Intelligent factory manufacturing optimization method and system based on energy consumption prediction

Abstract

The invention provides an intelligent factory manufacturing optimization method and system based on energy consumption prediction, and the method comprises the steps: carrying out the time sequence structural processing of historical energy consumption record data continuously collected in a factory manufacturing process, and generating an energy consumption sequence unit with a continuous time stamp through time window division and data correlation verification; constructing a manufacturing process-energy consumption association knowledge graph based on the energy consumption sequence unit, and establishing a structured knowledge representation containing node attributes and edge relationship weights by identifying a causal association relationship between process execution nodes and energy consumption fluctuation features; calling a pre-trained graph neural network model to carry out time sequence evolution prediction processing on the graph, and generating an energy consumption prediction sequence of a future preset manufacturing cycle; and according to the energy consumption prediction sequence, performing dynamic scheduling optimization on the execution sequence of the current manufacturing process and the resource allocation scheme, and generating an energy consumption optimization oriented manufacturing execution instruction. The energy consumption management level and the resource utilization efficiency in the manufacturing process of the intelligent factory are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing, and more specifically, to an intelligent factory manufacturing optimization method and system based on energy consumption prediction. Background Technology

[0002] With the development of intelligent manufacturing technology, energy consumption forecasting, as a key aspect of manufacturing optimization, provides crucial information for manufacturing optimization decisions by predicting energy consumption trends in future manufacturing cycles. Currently, the common approach is to acquire historical energy consumption data from the factory, use time series forecasting models to extrapolate these data, generate energy consumption forecasts for future periods, and then make empirical adjustments to manufacturing processes or resource allocation based on these forecasts. However, because forecasting relies solely on the statistical regularities of historical energy consumption data, it fails to effectively extract the intrinsic correlation between the execution of manufacturing processes and energy consumption fluctuations. Consequently, the energy consumption forecasts often fail to accurately reflect the actual impact of process execution on energy consumption changes, and manufacturing optimization schemes based on such forecasts frequently lack specificity and precision. Summary of the Invention

[0003] This invention provides a smart factory manufacturing optimization method and system based on energy consumption prediction.

[0004] In a first aspect, embodiments of the present invention provide a smart factory manufacturing optimization method based on energy consumption prediction. The method includes: performing time-series structured processing on historical energy consumption records continuously collected during the factory manufacturing process; generating energy consumption sequence units with continuous time markers through time window division and data correlation verification; the energy consumption sequence units include the energy consumption fluctuation characteristics of the manufacturing process execution period and the corresponding period; constructing a manufacturing process-energy consumption association knowledge graph based on the energy consumption sequence units; establishing a structured knowledge representation including node attributes and edge relationship weights by identifying the causal relationship between process execution nodes and energy consumption fluctuation characteristics; calling a pre-trained graph neural network model to perform time-series evolution prediction processing on the manufacturing process-energy consumption association knowledge graph; generating an energy consumption prediction sequence for a future preset manufacturing cycle by combining the changing trends of energy consumption sequence units in historical periods; and dynamically scheduling and optimizing the execution order and resource allocation scheme of the current manufacturing process according to the energy consumption prediction sequence, generating an energy consumption optimization-oriented manufacturing execution instruction including process adjustment priority and resource allocation scheme.

[0005] Secondly, embodiments of the present invention provide a computer system, including: a memory storing a computer program; and a processor for loading the computer program to implement the intelligent factory manufacturing optimization method based on energy consumption prediction as described above.

[0006] This invention provides a smart factory manufacturing optimization method based on energy consumption prediction. By performing time-series structured processing on continuously collected historical energy consumption records during the factory manufacturing process, it generates energy consumption sequence units with continuous time stamps. This achieves precise binding of energy consumption data with the execution period of manufacturing processes, avoiding the problem of missing time-process correlation in traditional energy consumption data processing. By constructing a manufacturing process-energy consumption correlation knowledge graph based on energy consumption sequence units, and establishing a structured knowledge representation including node attributes and edge relationship weights, it can effectively extract the causal relationship between process execution nodes and energy consumption fluctuation characteristics. This overcomes the limitation of traditional statistical models that can only reflect surface correlations, providing interpretable correlation evidence for energy consumption prediction. A pre-trained graph neural network model is used to perform time-series evolution prediction processing on the manufacturing process-energy consumption correlation knowledge graph. Combined with the changing trends of energy consumption sequence units in historical periods, it generates an energy consumption prediction sequence for the future preset manufacturing cycle. This achieves effective integration of the structured correlation of the knowledge graph and the time-series changes of energy consumption data, improving the accuracy and reliability of energy consumption prediction. Based on the energy consumption prediction sequence, the execution order and resource allocation scheme of the current manufacturing process are dynamically scheduled and optimized, generating energy consumption optimization-oriented manufacturing execution instructions that include process adjustment priorities and resource allocation schemes. This ensures that the energy consumption prediction results can be directly transformed into specific operations at the manufacturing execution level, thereby improving the energy consumption management level and resource utilization efficiency of the smart factory manufacturing process. Attached Figure Description

[0007] Figure 1 This is a flowchart of a smart factory manufacturing optimization method based on energy consumption prediction provided by an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of the composition of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0009] Please see Figure 1 , Figure 1 A flowchart illustrating a smart factory manufacturing optimization method based on energy consumption prediction, provided as an embodiment of the present invention, is included. This method can be executed by a computer system and may include the following steps: Step S100: Perform time series structuring on the historical energy consumption records continuously collected during the factory manufacturing process, and generate energy consumption sequence units with continuous time markers by dividing the time window and verifying the data correlation. The energy consumption sequence units include the energy consumption fluctuation characteristics of the manufacturing process execution period and the corresponding period.

[0010] Historical energy consumption records are continuous energy consumption data collected by smart factories during the product manufacturing process. This data reflects the factory's energy consumption at different times and in different manufacturing stages. Time window segmentation involves dividing continuous historical energy consumption records into windows of fixed length according to certain rules, ensuring data continuity within each window. Data correlation verification checks whether there are reasonable logical relationships between data in adjacent windows, thereby ensuring data accuracy and reliability.

[0011] In one implementation, step S100 may specifically include the following steps S110 to S160: Step S110: Perform data cleaning on the historical energy consumption records continuously collected during the factory manufacturing process to remove non-numerical interference information from the historical energy consumption records. Perform timestamp calibration on the cleaned historical energy consumption records to ensure that the timestamp accuracy of all energy consumption records is consistent, and generate calibrated energy consumption data.

[0012] During data acquisition, various factors such as sensor malfunctions and data transmission errors may introduce non-numerical interference information, including text descriptions, symbols, or incorrect characters. This interference can severely impact subsequent data analysis and processing. Therefore, it is necessary to remove this non-numerical interference from historical energy consumption records. Timestamp calibration involves adjusting and unifying the timestamps of the cleaned data. The calibrated energy consumption data is accurate and consistent energy consumption data obtained after data cleaning and timestamp calibration.

[0013] Step S120: The calibration energy consumption data is divided into time windows using a sliding time window algorithm. The window size of the sliding time window is determined according to the typical execution time of the factory manufacturing process. The calibration energy consumption data is continuously extracted according to the preset window sliding step size to generate multiple initial window data blocks with time continuity.

