Chemical coal distribution method and device

Abstract

The invention relates to a chemical coal distribution method. The method comprises the following steps: obtaining link identifiers between each coal source and a plurality of boilers, and path information and coal blending information matched with the link identifiers; screening a plurality of target link identifiers from the plurality of link identifiers; according to the coal blending information, the risk index and the credibility index corresponding to the target link identifier, obtaining an economic priority corresponding to the target link identifier; based on the economic priority and the coal blending information corresponding to the target link identifier, establishing a weighted logistics directed graph matched with a plurality of transportation points corresponding to the target link identifier; dividing the weighted logistics directed graph into a first sub-graph and a second sub-graph; based on a first processing result of the target space-time diagram convolutional network on the first sub-graph and an aggregation result of a second processing result of the target space-time diagram convolutional network on the second sub-graph, obtaining a furnace arrival probability corresponding to the target link identifier; and distributing the chemical coal based on the furnace arrival probability and the economic priority corresponding to each target link identifier.

Description

Technical Field

[0001] This application relates to the field of coal blending technology, and in particular to a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for distributing coal for chemical use. Background Technology

[0002] In existing coal blending schemes, coal blending optimization and logistics forecasting are typically separated into two independent processes. One side focuses on the economic aspect, first summarizing the daily port market conditions, coal quality inspection results, and boiler sulfur and ash control limits within the procurement system. Then, linear programming or genetic algorithms are used to find the static coal blending combination that meets the quality requirements and minimizes costs. This combination is recorded as a fixed priority for the furnace type-coal source pair and issued during the daily planning stage. The other side focuses on the uncertainties in the transportation process. The industry often uses regression models based on historical distributions or spatiotemporal graph convolutional networks constructed in a node-edge format to segmentally predict the delay probabilities of port loading, railway trunk lines, and in-plant transshipment, outputting the arrival time distribution at the furnace.

[0003] In actual operation, the two results mentioned above are simply merged through rule weighting or threshold filtering. For example, the scheduling system first sorts and selects several high-priority links according to static cost, and then checks the corresponding arrival probability. If it is lower than the preset safety value, its order is adjusted or replaced with the second-best cost link. At the same time, independent rolling schedulers are established for the three types of resources: port, railway, and factory. The scheduler reads the latest predicted delay and arranges the operation according to the principle of first-come, first-served or shortest processing time. Resource conflicts are resolved by local rollback, while cross-segment bottlenecks need to wait for the next prediction cycle to be recalculated.

[0004] However, when using the above method, the link priority is simply superimposed on the logistics forecast in the form of rule thresholds after the fact, without considering the correlation between the two, resulting in low accuracy of the final coal blending. Summary of the Invention

[0005] Therefore, it is necessary to provide a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for distributing chemical coal that can improve the accuracy of coal blending, in order to address the above-mentioned technical problems.

[0006] In a first aspect, this application provides a method for allocating coal for chemical use. The method includes: for multiple coal sources, obtaining link identifiers between each coal source and multiple boilers, as well as path information and coal blending information matching the link identifiers; the path information includes multiple transportation points; based on the coal blending information corresponding to each link identifier, selecting multiple target link identifiers from the multiple link identifiers; for each target link identifier, obtaining the economic priority corresponding to the target link identifier based on the coal blending information, risk indicators, and credibility indicators corresponding to the target link identifier; based on the economic priority and coal blending information corresponding to the target link identifier, establishing a weighted logistics directed graph matching multiple transportation points corresponding to the target link identifier; dividing the weighted logistics directed graph into a first subgraph and a second subgraph; the edge attributes corresponding to the first edge between the first nodes in the first subgraph satisfy a preset condition, and the edge attributes corresponding to the second edge between the second nodes in the second subgraph do not satisfy the preset condition; based on the aggregation result of the first processing result of the target spatiotemporal graph convolutional network on the first subgraph and the second processing result of the target spatiotemporal graph convolutional network on the second subgraph, obtaining the arrival probability of the target link identifier corresponding to the boiler; and allocating coal for chemical use based on the arrival probability and economic priority corresponding to each target link identifier.

[0007] The aforementioned method for allocating coal for chemical use involves multiple coal sources. It obtains the link identifiers between each coal source and multiple boilers, along with path information and coal blending information matching these link identifiers. The path information includes multiple transportation points. Based on the coal blending information corresponding to each link identifier, multiple target link identifiers are selected from these. For each target link identifier, an economic priority is obtained based on the corresponding coal blending information, risk indicators, and reliability indicators. Based on the economic priority and coal blending information, a weighted directed logistics graph matching multiple transportation points corresponding to the target link identifier is established. This graph is then divided into a first subgraph and a second subgraph. The edge attributes of the first edge between the first nodes in the first subgraph satisfy a preset condition, while the edge attributes of the second edge between the second nodes in the second subgraph do not. The arrival probability of the target link identifier is obtained by aggregating the first processing result of the target spatiotemporal graph convolutional network on the first subgraph and the second processing result of the target spatiotemporal graph convolutional network on the second subgraph. Finally, based on the arrival probability and economic priority of each target link identifier, coal for chemical use is allocated. Therefore, this application establishes a weighted logistics directed graph based on economic priority and coal allocation information, which enables the allocation of coal for chemical use to be linked with economic priority and weighted logistics directed graph, thereby improving the accuracy of coal allocation for chemical use. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart illustrating a method for distributing coal for chemical use in one embodiment;

[0010] Figure 2 This is a flowchart illustrating the process of obtaining the economic priority corresponding to the target link identifier based on the coal blending information, risk indicators, and credibility indicators corresponding to the target link identifier in one embodiment.

[0011] Figure 3 This is a flowchart illustrating the process of establishing a weighted logistics directed graph that matches multiple transportation points corresponding to the target link identifier based on the economic priority and coal blending information corresponding to the target link identifier in one embodiment.

[0012] Figure 4 This is a schematic diagram of the network architecture of a target spatiotemporal graph convolutional network in one embodiment.

[0013] Figure 5 This is a flowchart illustrating the coal distribution method for chemical applications in another embodiment;

[0014] Figure 6 This is a structural block diagram of a chemical coal distribution device in one embodiment;

[0015] Figure 7 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0017] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0018] For existing coal blending schemes, at the information organization level, most schemes use a centralized database to store static cost priorities, while logistics forecasting results exist in the form of message streams or cache tables, with the two linked by a primary key mapping. The scheduling engine extracts static priorities and the latest delay probabilities in each rolling window, and generates a job list using weighted scoring or hierarchical filtering methods. When calculating resource allocation, shunting, loading, and stacking / reclaiming are treated as independent constraints, and feasible schedules are obtained through heuristic search or integer programming. The entire framework can balance cost and risk to some extent, but the update frequency of static cost weights is far lower than the delay changes on the transportation side, and bottleneck information mainly relies on rule inference in the post-processing stage rather than directly coupling economic value within the forecasting model.

[0019] As shown above, although existing solutions can output both the static coal blending result that minimizes costs and the transportation delay prediction in the spatiotemporal dimension, the two are disconnected in terms of time scale, data structure, and weight update mechanism. Static coal blending priority is usually refreshed at the daily level, while the logistics delay probability rolls on a minute or hourly cycle. This asynchronous rhythm causes scheduling decisions to lag behind the actual economic situation under highly volatile market conditions. Secondly, the fusion of the two results mostly relies on threshold filtering or ex-post weighting, lacking an explicit characterization of the cost-risk coupling relationship within the model. The scheduling algorithm cannot distinguish which bottleneck links have a greater impact on overall procurement expenditure, nor can it quickly shift attention from the blocked main line to alternative channels when the delay risk surges. In addition, resource schedulers usually treat shipping, rail, and in-plant transshipment as independent constraints, lacking a unified quantitative indicator oriented towards economic value. This leads to the average distribution of computing power and warehousing resources, significantly weakening the guarantee of high-value links. In summary, existing technologies have not yet provided a complete closed loop that injects compliant and cost-effective coal blending priorities into spatiotemporal transportation forecasts in real time and drives multi-segment scheduling with uniform weights.

[0020] Therefore, in chemical plant scenarios with multiple coal sources, drastic price fluctuations, and uncertainties in multiple transportation segments, it is crucial to prioritize static coal blending that meets process red lines for sulfur and ash and has the lowest cost. This priority should be injected in real time into a predictive model that includes the spatiotemporal risks of ports, railways, and internal plant links. This would enable the scheduling system to dynamically concentrate limited computing power, warehousing, and loading / unloading resources on the high-risk transportation links that have the greatest impact on overall costs, thereby simultaneously reducing procurement costs and significantly reducing the probability of warehouse overflows or material shortages.

[0021] In view of this, this application provides a method for distributing coal for chemical use. This method can be applied to a server in a coal distribution system for chemical use. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. Specifically, such as... Figure 1 As shown, it includes the following steps:

[0022] S102, for multiple coal sources, obtain the link identifier between each coal source and multiple boilers, as well as the path information and coal blending information matching the link identifier.

[0023] In this context, any one of the multiple coal sources can be associated with any one of the multiple boilers. This association is represented by a link identifier, which uniquely identifies the relationship between the coal source and the boiler. For example, if k represents the coal source and b represents the boiler, then the link identifier between the coal source and the boiler can be represented as ID. k,b In some cases, the link identifier between the coal source and the boiler can also be associated with the boiler number, physical path number, and time. The physical path number represents the number of the path taken to transport the coal source to the boiler.

