A kind of automatic warehouse management system of metering collection fuel of separate furnace

Abstract

The application discloses a kind of separate stove metering collection fuel automatic warehouse-out management system, including following module: data acquisition module, for generating original running dataset;Data processing module, for forming boiler time series running data;Boiler behavior modeling module, for dynamic phase division, generates boiler behavior enhancement data;State feature coding module, for constructing boiler state consumption feature set;Fuel consumption prediction module, for based on the improved TabNet prediction model of introduction flow layer mechanism, construct feature flow path, output fuel consumption prediction result;Warehouse-out determination module, for generating warehouse-out determination result;Automatic warehouse-out execution module, for generating fuel warehouse-out instruction, execute fuel warehouse-out operation;ERP integrated module, for generating warehouse-out business data and completing inventory data update.The application realizes the reliable prediction of boiler fuel consumption and automatic warehouse-out decision by dynamic phase division and improved TabNet prediction model.

Description

Technical Field

[0001] This invention relates to the field of industrial automation control technology, and in particular to a fuel metering and automatic fuel dispensing management system for individual furnaces. Background Technology

[0002] As large industrial enterprises increasingly demand higher precision and automation in boiler fuel management, traditional fuel dispensing methods relying on manual recording and experience-based judgment are no longer sufficient to meet the dual requirements of efficiency and safety in modern fuel distribution. In the industrial sector, multiple boilers typically operate simultaneously to meet energy needs. Traditional fuel dispensing management relies heavily on manual operation, involving manual recording of fuel consumption for each boiler, judgment of the required fuel quantity based on experience, and manual operation of dispensing equipment. This method lacks the ability to model the evolution of boiler operating conditions and struggles to respond promptly to changes in fuel demand.

[0003] In existing systems, fuel consumption prediction typically relies on static characteristics or linear trend inferences, failing to incorporate time-series modeling mechanisms. This results in a weak ability to perceive boiler load fluctuations and operational phase transitions. Furthermore, the models are highly sensitive to state disturbances during actual deployment, leading to unstable prediction results and frequent misjudgments. They also lack interpretability and credibility feedback mechanisms. In addition, current systems generally do not quantify the stability or credibility of prediction results, which may result in erroneous issuance of warehousing instructions even under conditions of prediction fluctuations or sufficient raw materials, increasing material waste and equipment execution risks.

[0004] Therefore, how to provide a separate furnace metering and automatic fuel dispensing management system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an automatic fuel dispensing management system based on boiler metering and data acquisition. This invention introduces an improved TabNet prediction model with a flow stratification mechanism, combined with enhanced boiler behavior data and state feature vectors, to achieve accurate prediction of boiler fuel consumption. It also establishes a joint judgment mechanism for prediction stability index and credibility to generate fuel dispensing instructions. This system has the advantages of stable prediction, timely response, and high dispensing security, and can effectively cope with the prediction challenges caused by frequent fluctuations in boiler operating status, thereby improving the automation and intelligence level of fuel management.

[0006] According to an embodiment of the present invention, a separate furnace metering and automatic fuel dispensing management system includes the following modules:

[0007] The data acquisition module is used to collect multi-source operating data from the fuel delivery system, and add boiler and time identifiers to generate the original operating dataset.

[0008] The data processing module is used to perform time alignment processing on the original operating dataset and group the data based on the boiler identifier to form boiler time-series operating data corresponding to each boiler.

[0009] The boiler behavior modeling module is used to dynamically divide the phases based on the boiler time-series operation data, determine the switching points of each phase, construct the operation phase, generate the boiler behavior status information corresponding to the operation phase, and splice it with the boiler time-series operation data to generate boiler behavior enhancement data.

[0010] The status feature encoding module is used to perform feature encoding processing based on the enhanced boiler behavior data to construct a boiler status consumption feature set.

[0011] The fuel consumption prediction module is used to input the boiler state consumption feature set into the improved TabNet prediction model that introduces the flow layer mechanism, construct feature flow paths according to preset flow direction rules between multiple decision steps, and process the feature information step by step along the feature flow path, and output the fuel consumption prediction results of each boiler within the preset prediction time window.

[0012] The outbound judgment module is used to record the characteristic flow path formed during the fuel consumption prediction process, calculate the prediction stability index, generate the prediction confidence identifier, and jointly judge the prediction confidence identifier with the available fuel balance of the corresponding boiler to generate the outbound judgment result.

[0013] The automatic fuel delivery execution module is used to generate a fuel delivery instruction based on the delivery determination result and the corresponding boiler fuel consumption prediction result, and to control the fuel delivery equipment to perform the fuel delivery operation according to the fuel delivery instruction.

[0014] The ERP integration module is used to receive the actual outbound data generated by the automatic outbound execution module, generate the corresponding outbound business data, and complete the inventory data update.

[0015] A method for automatic fuel dispensing management based on fuel metering data from different furnaces, according to an embodiment of the present invention, includes the following steps:

[0016] Step 1: Collect multi-source operating data from the fuel delivery system, and add boiler identifiers and time identifiers to the multi-source operating data to generate the original operating dataset;

[0017] Step 2: Perform time alignment processing on the original operating dataset, and group the data based on the boiler identifier to form boiler time-series operating data corresponding to each boiler;

[0018] Step 3: Based on the boiler time-series operation data, perform dynamic phase division, determine the switching points of each phase, construct the operation phase, generate the boiler behavior status information corresponding to the operation phase, and splice it with the boiler time-series operation data to generate boiler behavior enhancement data;

[0019] Step 4: Perform feature encoding processing based on the enhanced boiler behavior data to construct a boiler state consumption feature set;

[0020] Step 5: Input the boiler state consumption feature set into the improved TabNet prediction model that introduces the flow layer mechanism, construct feature flow paths between multiple decision steps according to the preset flow direction rules, and process the feature information step by step along the feature flow path to output the fuel consumption prediction results of each boiler within the preset prediction time window.

[0021] Step 6: Record the characteristic flow paths formed during the fuel consumption prediction process, calculate the prediction stability index, generate a prediction confidence identifier corresponding to the fuel consumption prediction result, and make a joint judgment based on the prediction confidence identifier and the available fuel balance of the corresponding boiler to generate an outbound judgment result.

[0022] Step 7: Based on the outbound determination result and the corresponding boiler fuel consumption prediction result, generate a fuel outbound instruction;

[0023] Step 8: Control the fuel delivery equipment to perform the fuel outbound operation according to the fuel outbound instruction, and send the actual outbound data to the ERP integration module to generate outbound business data and complete the inventory data update.

[0024] Optionally, the multi-source operating data specifically includes the cumulative metering data of the coal feeder corresponding to each boiler, the real-time material level data of the coal bunker, and the operating load data.

