An intelligent sales forecasting system based on multi-modal data fusion

Abstract

The application discloses a kind of intelligent sales forecasting system based on multi-modal data fusion, it is related to intelligent management technical field, including, feature extraction module, through convolution calculation method, using first convolution layer and second convolution layer, one-dimensional convolution calculation is carried out to sales data input set, extract short-term time series feature and long-term time series feature, and using residual superposition method is fused, generates sales data time series correlation feature;Semantics extraction module, channel sales record and external text data are classified, segmented and context semantic encoding, generate trend report feature vector;Feature fusion module, sales data time series correlation feature and trend report feature vector are nonlinearly fused and encoded, and generate decoding feature matrix;Decoding regression module, using regression decoding method, decoding regression is carried out to decoding feature matrix, and generates sales volume prediction result.The application realizes the high accuracy and stability of sales volume prediction result.

Description

Technical Field

[0001] This invention belongs to the technical field of intelligent management, and in particular to an intelligent sales forecasting system based on multimodal data fusion. Background Technology

[0002] In modern enterprise management, sales forecasting is widely considered a key element for accurate decision-making and optimized resource allocation. With the development of information technology and data analytics, traditional sales forecasting methods have gradually evolved from simple historical data statistical analysis to forecasting methods based on time series models (such as ARIMA) and regression analysis. These methods utilize past sales records to predict future sales trends and provide references for inventory control, marketing planning, and resource allocation. As the market environment becomes increasingly complex and diverse, enterprises acquire data from a wider range of sources, including online sales records, channel data, social media texts, and industry reports.

[0003] To address the shortcomings of existing technologies, some studies have attempted to enhance predictive capabilities by introducing external data or improving feature engineering. However, problems remain, including insufficient capture of feature associations, limited ability to model nonlinear relationships, and biases in predictions under complex market environments. Traditional models struggle to respond quickly and make effective adjustments to the impact of sales fluctuations caused by short-term promotional activities or changes in external public opinion on sales trends.

[0004] The intelligent sales forecasting method based on multimodal data fusion integrates historical sales data, channel records, and external text information to achieve deep fusion of temporal feature extraction and semantic information encoding. This can improve forecast accuracy while generating management execution results that can be used for inventory optimization, marketing strategy adjustment, and resource allocation, thereby directly solving the limitations of existing technologies in terms of forecast accuracy and decision support. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes an intelligent sales forecasting system based on multimodal data fusion to resolve the issues of insufficient sales forecasting accuracy and limited support for management decisions. The specific solution is as follows: An intelligent sales forecasting system based on multimodal data fusion includes, The data acquisition module collects channel sales records and external text data, performs preprocessing, and generates a sales data input set. The feature extraction module uses a convolution calculation method, employing the first and second convolutional layers to perform one-dimensional convolution calculations on the sales data input set, extracting short-term and long-term time-series features, and then using a residual superposition method to fuse them, generating time-series correlation features of the sales data. The semantic extraction module classifies, segments, and encodes contextual semantics of channel sales records and external text data to generate trend report feature vectors. The feature fusion module performs non-linear fusion encoding on the time-series correlation features of sales data and the feature vectors of trend reports to generate a decoded feature matrix. The decoding and regression module uses regression decoding methods to perform decoding regression on the decoded feature matrix and generate sales volume prediction results. The decision execution module performs inventory optimization, marketing strategy adjustment, and resource allocation operations based on sales volume forecast results, generating management execution results.

[0006] Preferably, the channel sales records include online order data, offline store sales data, and distributor purchase records; The external text data includes industry policy information, holiday information, and user review text.

[0007] Preferably, generating the sales data input set includes the following steps: Entity parsing, deduplication and alignment, and timestamp standardization are performed on channel sales records and external text data to generate a time-aligned multimodal dataset; Semantic parsing and temporal completion are performed on time-aligned multimodal datasets to extract text trend features. Missing data is also filled in and anomaly annotation is performed to generate a multimodal augmented dataset. By using causal feature enhancement methods, feature standardization is performed on the multimodal augmentation dataset to obtain a multimodal association feature set. The data is then organized using an efficient index storage method to generate a sales data input set.

[0008] Preferably, the extraction of short-term and long-term time-series features includes the following steps: The sales data input set is sliced ​​according to time series, and then input into the first convolutional layer for one-dimensional convolution calculation. Short-term local change patterns within the window are extracted to generate short-term time series features. Short-term temporal features are input into the second convolutional layer, and one-dimensional convolution is performed using an expanded convolutional kernel to capture long-term dependencies across time periods and generate long-term temporal features.

[0009] As a preferred embodiment of the intelligent sales forecasting system based on multimodal data fusion described in this invention, the generation of time-series correlation features of sales data includes the following steps: Short-term time-series features are processed by channel mapping and batch normalization, and long-term time-series features are projected and mapped to the same dimension and then residually superimposed with short-term time-series features to generate a preliminary fused feature sequence. Using the channel attention method, the preliminary fused feature sequence is weighted by time steps to obtain the weighted fused feature sequence, and local frequency domain analysis is performed to generate the frequency-time fused feature sequence; Gated fusion, projection dimensionality reduction and normalization are performed on the frequency-time fusion feature sequences to generate time-series correlation features of sales data.

[0010] Preferably, generating the trend report feature vector includes the following steps: Perform thematic classification and tagging of channel sales records and external text data to generate a set of categorized and labeled texts; The classified and labeled text set is segmented into words, and each segmented word is mapped into a high-dimensional word vector sequence while maintaining the original context order, thus generating a preliminary text feature sequence. Contextual semantic encoding is performed on the preliminary text feature sequence to obtain the contextual semantic representation result; Global semantic modeling and contextual information integration are performed on high-dimensional word vector sequences to generate global semantic features; The contextual semantic representation results are fused with global semantic features to generate a trend report feature vector.