[0014] The window size of the sliding time window is determined based on the typical execution duration of the factory manufacturing process. This ensures that the divided windows match the execution period of the manufacturing process, allowing the data within each window to better reflect the energy consumption of a complete manufacturing process. The preset window sliding step size determines the distance the window slides each time. By continuously extracting calibration energy consumption data according to this step size, multiple initial window data blocks with temporal continuity can be generated.

[0015] For example, different automotive manufacturing processes have different typical execution times. For instance, the welding process in automobiles typically requires a certain amount of time to complete. Based on this typical execution time, the size of the sliding time window is determined. Assuming the typical execution time of the welding process is several hours, the window size can be set to be close to this duration. The preset window sliding step size can be set according to actual needs, such as half the window size. When processing calibration energy consumption data, the window starts from the beginning of the data, extracting a window-sized data block at a time. Then, the window slides to the right by the preset step size, extracting another window-sized data block, and so on, until the entire calibration energy consumption data is covered. This ultimately generates multiple initial window data blocks with temporal continuity, providing the foundation for subsequent data verification and analysis.

[0016] Step S130: Perform data integrity verification on the initial window data blocks, count the proportion of effective energy consumption record data in each initial window data block, mark the initial window data blocks with the proportion of effective energy consumption record data below the preset integrity threshold as invalid data blocks and remove them, and retain the valid window data blocks.

[0017] Effective energy consumption records refer to normal data that accurately reflects actual energy consumption, excluding missing, erroneous, or outlier values. Statistical analysis of the percentage of effective energy consumption records in each initial window data block is used to assess the completeness of that window data block. A preset completeness threshold is a pre-defined standard used to determine the validity of a window data block. If the percentage of effective energy consumption records in an initial window data block is lower than this threshold, it indicates poor completeness of the window data block, potentially containing a large amount of missing or erroneous data. This block is marked as invalid and removed, leaving only valid window data blocks to ensure the quality of subsequent data analysis.

[0018] For example, the initial window data block may contain missing energy consumption records due to sensor malfunctions, data transmission interruptions, or other reasons. When performing data integrity verification, the range of valid energy consumption records must first be clearly defined; for example, data within a reasonable energy consumption range can be defined as valid data. Then, the number of valid data records in each initial window data block is counted, and their proportion of the total data volume in that window data block is calculated. If a high integrity threshold is preset, when the proportion of valid energy consumption records in a window data block falls below this threshold, that window data block is marked as invalid and removed. The remaining valid window data blocks will be used for further processing.

[0019] Step S140: Perform data correlation verification on adjacent valid window data blocks, calculate the average energy consumption change rate of adjacent valid window data blocks, and mark adjacent window pairs whose average energy consumption change rate exceeds the preset correlation threshold as abnormal correlation window pairs.

[0020] Data correlation verification checks whether the data between adjacent valid window data blocks conforms to normal logical correlation. The average energy consumption change rate is an important indicator for measuring the degree of energy consumption variation between adjacent window data blocks. By calculating the average energy consumption change rate of adjacent valid window data blocks, it can be determined whether their energy consumption variation is within a reasonable range. A preset correlation threshold is a pre-defined standard used to determine whether the correlation between adjacent window pairs is normal. If the average energy consumption change rate of adjacent valid window data blocks exceeds this threshold, it indicates abnormal energy consumption variation between the two window data blocks, which may indicate data acquisition errors, equipment malfunctions, or production process adjustments. These pairs are marked as having abnormal correlation and require further processing.

[0021] For example, adjacent valid window data blocks record energy consumption over different time periods. To calculate the rate of change of the average energy consumption of adjacent window data blocks, the average energy consumption of each window data block is calculated separately. Then, by comparing the average energy consumption of adjacent windows, their rate of change is calculated. For instance, if one window data block records energy consumption in the morning and another adjacent window data block records energy consumption in the afternoon, the rate of change of the average energy consumption of these two window data blocks is calculated. If the preset correlation threshold is within a reasonable range, when the calculated rate of change of the average energy consumption exceeds this range, these two window pairs are marked as abnormally correlated window pairs.

[0022] Step S150: Perform data interpolation repair processing on the correlation anomaly window pairs. Use cubic spline interpolation algorithm to smooth the missing or abnormal data points in the correlation anomaly window pairs and generate data blocks of continuous correlation repair windows.

[0023] Data interpolation repair is an operation that repairs missing or anomalous data points in a pair of correlated anomaly windows. Cubic spline interpolation achieves a smooth transition of data by constructing a cubic polynomial between data points. In a pair of correlated anomaly windows, data may be missing or anomalous for various reasons, affecting the continuity and accuracy of the data. Using cubic spline interpolation to process these missing or anomalous data points allows for smooth changes in the data at these points, avoiding abrupt changes. Ultimately, a continuous correlated window data block is generated, containing more accurate and continuous data that better reflects actual energy consumption.

[0024] Step S160: Arrange the valid window data blocks after data correlation verification and repair in chronological order, add corresponding time stamp information to each valid window data block, including the window start timestamp and the window end timestamp, and define the valid window data block after adding time stamp information as an energy consumption sequence unit.

[0025] The valid window data blocks, after undergoing data correlation verification and repair, are arranged sequentially by time to ensure the temporal continuity and order of the data. Time stamp information is used to clearly define the specific temporal position of each valid window data block, including the window start timestamp and window end timestamp. These two timestamps accurately define the time interval corresponding to that window data block. Adding time stamp information provides a clearer understanding of the time range of each window data block. The energy consumption sequence unit is the final data unit obtained after the above series of processes, containing the manufacturing process execution period and the corresponding energy consumption fluctuation characteristics, and also has clear time stamp information.

[0026] Step S200: Construct a knowledge graph of manufacturing process-energy consumption association based on energy consumption sequence units. By identifying the causal relationship between process execution nodes and energy consumption fluctuation characteristics, establish a structured knowledge representation that includes node attributes and edge relationship weights.

[0027] A manufacturing process-energy consumption association knowledge graph is a knowledge model used to represent the relationship between manufacturing processes and energy consumption. It displays process execution nodes, energy consumption fluctuation characteristics, and their relationships in graph form. Process execution nodes represent each specific process execution unit in the manufacturing process, with each node corresponding to a process. Energy consumption fluctuation characteristics reflect the changes in energy consumption during process execution, including characteristics of different stages such as energy consumption increase, decrease, and stabilization. Causal associations refer to the causal connections between process execution nodes and energy consumption fluctuation characteristics. Identifying these relationships allows for a deeper understanding of the energy consumption generation mechanisms in the manufacturing process. Node attributes describe characteristic information about process execution nodes, such as process name, execution time, and importance. Edge relationship weights are indicators that measure the strength of the association between nodes. By establishing a structured knowledge representation that includes node attributes and edge relationship weights, the relationship between manufacturing processes and energy consumption can be represented more accurately and comprehensively.

[0028] In one implementation, step S200 may specifically include the following steps S210 to S250: Step S210: Extract process execution nodes from the energy consumption sequence unit. Divide the process execution nodes into independent process execution nodes according to the start and end times of the manufacturing process. Each process execution node corresponds to a unique process identifier and time interval.