[0025] Optionally, step two specifically includes:

[0026] For each type of data in the original running dataset, the corresponding time identifier field is extracted, and the time format is standardized and uniformly converted into minute-level time granularity. A continuous standard time axis is constructed based on whole minutes to generate the initial time series.

[0027] For data sampled from multiple data sources at the same time point, the multi-value samples are compressed by taking the arithmetic mean within each minute corresponding to the standard time axis to obtain a unique corresponding value; for time points in the initial time series with missing data, the sliding linear interpolation method is applied to complete the time axis and generate a complete standard time series.

[0028] Based on the boiler identification field contained in each data entry, the original operating dataset is grouped according to the boiler identification. The cumulative metering data of the coal feeder, the real-time material level data of the coal bunker, and the operating load data corresponding to each boiler are then spliced ​​together according to the complete standard time series to form boiler time-series operating data with time as the index, indicators as column fields, and boilers as the data grouping unit.

[0029] Optionally, step three specifically includes:

[0030] Based on the boiler time-series operation data and using the complete standard time series as the time reference, differential calculations are performed on the cumulative metering data of the coal feeder, the real-time material level data of the coal bunker, and the operating load data of each boiler at adjacent time points to obtain the corresponding coal consumption change sequence, material level change sequence, and load change sequence. The coal consumption change sequence represents the change in the cumulative metering data of the coal feeder at adjacent time points, the material level change sequence represents the change in the real-time material level data of the coal bunker at adjacent time points, and the load change sequence represents the change in the operating load data at adjacent time points.

[0031] In the coal consumption change sequence, when the coal consumption change at three consecutive time points is greater than a preset first change threshold, the third time point is determined as the coal consumption change trigger phase switching point; in the load change sequence, when the load change at three consecutive time points is greater than a preset second change threshold, the third time point is determined as the load change trigger phase switching point; when the direction of change in the material level change sequence reverses, the corresponding time point is determined as the material level change trigger phase switching point.

[0032] The time points determined by the phase switching points triggered by changes in coal consumption, load, and material level are merged, and the time interval between adjacent phase switching points is divided into an operating phase. The continuous operation process of the boiler is divided into multiple operating phases arranged in chronological order.

[0033] Within each operating phase, the cumulative phase change and average phase change rate of the cumulative metering data of the coal feeder, the cumulative phase change and average phase change rate of the real-time material level data of the coal bunker, and the cumulative phase change and average phase change rate of the operating load data are calculated respectively. The cumulative phase change and average phase change rate are combined to generate the boiler behavior status information corresponding to the operating phase.

[0034] The boiler behavior status information corresponding to the operating phase is mapped to all time points within the operating phase interval, and then spliced ​​with the boiler time-series operating data at the corresponding time points to generate enhanced boiler behavior data.

[0035] Optionally, step four specifically includes:

[0036] Based on the enhanced boiler behavior data, extract the boiler behavior status information corresponding to each time point.

[0037] For each boiler, feature encoding is performed on the boiler behavior status information to obtain the cumulative phase change and average phase change rate calculated based on the cumulative metering data of the coal feeder, the real-time material level data of the coal bunker, and the operating load data. The cumulative phase change and average phase change rate of each phase are then uniformly converted into a state feature vector of a set dimension.

[0038] Arrange the state feature vectors corresponding to each boiler in chronological order to generate boiler state consumption features for each boiler, and construct a boiler state consumption feature set.

[0039] Optionally, step five specifically includes:

[0040] The boiler state consumption feature set is used as the prediction input, where the boiler state consumption features corresponding to each boiler constitute a time series arranged in chronological order; each time point in the time series is predicted sequentially, so that the state feature vector corresponding to each time point is used as an independent prediction starting point and input into the improved TabNet prediction model that introduces the flow layer mechanism, thus forming a fuel consumption prediction process that is updated over time.

[0041] The improved TabNet prediction model consists of an input layer, multiple decision steps arranged in sequence, and an output layer. Each decision step is structurally identical and connected sequentially. Each decision step includes a feature transformation layer and a flow layer. The flow layer is located at the output of the feature transformation layer of the corresponding decision step and serves as the unique feature transfer node between the decision step and the next decision step with the adjacent number.

[0042] In each prediction process, the input layer inputs the state feature vector corresponding to the current time point to the feature transformation layer of the first decision step. The feature transformation layer performs linear transformation, Tanh activation and Sigmoid activation on the state feature vector in sequence to obtain the intermediate feature representation of the first decision step, and writes the intermediate feature representation of the first decision step into the flow layer corresponding to the first decision step.

[0043] Between the nth decision step and the (n+1th)th decision step, the feature transformation layer of the (n+1th)th decision step reads only the currently retained intermediate feature representation from the flow layer corresponding to the nth decision step, and filters out a set of valid intermediate feature representations from the read intermediate feature representations according to a preset flow rule; the set of valid intermediate feature representations is concatenated and combined with the state feature vector corresponding to the current time point, and used as the input of the feature transformation layer of the (n+1th)th decision step to obtain the intermediate feature representation of the (n+1th)th decision step, and written into the flow layer corresponding to the (n+1th)th decision step, so that the intermediate feature representation evolves step by step in multiple decision steps; where n represents the decision step number;

[0044] For each intermediate feature representation, a set of decision step numbers is determined to be included in the set of effective intermediate feature representations, and the set of decision step numbers constitutes the feature flow path of the intermediate feature representation.

[0045] After completing the feature transfer of all decision steps, the intermediate feature representation retained in the last decision step flow layer is input to the output layer to predict the fuel consumption within a set time interval extending into the future from the current time point. The set time interval constitutes a preset prediction time window, and the corresponding boiler fuel consumption prediction result within the preset prediction time window is output.

[0046] Optionally, the preset flow rules specifically include: recording the initial decision step number where each intermediate feature representation is generated; setting a maximum decision step span for each intermediate feature representation to participate in feature transformation; if the difference between the current decision step number and the initial decision step number is not greater than the maximum decision step span, the corresponding intermediate feature representation is included in the set of valid intermediate feature representations; otherwise, the corresponding intermediate feature representation is removed from the flow layer.

[0047] Optionally, step six specifically includes:

[0048] In the process of fuel consumption prediction, the feature flow path of the intermediate feature representation formed by each decision step in the improved TabNet prediction model is recorded, and the initial decision step number, the set of decision step numbers that actually participate in feature transformation and the feature transfer length corresponding to each intermediate feature representation are obtained. The feature transfer length is defined as the number of decision steps that the intermediate feature representation actually participates in feature transformation.

[0049] Based on the aforementioned characteristic flow path, all intermediate feature representations involved in generating fuel consumption prediction results are statistically analyzed to obtain the total number of intermediate feature representations. The feature transfer length of each intermediate feature representation is calculated, and the arithmetic mean of the feature transfer lengths of each intermediate feature representation is obtained as a prediction stability index.

