Regional market-oriented sand and gravel demand forecasting method, device, equipment and storage medium

Abstract

The application discloses a sand and gravel demand prediction method and device for a regional market, equipment and a storage medium, relates to the technical field of data prediction, and comprises the following steps: collecting historical engineering types, areas and sand and gravel consumption amounts of each construction stage in each month in a preset period of a target region, and calculating sand and gravel consumption coefficients of each engineering type in each stage; obtaining the number, type, area and construction stage of a current engineering in a target time dimension; inputting the number and type of the current engineering into a period demand prediction model to obtain a period total demand prediction value; inputting the type, area, construction stage of the current engineering and the sand and gravel consumption coefficient of each unit area into a stage demand prediction model to obtain a stage total demand prediction value; and performing weighted fusion on the period and stage demand prediction values to obtain a sand and gravel demand prediction result. The application can simultaneously consider the macro period law of regional sand and gravel demand and the micro sand consumption characteristics of engineering construction stages, and the accuracy of sand and gravel demand prediction is improved.

Description

Technical Field

[0001] This application relates to the field of data forecasting technology, and in particular to methods, apparatus, equipment and storage media for forecasting sand and gravel demand in regional markets. Background Technology

[0002] Currently, for individual projects, the sand consumption is calculated by adding up the building area and the empirical value of 0.35–0.66 m³ of concrete sand per square meter. Finally, the total amount for the region and the total amount for the project are manually adjusted to form the sales plan for the next month.

[0003] However, the project-level experience coefficient method is only effective within the red line area and lacks the linkage calculation of the number of projects under construction simultaneously, the pace of infrastructure construction, and the release of competing mine capacity. This leads to the deviation between the total regional sand consumption and the micro-project sand consumption being superimposed in the same direction, with prediction errors often exceeding ±30%, making it difficult to support daily dispatching at the 10,000-ton level. Therefore, how to simultaneously take into account the macro-cyclical patterns of regional sand and gravel demand and the micro-sand consumption characteristics of the engineering construction stage, and improve the accuracy of sand and gravel demand prediction, has become an urgent problem to be solved.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The purpose of this application is to provide a method, apparatus, equipment and storage medium for predicting sand and gravel demand in regional markets, aiming to solve the technical problem of how to simultaneously take into account the macro-cyclical patterns of regional sand and gravel demand and the micro-scale sand consumption characteristics during the engineering construction stage, thereby improving the accuracy of sand and gravel demand prediction.

[0006] To achieve the above objectives, this application proposes a method for forecasting sand and gravel demand in regional markets, the method comprising:

[0007] Obtain the historical project types, historical project areas, and sand and gravel consumption of historical projects at each construction stage for each month within a preset period in the target area;

[0008] Calculate the sand and gravel consumption coefficient per unit area for each type of project at each construction stage based on the sand and gravel consumption, the historical project area, and the historical project type.

[0009] Obtain the current number of projects, current project type, current project area, and current construction stage of the project within the target time dimension of the target area;

[0010] Input the current project quantity and the current project type into the periodic demand forecasting model to obtain the total periodic demand forecast;

[0011] Input the current project type, the current project area, the current project construction stage, and the sand and gravel consumption coefficient per unit project area corresponding to the current project type into the stage demand prediction model to obtain the total stage demand prediction value;

[0012] The total demand forecast value for the period and the total demand forecast value for the stage are weighted and fused according to preset weights to obtain the sand and gravel demand forecast result for the target area in the target time dimension.

[0013] In one embodiment, the periodic demand forecasting model includes an input layer, an engineering type weight embedding layer, an autoregressive module, a difference module, a moving average module, a seasonal autoregressive module, a sub-regional periodic dynamic adjustment layer, a seasonal difference module, a seasonal moving average module, and an output layer.

[0014] The step of inputting the current project quantity and the current project type into the periodic demand forecasting model to obtain the periodic total demand forecast includes:

[0015] The input layer decomposes the current project quantity into the current project quantity distribution of each sub-region, and converts the current project type into the current project type distribution of each sub-region.

[0016] By calling the type-weight correspondence through the project type weight embedding layer, the demand contribution weight of each type of project is calculated based on the current project type distribution. The current project quantity distribution is multiplied by the corresponding demand contribution weight to obtain the sub-region weighted feature data.

[0017] The weighted feature data of the sub-regions are sequentially passed to the autoregression module, the difference module, and the moving average module. After autoregression calculation, difference processing, and moving average smoothing, non-seasonal component processing data is obtained.

[0018] The non-seasonal component data is processed by the seasonal autoregressive module, the sub-region periodic dynamic adjustment layer, the seasonal difference module, the seasonal moving average module, and the output layer to obtain the periodic prediction sub-values ​​for each sub-region.

[0019] Based on the proportion of historical demand in each sub-region to the total demand in the target region, the periodic predicted sub-values ​​of each sub-region are multiplied by the corresponding proportions and then summed to obtain the periodic total demand prediction value.

[0020] In one embodiment, the step of processing the non-seasonal component data through the seasonal autoregressive module, the sub-region periodic dynamic adjustment layer, the seasonal differencing module, the seasonal moving average module, and the output layer to obtain the periodic prediction sub-values ​​of each sub-region includes:

[0021] The seasonal autoregression module performs seasonal autoregression calculations on the non-seasonal component data to obtain seasonal autoregression data.

[0022] The seasonal autoregressive data is corrected by calling the pre-configured seasonal cycle parameters for each sub-region through the sub-region periodic dynamic adjustment layer to obtain seasonally corrected data.

[0023] The seasonal difference module and the seasonal moving average module perform seasonal difference processing and seasonal moving average smoothing processing on the seasonally corrected data, and the output layer outputs the periodic prediction sub-values ​​of each sub-region.

[0024] In one embodiment, the stage demand prediction model includes an input layer, a feature embedding layer, a construction stage adaptation layer, an LSTM unit layer, a cross-type feature transfer layer, a fully connected layer, and an output layer.

[0025] The step of inputting the current project type, the current project area, the current project construction stage, and the unit project area sand and gravel consumption coefficient corresponding to the current project type into the stage demand forecasting model to obtain the stage total demand forecast includes:

[0026] The input layer and the feature embedding layer are used to preprocess and transform the current project type, the current project area, the current project construction stage, and the sand and gravel consumption coefficient per unit project area to obtain a fused feature vector.

[0027] By calling the phase-weight mapping matrix through the construction phase adaptation layer, dynamic weights are assigned to the numerical features in the fused feature vector according to the current construction phase of the project, resulting in a feature vector with phase weights.

[0028] The feature vector with stage weights is input into the LSTM unit layer. After the cell state is updated by the forget gate and the input gate, the hidden layer features are generated through the output gate.

[0029] The cross-type feature transfer layer calls the type similarity calculation module to calculate the process feature similarity between the current project type and the trained project type. If the process feature similarity is higher than the preset similarity threshold, the feature parameters of the corresponding trained type are extracted from the parameter sharing matrix, and the hidden layer features are corrected according to the feature parameters and the similarity ratio.

[0030] The modified hidden layer features are nonlinearly transformed by the fully connected layer, and a single stage demand prediction sub-value for the current project is output through the output layer.

[0031] The total stage demand forecast value is obtained by summing the stage demand forecast sub-values ​​of all current projects within the target area.

[0032] In one embodiment, the step of preprocessing and transforming the current project type, the current project area, the current project construction stage, and the sand and gravel consumption coefficient per unit project area through the input layer and the feature embedding layer to obtain a fused feature vector includes:

[0033] The input layer standardizes the current project type, the current project area, the current project construction stage, and the sand and gravel consumption coefficient per unit project area.

[0034] After the standardization process is completed, the current project type is converted into a type feature vector through the feature embedding layer;

[0035] Convert the current project area and the sand and gravel consumption coefficient per unit project area into a numerical feature vector;

[0036] The type feature vector and the numerical feature vector are merged to obtain a fused feature vector.

[0037] In one embodiment, the SARIMA model includes an input layer, an autoregressive module, a differencing module, a moving average module, a seasonal autoregressive module, a seasonal differencing module, a seasonal moving average module, and an output layer.