[0024] The route information includes multiple transport points, which can refer to ports, railways, and within the plant, etc. In other words, the coal source can be transported to the corresponding boiler through multiple transport points according to the transport sequence.

[0025] Coal blending information refers to information related to the allocation of coal for chemical industry. Coal blending information includes, but is not limited to: coal quality indicators (i.e., coal source quality indicators), boiler sulfur and ash limits, and real-time throughput of transportation points corresponding to coal sources.

[0026] S104, based on the coal blending information corresponding to each link identifier, select multiple target link identifiers from multiple link identifiers.

[0027] Specifically, for each link identifier, coal quality indicators are extracted from the coal blending information of the link identifier; when the coal quality indicator is less than or equal to the indicator threshold, the compliance mask of the link identifier is set to the first value; when the coal quality indicator is greater than the indicator threshold, the compliance mask of the link identifier is set to the second value; from multiple link identifiers, the link identifier with the compliance mask set to the first value is determined as the target link identifier.

[0028] For example, coal quality indicators may include sulfur content. Ash content volatile matter The indicator thresholds include the first judgment threshold. Second judgment threshold Second determination threshold The first to the third judgment thresholds can be determined based on the upper limit corresponding to the boiler.

[0029] Specifically, the compliance mask for the link identifier satisfies:

[0030]

[0031] in, This represents a compliance mask indicating the link identifier between coal source k and boiler b. To satisfy the indicator function that is 1 otherwise 0, the link identifier is determined as the target link identifier when the compliance mask is the first value (e.g., 1). Therefore, by eliminating link identifiers with a compliance mask of the second value (e.g., 0), the interference of invalid paths with priority ranking can be avoided, thus improving the accuracy of coal allocation for chemical industries.

[0032] S106. For each target link identifier, the economic priority corresponding to the target link identifier is obtained based on the coal blending information, risk indicators, and credibility indicators corresponding to the target link identifier.

[0033] Among them, the risk indicator corresponding to the target link identifier is used to characterize the price fluctuation range of the target link identifier in the past three days. The reliability indicator corresponding to the target link identifier is used to characterize the static arrival reliability of the target link identifier in the past week, that is, the seven-day on-time rate of the target link identifier.

[0034] In one embodiment, the risk indicator corresponding to the target link identifier satisfies:

[0035]

[0036] in, This represents the risk indicator corresponding to the target link identifier between coal source k and boiler b. T3 represents the first three days of the current production window, and t represents the time within the first three days. Let represent the transaction price per unit at time t. This indicates the highest single transaction price within the previous three days. This indicates the lowest transaction price within the previous three days.

[0037] In one embodiment, the credibility index corresponding to the target link identifier satisfies:

[0038]

[0039] in, This represents the credibility index corresponding to the target link identifier between coal source k and boiler b. This represents the actual time taken to reach the furnace at time t. This indicates the maximum acceptable power generation delay limit for boiler b. T7 represents the previous seven days of the current production window, and t represents the time within those seven days. This refers to the number of samples within a seven-day window, that is, the total number of samples that arrive at the furnace on time within seven days.

[0040] S108. Based on the economic priority and coal blending information corresponding to the target link identifier, a weighted directed logistics graph matching multiple transportation points corresponding to the target link identifier is established.

[0041] S110, the weighted logistics directed graph is divided into a first subgraph and a second subgraph; the edge attributes corresponding to the first edge between the first nodes in the first subgraph meet the preset conditions, and the edge attributes corresponding to the second edge between the second nodes in the second subgraph do not meet the preset conditions.

[0042] In this diagram, the first edge in the first subgraph and the second edge in the second subgraph can represent the main line segment, the loading segment, or the segment within the factory. The edge attributes in the subgraphs represent the corresponding transportation distance d. e and craft rhythm The cycle time refers to the transportation time along the path corresponding to the edge.

[0043] Specifically, in the weighted logistics directed graph, the edge corresponds to the transportation distance d. e Greater than or equal to distance threshold D thr or the rhythm of the process Greater than or equal to the time threshold T thr In the case of [condition], that edge is designated as the first edge, meaning the edge attribute corresponding to that edge satisfies the preset conditions. The distance threshold is measured in kilometers, and the time threshold is measured in minutes.

[0044] Specifically, in the weighted logistics directed graph, the edge corresponds to the transportation distance d. e Less than distance threshold D thr And the production cycle Less than the time threshold T thr In the case of this, the edge is determined as the second edge, meaning that the edge attribute corresponding to this edge does not meet the preset conditions.

[0045] for example, This represents the first subgraph. If we represent the second subgraph, then the following condition must be met:

[0046]

[0047] Where 'e' represents an edge in the weighted logistics directed graph. This indicates the granularity of edge e; when the edge attributes corresponding to edge e meet preset conditions, Therefore, a first subgraph can be formed by multiple edges with a granularity of 1 in the weighted logistics directed graph; the edge with a granularity of 1 in the weighted logistics directed graph is called the first edge in the first subgraph. When the edge attribute corresponding to edge e does not meet the preset conditions, In other words, a second subgraph can be formed by multiple edges with a granularity of 0 in the weighted logistics directed graph; the edges with a granularity of 0 in the weighted logistics directed graph are called second edges in the second subgraph.

[0048] S112, based on the aggregation result of the first processing result of the target spatiotemporal graph convolutional network on the first subgraph and the second processing result of the target spatiotemporal graph convolutional network on the second subgraph, the arrival probability corresponding to the target link identifier is obtained.

[0049] In one embodiment, an event-driven gating unit is inserted into each layer of the spatiotemporal graph convolutional network, and conditions such as tidal restrictions, temporary traffic restrictions, and equipment failures are injected to obtain the target spatiotemporal graph convolutional network. Thus, when the target spatiotemporal graph convolutional network processes the first subgraph or the second subgraph, the event-driven gating unit can immediately write the event tensor ε to the edges of the first or second subgraph. e (t) and calculate the gating coefficient g. e (t)=1+σ ε e (t) is used to amplify or suppress the convolutional activation values ​​of each layer and the next layer of the target spatiotemporal graph convolutional network.

[0050] In the context of coal use in chemical industries, tidal constraints narrow the periodic operating window of the port loading section: when the tide level is below the draft threshold, the loading capacity drops instantaneously, and the event tensor is denoted as ε. e (t)<0, gating coefficient g e (t)=1+σ ε e (t) causes the convolution activation to decay proportionally; ε during tide rise e When (t) approaches 0, the gate is lifted, and the port section returns to normal. Temporary traffic restrictions correspond to temporary closures, speed limits, and construction windows on main railway lines or stations, directly affecting the edges in the first subgraph. During the event's effective period, it also affects g. e If (t) < 1, reducing the effective throughput of this segment and increasing the propagation intensity of the delay potential energy can degating the backgating and returning to normal. Equipment failure corresponds to abnormalities in key equipment such as stacker-reclaimers, belt conveyors, tippers, and unloaders on the short chain within the plant, and their impact on real-time throughput is expressed as ε. e Inject the corresponding edge if (t) < 0, and gate g. e (t) Immediately suppress the activation of this channel.

[0051] S114: Based on the arrival probability and economic priority corresponding to each target link identifier, coal for chemical industry is allocated.

[0052] Using the method described in the above embodiments, for multiple coal sources, the method obtains the link identifiers between each coal source and multiple boilers, as well as the path information and coal blending information matching the link identifiers. The path information includes multiple transportation points. Based on the coal blending information corresponding to each link identifier, multiple target link identifiers are selected from the multiple link identifiers. For each target link identifier, the economic priority corresponding to the target link identifier is obtained according to the coal blending information, risk indicators, and credibility indicators corresponding to the target link identifier. Based on the economic priority and coal blending information corresponding to the target link identifier, a weighted logistics directed graph matching multiple transportation points corresponding to the target link identifier is established, and the weighted logistics directed graph corresponding to the target link identifier is divided into a first subgraph and a second subgraph. The edge attributes corresponding to the first edge between the first nodes in the first subgraph meet the preset conditions, and the edge attributes corresponding to the second edge between the second nodes in the second subgraph do not meet the preset conditions. Based on the aggregation results of the first processing result of the target spatiotemporal graph convolutional network on the first subgraph and the second processing result of the target spatiotemporal graph convolutional network on the second subgraph, the arrival probability of the target link identifier is obtained. Then, based on the arrival probability and economic priority corresponding to each target link identifier, the coal for chemical use is allocated. Therefore, this application establishes a weighted logistics directed graph based on economic priority and coal allocation information, which enables the allocation of coal for chemical use to be linked with economic priority and weighted logistics directed graph, thereby improving the accuracy of coal allocation for chemical use.

[0053] In one embodiment, such as Figure 2 As shown, based on the coal blending information, risk indicators, and reliability indicators corresponding to the target link identifier, the economic priority corresponding to the target link identifier is obtained, including:

[0054] S202, for the coal blending information corresponding to the target link identifier, based on the preset process range corresponding to each of the multiple continuous fields in the coal blending information, obtain the dimensionless ratio value corresponding to each continuous field.