[0029] Process execution node extraction is the process of accurately identifying the execution nodes of each process from the energy consumption sequence unit. Dividing the manufacturing process into independent execution nodes based on its start and end times clarifies the specific execution interval of each process in time. Each execution node corresponds to a unique process identifier and time interval. The process identifier distinguishes different processes, while the time interval precisely defines the execution time range of that process.

[0030] Step S220: Perform fluctuation pattern classification processing on the energy consumption fluctuation characteristics in the energy consumption sequence unit, identify the energy consumption rising segment, stable segment and falling segment corresponding to different process execution nodes, and generate the time distribution curve of energy consumption fluctuation characteristics.

[0031] Fluctuation pattern classification is an operation that systematically classifies the energy consumption fluctuation characteristics within energy consumption sequence units, aiming to better analyze and understand the patterns of energy consumption changes. An energy consumption rise phase refers to a stage where energy consumption shows an upward trend over a certain period, such as due to the start of a process, equipment startup, or increased workload. A stable phase refers to a stage where energy consumption is relatively stable with minimal changes, possibly indicating that the process is operating normally and the equipment's energy consumption remains at a relatively stable level. A decline phase refers to a stage where energy consumption shows a downward trend, possibly caused by the completion of a process, equipment shutdown, or reduced workload. By identifying the energy consumption rise, stable, and decline phases corresponding to different process execution nodes, a detailed understanding of the energy consumption changes during the execution of each process can be obtained. Generating time distribution curves of energy consumption fluctuation characteristics visually displays the changes in energy consumption fluctuation characteristics over time, helping analysts to more clearly observe energy consumption fluctuation patterns.

[0032] Step S230: Calculate the temporal adjacency relationship between execution nodes of different processes, determine the sequential order and time interval of consecutive execution processes, and establish temporal association edges of process execution nodes.

[0033] Calculating the temporal adjacency relationships between execution nodes of different processes is to clarify the temporal sequence and intervals of each process. Determining the sequence and time intervals of consecutively executed processes provides a clear understanding of the manufacturing process flow and rhythm. Temporal association edges between process execution nodes represent the temporal order relationships between different process execution nodes. By establishing these edges, the temporal relationships between processes can be visually represented in a knowledge graph.

[0034] For example, different process execution nodes include stamping, welding, painting, and assembly. By analyzing the time information of these process execution nodes, their temporal adjacency relationships can be calculated. For instance, after the stamping process is completed, the welding process immediately follows. By determining the execution times of these two processes, the time interval between them can be calculated. In the knowledge graph, temporal association edges are established for these two process execution nodes. The direction of the edge indicates the sequence of processes, and the edge attributes can record information such as the time interval. By establishing these temporal association edges, the time sequence and flow of automobile manufacturing processes can be clearly displayed.

[0035] Step S240: Analyze the differences in energy consumption fluctuation characteristics of the same process execution node in different historical periods, and extract the mean change rate and peak occurrence time of energy consumption fluctuation characteristics as node attributes.

[0036] Analyzing the differences in energy consumption fluctuation characteristics of the same process execution node at different historical periods is to gain a deeper understanding of the energy consumption changes of that process over time. The mean change rate of energy consumption fluctuation characteristics is an indicator that measures the degree of change in the mean energy consumption over different periods, reflecting the stability and trend of process energy consumption. The peak occurrence time refers to the point at which energy consumption reaches its maximum value, and this point in time may be related to the operation of the process or the operating status of the equipment.

[0037] For example, for a specific process execution node, such as the painting process, we analyze its energy consumption fluctuation characteristics across different historical periods. By comparing energy consumption data from different time periods, we calculate the mean rate of change of energy consumption fluctuation characteristics. If the mean rate of change of energy consumption for the painting process is large within a certain period, it indicates that the energy consumption of this process is unstable, potentially due to equipment failure or process adjustments. Simultaneously, we record the times when energy consumption peaks occur in each historical period; for example, in the painting process, energy consumption may peak at a specific painting step. These mean rates of change and peak times are added to the knowledge graph as attributes of the painting process execution node to better analyze the energy consumption of this process.

[0038] Step S250: Based on the time distribution curve and node attributes of energy consumption fluctuation characteristics, calculate the correlation strength value between the process execution node and the energy consumption fluctuation characteristics through the causal relationship identification algorithm, and fill the edge attributes of the knowledge graph as the edge relation weight to obtain the manufacturing process-energy consumption correlation knowledge graph containing node attributes and edge relation weights.

[0039] Causal relationship identification algorithms are used to identify the causal relationships between process execution nodes and energy consumption fluctuation characteristics. By deeply analyzing the time distribution curves and node attributes of energy consumption fluctuation characteristics, the algorithm calculates the correlation strength value between process execution nodes and energy consumption fluctuation characteristics. The correlation strength value is an indicator that measures the tightness of the association between a node and a feature. Using it as an edge weight in the edge attributes of the knowledge graph can more accurately represent the correlation between nodes and features. The resulting manufacturing process-energy consumption correlation knowledge graph includes node attributes and edge weights, providing a more comprehensive and accurate reflection of the relationship between manufacturing processes and energy consumption.

[0040] For example, based on the time distribution curve and node attributes of energy consumption fluctuation characteristics, causal relationship identification algorithms, such as the Granger causality test algorithm, are used to calculate the correlation strength value between the process execution node and the energy consumption fluctuation characteristics. For instance, for the welding process execution node and the corresponding energy consumption increase characteristic, the correlation strength value is calculated by analyzing the node attributes such as the welding process execution time and equipment power, as well as the time distribution curve of the energy consumption increase segment. If the correlation strength value is high, it indicates a strong causal relationship between the welding process execution and the energy consumption increase characteristic. This correlation strength value is then used as the edge relation weight and filled into the attribute of the corresponding edge in the knowledge graph.

[0041] In another implementation, step S200 may specifically include the following steps S201 to S205: Step S201: Perform outlier filtering on the energy consumption sequence units, delete abnormal data points that exceed the preset energy consumption fluctuation range, and retain valid energy consumption sequence units that conform to the normal energy consumption characteristics of the process.

[0042] Outlier filtering is the process of screening and purifying data within energy consumption sequence units. During data acquisition and recording, some outlier data points may appear that exceed the normal range. These outlier data points may be caused by sensor malfunctions, data transmission errors, or external interference. A preset energy consumption fluctuation range is a reasonable range of energy consumption fluctuations pre-defined based on historical data and experience, used to determine whether a data point is normal. Deleting outlier data points that exceed the preset energy consumption fluctuation range, and retaining only valid energy consumption sequence units that conform to the normal energy consumption characteristics of the process, ensures the accuracy and reliability of the data. For example, an energy consumption sequence unit records the energy consumption of an automobile production process. Based on historical data and experience, a preset energy consumption fluctuation range is determined. For example, for a certain process, its normal energy consumption fluctuation should be within a relatively stable range. When performing outlier filtering on the energy consumption sequence unit, each data point is checked; if the energy consumption value of a data point exceeds the preset energy consumption fluctuation range, it is deleted as an outlier data point.

[0043] Step S202: Align the process execution time period and energy consumption value in the effective energy consumption sequence unit with the time axis to ensure that the time interval of each process execution node is completely matched with the collection time of the corresponding energy consumption fluctuation characteristics.

