal graph during data growth, an online learning algorithm is used to incrementally update the causal edges. The specific process is as follows: when a new batch of multi-modal input data is received, the edge set of the original causal graph is scanned. For newly appearing node pairs, calculate their conditional dependence relationship and judge whether a new edge should be added; for the original edge, recalculate the dependence strength and direction. When the amount of new data is small, only the significantly changed node pairs are updated locally to reduce computational redundancy. In the calculation of the weight of the directed edge, the maximum likelihood estimation method is used, based on the observed joint probability of the node state in the sample data, to estimate the weight value of each edge. The weight value is set to be between 0 and 1, and the higher the value, the stronger the causal dependence between nodes. To avoid weight bias caused by sample imbalance, a sliding window method is used to update the sample time period smoothly, that is, after receiving a batch of new data, the weighted average of the last three data samples is taken as the new likelihood estimation input, so that the weight value can dynamically reflect the latest fund flow environment. The updated upper and lower causal graphs together form a complete hierarchical causal structure, and form a traceable causal path system between enterprise fund flow and multi-modal driving factors.

[0064] In S3, the multi-modal data is input into a multi-modal Transformer architecture, and cross-modal feature alignment is performed in a unified semantic space to output corresponding modal feature vectors.

[0065] The pre-processed text data, image data, and time series data are respectively input into the corresponding input channels of the multi-modal Transformer model. The model sets up independent modal encoding units for different modalities. The role of each modal encoding unit is to map the original data into a high-dimensional vector sequence of uniform dimensions, which is used for semantic alignment in subsequent layers. The input of the text modality is the structured parsed policy text, industry analysis report, and market comment content. After input, each word or phrase is mapped to a fixed-length vector representation through an embedding layer, such as a 512-dimensional feature vector to express the semantic features of the text. The input of the image modality is a sequence of image blocks after normalization and region division. Each image block is processed by convolution and linear projection to form a set of visual feature vectors, each corresponding to the spatial semantic features of a local region. The input of the time series modality is a time slice generated by continuous sampling. Each time slice is mapped to a time feature vector by an embedding layer and position encoding to represent the time-varying features such as fund flow and market fluctuation. The high-dimensional feature sequences output by the above three types of modal encoding all have time indices or semantic position identifiers and are input into the unified feature space of the Transformer in batches to form the modal feature vector set to be aligned.

[0066] In the unified semantic space, the multi-head attention layer of the model is responsible for calculating the semantic correlation between different modal feature vector sequences to depict the potential association strength between modalities. This process is achieved by calculating the similarity index between feature vectors in each attention head. The vector similarity measure, such as cosine similarity, is used to measure the matching degree of different modal features. During the calculation process, the feature vectors of each modality are projected into a shared attention subspace, and the semantic interaction between different modal features is automatically learned by the attention mechanism. Taking the text modality and the image modality as an example, if the semantic phrase "infrastructure investment growth" in the policy text has a high similarity with "factory expansion image" in the image features, the attention layer will give this pair of features a higher association strength value. After the correlation calculation between all modal features is completed, the model generates a set of cross-modal weight distributions in each attention head, which reflects the contribution proportion of different modal features to the current fusion task. The numerical sum of the cross-modal weight distribution is normalized to 1 to ensure that the influence of different modal features on the fusion result is interpretable and additive.

[0067] After obtaining the cross-modal weight distribution, the model realizes feature fusion and semantic alignment through weight adjustment and weighted summation operation. Specifically, the feature vector of each modality will recalibrate its representation strength in the unified semantic space according to the cross-modal weight. The weight enhancement of the text feature vector represents that the policy text content has a greater impact in the current semantic context, and the weight enhancement of the image feature vector represents that the visual change is more indicative in the current prediction task. Then, the weighted summation operation is performed on the weighted multi-modal feature vector to generate the fused and aligned feature representation. The weighted summation is performed independently at each time step or semantic position to ensure the semantic consistency between local features. After fusion processing, the text, image and time series feature vectors have equivalent expression forms in the unified semantic space. The model outputs the aligned multi-modal feature sequence, and each vector element has comprehensively reflected the joint information of semantic content, spatial state and time change, providing a standardized input basis for subsequent causal relationship calculation and dynamic weight adjustment of the hierarchical causal graph.

[0068] In S4, the global weights of each modality in the hierarchical causal graph are dynamically adjusted based on the change pattern of the multi-modal feature vector.