[0055] In this context, continuous fields in coal blending information refer to numerical variables, such as coal quality indicators, coal source prices, and throughput capacity. Discrete fields in coal blending information refer to categorical variables, such as coal source type, supplier, port, pipeline, and equipment status. In some embodiments, for each discrete field, the master data dictionary can be invoked to assign a unique integer label. z In addition, a business domain prefix can be added before the integer label, for example, the supplier prefix is ​​S_, the port prefix is ​​P_, and finally the ⟨ prefix is ​​used. z > Represents discrete fields.

[0056] Specifically, for each continuous field, the dimensionless ratio value corresponding to the continuous field satisfies:

[0057]

[0058] in, This indicates a continuous field, where r represents the dimensionless proportion corresponding to the continuous field. Indicates the lower limit of the preset process range. This indicates the upper limit of the preset process range.

[0059] S204. The dimensionless ratio values ​​are sequentially concatenated with the discrete fields in the coal blending information to obtain the first concatenation vector corresponding to the target link identifier.

[0060] In one embodiment, when the proportional value sequence is composed of dimensionless proportional values ​​and the discrete label sequence is composed of discrete fields, the proportional value sequence and the discrete label sequence are concatenated into a row vector according to a fixed field order to obtain a first concatenated vector. The first concatenated vector is used to describe the static attributes and instantaneous state of the target link identifier in the current production window.

[0061] S206, associate the target link identifier with the first concatenation vector to obtain the cross-domain feature corresponding to the target link identifier.

[0062] In one scenario, by using the target link identifier as the row index and combining the cross-domain features corresponding to multiple target link identifiers, a cross-domain feature matrix can be obtained.

[0063] S208. Based on the cross-domain characteristics, risk indicators, and credibility indicators corresponding to the target link identifier, the economic priority corresponding to the target link identifier is obtained.

[0064] In this process, the economic priority of each target link is combined to obtain the economic priority weight matrix.

[0065] In one embodiment, the economic priority corresponding to the target link identifier is obtained based on the cross-domain characteristics, risk indicators, and credibility indicators corresponding to the target link identifier, including the following steps:

[0066] Step 1: Combine the cross-domain features, risk indicators, and credibility indicators corresponding to the target link identifier to obtain the link feature vector corresponding to the target link identifier.

[0067] Step 2: Input the link feature vector corresponding to the target link identifier into the first-stage self-searching factor decomposition machine to obtain the procurement cost baseline corresponding to the target link identifier.

[0068] The first-stage self-searching factorization machine can adopt a standard structure with fixed embedding dimensions and interaction with second-order features, without performing architecture search.

[0069] Step 3: Sort the procurement cost baselines corresponding to the multiple target link identifiers in ascending order to obtain the cost sequence number corresponding to the target link identifier.

[0070] Step 4: Normalize and combine the cost sequence number, risk index and credibility index corresponding to the target link identifier to obtain the three-dimensional interaction risk vector corresponding to the target link identifier.

[0071] Specifically, by normalizing the cost sequence number, risk index, and credibility index corresponding to the target link identifier and then combining them bit by bit, a three-dimensional interaction risk vector corresponding to the target link identifier can be obtained.

[0072] for example, This represents the cost sequence number corresponding to the i-th target link identifier. This represents the risk indicator corresponding to the i-th target link identifier. Let represent the credibility index corresponding to the i-th target link identifier. Then, the three-dimensional interaction risk vector corresponding to the target link identifier satisfies:

[0073]

[0074] in, This represents the three-dimensional interaction risk vector corresponding to the i-th target link identifier. This represents the feasible set, which includes multiple target link identifiers. This represents the total number of target link identifiers in the feasible set. ε is a very small positive number to prevent the denominator from being zero.

[0075] Step 5: Concatenate the three-dimensional interactive risk vector with the original feature embedding results corresponding to the cross-domain features to obtain the second concatenated vector, and input the second concatenated vector into the second-stage self-searching factor decomposition machine to obtain the comprehensive ranking weight corresponding to the target link identifier.

[0076] Specifically, the method for obtaining the original feature embedding result corresponding to the cross-domain feature includes: normalizing the cross-domain feature to obtain the original feature embedding result corresponding to the cross-domain feature.

[0077] for example, The comprehensive ranking weight corresponding to the i-th target link identifier satisfies:

[0078]

[0079] in, This indicates the second-stage self-searching factorization machine, which is a factorization machine with searchable embedded dimensions and weights. This represents the result of the original feature embedding.

[0080] Step 6: Perform a weighted summation of the procurement cost baseline, comprehensive ranking weight, and interactive risk vector contribution corresponding to the target link identifier, to obtain the economic priority corresponding to the target link identifier.

[0081] Specifically, the product of the transpose of the three-dimensional risk projection vector and the three-dimensional interactive risk vector is determined as the contribution of the interactive risk vector.

[0082] for example, The economic priority corresponding to the i-th target link identifier is:

[0083]

[0084] in, , , These are the three weighting coefficients. ; represents the procurement cost baseline corresponding to the i-th target link identifier; u represents the three-dimensional risk projection vector, which is automatically learned by architecture search. This represents the contribution of the interaction risk vector.

[0085] Using the above Figure 2 The method, considering the risk and credibility indicators corresponding to the target link identifier, uses a self-searching factor decomposition machine to obtain economic priority, which can improve the accuracy of the analysis.

[0086] In one embodiment, such as Figure 3 As shown, based on the economic priority and coal blending information corresponding to the target link identifier, a weighted directed logistics graph matching multiple transportation points corresponding to the target link identifier is established, including:

[0087] S302. Based on the transportation direction between the coal source and the boiler corresponding to the target link identifier, establish a directed link graph that matches multiple transportation points corresponding to the target link identifier.

[0088] S304. For each transport edge in the directed graph corresponding to the target link identifier, extract the real-time available throughput of the transport edge at the target time from the coal distribution information corresponding to the target link identifier.

[0089] In this context, a transport edge refers to an edge between two adjacent nodes in a directed graph of links.

[0090] S306: Based on the real-time available throughput and design throughput at the target time, obtain the bottleneck transfer ratio corresponding to the transport edge at the target time.

[0091] Specifically, the ratio between the real-time available throughput corresponding to the transport edge and the designed throughput is determined as the bottleneck transfer ratio corresponding to the transport edge.

[0092] for example, The transport edge is represented by the following: when s=1, the transport edge represents the loading section; when s=2, the transport edge represents the main line section; and when s=3, the transport edge represents the section within the factory. The bottleneck transfer ratio of the transport edge at time t is expressed as follows:

[0093]

[0094] in, This represents the real-time available throughput of the transport edge at time t. This represents the designed throughput corresponding to the transport edge.

[0095] In some embodiments, when the design throughput corresponding to a transport edge is missing, a preset throughput is determined as the design throughput corresponding to the transport edge to maintain the continuity of the ratio calculation.

[0096] S308. Based on the bottleneck transfer ratio of the transport edge at the target time and the economic priority corresponding to the target link identifier, the economic edge weight of the transport edge at the target time is obtained.

[0097] Specifically, determine the first difference between 1 and the bottleneck transfer ratio corresponding to the transport edge; determine the statistical value of the first difference corresponding to each transport edge; and determine the ratio of the first difference to the statistical value as the economic edge weight corresponding to the transport edge.

[0098] for example, The economic edge weight corresponding to the transportation edge satisfies:

[0099]

[0100] in, This indicates the total sensitivity corresponding to the target link identifier where the transport edge is located. As can be seen from the above formula, according to the principle that the more severe (lower) the bottleneck transfer ratio, the higher the weight is assigned, mapping the static economic value to the segmented edges can lay the initial distribution for bottleneck reinforcement and dynamic weight leakage.

[0101] S310 combines the economic edge weight, bottleneck transfer ratio, and furnace arrival time of the transport edge at the target time to obtain the edge attributes corresponding to the transport edge.

[0102] S312, associate the edge attributes corresponding to the transportation edge with the transportation edge in the directed graph of the link to obtain a weighted logistics directed graph that matches multiple transportation points corresponding to the target link identifier.

[0103] use Figure 3The content shown directly solidifies economic priorities onto the edge weights of the logistics graph through the inverse mapping of the bottleneck transfer ratio. As a result, economic value is no longer retroactively added in the form of rule thresholds, thus improving the accuracy of coal allocation for chemical industry.

[0104] In one embodiment, the economic edge weight and arrival time limit corresponding to the transport edge can be determined based on the relationship between the bottleneck transfer ratio corresponding to the transport edge and the business threshold.

[0105] Specifically, the bottleneck transfer ratio at the transportation edge is less than the bottleneck determination threshold. In this case, update the bottleneck transfer ratio and furnace arrival time corresponding to the transport edge. That is, calculate the strengthening coefficient corresponding to the transport edge. At the same time, the time limit for this section to arrive at the furnace will be tightened. The two are related by the same bottleneck function as follows:

[0106]

[0107]

[0108] Among them, when This indicates a need to increase economic focus on the transportation sector. This indicates the arrival time limit of the furnace corresponding to the transport side. This indicates the arrival time limit of the furnace corresponding to the updated transport edge. and These are the magnification factor and the first reduction factor, respectively. The value is assigned only if the content within the parentheses is positive; otherwise, it is 0.