[0044] Time axis alignment is an operation that precisely aligns the execution time periods and energy consumption values ​​of processes within an effective energy consumption sequence unit on the time axis. In actual data acquisition and processing, the acquisition times of process execution periods and energy consumption fluctuation characteristics may not be completely consistent, which can affect the accurate analysis of the relationship between processes and energy consumption. Time axis alignment ensures that the time interval of each process execution node perfectly matches the acquisition time of the corresponding energy consumption fluctuation characteristic, guaranteeing accurate time correspondence of the data when analyzing the relationship between processes and energy consumption and avoiding time deviations.

[0045] For example, the effective energy consumption sequence unit records the energy consumption during different process execution periods. When performing timeline alignment, the time interval of each process execution node and the corresponding energy consumption value collection time must be carefully checked. For instance, the time interval of a certain process execution node might be from a certain time in the morning to a certain time in the afternoon, and the corresponding energy consumption value collection time should also be within this time period. If any time mismatch is found, adjustments need to be made to ensure complete consistency.

[0046] Step S203: Define the entity for the process execution node, and use the process identifier, time interval and node attributes as the basic description information of the entity to generate a set of nodes for the knowledge graph.

[0047] Defining process execution nodes as entities is the process of abstracting these nodes into entities within a knowledge graph. Process identifiers, time ranges, and node attributes serve as the basic descriptive information for each entity. Process identifiers distinguish different processes, time ranges define the execution timeframe, and node attributes describe characteristic information about the process, such as its importance and equipment usage. By using this information as the basic descriptive information for entities, a set of nodes in the knowledge graph is generated, providing the node foundation for constructing the knowledge graph.

[0048] Step S204: Calculate the similarity of energy consumption fluctuation characteristics of different process execution nodes, classify energy consumption fluctuation characteristics with similarity higher than a preset threshold into the same feature category, and establish the association edge between the feature category and the process execution node.

[0049] A preset threshold is a pre-defined standard used to determine whether two energy consumption fluctuation features are similar. Grouping energy consumption fluctuation features with a similarity higher than the preset threshold into the same feature category is for the purpose of classifying and managing energy consumption fluctuation features, facilitating the analysis and understanding of energy consumption relationships between different processes. Establishing association edges between feature categories and process execution nodes represents the relationship between feature categories and process execution nodes in the knowledge graph, thus more clearly demonstrating the energy consumption similarity between different processes.

[0050] For example, similarity calculations are performed on the energy consumption fluctuation characteristics of different process execution nodes. For instance, comparing the energy consumption fluctuation characteristics of the stamping and forging processes, by calculating their similarity (such as Euclidean distance, cosine similarity, etc.), it is found that their energy consumption fluctuation patterns are quite similar. A preset threshold is a reasonable standard; if the similarity of the energy consumption fluctuation characteristics of the two processes is higher than the preset threshold, their energy consumption fluctuation characteristics are classified into the same feature category. In the knowledge graph, association edges are established between this feature category and the execution nodes of the stamping and forging processes to represent the relationships between them.

[0051] Step S205: Set the initial value of the edge relationship weight according to the type of the associated edge, and adjust the edge relationship weight through an iterative optimization algorithm so that the association relationship between nodes and edges in the knowledge graph conforms to the actual fluctuation pattern of historical energy consumption data.

[0052] Setting initial values ​​for edge relationship weights based on the type of associated edges is to assign initial weights to the edges in the knowledge graph. Different types of associated edges reflect different relationships, thus requiring different initial weights. Iterative optimization algorithms are used to adjust these edge relationship weights. Through continuous iteration, the relationships between nodes and edges in the knowledge graph are made to better reflect the actual fluctuations in historical energy consumption data. Initially, the edge relationship weights may only be a rough estimate. Through iterative optimization algorithms, the edge relationship weights can be continuously adjusted and optimized based on the actual situation of historical energy consumption data, enabling the knowledge graph to more accurately reflect the relationship between manufacturing processes and energy consumption.

[0053] For example, the types of associated edges include causal, temporal, and similarity-based edges. For causal edges, the initial edge relationship weights are set to relatively high values ​​because causal relationships typically represent a close connection between processes and energy consumption. For temporal edges, the initial edge relationship weights are set based on the time intervals and order of processes. Then, iterative optimization algorithms, such as an improved version of gradient descent, are used to continuously adjust the edge relationship weights. In each iteration, based on the actual fluctuation patterns of historical energy consumption data, the error between the current node-edge relationships in the knowledge graph and the actual situation is calculated. The edge relationship weights are then adjusted based on this error to gradually reduce it. After multiple iterations, the final association between nodes and edges in the knowledge graph conforms to the actual fluctuation patterns of historical energy consumption data.

[0054] Step S300: Call the pre-trained graph neural network model to perform time-series evolution prediction processing on the knowledge graph of manufacturing process-energy consumption association, and generate the energy consumption prediction sequence for the future preset manufacturing cycle by combining the energy consumption sequence unit change trend of historical periods.

[0055] A pre-trained graph neural network model is a neural network model pre-trained for processing graph data, capable of learning the features of nodes and edges in the graph and the relationships between them. Temporal evolution prediction processing of the manufacturing process-energy consumption association knowledge graph utilizes the graph neural network model to predict the evolution of the knowledge graph over time, i.e., how the relationship between processes and energy consumption will change over time. Incorporating the changing trends of energy consumption sequence units over historical periods is to account for the changing patterns of historical energy consumption data, thereby more accurately generating the energy consumption prediction sequence for the future preset manufacturing cycle. The energy consumption prediction sequence for the future preset manufacturing cycle is the prediction result of the energy consumption of the manufacturing process over a future period, including the total predicted energy consumption value for each time unit and the energy consumption proportion of each process.

[0056] For example, a pre-trained graph neural network model, such as an improved version of the GraphSAGE model, can be used to process the knowledge graph of the relationship between manufacturing processes and energy consumption. This model has been pre-trained on a large amount of historical data, learning the relationship and changing patterns between processes and energy consumption. This is combined with the changing trends of energy consumption sequence units over historical periods.

[0057] In one implementation, step S300 may specifically include the following steps S310 to S350: Step S310: Input the knowledge graph of manufacturing process-energy consumption association into the graph embedding layer of the graph neural network model, and vectorize the node attributes and edge relationship weights to generate low-dimensional dense node embedding vectors and edge embedding vectors.

[0058] The role of the graph embedding layer is to vectorize the node attributes and edge weights in the manufacturing process-energy consumption association knowledge graph. Node attributes describe the feature information of the process execution node, such as process name, execution time, and importance; edge weights are indicators that measure the strength of the association between nodes. Vectorization converts this non-numerical information into numerical vectors so that the graph neural network model can process it. Generating low-dimensional dense node embedding vectors and edge embedding vectors maps high-dimensional, sparse node and edge information into a low-dimensional, dense vector space. This reduces the dimensionality of the data, improves the model's processing efficiency and performance, and also better captures the potential relationships between nodes and edges.

[0059] For example, a knowledge graph relating manufacturing processes and energy consumption can be input into the graph embedding layer of a graph neural network model. The knowledge graph contains multiple process execution nodes and the edges connecting them. Each node has node attributes, and each edge has an edge relation weight. The graph embedding layer uses embedding algorithms, such as deep learning-based embedding algorithms, to convert node attributes and edge relation weights into vector representations. For example, for the process name in the node attributes, word embedding techniques are used to convert it into a vector; for the edge relation weights, they are directly used as one dimension of the vector. This ultimately generates low-dimensional, dense node embedding vectors and edge embedding vectors, effectively mapping the originally high-dimensional, sparse node information into a relatively low-dimensional vector space.

