A method and system for time-series demand sensing in the poultry industry supply chain
By constructing a cross-modal common knowledge base and using an attention mechanism to fuse features, the problem of insufficient structured modeling of cross-scenario common knowledge in poultry sales forecasting is solved, achieving more accurate and interpretable sales forecasting and supporting intelligent decision-making and refined management of the poultry supply chain.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- WENS FOODSTUFF GROUP CO LTD
- Filing Date
- 2026-03-12
- Publication Date
- 2026-05-26
AI Technical Summary
Poultry sales forecasting faces challenges in complex market environments with multiple varieties, regions, and customer types. These challenges include insufficient cross-scenario common knowledge structure modeling, insufficient fusion of multimodal semantic information, simplistic and rigid integration mechanisms for individual and common aspects, and weak model interpretability, leading to difficulties in business implementation.
By acquiring historical time-series sales data and semantic attribute data, encoding and aligning them to a unified semantic space, a cross-modal common knowledge base is constructed. Attention mechanisms are used to fuse dominant features and reference features to generate a future time-series demand perception sequence.
It improves the accuracy and interpretability of poultry sales forecasts, enhances intelligent decision-making capabilities in complex market environments, supports refined management of the poultry supply chain, reduces operating costs, and improves market responsiveness.
Smart Images
Figure CN121836791B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of time-series data analysis technology, and more specifically, to a method and system for time-series demand perception in the poultry industry supply chain. Background Technology
[0002] The poultry farming industry is developing towards large-scale and intensive operations, and accurate sales forecasting plays a crucial role in enterprises' production scheduling, inventory management, market layout, and business decisions. However, poultry sales are characterized by complexity, involving multiple breeds, regions, customer types, and organizational levels. Sales are influenced by a combination of factors, including breed attributes, market environment, seasonal cycles, regional consumption habits, holiday effects, and customer preferences, exhibiting high nonlinearity and dynamic volatility. Furthermore, when launching new breeds, exploring new markets, or responding to sudden market changes, enterprises often face a "cold start" problem due to insufficient historical data, while a wealth of accumulated sales experience across breeds, regions, and time periods has not been systematically refined and reused.
[0003] Currently, there are various technical methods in the field of poultry sales forecasting, mainly including traditional time series models (such as ARIMA), multiple regression analysis, and deep learning-based sequence forecasting methods (such as LSTM, Transformer, etc.). These methods can capture the trend, seasonality, and periodicity of sales sequences to a certain extent, but they still have the following limitations: (1) lack of structured modeling and reuse mechanisms for common knowledge across scenarios; (2) insufficient fusion of multimodal semantic information; (3) simple and rigid integration mechanisms for individual and commonalities; (4) weak model interpretability and difficulty in business implementation. Therefore, there is an urgent need for a time series forecasting method that can systematically integrate historical commonalities, fuse multimodal semantic information, and possess a dynamic individual-commonality fusion mechanism. Summary of the Invention
[0004] To address the aforementioned technical issues, this invention proposes a time-series demand perception method and system for the poultry industry supply chain, which improves the accuracy, generalization ability, and interpretability of poultry sales forecasting, and supports intelligent decision-making and refined management of the poultry industry supply chain in a complex and ever-changing market environment.
[0005] The first aspect of this invention provides a time-series demand sensing method for the poultry industry supply chain, comprising the following steps:
[0006] Historical time-series sales data and corresponding semantic attribute data are acquired. The time-series sales data and semantic attribute data are encoded respectively. Through comparative learning, the two types of encoded vectors are aligned to a unified semantic space and clustered to form a set of common patterns with semantic labels, and a cross-modal common knowledge base is constructed.
[0007] The current temporal data and current semantic attributes of the object to be perceived are obtained and mapped to the unified semantic space to obtain the current joint vector;
[0008] Based on the semantic similarity between the current joint vector and each common pattern, an initial screening is performed. Then, based on the dynamic time warping algorithm, the sequence similarity between the current time series data and the historical sequences corresponding to each pattern in the initial candidate pattern set is calculated to select at least one reference common pattern.
[0009] The vector encoded from the current time series data is used as the dominant feature, and the vector corresponding to the reference common pattern is used as the reference feature.
[0010] The dominant and reference features are fused through an attention mechanism, and the enhanced fused features are adaptively weighted and output. The enhanced fused features are then input into the decoder to generate a time-series demand perception sequence for the future target period.
[0011] In this solution, historical time-series sales data and corresponding semantic attribute data are obtained, and the time-series sales data and semantic attribute data are encoded respectively, including:
[0012] Historical time-series sales data and associated semantic attribute data are obtained from the poultry industry supply chain system. The time-series sales data includes a sequence of sales values arranged in chronological order, and the semantic attribute data includes product attributes, regional attributes, customer attributes, and time semantic tags. The time-series sales data and semantic attribute data are paired and aligned according to timestamps.