[0109] Specifically, the updated economic edge weight is the product of the reinforcement coefficient and the economic edge weight. Therefore, by constructing a weighted logistics directed graph based on the updated economic edge weight, the downstream spatiotemporal graph convolutional network can not only calculate higher economic losses when encountering severe bottlenecks, but also complete delayed propagation within a narrower time window.

[0110] In one embodiment, the method further includes the following steps:

[0111] Step 1: Based on the path information corresponding to the target link identifier, filter the boundary nodes that are associated with the first subgraph and the second subgraph from each first node in the first subgraph and each second node in the second subgraph corresponding to the target link identifier.

[0112] The transport edges pointed to by the boundary nodes include the first edge in the first subgraph and the second edge in the second subgraph.

[0113] Step 2: The bottleneck transfer ratio corresponding to the upstream edge of the boundary node is less than or equal to the alarm threshold. At that time, the leakage amount is extracted from the economic edge weight corresponding to the upstream edge to obtain the updated economic edge weight corresponding to the upstream edge.

[0114] The upstream edge is a transport edge pointing to the boundary node. Specifically, Indicates the leakage amount, satisfying:

[0115]

[0116] in, This represents the leakage ratio coefficient. This represents the economic edge right corresponding to the upstream edge.

[0117] Step 3: For the downstream edge corresponding to the boundary node, obtain the updated economic edge weight of the downstream edge based on the economic edge weight of the downstream edge, the updated economic edge weight of the upstream edge, and the leakage amount.

[0118] The number of downstream edges is at least two, with at least one downstream edge existing in the first subgraph and at least one downstream edge existing in the second subgraph.

[0119] for example, Let the updated economic edge weight corresponding to the j-th downstream edge satisfy:

[0120]

[0121] in, This represents the economic edge weight corresponding to the j-th downstream edge. Let J represent the bottleneck transfer ratio corresponding to the j-th downstream edge, and J represent the total number of downstream edges. Therefore, by automatically injecting economic edge weights according to the reverse principle that the smoother the branch (the higher the bottleneck transfer ratio), the less weight is allocated, and the more congested the branch, the more weight is allocated, this prompts the scheduling attention to quickly and smoothly shift from the blocked main line to the branch with the best chance of ensuring delivery to the furnace, responding to sudden changes in the transportation pattern without waiting for model retraining.

[0122] In one embodiment, the first processing result includes the first state vector corresponding to each first node in the first subgraph at the next time step at the target time. The method further includes: inserting event-driven gating units into each layer of the spatiotemporal graph convolutional network, and injecting conditions such as tidal restrictions, temporary traffic restrictions, and equipment failures to obtain the target spatiotemporal graph convolutional network. Specifically, the processing method of the target spatiotemporal graph convolutional network for the first subgraph includes the following steps:

[0123] Step 1: For the first edge between adjacent first nodes in the first subgraph, obtain the queuing depth of the first edge at the target time, and obtain the delay potential energy of the first edge at the target time based on the product of the queuing depth of the first edge at the target time and the economic edge weight.

[0124] Specifically, the product of the queuing depth corresponding to the first side and the economic edge weight is determined as the delay potential energy corresponding to the first side. For example, Let the delayed potential energy of the first side at time t be such that:

[0125]

[0126] in, This represents the queuing depth of the first side at time t. This represents the economic edge weight of the first edge at time t.

[0127] Step 2: For each first node in the first subgraph, determine the first upstream edge pointing to the first node from each first edge.

[0128] Step 3: Determine the delayed potential energy corresponding to the first upstream edge at the target time, and obtain the time limit value after time limit normalization by the time limit of the furnace arrival corresponding to the first upstream edge.

[0129] Specifically, the ratio of the delayed potential energy corresponding to the first upstream edge to the furnace arrival time corresponding to the first upstream edge is determined as the time limit value corresponding to the first upstream edge.

[0130] for example. Let the time limit value of the first upstream edge at time t be such that:

[0131]

[0132] in, This indicates the arrival time limit for the furnace corresponding to the first upstream edge. This arrival time limit can refer to the updated road time limit. Therefore, by normalizing the delay potential energy, the node state can be automatically adjusted according to the principle that the greater the impact of the same potential energy on the more pressing time limit, so that the slow variable congestion across ports and railways is fully captured on an hourly scale, without diluting the amplifying effect of economic weight on delay.

[0133] Step 4: Based on the time limit value corresponding to the first upstream edge at the target time, obtain the first state vector corresponding to the first node at the next time after the target time.

[0134] Among them, based on the time limit value corresponding to the first upstream edge at the target time, an explicit injection method that implements slow variable blocking can be adopted to obtain the first state vector corresponding to the next time of the first node at the target time.

[0135] Specifically, This indicates that the first state vector of the first node v at time t+1 satisfies:

[0136]

[0137] in, Here, W1 and W2 are non-linear activation functions, and W1 and W2 are the weight matrices to be learned. This represents the first state vector of the first node v at time t. This indicates the upstream node of the first node, meaning the upstream node points to the first node; Indicates upstream node The first state vector at time t.

[0138] By employing the method described in the above embodiments and normalizing the delayed potential energy over a time limit, the accuracy of the analysis of the first state vector can be improved, and further, the accuracy of coal allocation for chemical applications can be improved.

[0139] In one embodiment, the second processing result includes the second state vector corresponding to each second node in the second subgraph at the next time step after the target time step. Specifically, the processing method of the target spatiotemporal graph convolutional network for the second subgraph includes the following steps:

[0140] Step 1: For the second edge between adjacent second nodes in the second subgraph, determine the cost-sensitive gate corresponding to the second edge at the target time based on the relationship between the economic edge weight and the quantile rank threshold of the second edge at the target time.

[0141] The quantile rank threshold refers to the threshold set for the quantile rank in a weighted logistics directed graph. For example, the quantile rank threshold can be 0.8. Specifically, if the economic edge weight corresponding to the second edge is greater than or equal to the quantile rank threshold, the cost sensitivity threshold corresponding to the second edge is set to 1; otherwise, it is set to a reduction coefficient. The reduction coefficient is greater than 0 and less than 1.

[0142] Step 2: For each second node in the second subgraph, determine the second upstream edge pointing to the second node from each second edge.

[0143] Step 3: Determine the queuing potential energy represented by the product of the queuing depth and economic edge weight of the second upstream edge at the target time and the scaling factor, and normalize it to a minute-level value according to the furnace arrival time limit corresponding to the second upstream edge.

[0144] Specifically, the product of the queuing depth and economic edge weight corresponding to the second side and the scaling factor is determined as the queuing potential energy corresponding to the second side.

[0145] for example, Let the minute-level value of the second upstream edge at time k satisfy:

[0146]

[0147] in, This represents the queuing depth of the second side at time k. Let k represent the economic edge weight of the second edge at time k. k2 represents the second reduction factor.

[0148] Step 4: Based on the minute-level value of the second upstream edge at the target time and the cost-sensitive gate, obtain the second state vector of the second node at the next time after the target time.

[0149] Specifically, Let the second state vector of the second node w at time k+1 satisfy:

[0150]

[0151] Where W3 and W4 are the weight matrices to be learned. Let w represent the second state vector of the second node w at time k. Indicates the upstream node of the second node. Indicates upstream node The second state vector corresponding to time k. This represents the cost-sensitive gate on the second side, which determines the opening and closing strength of the convolution channel.

[0152] By normalizing the queuing potential energy to minute-level values ​​using the method described in the above embodiments, the accuracy of the first state vector analysis can be improved, which in turn can improve the accuracy of coal allocation for chemical production. Furthermore, based on the dynamic adjustment of cost-sensitive gates, it is possible to proactively focus on the shortest intra-plant chains (i.e., the chains corresponding to the second edge) with the highest economic edge weight and the most significant queuing delay, while retaining only basic channels for low-value chains. This saves computational resources and avoids averaging high-risk paths, ensuring that forecasts prioritize serving high-cost, high-risk, and critical chains prone to material shortages.

[0153] In one embodiment, the arrival probability corresponding to the target link identifier is obtained by aggregating the first processing result of the target spatiotemporal graph convolutional network on the first subgraph and the second processing result of the target spatiotemporal graph convolutional network on the second subgraph, including the following steps:

[0154] Step 1: Based on the path information corresponding to the target link identifier, filter the boundary nodes that are associated with the first subgraph and the second subgraph from each first node in the first subgraph and each second node in the second subgraph.

[0155] Step 2: Perform time alignment and concatenation on the first state vector corresponding to the first side pointed to by the boundary node at the target time and the second state vector corresponding to the second side pointed to by the boundary node at the target time to obtain the third concatenated vector corresponding to the boundary node at the target time.

[0156] Step 3: Based on the third concatenation vector corresponding to the boundary node at the target time, obtain the probability of the target link identifier arriving at the furnace at the next time.

[0157] Specifically, the product of the transpose of the weight vector and the third concatenation vector, and the sum of the product with the bias vector, are used to determine the target vector. The target vector is then processed using a nonlinear activation function to obtain the arrival probability of the target link identifier.