[0011] Preferably, generating the decoding feature matrix includes the following steps: The time-series correlation features of sales data and the feature vectors of trend reports are mapped to a unified feature space to generate a mapped feature set. By using a weighted nonlinear mapping method, the fusion ratio of the time-series correlation features of sales data and the feature vectors of trend reports in the mapped feature set is adjusted to generate a weighted feature set. The weighted feature set is integrated to capture the non-linear relationship between the time-series correlation features of sales data and the feature vector of trend report, and a decoded feature matrix is ​​generated.

[0012] Preferably, generating the sales volume forecast result includes the following steps: The time-series correlation features of sales data and the feature vector of trend report in the decoded feature matrix are weighted and decoded to generate a preliminary prediction vector; The initial prediction vector is corrected for errors and trends. Short-term fluctuations and long-term trends are captured through residual calculation and feedback processing to generate an optimized prediction vector. The optimized forecast vector is mapped to obtain a sales trend representation, and time series smoothing and confidence interval calculation are performed to generate sales volume forecast results.

[0013] More preferably, the step of performing error correction and trend correction on the initial prediction vector, and capturing short-term fluctuations and long-term trends through residual calculation and feedback processing to generate an optimized prediction vector, includes the following steps: Historical error analysis is performed on the preliminary forecast vector to calculate the residual sequence between the forecast value and the actual sales value, and to identify short-term fluctuations and long-term trends. Multi-scale trend decomposition is performed on the residual sequence to separate short-term fluctuations from long-term trends and weighted adjustments are made to generate a trend correction vector. The residual feedback integration method is adopted to integrate the trend correction vector with residual feedback, balance the influence of short-term fluctuations and long-term trends, and generate an optimized prediction vector.

[0014] Preferably, the generation of management execution results includes the following steps: Based on sales volume forecasts, analyze inventory demand, assess safety stock levels and replenishment plans, and generate inventory adjustment schemes. By combining inventory adjustment plans with sales forecasts, the impact of different marketing strategies on sales can be quantified, and optimized marketing plans can be generated. Based on the inventory adjustment plan and marketing optimization plan, optimize the allocation of resources for inventory, sales channels and marketing resources, and generate management execution results.

[0015] The beneficial effects of this invention are: 1. The preliminary fused feature sequence is weighted by time step using the channel attention method, and periodic patterns are extracted by combining local frequency domain analysis. Then, the final sales data time-series correlation features are generated by gating fusion, projection dimensionality reduction and normalization, which realizes differentiated processing and feature enhancement of short-term fluctuations and long-term trends.

[0016] 2. It can automatically highlight important signals in key time periods while suppressing noise, improving the effectiveness and interpretability of feature representation. This allows subsequent multimodal fusion modules to more accurately integrate sales data and textual trend information in a unified feature space, achieving high accuracy and stability in sales volume prediction results. Attached Figure Description

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. However, it should be understood that these drawings are designed for illustrative purposes only and are not intended to limit the scope of the present invention. Furthermore, unless specifically indicated, these drawings are intended only to conceptually illustrate the structural construction described herein and are not necessarily drawn to scale.

[0018] Figure 1 This is a flowchart illustrating the process of an intelligent sales forecasting system based on multimodal data fusion. Figure 2 A flowchart for generating the sales data input set; Figure 3 A flowchart for generating time-series correlation features of sales data; Figure 4A flowchart for generating sales volume forecast results. Detailed Implementation

[0019] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0020] The following is in conjunction with the appendix Figure 1 ~Appendix Figure 4 This invention will be described in detail.

[0021] As attached Figure 1 ~Appendix Figure 4 As shown in the figure, an intelligent sales forecasting system based on multimodal data fusion provided by an embodiment of the present invention includes: The data acquisition module collects channel sales records and external text data, performs preprocessing, and generates a sales data input set.

[0022] In this embodiment, channel sales records include online order data, offline store sales data, and distributor purchase records.

[0023] Specifically, online order data is collected item by item, recording fields such as order number, order time, product category, sales quantity, and payment amount; offline store sales data is collected and structured according to fields such as store number, sales date, product category, and sales quantity; distributor purchase records are summarized, extracting fields such as purchase date, product batch, purchase quantity, and supply amount; and online order data, offline store sales data, and distributor purchase records are aligned according to time sequence and a unified field format to form channel sales records.

[0024] External text data includes industry policy information, holiday information, and user review text.

[0025] Specifically, industry policy information is collected, and fields such as policy release time, policy title, and policy text content are extracted and formatted; holiday information is organized, and fields such as holiday name, start date, end date, and duration are recorded; user review text is captured and extracted, and fields such as review release time, review content, user identifier, and product identifier are retained; industry policy information, holiday information, and user review text are stored in a unified text format and arranged in chronological order to form external text data.

[0026] In this embodiment, generating the sales data input set includes the following steps: 1. Perform entity parsing, deduplication and alignment, and timestamp standardization on channel sales records and external text data to generate a time-aligned multimodal dataset.

[0027] Specifically, entity parsing is performed on channel sales records and external text data. Predefined fields are extracted from online order data, offline store sales data, distributor purchase records, industry policy information, holiday information, and user review text. These fields include timestamps, product identifiers, channel identifiers, order numbers, transaction quantities, transaction amounts, text content, and author identifiers. The field names are then standardized into a unified field set. The entity parsing results are deduplicated by merging duplicate records through primary key mapping and field similarity matching. Product identifiers and channel identifiers are encoded and mapped to achieve cross-data source field alignment. The timestamps are converted to a unified time zone and ISO 8601 format and rounded and aggregated at the daily granularity to generate sales records and text entries indexed by date. The aligned structured sales fields and text fields are then linked and merged by date and product identifier to form a time-aligned multimodal dataset.