[0013] The time-series sales data is divided into a fixed-length sliding window sequence. A time-series encoder is used to encode each window sequence. The time-series encoder is a recurrent neural network, a time-series convolutional network, or a Transformer encoding structure. The trend, periodicity, and volatility features in the time-series sales data are extracted and output as a time-series feature vector.
[0014] The categorical attributes in the semantic attribute data are one-hot encoded, the tag attributes are multi-tag binary representation, and the descriptive text is text encoded to obtain a preliminary semantic vector.
[0015] The initial semantic vector is fused to obtain a semantic feature vector, and the temporal feature vector and the semantic feature vector are used as the encoded representation.
[0016] In this scheme, two types of encoding vectors are aligned to a unified semantic space and clustered through contrastive learning to form a set of common patterns with semantic labels, thus constructing a cross-modal common knowledge base, including:
[0017] Positive sample pairs are constructed by combining temporal feature vectors and semantic feature vectors from the same sales scenario, while negative sample pairs are constructed by combining temporal vectors and semantic vectors from different sales scenarios.
[0018] The temporal feature vector and semantic feature vector are mapped to the same high-dimensional semantic space by a projection network. The projection network is trained using a contrastive learning objective function based on InfoNCE loss, so that the distance between positive sample pairs in the high-dimensional semantic space approaches and the distance between negative sample pairs moves away.
[0019] The temporal and semantic feature vectors of all historical samples are mapped to the high-dimensional semantic space through a trained projection network to generate a joint alignment vector for each sample.
[0020] Clustering algorithms are used to cluster the joint alignment vector to obtain several clusters. The cluster center vector of each cluster is calculated as a common pattern vector. The semantic attributes common to the samples in each cluster are analyzed to extract structured semantic tags, and the semantic tags are associated and bound with the common pattern vector.
[0021] The common pattern vectors, semantic tags, and the most representative historical time-series samples selected from each cluster are collectively encapsulated into knowledge entries. An index structure is built based on the semantic tags, and all knowledge entries are constructed into a cross-modal common knowledge base.
[0022] In this scheme, the current temporal data and current semantic attributes of the object to be perceived are obtained and mapped to the unified semantic space to obtain the current joint vector. Preliminary screening is then performed based on the semantic similarity between the current joint vector and each common pattern, including:
[0023] Obtain the current time-series sales data and associated current semantic attribute data of the object to be perceived within the most recent preset continuous time period, encode the current time-series sales data and the current semantic attribute data, and generate the current time-series feature vector and the current semantic feature vector;
[0024] The current temporal feature vector and the current semantic feature vector are input into the trained projection network and mapped to the unified semantic space to generate the current joint vector.
[0025] Extract all common pattern vectors from the cross-modal common knowledge base, calculate the semantic similarity between the current joint vector and each common pattern vector, and based on the semantic similarity, filter out common patterns with similarity higher than a preset threshold to form an initial candidate pattern set;
[0026] Each candidate pattern in the initial screening candidate pattern set is accompanied by a common pattern vector, semantic label, and historical time sequence samples.
[0027] In this scheme, based on the dynamic time warping algorithm, the sequence similarity between the current time series data and the historical sequences corresponding to each pattern in the initially screened candidate pattern set is calculated, and at least one reference common pattern is selected, including:
[0028] For all candidate patterns in the initial candidate pattern set obtained from the initial screening, extract representative historical time series samples associated with the patterns;
[0029] The dynamic time warping algorithm is used to calculate the optimal alignment path and cumulative distance between the current time series data and the historical reference sequence of the candidate pattern, and the cumulative distance is converted into a sequence similarity score.
[0030] All candidate patterns are sorted based on the sequence similarity score, and at least one reference common pattern is selected from the sorted candidate patterns according to the preset screening rules.
[0031] Output the reference common pattern set, each reference common pattern includes its corresponding common pattern vector, semantic label, historical time series sequence sample and sequence similarity score.
[0032] In this scheme, the dominant feature and reference feature are fused through an attention mechanism, and the fused feature is adaptively weighted and output to enhance the fused feature, including:
[0033] The dominant feature is used as the query vector, and the reference feature is used as the key vector and value vector, and then input into the attention fusion module.
[0034] The correlation between the query vector and each key vector is calculated, and the attention weights corresponding to each reference feature are obtained after normalization.
[0035] The value vector is weighted and summed according to the attention weights to generate an aggregated common reference representation;
[0036] Dynamic fusion weight coefficients are generated using learnable dynamic fusion units based on the dominant features and the common reference representation.