[0158] for example, Let the probability of reaching the furnace corresponding to the i-th target link identifier be satisfied as follows:

[0159]

[0160] in, This represents the bias vector. Represents the weight vector. This represents the transpose of the weight vector. This represents the third concatenation vector corresponding to the target link identifier at time t.

[0161] like Figure 4 As shown, a schematic diagram of a target spatiotemporal graph convolutional network is provided, wherein, based on Figure 4 The network structure shown allows us to obtain the probability of reaching the furnace corresponding to the target link identifier.

[0162] Based on the above, it can be seen that the first subgraph or the second subgraph can be processed based on the target spatiotemporal graph convolutional network. In some embodiments, the target spatiotemporal graph convolutional network can also be optimized, and the final probability of arrival at the furnace can be obtained based on the optimized target spatiotemporal graph convolutional network. Specifically, the optimization method includes the following steps:

[0163] Step 1: For each target link identifier, based on the arrival probability and measured arrival probability of the target link identifier, obtain the basic prediction error corresponding to the target link identifier.

[0164] Specifically, the absolute value of the difference between the measured probability of furnace arrival and the actual probability of furnace arrival is determined as the basic prediction error.

[0165] Step 2: Based on the furnace arrival mark corresponding to the measured furnace arrival probability and the basic prediction error, obtain the penalty enhancement error corresponding to the target link identifier.

[0166] Specifically, The penalty enhancement error corresponding to the i-th target link identifier satisfies:

[0167]

[0168] in, This indicates the economic priority corresponding to the i-th target link identifier. This represents the basic prediction error corresponding to the i-th target link identifier. The amplification factor is a fixed factor, set by the business side. This represents the transport tag corresponding to the i-th target link identifier. , This represents the arrival marker corresponding to the measured arrival probability. Specifically, when the transport marker characterizes the delay, Conversely, when transport marks indicate punctuality, .

[0169] It is understandable that multiplying by the linearity error beforehand is acceptable. And additional stacking when delay occurs. This can amplify the gradient contribution of samples with high economic weights and predicted failures, allowing the model to prioritize correcting high-risk paths.

[0170] Step 3: Normalize the penalty enhancement error to obtain the normalized result, and input the normalized result into the optimizer in the target spatiotemporal graph convolutional network to complete gradient backpropagation and weight update.

[0171] Specifically, determine the maximum penalty enhancement error among the penalty enhancement errors corresponding to each target link identifier;

[0172] The ratio of the penalty enhancement error corresponding to the target link identifier to the maximum penalty enhancement error is determined as the normalized result.

[0173] for example, The normalized result corresponding to the i-th target link identifier satisfies:

[0174]

[0175] in, This represents the maximum penalty-enhanced error.

[0176] Understandably, clearing the transport marker set immediately after the data return is completed ensures that the next training batch is recalculated based solely on the real-time economic priority and the latest transport markers. This continuously directs resources toward the critical links with the highest economic value and greatest transport risk throughout the training process, thereby improving the accuracy of coal allocation for chemical production.

[0177] Updating the target spatiotemporal graph convolutional network using the above steps can improve the accuracy of the analysis of the arrival probability corresponding to the target link identifier. Therefore, in the embodiments related to this application, the arrival probability corresponding to the target link identifier can be obtained based on the optimized target spatiotemporal graph convolutional network.

[0178] In one embodiment, coal for chemical production is allocated based on the arrival probability and economic priority corresponding to each target link identifier, including the following steps:

[0179] Step 1: Determine the risk controllability correction factor that is positively correlated with the probability of arrival at the furnace and the remaining time limit of arrival at the furnace corresponding to the target link identifier.

[0180] Specifically, The risk controllability correction factor corresponding to the i-th target link identifier satisfies:

[0181]

[0182] in, This represents the probability of reaching the furnace corresponding to the i-th target link identifier. This represents the remaining time limit for the i-th target link identifier to reach the furnace. This represents the maximum value among the remaining time limits for each target link identifier to reach the furnace.

[0183] Step 2: Based on the matching results between the safety operation information and safety operation conditions corresponding to the target link identifier within the current time window, determine the constraint coefficient corresponding to the target link identifier.

[0184] The safety operation information corresponding to the target link identifier includes cumulative loading and unloading intensity and remaining storage. Remaining storage represents the storage remaining after the chemical coal allocation is performed on the link corresponding to the target link identifier. Safety operation conditions include loading and unloading peak thresholds. and warehouse safety minimum .

[0185] Specifically, if the cumulative loading and unloading intensity is less than or equal to the peak loading and unloading threshold... And the remaining storage is greater than or equal to the storage safety lower limit. If the safety operation information corresponding to the target link identifier meets the safety operation conditions, then the constraint coefficient corresponding to the target link identifier is set to 1; otherwise, it is set to 0.

[0186] Step 3: The product of the risk controllability correction factor, economic priority and constraint coefficient corresponding to each target link identifier is determined as the scheduling score of each target link identifier.

[0187] Step 4: Allocate coal for chemical production based on the scheduling score of each target link identifier.

[0188] In one embodiment, coal for chemical production is allocated based on the scheduling score of each target link identifier, including the following steps:

[0189] Step 41: Sort the scheduling scores of each target link identifier in descending order to obtain the sorted target link identifiers.

[0190] Step 42: Select the link corresponding to the first target link identifier among the sorted target link identifiers as the candidate link.

[0191] Step 43: In the case of transporting coal for chemical use based on candidate links within a rolling window, predict whether the available inventory of the boiler corresponding to the target link is lower than the safety threshold at the end of the rolling window.

[0192] The duration represented by the current time window is greater than the duration represented by the scrolling window.

[0193] Step 44: If yes, then the link corresponding to the second-ranked target link identifier in the sorted target link identifiers will be taken as the new candidate link, and the end time will be taken as the start time of the new rolling window.

[0194] Step 45: Return to the step of predicting whether the available inventory of the boiler corresponding to the candidate link is lower than the safety threshold at the end of the rolling window when transporting coal for chemical use based on the candidate link within the rolling window, until the allocation of coal for chemical use is completed.

[0195] By employing the method described in the above embodiments, and by considering the arrival probability and economic priority corresponding to each target link identifier, coal for chemical use can be allocated, thereby improving the accuracy of allocation to chemical users.

[0196] It is understood that the arrival probability of the target link identifier in this application can refer to the arrival probability of the target link identifier at the next moment. Therefore, in the actual chemical coal allocation process, within the daily production window, according to the rolling window method, the chemical coal allocation strategy can be adjusted in time at the next moment by predicting the arrival probability at the next moment, thereby improving the allocation accuracy.

[0197] In summary, such as Figure 5 As shown, a method for distributing coal for chemical use is provided. Taking the application of this method in a chemical coal distribution system as an example, the method may include the following steps:

[0198] S502, for multiple coal sources, obtains the link identifier between each coal source and multiple boilers, as well as the path information and coal blending information matching the link identifier; the path information includes multiple transportation points.

[0199] S504, based on the coal blending information corresponding to each link identifier, selects multiple target link identifiers from multiple link identifiers.

[0200] S506, for the coal blending information corresponding to the target link identifier, according to the preset process range corresponding to each of the multiple continuous fields in the coal blending information, obtain the dimensionless ratio value corresponding to each continuous field.

[0201] S508, sequentially concatenate each dimensionless ratio value with each discrete field in the coal blending information to obtain the first concatenation vector corresponding to the target link identifier.

[0202] S510, associate the target link identifier with the first concatenation vector to obtain the cross-domain feature corresponding to the target link identifier.

[0203] S512, based on the cross-domain characteristics, risk indicators and credibility indicators corresponding to the target link identifier, obtain the economic priority corresponding to the target link identifier.

[0204] S514, based on the economic priority and coal blending information corresponding to the target link identifier, establish a weighted logistics directed graph that matches multiple transportation points corresponding to the target link identifier.

[0205] S516, divide the weighted logistics directed graph into a first subgraph and a second subgraph; the edge attributes corresponding to the first edge between the first nodes in the first subgraph satisfy the preset conditions, and the edge attributes corresponding to the second edge between the second nodes in the second subgraph do not satisfy the preset conditions.

[0206] S518, based on the aggregation result of the first processing result of the target spatiotemporal graph convolutional network on the first subgraph and the second processing result of the target spatiotemporal graph convolutional network on the second subgraph, the arrival probability corresponding to the target link identifier is obtained.

[0207] S520 allocates coal for chemical production based on the arrival probability and economic priority corresponding to each target link identifier.

[0208] The contents of S502 to S520 can be referred to the aforementioned content adaptation description, and will not be repeated here.

[0209] Based on the above, the following experiments were conducted using the method provided in this application. The experimental details are as follows:

[0210] Since year M, a large coal chemical base in Province Z has frequently experienced cost overruns in procurement and material shortages in the furnace during peak production periods. The main reason is that the daily fluctuation of coal prices exceeds 8%, while the delays in transportation at ports, railways, and within the plant exhibit significant asynchronous characteristics. In year M+1, a scientific team conducted end-to-end verification using the methods provided in this application, with real-world operational data from February to April of year M+1 as a pilot project.