[0028] 2. Perform semantic parsing and temporal completion on the time-aligned multimodal dataset, extract text trend features, and simultaneously perform missing data completion and anomaly labeling to generate a multimodal augmented dataset.

[0029] Specifically, word segmentation and entity recognition are performed on the text field, and the segmentation results are converted into word vector sequences through pre-trained word vector mapping to generate text records with word vector annotations. Taking the text records with word vector annotations as input, the word vector sequences are encoded using a Transformer encoder, and the topic distribution is extracted based on the TF-IDF method and the sentiment score is calculated based on the sentiment dictionary to generate text trend features. At the same time, the time series field in the time-aligned multimodal dataset is taken as input, and missing data completion is performed. The nearest neighbor window interpolation and Bayesian smoothing methods are used to obtain the completed time series, and a confidence label is attached to each completion point. Taking the completed time series and text trend features as input, anomaly annotation is carried out. Seasonal decomposition is performed on the time series to obtain residuals, and anomalies are identified based on the statistical distribution characteristics of the residuals. Meanwhile, Z-score burst detection is applied to the time series of text topics to identify topic anomalies, and an anomaly annotation table is generated. The dataset is composed of the completed time series data, text trend features, confidence labels, and anomaly labels as input. It is then merged into a multimodal augmented dataset based on predefined fields, including: time index, product identifier, completed sales quantity, text trend features, sentiment score, anomaly label, and completed confidence.

[0030] It should also be noted that pre-trained word vectors include word vector models trained on large-scale public corpora, such as corpora constructed using Wikipedia text, news corpora, and product review text, trained using the FastText method to generate high-dimensional word vectors with contextual semantic relationships; in the specific implementation process, existing pre-trained word vector models, such as 300-dimensional Word2Vec word vectors or sub-word vectors trained based on FastText, can be called to map the text segmentation results to the corresponding high-dimensional word vector representations; The settings of predefined fields are determined based on business analysis needs and model training requirements, combined with time index, product identifier, completed sales quantity, text trend features, sentiment score, anomaly labeling, completion confidence, and analysis objectives. For example, the time index is used for sequence alignment and time series modeling, the product identifier is used to distinguish different objects, the completed sales quantity reflects core numerical features, the text trend features and sentiment score represent auxiliary information, the anomaly labeling is used for anomaly detection and model supervision, and the completion confidence is used to measure data reliability, thereby ensuring that the multimodal augmented dataset has a complete structure and can be used for subsequent analysis and modeling.

[0031] 3. Using causal feature enhancement methods, feature standardization is performed on the multimodal augmentation dataset to obtain a multimodal association feature set. The data is then organized using an efficient index storage method to generate a sales data input set.

[0032] Specifically, the mean and standard deviation of each type of numerical feature in the multimodal augmentation dataset are calculated, and the features are normalized based on the mean and standard deviation. For example, the sales quantity feature is converted into a form with zero mean and unit variance. The categorical features in the multimodal augmentation dataset are one-hot encoded, and feature interaction processing is introduced into the causal feature augmentation method. For example, sales channel features are interactively encoded with holiday information to generate new interactive feature vectors. The multimodal features after standardization and interaction augmentation are merged into a multimodal associated feature set. The multimodal associated feature set is organized through an efficient index storage method. For example, a multi-level index table is built according to the timestamp field and the channel sales record field to achieve fast retrieval and access, generating a sales data input set.

[0033] The feature extraction module uses a convolution calculation method, employing the first and second convolutional layers to perform one-dimensional convolution calculations on the sales data input set, extracting short-term and long-term time-series features, and then using a residual superposition method to fuse them, generating time-series correlation features of the sales data.

[0034] In this embodiment, the extraction of short-term and long-term time-series features includes the following steps: 1. Slice the sales data input set according to time series, input it into the first convolutional layer for one-dimensional convolution calculation, extract short-term local change patterns within the window, and generate short-term time series features.

[0035] Specifically, the sales data input set is sliced ​​into time series segments with a fixed window length, using a sliding window approach with a set step size to generate a windowed time segment sequence; standardization is calculated and applied to each time segment; the standardized time segments are stacked in chronological order into a three-dimensional tensor; the three-dimensional tensor is input into the first convolutional layer, which uses a one-dimensional convolution kernel to perform convolution operations on the time dimension, capturing local numerical fluctuations and short-term trends within the window to form a local pattern response; the one-dimensional convolution output is processed by normalization and a nonlinear activation function (e.g., ReLU), and local max pooling can be optionally used to compress the time dimension and suppress noise; the processed time step features are reconstructed into a serialized representation according to the time series to generate short-term time series features.

[0036] 2. Input the short-term temporal features into the second convolutional layer, and use the extended convolutional kernel to perform one-dimensional convolution to capture long-term dependencies across time periods and generate long-term temporal features.

[0037] Specifically, short-term time-series features are stacked into a three-dimensional tensor along the time dimension and input into the second convolutional layer. The second convolutional layer uses an extended convolutional kernel to perform one-dimensional convolution along the time dimension to extract long-term patterns across the window using a sliding window approach, outputting a long-term feature mapping tensor. Channel mapping is performed on the long-term feature mapping tensor to adjust the output channel dimension, and batch normalization and nonlinear activation processing are implemented. When the output channel dimension is inconsistent with the channel dimension of the short-term time-series features, the short-term time-series features are first linearly projected to match the channel dimension. The projected short-term time-series features are then fused with the processed long-term feature mapping through the residual path to maintain information flow. The fused time series is reconstructed in chronological order, and its length is adjusted as needed through time dimension interpolation to generate long-term time-series features.