[0037] The dominant feature and the common reference representation are weighted and summed using the dynamic fusion weight coefficients to generate an enhanced feature vector.
[0038] In this scheme, the enhanced fusion features are input into the decoder to generate a time-series demand-aware sequence for the future target period, including:
[0039] The enhanced fusion features are used as the initial context input of the decoder, and the decoder iteratively generates the demand perception value at each time point in the future target period in an autoregressive manner based on the enhanced fusion features.
[0040] The decoder outputs a preliminary demand perception sequence and performs a rationality calibration on the preliminary demand perception sequence. After calibration, it outputs a final temporal demand perception sequence and associates the semantic tags and similarity information of the reference common patterns on which it is based, as an interpretable basis for the perception results.
[0041] The second aspect of the present invention provides a time-series demand perception system for the poultry industry supply chain. The system includes: a multi-source data acquisition and preprocessing module, a multimodal encoding and alignment module, a common knowledge base module, a retrieval module, an adaptive feature fusion module, and a time-series decoding and perception generation module.
[0042] The multi-source data acquisition and preprocessing module automatically collects historical time-series sales data and corresponding semantic attribute data from each link of the poultry industry supply chain, and performs data cleaning, alignment, standardization and structuring operations to convert heterogeneous data into paired sample data.
[0043] The multimodal encoding and alignment module encodes time-series sales data and semantic attribute data respectively, and aligns the two types of encoding vectors to a unified semantic space through comparative learning;
[0044] The common knowledge base module performs clustering in a unified semantic space to form a set of common patterns with semantic labels, and constructs a cross-modal common knowledge base. The cross-modal common knowledge base supports an incremental update mechanism, which incorporates new data periodically or as needed.
[0045] The retrieval module obtains the current time-series data and current semantic attributes of the object to be perceived and maps them to the unified semantic space to obtain the current joint vector; it performs initial screening based on the semantic similarity between the current joint vector and each common pattern, and then calculates the sequence similarity between the current time-series data and the historical sequences corresponding to each pattern in the initial candidate pattern set based on the dynamic time warping algorithm, and selects at least one reference common pattern.
[0046] The adaptive feature fusion module uses the vector encoded from the current time-series data as the dominant feature and the vector corresponding to the reference common pattern as the reference feature; it fuses the dominant feature and the reference feature through an attention mechanism and introduces a learnable dynamic fusion coefficient to adaptively weight and output an enhanced fusion feature.
[0047] The time-series decoding and perception generation module inputs the enhanced fusion features into the decoder to generate a time-series demand perception sequence for the future target time period.
[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0049] This invention forms a reusable knowledge base by mining and storing common sales patterns across product types and regions. When faced with scenarios with few samples or drastic fluctuations, the system can automatically retrieve similar historical patterns and adaptively fuse them with the personalized information of the current object, thereby achieving a more accurate and stable perception of demand trends, and especially enhancing the generalization ability in new scenarios.
[0050] This invention, while outputting demand perception sequences, explicitly provides historical common patterns and their semantic tags, significantly increasing business personnel's trust in the model results. This facilitates decision-makers' judgment based on experience, driving the intelligent system from an auxiliary tool to a trusted decision-making partner. The demand perception results provide reliable decision-making basis for all links in the poultry industry supply chain. By improving the efficiency of supply and demand matching across the entire chain, it helps reduce operating costs, minimize resource waste, enhance market responsiveness, and ultimately promote the evolution of the entire supply chain towards data-driven and collaborative intelligence. Attached Figure Description
[0051] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.
[0052] Figure 1 A flowchart of a time-series demand perception method for the poultry industry supply chain according to the present invention is shown;
[0053] Figure 2 A flowchart for constructing a cross-modal common knowledge base is shown;
[0054] Figure 3 A flowchart for filtering common reference patterns is shown;
[0055] Figure 4 A block diagram of a time-series demand sensing system for the poultry industry supply chain according to the present invention is shown. Detailed Implementation
[0056] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.
[0057] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.
[0058] like Figure 1 As shown, the first embodiment of the present invention provides a time-series demand sensing method for the poultry industry supply chain, including:
[0059] Historical time-series sales data and corresponding semantic attribute data are acquired. The time-series sales data and semantic attribute data are encoded respectively. Through comparative learning, the two types of encoded vectors are aligned to a unified semantic space and clustered to form a set of common patterns with semantic labels, and a cross-modal common knowledge base is constructed.
[0060] The current temporal data and current semantic attributes of the object to be perceived are obtained and mapped to the unified semantic space to obtain the current joint vector;
[0061] Based on the semantic similarity between the current joint vector and each common pattern, an initial screening is performed. Then, based on the dynamic time warping algorithm, the sequence similarity between the current time series data and the historical sequences corresponding to each pattern in the initial candidate pattern set is calculated to select at least one reference common pattern.