[0069] The change pattern of each modality feature over time is analyzed to identify the macro-environmental state of the current enterprise. The multi-modal feature vector comes from the text, image and time series modalities, among which the text modality feature reflects the change trend of policy and industry semantics, the image modality feature reflects the evolution of physical or spatial state, and the time series modality feature reflects the dynamic fluctuation of fund flow and market indicators. To identify the environmental state, the distribution offset of each modality feature vector is calculated within a sliding time window, and the fluctuation amplitude in the time series is calculated. When the text modality feature appears concentrated offset in a short time, and the semantic clustering center of the policy keyword is similar, it is determined as the policy release period; when the spatial features related to resource allocation or capacity expansion in the image modality feature show continuous change, it is determined as the industry transformation period; when the fluctuation frequency and amplitude of the time series modality feature significantly increase, represented by the significant deviation of fund flow rate and market fluctuation indicators from the historical mean, it is identified as the market fluctuation period. The identification result of the environmental state is used as the state input of the reinforcement learning algorithm to guide the subsequent adaptive adjustment of the modality weight.

[0070] In the reinforcement learning phase, the importance score of each modality in the hierarchical causal graph is calculated. The importance score is used to measure the degree of contribution of the modality to the accuracy of causal reasoning in the current environmental state. The reinforcement learning process defines a reward function based on prediction error minimization, which is used to evaluate the improvement of different combinations of modality features on the prediction results of fund flow. For example, when the text modality feature is input during the policy release period, the prediction error is reduced by more than 10% compared to the case without this modality, and the text modality is assigned a higher reward value; if the change of time series features during market fluctuations is more significant in improving prediction accuracy, the time series modality will obtain a higher reward score. Through multiple iterations, the algorithm determines the importance score of each modality based on the cumulative reward value. To ensure the stability of the calculation, the maximum value of the reward function is set to 1 and the minimum value is set to 0, and the score is updated through a smoothing average mechanism, so that the weight adjustment remains stable in continuous time steps.

[0071] According to the modality importance score output by reinforcement learning, the global weight distribution of each modality node in the hierarchical causal graph is adjusted in real time. The weight adjustment of the upper layer causal graph mainly affects the dependence strength between the enterprise fund flow node and the direct factor node, and the weight adjustment of the lower layer causal graph changes the causal influence degree of the multi-modal driving factor node on the direct factor node. For example, during the policy release period, the global weight of the text modality node will be increased to 0.6 or above, while the weights of the image modality node and the time series modality node will decrease accordingly, in order to enhance the dominant role of policy semantics on the causal structure; during market fluctuations, the global weight of the time series modality node may increase to about 0.7, in order to strengthen the influence of transaction volatility features. The adjustment of weight distribution is completed through a parameter update mechanism, and the algorithm recalculates the global weight of each modality node at each time step and updates it to the edge weight structure of the hierarchical causal graph, forming a causal dependence configuration that matches the macro-environmental state. The dynamically adjusted hierarchical causal graph can keep pace with the changes of multi-modal features, so that the causal modeling has the ability of structure update with time adaptation, providing accurate input basis for subsequent time series modeling and counterfactual reasoning.

[0072] In S5, time series causal convolution and memory network are applied to process multi-modal feature vectors, capture long-period time series dependence, and output time series enhanced multi-modal representation.

[0073] In the multi-modal feature input stage, the feature sequence after semantic alignment is divided by time index to ensure the comparability of each modality on the unified time axis. Then, the time series causal convolution is applied to process the multi-modal feature vector. This process is constrained by time continuity, and the convolution kernel is used to perform local feature extraction on the feature sequence of adjacent time steps within the sliding time window. The size of the convolution kernel is adaptively adjusted according to the business cycle or the period characteristics of market data, for example, a convolution window of length 90 is used for quarterly operating data, and a convolution window of length 7 is used for daily trading fluctuations. Convolution operation only uses the feature vector of the current time and before to prevent future information leakage while maintaining the causality in the time direction. By continuously sliding the convolution kernel, the model can identify local change patterns within the time window, such as accelerated capital inflow, concentrated sales peak, or frequent inventory fluctuations. The result of convolution output is a set of local feature maps, representing the dynamic change trend of short-period behavior in the time dimension.