[0038] The training steps for the cyclical demand forecasting model include:

[0039] An improvement is made to the seasonal autoregressive module of the SARIMA model by adding a sub-regional periodic dynamic adjustment layer between the seasonal autoregressive module and the seasonal difference module. The sub-regional periodic dynamic adjustment layer contains a periodic parameter mapping table, which is used to dynamically match seasonal periodic parameters according to the input regional features.

[0040] The input layer of the SARIMA model is improved by adding an engineering type weight embedding layer between the input layer and the autoregressive module. The engineering type weight embedding layer contains a type-weight correspondence and is used to convert the proportion of engineering types into weight coefficients of input features.

[0041] The historical total demand for sand and gravel, the number of historical projects, and the proportion of historical project types for each sub-region within the target area are obtained for each month within the preset period to form hierarchical training data.

[0042] The hierarchical training data is input into the improved SARIMA model. The periodic parameter mapping table is called through the sub-region periodic dynamic adjustment layer to match the corresponding seasonal periodic parameters for each sub-region. The type-weight correspondence is called through the engineering type weight embedding layer to configure input weights for the proportion of engineering types in each sub-region.

[0043] Based on the historical total demand for sand and gravel, the SARIMA model with configured parameters is trained to minimize the error between the model output and the actual demand, thereby obtaining the basic prediction model for each sub-region.

[0044] The proportion of historical demand in each sub-region to the total demand in the target region is extracted from the hierarchical training data. The output results of the basic prediction models of each sub-region are weighted and integrated according to the proportion to obtain the periodic demand prediction model.

[0045] In one embodiment, the LSTM model includes an input layer, a feature embedding layer, an LSTM unit layer, a fully connected layer, and an output layer;

[0046] The training steps for the stage demand forecasting model include:

[0047] An improvement is made to the feature embedding layer of the LSTM model by adding a construction stage adaptation layer between the feature embedding layer and the LSTM unit layer. The construction stage adaptation layer contains a stage-weight mapping matrix, which is used to allocate dynamic weights according to the construction stage.

[0048] The fully connected layer of the LSTM model is improved by adding a cross-type feature transfer layer between the LSTM unit layer and the fully connected layer. The cross-type feature transfer layer includes a type similarity calculation module, which is used to share parameters of the hidden layer features of engineering types with similar processes.

[0049] Based on the unit project area sand and gravel consumption coefficient, the construction stage information of the historical project, the area of ​​the historical project, the sand and gravel consumption and process characteristic data, multiple types of training data are obtained;

[0050] The multi-type training data is input into the improved LSTM model, and the stage-weight mapping matrix is ​​called through the construction stage adaptation layer to configure corresponding weights for different construction stages.

[0051] The cross-type feature transfer layer calls the type similarity calculation module to calculate the process feature similarity between the current project type and the trained project type. If the process feature similarity is higher than the preset similarity threshold, the hidden layer feature parameters of the trained project type are extracted from the parameter sharing matrix, and the reuse weight of the hidden layer feature parameters is allocated according to the process feature similarity.

[0052] Based on the historical sand and gravel consumption in the multi-type training data, the LSTM model with the parameters configured is trained to minimize the error between the model output and the actual consumption, and a basic prediction model for each type of project is obtained.

[0053] By integrating the basic prediction models of the various project types, a phased demand prediction model is obtained.

[0054] Furthermore, to achieve the above objectives, this application also proposes a sand and gravel demand forecasting device for regional markets, the device comprising:

[0055] The historical data acquisition module is used to acquire the historical project types, historical project areas, and sand and gravel consumption of historical projects at each construction stage for each month within a preset period in the target area.

[0056] The coefficient calculation module is used to calculate the unit project area sand and gravel consumption coefficient for various project types at each construction stage based on the sand and gravel consumption, the historical project area, and the historical project type.

[0057] The current data acquisition module is used to acquire the current number of projects, current project type, current project area, and current project construction stage within the target time dimension of the target area;

[0058] The cycle prediction module is used to input the current project quantity and the current project type into the cycle demand prediction model to obtain the cycle total demand prediction value;

[0059] The phase prediction module is used to input the current project type, the current project area, the current project construction stage, and the sand and gravel consumption coefficient per unit area of ​​the current project type into the phase demand prediction model to obtain the total phase demand prediction value.

[0060] The result fusion module is used to perform weighted fusion of the predicted total demand for the period and the predicted total demand for the stage according to preset weights, so as to obtain the predicted sand and gravel demand of the target area in the target time dimension.

[0061] Furthermore, to achieve the above objectives, this application also proposes a sand and gravel demand forecasting device for regional markets, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the sand and gravel demand forecasting method for regional markets as described above.

[0062] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the sand and gravel demand forecasting method for regional markets as described above.

[0063] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the sand and gravel demand forecasting method for regional markets as described above.

[0064] One or more technical solutions proposed in this application have at least the following technical effects:

[0065] First, the historical project types, areas, and phased sand and gravel consumption for each month within a preset period are obtained. Based on this, the sand and gravel consumption coefficient per unit area for each project type at each construction stage is calculated, directly mapping scale to material usage and eliminating systematic biases caused by empirical coefficients. Second, the current project quantity, type, area, and construction stage within the target time dimension are captured in real time. The quantity and type proportions are fed into a cyclical demand forecasting model to obtain the macro total, used to lock in annual investment fluctuation trends. Then, the type, area, stage, and corresponding coefficient of each project are fed into a phased demand forecasting model to obtain the micro phase total, used to capture weekly construction pulses. The two are then merged according to preset weights to output the final forecast result, enabling mines to accurately schedule production in advance and reduce inventory risks. This application can simultaneously take into account the macro-cyclical patterns of regional sand and gravel demand and the micro-sand consumption characteristics of project construction stages, improving the accuracy of sand and gravel demand forecasting. Attached Figure Description

[0066] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0067] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0068] Figure 1 This is a flowchart illustrating an embodiment of the sand and gravel demand forecasting method for regional markets in this application.

[0069] Figure 2 This is a flowchart illustrating Embodiment 2 of the sand and gravel demand forecasting method for regional markets provided in this application;

[0070] Figure 3This is a schematic diagram of the module structure of the sand and gravel demand forecasting device for regional markets according to an embodiment of this application;

[0071] Figure 4 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the sand and gravel demand forecasting method for regional markets in this application embodiment.

[0072] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0073] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0074] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0075] It should be noted that the executing entity of this application embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of realizing the above functions, such as a sand and gravel demand forecasting system. The following uses a sand and gravel demand forecasting system as an example to describe this embodiment and the following embodiments.

[0076] Based on this, the embodiments of this application provide a method for forecasting sand and gravel demand in regional markets, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the sand and gravel demand forecasting method for regional markets in this application.

[0077] In this embodiment, the method for forecasting sand and gravel demand for regional markets includes steps S10 to S60:

[0078] Step S10: Obtain the historical project type, historical project area, and sand and gravel consumption of historical projects at each construction stage for each month within a preset period in the target area.

[0079] It should be noted that the target area refers to a geographical unit where sand and gravel demand forecasting is to be conducted. This can be a single city, metropolitan area, single province, or river basin, defined by administrative boundaries or economic ties to define the spatial scope of the forecast. The preset period refers to a unified backtesting time window used to train and validate the forecasting model, typically the past 5 years, but can be extended to 3-7 years depending on data completeness, ensuring coverage of at least one infrastructure investment cycle. Historical project types refer to completed or ongoing construction projects within the target area that can be categorized by structural characteristics, including multi-story residential buildings, high-rise office buildings, standard factory buildings, newly built roads, urban secondary roads, water plants, etc., used to differentiate the sand and gravel consumption intensity per unit area. Historical project area refers to the converted area of ​​the building area or road length recorded at the time of project initiation or planning approval for each historical project, measured in tens of thousands of square meters or kilometers, reflecting the project scale. The amount of sand and gravel consumed in each construction stage of a historical project refers to the actual amount of sand and gravel used in each stage after the project is divided into four stages: earthwork, foundation, main structure, and decoration. This amount is calculated using mix proportions or material lists and recorded in tens of thousands of tons or tens of thousands of cubic meters to depict the time pulse characteristics of demand.

[0080] Step S20: Calculate the unit project area sand and gravel consumption coefficient for each project type at each construction stage based on the sand and gravel consumption, the historical project area, and the historical project type.