[0062] The vector encoded from the current time series data is used as the dominant feature, and the vector corresponding to the reference common pattern is used as the reference feature.
[0063] The dominant and reference features are fused through an attention mechanism, and the enhanced fused features are adaptively weighted and output. The enhanced fused features are then input into the decoder to generate a time-series demand perception sequence for the future target period.
[0064] It should be noted that historical time-series sales data and associated semantic attribute data are obtained from the poultry industry supply chain system. The time-series sales data includes a sequence of sales values arranged in chronological order, and the semantic attribute data includes product attributes such as breed name, regional attributes such as sales area, customer attributes such as customer type, and time semantic tags such as holiday tags. The time-series sales data and semantic attribute data are paired and aligned according to timestamps, missing values are imputed, and outliers are smoothed or removed to ensure the integrity and consistency of the data.
[0065] The time-series sales data is divided into a fixed-length sliding window sequence, with each window serving as an independent time-series sample. A time-series encoder, such as a recurrent neural network, temporal convolutional network, or Transformer encoding structure, is used to encode each window sequence, extracting trend, periodic, and volatility features from the time-series sales data and outputting them as time-series feature vectors. Categorical attributes such as variety and region in the semantic attribute data are one-hot encoded, label attributes such as holidays are represented using multi-label binarization, and descriptive text such as weather event descriptions is text-encoded using methods like BERT to obtain preliminary semantic vectors. These preliminary semantic vectors are then fused, and a pre-trained language model or lightweight text encoder is used to perform deep encoding on the fused semantic features to obtain semantic feature vectors. The time-series feature vectors and semantic feature vectors are then used as aligned, encoded sample representations.
[0066] Preferably, when acquiring historical time-series sales data and corresponding semantic attribute data, and encoding the time-series sales data and semantic attribute data respectively, a relationship-aware graph attention network is constructed to generate time-series features enhanced with spatiotemporal context. A dynamic multi-relationship graph is constructed, where nodes represent sales entities, supply entities, or market environment entities in the poultry industry supply chain, and edges represent spatial proximity relationships, category association relationships, or supply chain flow relationships between entities. For each node, its historical time-series sales data and semantic attribute data are embedded to form the initial node features. The multi-relationship graph attention network is used to aggregate the neighbor information of each node and dynamically update the node representation to capture cross-entity collaborative effects. The updated sales entity node features are used as entity-enhanced time-series sales features for subsequent encoding and common knowledge base construction.
[0067] The sales nodes correspond to combinations of regions, varieties, and customer types; supply nodes correspond to production units such as farms and breeding farms; and environmental nodes correspond to weather monitoring points and raw material price indices. Spatial proximity edges are based on geographical distance; category association edges are constructed based on the similarity of consumption scenarios; supply chain flow edges are based on the upstream and downstream relationships determined by the supply chain; and temporal influence edges are based on historical influence relationships determined by Granger causality tests or lagged correlation analysis. The initial features of each node at time t are concatenated from recent temporal feature vectors and semantic attribute feature vectors, and information aggregation is performed using a multi-relationship graph attention network. For sales nodes as the objects to be perceived, the updated features are encoded as temporal sales features enhanced with spatiotemporal context. The features not only include the node's own historical information but also incorporate collaborative or inhibitory signals from related supply nodes and market environment nodes, more comprehensively reflecting the true state in the supply chain network.
[0068] It should be noted that, as Figure 2As shown, temporal feature vectors and semantic feature vectors from the same sales scenario are constructed as positive sample pairs, while temporal vectors and semantic vectors from different sales scenarios are constructed as negative sample pairs. Negative samples exhibit significant semantic differences from positive samples at the regional, product, and seasonal levels. A projection network maps the temporal feature vectors and semantic feature vectors to the same high-dimensional semantic space. In this high-dimensional semantic space, a contrastive learning objective function based on InfoNCE loss is used to train the projection network, causing positive sample pairs to converge in distance and negative sample pairs to converge in distance within the high-dimensional semantic space.
[0069] After comparative learning training is completed, the temporal and semantic feature vectors of all historical samples are mapped to the high-dimensional semantic space through the trained projection network to generate a joint alignment vector for each sample. Clustering algorithms such as K-means, DBSCAN, or hierarchical clustering are used to cluster the joint alignment vector, obtaining several clusters. The cluster center vector of each cluster is calculated as a common pattern vector. Structured semantic labels are extracted from the semantic attributes shared by samples within each cluster, and these semantic labels are associated with the common pattern vector. The common pattern vector, semantic label, and the most representative historical temporal sequence samples selected from each cluster are collectively encapsulated as knowledge entries. An index structure is built based on these semantic labels, constructing a cross-modal common knowledge base. A knowledge base update cycle or triggering condition is set. When new data enters, after encoding and alignment, the distance to the existing cluster centers is calculated to determine whether it is assigned to an existing pattern or forms a new pattern, thus achieving continuous evolution of the knowledge base.