[0074] After obtaining the short-term pattern, the memory network is used to store the historical multi-modal feature vector sequence and extract long-period dependencies. The storage unit of the memory network records the historical state of the multi-modal feature sequence with time as the primary key. When the feature vector of the current time step is input, the most relevant historical feature vectors are searched in the stored historical feature set through a similarity retrieval mechanism. The similarity measurement standard is the vector distance in the feature space, such as cosine similarity or Euclidean distance, and when the similarity exceeds 0.85, it is considered as a related sample. Through the retrieval results, the model can trace back to the similar stage of the enterprise's history as the current operating state, such as the capital tightening period, the market overheating period, or the policy adjustment period, to extract potential long-term dependency patterns. The memory network update follows the first-in-first-out principle, and each time a new feature record is inserted, the oldest historical record is removed to maintain the storage capacity within a fixed range, such as a time span of one year. This updating process ensures the sensitivity of the model to recent changes while preserving the continuity of historical period characteristics.

[0075] After the short-term local pattern and long-period dependence extraction, the outputs of the causal convolution and the memory network are integrated. The integration process is completed through a fully connected layer, and the input includes local pattern feature maps and historical dependence feature vectors. The parameters of the fully connected layer are trained to adaptively allocate the fusion ratio of the two types of information. To balance the influence of local and global information, the initial weight ratio of the two types of features in the fusion layer is set to 1:1, and the ratio is automatically adjusted according to the prediction error in the iterative training. The output of the integrated fusion is a time-series enhanced multi-modal representation, which is a set of high-dimensional vector sequences, each of which contains short-term fluctuation features and long-term periodic dependence information. This enhanced representation not only retains the local dynamic characteristics of the changes in the financial receipts and expenditures, but also reflects the trend continuity at different time scales, providing a deep input basis in the time dimension for the subsequent counterfactual reasoning link, enabling the model to accurately infer the long-term response law of the enterprise's financial receipts and expenditures under external event scenarios.

[0076] In S6, counterfactual reasoning is performed based on the weight-adjusted hierarchical causal graph and the time-series enhanced multi-modal representation to generate long-period prediction data of the enterprise's financial receipts and expenditures.

[0077] The trained and dynamically updated hierarchical causal graph structure and global weights of the modal nodes are obtained. The hierarchical causal graph includes enterprise financial receipts and expenditures nodes, direct factor nodes, and multi-modal driving factor nodes. The directed edges between the nodes reflect the causal dependence path, and the weights of the edges represent the influence strength. After receiving the external event description information, the affected node set is identified from the hierarchical causal graph based on the semantic features, time attributes, and impact areas of the event. For example, when the external event is “macro policy adjustment”, the identification result is mainly concentrated in the text modal driving nodes related to the semantic features of the policy and the direct factor nodes related to the policy direction; if the event is “significant fluctuation in raw material prices”, the identification result is concentrated in the time-series modal nodes reflecting the state of the supply chain and the industry demand nodes related to the cost structure. The identification process integrates the semantic labels, modal weights, and historical event response records of the nodes. When the semantic relevance of a node is higher than 0.8 and there is a direct or indirect connection relationship between the node and the financial receipts and expenditures node in the historical causal path, the node is confirmed as an affected node. Subsequently, the corresponding counterfactual scenario is constructed based on the identification result, and the change direction and amplitude of the node state under the action of the external event are determined, for example, the state of the policy node is set from “stable” to “tightening”, and the state of the industry demand node is set from “stable” to “contraction”.

[0078] After constructing the counterfactual scenario, an intervention operation is applied to the identified node. The core of the intervention operation is to set the value of the affected node to the counterfactual condition, and propagate it to the fund flow node layer by layer through the causal path. In specific implementation, first, lock the state value of the node at the current time, and then set the new node value according to the event type. For example, in the market impact event, the characteristic value of the "market fluctuation node" is increased by 20% to simulate the short-term fluctuation enhancement; in the policy tightening situation, the characteristic value of the "policy direction node" is reduced by 15% to reflect the increase of regulatory strength. The time series enhanced multi-modal representation provides historical dependent information such as fund flow, inventory, and capacity utilization, ensuring time consistency during propagation. The intervention propagation is performed from bottom to top according to the order of the directed edges of the causal graph, and the downstream nodes directly connected to the intervened node are updated first, and then the indirect nodes are updated in turn, until the state value of the fund flow node is recalculated. The update amplitude of each propagation path is allocated according to the edge weight proportion. When the edge weight value is 0.6, it means that the path accounts for 60% of the total influence, and the downstream node change amount is adjusted accordingly to ensure that the influence transmission conforms to the causal structure logic.