[0038] It should also be noted that the channel mapping of the long-term feature mapping tensor performs a linear transformation in the channel dimension through a convolution kernel, mapping the linear combination of input channels to a specified number of output channels; during the mapping process, each output channel is formed by the weighted sum of the input channels, batch normalization is performed on the mapping results to stabilize the numerical distribution, and nonlinear activation functions (such as ReLU) are used to enhance the nonlinearity and improve the feature representation ability.

[0039] The generation of time-series correlation features for sales data includes the following steps: (1) Channel mapping and batch normalization are performed on the short-term time series features, and the long-term time series features are projected and mapped to the same dimension and then superimposed with the short-term time series features to generate a preliminary fusion feature sequence.

[0040] Specifically, channel mapping is performed on short-term time-series features to adjust the channel dimension and batch normalization is applied to generate short-term time-series feature representations that have undergone channel mapping and batch normalization. Projection mapping is applied to long-term time-series features to make the channel dimension of the long-term time-series features consistent with the short-term time-series feature representations that have undergone channel mapping and batch normalization. Projection mapping can be achieved through one-dimensional convolution or linear transformation to generate projected long-term time-series features. Using the short-term time-series feature representations that have undergone channel mapping and batch normalization and the projected long-term time-series features as inputs, residual superposition is performed in an element-wise addition manner to generate a preliminary fused feature sequence.

[0041] (2) Using the channel attention method, the preliminary fused feature sequence is weighted by time step to obtain the weighted fused feature sequence, and local frequency domain analysis is performed to generate the frequency-time fused feature sequence.

[0042] Specifically, global average pooling is performed along the time dimension on the preliminary fused feature sequence to obtain the statistics for each channel. The channel statistics are input into two fully connected layers and mapped by nonlinear activation and the Sigmoid function to generate channel weights. Channel response modulation is performed on the preliminary fused feature sequence according to the channel weight vector. The weighted fused feature sequence is obtained by proportionally adjusting the feature values ​​of the corresponding channels in each time step according to the channel weights. Short-time Fourier transform is performed on the weighted fused feature sequence according to the sliding time window to extract the local frequency domain features of each window. The time domain features of each time step and the frequency domain description of the corresponding window are aligned by channel and then fused by splicing or linear mapping to generate the frequency-time fused feature sequence.

[0043] (3) Perform gated fusion, projection dimensionality reduction and normalization on the frequency-time fusion feature sequence to generate time-series correlation features of sales data.

[0044] Specifically, a gating coefficient sequence with the same length as the frequency-time fusion feature sequence is generated in the time dimension. The gating coefficient sequence is then used to filter and modulate the frequency-time fusion feature sequence step by step to form a gating output sequence. The gating output sequence is then input into a fully connected mapping layer for projection dimensionality reduction, mapping the high-dimensional features to the target low-dimensional feature representation. Finally, batch normalization processing is performed on the projected dimensionality reduction features in both the channel dimension and the time dimension to eliminate numerical distribution differences and maintain scale consistency, thereby generating time-series correlation features of sales data.

[0045] It should also be noted that by using multi-layer convolution and residual stacking, nonlinear and multi-scale features can be modeled simultaneously, thereby improving the ability to capture complex fluctuations and long-term trends in sales data. This achieves effective integration of short-term fluctuations and long-term trends in sales data, providing high-quality and structured temporal feature representations for subsequent multimodal feature fusion and sales volume prediction, significantly improving prediction accuracy and robustness, and enhancing enterprises' responsiveness to sales trends in complex market environments.

[0046] The semantic extraction module classifies, segments, and encodes contextual semantics of channel sales records and external text data to generate trend report feature vectors.

[0047] In this embodiment, generating the trend report feature vector includes the following steps: 1. Perform thematic classification and tagging on channel sales records and external text data to generate a collection of categorized and labeled texts.

[0048] Specifically, the text is preprocessed, including removing stop words, punctuation marks, and meaningless characters, and uniformly encoding the text into UTF-8 format; the text is then matched to a predefined topic classification thesaurus, for example, text involving price adjustments is labeled as "price change", text involving holiday promotions is labeled as "holiday promotion", and text involving user feedback is labeled as "user review"; for each text, topic tags are attached as category tags, and a set of category-labeled text is generated based on the tag numbers, where each text corresponds to the example format {text content, topic tag number}, generating a set of category-labeled text.

[0049] It should also be noted that the predefined topic classification thesaurus is set according to the core topics of business concern and the semantic features of the text, and each topic corresponds to a set of representative keywords or phrases; the construction of the topic classification thesaurus can be determined based on the statistical analysis of historical text data and business rules to ensure that it can cover the typical expressions of the target topic and distinguish the semantic differences between different topics.

[0050] 2. Segment the labeled text set into words and map each segment into a high-dimensional word vector sequence while maintaining the original context order to generate a preliminary text feature sequence.

[0051] Specifically, each text is segmented according to natural language segmentation rules. For example, Chinese text is segmented using Jieba segmentation or dictionary-based segmentation methods, while English text is segmented using spaces or punctuation marks as separators, resulting in a segmentation sequence. Each segment is then used to find its corresponding high-dimensional word vector in a pre-trained word vector table. For example, each word vector has an example dimension of 300. The segmentation order is kept unchanged and arranged to form a high-dimensional word vector sequence. The high-dimensional word vector sequences of each text are then integrated according to the original context order to generate a preliminary text feature sequence, where the preliminary text feature sequence for each text maintains the same word order as the original text.

[0052] It should also be noted that the pre-trained word vector table comes from publicly available natural language processing corpora. For example, Chinese text can use word vector tables trained from Chinese Wikipedia corpora, and English text can use word vector tables trained from English Wikipedia or general corpora (such as Google News corpora). The word vectors are generated using existing word embedding training methods such as Word2Vec, GloVe, or FastText to ensure that each word segment corresponds to a unique high-dimensional vector in the table, which is used for subsequent text feature sequence construction.