[0081] It should be noted that the unit project area sand and gravel consumption coefficient refers to the average amount of sand and gravel used in a specific construction stage for a certain type of project, under each unit building area (or road area). It is used to quickly convert "project scale" into "stage sand and gravel demand".

[0082] Understandably, the sand and gravel demand forecasting system first groups sand and gravel by "project type + construction stage", sums the historical sand and gravel consumption fields to obtain the total amount for each stage, and sums the historical project area fields within the corresponding group to obtain the total area. Then, the total amount for each stage is divided by the total area to obtain the unit area consumption value for that group, and records with zero or missing area are removed before calculation to avoid division by zero errors. Finally, the results are written into a coefficient table for direct use in subsequent forecasts.

[0083] Step S30: Obtain the current number of projects, current project type, current project area, and current project construction stage within the target time dimension of the target area.

[0084] It should be noted that the target time dimension refers to the specific period targeted by the predicted demand, usually accurate to the month or week, used to define a "looking from the present to the future" time segment. The current number of projects refers to the total number of projects that have obtained construction permits, have actually started construction, and are not yet completed within this time dimension. The current project type refers to the design functional category corresponding to these projects, such as multi-story residential buildings, urban main roads, standard factory buildings, etc. The current project area refers to the sum of the building area or road length stated in the planning permits for the same batch of projects, reflecting the overall construction scale. The current construction stage refers to one of the four stages (earthwork, foundation, main structure, and decoration) that each project is currently in as of the target time dimension, used to determine the short-term aggregate consumption rhythm.

[0085] Step S40: Input the current project quantity and the current project type into the periodic demand forecasting model to obtain the total periodic demand forecast value.

[0086] It should be noted that the cyclical demand forecasting model refers to a statistical or machine learning model trained on historical time series data of "project type - area - stage - sand and gravel consumption" with monthly, quarterly, or annual steps. It is used to map "current project quantity + current project type" to the total regional sand and gravel demand over a future period. The total cyclical demand forecast is a single value or sequence output by this model, representing the total sand and gravel demand generated by all projects under construction within the target time period (e.g., next month, next quarter, or next year), typically expressed in tens of thousands of tons.

[0087] Understandably, the sand and gravel demand forecasting system first splits the current project quantity obtained from statistics within the target time dimension into a one-dimensional vector according to project type, and normalizes it to the same dimension range as the training samples. Then, it calls the pre-trained periodic demand forecasting model, feeds this vector into the network as an input sequence, and the model directly outputs the total regional sand and gravel demand for the next period by memorizing the area weights and consumption rhythms of various types of projects in the same historical period. Finally, the system denormalizes the dimensionless values ​​returned by the model to obtain the periodic total demand forecast value in tens of thousands of tons.

[0088] As an example, the SARIMA model includes an input layer, an autoregressive module, a differencing module, a moving average module, a seasonal autoregressive module, a seasonal differencing module, a seasonal moving average module, and an output layer. The training steps of the cyclical demand prediction model include: improving the seasonal autoregressive module of the SARIMA model by adding a sub-regional periodic dynamic adjustment layer between the seasonal autoregressive module and the seasonal differencing module. This sub-regional periodic dynamic adjustment layer contains a periodic parameter mapping table, used to dynamically match seasonal periodic parameters based on the input regional features; improving the input layer of the SARIMA model by adding an engineering type weight embedding layer between the input layer and the autoregressive module. This engineering type weight embedding layer contains a type-weight correspondence, used to convert the proportion of engineering types into weight coefficients of the input features; and obtaining the weights of each sub-region within the target region according to a preset... The historical total demand for sand and gravel, the number of historical projects, and the proportion of historical project types for each month within a given period are used to form hierarchical training data. This hierarchical training data is then input into the improved SARIMA model. The sub-regional periodic dynamic adjustment layer calls the periodic parameter mapping table to match corresponding seasonal periodic parameters for each sub-region. The project type weight embedding layer calls the type-weight correspondence to configure input weights for the proportion of project types in each sub-region. Based on the historical total demand for sand and gravel, the parameter-configured SARIMA model is trained to minimize the error between the model output and the actual demand, resulting in a basic prediction model for each sub-region. The proportion of historical demand in each sub-region to the total demand in the target region is extracted from the hierarchical training data. The outputs of the basic prediction models for each sub-region are then weighted and integrated according to this proportion to obtain a periodic demand prediction model.

[0089] The sub-region periodic dynamic adjustment layer refers to a callable network unit inserted between the seasonal autoregression and seasonal differencing in the SARIMA (Seasonal Autoregressive Integrated Moving Average Model). It automatically selects the period length for the corresponding sub-region based on the input tensor, achieving differentiated modeling of "one period per region." The periodic parameter mapping table is a lookup table using the sub-region ID as the key and the optimal seasonal order as the value. It is determined by grid search before training and persisted in memory for real-time querying by the dynamic adjustment layer. The seasonal periodic parameter is an integer describing the length of repeated demand fluctuations within the year; for example, 12 represents the annual period in monthly data, and 6 represents the semi-annual period. It is used to set the order of seasonal differencing and seasonal moving averages. The project type weight embedding layer is an embedding module placed between the input layer and the autoregressive module. It maps the project type proportion vector into a weight vector of the same dimension as the input sequence, achieving adaptive weighting of "type-demand intensity." The type-weight correspondence refers to a dictionary obtained by normalizing historical regression coefficients, recording the marginal contribution ratio of each type of project to the demand for sand and gravel. The embedding layer converts this proportion into weight coefficients. The project type proportion refers to the ratio vector of the area of ​​each type of project under construction in a sub-region to the total area of ​​that region in a single month, reflecting the combination of structural demand. The weight coefficients of the input features (input weights) are the element-wise multipliers output by the embedding layer, multiplied point-to-point with the original demand sequence, giving high sand-consuming types a larger model input amplitude. The historical total sand and gravel demand refers to the total amount of sand and gravel actually consumed by the sub-region each month within a preset period, in tons, used for supervised learning. The historical number of projects refers to the number of projects under construction in the sub-region each month during the same period, used to construct the sample size feature. The historical project type proportion refers to the ratio vector of the area of ​​each functional type of project to the total area in the above-mentioned months, used to generate weight embeddings. The hierarchical training data refers to a data table organized by a three-level index of "sub-region-month," containing three columns: total demand, number of projects, and project type proportion, ensuring clear sample hierarchy during training. Each sub-region's basic prediction model refers to an independent SARIMA instance trained and stored using the corresponding stratified data. Its seasonal order and type weights are localized and specific to that sub-region. The historical demand for each sub-region refers to the monthly total sand and gravel demand sequence for that sub-region in the stratified training data. The total demand for the target region is the sum of the historical demands of all sub-regions in the same month, used to calculate the sub-region demand weights and complete the final integration.

[0090] First, the sand and gravel demand forecasting system inserts a sub-regional periodic dynamic adjustment layer between the seasonal autoregressive and seasonal differencing methods. After reading the regional feature fields, it looks up a table to replace the fixed seasonal order with the sub-regional specific order and rewrites the model configuration in real time. Next, the system adds an engineering type weight embedding layer between the input layer and the autoregressive module. It converts the input engineering type proportion vector into weight coefficients according to a pre-stored correspondence and multiplies it point-by-point with the original sequence to complete the weighting. Then, the system captures the monthly historical total sand and gravel demand, engineering quantity, and engineering type proportion for each sub-region within a preset period, and merges them into hierarchical training data according to the sub-region-month index. The data is fed into the improved SARIMA. First, the periodic adjustment layer sets seasonal periodic parameters for each sub-region. Then, the embedding layer generates input weights for each sample, so that the model input carries both regional periodic differences and type intensity differences. Subsequently, using the historical total demand for sand and gravel as a label, the model parameters are iteratively optimized using maximum likelihood estimation until the prediction error is minimized, resulting in an independent basic prediction model for each sub-region. Finally, the system calculates the proportion of historical demand in each sub-region to the total demand in the target region. This proportion is used to weight and sum the monthly sub-region prediction values ​​output by each basic model. The combined overall sequence is the periodic demand prediction model.

[0091] Step S50: Input the current project type, the current project area, the current project construction stage, and the sand and gravel consumption coefficient per unit area corresponding to the current project type into the stage demand prediction model to obtain the stage total demand prediction value.