[0050] The predicted stability index is associated with the fuel consumption prediction result to generate a prediction confidence identifier: when the predicted stability index is not less than a set stability threshold, the corresponding fuel consumption prediction result is marked as a high confidence prediction result; when the predicted stability index is less than the set stability threshold, the corresponding fuel consumption prediction result is marked as a low confidence prediction result.

[0051] Obtain real-time coal bunker level data for each boiler, and based on the real-time coal bunker level data, obtain the available fuel balance of the corresponding boiler at the current time point;

[0052] The prediction confidence indicator is jointly judged with the available fuel balance of the corresponding boiler. When the fuel consumption prediction result of the corresponding boiler is marked as a high confidence prediction result, and the available fuel balance is less than the predicted fuel consumption demand calculated based on the fuel consumption prediction result within a preset prediction time window, the corresponding boiler's outbound judgment result is generated.

[0053] Optionally, step seven specifically includes:

[0054] When the out-of-warehouse determination result indicates that the corresponding boiler needs to be refueled, a fuel out-of-warehouse instruction is generated based on the fuel consumption prediction result of the corresponding boiler. The fuel out-of-warehouse instruction includes the target boiler identifier, the amount of fuel to be out-of-warehouse, and the out-of-warehouse time window.

[0055] The target boiler identifier is used to indicate the specific boiler that needs to be refueled.

[0056] The amount of fuel dispensed is consistent with the predicted fuel consumption demand within the preset prediction time window as indicated by the fuel consumption prediction result.

[0057] The outbound time window is consistent with the preset prediction time window corresponding to the fuel consumption prediction result, and is used to limit the execution time range of the fuel outbound operation.

[0058] The beneficial effects of this invention are:

[0059] This invention introduces an improved TabNet prediction model with a flow layer mechanism, combined with dynamic modeling of boiler operating phase state characteristics and fuel consumption behavior, to construct a high-precision fuel consumption prediction framework with historical state memory and current state response capabilities. The boiler state consumption feature set is progressively transferred between multiple decision steps through feature flow paths, and an independent flow layer is set at each decision step to record intermediate feature representations and their initial decision step numbers, realizing structured management of multi-step feature evolution. In feature flow path control, by setting a maximum decision step span, the lifecycle of intermediate feature representations participating in feature transformation is limited, effectively suppressing early state interference and improving the stage sensitivity and response speed of the prediction results.

[0060] In terms of prediction stability assessment, the system calculates the feature transfer length of all intermediate features involved in generating the current prediction result, extracts their continuous participation in feature transformation within the model, and constructs a prediction stability index based on the average feature transfer length, thereby forming a high / low confidence level judgment for fuel consumption prediction results. The outbound judgment module jointly judges the prediction confidence level with the real-time coal bunker level data, and triggers the outbound command only when the prediction result is stable and the available fuel balance is insufficient to cover the demand within the prediction time window, significantly improving the accuracy and safety of outbound judgment.

[0061] Ultimately, based on highly reliable prediction results, the system automatically generates fuel delivery instructions that include the target boiler identifier, the amount of fuel to be delivered, and the delivery time window. This drives the fuel delivery equipment to accurately execute the delivery operation. The system also synchronizes delivery business data with inventory information through the ERP integration module, realizing a prediction-driven delivery control closed loop. This improves the intelligence, real-time performance, and business integration of the delivery process, making it suitable for the refined fuel management and intelligent scheduling needs in scenarios with multiple boilers operating in parallel. Attached Figure Description

[0062] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0063] Figure 1 This is a schematic diagram of the automatic fuel dispensing management system for separate furnace metering and collection proposed in this invention;

[0064] Figure 2 This is an overall flowchart of a method for automatic fuel dispensing management based on the metering and collection of fuel in a separate furnace, as proposed in this invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0066] refer to Figure 1 A separate furnace metering and automatic fuel dispensing management system includes the following modules:

[0067] The data acquisition module is used to collect multi-source operating data from the fuel delivery system, and add boiler and time identifiers to generate the original operating dataset.

[0068] The data processing module is used to perform time alignment processing on the original operating dataset and group the data based on the boiler identifier to form boiler time-series operating data corresponding to each boiler.

[0069] The boiler behavior modeling module is used to dynamically divide the phases based on the boiler time-series operation data, determine the switching points of each phase, construct the operation phase, generate the boiler behavior status information corresponding to the operation phase, and splice it with the boiler time-series operation data to generate boiler behavior enhancement data.

[0070] The status feature encoding module is used to perform feature encoding processing based on the enhanced boiler behavior data to construct a boiler status consumption feature set.

[0071] The fuel consumption prediction module is used to input the boiler state consumption feature set into the improved TabNet prediction model that introduces the flow layer mechanism, construct feature flow paths according to preset flow direction rules between multiple decision steps, and process the feature information step by step along the feature flow path, and output the fuel consumption prediction results of each boiler within the preset prediction time window.

[0072] The outbound judgment module is used to record the characteristic flow path formed during the fuel consumption prediction process, calculate the prediction stability index, generate the prediction confidence identifier, and jointly judge the prediction confidence identifier with the available fuel balance of the corresponding boiler to generate the outbound judgment result.

[0073] The automatic fuel delivery execution module is used to generate a fuel delivery instruction based on the delivery determination result and the corresponding boiler fuel consumption prediction result, and to control the fuel delivery equipment to perform the fuel delivery operation according to the fuel delivery instruction.

[0074] The ERP integration module is used to receive the actual outbound data generated by the automatic outbound execution module, generate the corresponding outbound business data, and complete the inventory data update.

[0075] refer to Figure 2 A method for automatic fuel dispensing management based on fuel metering in separate furnaces includes the following steps:

[0076] Step 1: Collect multi-source operating data from the fuel delivery system, and add boiler identifiers and time identifiers to the multi-source operating data to generate the original operating dataset;

[0077] Step 2: Perform time alignment processing on the original operating dataset, and group the data based on the boiler identifier to form boiler time-series operating data corresponding to each boiler;

[0078] Step 3: Based on the boiler time-series operation data, perform dynamic phase division, determine the switching points of each phase, construct the operation phase, generate the boiler behavior status information corresponding to the operation phase, and splice it with the boiler time-series operation data to generate boiler behavior enhancement data;

[0079] Step 4: Perform feature encoding processing based on the enhanced boiler behavior data to construct a boiler state consumption feature set;

[0080] Step 5: Input the boiler state consumption feature set into the improved TabNet prediction model that introduces the flow layer mechanism, construct feature flow paths between multiple decision steps according to the preset flow direction rules, and process the feature information step by step along the feature flow path to output the fuel consumption prediction results of each boiler within the preset prediction time window.

[0081] Step 6: Record the characteristic flow paths formed during the fuel consumption prediction process, calculate the prediction stability index, generate a prediction confidence identifier corresponding to the fuel consumption prediction result, and make a joint judgment based on the prediction confidence identifier and the available fuel balance of the corresponding boiler to generate an outbound judgment result.