[0060] Step S320: Through the temporal feature injection module of the graph neural network model, the energy consumption sequence units of the historical period are transformed into a temporal feature matrix in chronological order. The temporal feature matrix and the node embedding vector are associated element by element to generate a temporal enhanced node embedding vector that integrates temporal information.

[0061] The temporal feature injection module is a dedicated module in the graph neural network model for processing temporal information. It transforms historical energy consumption sequence units into a temporal feature matrix in chronological order. The historical energy consumption sequence units record energy consumption over a past period. Arranging these units chronologically yields a sequence of energy consumption changes over time. These sequences are then transformed into a temporal feature matrix, where rows represent different time points and columns represent different energy consumption features, such as average energy consumption and rate of change of energy consumption. Element-wise association between the temporal feature matrix and the node embedding vectors integrates temporal information into the node embedding vectors, making them contain more temporal information and thus better reflecting the feature changes of nodes over time.

[0062] In one implementation, step S320 may specifically include the following steps S321 to S325: Step S321: Perform time granularity unification processing on the energy consumption sequence units of historical periods, convert energy consumption data with different collection frequencies into energy consumption value sequences with the same time interval, and generate an energy consumption time series with equal intervals.

[0063] In actual data acquisition, the frequency of energy consumption data collection may vary due to different acquisition devices or strategies; some data may be collected hourly, while others may be collected daily. This difference in collection frequency can complicate subsequent data analysis and processing. By converting energy consumption data with different collection frequencies into energy consumption value sequences with the same time intervals, generating equally spaced energy consumption time series, the data can be made comparable over time.

[0064] Step S322: Input the energy consumption time series into the time series feature extractor, extract the periodic and trend features of energy consumption fluctuations through the sliding window Fourier transform, and combine them to generate a time series feature matrix containing periodic and trend components.

[0065] A time-series feature extractor is a tool used to extract features from time-series data, extracting useful feature information from energy consumption time series. The sliding window Fourier transform extracts the periodic and trend features of energy consumption fluctuations by sliding a fixed-size window across the energy consumption time series and performing a Fourier transform on the data within the window. Periodic features refer to the periodic regularity of energy consumption fluctuations, such as daily, weekly, or monthly energy consumption cycles; trend features refer to the overall trend of energy consumption over a period of time, such as whether energy consumption is increasing or decreasing. Combining these features generates a time-series feature matrix containing periodic and trend components, representing the periodic and trend characteristics of energy consumption in matrix form.

[0066] Step S323: Perform dimensional expansion processing on the node embedding vector. Adjust the dimension of the node embedding vector to be consistent with the column dimension of the temporal feature matrix by zero padding, and generate a dimension-matched expanded node embedding vector.

[0067] Dimension expansion is an operation that adjusts the dimensions of the node embedding vectors to match the column dimensions of the temporal feature matrix. The temporal feature matrix generated in step S322 has a certain column dimension, while the node embedding vectors may have a different dimension. The dimension of the node embedding vectors is adjusted to match the column dimension of the temporal feature matrix using zero-padding; that is, zero elements are added to the end of the node embedding vectors to increase their dimension. Generating dimension-matched expanded node embedding vectors ensures that the dimensions of the temporal feature matrix and the node embedding vectors match during subsequent element-wise association processing, allowing for successful computation.

[0068] For example, the node embedding vector might have a set dimension, while the temporal feature matrix might have a different column dimension. During dimension expansion, zero-padding is used to add zero elements to the end of the node embedding vector to make its dimension match the column dimension of the temporal feature matrix. For instance, if the dimension of the node embedding vector is smaller than the column dimension of the temporal feature matrix, enough zero elements are added to the end of the node embedding vector until its dimension matches the column dimension, ultimately generating a dimension-matched expanded node embedding vector.

[0069] Step S324: Perform element-wise association processing between the temporal feature matrix and the extended node embedding vector, calculate the association result of each element as the fused feature value, and generate a preliminary fused feature matrix.

[0070] Element-wise association between the temporal feature matrix and the extended node embedding vector involves associating each element in the temporal feature matrix with the corresponding element in the extended node embedding vector. This association operation can be multiplication, addition, or other custom operations, with the aim of fusing temporal and node feature information. The association result for each element is calculated as the fused feature value, and these feature values ​​are combined to form a preliminary fused feature matrix. This preliminary fused feature matrix contains the initial fusion results of temporal and node feature information, providing data for subsequent feature selection processing.

[0071] For example, the temporal feature matrix and the extended node embedding vector are multiplied element-wise. For instance, an element in the temporal feature matrix represents the energy consumption cycle feature value within a certain time window, and the corresponding element in the extended node embedding vector represents the feature value of a certain node. Multiplying these two elements yields the fused feature value. This operation is performed on each element of both the matrix and the vector, ultimately generating a preliminary fused feature matrix. The elements in this matrix are the fused feature values, which combine temporal and node feature information.

[0072] Step S325: The preliminary fused feature matrix is ​​processed by the gating mechanism of the temporal feature injection module to select features, retain the feature components whose contribution to energy consumption prediction is greater than the contribution threshold, and generate the final temporal enhancement node embedding vector. The gating mechanism dynamically adjusts the retention weight of each feature component through the activation function.

[0073] The gating mechanism filters the initial fused feature matrix, removing feature components that contribute little to energy consumption prediction and retaining those whose contribution exceeds a contribution threshold. The contribution threshold is a pre-defined standard used to determine the importance of a feature component to energy consumption prediction. By dynamically adjusting the retention weights of each feature component through an activation function, the gating mechanism adaptively determines the retention level of each feature component based on its importance. The resulting temporal augmentation node embedding vector contains the most useful temporal and node feature information for energy consumption prediction, providing high-quality data for subsequent graph convolutional layer processing.

[0074] For example, the gating mechanism of the temporal feature injection module processes the initial fused feature matrix. Assuming the contribution threshold is a reasonable standard, for each feature component in the initial fused feature matrix, its retention weight is calculated using an activation function, such as an improved version of the sigmoid function. If the retention weight of a feature component is higher than the contribution threshold, the feature component is retained; if it is lower than the contribution threshold, the feature component is removed. After such filtering, the final temporal augmentation node embedding vector containing the most useful information is generated for subsequent graph convolutional layer processing.

[0075] Step S330: The graph convolutional layer of the graph neural network model is used to aggregate the neighbor node information of the temporal enhancement node embedding vector. The contribution of the neighbor nodes is weighted and summed by the edge embedding vector to generate aggregated node features containing local neighborhood association information.

[0076] Neighbor nodes are nodes adjacent to a given node, connected by edges. In a graph structure, the relationships between nodes can be represented by edges, and edge embedding vectors represent the features and weights of those edges. Weighting the contributions of neighbor nodes using edge embedding vectors takes into account the varying degrees of contribution from different neighbor nodes to the central node; a larger edge weight indicates a greater influence of that neighbor node on the central node. Generating aggregated node features that include local neighborhood association information involves fusing information from the central node and its neighbors to obtain a feature vector containing local neighborhood association information. This feature vector more comprehensively reflects the contextual information of a node within the graph structure.