[0070] It should be noted that the current time-series sales data and associated current semantic attribute data of the object to be sensed within the most recent preset continuous time period are obtained from the poultry industry supply chain system. This includes product variety, sales region, customer type, current holidays or promotional tags, etc., ensuring consistency with historical data definitions. Using the same time-series encoder and text encoder trained during knowledge base construction, the current time-series sales data and the current semantic attribute data are encoded to generate a current time-series feature vector and a current semantic feature vector. These vectors are then input into a trained projection network to map onto the unified semantic space, generating a current joint vector.
[0071] All common pattern vectors are extracted from the cross-modal common knowledge base. The semantic similarity between the current joint vector and each common pattern vector is calculated. Based on the semantic similarity, common patterns with similarity higher than a preset threshold are selected to form an initial candidate pattern set. The similarity measure can be a vector space metric such as cosine similarity, the inverse of Euclidean distance, or the inner product. This ensures that the retrieval candidates are consistent with the object to be perceived in the business context. Each candidate pattern in the initial candidate pattern set is accompanied by a common pattern vector, semantic label, and historical time-series samples, which narrows the calculation range of exact matching, saves computational resources, and ensures contextual consistency.
[0072] It should be noted that, as Figure 3 As shown, for all candidate patterns in the initial candidate pattern set obtained from the initial screening, representative historical time series samples associated with the patterns are extracted. These representative historical time series samples are sequences that actually occurred in history for the corresponding patterns and best reflect common trends. The current time series sales data is standardized in length or aligned with a sliding window with each historical time series sample to ensure comparability across time spans.
[0073] The Dynamic Time Warping (VTW) algorithm is used to calculate the optimal alignment path and cumulative distance between the current time series data and the historical reference sequence of the candidate pattern. The cumulative distance is then converted into a sequence similarity score. The VTW algorithm constructs a cost matrix to find the alignment path that best matches the shape of the two sequences, allowing for non-linear scaling and alignment of sequences along the time axis. The VTW algorithm effectively captures the similarity of common phase shifts, rhythm differences, and local fluctuation patterns in sales sequences. The specific calculation of the VTW algorithm is as follows:
[0074] Given the current time series data Historical reference sequence of candidate patterns Construct the cost matrix of the dynamic time warping algorithm :
[0075] ,
[0076] Filter the alignment path that minimizes the cumulative distance. , represented as
[0077] ,
[0078] in, This is the index of a time point in the current time series data. For time point indices in the historical reference sequence, For the current time series data in the 1st... The values at each time point For the historical reference sequence in the 1st The values at each time point Position in the cost matrix The cumulative minimum distance at the location.
[0079] Based on the sequence similarity scores, all candidate patterns are sorted. According to preset filtering rules, at least one reference common pattern is selected from the sorted candidate patterns. The set of reference common patterns is output, where each reference common pattern includes its corresponding common pattern vector, semantic label, historical time-series sequence sample, and sequence similarity score. The common patterns in the reference common pattern set are not only consistent with the current scenario in semantic context, but also highly similar to the current data in specific sales fluctuation patterns and trend directions.
[0080] It should be noted that the current time-series sales data is encoded to obtain the current time-series feature vector. The current time-series feature vector is used as the dominant feature, representing the unique historical sales patterns and trend information of the object to be perceived. From the set of reference common patterns, the common pattern vector corresponding to each reference pattern is extracted. The common pattern vector is used as the reference feature, representing historical common sales knowledge that is similar to the current scenario in both semantics and form.
[0081] Using the dominant feature as the query vector and the reference features as the key and value vectors, the input is given to the attention fusion module. The correlation between the query vector and each key vector is calculated, and after normalization, the attention weights corresponding to each reference feature are obtained. The value vectors are weighted and summed according to the attention weights to generate an aggregated common reference representation. A learnable dynamic fusion unit generates dynamic fusion weight coefficients based on the dominant feature and the common reference representation. For example, a small neural network or gating unit is used to analyze the confidence of the dominant feature or the correlation of the reference representation, outputting fusion weight coefficients between 0 and 1. The dynamic fusion weight coefficients are used to weight and sum the dominant feature and the common reference representation to generate an enhanced feature vector. The weighting formula is: Enhanced Feature = (1-α)*Dominant Feature + α*Common Reference Representation, where α is the dynamic fusion weight coefficient. When the object to be perceived has sufficient historical data and obvious patterns, the dynamic fusion weight coefficient is small, and the dominant feature is mainly relied upon. When in a cold start, data sparse, or highly fluctuating scenario, the dynamic fusion weight coefficient increases, and historical common knowledge is used for enhancement.