[0092] It should be noted that the phased demand forecasting model refers to a machine learning model trained using project type, construction phase, and unit area consumption coefficient as category features, and project area as numerical feature. This model can accurately output the phased demand for sand and gravel. The total phased demand forecast is the sum of sand and gravel usage for all ongoing projects within the target time frame, as given by the model during inference; the unit is typically 10,000 tons.

[0093] As an example, the LSTM model includes an input layer, a feature embedding layer, an LSTM unit layer, a fully connected layer, and an output layer. The training steps of the stage demand prediction model include: improving the feature embedding layer of the LSTM model by adding a construction stage adaptation layer between the feature embedding layer and the LSTM unit layer. The construction stage adaptation layer contains a stage-weight mapping matrix for dynamically assigning weights according to the construction stage. Improving the fully connected layer of the LSTM model by adding a cross-type feature transfer layer between the LSTM unit layer and the fully connected layer. The cross-type feature transfer layer contains a type similarity calculation module for parameter sharing of hidden layer features of similar engineering types. Based on the unit project area sand and gravel consumption coefficient, the construction stage information of historical projects, the area of ​​historical projects, the sand and gravel consumption, and process feature data, multiple types are obtained. Training data; inputting the multi-type training data into the improved LSTM model, calling the stage-weight mapping matrix through the construction stage adaptation layer to configure corresponding weights for different construction stages; calling the type similarity calculation module through the cross-type feature transfer layer to calculate the process feature similarity between the current project type and the trained project type; if the process feature similarity is higher than a preset similarity threshold, extracting the hidden layer feature parameters of the trained project type from the parameter sharing matrix, and allocating reuse weights of the hidden layer feature parameters according to the process feature similarity; training the LSTM model with configured parameters based on the historical sand and gravel consumption in the multi-type training data, minimizing the error between the model output and the actual consumption, and obtaining the basic prediction model for each project type; integrating the basic prediction models for each project type to obtain the stage demand prediction model.

[0094] The construction stage adaptation layer is a learnable subnetwork inserted between the feature embedding layer and the unit layer of the LSTM (Long Short-Term Memory Neural Network). It generates a set of weights based on the construction stage number, scaling the input features dimension-wise to prioritize stage-sensitive variables. The stage-weight mapping matrix is ​​a trainable parameter table of shape "number of stages × feature dimension," with each row corresponding to a construction stage, and the row vector serving as the weight template for that stage. Dynamic weights refer to the weights extracted from the mapping matrix by the construction stage adaptation layer based on the actual stage number of the current sample, changing in real-time as the sample changes. The cross-type feature transfer layer is located between the LSTM unit layer and the fully connected layer. It determines whether to reuse the implicit parameters of the trained type by calculating the process similarity between types, thus achieving knowledge transfer. The type similarity calculation module takes process feature vectors (water-cement ratio, number of layers, structural form, etc.) as input and outputs a cosine similarity score of 0–1 to measure the degree of process similarity between engineering types. Similar engineering types refer to existing types that differ from the current type to be predicted in less than a threshold in terms of concrete strength grade, structural system, and construction procedures, such as high-rise residential buildings and apartment hotels. Hidden layer features are fixed-length vectors output by LSTM unit layers, condensing the temporal information of the input sequence and used by subsequent layers as the basis for demand prediction. Multi-type training data is a time-series sample set composed of unit area coefficients, stages, areas, actual consumption, and technological features of several engineering types such as historical high-rise residential buildings, factories, and roads. The current engineering type is the functional category to which the project to be predicted belongs during the inference stage, such as "urban secondary arterial road." Trained engineering types are categories that have already completed LSTM training and whose parameters have been saved, such as "standard factory building 1." Technological feature similarity is the cosine value output by the type similarity calculation module; a higher value indicates that the two engineering types are more similar at the technological level. The preset similarity threshold is a manually set threshold, such as 0.85; values ​​exceeding this trigger parameter sharing. The parameter sharing matrix is ​​a "type pair × hidden dimension" tensor maintained by the cross-type transfer layer, storing the optimal hidden layer parameters for all trained types. The hidden layer feature parameters, namely the weights and biases of the LSTM unit layers, are considered reusable knowledge. The reuse weights are 0–1 coefficients dynamically assigned by the transfer layer based on the similarity of process features, determining the contribution ratio of shared parameters in the current model update. The basic prediction models for each engineering type refer to model files trained using the improved LSTM and saved separately for each engineering type, whose parameters have incorporated stage adaptation and cross-type transfer results.

[0095] First, after the feature embedding layer outputs, the system inserts a construction stage adaptation layer. Based on the stage number carried by the sample, it extracts the corresponding row vector from the stage-weight mapping matrix and performs dynamic weighting by multiplying it element-by-element with the embedded features. This allows the LSTM to prioritize the variables most sensitive to that stage (e.g., increasing the weight of "earthwork depth" in the foundation stage and increasing the weight of "number of layers" in the main body stage). The weighted features are then fed into the LSTM unit layer. Second, a cross-type feature transfer layer is inserted between the LSTM unit layer and the fully connected layer. The current type's process feature vector is input, and the type similarity calculation module compares it with the trained types using cosine distance. If the similarity is higher than a preset threshold, the hidden layer parameters of the corresponding type are extracted from the parameter sharing matrix and adjusted according to the similarity ratio. Weighted reuse reduces the risk of overfitting for small sample types (e.g., high-rise residential buildings and apartment hotels have a similarity of 0.88, so 88% of the hidden layer parameters of the former are directly reused), while allowing the network to fine-tune the remaining 12% through backpropagation. Then, the system concatenates the unit project area sand and gravel consumption coefficient, historical construction stage, area, actual consumption, and process characteristics into multi-type training data according to type, ensuring that each row of samples contains a complete time series, stage label, and process vector. After feeding the data into the improved LSTM, the construction stage adaptation layer dynamically configures the stage weights for each sample, and the cross-type transfer layer synchronously completes parameter reuse and gradient update. Using historical sand and gravel consumption as the label, the Adam (AdaptiveMoment) algorithm is used to perform the following steps: The Estimation (Adaptive Moment Estimation) optimizer minimizes the MAE (Mean Absolute Error) loss. Training stops early when the error on the validation set fails to decrease for three consecutive rounds. The weights specific to each type are saved to obtain the basic prediction model for each project type. Finally, all basic models are routed and merged according to the current project type identifier to form a unified stage demand prediction model. During inference, the corresponding weights are automatically called according to the input type and the total stage demand is output.

[0096] As an example, the stage demand prediction model includes an input layer, a feature embedding layer, a construction stage adaptation layer, an LSTM unit layer, a cross-type feature transfer layer, a fully connected layer, and an output layer. The step of inputting the current project type, the current project area, the current project construction stage, and the sand and gravel consumption coefficient per unit area corresponding to the current project type into the stage demand prediction model to obtain the predicted total demand value for the stage includes: preprocessing and feature transformation of the current project type, current project area, current project construction stage, and sand and gravel consumption coefficient per unit area through the input layer and the feature embedding layer to obtain a fused feature vector; and using the construction stage adaptation layer to call the stage-weight mapping matrix and assign dynamic weights to the numerical features in the fused feature vector according to the current project construction stage. The process involves: generating a feature vector with stage weights; inputting the stage-weighted feature vector into the LSTM unit layer, updating the cell state through a forget gate and an input gate, and generating hidden layer features through an output gate; calling the type similarity calculation module through the cross-type feature transfer layer to calculate the process feature similarity between the current project type and the trained project type; if the process feature similarity is higher than a preset similarity threshold, extracting the feature parameters of the corresponding trained type from the parameter sharing matrix, and correcting the hidden layer features according to the feature parameters and the similarity ratio; performing a nonlinear transformation on the corrected hidden layer features through the fully connected layer, and outputting the stage demand prediction sub-value of a single current project through the output layer; summing the stage demand prediction sub-values ​​of all current projects within the target area to obtain the total stage demand prediction value.

