An emergency material demand prediction method based on deep learning
Patent Information
- Application Number
- CN202611066653.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-17
- Publication Date
- 2026-10-20
AI Technical Summary
[0003]传统的应急物资需求预测多依赖历史经验和简单的统计分析方法，随着应急事件复杂度的提升和数据量的增长，这类方法已难以满足现代应急管理的需求；
[0027]1、本发明能够整合多维度异构数据，全面提取与应急物资需求相关的特征信息，同时捕捉物资需求的时间依赖与空间依赖关系，实现多类型应急物资的同步预测，提升预测结果的准确性和全面性，采用自适应的特征融合方式，能够自动学习不同时间点和不同区域对物资需求的影响程度，减少人工特征设计带来的局限性，增强模型的泛化能力。
Smart Images
Figure FT_1 
Figure SMS_1 
Figure SMS_10
Abstract
Description
Technical Field
[0001] This invention relates to the field of emergency management technology, and in particular to a method for predicting emergency material demand based on deep learning. Background Technology
[0002] With the frequent occurrence of various public emergencies, the accurate allocation of emergency supplies has become one of the core aspects of emergency response work; the accuracy of emergency supply demand forecasting, as a prerequisite for supply allocation, directly affects the efficiency and effectiveness of emergency rescue.
[0003] Traditional emergency material demand forecasting relies heavily on historical experience and simple statistical analysis methods. However, with the increasing complexity of emergency events and the growth of data volume, these methods are no longer sufficient to meet the needs of modern emergency management.
[0004] In recent years, deep learning technology has been widely used in the field of time series forecasting, providing a new technical path for forecasting emergency material demand. However, existing deep learning-based forecasting methods still have many shortcomings, which limit their application effect in actual emergency scenarios. When conducting emergency material demand forecasting, existing technologies mostly use a single data source for analysis, which cannot effectively integrate multi-dimensional heterogeneous data such as emergency events, geospatial data, and meteorological and hydrological data, resulting in incomplete feature information. Summary of the Invention
[0005] The purpose of this invention is to provide a deep learning-based method for predicting emergency material demand, in order to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] A deep learning-based method for predicting emergency supplies demand includes the following steps:
[0008] S1. Multi-source heterogeneous data acquisition and preprocessing stage: acquire raw data from multiple sources, process missing and outlier values in the raw data, and unify the dimension and time granularity of different data.
[0009] S2, Multi-dimensional Feature Engineering Stage: Extract periodic time features, convert geographic location information into numerical features, quantify the type and intensity features of emergency events, extract the inherent attribute features of various emergency supplies, screen highly relevant features and reduce feature dimensionality.
[0010] S3, the core stage of spatiotemporal attention fusion prediction, captures the long-term temporal dependency relationship of material demand data, constructs regional association map and captures the spatial dependency relationship between different regions, automatically learns the influence weight of different time points and different regions on the current material demand, and outputs the demand prediction results of various emergency materials.
[0011] S4. In the model training and validation phase, the dataset is divided into training set, validation set and test set. The error between the model's predicted value and the true value is calculated, the model parameters are updated to minimize the loss function, and the generalization ability of the model is evaluated.
[0012] S5. Dynamic rolling prediction stage: Obtain real-time development information of emergency events, continuously update prediction results as events develop, and quantify the uncertainty of prediction results;
[0013] S6. In the forecast result output and feedback stage, the forecast results and the supply and demand of materials are displayed. The forecast results are adjusted in combination with expert experience, and the actual consumption data is fed back to the model for continuous optimization.
[0014] As a further improvement to this technical solution: the S1 multi-source heterogeneous data acquisition and preprocessing stage specifically includes a data acquisition submodule, a data cleaning submodule, and a data standardization submodule.
[0015] As a further improvement to this technical solution: the S2 multi-dimensional feature engineering stage specifically includes a time feature extraction submodule, a spatial feature extraction submodule, an event feature extraction submodule, a material feature extraction submodule, and a feature screening and dimensionality reduction submodule.