[0082] Preferably, while fusing the dominant feature and reference feature through an attention mechanism and adaptively weighting the output to enhance the fused feature, structured events such as holidays and promotional calendars, and unstructured events such as industry news and weather reports are extracted from historical data, and the event information is aligned with historical time-series sales data. During the attention mechanism fusion process, an event attention head is added to calculate the attention weight of the current dominant feature on historical events. Based on the event attention weight, key historical events with weights exceeding a preset threshold are identified. Based on the semantic tags and similarity information of the key historical events and the reference common patterns, a natural language description or visual attribution report is generated to explain the time-series demand perception sequence. In the event attention head, the same dominant feature is used as the query vector, and all historical event vectors are used as the value vector and key vector. The attention weight of the dominant feature on each historical event is calculated, and events with attention weights exceeding a preset threshold are selected as key impact events for this perception. Based on the retrieved reference common patterns and their similarities, as well as the identified key historical events and their attention weights, a predefined explanation template is filled to generate a natural language explanation report.
[0083] It should be noted that the enhanced fusion features are used as the initial context input of the decoder. Preferably, the decoder is based on a recurrent neural network, a temporal convolutional network, or a Transformer decoder. The corresponding parameters have been trained together with the encoder and fusion module in the early stage, and it can understand the trend, periodicity, and common pattern information contained in the enhanced fusion features. Based on the enhanced fusion features, the decoder iteratively generates the demand perception value at each time point in the future target period in an autoregressive manner. During the generation process, the decoder uses the enhanced feature vector as a global context to ensure that each step of the prediction is constrained by the original fusion information, and maintains the consistency of the prediction sequence with the input background in terms of semantics and trend.
[0084] The decoder outputs a preliminary demand perception sequence and performs a rationality calibration on it. After calibration, it outputs a final time-series demand perception sequence, along with the semantic tags and similarity information of the reference common patterns on which it is based, serving as the interpretable basis for the perception results. The time-series demand perception sequence and its interpretability annotations are provided to the poultry industry supply chain management system to support subsequent decision-making processes such as production planning, inventory allocation, and logistics arrangements.
[0085] In the process of perceiving demand over time, historical time series determine the main trend of the forecast and serve as the primary basis for the model to infer future sales changes. The retrieved common reference set is fine-tuned and supplemented with the main trend through an attention mechanism, providing the model with regular references such as holiday peaks, seasonal fluctuations, or regional differences. At the same time, textual information not only provides the model with background context constraints, making the forecast reasonable at the levels of region, category, and customer type, but also serves as an explanatory basis to enhance comprehensibility. Through this design, the forecasting process does not sacrifice the authenticity and specificity of individual samples, while fully absorbing the statistical regularities in common patterns. This makes the final forecast results more stable, accurate, and more valuable for business interpretation in scenarios with sparse data, cold starts, or seasonal fluctuations.
[0086] like Figure 4 As shown, the second aspect of the present invention provides a time-series demand perception system for the poultry industry supply chain. The system includes: a multi-source data acquisition and preprocessing module, a multimodal encoding and alignment module, a common knowledge base module, a retrieval module, an adaptive feature fusion module, and a time-series decoding and perception generation module.
[0087] The multi-source data acquisition and preprocessing module automatically collects historical time-series sales data and corresponding semantic attribute data from each link of the poultry industry supply chain, and performs data cleaning, alignment, standardization and structuring operations to convert heterogeneous data into paired sample data.
[0088] The multimodal encoding and alignment module encodes time-series sales data and semantic attribute data respectively, and aligns the two types of encoding vectors to a unified semantic space through comparative learning;
[0089] The common knowledge base module performs clustering in a unified semantic space to form a set of common patterns with semantic labels, and constructs a cross-modal common knowledge base. The cross-modal common knowledge base supports an incremental update mechanism, which incorporates new data periodically or as needed.
[0090] The retrieval module obtains the current time-series data and current semantic attributes of the object to be perceived and maps them to the unified semantic space to obtain the current joint vector; it performs initial screening based on the semantic similarity between the current joint vector and each common pattern, and then calculates the sequence similarity between the current time-series data and the historical sequences corresponding to each pattern in the initial candidate pattern set based on the dynamic time warping algorithm, and selects at least one reference common pattern.
[0091] The adaptive feature fusion module uses the vector encoded from the current time-series data as the dominant feature and the vector corresponding to the reference common pattern as the reference feature; it fuses the dominant feature and the reference feature through an attention mechanism and introduces a learnable dynamic fusion coefficient to adaptively weight and output an enhanced fusion feature.