[0211] During the data acquisition phase, real-time price quotes for thermal coal and coking coal from the three major ports were captured from interfaces of five supplier traders, with a time resolution of 5 minutes. Simultaneously, data was accessed from various business systems to obtain transportation status data such as loading efficiency, actual train operation times, and stacking / reclaiming operation trajectories, covering a total of 142 MTR combined routes and 7 boiler groups with a daily average coal demand of 98,000 tons. All discrete fields were first mapped to unique codes using the master data dictionary. Continuous fields were then transformed into dimensionless values ​​according to process guidelines and concatenated with the original codes to generate a 42-dimensional cross-domain feature vector, which was then written to a Kafka stream.

[0212] Subsequently, the first-stage self-searching factor decomposition machine, using three-day price fluctuations, seven-day on-time delivery rates, and real-time coal quality parameters as inputs, completes the structural search and outputs a static procurement baseline within 20 minutes. The second-stage self-searching factor decomposition machine receives the embedded vectors from the previous stage, along with a three-dimensional interactive risk vector of cost, fluctuations, and on-time delivery rates. Through an interpretable linear synthesis strategy, it obtains a single economic priority and writes it into the economic priority weight matrix in real time. The system automatically calculates the bottleneck transfer ratio and inversely maps the economic priority to the edge weights of multiple topologies within ports, railways, and factories, forming a weighted directed logistics graph with dynamic time limits and economic edge weights.

[0213] Based on a weighted logistics directed graph, the first subgraph uses hourly spatiotemporal graph convolution to inject delay potential, while the second subgraph enables cost-sensitive gating at the minute-level granularity to weaken low-value link channels. The outputs of the two subgraphs are concatenated after node-level alignment to form the arrival probability and timeout warning vector. During the training phase, economic weighting is introduced to enhance the error, focusing on amplifying the gradient of high-value link prediction failures. The model converged after 12 hours of training on 300,000 historical samples. The system was officially launched in the pilot data center, with GPU real-time inference latency controlled within 2.3 seconds.

[0214] The system completed closed-loop scheduling of coal blending tasks for 7 furnace groups totaling 1.8 million tons, with the average cost per ton of coal purchased decreasing by 6.8% compared to the same period in M ​​years, equivalent to saving approximately 18.7 million yuan; the number of alarms when the safety stock in front of the furnace was below 24 hours decreased from an average of 17 times per month to 3 times, and there were no furnace shutdowns due to material shortages; the peak occupancy of the port yard decreased by 11%, the railway station delay time was shortened by 18%, and the shift output of loading and unloading machinery increased by 9.4%.

[0215] To quantify the performance of the method provided in this application, data from February to April of year M+1 were selected to compare the key metrics of the three schemes, as shown in Table 1.

[0216] Table 1

[0217]

[0218] In Table 1, the model framework of the method provided in this application is a combination of a Field-aware Factorization Machine (FFM) and a Spatio-Temporal Graph Convolutional Network (ST-GCN). Comparison Method 1 refers to the method combining static coal blending with rule-weighted logistics. Comparison Method 2 refers to the method combining an Independent Gradient Boosting Decision Tree (GBDT) cost model with a spatio-temporal graph convolutional network; the model framework used in Comparison Method 2 is a combination of GBDT and a single-granularity spatio-temporal graph convolutional network.

[0219] As can be seen from the table above, the method provided in this application outperforms the comparative methods in both the core indicators of procurement savings rate and material shortage control, balancing cost and risk within an acceptable inference latency. In particular, because the economic weights are deeply injected into the graph structure before the spatiotemporal graph convolution, the model can focus computational power on high-value links at an early stage of transportation bottlenecks, achieving dynamic tilting of scheduling resources.

[0220] Based on the above, it can be seen that the chemical coal utilization method provided in this application, compared with the existing schemes that separate static coal blending optimization and transportation delay prediction, first obtains a single economic priority coupled with cost, risk, and reliability through a self-searching factor decomposition mechanism, and obtains economic priority at the minute level. Then, through the inverse mapping of bottleneck transfer ratio, the economic priority is directly solidified onto the edge weights of the logistics graph. As a result, economic value is no longer superimposed ex-post in the form of rule thresholds, but has become a learnable core feature before the spatiotemporal graph convolution begins. This allows cost changes to propagate synchronously with real-time risks in ports, railways, and various sections within the plant in the same data structure, reducing the time lag between static daily planning and dynamic prediction.

[0221] At the model level, the first subgraph explicitly injects slow variable congestion by normalizing the delayed potential energy through the furnace time limit, while the second subgraph uses cost-sensitive gating to retain complete channels at high-quantile economic edge weights and implement weighted reduction on low-value links. This dual-granularity convolutional structure retains the capture of cross-port and railway congestion at the hourly scale while ensuring that computing power at the minute scale is concentrated on the shortest links that have the greatest impact on procurement costs, overcoming the problem of insufficient attention to critical links in traditional uniform-granularity convolution.

[0222] The economically weighted enhancement error introduced during the training phase amplifies the gradients of high-value samples with predicted failures and recalculates the weights in each batch, prompting the model to continuously correct the links that have the greatest impact on procurement expenditures. The scheduling end uses the same economic edge weight as the sovereign weight to run through the rolling queue process of loading, dispatching, unloading, and in-plant transfer. When a bottleneck triggers an enhancement or leakage mechanism, resource locking can be adjusted immediately without waiting for the next round of optimization or retraining. Through these structural and principle improvements, this application proposal achieves a deep coupling between static cost minimization and dynamic transportation risk while ensuring the process red lines for sulfur and ash, providing a faster-responding and more focused overall solution for multi-source coal blending scenarios.

[0223] In summary, this patent aims to solve the challenges of chemical coal blending, which involves multiple coal sources, drastic price fluctuations, and uncertainties in segmented transportation. The core approach is to first use a self-searching factor decomposition machine to quantify the static coal blending priority that meets the sulfur and ash red line and has the lowest cost into a single economic weight. Then, through the inverse mapping of bottleneck transfer ratios, this weight is written into the edge weights of the port, railway, and plant logistics graphs, enabling spatiotemporal graph convolution to simultaneously learn economic losses, transportation delays, and dynamic time limits within the same graph structure. Combined with dual-granularity convolution of the first and second subgraphs, cost-sensitive gating, and a gain gradient that strengthens the economic weights, computing power and loading / unloading resources can be continuously focused on the high-risk links that have the greatest impact on procurement costs within a minute-level rolling window.

[0224] If a third party intends to circumvent the method provided in this application, a common approach is to replace the self-searching factorization machine with a traditional gradient boosting tree or a deep multilayer perceptron, thereby avoiding features whose embedding dimensions and feature interaction terms are determined by Neural Architecture Search (NAS). However, such models struggle to maintain a unified vectorization of discrete labels and dimensionless proportions within the same sparse embedding space. Economic weights often rely on post-processing to be implemented, lacking an interpretable linear synthesis path. Under rapidly iterating market conditions, weight drift is prone to occur, leading to a disconnect between scheduling focus and actual costs.

[0225] Another potential workaround is to retain the graph convolution framework but add economic factors to the prediction output through external weighting or threshold filtering, instead of writing them into the edge weights. This seems to circumvent the feature of injecting economic weights into the graph structure, but the coupling between prediction and economics decreases, the delay potential cannot be synchronously modulated by the cost amplification function, the model lacks recursive information about economic value upstream of the bottleneck segment, and once cross-segment blockage occurs, additional heuristic corrections are often required, which limits both real-time performance and algorithm stability.

[0226] Furthermore, some manufacturers may attempt to replace two-level factorization and graph convolution with reinforcement learning as a whole, setting economic losses as negative rewards to directly learn end-to-end scheduling strategies. While theoretically this can cover the patent process, reinforcement learning often converges slowly in high-dimensional sparse state spaces, requiring a large amount of historical data and online trial-and-error opportunities. In scenarios such as loading and unloading, where deviations from the reward could result in actual material shortages, the cost of strategy exploration far exceeds the acceptable threshold, posing significant risks for engineering implementation.

[0227] Furthermore, some might treat the trunk line and intra-plant links as completely independent subgraphs, using multiple models trained in parallel to circumvent the cross-granularity alignment design in the patent. While this approach avoids dual-granularity convolution and node-level alignment, the two models lack a shared state vector, and economic weights cannot be uniformly measured between the trunk line and intra-plant. This necessitates manual trade-offs in resource scheduling at the decision-making level, forcing a break in the overall end-to-end closed loop and failing to achieve the real-time integrated performance of this patent.

[0228] Based on the above analysis, it is evident that the aforementioned alternatives have shortcomings in both model coupling depth and system real-time performance: either the economic weights are separated from transportation risks, making it impossible to accurately locate key bottlenecks; or the training and deployment costs increase significantly, making it difficult to meet minute-level rolling windows; or due to the lack of a unified metric standard, the scheduler needs to rely on additional rules, easily leading to multi-objective conflicts. These deficiencies make it difficult to achieve the same procurement savings and material shortage suppression effects as this patent in actual deployment, even if theoretically feasible. Therefore, although there are several workarounds, they either increase system complexity or weaken real-time performance and resource focusing capabilities, making it difficult to fully replace the method provided in this application in terms of performance, reliability, and commercial viability. The method provided in this application maintains its unique advantages due to its deep coupling of economic weights and the overall architecture of dual-granularity spatiotemporal convolution.