[0016] As a further improvement to this technical solution: the S3 spatiotemporal attention fusion prediction core stage specifically includes an LSTM temporal coding submodule, a GCN spatial coding submodule, a spatiotemporal attention fusion submodule, and a multi-task output submodule.
[0017] As a further improvement to this technical solution: the spatiotemporal attention fusion submodule first calculates the attention weight of each region at each time step, using the following formula:
[0018]
[0019] in Indicates the first The region in the first Attention weights for the current prediction at each time step. This represents the temporal feature vector at the current time step. This represents the spatial feature vector of the i-th region. This represents the learnable attention weight matrix. The dimension of the feature vector. Represents the normalization function;
[0020] Then, the temporal and spatial features are weighted and fused based on the calculated attention weights, using the following formula:
[0021]
[0022] in This represents the final feature vector after fusion. This represents the temporal feature matrix output by the LSTM module. Indicates the output of the GCN module. Spatial feature matrix of each region Indicates the total number of regions. ∑ represents the feature concatenation operation, and ∑ represents the weighted summation operation.
[0023] As a further improvement to this technical solution: the S4 model training and validation stage specifically includes a dataset partitioning submodule, a loss calculation submodule, an optimizer submodule, and a cross-validation submodule.
[0024] As a further improvement to this technical solution: the S5 dynamic rolling prediction stage specifically includes a real-time data access submodule, a rolling prediction execution submodule, and a confidence interval calculation submodule.
[0025] As a further improvement to this technical solution: the S6 prediction result output feedback stage specifically includes a result output submodule, a manual correction submodule, and a model feedback submodule.
[0026] Compared with the prior art, the beneficial effects of the present invention are:
[0027] 1. This invention can integrate multi-dimensional heterogeneous data, comprehensively extract feature information related to emergency material demand, and capture the time and space dependencies of material demand, so as to realize the synchronous prediction of multiple types of emergency materials, improve the accuracy and comprehensiveness of prediction results, and adopt an adaptive feature fusion method to automatically learn the degree of influence of different time points and different regions on material demand, reduce the limitations of manual feature design, and enhance the generalization ability of the model.
[0028] 2. This invention supports dynamic rolling prediction, which can adjust the prediction results in a timely manner according to the real-time development of emergency events, adapt to the dynamic evolution characteristics of emergency events, quantify the uncertainty of prediction results, provide a more comprehensive reference for emergency decision-making, establish a model feedback mechanism, which can continuously optimize model performance using actual material consumption data, and adjust the prediction results by combining expert experience, taking into account the advantages of model automation and human experience, and improving the practicality of the prediction method.
[0029] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it according to the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Specific embodiments of the present invention are given in detail below with reference to the accompanying drawings. Attached Figure Description
[0030] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:
[0031] Figure 1 This is a schematic diagram of the method structure of an emergency material demand prediction method based on deep learning. Detailed Implementation
[0032] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are for illustrative purposes only and are not intended to limit the scope of the invention. The invention is described more specifically in the following paragraphs by way of example with reference to the accompanying drawings. It should be noted that the drawings are in a very simplified form and use non-precise proportions, and are only used to facilitate and clarify the illustration of the embodiments of the present invention.
[0033] Please see Figure 1 In this embodiment of the invention, an emergency supplies demand prediction method based on deep learning includes the following steps:
[0034] S1. Multi-source heterogeneous data acquisition and preprocessing stage: acquire raw data from multiple sources, process missing and outlier values in the raw data, and unify the dimension and time granularity of different data.
[0035] S2, Multi-dimensional Feature Engineering Stage: Extract periodic time features, convert geographic location information into numerical features, quantify the type and intensity features of emergency events, extract the inherent attribute features of various emergency supplies, screen highly relevant features and reduce feature dimensionality.
[0036] S3, the core stage of spatiotemporal attention fusion prediction, captures the long-term temporal dependency relationship of material demand data, constructs regional association map and captures the spatial dependency relationship between different regions, automatically learns the influence weight of different time points and different regions on the current material demand, and outputs the demand prediction results of various emergency materials.
[0037] S4. In the model training and validation phase, the dataset is divided into training set, validation set and test set. The error between the model's predicted value and the true value is calculated, the model parameters are updated to minimize the loss function, and the generalization ability of the model is evaluated.