[0097] The fused feature vector is a fixed-dimensional floating-point vector formed by one-hot encoding, standardization, and alignment of the current project type, area, construction stage, and unit area consumption coefficient. It can be directly fed into subsequent networks. Numerical features are continuous elements in the fused feature vector representing comparable values ​​such as area and unit area consumption coefficient, distinct from the 0 / 1 bits of the category encoding. The feature vector with stage weights is the result of multiplying the fused feature vector element-wise with the corresponding weights extracted from the stage-weight mapping matrix after passing through the construction stage adaptation layer, amplifying key features and compressing secondary features. Hidden layer features are the hidden state vectors output by the LSTM unit layer after performing temporal modeling on the feature vector with stage weights, condensing project scale and stage information. The stage demand prediction sub-value for a single current project refers to the predicted sand / gravel usage for that stage, in tons, obtained by regression analysis of a specific project by the output layer, used for subsequent regional aggregation.

[0098] First, the sand and gravel demand prediction system feeds the current project type, current project area, current construction stage, and sand and gravel consumption coefficient per unit area into the input layer. The type and stage are one-hot encoded, concatenated with the area and coefficient, and then Z-score standardized to obtain a fused feature vector, ensuring all features are on the same scale. Second, the construction stage adaptation layer extracts the corresponding weight vector from the stage-weight mapping matrix based on the current construction stage number. This weight vector is then multiplied element-wise with the numerical features in the fused feature vector (e.g., the weight vector for the basic stage is 1.2 at the area position and 0.9 at the unit coefficient position), resulting in a feature vector with stage weights, allowing the network to focus on stage-sensitive variables. Finally, the feature vector with stage weights is input into the LSTM unit layer, where a forget gate discards the unweighted features from the previous time step. After the input gate updates the cell state, the output gate generates hidden layer features to capture the temporal evolution of the project scale. Subsequently, the cross-type feature transfer layer calls the type similarity calculation module to calculate the cosine similarity of the process features of the current project type and the trained types. If it is higher than a preset threshold (e.g., 0.85), the hidden layer feature parameters of the corresponding type are extracted from the parameter sharing matrix, and the current hidden layer features are corrected by weighting according to the similarity ratio to reduce the error caused by insufficient training of small sample types. Next, the fully connected layer performs a ReLU nonlinear transformation on the corrected hidden layer features, and the output layer gives the stage demand prediction sub-value of a single current project after linear activation. Finally, the system accumulates the prediction sub-values ​​of all current projects in the target area according to the project ID to obtain the stage total demand prediction value.

[0099] As an example, the step of preprocessing and transforming the current project type, current project area, current project construction stage, and sand and gravel consumption coefficient per unit project area through the input layer and the feature embedding layer to obtain a fused feature vector includes: standardizing the current project type, current project area, current project construction stage, and sand and gravel consumption coefficient per unit project area through the input layer; after the standardization is completed, converting the current project type into a type feature vector through the feature embedding layer; converting the current project area and the sand and gravel consumption coefficient per unit project area into numerical feature vectors; and merging the type feature vector and the numerical feature vector to obtain a fused feature vector.

[0100] Type feature vectors are low-dimensional dense vectors generated by one-hot encoding or embedding mapping of the current project type, used to characterize the semantic information of the project category. Numerical feature vectors are floating-point vectors obtained by concatenating two continuous variables, the current project area and the sand and gravel consumption coefficient per unit area, and then standardizing them, used to directly reflect the project scale and material intensity.

[0101] First, the sand and gravel demand prediction system uses an input layer to concatenate the current project type, current project area, current project construction stage, and sand and gravel consumption coefficient per unit area into an original vector. The area and coefficient are then Z-score standardized to achieve a mean of zero and a variance of one. Next, a feature embedding layer performs one-hot encoding on the current project type and uses a linear embedding matrix to reduce its dimensionality, resulting in a low-dimensional type feature vector. Simultaneously, the area and coefficient are directly concatenated into a numerical feature vector. Finally, the type feature vector and the numerical feature vector are concatenated column-wise to form a unified-dimensional fused feature vector for use by subsequent networks.

[0102] Step S60: The total demand forecast value of the period and the total demand forecast value of the stage are weighted and fused according to the preset weights to obtain the sand and gravel demand forecast result of the target area in the target time dimension.

[0103] It should be noted that the preset weights refer to fixed coefficients pre-set by the system, used to linearly weight and fuse the total demand forecast for the period and the total demand forecast for the stage in a proportional ratio (6:4 in this embodiment), ensuring that the long-term trend dominates and short-term fluctuations are appropriately corrected. The sand and gravel demand forecast result is a single value or sequence output after this weighted fusion, representing the final predicted total demand for sand and gravel in the target area within the target time dimension, in tens of thousands of tons, which can be directly used for mine production scheduling and logistics dispatching.

[0104] This embodiment provides a method for forecasting sand and gravel demand in a regional market. First, it obtains historical project types, areas, and phased sand and gravel consumption data for each month within a preset period. Based on this, it calculates the sand and gravel consumption coefficient per unit area for each project type at each construction stage, directly mapping scale to material usage and eliminating systematic biases caused by empirical coefficients. Second, it captures the current number, type, area, and construction stage of projects within the target time dimension in real time. The proportion of quantity and type is fed into a cyclical demand forecasting model to obtain a macro total, used to lock in annual investment fluctuation trends. Then, the type, area, stage, and corresponding coefficient of each project are fed into a phased demand forecasting model to obtain a micro phase total, used to capture weekly construction pulses. The two are then merged according to preset weights to output the final forecast result, enabling mines to accurately schedule production in advance and reduce inventory risk. This embodiment can simultaneously consider the macro-cyclical patterns of regional sand and gravel demand and the micro-sand consumption characteristics of project construction stages, improving the accuracy of sand and gravel demand forecasting.

[0105] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 2 , Figure 2This is a flowchart illustrating the second embodiment of the sand and gravel demand forecasting method for regional markets according to this application. The periodic demand forecasting model includes an input layer, a project type weight embedding layer, an autoregressive module, a difference module, a moving average module, a seasonal autoregressive module, a sub-regional periodic dynamic adjustment layer, a seasonal difference module, a seasonal moving average module, and an output layer. Step S40 of the sand and gravel demand forecasting method for regional markets includes steps S41 to S45:

[0106] Step S41: The current project quantity is decomposed into the current project quantity distribution of each sub-region through the input layer, and the current project type is converted into the current project type distribution of each sub-region.

[0107] It should be noted that the current project quantity distribution refers to the one-dimensional vector of "sub-region - number of projects" obtained by splitting the total number of projects under construction in the target area according to the granularity of sub-regions, which is used to predict the demand of each sub-region separately. The current project type distribution refers to the two-dimensional matrix of "sub-region - project type - area percentage" formed by further statistically analyzing the above projects by type, which is used to characterize the combination of structural demand within each sub-region.

[0108] Understandably, the sand and gravel demand forecasting system first scans the global current project list through the input layer, accumulates the project ID according to the sub-region coding field, and obtains the current project quantity distribution in the form of a "sub-region-project number" vector, avoiding the loss of spatial differences by directly feeding the total amount of the region into the model; then, it continues to accumulate the building area corresponding to each type within the same group and calculates the proportion, generating a "sub-region-type-area proportion" matrix as the current project type distribution, so that the model can subsequently perceive the local structural differences such as "more residential buildings or more factories".

[0109] Step S42: By calling the type-weight correspondence through the project type weight embedding layer, the demand contribution weight of each type of project is calculated based on the current project type distribution. The current project quantity distribution is multiplied by the corresponding demand contribution weight to obtain the sub-region weighted feature data.

[0110] It should be noted that the demand contribution weight refers to the row vector obtained by the project type weight embedding layer based on the "type-weight correspondence". Its value reflects the marginal contribution intensity of each type of project to the regional sand and gravel demand. The higher the area proportion and the larger the weight, the higher the contribution.

[0111] Subregion weighted feature data refers to the same-dimensional vector obtained by multiplying the current project quantity distribution with the aforementioned demand contribution weights element by element. It transforms the "number of projects" into "demand equivalents", so that subsequent model inputs can directly reflect the degree to which the internal structural differences of the subregion affect the consumption of sand and gravel.

[0112] Understandably, firstly, the project type weight embedding layer queries the type-weight correspondence according to the category number in the current project type distribution, and extracts the demand contribution weight vector for each type of project; then, it multiplies each element of the current project quantity distribution with the corresponding type's demand contribution weight element by element, so that the quantity of different types of projects such as residential buildings and factories is converted into demand equivalents; finally, the product results are combined into sub-region weighted feature data, which is used by the subsequent network to directly perceive the pulling strength of structural differences on sand and gravel consumption.