[0053] 3. Perform contextual semantic encoding on the preliminary text feature sequence to obtain the contextual semantic representation result.

[0054] Specifically, the initial text feature sequence is input into an existing context encoding method, such as a bidirectional long short-term memory network, according to its original context order. Each word segmentation vector is then used to generate a context-dependent representation based on the preceding and following word segmentation information, and the encoding results are standardized to ensure numerical stability. All word segmentation context-dependent representations are then aggregated in the original text order to form a context semantic representation. During processing, for text with insufficient sequence length, a sequence padding method is used to fill in the gaps, for example, to a fixed example length. For text exceeding the fixed length, a truncation method is used to retain the length of the front or back example, generating a context semantic representation. For example, the text "During the promotion, the store adjusted prices and launched a limited-time discount activity, which received positive user feedback, but there were many logistics delays" is truncated to the first 10 words "During the promotion, the store adjusted prices and launched a limited-time discount activity," and the resulting text context semantic representation is obtained through text encoding. This can be used for subsequent topic classification, text clustering, or anomaly detection.

[0055] 4. Perform global semantic modeling and contextual information integration on high-dimensional word vector sequences to generate global semantic features.

[0056] Specifically, the high-dimensional word vector sequence is input into existing global encoding methods, such as multi-head self-attention mechanisms or global context encoding networks, in the order of the original text. Global dependencies are captured based on the correlation between each word vector in the sequence and other word vectors in the sequence, and the encoding results are standardized to ensure numerical stability. The global dependency representations of each word are integrated into a unified sequence representation, and the integrated sequence is subjected to dimensional unification and necessary padding, such as padding to the example length. The integrated global sequence representation is used to generate global semantic features.

[0057] 5. The contextual semantic representation results are fused with global semantic features to generate a trend report feature vector.

[0058] Specifically, the contextual semantic representation results are aligned with the global semantic features by time or word order to ensure consistency in sequence length and feature dimension; projection mapping is applied to the global semantic features to match the dimensions of the contextual semantic representation results, for example, by adjusting to the example dimension through linear transformation; the projected global semantic features and contextual semantic representation results are combined sequentially according to the corresponding time step or word position to generate a preliminary fusion sequence; normalization processing and necessary nonlinear mappings, such as example activation functions, are performed on the fusion sequence to ensure numerical stability; the processed fusion sequence is arranged in chronological order to form a trend report feature vector.

[0059] The feature fusion module performs non-linear fusion encoding on the time-series correlation features of sales data and the feature vectors of trend reports to generate a decoded feature matrix.

[0060] In this embodiment, generating the decoding feature matrix includes the following steps: 1. Map the time-series correlation features of sales data and the feature vectors of trend reports to a unified feature space to generate a mapped feature set.

[0061] Specifically, the time-series correlation features of sales data and the feature vectors of trend reports are input into a linear projection operation to adjust the dimensions of the time-series correlation features and trend report feature vectors to a unified instance dimension, for example, mapping to 128 dimensions; the mapped time-series correlation features and trend report feature vectors are aligned sequentially by time step or feature position to ensure consistent sequence length; the two aligned feature sequences are combined sequentially according to their corresponding positions to form a preliminary mapped feature sequence; batch normalization and nonlinear mapping processing, such as instance activation functions, are performed on the preliminary mapped feature sequence to ensure the stability of feature values; the processed sequence is then organized to generate a mapped feature set.

[0062] 2. By using a weighted nonlinear mapping method, the fusion ratio of the time-series correlation features of sales data and the feature vector of trend reports in the mapped feature set is adjusted to generate a weighted feature set.

[0063] Specifically, initial fusion weights are assigned to the sales data time-series correlation features and trend report feature vectors in the mapped feature set; nonlinear mapping processing is performed on the sales data time-series correlation features and trend report feature vectors respectively, for example, using the ReLU function to obtain nonlinear transformed features; based on the set initial fusion weights, the nonlinearly transformed sales data time-series correlation features and trend report feature vectors are fused and adjusted according to their corresponding positions to form a weighted fusion sequence; batch normalization processing is performed on the weighted fusion sequence to ensure the stability of feature values; the processed sequence is then organized to generate a weighted feature set.

[0064] It should also be explained that the specific steps for setting the initial fusion weights are as follows: Based on the contribution ratio of historical sales data and trend report data, determine the example initial weights, for example, set the sales data time-series correlation feature to 0.6 and the trend report feature vector to 0.4; align the sales data time-series correlation feature and the trend report feature vector according to the time step or corresponding dimension; at each alignment position, mark the sales data time-series correlation feature and the trend report feature vector with the corresponding initial fusion weight value; bind the marked weight information with the corresponding feature to form an initial fusion weight sequence that can be used for subsequent weighted nonlinear mapping.

[0065] 3. Integrate the weighted feature set to capture the non-linear relationship between the time-series correlation features of sales data and the feature vector of trend report, and generate a decoded feature matrix.

[0066] Specifically, the weighted feature set is subjected to nonlinear integration processing, including applying nonlinear mapping to the time-series correlation features of sales data and the feature vector of trend report to unify the feature dimensions. Through feature sequence splicing and serialization, the time-series correlation features of sales data and the feature vector of trend report at the corresponding time step are combined accordingly. Then, the combined feature sequence is normalized and regularized to balance the feature scale and generate a decoded feature matrix. For example, after mapping the example dimension 128 of the time-series correlation features of sales data and the example dimension 64 of the trend report feature vector to the unified dimension 128, a decoded feature matrix is ​​formed.

[0067] The decoding and regression module uses regression decoding methods to perform decoding regression on the decoded feature matrix and generate sales volume prediction results.

[0068] In this embodiment, generating the sales volume forecast result includes the following steps: 1. Perform weighted decoding on the time-series correlation features of sales data and the feature vector of trend report in the decoded feature matrix to generate a preliminary prediction vector.