[0038] S5. Dynamic rolling prediction stage: Obtain real-time development information of emergency events, continuously update prediction results as events develop, and quantify the uncertainty of prediction results;
[0039] S6. Forecast result output and feedback stage: display the forecast results and the supply and demand of materials, adjust the forecast results based on expert experience, and feed back the actual consumption data to the model for continuous optimization.
[0040] Specifically, the S1 multi-source heterogeneous data acquisition and preprocessing stage completes the acquisition and basic processing of raw data, providing a unified and standardized data foundation for subsequent feature extraction;
[0041] In the S2 multi-dimensional feature engineering stage, effective features are extracted from four dimensions: time, space, event, and material. The feature dimensions are then filtered and reduced to improve the quality of the model input.
[0042] The S3 spatiotemporal attention fusion prediction core stage simultaneously captures the temporal and spatial dependencies of resource demand, achieves adaptive feature fusion through an attention mechanism, and outputs multi-resource prediction results.
[0043] In the S4 model training and validation phase, the dataset is divided and the model parameters are trained to evaluate the model's generalization ability.
[0044] In the S5 dynamic rolling prediction phase, prediction results are updated by combining real-time event information to quantify prediction uncertainty.
[0045] In the S6 prediction result output feedback stage, the results are displayed and manually adjusted, and the model is continuously optimized based on actual data feedback.
[0046] The S1 multi-source heterogeneous data acquisition and preprocessing stage specifically includes a data acquisition submodule, a data cleaning submodule, and a data standardization submodule.
[0047] Specifically, the data acquisition submodule is used to acquire historical emergency event data, historical material consumption data, geospatial data, and meteorological and hydrological data.
[0048] The data cleaning submodule is used to handle missing and outlier values in the raw data;
[0049] The data standardization submodule is used to unify the dimensions and time granularity of different data.
[0050] The S2 multi-dimensional feature engineering stage specifically includes a time feature extraction submodule, a spatial feature extraction submodule, an event feature extraction submodule, a material feature extraction submodule, and a feature screening and dimensionality reduction submodule.
[0051] Specifically, the time feature extraction submodule is used to extract periodic time features such as year, month, day, week, and season;
[0052] The spatial feature extraction submodule is used to convert geographic location information into numerical features and calculate the spatial distance and accessibility between regions;
[0053] The event feature extraction submodule is used to encode emergency event types and quantify the intensity and impact of events.
[0054] The material feature extraction submodule is used to extract the inherent attribute features of various emergency materials, such as shelf life, transportation difficulty, and storage requirements.
[0055] The feature filtering and dimensionality reduction submodule is used to filter features that are highly relevant to material demand and reduce the feature dimensionality.
[0056] The core stage of S3 spatiotemporal attention fusion prediction specifically includes the LSTM temporal coding submodule, the GCN spatial coding submodule, the spatiotemporal attention fusion submodule, and the multi-task output submodule;
[0057] Specifically, the LSTM time-series coding submodule is used to encode time-series features and capture the long-term temporal dependencies of material demand data.
[0058] The GCN spatial coding submodule is used to construct regional association maps based on geospatial data, encode spatial features, and capture spatial dependencies between different regions.
[0059] The spatiotemporal attention fusion submodule is used to calculate the attention weights at different time points and in different regions, so as to achieve adaptive fusion of temporal features and spatial features;
[0060] The multi-task output submodule is used to simultaneously output the demand forecast results for multiple emergency supplies.