[0113] Step S43: The weighted feature data of the sub-region is sequentially passed to the autoregression module, the difference module, and the moving average module. After autoregression calculation, difference processing, and moving average smoothing, non-seasonal component processing data is obtained.

[0114] It should be noted that non-seasonal component processing data refers to the sub-regional weighted feature data that has passed through autoregression, differencing, and moving average modules in sequence, and then a smoothed sequence that removes seasonal fluctuations and retains only trends and random disturbances, which is used as an intermediate input before the subsequent seasonal module overlay.

[0115] Understandably, firstly, the sand and gravel demand forecasting system slices the weighted feature data of sub-regions into time windows and sends it to the autoregressive module. It then uses p-period data to fit linear coefficients and extrapolates the current value (p is a preset value; when p is 12, it uses data from the past 12 months to predict this month), capturing local demand inertia. Secondly, the differencing module performs first-order differencing on the fitting results to eliminate trends and make the sequence mean zero, avoiding distortion of seasonal parameters due to non-stationarity. Finally, the moving average module takes the moving average of the previous q periods to smooth out random fluctuations (q is a preset value) and outputs non-seasonal component processed data to reduce noise interference with subsequent seasonal overlays.

[0116] Step S44: The non-seasonal component processing data is processed by the seasonal autoregressive module, the sub-region periodic dynamic adjustment layer, the seasonal difference module, the seasonal moving average module, and the output layer to obtain the periodic prediction sub-values ​​of each sub-region.

[0117] It should be noted that the periodic forecast sub-value refers to the sand and gravel demand forecast component generated by the output layer for each sub-region after joint processing by the seasonal autoregression, sub-regional periodic dynamic adjustment, seasonal difference and seasonal moving average modules, which contains only seasonal fluctuation components.

[0118] As an example, the step of processing the non-seasonal component data through the seasonal autoregression module, the sub-region periodic dynamic adjustment layer, the seasonal difference module, the seasonal moving average module, and the output layer to obtain the periodic prediction sub-values ​​for each sub-region includes: performing seasonal autoregression calculations on the non-seasonal component data through the seasonal autoregression module to obtain seasonal autoregression data; calling the seasonal periodic parameters pre-configured for each sub-region through the sub-region periodic dynamic adjustment layer to perform seasonal periodic correction on the seasonal autoregression data to obtain seasonally corrected data; performing seasonal difference processing and seasonal moving average smoothing processing on the seasonally corrected data through the seasonal difference module and the seasonal moving average module, and outputting the periodic prediction sub-values ​​for each sub-region through the output layer.

[0119] Seasonal autoregressive data refers to the series obtained by linearly extrapolating the non-seasonal component data using the first p historical values ​​of the same period in the seasonal autoregressive module. It only contains fluctuation estimates at the annual cycle level.

[0120] The seasonal cycle parameter refers to the seasonal order and cycle length that the sub-regional cycle dynamic adjustment layer pre-stores for each sub-region, which is used to correct the global seasonal pattern to the local rhythm.

[0121] Seasonally corrected data refers to a new sequence obtained by multiplying or interpolating seasonal autoregressive data with sub-region-specific seasonal cycle parameters. It scales the "one-size-fits-all" seasonal waveform into a waveform that conforms to the seasonal patterns of infrastructure activity in the region, allowing subsequent seasonal differences and moving averages to continue smoothing.

[0122] First, the system aligns the non-seasonal component data according to the "same period of the year" (the system extracts data points from the same month or week of each year, arranges them in chronological order, and compares them with the corresponding month of the current year). It then fits linear coefficients using the first p data points from the same month and extrapolates them to obtain seasonal autoregressive data, preserving the annual cycle shape. Second, it calls the sub-regional periodic dynamic adjustment layer to scale the length and multiply the amplitude of the seasonal autoregressive data according to the pre-configured seasonal cycle parameters for that region (e.g., shortening the cycle length by 10% and multiplying the amplitude by 10% in the southern region), generating seasonally corrected data that fits the local seasonal rhythm and avoids distortion of the global seasonal waveform. Finally, the system performs first-order seasonal differencing on the seasonally corrected data to eliminate residual trends, then uses a q-period moving average to filter out random spikes. The output layer then obtains the cycle prediction sub-values ​​for each sub-region, providing a clean seasonal increment for subsequent superposition with the trend component.

[0123] Step S45: Based on the proportion of historical demand in each sub-region to the total demand in the target region, multiply the periodic predicted sub-value of each sub-region by the corresponding proportion and sum them to obtain the periodic total demand prediction value.

[0124] Understandably, the system first retrieves the historical sand and gravel demand of each sub-region within a preset period, calculates the percentage of the cumulative demand of each sub-region to the total demand of the target region as a weighting coefficient, then multiplies the periodic predicted sub-value of each sub-region with the corresponding weighting coefficient element by element, and then sums the multipliers to obtain the fused periodic total demand prediction value, ensuring that the high-demand sub-regions contribute more and the low-demand sub-regions contribute less, thus achieving synchronous calibration of the total regional amount and the sub-region structure.

[0125] This embodiment first breaks down the current project quantity into the current project quantity distribution of each sub-region according to administrative or grid coding. Simultaneously, it statistically analyzes the project type and area to form the current project type distribution for each sub-region, making spatial differences readily apparent. Then, through the project type weight embedding layer, it calls the pre-stored type-weight correspondence to convert the type area ratio into a demand contribution weight, and multiplies it element-by-element with the current project quantity distribution to obtain the weighted feature data for each sub-region. This transforms the "number of projects" into "demand equivalent," avoiding the neglect of sand consumption intensity differences by only using quantity. Next, the weighted feature data is sequentially fed into the autoregressive module, the difference module, and the moving average module to complete trend fitting, stabilization, and noise filtering, respectively. The non-seasonal component data is processed to ensure that the subsequent seasonal module is not affected by trends and random fluctuations. Then, the seasonal autoregressive module extracts the same-year pattern, the sub-regional periodic dynamic adjustment layer corrects the waveform with local seasonal parameters, and the seasonal difference and seasonal moving average modules further smooth it. The output layer provides the periodic prediction sub-values ​​for each sub-region, realizing refined prediction of "one region, one season". Finally, the periodic prediction sub-values ​​are weighted and summed according to the historical demand ratio of each sub-region to obtain the predicted value of the total periodic demand of the target region. This makes the macro total quantity and micro spatial distribution accurately aligned at the same time, and can take into account both the macro-periodic pattern of regional sand and gravel demand and the micro-sand consumption characteristics of the engineering construction stage, thus improving the accuracy of sand and gravel demand prediction.

[0126] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the sand and gravel demand forecasting method for regional markets. Any simple modifications based on this technical concept are within the scope of protection of this application.

[0127] This application also provides a sand and gravel demand forecasting device for regional markets; please refer to... Figure 3 The sand and gravel demand forecasting device for regional markets includes:

[0128] The historical data acquisition module 10 is used to acquire the historical project types, historical project areas, and sand and gravel consumption of historical projects at each construction stage for each month within a preset period in the target area.

[0129] The coefficient calculation module 20 is used to calculate the unit area sand and gravel consumption coefficient of various project types at each construction stage based on the sand and gravel consumption, the historical project area and the historical project type.

[0130] The current data acquisition module 30 is used to acquire the current number of projects, current project type, current project area, and current project construction stage within the target time dimension of the target area;

[0131] The cycle prediction module 40 is used to input the current project quantity and the current project type into the cycle demand prediction model to obtain the cycle total demand prediction value;

[0132] The stage prediction module 50 is used to input the current project type, the current project area, the current project construction stage, and the sand and gravel consumption coefficient per unit project area corresponding to the current project type into the stage demand prediction model to obtain the total demand prediction value for the stage.

[0133] The result fusion module 60 is used to perform weighted fusion of the predicted total demand for the period and the predicted total demand for the stage according to preset weights, so as to obtain the predicted sand and gravel demand of the target area in the target time dimension.

[0134] The sand and gravel demand forecasting device for regional markets provided in this application employs the sand and gravel demand forecasting method for regional markets described in the above embodiments. This addresses the technical problem of simultaneously considering the macro-cyclical patterns of regional sand and gravel demand and the micro-level sand consumption characteristics during engineering construction, thereby improving the accuracy of sand and gravel demand forecasting. Compared with the prior art, the beneficial effects of the sand and gravel demand forecasting device for regional markets provided in this application are the same as those of the sand and gravel demand forecasting method for regional markets provided in the above embodiments. Furthermore, other technical features of the sand and gravel demand forecasting device for regional markets are the same as those disclosed in the methods of the above embodiments, and will not be elaborated upon here.