[0092] The time-series decoding and perception generation module inputs the enhanced fusion features into the decoder to generate a time-series demand perception sequence for the future target time period.
[0093] A third embodiment of the present invention provides a computer-readable storage medium, which includes a time-series demand sensing method program for the poultry industry supply chain. When the time-series demand sensing method program for the poultry industry supply chain is executed by a processor, it implements the steps of a time-series demand sensing method for the poultry industry supply chain.
[0094] In the several embodiments provided in this application, it should be understood that the disclosed methods can be implemented in other ways. The embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another system, or some features may be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed may be through some interfaces, indirect coupling or communication connection of devices or units, and may be electrical, mechanical, or other forms.
[0095] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0096] Alternatively, if the integrated units of the present invention are implemented as software functional units and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.
[0097] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for sensing time-series demand in the poultry industry supply chain, characterized in that, Includes the following steps: Historical time-series sales data and corresponding semantic attribute data are acquired. The time-series sales data and semantic attribute data are encoded respectively. Through comparative learning, the two types of encoded vectors are aligned to a unified semantic space and clustered to form a set of common patterns with semantic labels, and a cross-modal common knowledge base is constructed. Obtain the current temporal data and current semantic attributes of the object to be perceived, and map them to the unified semantic space to obtain the current joint vector; Based on the semantic similarity between the current joint vector and each common pattern, an initial screening is performed. Then, based on the dynamic time warping algorithm, the sequence similarity between the current time series data and the historical sequences corresponding to each pattern in the initial candidate pattern set is calculated to select at least one reference common pattern. The vector encoded from the current time series data is used as the dominant feature, and the vector corresponding to the reference common pattern is used as the reference feature. The dominant feature and reference feature are fused through an attention mechanism, and the enhanced fused feature is output with adaptive weighting. The enhanced fused feature is then input into the decoder to generate a time-series demand perception sequence for the future target period. Based on the dynamic time warping algorithm, the sequence similarity between the current time series data and the historical sequences corresponding to each pattern in the initially screened candidate pattern set is calculated, and at least one reference common pattern is selected, including: For all candidate patterns in the initial candidate pattern set obtained from the initial screening, extract representative historical time series samples associated with the patterns; The dynamic time warping algorithm is used to calculate the optimal alignment path and cumulative distance between the current time series data and the historical reference sequence of the candidate pattern, and the cumulative distance is converted into a sequence similarity score. All candidate patterns are sorted based on the sequence similarity score, and at least one reference common pattern is selected from the sorted candidate patterns according to the preset screening rules. Output the reference common pattern set, each reference common pattern includes its corresponding common pattern vector, semantic label, historical time series sequence sample and sequence similarity score.
2. The time-series demand sensing method for the poultry industry supply chain according to claim 1, characterized in that, Obtain historical time-series sales data and corresponding semantic attribute data, and encode the time-series sales data and semantic attribute data respectively, including: Historical time-series sales data and associated semantic attribute data are obtained from the poultry industry supply chain system. The time-series sales data includes a sequence of sales values arranged in chronological order, and the semantic attribute data includes product attributes, regional attributes, customer attributes, and time semantic tags. The time-series sales data and semantic attribute data are paired and aligned according to timestamps. The time-series sales data is divided into a fixed-length sliding window sequence. A time-series encoder is used to encode each window sequence. The time-series encoder is a recurrent neural network, a time-series convolutional network, or a Transformer encoding structure. The trend, periodicity, and volatility features in the time-series sales data are extracted and output as a time-series feature vector. The categorical attributes in the semantic attribute data are one-hot encoded, the tag attributes are multi-tag binary representation, and the descriptive text is text encoded to obtain a preliminary semantic vector. The initial semantic vector is fused to obtain a semantic feature vector, and the temporal feature vector and the semantic feature vector are used as the encoded representation.
3. The time-series demand sensing method for the poultry industry supply chain according to claim 1, characterized in that, By contrastive learning, two types of encoded vectors are aligned to a unified semantic space and clustered to form a set of common patterns with semantic labels, thus constructing a cross-modal common knowledge base, including: Positive sample pairs are constructed by combining temporal feature vectors and semantic feature vectors from the same sales scenario, while negative sample pairs are constructed by combining temporal vectors and semantic vectors from different sales scenarios. The temporal feature vector and semantic feature vector are mapped to the same high-dimensional semantic space by a projection network. The projection network is trained using a contrastive learning objective function based on InfoNCE loss, so that the distance between positive sample pairs in the high-dimensional semantic space approaches and the distance between negative sample pairs moves away. The temporal and semantic feature vectors of all historical samples are mapped to the high-dimensional semantic space through a trained projection network to generate a joint alignment vector for each sample. Clustering algorithms are used to cluster the joint alignment vector to obtain several clusters. The cluster center vector of each cluster is calculated as a common pattern vector. The semantic attributes common to the samples in each cluster are analyzed to extract structured semantic tags, and the semantic tags are associated and bound with the common pattern vector. The common pattern vectors, semantic tags, and the most representative historical time-series samples selected from each cluster are collectively encapsulated into knowledge entries. An index structure is built based on the semantic tags, and all knowledge entries are constructed into a cross-modal common knowledge base.