[0061] The spatiotemporal attention fusion submodule first calculates the attention weight of each region at each time step, using the following formula:
[0062]
[0063] in Indicates the first The region in the first Attention weights for the current prediction at each time step. This represents the temporal feature vector at the current time step. This represents the spatial feature vector of the i-th region. This represents the learnable attention weight matrix. The dimension of the feature vector. Represents the normalization function;
[0064] Then, the temporal and spatial features are weighted and fused based on the calculated attention weights, using the following formula:
[0065]
[0066] in This represents the final feature vector after fusion. This represents the temporal feature matrix output by the LSTM module. Indicates the output of the GCN module. Spatial feature matrix of each region Indicates the total number of regions. ∑ represents the feature concatenation operation, and ∑ represents the weighted summation operation;
[0067] Specifically, first, the attention weight of each region at each time step is calculated using the following formula: ,in Indicates the first The region in the first Attention weights for the current prediction at each time step. This represents the temporal feature vector at the current time step. This represents the spatial feature vector of the i-th region. This represents the learnable attention weight matrix. The dimension of the feature vector. The formula represents the normalization function. By calculating the similarity between temporal and spatial features, it obtains the influence weight of each region at the corresponding time step, thereby achieving adaptive differentiation of the importance of different regions.
[0068] Then, the temporal and spatial features are weighted and fused based on the calculated attention weights, using the following formula: ,in This represents the final feature vector after fusion. This represents the temporal feature matrix output by the LSTM module. Indicates the output of the GCN module. Spatial feature matrix of each region Indicates the total number of regions. The formula represents the feature concatenation operation, and ∑ represents the weighted summation operation. The formula concatenates the weighted spatial features with the original temporal features to obtain a fused feature that contains both temporal and spatial information, which is then input into the subsequent multi-task output layer.
[0069] The S4 model training and validation phase specifically includes a dataset partitioning submodule, a loss calculation submodule, an optimizer submodule, and a cross-validation submodule.
[0070] Specifically, the dataset partitioning submodule is used to divide the preprocessed dataset into training set, validation set and test set according to the proportions.
[0071] The loss calculation submodule is used to calculate the error between the model's predicted values and the actual values;
[0072] The optimizer submodule is used to update the model parameters based on the gradient of the loss function and minimize the loss function;
[0073] The cross-validation submodule is used to evaluate the model's generalization ability by splitting the dataset multiple times.
[0074] The S5 dynamic rolling prediction stage specifically includes a real-time data access submodule, a rolling prediction execution submodule, and a confidence interval calculation submodule.
[0075] Specifically, the real-time data access submodule is used to obtain real-time development information of emergency events and the current status of material inventory;
[0076] The rolling prediction execution submodule is used to continuously update the input data and re-execute the prediction process as the event progresses.
[0077] The confidence interval calculation submodule is used to calculate the confidence interval of the prediction result and quantify the uncertainty of the prediction result.
[0078] The S6 prediction result output feedback stage specifically includes a result output submodule, a manual correction submodule, and a model feedback submodule.
[0079] Specifically, the results output submodule is used to visualize the prediction results and the supply and demand of materials;
[0080] The manual correction submodule is used to adjust the model's prediction results by incorporating expert experience;
[0081] The model feedback submodule is used to feed back actual material consumption data to the model for subsequent training and optimization.
[0082] The method of use and working principle of this invention are as follows:
[0083] Usage: First, collect raw data from multiple sources and preprocess it to handle missing and outlier values, and unify the data dimensionality and time granularity. Then, extract effective features from multiple dimensions such as time, space, event, and resources, filter highly relevant features and reduce feature dimensionality. Divide the processed dataset into training, validation, and test sets. Build a spatiotemporal attention fusion prediction model and train and validate it. Evaluate the model's generalization ability. When an emergency occurs, access real-time information on the event's development, perform dynamic rolling predictions using the trained model, quantify the uncertainty of the prediction results, display the prediction results and resource supply and demand status, adjust the prediction results based on expert experience, and feed back actual resource consumption data to the model to achieve continuous model optimization.
[0084] Working principle: The LSTM temporal coding submodule encodes temporal features to capture the long-term dependencies of material demand data over time. The GCN spatial coding submodule constructs a regional correlation map based on geospatial data and encodes spatial features to capture spatial dependencies between different regions. The spatiotemporal attention fusion submodule adaptively fuses temporal and spatial features by calculating attention weights at different time points and in different regions to obtain fused features that simultaneously contain both temporal and spatial information. The multi-task output submodule outputs multiple emergency material demand prediction results based on the fused features. During the prediction process, the real-time data access submodule obtains the latest information on emergency events and uses a rolling prediction mechanism to continuously update the input data and re-execute the prediction process. The confidence interval calculation submodule quantifies the uncertainty of the prediction results. Finally, the model feedback submodule uses actual consumption data for subsequent model training to continuously improve the model's prediction performance.