[0135] This application provides a sand and gravel demand forecasting device for regional markets. The sand and gravel demand forecasting device for regional markets includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the sand and gravel demand forecasting method for regional markets in the first embodiment described above.

[0136] The following is for reference. Figure 4This document illustrates a structural schematic diagram of a sand and gravel demand forecasting device suitable for implementing embodiments of this application. The sand and gravel demand forecasting device for regional markets in embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 4 The sand and gravel demand forecasting device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0137] like Figure 4 As shown, a sand and gravel demand forecasting device for a regional market may include a processing unit 1001 (e.g., a central processing unit, a graphics processor, etc.) that can perform various appropriate actions and processes according to a program stored in ROM (Read Only Memory) 1002 or a program loaded from storage device 1003 into RAM (Random Access Memory) 1004. RAM 1004 also stores various programs and data required for the operation of the sand and gravel demand forecasting device for the regional market. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via bus 1005. Input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, LCDs (Liquid Crystal Displays), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the regional market sand and gravel demand forecasting equipment to exchange data with other devices wirelessly or via wired communication. Although the figure shows a regional market sand and gravel demand forecasting equipment with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0138] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0139] The sand and gravel demand forecasting device for regional markets provided in this application, employing the sand and gravel demand forecasting method for regional markets described in the above embodiments, can solve the technical problem of how to simultaneously consider the macro-cyclical patterns of regional sand and gravel demand and the micro-scale sand consumption characteristics during the engineering construction phase, thereby improving the accuracy of sand and gravel demand forecasting. Compared with the prior art, the beneficial effects of the sand and gravel demand forecasting device for regional markets provided in this application are the same as those of the sand and gravel demand forecasting method for regional markets provided in the above embodiments, and other technical features of this sand and gravel demand forecasting device for regional markets are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0140] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0141] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0142] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, the computer-readable program instructions being used to execute the sand and gravel demand forecasting method for regional markets in the above embodiments.

[0143] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, RAM (Random Access Memory), ROM (Read Only Memory), EPROM (Erasable Programmable Read Only Memory or Flash Memory), optical fibers, CD-ROM (CD-Read Only Memory), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0144] The aforementioned computer-readable storage medium may be included in a regional market-oriented sand and gravel demand forecasting device; or it may exist independently and not be assembled into a regional market-oriented sand and gravel demand forecasting device.

[0145] The aforementioned computer-readable storage medium carries one or more programs. When these programs are executed by a sand and gravel demand forecasting device for a regional market, the device causes the following: it acquires the historical project types, historical project areas, and sand and gravel consumption at each construction stage of historical projects for each month within a preset period in the target region; it calculates the unit project area sand and gravel consumption coefficient for each project type at each construction stage based on the sand and gravel consumption, the historical project area, and the historical project type; it acquires the current project quantity, current project type, current project area, and current project construction stage within the target region's target time dimension; it inputs the current project quantity and current project type into a periodic demand forecasting model to obtain a periodic total demand forecast; it inputs the current project type, current project area, current project construction stage, and the unit project area sand and gravel consumption coefficient corresponding to the current project type into a stage demand forecasting model to obtain a stage total demand forecast; and it performs a weighted fusion of the periodic total demand forecast and the stage total demand forecast according to preset weights to obtain the sand and gravel demand forecast result for the target region within the target time dimension.

[0146] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including LAN (Local Area Network) or WAN (Wide Area Network)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0147] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0148] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0149] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described method for forecasting sand and gravel demand in a regional market. This method addresses the technical problem of simultaneously considering the macro-cyclical patterns of regional sand and gravel demand and the micro-level sand consumption characteristics during engineering construction, thereby improving the accuracy of sand and gravel demand forecasting. Compared to existing technologies, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the sand and gravel demand forecasting method for a regional market provided in the above embodiments, and will not be elaborated upon here.

[0150] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method for forecasting sand and gravel demand in a regional market.

[0151] The computer program product provided in this application can solve the technical problem of how to simultaneously consider the macro-cyclical patterns of regional sand and gravel demand and the micro-scale sand consumption characteristics during the engineering construction phase, thereby improving the accuracy of sand and gravel demand forecasting. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the sand and gravel demand forecasting method for regional markets provided in the above embodiments, and will not be repeated here.

[0152] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for sand and gravel demand forecasting for regional markets, characterized by, The method comprises: obtaining the historical engineering type, the historical engineering area and the historical engineering sand and gravel consumption at each construction stage of each month in a preset period of a target area; calculating the sand and gravel consumption coefficient per unit engineering area of various engineering types at each construction stage according to the sand and gravel consumption, the historical engineering area and the historical engineering type; obtaining the current engineering quantity, the current engineering type, the current engineering area and the current engineering construction stage in a target time dimension of the target area; inputting the current engineering quantity and the current engineering type into a cycle demand prediction model to obtain a cycle total demand prediction value; inputting the current engineering type, the current engineering area, the current engineering construction stage and the unit engineering area sand and gravel consumption coefficient corresponding to the current engineering type into a stage demand prediction model to obtain a stage total demand prediction value; weighting and fusing the cycle total demand prediction value and the stage total demand prediction value according to a preset weight to obtain a sand and gravel demand prediction result of the target area in the target time dimension; The SARIMA model comprises an input layer, an autoregressive module, a difference module, a moving average module, a seasonal autoregressive module, a seasonal difference module, a seasonal moving average module and an output layer. The training step of the cycle demand prediction model comprises: improving the seasonal autoregressive module of the SARIMA model, adding a sub-area cycle dynamic adjustment layer between the seasonal autoregressive module and the seasonal difference module, wherein the sub-area cycle dynamic adjustment layer comprises a cycle parameter mapping table for dynamically matching seasonal cycle parameters according to input area characteristics; improving the input layer of the SARIMA model, adding an engineering type weight embedding layer between the input layer and the autoregressive module, wherein the engineering type weight embedding layer comprises a type-weight correspondence relationship for converting engineering type proportions into weight coefficients of input features; obtaining the historical sand and gravel total demand, the historical engineering quantity and the historical engineering type proportion of each sub-area in the target area in each month in the preset period to form layered training data; inputting the layered training data into the improved SARIMA model, calling the cycle parameter mapping table through the sub-area cycle dynamic adjustment layer to match corresponding seasonal cycle parameters for each sub-area, and calling the type-weight correspondence relationship through the engineering type weight embedding layer to configure input weights for the engineering type proportion of each sub-area; training the SARIMA model after parameter configuration according to the historical sand and gravel total demand to minimize the error between the model output and the actual demand to obtain a basic prediction model of each sub-area; extracting the proportion of the historical demand of each sub-area to the total demand of the target area from the layered training data, weighting and integrating the output results of the basic prediction models of the sub-areas according to the proportion to obtain a cycle demand prediction model.

2. The method of claim 1, wherein, The periodic demand prediction model comprises an input layer, an engineering type weight embedding layer, an autoregressive module, a difference module, a moving average module, a seasonal autoregressive module, a sub-region periodic dynamic adjustment layer, a seasonal difference module, a seasonal moving average module and an output layer; The step of inputting the current engineering quantity and the current engineering type into the periodic demand prediction model to obtain a periodic total demand prediction value comprises: The current engineering quantity is decomposed into a current engineering quantity distribution of each sub-region through the input layer, and the current engineering type is converted into a current engineering type distribution of each sub-region; The engineering type weight embedding layer is used to call a type-weight correspondence relationship, the demand contribution weight of each type of engineering is calculated based on the current engineering type distribution, and the current engineering quantity distribution is multiplied by the corresponding demand contribution weight to obtain sub-region weighted feature data; The sub-region weighted feature data is sequentially input into the autoregressive module, the difference module and the moving average module, and after autoregressive calculation, difference processing and moving average smoothing processing, non-seasonal component processing data is obtained; The non-seasonal component processing data is processed through the seasonal autoregressive module, the sub-region periodic dynamic adjustment layer, the seasonal difference module, the seasonal moving average module and the output layer to obtain a periodic prediction sub-value of each sub-region; The periodic prediction sub-value of each sub-region is multiplied by the corresponding proportion and then summed to obtain a periodic total demand prediction value according to the proportion of the historical demand of each sub-region in the total demand of the target region.