[0069] Specifically, sample weighting coefficients are set for the time-series correlation features of sales data and the feature vectors of trend reports; the decoded feature matrix is ​​divided by time steps, and nonlinear mapping is applied to the time-series correlation features of sales data and the feature vectors of trend reports for each time step to match the output dimension; the mapped time-series correlation features of sales data and the feature vectors of trend reports are fused according to the set weighting coefficients to form a weighted feature representation for each time step; the weighted feature representations of all time steps are arranged in chronological order to generate a preliminary prediction vector; for example, the unified dimension of the time-series correlation features of sales data and the feature vectors of trend reports can be set to 128, and the sample length of the preliminary prediction vector is the sequence of predicted values ​​for the next 7 time steps.

[0070] It should also be explained that the specific steps for setting the example weighting coefficients are as follows: determine the names of the features to be weighted, such as the sales data time-series correlation feature and the trend report feature vector; set the initial example values ​​according to the importance or historical contribution of the features, such as a weight of 0.7 for the sales data time-series correlation feature and a weight of 0.3 for the trend report feature vector; normalize the example weights so that the sum of the weights of all features to be weighted is 1; record the normalized example weights and associate them with the corresponding feature names for use in the weighted fusion or weighted decoding process.

[0071] 2. The initial prediction vector is corrected for errors and trends. Short-term fluctuations and long-term trends are captured through residual calculation and feedback processing to generate an optimized prediction vector.

[0072] Furthermore, historical error analysis is performed on the preliminary prediction vector to calculate the residual sequence between the predicted value and the actual sales value, thereby identifying short-term fluctuations and long-term trends.

[0073] Specifically, the preliminary prediction vector and the actual sales value at the corresponding time point are read in chronological order, and each prediction value and the corresponding actual sales value are used to form a time-referenced sequence. For each pair of time-referenced sequences, residual records are generated to record the difference between the prediction and the actual value at each time point, for example, by marking example residual values ​​on the time series. The residual series is smoothed to identify long-term trends, for example, by using the example moving average method or exponential smoothing method to generate long-term trends, and the trend curve is compared with the residual series. Based on the trend curve, short-term fluctuations are separated from the residual series, for example, by statistically analyzing the range of residual changes within each window through a fixed-length sliding window to form short-term fluctuations.

[0074] Multi-scale trend decomposition is performed on the residual sequence to separate short-term fluctuations from long-term trends and weighted adjustments are made to generate a trend correction vector.

[0075] Specifically, the residual sequence is organized into a continuous time point sequence in chronological order; based on the example time window length or multi-scale decomposition parameters, the residual sequence is divided into time segments of different scales, and a local residual subsequence is generated for each time segment; exponential smoothing is applied to the local residual subsequence to obtain the long-term trend component of the corresponding time segment, and the remaining part is used as the short-term fluctuation component; the long-term trend component and the short-term fluctuation component are weighted and adjusted according to the example weighting coefficient to generate a weighted trend component and a weighted fluctuation component; the weighted trend component and the weighted fluctuation component are recombined in chronological order to generate a trend correction vector arranged by time step, with each record containing the example time point, weighted trend value, and weighted fluctuation value, which are used for subsequent prediction vector correction.

[0076] The residual feedback integration method is adopted to integrate the trend correction vector with residual feedback, balance the influence of short-term fluctuations and long-term trends, and generate an optimized prediction vector.

[0077] Specifically, the trend correction vector is arranged into a continuous sequence by time step; based on the example residual feedback coefficient, the short-term fluctuation component and the long-term trend component in the trend correction vector are adjusted to balance the influence of fluctuation and trend at each time step; the adjusted data of each time step are accumulated sequentially to generate a residual feedback sequence, with each record containing weighted short-term fluctuation and long-term trend information corresponding to the example time point; a sequence integration operation is performed on the residual feedback sequence to form an optimized time step data sequence; the optimized time step data sequence is organized to generate an optimized prediction vector, with each record corresponding to the predicted value at the example prediction time point, and maintaining the same time step order as the initial prediction vector, for subsequent prediction result output.

[0078] 3. Map the optimized forecast vector to obtain a sales trend representation, and perform time series smoothing and confidence interval calculation to generate sales volume forecast results.

[0079] Specifically, the optimized prediction vector is mapped to the corresponding position in the sales trend representation according to the time step to form a continuous time series; the sales trend representation is smoothed, for example, by using a sample moving average or exponential smoothing method to reduce short-term fluctuations between time steps, generating a smoothed sales trend series; based on the smoothed sales trend series, the confidence interval for each time step is calculated, for example, by using the sample historical error distribution or standard deviation to define the upper and lower bounds, generating a confidence interval record for each time step; the smoothed sales trend series and confidence intervals are arranged in chronological order to form the sales volume prediction result, with each record containing the predicted value and confidence interval for the sample time step.

[0080] It should also be noted that by weighting the decoding of various features in the feature matrix, the impact of short-term temporal fluctuations, long-term trends, and external textual semantic information on sales volume can be considered simultaneously, thereby more accurately capturing the nonlinear relationship of sales patterns. Effective decoding and quantization mapping of multimodal features are achieved, enabling sales volume prediction results to not only reflect historical sales data patterns but also incorporate external factors such as market trends, holiday influences, and user reviews. This significantly improves prediction accuracy and reliability, providing a basis for decision-making in inventory management, marketing strategy optimization, and resource allocation.

[0081] The decision execution module performs inventory optimization, marketing strategy adjustment, and resource allocation operations based on sales volume forecast results, generating management execution results.

[0082] In this embodiment, generating the management execution result includes the following steps: Based on sales volume forecasts, analyze inventory demand, assess safety stock levels and replenishment plans, and generate inventory adjustment schemes.