[0082] Step 7: Based on the outbound determination result and the corresponding boiler fuel consumption prediction result, generate a fuel outbound instruction;

[0083] Step 8: Control the fuel delivery equipment to perform the fuel outbound operation according to the fuel outbound instruction, and send the actual outbound data to the ERP integration module to generate outbound business data and complete the inventory data update.

[0084] In this embodiment, the multi-source operating data specifically includes the cumulative metering data of the coal feeder corresponding to each boiler, the real-time material level data of the coal bunker, and the operating load data.

[0085] In this embodiment, step two specifically includes:

[0086] For each type of data in the original running dataset, the corresponding time identifier field is extracted, and the time format is standardized and uniformly converted into minute-level time granularity. A continuous standard time axis is constructed based on whole minutes to generate the initial time series.

[0087] For data sampled from multiple data sources at the same time point, the multi-value samples are compressed by taking the arithmetic mean within each minute corresponding to the standard time axis to obtain a unique corresponding value; for time points in the initial time series with missing data, the sliding linear interpolation method is applied to complete the time axis and generate a complete standard time series.

[0088] Based on the boiler identification field contained in each data entry, the original operation dataset is grouped according to the boiler identification, and the cumulative metering data of the coal feeder, the real-time material level data of the coal bunker, and the operating load data corresponding to each boiler are spliced ​​together according to the complete standard time series to form boiler time-series operation data with time as index, indicators as column fields, and boiler as data grouping unit.

[0089] In this invention, by uniformly parsing and standardizing the time signatures of data from different sources in the original operating dataset, various sampling times are uniformly mapped to minute-level time granularity. A continuous standard time axis is constructed with 1 minute as the basic time unit. The standard time axis covers the entire time interval during the continuous operation of the boiler, ensuring that data with different sampling frequencies can be aligned under a unified time reference.

[0090] When multiple sampling data exist for the same boiler within the same minute, the arithmetic mean is used to compress and fuse the multi-value samples to avoid interference from instantaneous fluctuations in the analysis. For time points with missing data in the time axis, the sliding linear interpolation method is used to complete the data based on the data of the adjacent valid time points before and after. The interpolation window does not exceed two adjacent time units to ensure the continuity and physical rationality of data changes.

[0091] After time alignment and completion, the data is grouped according to the boiler identifier, and the cumulative metering data of the coal feeder, the real-time material level data of the coal bunker and the operating load data of each boiler are spliced ​​together according to the complete standard time series to form boiler time series operation data with consistent structure and continuous time, providing a stable data foundation for phase division and predictive modeling.

[0092] In this embodiment, step three specifically includes:

[0093] Based on the boiler time-series operation data and using the complete standard time series as the time reference, differential calculations are performed on the cumulative metering data of the coal feeder, the real-time material level data of the coal bunker, and the operating load data of each boiler at adjacent time points to obtain the corresponding coal consumption change sequence, material level change sequence, and load change sequence. The coal consumption change sequence represents the change in the cumulative metering data of the coal feeder at adjacent time points, the material level change sequence represents the change in the real-time material level data of the coal bunker at adjacent time points, and the load change sequence represents the change in the operating load data at adjacent time points.

[0094] In the coal consumption change sequence, when the coal consumption change at three consecutive time points is greater than a preset first change threshold, the third time point is determined as the coal consumption change trigger phase switching point; in the load change sequence, when the load change at three consecutive time points is greater than a preset second change threshold, the third time point is determined as the load change trigger phase switching point; when the direction of change in the material level change sequence reverses, the corresponding time point is determined as the material level change trigger phase switching point.

[0095] The time points determined by the phase switching points triggered by changes in coal consumption, load, and material level are merged, and the time interval between adjacent phase switching points is divided into an operating phase. The continuous operation process of the boiler is divided into multiple operating phases arranged in chronological order.

[0096] Within each operating phase, the cumulative phase change and average phase change rate of the cumulative metering data of the coal feeder, the cumulative phase change and average phase change rate of the real-time material level data of the coal bunker, and the cumulative phase change and average phase change rate of the operating load data are calculated respectively. The cumulative phase change and average phase change rate are combined to generate the boiler behavior status information corresponding to the operating phase.

[0097] The boiler behavior status information corresponding to the operating phase is mapped to all time points within the operating phase interval, and then spliced ​​with the boiler time-series operating data at the corresponding time points to generate enhanced boiler behavior data.

[0098] During the dynamic phase division process, independent thresholds are set for different change sequences. Preferably, the first change threshold is set to 5% to 10% of the change in the coal feeder per minute, and the second change threshold is set to 3% to 8% of the boiler's rated load. When three consecutive time points meet the corresponding threshold conditions, the third time point is used as the phase switching point to eliminate interference caused by instantaneous fluctuations. For changes in coal bunker level, the substantial change in operating status is identified by detecting the reversal of the positive and negative signs of the change direction.

[0099] Within each operating phase, the cumulative phase change and average phase change rate of the three types of operating data are calculated to characterize the overall characteristics of boiler operation within that phase. Based on this, phase-level boiler behavior status information is generated. This boiler behavior status information is mapped to all time points within the operating phase according to the operating phase interval, so that the boiler time-series operating data at each time point is associated with a clear behavior status identifier, thereby forming boiler behavior enhancement data containing operating data and behavior status, providing a stable and structured input basis for predictive modeling.

[0100] In this embodiment, step four specifically includes:

[0101] Based on the enhanced boiler behavior data, extract the boiler behavior status information corresponding to each time point.

[0102] For each boiler, feature encoding is performed on the boiler behavior status information to obtain the cumulative phase change and average phase change rate calculated based on the cumulative metering data of the coal feeder, the real-time material level data of the coal bunker, and the operating load data. The cumulative phase change and average phase change rate of each phase are then uniformly converted into a state feature vector of a set dimension.

[0103] Arrange the state feature vectors corresponding to each boiler in chronological order to generate boiler state consumption features for each boiler and construct a boiler state consumption feature set.

[0104] In the feature encoding process, the cumulative phase change and average phase change rate corresponding to each boiler are uniformly mapped and preferably converted into a 16-dimensional state feature vector to ensure the consistency of the feature structure between different boilers. The state feature vectors corresponding to each boiler are arranged in chronological order to form the boiler state consumption features of each boiler. The boiler state consumption features of each boiler are combined to form the boiler state consumption feature set, providing a structurally unified feature basis for improving the input of the TabNet prediction model.

[0105] In this embodiment, step five specifically includes:

[0106] The boiler state consumption feature set is used as the prediction input, where the boiler state consumption features corresponding to each boiler constitute a time series arranged in chronological order; each time point in the time series is predicted sequentially, so that the state feature vector corresponding to each time point is used as an independent prediction starting point and input into the improved TabNet prediction model that introduces the flow layer mechanism, thus forming a fuel consumption prediction process that is updated over time.