[0077] In one implementation, step S330 may specifically include the following steps S331 to S335: Step S331: Construct the adjacency matrix of the manufacturing process-energy consumption association knowledge graph, determine the search range of neighbor nodes according to the weight of edge relationships, and define nodes with edge relationship weights greater than a preset threshold as direct neighbor nodes.

[0078] An adjacency matrix is ​​used to represent the connection relationships between nodes in a graph. In this step, an adjacency matrix for the manufacturing process-energy consumption association knowledge graph is constructed. The elements of the matrix represent whether nodes are connected by edges and the weight of the edges. Determining the search range of neighboring nodes based on the magnitude of the edge relationship weights is to filter out neighboring nodes closely related to the central node. A preset threshold is a pre-defined standard; nodes with edge relationship weights greater than this threshold are defined as direct neighbors. Considering only the information of these direct neighbors reduces computational load and noise, improving the efficiency and accuracy of the model.

[0079] For example, construct an adjacency matrix for a knowledge graph linking manufacturing processes and energy consumption. Assume the rows and columns of the matrix represent different process execution nodes, and the elements in the matrix represent the edge weights between nodes. A preset threshold is a reasonable standard: if the edge weight between a node and the central node is greater than this threshold, the node is defined as a direct neighbor; if the edge weight is less than the threshold, it is not considered a direct neighbor. Through this process, the set of direct neighbors of the central node is determined.

[0080] Step S332: Perform directional labeling on the temporal augmentation node embedding vectors of each node's direct neighbors, and add directional features to the neighbor node embedding vectors according to the type of edge relationship.

[0081] Direction labeling is an operation that labels the temporal augmented node embedding vectors of each node's direct neighbors to reflect the directional information of the edges. Edge relationships can be causal, temporal, or other types, each with different directional meanings. Adding directional features to the neighbor node embedding vectors based on the edge relationship type allows the neighbor node's information to more accurately reflect its relationship with the central node, and better considers the directional factors of edges during information aggregation.

[0082] For example, for a direct neighbor node of a central node, directional features are added to its temporal augmented node embedding vector based on the type of edge relationship. If the edge relationship is causal and the neighbor node is the cause of the central node's energy consumption change, then a feature representing the causal direction is added to the neighbor node's temporal augmented node embedding vector; for example, a set value is set at a certain position in the vector to represent the causal direction. If the edge relationship is temporal and the neighbor node precedes the central node in time, then corresponding temporal directional features are added.

[0083] Step S333: Using the edge embedding vector as a weight coefficient, perform weighted summation on the neighbor node embedding vectors with directional features, calculate the contribution of each neighbor node to the center node, and generate a neighbor contribution vector.

[0084] Edge embedding vectors are used as weighting coefficients because they represent the features and weights of edges, reflecting the strength of the association between neighboring nodes and the central node. The weighted summation of the neighbor embedding vectors with directional features involves multiplying each neighbor's embedding vector by the weighting coefficient corresponding to the edge embedding vector, and then adding these results to obtain the contribution of each neighbor to the central node. The generated neighbor contribution vector represents the degree of contribution of all neighboring nodes to the central node, comprehensively considering the features of neighboring nodes, edge weights, and directional information.

[0085] As one implementation method, in step S333, the contribution of each neighbor node to the central node is calculated, and a neighbor contribution vector is generated. Specifically, this may include the following steps S3331~S3336: Step S3331: Perform relation type decoding on the edge embedding vector, call the basic weight coefficients corresponding to the edge relation type from the preset relation weight library, and generate the edge type weight vector.

[0086] Relation type decoding is the process of parsing edge embedding vectors to identify the relationship type of the edges. The preset relationship weight library is a database that pre-stores the basic weight coefficients corresponding to different edge relationship types. Retrieving the basic weight coefficients corresponding to the edge relationship type from the preset relationship weight library is to determine the basic weight of the edge based on its relationship type. Generating an edge type weight vector represents the basic weight coefficients of different edge relationship types in vector form.

[0087] For example, the edge embedding vector contains the feature information of the edge. Through relation type decoding, the relation type of the edge is identified, such as causal association, temporal sequence, etc. Assume that in the preset relation weight library, the basic weight coefficient for causal association is a higher value, and the basic weight coefficient for temporal sequence is a lower value. When an edge is identified as having a causal association, the basic weight coefficient is retrieved from the preset relation weight library to generate an edge type weight vector. The value at the position corresponding to the edge in the vector is that basic weight coefficient.

[0088] Step S3332: Combine the directional features with the edge type weight vector to calculate the directional weighting coefficient.

[0089] The directional feature is a feature representing the edge direction added to the neighbor node embedding vector in step S332, and the edge type weight vector is a vector representing the basic weight of the edge relationship type generated in step S3331. Combining the directional feature with the edge type weight vector combines the directional information and the basic edge weights to calculate a weighted coefficient that considers the directional factor. The directional weighting coefficient is a coefficient that comprehensively considers the edge relationship type and directional information, used to more accurately calculate the contribution of neighbor nodes to the center node. For example, the directional feature of a neighbor node might be a set of preset values, and the edge type weight vector might be another set of values. Combining the directional feature with the edge type weight vector, such as multiplying corresponding elements to obtain the product value of each element, and then summing these products or performing other operations, yields the directional weighting coefficient. This directional weighting coefficient comprehensively considers the edge relationship type and directional information, providing a more accurate weight for subsequent calculations of the contribution of neighbor nodes.

[0090] Step S3333: Extract the dynamic change features from the temporal augmentation node embedding vector of neighboring nodes, and calculate the contribution decay coefficient of historical neighboring nodes to the current central node through the time decay function.

[0091] Dynamic change characteristics are features reflected in the temporal augmentation node embedding vectors of neighboring nodes, such as the rate of change and trend of energy consumption. The time decay function is used to calculate the degree of attenuation of the contribution of historical neighboring nodes to the current central node; the independent variable of the function is the difference in timestamps between the current central node and historical neighboring nodes. Over time, the influence of historical neighboring nodes on the current central node gradually decreases. Calculating the contribution decay coefficient using the time decay function allows for a more accurate consideration of the impact of time factors on the contribution of neighboring nodes.

[0092] For example, we can extract dynamic features from the temporal augmentation embedding vectors of neighboring nodes, such as the energy consumption trend of a particular neighboring node over a past period. Assuming the time decay function is an exponential decay function, the timestamp of the current center node is the current time, and the timestamp of a historical neighboring node is a past time. We calculate the difference between the timestamps of the current center node and the historical neighboring node, and substitute this difference into the time decay function to obtain the contribution decay coefficient. This coefficient represents the degree to which the contribution of the historical neighboring node to the current center node decays over time.

[0093] Step S3334: Perform a comprehensive calculation on the edge type weight vector, direction weighting coefficient, and contribution attenuation coefficient to generate a comprehensive contribution coefficient.

[0094] The comprehensive contribution coefficient is generated by combining the edge type weight vector, direction weighting coefficient, and contribution attenuation coefficient, taking into account the basic weights, direction information, and time factors of the edges. This comprehensive contribution coefficient measures the degree of contribution of neighboring nodes to the central node. For example, the comprehensive contribution coefficient of each neighboring node is obtained by multiplying these three vectors or coefficients. This coefficient comprehensively considers the relationship type, direction, and time factors of the edges, providing accurate weights for the subsequent weighted summation of the embedding vectors of neighboring nodes.