[0085] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Those skilled in the art can readily implement the present invention based on the description and drawings above. However, any modifications, alterations, and variations made by those skilled in the art without departing from the scope of the present invention using the disclosed technical content are equivalent embodiments of the present invention. Furthermore, any modifications, alterations, and variations made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the present invention.
Claims
1. A method for predicting emergency material demand based on deep learning, characterized in that, Includes the following steps: S1. Multi-source heterogeneous data acquisition and preprocessing stage: acquire raw data from multiple sources, process missing and outlier values in the raw data, and unify the dimension and time granularity of different data. S2, Multi-dimensional Feature Engineering Stage: Extract periodic time features, convert geographic location information into numerical features, quantify the type and intensity features of emergency events, extract the inherent attribute features of various emergency supplies, screen highly relevant features and reduce feature dimensionality. S3, the core stage of spatiotemporal attention fusion prediction, captures the long-term temporal dependency relationship of material demand data, constructs regional association map and captures the spatial dependency relationship between different regions, automatically learns the influence weight of different time points and different regions on the current material demand, and outputs the demand prediction results of various emergency materials. S4. In the model training and validation phase, the dataset is divided into training set, validation set and test set. The error between the model's predicted value and the true value is calculated, the model parameters are updated to minimize the loss function, and the generalization ability of the model is evaluated. S5. Dynamic rolling prediction stage: Obtain real-time development information of emergency events, continuously update prediction results as events develop, and quantify the uncertainty of prediction results; S6. In the forecast result output and feedback stage, the forecast results and the supply and demand of materials are displayed. The forecast results are adjusted in combination with expert experience, and the actual consumption data is fed back to the model for continuous optimization.
2. The emergency material demand forecasting method based on deep learning according to claim 1, characterized in that, The S1 multi-source heterogeneous data acquisition and preprocessing stage specifically includes a data acquisition submodule, a data cleaning submodule, and a data standardization submodule.
3. The emergency material demand forecasting method based on deep learning according to claim 1, characterized in that, The S2 multi-dimensional feature engineering stage specifically includes a time feature extraction submodule, a spatial feature extraction submodule, an event feature extraction submodule, a material feature extraction submodule, and a feature screening and dimensionality reduction submodule.
4. The emergency material demand forecasting method based on deep learning according to claim 1, characterized in that, The S3 spatiotemporal attention fusion prediction core stage specifically includes an LSTM temporal coding submodule, a GCN spatial coding submodule, a spatiotemporal attention fusion submodule, and a multi-task output submodule.
5. The emergency material demand forecasting method based on deep learning according to claim 4, characterized in that, The spatiotemporal attention fusion submodule first calculates the attention weight of each region at each time step, using the following formula: in Indicates the first The region in the first Attention weights for the current prediction at each time step. This represents the temporal feature vector at the current time step. This represents the spatial feature vector of the i-th region. This represents the learnable attention weight matrix. The dimension of the feature vector. Represents the normalization function; Then, the temporal and spatial features are weighted and fused based on the calculated attention weights, using the following formula: in This represents the final feature vector after fusion. This represents the temporal feature matrix output by the LSTM module. This indicates the output of the GCN module. Spatial feature matrix of each region Indicates the total number of regions. ∑ represents the feature concatenation operation, and ∑ represents the weighted summation operation.
6. The emergency material demand forecasting method based on deep learning according to claim 1, characterized in that, The S4 model training and validation phase specifically includes a dataset partitioning submodule, a loss calculation submodule, an optimizer submodule, and a cross-validation submodule.
7. The emergency material demand forecasting method based on deep learning according to claim 1, characterized in that, The S5 dynamic rolling prediction stage specifically includes a real-time data access submodule, a rolling prediction execution submodule, and a confidence interval calculation submodule.
8. The emergency material demand forecasting method based on deep learning according to claim 1, characterized in that, The S6 prediction result output feedback stage specifically includes a result output submodule, a manual correction submodule, and a model feedback submodule.