[0229] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0230] Based on the same inventive concept, this application also provides a chemical coal distribution device for implementing the above-mentioned chemical coal distribution method. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the chemical coal distribution device provided below can be found in the limitations of the chemical coal distribution method above, and will not be repeated here.

[0231] In one exemplary embodiment, such as Figure 6 As shown, a chemical coal distribution device is provided, comprising: an acquisition module 602, a screening module 604, an analysis module 606, a first processing module 608, a division module 610, a second processing module 612, and a distribution processing module 614, wherein:

[0232] The acquisition module 602 is used to obtain the link identifiers between each coal source and multiple boilers, as well as the path information and coal blending information matching the link identifiers, for multiple coal sources; the path information includes multiple transportation points; the filtering module 604 is used to filter multiple target link identifiers from the multiple link identifiers based on the coal blending information corresponding to each link identifier; the analysis module 606 is used to obtain the economic priority corresponding to each target link identifier based on the coal blending information, risk indicators, and credibility indicators corresponding to the target link identifier; the first processing module 608 is used to establish multiple transportation points corresponding to the target link identifier based on the economic priority and coal blending information corresponding to the target link identifier. A weighted directed graph of logistics with input point matching; a partitioning module 610, used to partition the weighted directed graph of logistics into a first subgraph and a second subgraph; the edge attributes corresponding to the first edge between the first nodes in the first subgraph meet the preset conditions, and the edge attributes corresponding to the second edge between the second nodes in the second subgraph do not meet the preset conditions; a second processing module 612, used to aggregate the first processing result of the target spatiotemporal graph convolutional network on the first subgraph and the second processing result of the target spatiotemporal graph convolutional network on the second subgraph to obtain the arrival probability of the target link identifier; an allocation processing module 614, used to allocate coal for chemical use based on the arrival probability and economic priority corresponding to each target link identifier.

[0233] In one embodiment, the analysis module is further configured to: for the coal blending information corresponding to the target link identifier, obtain the dimensionless ratio value corresponding to each of the multiple continuous fields in the coal blending information according to the preset process range corresponding to each of the multiple continuous fields in the coal blending information; sequentially concatenate each dimensionless ratio value with each discrete field in the coal blending information to obtain the first concatenation vector corresponding to the target link identifier; associate the target link identifier with the first concatenation vector to obtain the cross-domain feature corresponding to the target link identifier; and obtain the economic priority corresponding to the target link identifier based on the cross-domain feature, risk index, and credibility index corresponding to the target link identifier.

[0234] In one embodiment, the analysis module is further configured to: combine the cross-domain features, risk indicators, and credibility indicators corresponding to the target link identifier to obtain the link feature vector corresponding to the target link identifier; input the link feature vector corresponding to the target link identifier into the first-stage self-searching factor decomposition machine to obtain the procurement cost baseline corresponding to the target link identifier; sort the procurement cost baselines corresponding to multiple target link identifiers in ascending order to obtain the cost sequence number corresponding to the target link identifier; normalize and combine the cost sequence number, risk indicators, and credibility indicators corresponding to the target link identifier to obtain the three-dimensional interaction risk vector corresponding to the target link identifier; concatenate the three-dimensional interaction risk vector with the original feature embedding results corresponding to the cross-domain features to obtain the second concatenation vector, and input the second concatenation vector into the second-stage self-searching factor decomposition machine to obtain the comprehensive ranking weight corresponding to the target link identifier; and perform weighted summation processing on the procurement cost baseline, comprehensive ranking weight, and interaction risk vector contribution corresponding to the three-dimensional interaction risk vector corresponding to the target link identifier to obtain the economic priority corresponding to the target link identifier.

[0235] In one embodiment, the first processing module is further configured to: establish a directed graph of links matching multiple transportation points corresponding to the target link identifier according to the transportation direction between the coal source and the boiler corresponding to the target link identifier; for each transportation edge in the directed graph of links corresponding to the target link identifier, extract the real-time available throughput of the transportation edge at the target time from the coal matching information corresponding to the target link identifier; obtain the bottleneck transfer ratio corresponding to the transportation edge at the target time based on the real-time available throughput and the design throughput at the target time; obtain the economic edge weight corresponding to the transportation edge at the target time based on the bottleneck transfer ratio corresponding to the transportation edge at the target time and the economic priority corresponding to the target link identifier; combine the economic edge weight, bottleneck transfer ratio and arrival time limit of the transportation edge at the target time to obtain the edge attribute corresponding to the transportation edge; associate the edge attribute corresponding to the transportation edge with the transportation edge in the directed graph of links to obtain a weighted directed graph of logistics matching multiple transportation points corresponding to the target link identifier.

[0236] In one embodiment, the first processing result includes a first state vector corresponding to each first node in the first subgraph at the next time step of the target time; the second processing module is further configured to: insert an event-driven gating unit into each layer of the spatiotemporal graph convolutional network and inject tidal restrictions, temporary traffic restrictions, and equipment failure scenarios to obtain a target spatiotemporal graph convolutional network; for the first edge between adjacent first nodes in the first subgraph, obtain the queuing depth corresponding to the first edge at each time step, and obtain the delay potential energy corresponding to the first edge at the target time step based on the product between the queuing depth corresponding to the first edge at the target time step and the economic edge weight; for each first node in the first subgraph, determine a first upstream edge pointing to the first node from each first edge; determine the delay potential energy corresponding to the first upstream edge at the target time step, and obtain the time limit value after time limit normalization by the time limit corresponding to the furnace arrival time step of the first upstream edge; and obtain the first state vector corresponding to the first node at the next time step of the target time step based on the time limit value corresponding to the first upstream edge at the target time step.

[0237] In one embodiment, the second processing result includes a second state vector corresponding to each second node in the second subgraph at the next time step of the target time; the second processing module is further configured to: for the second edge between adjacent second nodes in the second subgraph, determine the cost-sensitive gate corresponding to the second edge at the target time based on the relationship between the economic edge weight and the quantile rank threshold corresponding to the second edge at each time step; for each second node in the second subgraph, determine a second upstream edge pointing to the second node from each second edge; determine the queuing potential energy represented by the product of the queuing depth and economic edge weight corresponding to the second upstream edge at the target time and the scaling factor, and normalize it to a minute-level value according to the furnace arrival time limit corresponding to the second upstream edge; and obtain the second state vector corresponding to the second node at the next time step of the target time based on the minute-level value of the second upstream edge corresponding to the target time and the cost-sensitive gate.

[0238] In one embodiment, the second processing module is further configured to: based on the path information corresponding to the target link identifier, filter the boundary nodes associated with the first subgraph and the second subgraph from each first node in the first subgraph and each second node in the second subgraph; perform time alignment and splicing processing on the first state vector corresponding to the first edge pointed to by the boundary node at the target time and the second state vector corresponding to the second edge pointed to by the boundary node to obtain the third spliced ​​vector corresponding to the boundary node at the target time; and based on the third spliced ​​vector corresponding to the boundary node at the target time, obtain the arrival probability of the target link identifier at the next time.

[0239] In one embodiment, the allocation processing module is further configured to: determine a risk controllability correction factor that is positively correlated with the arrival probability and remaining arrival time of the target link identifier; determine a constraint coefficient corresponding to the target link identifier based on the matching result between the safety operation information and safety operation conditions corresponding to the target link identifier within the current time window; determine the scheduling score of each target link identifier by multiplying the risk controllability correction factor, economic priority, and constraint coefficient corresponding to each target link identifier; and allocate coal for chemical use based on the scheduling scores of each target link identifier.

[0240] In one embodiment, the allocation processing module is further configured to: sort the scheduling scores of each target link identifier in descending order to obtain sorted target link identifiers; select the link corresponding to the first target link identifier among the sorted target link identifiers as a candidate link; predict whether the available inventory of the boiler corresponding to the target link at the end of the rolling window is lower than the safety threshold when transporting chemical coal based on the candidate link within the rolling window; if so, select the link corresponding to the second target link identifier among the sorted target link identifiers as a new candidate link, and use the end time as the start time of the new rolling window; return to the step of predicting whether the available inventory of the boiler corresponding to the candidate link at the end of the rolling window is lower than the safety threshold when transporting chemical coal based on the candidate link within the rolling window, until the allocation of chemical coal is completed.

[0241] Each module in the aforementioned coal distribution device for chemical use can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0242] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data related to the chemical coal distribution process. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a chemical coal distribution method.

[0243] Those skilled in the art will understand that Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0244] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0245] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0246] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0247] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0248] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0249] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0250] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for distributing coal for chemical industry, characterized in that, The method includes: For multiple coal sources, obtain the link identifier between each coal source and multiple boilers, as well as the path information and coal blending information matching the link identifier; the path information includes multiple transportation points; Based on the coal blending information corresponding to each of the aforementioned link identifiers, multiple target link identifiers are selected from the multiple link identifiers; For each target link identifier, the economic priority corresponding to the target link identifier is obtained based on the coal blending information, risk indicators, and credibility indicators corresponding to the target link identifier. Based on the economic priority and coal blending information corresponding to the target link identifier, a weighted directed logistics graph matching multiple transportation points corresponding to the target link identifier is established. The weighted logistics directed graph is divided into a first subgraph and a second subgraph; the edge attributes corresponding to the first edge between the first nodes in the first subgraph satisfy the preset conditions, and the edge attributes corresponding to the second edge between the second nodes in the second subgraph do not satisfy the preset conditions. Based on the aggregation result of the first processing result of the target spatiotemporal graph convolutional network on the first subgraph and the second processing result of the target spatiotemporal graph convolutional network on the second subgraph, the arrival probability corresponding to the target link identifier is obtained. Based on the arrival probability and economic priority corresponding to each target link identifier, coal for chemical use is allocated.