[0083] Specifically, based on the sales volume forecast results, daily or periodic forecast sales volumes are organized according to time series to form a forecast sales volume sequence; combined with existing inventory levels and safety stock rules, the required safety stock level for each time step is calculated, for example, using historical sales fluctuations as an example to determine the safety stock benchmark; based on the safety stock level and the forecast sales volume sequence, a replenishment plan is formulated, including example replenishment time points and replenishment quantities, and inventory adjustment records for each time step are generated; the inventory adjustment records are organized in chronological order to form an inventory adjustment scheme, where each record contains the forecast sales volume for the example time step, the corresponding safety stock level, and the recommended replenishment quantity, which is used to guide inventory management and execution.

[0084] It should also be noted that the current inventory level refers to the actual quantity of various goods currently stored in the sales channels or warehouses, which is used to reflect the current inventory status; the safety stock rule refers to the inventory holding standard preset based on factors such as historical sales fluctuations, forecast errors and supply cycles. For example, the safety stock level is set at 1.2 times the average daily sales volume of the past 30 days or the upper limit of forecast error, so as to ensure that demand can still be met in the event of sales fluctuations or replenishment delays, thereby providing a reference for inventory adjustment and replenishment plans.

[0085] By combining inventory adjustment plans with sales forecasts, the impact of different marketing strategies on sales can be quantified, and optimized marketing plans can be generated.

[0086] Specifically, each product category and corresponding replenishment quantity in the inventory adjustment plan is recorded item by item, and the sales volume forecast results are associated with each product category by time step. After the example marketing strategy parameters, such as promotional discounts, activity duration, and channel coverage, are labeled by product category, they are combined with the inventory adjustment plan and sales volume forecast results. Based on the adjustment effect of different strategy parameters on the sales volume forecast results, a sales change sequence corresponding to each strategy is generated, and the sequence is sorted according to strategy effect evaluation indicators such as sales increase and inventory turnover rate to form an optimized marketing plan.

[0087] It should also be noted that different strategy parameters refer to the quantitative representation of key controllable elements in sales activities, including promotional discounts, activity duration, channel coverage, and inventory adjustment coefficients. For example, promotional discounts represent the magnitude of price reductions, which can be set to 5% to 20%; activity duration represents the duration of the promotional activity, which can be set to 1 to 7 days; channel coverage represents the number or type of sales channels covered by the sales activity, which can be set to 3 online channels and 2 offline channels; and inventory adjustment coefficients represent the proportion of inventory adjustment in the replenishment plan, which can be set to a range of 0.8 to 1.2. Each strategy parameter can be adjusted independently or in combination, and its impact on sales volume forecasts is quantified to evaluate the actual effectiveness of different marketing strategies. The strategy effectiveness evaluation index is formed by calculating quantitative indicators such as sales increase and inventory turnover rate based on historical sales data and inventory data, and is used to measure the effectiveness of marketing strategy execution.

[0088] Based on the inventory adjustment plan and marketing optimization plan, optimize the allocation of resources for inventory, sales channels and marketing resources, and generate management execution results.

[0089] Specifically, for the parameters of the inventory adjustment plan in the inventory adjustment plan and the parameters of the optimization marketing plan in the optimization marketing plan, the sales forecast, current inventory level, and available marketing resources of each sales channel are extracted. The sales channels are then sorted based on factors including the level of sales forecast, the degree of inventory shortage, and the availability of marketing resources. The sorting results are used as a reference for allocation priority. Based on the allocation priority, and according to the inventory capacity and total resource limits, the inventory quantity in the inventory adjustment plan is allocated to each sales channel to ensure that high-demand sales channels receive more inventory, while low-demand sales channels maintain the minimum inventory and resource retention value. The available marketing resources in the optimization marketing plan are then sequentially allocated to each sales channel to form the inventory allocation and marketing resource allocation for each sales channel. At the same time, a corresponding sales channel execution plan is generated. All allocations and execution plans are arranged in time series to form an inventory, sales channel, and marketing resource allocation matrix, generating the management execution results.

[0090] It should also be noted that the determination of high-demand and low-demand sales channels is based on historical sales data and projected sales volume. The actual or projected sales volume within a certain time period will be sorted or classified. Channels with higher sales volume or projected sales volume exceeding the reference value are identified as high-demand channels, while channels with lower sales volume or projected sales volume below the reference value are identified as low-demand channels. The reference value is a value set based on the historical average sales volume or projected sales volume, used to distinguish between high-demand and low-demand channels. For example, the median or mean of the sales volume distribution can be used as the reference value. Minimum inventory refers to the minimum inventory level that a sales channel maintains at any given time to ensure basic sales demand and supply continuity; resource reserve value refers to the minimum available quantity reserved when allocating marketing resources to each sales channel to ensure the continuous execution of marketing activities and to respond to sudden sales fluctuations.

[0091] In summary, this invention employs a channel attention method to weight the initial fused feature sequence over time steps, combines this with local frequency domain analysis to extract periodic patterns, and then generates the final sales data temporal correlation features through gated fusion, projection dimensionality reduction, and normalization. This achieves differentiated processing and feature enhancement for short-term fluctuations and long-term trends. It automatically highlights important signals in key time periods while suppressing noise, improving the effectiveness and interpretability of feature representation. This allows subsequent multimodal fusion modules to more accurately integrate sales data and textual trend information in a unified feature space, achieving high accuracy and stability in sales volume prediction results.