4. The time-series demand sensing method for the poultry industry supply chain according to claim 1, characterized in that, The current temporal data and current semantic attributes of the object to be perceived are obtained and mapped to the unified semantic space to obtain the current joint vector. Preliminary screening is then performed based on the semantic similarity between the current joint vector and each common pattern, including: Obtain the current time-series sales data and associated current semantic attribute data of the object to be perceived within the most recent preset continuous time period, encode the current time-series sales data and the current semantic attribute data, and generate the current time-series feature vector and the current semantic feature vector; The current temporal feature vector and the current semantic feature vector are input into the trained projection network and mapped to the unified semantic space to generate the current joint vector. Extract all common pattern vectors from the cross-modal common knowledge base, calculate the semantic similarity between the current joint vector and each common pattern vector, and based on the semantic similarity, filter out common patterns with similarity higher than a preset threshold to form an initial candidate pattern set; Each candidate pattern in the initial screening candidate pattern set is accompanied by a common pattern vector, semantic label, and historical time sequence samples.
5. A time-series demand sensing method for the poultry industry supply chain according to claim 1, characterized in that, The dominant and reference features are fused through an attention mechanism, and the fused features are adaptively weighted and enhanced, including: The dominant feature is used as the query vector, and the reference feature is used as the key vector and value vector, and then input into the attention fusion module. The correlation between the query vector and each key vector is calculated, and the attention weights corresponding to each reference feature are obtained after normalization. The value vector is weighted and summed according to the attention weights to generate an aggregated common reference representation; Dynamic fusion weight coefficients are generated using learnable dynamic fusion units based on the dominant features and the common reference representation. The dominant feature and the common reference representation are weighted and summed using the dynamic fusion weight coefficients to generate an enhanced feature vector.
6. The time-series demand sensing method for the poultry industry supply chain according to claim 1, characterized in that, The enhanced fusion features are input into the decoder to generate a time-series demand-aware sequence for the future target time period, including: The enhanced fusion features are used as the initial context input of the decoder, and the decoder iteratively generates the demand perception value at each time point in the future target period in an autoregressive manner based on the enhanced fusion features. The decoder outputs a preliminary demand perception sequence and performs a rationality calibration on the preliminary demand perception sequence. After calibration, it outputs a final temporal demand perception sequence and associates the semantic tags and similarity information of the reference common patterns on which it is based, as an interpretable basis for the perception results.
7. A time-series demand sensing system for the poultry industry supply chain, characterized in that, To implement a time-series demand perception method for poultry industry supply chain as described in any one of claims 1-6, the system includes: a multi-source data acquisition and preprocessing module, a multimodal encoding and alignment module, a common knowledge base module, a retrieval module, an adaptive feature fusion module, and a time-series decoding and perception generation module; The multi-source data acquisition and preprocessing module automatically collects historical time-series sales data and corresponding semantic attribute data from each link of the poultry industry supply chain, and performs data cleaning, alignment, standardization and structuring operations to convert heterogeneous data into paired sample data. The multimodal encoding and alignment module encodes time-series sales data and semantic attribute data respectively, and aligns the two types of encoding vectors to a unified semantic space through comparative learning; The common knowledge base module performs clustering in a unified semantic space to form a set of common patterns with semantic labels, and constructs a cross-modal common knowledge base. The cross-modal common knowledge base supports an incremental update mechanism, which incorporates new data periodically or as needed. The retrieval module obtains the current time-series data and current semantic attributes of the object to be perceived and maps them to the unified semantic space to obtain the current joint vector; it performs initial screening based on the semantic similarity between the current joint vector and each common pattern, and then calculates the sequence similarity between the current time-series data and the historical sequences corresponding to each pattern in the initial candidate pattern set based on the dynamic time warping algorithm, and selects at least one reference common pattern. The adaptive feature fusion module uses the vector encoded from the current time-series data as the dominant feature and the vector corresponding to the reference common pattern as the reference feature; it fuses the dominant feature and the reference feature through an attention mechanism and introduces a learnable dynamic fusion coefficient to adaptively weight and output an enhanced fusion feature. The time-series decoding and perception generation module inputs the enhanced fusion features into the decoder to generate a time-series demand perception sequence for the future target time period.