[0079] After the propagation is completed, the system uses a Bayesian network model to infer the probability distribution of the change amount of the fund flow node. The model takes the node state set under the counterfactual scenario as input, combines the prior distribution and conditional dependence relationship, and infers the value range and uncertainty of the fund flow under various external situations. During the inference process, a large number of possible trajectories of the fund flow are generated through Monte Carlo sampling, and the frequency of each trajectory is counted to form the posterior distribution. If in the "policy tightening" scenario, the probability of fund inflow reduction calculated by the model exceeds 70%, the output result identifies the high-confidence conclusion that the fund is tight under this situation. Finally, by integrating various external event scenarios, the fund flow prediction trajectory under different policy, market or industry impact is output, forming a long-period prediction dataset with time label. The dataset structure is in time series form, and each time point corresponds to a fund flow probability distribution, which is used to support subsequent enterprise decision analysis.

[0080] The above formulas are dimensionless values calculated, and the formulas are obtained by software simulation of a large amount of data to obtain a formula of the latest real situation. The preset parameters and threshold values in the formula are set by a person skilled in the art according to the actual situation.

[0081] The above-described embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented by software, the above-described embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are wholly or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center through a wired (for example, infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0082] Those of ordinary skill in the art can realize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0083] Those of ordinary skill in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, device, and module can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0084] In several embodiments provided in the present application, it should be understood that the disclosed system, device, and method can be implemented in other ways. For example, the above-described device embodiments are only schematic, for example, the division of the modules is only a logical function division, and actual implementation can have another division manner, for example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed ones can be indirect coupling or communication connection through some interfaces, devices, or modules, which can be electrical, mechanical, or other forms.

[0085] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed on multiple network modules. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment.

[0086] In addition, the functional modules in each embodiment of the present application can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module.

[0087] The functions, if realized in the form of software function modules and sold or used as independent products, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0088] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0089] Finally: the above is only the preferred embodiment of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application, should be included in the protection scope of the present application.

Claims

1. A method for predicting enterprise cash inflows and outflows based on AI large-scale models and multimodal interaction, characterized in that: Includes the following steps: S1. Collect multimodal data related to the enterprise's cash inflows and outflows, including text data, image data, and time-series data; S2. Construct a hierarchical causal graph based on the multimodal data, wherein the hierarchical causal graph includes an upper-level causal graph and a lower-level causal graph; S3. Input the multimodal data into the multimodal Transformer architecture, perform cross-modal feature alignment in a unified semantic space, and output the corresponding modal feature vector; S4. Dynamically adjust the global weights of each modality in the hierarchical causal graph based on the change patterns of the multimodal feature vectors; S5. Apply time-series causal convolution and memory network to process multimodal feature vectors, capture long-period temporal dependencies, and output temporally enhanced multimodal representations; S6. Based on the weighted hierarchical causal graph and time-enhanced multimodal representation, perform counterfactual reasoning to generate long-term forecast data for corporate cash inflows and outflows; In step S4, dynamically adjusting the global weights of each modality in the hierarchical causal graph based on the change patterns of the multimodal feature vectors specifically includes: The current environmental state is identified based on the change pattern of multimodal feature vectors. The environmental state includes the policy release period, the industry transformation period, and the market fluctuation period. The change pattern is identified through the distribution shift and temporal fluctuation of the modal feature vectors. The importance score of each modality in the hierarchical causal graph is calculated using a reinforcement learning algorithm. The importance score is used to calculate the contribution of the modality to the causal relationship through a reward function, which is set based on the principle of minimizing prediction error. The global weight allocation of modal nodes in the upper and lower causal graphs is adjusted in real time based on importance scores.

2. The enterprise cash inflow and outflow forecasting method based on AI large model and multimodal interaction as described in claim 1, characterized in that, In step S1, multimodal data related to the company's cash inflows and outflows are collected. This multimodal data includes text data, image data, and time-series data, specifically: Perform data extraction operations from a structured database of enterprise information to extract text data containing policy and regulatory texts and industry analysis reports; Image data is acquired from a geographic information platform, and the image data is a record of images reflecting changes in resource distribution and infrastructure layout. Time-series data is collected from financial transaction records and market data sources. The time-series data includes capital flow rate indicators and market fluctuation sequences. The time-series data is sampled through a sliding window to generate continuous time series segments. Data cleaning and data format standardization are performed on the extracted text data, image data and time series data to convert data from different sources into a unified time series format and output preprocessed multimodal data.