[0095] Step S3335: The temporal augmentation node embedding vectors of all direct neighbor nodes are weighted and summed according to the comprehensive contribution coefficient to generate a preliminary neighbor contribution vector.

[0096] The temporal augmentation node embedding vectors of all direct neighbor nodes are weighted and summed according to the comprehensive contribution coefficient. This is done by multiplying the temporal augmentation node embedding vector of each direct neighbor node by its corresponding comprehensive contribution coefficient, and then summing these results to obtain the preliminary neighbor contribution vector. The preliminary neighbor contribution vector is a preliminary result of the contribution of all direct neighbor nodes to the central node after considering multiple factors.

[0097] For example, there are multiple direct neighbor nodes, and their temporally augmented node embedding vectors represent their respective feature information. Each neighbor node corresponds to a comprehensive contribution coefficient. Multiplying the embedding vector of each neighbor node by the corresponding comprehensive contribution coefficient yields the contribution value of that neighbor node to the center node. This process is performed on all direct neighbor nodes, and then these contribution values ​​are summed to obtain the preliminary neighbor contribution vector.

[0098] Step S3336: Normalize the initial neighbor contribution vector to adjust its size to an appropriate range and generate a neighbor contribution vector that includes direction awareness and relationship type differences.

[0099] Normalization is an operation that adjusts the initial neighbor contribution vector to bring its size to a suitable range, making its elements comparable. By normalizing the initial neighbor contribution vector, scale differences between vector elements can be eliminated, making the neighbor contribution vector more stable and comparable. The generated neighbor contribution vector, which incorporates direction awareness and relationship type differences, is the normalized vector, comprehensively considering the directional information of neighbor nodes, differences in edge relationship types, and time factors.

[0100] For example, the initial neighbor contribution vector is obtained by weighting and summing the temporal augmentation node embedding vectors of all direct neighbor nodes according to their comprehensive contribution coefficients. This vector is then normalized, for example using L2 normalization, by dividing each element of the vector by its magnitude, adjusting the vector's size to an appropriate range. The final result is a neighbor contribution vector that incorporates direction awareness and relationship type differences. The elements of this vector are comparable and can be better concatenated with the temporal augmentation node embedding vector of the central node.

[0101] Step S334: Concatenate the temporal enhancement node embedding vector of the center node with the neighbor contribution vector, input the nonlinear transformation function of the graph convolutional layer, perform feature mapping through the activation function, and generate preliminary aggregated node features.

[0102] The concatenation of the temporal augmentation node embedding vector of the central node with the neighbor contribution vector merges the information of the central node itself and the neighbor nodes, resulting in a vector containing more information. The nonlinear transformation function of the input graph convolutional layer introduces nonlinearity to increase the model's expressive power. Feature mapping through the activation function maps the input vector to a new feature space, generating preliminary aggregated node features. This preliminary aggregated node feature integrates information from the central node and neighbor nodes and, having undergone nonlinear transformation, can better capture the complex relationships in the data.

[0103] For example, the temporal augmentation node embedding vector of the central node represents the feature information of the central node, and the neighbor contribution vector represents the contribution information of neighboring nodes to the central node. These two vectors are concatenated, for example, by appending the neighbor contribution vector to the temporal augmentation node embedding vector of the central node, resulting in a longer vector. This concatenated vector is then input into the nonlinear transformation function of the graph convolutional layer. Through an activation function, such as an improved version of the ReLU activation function, feature mapping is performed. The activation function nonlinearly transforms the elements in the input vector, changing elements less than zero to zero and leaving elements greater than zero unchanged, thus generating preliminary aggregated node features.

[0104] Step S335: Perform layer normalization on the initial aggregated node features to eliminate the dimensional differences between different node features and generate aggregated node features with uniform distribution characteristics.

[0105] Layer normalization is a method for normalizing the features of initially aggregated nodes. It eliminates the dimensional differences between features of different nodes, giving them a uniform distribution. Features of different nodes may have different dimensions and distributions due to different data sources and calculation methods. Layer normalization allows these features to be compared and processed on the same scale, improving the stability and performance of the model. For example, the initial aggregated node features are obtained by concatenating and nonlinearly transforming information from the center node and its neighbors. The initial aggregated node features of different nodes may have different numerical ranges and distributions. Layer normalization calculates the mean and standard deviation of the feature vector, then subtracts the mean from each element of the vector and divides it by the standard deviation, thereby adjusting the elements of the feature vector to a relatively uniform distribution range.

[0106] Step S340: Call the temporal prediction layer of the graph neural network model to perform long short-term memory modeling on the features of the aggregated nodes, extract the dependence of energy consumption fluctuation features in the time dimension, and generate the energy consumption prediction subsequence of each process execution node in the future preset manufacturing cycle.

[0107] The temporal prediction layer is a module in a graph neural network model specifically designed for temporal prediction. It utilizes Long Short-Term Memory (LSTM) modeling to capture the temporal dependencies of energy consumption fluctuations. In practice, aggregated node features are input to the LSTM units of the temporal prediction layer. Each LSTM unit contains input gates, forget gates, and output gates. These gating mechanisms control the inflow, retention, and output of information. The input gate determines how much of the current input information can enter the unit state, the forget gate determines how much information from the previous unit state needs to be forgotten, and the output gate determines how much information from the current unit state can be output. Through the coordinated operation of these gating mechanisms, the LSTM unit can effectively learn and retain the temporal dependencies of energy consumption fluctuations.

[0108] Step S350: Concatenate the energy consumption prediction sub-sequences of all process execution nodes according to the process execution order to generate an energy consumption prediction sequence covering the entire manufacturing cycle. The energy consumption prediction sequence includes the total energy consumption prediction value corresponding to each time unit and the energy consumption ratio of each process.

[0109] The purpose of concatenating the energy consumption prediction sub-sequences of all process execution nodes according to the execution order is to integrate the energy consumption prediction information of each process into a complete energy consumption prediction sequence. For example, different processes are executed sequentially in a certain order, such as stamping first, followed by welding, painting, and assembly. By concatenating the energy consumption prediction sub-sequences of each process according to this execution order, a continuous energy consumption prediction sequence covering the entire manufacturing cycle can be obtained.

[0110] During the assembly process, it is necessary to ensure that the time unit of each sub-sequence matches the time unit of the entire manufacturing cycle. For example, if the time unit of each sub-sequence is hours, then the energy consumption prediction sequence for the entire manufacturing cycle is also constructed with hours as the time unit. After assembly, the energy consumption prediction sequence includes the total energy consumption prediction value corresponding to each time unit and the energy consumption percentage of each process. The total energy consumption prediction value is the sum of the energy consumption of all processes within that time unit, while the energy consumption percentage of each process reflects the proportion of each process in the total energy consumption.

[0111] Step S400: Dynamically schedule and optimize the execution order and resource allocation scheme of the current manufacturing process based on the energy consumption prediction sequence, and generate an energy consumption optimization-oriented manufacturing execution instruction that includes process adjustment priority and resource allocation scheme.