[0107] The improved TabNet prediction model consists of an input layer, multiple decision steps arranged in sequence, and an output layer. Each decision step is structurally identical and connected sequentially. Each decision step includes a feature transformation layer and a flow layer. The flow layer is located at the output of the feature transformation layer of the corresponding decision step and serves as the unique feature transfer node between the decision step and the next decision step with the adjacent number.

[0108] In each prediction process, the input layer inputs the state feature vector corresponding to the current time point to the feature transformation layer of the first decision step. The feature transformation layer performs linear transformation, Tanh activation and Sigmoid activation on the state feature vector in sequence to obtain the intermediate feature representation of the first decision step, and writes the intermediate feature representation of the first decision step into the flow layer corresponding to the first decision step.

[0109] Between the nth decision step and the (n+1th)th decision step, the feature transformation layer of the (n+1th)th decision step reads only the currently retained intermediate feature representation from the flow layer corresponding to the nth decision step, and filters out a set of valid intermediate feature representations from the read intermediate feature representations according to a preset flow rule; the set of valid intermediate feature representations is concatenated and combined with the state feature vector corresponding to the current time point, and used as the input of the feature transformation layer of the (n+1th)th decision step to obtain the intermediate feature representation of the (n+1th)th decision step, and written into the flow layer corresponding to the (n+1th)th decision step, so that the intermediate feature representation evolves step by step in multiple decision steps; where n represents the decision step number;

[0110] For each intermediate feature representation, a set of decision step numbers is determined to be included in the set of effective intermediate feature representations, and the set of decision step numbers constitutes the feature flow path of the intermediate feature representation.

[0111] After completing the feature transfer for all decision steps, the intermediate feature representations retained in the last decision step's flow layer are input to the output layer. This input layer predicts fuel consumption over a set time interval extending into the future from the current time point. This set time interval constitutes a preset prediction time window, and the output layer outputs the corresponding boiler's fuel consumption prediction result within this window. The output layer employs a multi-layer feedforward neural network structure, including a feature aggregation unit, a fully connected mapping unit, and a prediction output unit connected sequentially. The feature aggregation unit unifies and organizes multiple intermediate feature representations from the flow layer, forming a fixed-dimensional comprehensive feature representation through vector concatenation to retain the state information of each intermediate feature in different decision steps. The fully connected mapping unit performs a linear mapping operation on the comprehensive feature representation to establish a mapping relationship between the intermediate feature representations and fuel consumption. It also adjusts the mapping result using a nonlinear activation function to enhance the model's ability to express nonlinear consumption patterns. The prediction output unit is set as a linear output structure, with its output dimension corresponding to the fuel consumption within the preset prediction time window. It generates a fuel consumption prediction result covering the prediction time window, starting from the current time point. Through the above structure, the output layer can effectively convert the intermediate feature representations obtained from the multi-decision-step evolution into fuel consumption prediction values ​​with actual physical meaning, providing a reliable basis for the outbound judgment; as time progresses, the above prediction processing is repeated at subsequent time points, so that the fuel consumption prediction results are continuously updated as the boiler operating status changes.

[0112] In this embodiment, the preset flow direction rule specifically includes: recording the initial decision step number where each intermediate feature representation is generated; setting the maximum decision step span allowed to participate in feature transformation for each intermediate feature representation; if the difference between the current decision step number and the initial decision step number is not greater than the maximum decision step span, the corresponding intermediate feature representation is included in the set of valid intermediate feature representations; otherwise, the corresponding intermediate feature representation is removed from the flow layer.

[0113] This invention introduces an improved TabNet prediction model with a flow layer mechanism in the fuel consumption prediction stage. Through structural improvements, it addresses the technical challenge of traditional prediction models simultaneously considering historical state memory and current state sensitivity. Unlike existing TabNet prediction models that only transfer features between adjacent decision steps, this invention sets up an independent flow layer in each decision step to store intermediate feature representations generated by different decision steps. The lifecycle of these intermediate feature representations is controlled by preset flow rules. Specifically, each intermediate feature representation records its initial decision step number at the time of generation, and a maximum decision step span is set as the effective transfer range. Preferably, the maximum decision step span is set to four decision steps, thereby limiting intermediate features to participate in feature transformation only within a limited decision step interval, avoiding long-term interference from early state features in subsequent predictions.

[0114] During the feature transfer process, each decision step concatenates the boiler state consumption features corresponding to the current time point with the intermediate feature representation set that is still within the effective life cycle. This allows the improved TabNet prediction model to enhance its state evolution modeling capabilities by introducing limited historical decision information while maintaining a high responsiveness to the current operating state.

[0115] By applying the aforementioned structural constraints, the improved TabNet prediction model can maintain prediction stability across consecutive time points during the rolling prediction process, while significantly reducing the risks of state drift and prediction lag. This enhances the adaptability of fuel consumption prediction results to actual boiler operation changes and provides a reliable data foundation for automatic fuel dispensing decisions.

[0116] In this embodiment, step six specifically includes:

[0117] In the process of fuel consumption prediction, the feature flow path of the intermediate feature representation formed by each decision step in the improved TabNet prediction model is recorded, and the initial decision step number, the set of decision step numbers that actually participate in feature transformation and the feature transfer length corresponding to each intermediate feature representation are obtained. The feature transfer length is defined as the number of decision steps that the intermediate feature representation actually participates in feature transformation.

[0118] Based on the aforementioned characteristic flow path, all intermediate feature representations involved in generating the fuel consumption prediction result are statistically analyzed to obtain the total number of intermediate feature representations. The feature transfer length of each intermediate feature representation is then calculated, and the arithmetic mean of the feature transfer lengths of each intermediate feature representation is taken as the prediction stability index. Specifically, for all intermediate feature representations involved in generating the current fuel consumption prediction result, the number of decision steps in which they actually participate in feature transformation in the decision step sequence is counted. The sum of these decision step numbers is then divided by the total number of intermediate feature representations to obtain the value. The prediction stability index is used to characterize the degree to which intermediate feature representations continuously participate in feature transformation in multiple decision steps during the prediction process.

[0119] The predicted stability index is associated with the fuel consumption prediction result to generate a prediction confidence identifier: when the predicted stability index is not less than a set stability threshold, the corresponding fuel consumption prediction result is marked as a high confidence prediction result; when the predicted stability index is less than the set stability threshold, the corresponding fuel consumption prediction result is marked as a low confidence prediction result.