3. The enterprise cash inflow and outflow forecasting method based on AI large model and multimodal interaction as described in claim 1, characterized in that, In step S2, a hierarchical causal graph is constructed based on the multimodal data. The hierarchical causal graph includes an upper-level causal graph and a lower-level causal graph, specifically comprising: Analyze the preprocessed multimodal data to identify the causal relationship between corporate cash inflows and outflows and direct factors. Generate a hierarchical causal graph based on conditional independence tests and causal edge direction inference. The hierarchical causal graph structure is divided into upper and lower layers. The hierarchical causal graph uses an online learning algorithm to incrementally update the graph structure, recalculating the weights and directions of causal edges when new multimodal data is input.

4. The enterprise cash inflow and outflow forecasting method based on AI large model and multimodal interaction according to claim 3, characterized in that, The hierarchical cause-effect graph structure is divided into upper and lower layers, specifically: The upper layer of the causal graph uses directed edges to represent the causal dependency between enterprise cash inflow and outflow nodes and direct factor nodes. The direct factors include policy adjustment factors and industry demand change factors. The lower layer of the cause-effect graph uses directed edges to represent the causal influence of multimodal driving factor nodes on direct factor nodes. The multimodal driving factor nodes are the modal nodes corresponding to the multimodal data. The weights of the directed edges in the upper and lower layers of the causal graph are calculated using maximum likelihood estimation.

5. The enterprise cash inflow and outflow forecasting method based on AI large model and multimodal interaction according to claim 1, characterized in that, In step S3, inputting the multimodal data into the multimodal Transformer architecture, performing cross-modal feature alignment within a unified semantic space, and outputting the corresponding modal feature vector specifically includes: Preprocessed text data, image data, and time-series data are input into the input channel of the multimodal Transformer model, and the model is set with independent modality coding units and a shared multi-head attention layer. The modality coding unit performs feature mapping on different modality data and outputs the corresponding high-dimensional feature vector sequence; The multi-head attention layer calculates the semantic relevance between the feature vector sequences of each modality in a unified semantic space as the association strength, and generates a cross-modal weight distribution based on the association strength; By adjusting the representation of each modality feature vector through cross-modal weight distribution, and using a weighted summation operation to fuse text feature vectors, image feature vectors, and temporal feature vectors, the data from different modalities are semantically aligned.

6. The enterprise cash inflow and outflow forecasting method based on AI large model and multimodal interaction according to claim 1, characterized in that, In step S5, the application of time-series causal convolution and memory networks to process multimodal feature vectors, capture long-period temporal dependencies, and output time-enhanced multimodal representations specifically includes: The time-series causal convolution is applied to process multimodal feature vectors. The size of the convolution kernel is adaptively adjusted according to the time period, and the output is a short-term local pattern of local changes in a continuous time window. Based on the memory network to store the sequence of historical multimodal feature vectors, long-term dependencies are extracted by retrieving historical feature vectors that are similar to the current multimodal feature vectors; By integrating the outputs of causal convolutions and memory networks, and using fully connected layers to fuse short-term local patterns and global dependency information, a temporally enhanced multimodal representation is generated.

7. The enterprise cash inflow and outflow forecasting method based on AI large model and multimodal interaction according to claim 1, characterized in that, In step S6, counterfactual reasoning is performed based on the weighted hierarchical causal graph and time-enhanced multimodal representation to generate long-term forecast data for corporate cash inflows and outflows. This specifically includes: Obtain the hierarchical cause-effect graph structure and global weights of modal nodes, identify direct factor nodes or multimodal driving factor nodes affected by external events, and construct corresponding counterfactual scenarios; Intervention operations are applied to the identified nodes, setting the node values ​​as counterfactual conditions. Based on the historical dependency information provided by the time-enhanced multimodal representation, the intervention effect is propagated layer by layer along the directed edge structure of the causal graph to the enterprise's cash inflow and outflow nodes. The Bayesian network model is used to infer the probability distribution of changes in the nodes of fund inflow and outflow, and to output the predicted trajectory of fund inflow and outflow under various external event scenarios.

Images