2. The method according to claim 1, characterized in that, The step of obtaining the economic priority corresponding to the target link identifier based on the coal blending information, risk indicators, and reliability indicators corresponding to the target link identifier includes: For the coal blending information corresponding to the target link identifier, according to the preset process range corresponding to each of the multiple continuous fields in the coal blending information, the dimensionless ratio value corresponding to each of the continuous fields is obtained. The dimensionless ratio values ​​are sequentially concatenated with the discrete fields in the coal blending information to obtain the first concatenation vector corresponding to the target link identifier. Associating the target link identifier with the first concatenation vector yields the cross-domain feature corresponding to the target link identifier; Based on the cross-domain characteristics, risk indicators, and credibility indicators corresponding to the target link identifier, the economic priority corresponding to the target link identifier is obtained.

3. The method according to claim 2, characterized in that, The step of obtaining the economic priority corresponding to the target link identifier based on the cross-domain characteristics, risk indicators, and credibility indicators corresponding to the target link identifier includes: The cross-domain features, risk indicators, and credibility indicators corresponding to the target link identifier are combined to obtain the link feature vector corresponding to the target link identifier; The link feature vector corresponding to the target link identifier is input into the first-stage self-searching factor decomposition machine to obtain the procurement cost baseline corresponding to the target link identifier; The procurement cost baselines corresponding to each of the multiple target link identifiers are sorted in ascending order to obtain the cost sequence number corresponding to the target link identifier. The cost sequence number, risk index and credibility index corresponding to the target link identifier are normalized and combined to obtain the three-dimensional interaction risk vector corresponding to the target link identifier. The three-dimensional interactive risk vector is concatenated with the original feature embedding result corresponding to the cross-domain feature to obtain a second concatenated vector. The second concatenated vector is then input into the second-stage self-searching factor decomposition machine to obtain the comprehensive ranking weight corresponding to the target link identifier. The economic priority corresponding to the target link identifier is obtained by weighted summation of the procurement cost baseline, the comprehensive ranking weight, and the interaction risk vector contribution corresponding to the three-dimensional interaction risk vector.

4. The method according to claim 1, characterized in that, The step of establishing a weighted directed logistics graph matching multiple transportation points corresponding to the target link identifier based on the economic priority and coal blending information corresponding to the target link identifier includes: According to the transportation direction between the coal source and the boiler corresponding to the target link identifier, establish a directed link graph that matches multiple transportation points corresponding to the target link identifier; For each transport edge in the directed graph corresponding to the target link identifier, the real-time available throughput of the transport edge at the target time is extracted from the coal distribution information corresponding to the target link identifier. Based on the real-time available throughput and designed throughput at the target time, the bottleneck transfer ratio of the transport edge at the target time is obtained; Based on the bottleneck transfer ratio of the transport edge at the target time and the economic priority of the target link identifier, the economic edge weight of the transport edge at the target time is obtained. By combining the economic edge weight, the bottleneck transfer ratio, and the furnace arrival time limit corresponding to the transport edge at the target time, the edge attribute corresponding to the transport edge is obtained. Associating the edge attributes corresponding to the transportation edge with the transportation edge in the directed link graph yields a weighted logistics directed graph that matches multiple transportation points corresponding to the target link identifier.

5. The method according to claim 1, characterized in that, The first processing result includes the first state vector of each first node in the first subgraph at the next time step corresponding to the target time step; the method further includes: By inserting event-driven gating units into each layer of the spatiotemporal graph convolutional network and injecting conditions such as tidal restrictions, temporary traffic restrictions, and equipment failures, the target spatiotemporal graph convolutional network is obtained. The processing method of the target spatiotemporal graph convolutional network for the first subgraph includes: For the first edge between adjacent first nodes in the first subgraph, the queuing depth of the first edge at the target time is obtained, and the delay potential energy of the first edge at the target time is obtained based on the product between the queuing depth of the first edge at the target time and the economic edge weight. For each first node in the first subgraph, determine a first upstream edge pointing to the first node from each of the first edges; Determine the delayed potential energy corresponding to the first upstream edge at the target time, and obtain the time limit value after time limit normalization by the time limit of the furnace corresponding to the first upstream edge; Based on the time limit value corresponding to the first upstream edge at the target time, the first state vector corresponding to the first node at the next time after the target time is obtained.

6. The method according to claim 5, characterized in that, The second processing result includes the second state vector corresponding to each second node in the second subgraph at the next time step after the target time step; the processing method of the target spatiotemporal graph convolutional network for the second subgraph includes: For the second edge between adjacent second nodes in the second subgraph, based on the relationship between the economic edge weight and the quantile rank threshold of the second edge at the target time, the cost-sensitive gate corresponding to the second edge at the target time is determined. For each second node in the second subgraph, determine a second upstream edge pointing to the second node from each of the second edges; Determine the queuing potential energy represented by the product of the queuing depth and economic edge weight of the second upstream edge at the target time and the scaling factor, and normalize it to a minute-level value according to the furnace arrival time limit corresponding to the second upstream edge. Based on the minute-level value of the second upstream edge at the target time and the cost-sensitive gate, the second state vector of the second node at the next time after the target time is obtained.

7. The method according to claim 6, characterized in that, The aggregation result of the first processing result of the target spatiotemporal graph convolutional network on the first subgraph and the second processing result of the target spatiotemporal graph convolutional network on the second subgraph yields the arrival probability corresponding to the target link identifier, including: Based on the path information corresponding to the target link identifier, the boundary nodes that are associated with the first subgraph and the second subgraph are filtered from each first node in the first subgraph and each second node in the second subgraph. The first state vector corresponding to the first side pointed to by the boundary node at the target time and the second state vector corresponding to the second side pointed to by the boundary node are time-aligned and spliced ​​to obtain the third spliced ​​vector corresponding to the boundary node at the target time. Based on the third concatenation vector corresponding to the boundary node at the target time, the arrival probability of the target link identifier at the next time is obtained.

8. The method according to claim 1, characterized in that, The allocation of coal for chemical production based on the arrival probability and economic priority corresponding to each target link identifier includes: Determine a risk controllability correction factor that is positively correlated with the furnace arrival probability and the remaining furnace arrival time corresponding to the target link identifier; Based on the matching result between the safety operation information and safety operation conditions corresponding to the target link identifier within the current time window, the constraint coefficient corresponding to the target link identifier is determined. The product of the risk controllability correction factor, economic priority, and constraint coefficient corresponding to each of the target link identifiers is determined as the scheduling score of each target link identifier. Based on the scheduling score of each target link identifier, coal for chemical production is allocated.

9. The method according to claim 8, characterized in that, The allocation of coal for chemical production based on the scheduling score of each target link identifier includes: The scheduling scores of each target link identifier are sorted in descending order to obtain the sorted target link identifiers. The link corresponding to the first target link identifier among the sorted target link identifiers is taken as the candidate link. In the case of transporting coal for chemical use based on the candidate link within a rolling window, predict whether the available inventory of the boiler corresponding to the target link is lower than the safety threshold at the end of the rolling window. If so, the link corresponding to the second-ranked target link identifier in the sorted target link identifiers will be taken as the new candidate link, and the end time will be taken as the start time of the new rolling window. Returning to the step of predicting whether the available inventory of the boiler corresponding to the candidate link is lower than the safety threshold at the end of the rolling window when transporting coal for chemical use based on the candidate link within the rolling window, until the allocation of coal for chemical use is completed.

10. A coal distribution device for chemical industry, characterized in that, The device includes: The acquisition module is used to obtain, for multiple coal sources, the link identifier between each coal source and multiple boilers, as well as the path information and coal blending information matching the link identifier; the path information includes multiple transportation points; The filtering module is used to filter multiple target link identifiers from multiple link identifiers based on the coal blending information corresponding to each link identifier; The analysis module is used to obtain the economic priority corresponding to each target link identifier based on the coal blending information, risk indicators and credibility indicators corresponding to the target link identifier. The first processing module is used to establish a weighted logistics directed graph that matches multiple transportation points corresponding to the target link identifier based on the economic priority and coal blending information corresponding to the target link identifier. The partitioning module is used to divide the weighted logistics directed graph into a first subgraph and a second subgraph; the edge attributes corresponding to the first edge between the first nodes in the first subgraph meet the preset conditions, and the edge attributes corresponding to the second edge between the second nodes in the second subgraph do not meet the preset conditions. The second processing module is used to obtain the arrival probability corresponding to the target link identifier based on the aggregation result of the first processing result of the target spatiotemporal graph convolutional network on the first subgraph and the second processing result of the target spatiotemporal graph convolutional network on the second subgraph. The allocation processing module is used to allocate coal for chemical production based on the arrival probability and economic priority corresponding to each target link identifier.

Images