[0120] Obtain real-time coal bunker level data for each boiler, and based on the real-time coal bunker level data, obtain the available fuel balance of the corresponding boiler at the current time point;

[0121] The prediction confidence indicator is jointly judged with the available fuel balance of the corresponding boiler. When the fuel consumption prediction result of the corresponding boiler is marked as a high confidence prediction result, and the available fuel balance is less than the predicted fuel consumption demand calculated based on the fuel consumption prediction result within a preset prediction time window, a departure determination result for the corresponding boiler is generated. Specifically, the fuel consumption prediction result is the predicted fuel consumption value covering each time point within the preset prediction time window, starting from the current time point. The predicted fuel consumption values ​​within the preset prediction time window are accumulated to obtain the predicted fuel consumption demand of the corresponding boiler within the preset prediction time window. By summarizing the overall consumption within the prediction time window, a fuel demand benchmark for departure determination can be obtained.

[0122] In this invention, a prediction reliability determination mechanism directly related to fuel consumption prediction results is constructed by recording and analyzing the feature flow trajectory within the prediction model. Specifically, during each fuel consumption prediction process, the system tracks the intermediate feature representations generated in each decision step of the improved TabNet prediction model, records its initial decision step number, the set of decision step numbers participating in feature transformation, and the corresponding feature transfer length. The feature transfer length reflects the degree to which a single intermediate feature continuously participates in decision-making within the model. Furthermore, all intermediate feature representations involved in generating the current fuel consumption prediction result are statistically analyzed, and the feature transfer length of each intermediate feature representation is calculated. The arithmetic mean of all feature transfer lengths is used as the prediction stability index. The larger the value of the prediction stability index, the more consistent the feature support for the current prediction result is formed by the model in multiple decision steps, thus resulting in higher prediction stability.

[0123] Preferably, the prediction stability threshold is set to 3. When the prediction stability index is not less than 3, the corresponding fuel consumption prediction result is marked as a high-confidence prediction result; otherwise, it is marked as a low-confidence prediction result. By jointly judging the prediction confidence index with the real-time material level data of the boiler coal bunker, the fuel delivery decision is triggered only when the prediction result is stable and the current available fuel balance is insufficient to cover the predicted fuel consumption demand within the preset prediction time window. This effectively avoids erroneous delivery operations due to unstable predictions or sufficient material levels, improves the safety and reliability of fuel delivery decisions, and provides a reliable data foundation for subsequent automatic delivery control.

[0124] In this embodiment, step seven specifically includes:

[0125] When the out-of-warehouse determination result indicates that the corresponding boiler needs to be refueled, a fuel out-of-warehouse instruction is generated based on the fuel consumption prediction result of the corresponding boiler. The fuel out-of-warehouse instruction includes the target boiler identifier, the amount of fuel to be out-of-warehouse, and the out-of-warehouse time window.

[0126] The target boiler identifier is used to indicate the specific boiler that needs to be refueled.

[0127] The amount of fuel dispensed is consistent with the predicted fuel consumption demand within the preset prediction time window as indicated by the fuel consumption prediction result.

[0128] The outbound time window is consistent with the preset prediction time window corresponding to the fuel consumption prediction result, and is used to limit the execution time range of the fuel outbound operation.

[0129] Example 1:

[0130] To verify the feasibility of this invention in practice, it was applied to the fuel management system of a large thermal power plant. The aim was to solve the problems of untimely replenishment, frequent erroneous fuel releases, and unstable prediction results in the existing boiler fuel management. The boiler operating load fluctuates greatly, and fuel consumption has obvious phased and nonlinear characteristics. Traditional fixed threshold release logic and simple historical average prediction models cannot accurately reflect the future fuel demand of the boiler, resulting in frequent manual intervention and unnecessary release instructions, which greatly affects fuel allocation efficiency and inventory safety.

[0131] In this embodiment, the system first collects multi-source operating data from the coal feeders, coal bunker level sensors, and main control system of each boiler through the data acquisition module. This includes cumulative metering data of the coal feeders, real-time coal bunker level data, and operating load data. The data sampling frequency is uniformly once per minute. Subsequently, the data processing module performs time alignment processing on the data from different sources and groups the data according to the boiler identification field to generate structured boiler time-series operating data.

[0132] Through the boiler behavior modeling module, the system performs dynamic phase division based on the boiler time-series operation data, successfully identifying multiple representative operation phases. Each phase is mapped to boiler behavior enhancement data with boiler behavior state information, and then encoded by the state feature encoding module into a 16-dimensional state feature vector, which is then input to the fuel consumption prediction module.

[0133] To verify the effectiveness of the improved TabNet prediction model, the system compared the prediction performance of the traditional LSTM model, the existing TabNet prediction model, and the improved TabNet prediction model with the introduction of a flow layer mechanism proposed in this invention. Table 1 shows the comparison results of the three models in terms of fuel consumption prediction accuracy under the same dataset and prediction window (30 minutes).

[0134] Table 1. Comparison of the accuracy of different models in predicting boiler fuel consumption.

[0135] Predictive Model Prediction accuracy (%) Relative error mean (%) Maximum prediction deviation (%) LSTM neural network model 84.6 7.8 18.4 Existing TabNet prediction model 89.7 5.9 12.6 Improved TabNet prediction model 95.3 3.6 7.9

[0136] The data analysis in Table 1 shows that the improved TabNet prediction model significantly outperforms the LSTM neural network model and the existing TabNet prediction model in three key performance indicators: prediction accuracy, mean relative error, and maximum prediction bias. Specifically, the improved TabNet prediction model achieves a prediction accuracy of 95.3%, higher than the LSTM neural network model's 84.6% and the existing TabNet prediction model's 89.7%; the improved TabNet prediction model has a mean relative error of 3.6%, lower than the LSTM neural network model's 7.8% and the existing TabNet prediction model's 5.9%; and the improved TabNet prediction model's maximum prediction bias is also controlled at 7.9%, significantly better than the comparative models' 18.4% and 12.6%. This difference is mainly due to the improved TabNet prediction model structure introduced in this invention, which combines a flow layer mechanism with a characteristic flow path control strategy. By setting an independent flow layer for each decision step and limiting the maximum decision step span, a balance is effectively achieved between short-term focus and long-term dependence of feature information. This avoids the prediction drift problem caused by over-reliance on early state features in traditional models. At the same time, the improved TabNet prediction model filters and combines historical intermediate feature representations in each decision step, enabling the prediction process to maintain high sensitivity to the current operating state while integrating stable and reliable historical feature support. This significantly enhances the model's adaptability to nonlinear fluctuations in boiler fuel consumption. This structural improvement improves prediction stability while ensuring prediction response speed, ultimately achieving higher accuracy and lower error levels.

[0137] In addition, the system introduces a prediction reliability discrimination mechanism in the outbound judgment module. Fuel outbound judgment is triggered only when the fuel consumption prediction result of the corresponding boiler is marked as a high reliability prediction result and the available fuel balance is less than the predicted fuel consumption demand calculated based on the fuel consumption prediction result within a preset prediction time window. This strategy effectively filters out uncertain results in the model prediction and avoids erroneous outbound operations due to short-term anomalies or misjudgments. The automatic outbound execution module controls the fuel conveying equipment to complete the corresponding operation and feeds back the outbound results and actual inventory changes through the ERP integration module.