[0092] The above embodiments have provided a detailed description of the present invention, but the content described is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the present invention. All equivalent variations and improvements made within the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. An intelligent sales forecasting system based on multimodal data fusion, characterized in that: include, The data acquisition module is used to collect channel sales records and external text data, and preprocess them to generate a sales data input set; The feature extraction module is used to perform one-dimensional convolution calculation on the sales data input set using the first and second convolutional layers to extract short-term and long-term time-series features, and then fuse them using the residual superposition method to generate time-series correlation features of sales data. The semantic extraction module is used to classify, segment, and encode contextual semantics of channel sales records and external text data to generate trend report feature vectors. The feature fusion module is used to perform non-linear fusion encoding on the time-series correlation features of sales data and the feature vectors of trend reports to generate a decoded feature matrix; The decoding and regression module is used to perform decoding and regression on the decoding feature matrix using regression decoding methods to generate sales volume prediction results. The decision execution module is used to optimize inventory, adjust marketing strategies, and allocate resources based on sales volume forecasts, generating management execution results.

2. The intelligent sales forecasting system based on multimodal data fusion as described in claim 1, characterized in that: The channel sales records include online order data, offline store sales data, and distributor purchase records; The external text data includes industry policy information, holiday information, and user review text.

3. The intelligent sales forecasting system based on multimodal data fusion as described in claim 1, characterized in that, The process of generating the sales data input set includes the following steps: Entity parsing, deduplication and alignment, and timestamp standardization are performed on channel sales records and external text data to generate a time-aligned multimodal dataset; Semantic parsing and temporal completion are performed on the time-aligned multimodal dataset to extract text trend features. At the same time, missing data completion and anomaly labeling are performed to generate a multimodal augmented dataset. The multimodal augmentation dataset is processed using a causal feature enhancement method to obtain a multimodal association feature set. The data is then organized using an efficient index storage method to generate a sales data input set.

4. The intelligent sales forecasting system based on multimodal data fusion as described in claim 1, characterized in that, The extraction of short-term and long-term time-series features includes the following steps: The sales data input set is sliced ​​according to time series, and then input into the first convolutional layer for one-dimensional convolution calculation. Short-term local change patterns within the window are extracted to generate short-term time series features. Short-term temporal features are input into the second convolutional layer, and one-dimensional convolution is performed using an expanded convolutional kernel to capture long-term dependencies across time periods and generate long-term temporal features.

5. The intelligent sales forecasting system based on multimodal data fusion as described in claim 1, characterized in that, The generation of time-series correlation features for sales data includes the following steps: Short-term time-series features are processed by channel mapping and batch normalization, and long-term time-series features are projected and mapped to the same dimension and then residually superimposed with short-term time-series features to generate a preliminary fused feature sequence. Using the channel attention method, the preliminary fused feature sequence is weighted by time steps to obtain the weighted fused feature sequence, and local frequency domain analysis is performed to generate the frequency-time fused feature sequence; Gated fusion, projection dimensionality reduction and normalization are performed on the frequency-time fusion feature sequences to generate time-series correlation features of sales data.

6. The intelligent sales forecasting system based on multimodal data fusion as described in claim 1, characterized in that, The generation of the trend report feature vector includes the following steps: Perform thematic classification and tagging of channel sales records and external text data to generate a set of categorized and labeled texts; The classified and labeled text set is segmented into words, and each segmented word is mapped into a high-dimensional word vector sequence while maintaining the original context order, thus generating a preliminary text feature sequence. Contextual semantic encoding is performed on the preliminary text feature sequence to obtain the contextual semantic representation result; Global semantic modeling and contextual information integration are performed on high-dimensional word vector sequences to generate global semantic features; The contextual semantic representation results are fused with global semantic features to generate a trend report feature vector.

7. The intelligent sales forecasting system based on multimodal data fusion as described in claim 1, characterized in that, The generation of the decoding feature matrix includes the following steps: The time-series correlation features of sales data and the feature vectors of trend reports are mapped to a unified feature space to generate a mapped feature set. By using a weighted nonlinear mapping method, the fusion ratio of the time-series correlation features of sales data and the feature vectors of trend reports in the mapped feature set is adjusted to generate a weighted feature set. The weighted feature set is integrated to capture the non-linear relationship between the time-series correlation features of sales data and the feature vector of trend report, and a decoded feature matrix is ​​generated.

8. The intelligent sales forecasting system based on multimodal data fusion as described in claim 1, characterized in that, The process of generating sales volume forecast results includes the following steps: The time-series correlation features of sales data and the feature vector of trend report in the decoded feature matrix are weighted and decoded to generate a preliminary prediction vector; The initial prediction vector is corrected for errors and trends. Short-term fluctuations and long-term trends are captured through residual calculation and feedback processing to generate an optimized prediction vector. The optimized forecast vector is mapped to obtain a sales trend representation, and time series smoothing and confidence interval calculation are performed to generate sales volume forecast results.

9. The intelligent sales forecasting system based on multimodal data fusion as described in claim 8, characterized in that, The process of error correction and trend adjustment of the initial prediction vector, capturing short-term fluctuations and long-term trends through residual calculation and feedback processing, and generating an optimized prediction vector includes the following steps: Historical error analysis is performed on the preliminary forecast vector to calculate the residual sequence between the forecast value and the actual sales value, and to identify short-term fluctuations and long-term trends. Multi-scale trend decomposition is performed on the residual sequence to separate short-term fluctuations from long-term trends and weighted adjustments are made to generate a trend correction vector. The residual feedback integration method is adopted to integrate the trend correction vector with residual feedback, balance the influence of short-term fluctuations and long-term trends, and generate an optimized prediction vector.

10. The intelligent sales forecasting system based on multimodal data fusion as described in claim 1, characterized in that: The process of generating management execution results includes the following steps: Based on sales volume forecasts, analyze inventory demand, assess safety stock levels and replenishment plans, and generate inventory adjustment schemes. By combining inventory adjustment plans with sales forecasts, the impact of different marketing strategies on sales can be quantified, and optimized marketing plans can be generated. Based on the inventory adjustment plan and marketing optimization plan, optimize the allocation of resources for inventory, sales channels and marketing resources, and generate management execution results.

Images