3. The method of claim 2, wherein, The step of processing the non-seasonal component processing data through the seasonal autoregressive module, the sub-region periodic dynamic adjustment layer, the seasonal difference module, the seasonal moving average module and the output layer to obtain a periodic prediction sub-value of each sub-region comprises: The non-seasonal component processing data is subjected to seasonal autoregressive calculation through the seasonal autoregressive module to obtain seasonal autoregressive data; The seasonal cycle parameters preconfigured for each sub-region are called through the sub-region periodic dynamic adjustment layer, and the seasonal autoregressive data is subjected to seasonal cycle correction to obtain seasonal correction data; The seasonal correction data is subjected to seasonal difference processing and seasonal moving average smoothing processing through the seasonal difference module and the seasonal moving average module, and the periodic prediction sub-value of each sub-region is output through the output layer.

4. The method of claim 1, wherein, The stage demand prediction model comprises an input layer, a feature embedding layer, a construction stage adaptation layer, an LSTM unit layer, a cross-type feature migration layer, a full connection layer and an output layer; The step of inputting the current engineering type, the current engineering area, the current engineering construction stage and the unit engineering area gravel consumption coefficient corresponding to the current engineering type into the stage demand prediction model to obtain a stage total demand prediction value comprises: The current engineering type, the current engineering area, the current engineering construction stage and the unit engineering area gravel consumption coefficient are preprocessed and feature-converted through the input layer and the feature embedding layer to obtain a fusion feature vector; The fusion feature vector is input into the construction stage adaptation layer, the LSTM unit layer, the cross-type feature migration layer, the full connection layer and the output layer in sequence, and after construction stage adaptation, LSTM unit calculation, cross-type feature migration, full connection and output layer processing, a stage total demand prediction value is obtained. The construction stage adaptive layer calls a stage-weight mapping matrix to assign a dynamic weight to a numerical feature in the fusion feature vector according to the current construction stage, so as to obtain a feature vector with a stage weight; The feature vector with the stage weight is input into the LSTM unit layer, and after the cell state is updated through the forget gate and the input gate, the hidden layer feature is generated through the output gate; The type similarity calculation module in the cross-type feature migration layer is called to calculate the process feature similarity between the current engineering type and the trained engineering type, and if the process feature similarity is higher than a preset similarity threshold, the feature parameters of the corresponding trained type are extracted from the parameter sharing matrix, and the hidden layer feature is corrected according to the feature parameters and the similarity ratio; The hidden layer feature after correction is subjected to nonlinear transformation through the full connection layer, and a stage demand prediction subvalue of a single current engineering is output through the output layer; The stage demand prediction subvalues of all current engineering in the target area are summed to obtain a stage total demand prediction value.

5. The method of claim 4, wherein, The step of preprocessing and feature conversion of the current engineering type, the current engineering area, the current construction stage and the unit engineering area stone consumption coefficient through the input layer and the feature embedding layer to obtain a fusion feature vector includes: The current engineering type, the current engineering area, the current construction stage and the unit engineering area stone consumption coefficient are standardized through the input layer; After the standardization is completed, the current engineering type is converted into a type feature vector through the feature embedding layer; The current engineering area and the unit engineering area stone consumption coefficient are converted into a numerical feature vector; The type feature vector and the numerical feature vector are merged to obtain a fusion feature vector.

6. The method of any one of claims 1 to 5, wherein, The LSTM model includes an input layer, a feature embedding layer, an LSTM unit layer, a full connection layer and an output layer; The training steps of the stage demand prediction model include: The feature embedding layer of the LSTM model is improved, and a construction stage adaptive layer is added between the feature embedding layer and the LSTM unit layer, the construction stage adaptive layer includes a stage-weight mapping matrix for assigning a dynamic weight according to a construction stage; The full connection layer of the LSTM model is improved, and a cross-type feature migration layer is added between the LSTM unit layer and the full connection layer, the cross-type feature migration layer includes a type similarity calculation module for parameter sharing of the hidden layer feature of a process similar engineering type; According to the unit engineering area stone consumption coefficient, the construction stage information of the historical engineering, the historical engineering area, the stone consumption amount and the process feature data, a plurality of types of training data are obtained; The plurality of types of training data are input into the improved LSTM model, the stage-weight mapping matrix is called through the construction stage adaptive layer, and corresponding weights are configured for different construction stages; The hidden layer feature is corrected according to the feature parameters and the similarity ratio. The type similarity calculation module is called through the cross-type feature migration layer to calculate a process feature similarity between a current engineering type and a trained engineering type, and if the process feature similarity is higher than a preset similarity threshold, hidden layer feature parameters of the trained engineering type are extracted from a parameter sharing matrix, and a reuse weight of the hidden layer feature parameters is allocated according to the process feature similarity; Based on historical sand and gravel consumption in the multi-type training data, the LSTM model after parameter configuration is trained to minimize the error between the model output and the actual consumption, and each engineering type basic prediction model is obtained; The stage demand prediction model is obtained by integrating the engineering type basic prediction models.

7. A sand and gravel demand forecasting device for regional markets, characterized in that, The device comprises: A historical data acquisition module is configured to acquire historical engineering types, historical engineering areas and sand and gravel consumption of historical engineering at each construction stage of a target area in each month of a preset period; A coefficient calculation module is configured to calculate sand and gravel consumption coefficients of various engineering types at each construction stage according to the sand and gravel consumption, the historical engineering area and the historical engineering type; A current data acquisition module is configured to acquire current engineering quantity, current engineering type, current engineering area and current engineering construction stage of the target area in a target time dimension. The period prediction module is configured to input the current project quantity and the current project type into a period demand prediction model to obtain a period total demand prediction value; the SARIMA model comprises an input layer, an autoregressive module, a difference module, a moving average module, a seasonal autoregressive module, a seasonal difference module, a seasonal moving average module and an output layer; the training steps of the period demand prediction model comprise: improving the seasonal autoregressive module of the SARIMA model, adding a sub-regional period dynamic adjustment layer between the seasonal autoregressive module and the seasonal difference module, wherein the sub-regional period dynamic adjustment layer comprises a period parameter mapping table configured to dynamically match seasonal period parameters according to input regional characteristics; improving the input layer of the SARIMA model, adding a project type weight embedding layer between the input layer and the autoregressive module, wherein the project type weight embedding layer comprises a type-weight correspondence relationship configured to convert the project type proportion to a weight coefficient of the input feature; obtaining the historical total sandstone demand, the historical project quantity and the historical project type proportion of each sub-region in the target region in each month within the preset period to form layered training data; inputting the layered training data into the improved SARIMA model, calling the period parameter mapping table through the sub-regional period dynamic adjustment layer to match corresponding seasonal period parameters for each sub-region, and calling the type-weight correspondence relationship through the project type weight embedding layer to configure input weights for the project type proportion of each sub-region; training the SARIMA model after parameter configuration according to the historical total sandstone demand to minimize the error between the model output and the actual demand, thereby obtaining a basic prediction model for each sub-region; extracting the proportion of the historical demand of each sub-region to the total demand of the target region from the layered training data, weighting and integrating the output results of the basic prediction models of each sub-region according to the proportion, and obtaining a period demand prediction model; The stage prediction module is configured to input the current project type, the current project area, the current project construction stage and the unit project area sandstone consumption coefficient corresponding to the current project type into a stage demand prediction model to obtain a stage total demand prediction value. The result fusion module is configured to perform weighted fusion on the period total demand prediction value and the stage total demand prediction value according to a preset weight to obtain a sandstone demand prediction result of the target region in the target time dimension.

8. An area market-oriented sand demand prediction device, characterized by, The device comprises a memory, a processor and a computer program stored on the memory and executable on the processor, and the computer program is configured to implement the steps of the regional market-oriented sandstone demand prediction method according to any one of claims 1 to 6.

9. A storage medium, characterized by The storage medium is a computer readable storage medium, and the storage medium stores a computer program, and the computer program is executed by the processor to implement the steps of the regional market-oriented sandstone demand prediction method according to any one of claims 1 to 6.

Images