[0138] This embodiment achieves high-precision rolling prediction of boiler fuel consumption by improving the TabNet prediction model and combining it with a flow layer mechanism. The model constructs a multi-step decision structure combining a feature transformation layer and a flow layer, enabling the orderly transfer and filtering of intermediate feature representations across multiple decision steps, thus improving the stability of feature expression and the ability to model deep-level correlations. Simultaneously, the system records and analyzes the feature flow paths of intermediate features during the prediction process, forming prediction confidence indicators, and dynamically determines whether to generate a delivery instruction based on real-time coal bunker level data, effectively improving the system's intelligent decision-making capability and operational reliability.

[0139] In terms of outbound decision-making, the introduction of a predictive stability index to quantify the reliability of the prediction results helps the system make robust judgments on uncertainties in actual operation, thereby avoiding false or missed triggering of fuel replenishment operations. This invention not only improves the accuracy and stability of fuel prediction, but also enhances the autonomy and response efficiency of the fuel scheduling process, and has good engineering applicability and promotion value.

[0140] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A separate furnace metering and automatic fuel dispensing management system, characterized in that, Includes the following modules: The data acquisition module is used to collect multi-source operating data from the fuel delivery system, and add boiler and time identifiers to generate the original operating dataset. The data processing module is used to perform time alignment processing on the original operating dataset and group the data based on the boiler identifier to form boiler time-series operating data corresponding to each boiler. The boiler behavior modeling module is used to dynamically divide the phases based on the boiler time-series operation data, determine the switching points of each phase, construct the operation phase, generate the boiler behavior status information corresponding to the operation phase, and splice it with the boiler time-series operation data to generate boiler behavior enhancement data. The status feature encoding module is used to perform feature encoding processing based on the enhanced boiler behavior data to construct a boiler status consumption feature set. The fuel consumption prediction module is used to input the boiler state consumption feature set into the improved TabNet prediction model that introduces the flow layer mechanism, construct feature flow paths according to preset flow direction rules between multiple decision steps, and process the feature information step by step along the feature flow path, and output the fuel consumption prediction results of each boiler within the preset prediction time window. The outbound judgment module is used to record the characteristic flow path formed during the fuel consumption prediction process, calculate the prediction stability index, generate the prediction confidence identifier, and jointly judge the prediction confidence identifier with the available fuel balance of the corresponding boiler to generate the outbound judgment result. The automatic fuel delivery execution module is used to generate a fuel delivery instruction based on the delivery determination result and the corresponding boiler fuel consumption prediction result, and to control the fuel delivery equipment to perform the fuel delivery operation according to the fuel delivery instruction. The ERP integration module is used to receive the actual outbound data generated by the automatic outbound execution module, generate the corresponding outbound business data, and complete the inventory data update.

2. The automatic fuel dispensing management system for separate furnace metering and data acquisition according to claim 1, characterized in that, The modules are connected in the following way: Step 1: Collect multi-source operating data from the fuel delivery system, and add boiler identifiers and time identifiers to the multi-source operating data to generate the original operating dataset; Step 2: Perform time alignment processing on the original operating dataset, and group the data based on the boiler identifier to form boiler time-series operating data corresponding to each boiler; Step 3: Based on the boiler time-series operation data, perform dynamic phase division, determine the switching points of each phase, construct the operation phase, generate the boiler behavior status information corresponding to the operation phase, and splice it with the boiler time-series operation data to generate boiler behavior enhancement data; Step 4: Perform feature encoding processing based on the enhanced boiler behavior data to construct a boiler state consumption feature set; Step 5: Input the boiler state consumption feature set into the improved TabNet prediction model that introduces the flow layer mechanism, construct feature flow paths between multiple decision steps according to the preset flow direction rules, and process the feature information step by step along the feature flow path to output the fuel consumption prediction results of each boiler within the preset prediction time window. Step 6: Record the characteristic flow paths formed during the fuel consumption prediction process, calculate the prediction stability index, generate a prediction confidence identifier corresponding to the fuel consumption prediction result, and make a joint judgment based on the prediction confidence identifier and the available fuel balance of the corresponding boiler to generate an outbound judgment result. Step 7: Based on the outbound determination result and the corresponding boiler fuel consumption prediction result, generate a fuel outbound instruction; Step 8: Control the fuel delivery equipment to perform the fuel outbound operation according to the fuel outbound instruction, and send the actual outbound data to the ERP integration module to generate outbound business data and complete the inventory data update.

3. The automatic fuel dispensing management system for separate furnace metering and data acquisition according to claim 2, characterized in that, The multi-source operating data specifically includes the cumulative metering data of the coal feeder for each boiler, the real-time material level data of the coal bunker, and the operating load data.

4. The automatic fuel dispensing management system for separate furnace metering and data acquisition according to claim 2, characterized in that, Step two specifically involves: For each type of data in the original running dataset, the corresponding time identifier field is extracted, and the time format is standardized and uniformly converted into minute-level time granularity. A continuous standard time axis is constructed based on whole minutes to generate the initial time series. For data sampled from multiple data sources at the same time point, the multi-value samples are compressed by taking the arithmetic mean within each minute corresponding to the standard time axis to obtain a unique corresponding value; for time points in the initial time series with missing data, the sliding linear interpolation method is applied to complete the time axis and generate a complete standard time series. Based on the boiler identification field contained in each data entry, the original operating dataset is grouped according to the boiler identification. The cumulative metering data of the coal feeder, the real-time material level data of the coal bunker, and the operating load data corresponding to each boiler are then spliced ​​together according to the complete standard time series to form boiler time-series operating data with time as the index, indicators as column fields, and boilers as the data grouping unit.

5. The automatic fuel dispensing management system for separate furnace metering and data acquisition according to claim 2, characterized in that, Step three specifically involves: Based on the boiler time-series operation data and using the complete standard time series as the time reference, differential calculations are performed on the cumulative metering data of the coal feeder, the real-time material level data of the coal bunker, and the operating load data of each boiler at adjacent time points to obtain the corresponding coal consumption change sequence, material level change sequence, and load change sequence. The coal consumption change sequence represents the change in the cumulative metering data of the coal feeder at adjacent time points, the material level change sequence represents the change in the real-time material level data of the coal bunker at adjacent time points, and the load change sequence represents the change in the operating load data at adjacent time points. In the coal consumption change sequence, when the coal consumption change at three consecutive time points is greater than a preset first change threshold, the third time point is determined as the coal consumption change trigger phase switching point; in the load change sequence, when the load change at three consecutive time points is greater than a preset second change threshold, the third time point is determined as the load change trigger phase switching point; when the direction of change in the material level change sequence reverses, the corresponding time point is determined as the material level change trigger phase switching point. The time points determined by the phase switching points triggered by changes in coal consumption, load, and material level are merged, and the time interval between adjacent phase switching points is divided into an operating phase. The continuous operation process of the boiler is divided into multiple operating phases arranged in chronological order. Within each operating phase, the cumulative phase change and average phase change rate of the cumulative metering data of the coal feeder, the cumulative phase change and average phase change rate of the real-time material level data of the coal bunker, and the cumulative phase change and average phase change rate of the operating load data are calculated respectively. The cumulative phase change and average phase change rate are combined to generate the boiler behavior status information corresponding to the operating phase. The boiler behavior status information corresponding to the operating phase is mapped to all time points within the operating phase interval, and then spliced ​​with the boiler time-series operating data at the corresponding time points to generate enhanced boiler behavior data.

6. The automatic fuel dispensing management system for separate furnace metering and data acquisition according to claim 2, characterized in that, Step four specifically involves: Based on the enhanced boiler behavior data, extract the boiler behavior status information corresponding to each time point. For each boiler, feature encoding is performed on the boiler behavior status information to obtain the cumulative phase change and average phase change rate calculated based on the cumulative metering data of the coal feeder, the real-time material level data of the coal bunker, and the operating load data. The cumulative phase change and average phase change rate of each phase are then uniformly converted into a state feature vector of a set dimension. Arrange the state feature vectors corresponding to each boiler in chronological order to generate boiler state consumption features for each boiler, and construct a boiler state consumption feature set.

7. The automatic fuel dispensing management system for separate furnace metering and data acquisition according to claim 2, characterized in that, Step five specifically involves: The boiler state consumption feature set is used as the prediction input, where the boiler state consumption features corresponding to each boiler constitute a time series arranged in chronological order; each time point in the time series is predicted sequentially, so that the state feature vector corresponding to each time point is used as an independent prediction starting point and input into the improved TabNet prediction model that introduces the flow layer mechanism, thus forming a fuel consumption prediction process that is updated over time. The improved TabNet prediction model consists of an input layer, multiple decision steps arranged in sequence, and an output layer. Each decision step is structurally identical and connected sequentially. Each decision step includes a feature transformation layer and a flow layer. The flow layer is located at the output of the feature transformation layer of the corresponding decision step and serves as the unique feature transfer node between the decision step and the next decision step with the adjacent number. In each prediction process, the input layer inputs the state feature vector corresponding to the current time point to the feature transformation layer of the first decision step. The feature transformation layer performs linear transformation, Tanh activation and Sigmoid activation on the state feature vector in sequence to obtain the intermediate feature representation of the first decision step, and writes the intermediate feature representation of the first decision step into the flow layer corresponding to the first decision step. Between the nth decision step and the (n+1th)th decision step, the feature transformation layer of the (n+1th)th decision step reads only the currently retained intermediate feature representation from the flow layer corresponding to the nth decision step, and filters out a set of valid intermediate feature representations from the read intermediate feature representations according to a preset flow rule; the set of valid intermediate feature representations is concatenated and combined with the state feature vector corresponding to the current time point, and used as the input of the feature transformation layer of the (n+1th)th decision step to obtain the intermediate feature representation of the (n+1th)th decision step, and written into the flow layer corresponding to the (n+1th)th decision step, so that the intermediate feature representation evolves step by step in multiple decision steps; where n represents the decision step number; For each intermediate feature representation, a set of decision step numbers is determined to be included in the set of effective intermediate feature representations, and the set of decision step numbers constitutes the feature flow path of the intermediate feature representation. After completing the feature transfer of all decision steps, the intermediate feature representation retained in the last decision step flow layer is input to the output layer to predict the fuel consumption within a set time interval extending into the future from the current time point. The set time interval constitutes a preset prediction time window, and the corresponding boiler fuel consumption prediction result within the preset prediction time window is output.

8. The automatic fuel dispensing management system for separate furnace metering and data acquisition according to claim 7, characterized in that, The preset flow rules specifically include: recording the initial decision step number where each intermediate feature representation is generated; setting a maximum decision step span for each intermediate feature representation to participate in feature transformation; if the difference between the current decision step number and the initial decision step number is not greater than the maximum decision step span, the corresponding intermediate feature representation is included in the set of valid intermediate feature representations; otherwise, the corresponding intermediate feature representation is removed from the flow layer.

9. The automatic fuel dispensing management system for separate furnace metering and data acquisition according to claim 2, characterized in that, Step six specifically involves: In the process of fuel consumption prediction, the feature flow path of the intermediate feature representation formed by each decision step in the improved TabNet prediction model is recorded, and the initial decision step number, the set of decision step numbers that actually participate in feature transformation and the feature transfer length corresponding to each intermediate feature representation are obtained. The feature transfer length is defined as the number of decision steps that the intermediate feature representation actually participates in feature transformation. Based on the aforementioned characteristic flow path, all intermediate feature representations involved in generating fuel consumption prediction results are statistically analyzed to obtain the total number of intermediate feature representations. The feature transfer length of each intermediate feature representation is calculated, and the arithmetic mean of the feature transfer lengths of each intermediate feature representation is obtained as a prediction stability index. The predicted stability index is associated with the fuel consumption prediction result to generate a prediction confidence identifier: when the predicted stability index is not less than a set stability threshold, the corresponding fuel consumption prediction result is marked as a high confidence prediction result; when the predicted stability index is less than the set stability threshold, the corresponding fuel consumption prediction result is marked as a low confidence prediction result. Obtain real-time coal bunker level data for each boiler, and based on the real-time coal bunker level data, obtain the available fuel balance of the corresponding boiler at the current time point; The prediction confidence indicator is jointly judged with the available fuel balance of the corresponding boiler. When the fuel consumption prediction result of the corresponding boiler is marked as a high confidence prediction result, and the available fuel balance is less than the predicted fuel consumption demand calculated based on the fuel consumption prediction result within a preset prediction time window, the corresponding boiler's outbound judgment result is generated.

10. The automatic fuel dispensing management system for separate furnace metering and data acquisition according to claim 2, characterized in that, Step seven specifically involves: When the out-of-warehouse determination result indicates that the corresponding boiler needs to be refueled, a fuel out-of-warehouse instruction is generated based on the fuel consumption prediction result of the corresponding boiler. The fuel out-of-warehouse instruction includes the target boiler identifier, the amount of fuel to be out-of-warehouse, and the out-of-warehouse time window. The target boiler identifier is used to indicate the specific boiler that needs to be refueled. The amount of fuel dispensed is consistent with the predicted fuel consumption demand within the preset prediction time window as indicated by the fuel consumption prediction result. The outbound time window is consistent with the preset prediction time window corresponding to the fuel consumption prediction result, and is used to limit the execution time range of the fuel outbound operation.

Images