[0112] Energy consumption forecasting sequences provide factories with predictions of energy consumption during future manufacturing cycles. Based on these predictions, the execution sequence of current manufacturing processes and resource allocation schemes can be dynamically optimized. Dynamic scheduling optimization involves adjusting the execution sequence of processes and resource allocation in real time based on actual energy consumption and forecast results, in order to reduce energy consumption and improve resource utilization.

[0113] For example, different execution sequences and resource allocation methods can have varying impacts on energy consumption. For instance, the energy consumption of certain processes may differ significantly depending on when they are executed at different times. Reasonable allocation of resources, such as equipment, manpower, and raw materials, can improve production efficiency and reduce energy consumption. By analyzing energy consumption prediction sequences, we can identify periods and processes with high energy consumption, as well as areas where resource allocation is unreasonable, and then make corresponding adjustments.

[0114] The generated energy-optimized manufacturing execution instructions include process adjustment priorities and resource allocation plans. Process adjustment priorities are determined based on factors such as energy consumption and importance, prioritizing adjustments to processes with high energy consumption but minimal impact on overall production. Resource allocation plans, based on process requirements and energy consumption forecasts, rationally allocate resources to ensure optimal resource utilization. This instruction provides clear guidance for factory production execution, contributing to energy optimization and improved production efficiency.

[0115] In one implementation, step S400 may specifically include the following steps S410 to S450: Step S410: Compare and analyze the energy consumption prediction sequence with the preset energy consumption threshold range, identify the process execution time period and the corresponding process execution node where the energy consumption prediction value exceeds the threshold range, and mark them as process nodes to be optimized.

[0116] The preset energy consumption threshold range is a reasonable energy consumption interval pre-set based on factors such as the factory's historical energy consumption data, production processes, and energy management goals. By comparing and analyzing the energy consumption prediction sequence with this threshold range, it is possible to identify the periods in which the predicted energy consumption values ​​exceed the normal range and the corresponding process execution nodes.

[0117] For example, each process has a reasonable energy consumption range under normal conditions. For instance, the energy consumption of the stamping process should be within a reasonable range under normal production speed and equipment operating conditions. If the energy consumption prediction sequence shows that the predicted energy consumption value for a certain stamping process execution period exceeds the preset threshold range, then this period and the corresponding process execution node will be marked as a process node to be optimized.

[0118] Step S420: Perform energy consumption sensitivity assessment on the process node to be optimized, calculate the product of the magnitude by which the predicted energy consumption exceeds the threshold and the process importance coefficient, and use the product result as a quantitative indicator for process adjustment priority.

[0119] Energy sensitivity assessment is a process of in-depth analysis of each process node to be optimized, aiming to determine the importance and priority of each process in energy consumption optimization. The extent to which the predicted energy consumption value exceeds the threshold reflects the current degree of energy consumption anomaly in that process; the greater the exceedance, the more serious the energy consumption problem in that process. The process importance coefficient is determined comprehensively based on factors such as the process's impact on the final product quality, the process's irreplaceability, and the process's execution time. For example, some processes are crucial to product quality, such as engine assembly, and have relatively high importance coefficients; while some auxiliary processes, such as parts cleaning, have relatively low importance coefficients. For a process node to be optimized, if its predicted energy consumption value exceeds the threshold significantly while its importance coefficient is low, then that process should be prioritized for adjustment. By calculating the product of the extent to which the predicted energy consumption value exceeds the threshold and the process importance coefficient, a quantitative index is obtained, which can be used to determine the priority of process adjustments.

[0120] In one implementation, step S420 may specifically include: Step S421: Extract the difference between the predicted energy consumption value of the process node to be optimized and the preset energy consumption threshold, and define the ratio of the difference to the preset energy consumption threshold as the energy consumption exceedance.

[0121] Extracting the difference between the predicted energy consumption value and the preset energy consumption threshold for the process node to be optimized is to clarify the specific extent to which the energy consumption of that process exceeds the normal range. Comparing this difference with the preset energy consumption threshold yields a ratio, which represents the extent of energy consumption exceedance. The larger the extent of energy consumption exceedance, the more severe the energy consumption anomaly of that process.

[0122] Step S422: Determine the process importance coefficient of each process execution node through statistical analysis of historical manufacturing data.

[0123] The importance coefficient of a process is determined comprehensively based on its impact on the quality of the final product, its irreplaceability, and its execution time. Statistical analysis of historical manufacturing data reveals the role and impact of each process throughout the production process. The degree of impact on the quality of the final product is a key factor; if a process directly affects important quality indicators such as product performance and safety, its importance coefficient will be high. The irreplaceability of a process is also important; if a process can only be performed by specific equipment or personnel and cannot be replaced by other processes, its importance coefficient will increase accordingly. The execution time of a process also affects its importance coefficient; processes with longer execution times occupy more time and resources in the overall production process, and their importance is relatively higher.

[0124] For example, by analyzing historical data, the impact of each process on product quality can be statistically analyzed, the irreplaceability of each process can be assessed, and the execution time of each process can be recorded. Taking all these factors into account, a process importance coefficient is determined for each process execution node. For instance, the engine manufacturing process is crucial to the performance and quality of automobiles, has high irreplaceability, and a long execution time, so its process importance coefficient will be relatively high; while some simple packaging processes have less impact on product quality, are easily substitutable, and have short execution times, so their process importance coefficient will be lower.

[0125] Step S423: Standardize the energy consumption exceedance and process importance coefficient, converting them into dimensionless values ​​with the same dimensions.

[0126] For example, the calculation methods and value ranges of energy consumption exceedance and process importance coefficient may vary greatly. Through standardization, such as using normalization methods, the values ​​of energy consumption exceedance and process importance coefficient can be mapped to the same interval, such as [0,1].

[0127] Step S424: Calculate the product of the standardized energy consumption excess and the process importance coefficient, and use the product result as the initial quantitative indicator for process adjustment priority.

[0128] After standardization, the standardized energy consumption exceedance is multiplied by the process importance coefficient. The resulting product is the initial quantitative indicator for process adjustment priority. This indicator comprehensively considers the degree of energy consumption anomaly and importance of the process, and can accurately reflect the priority of the process in energy consumption optimization.

[0129] For example, for each process node to be optimized, its standardized energy consumption exceedance is multiplied by the process importance coefficient. For instance, if a process has a high standardized energy consumption exceedance but a low process importance coefficient, the resulting initial quantification index will be relatively high, indicating that this process should be prioritized for adjustment. This initial quantification index allows for a preliminary ranking of all process nodes to be optimized.

[0130] Step S425: Sort the initial quantitative indicators of all process nodes to be optimized, and use the sorting results as the final process adjustment priority. The higher the priority value, the more priority the process node needs to be adjusted.

[0131] By ranking the initial quantitative indicators of all process nodes to be optimized, the adjustment priority of each process can be clearly determined. The ranking result will serve as the final process adjustment priority. The higher the priority value of a process node, the more serious its energy consumption problem is and the relatively smaller its impact on overall production; therefore, it should be adjusted first.

[0132] For example, all process nodes to be optimized are sorted from largest to smallest according to the initial quantitative indicators. After sorting, the process nodes at the top are the processes that need to be adjusted first. For example, if the sorting results show that a certain painting process has the highest priority value, then when adjusting the process, the first thing to consider is to optimize that painting process, such as adjusting its execution time, replacing equipment, or optimizing the process, in order to reduce its energy consumption.

[0133] Step S430: Based on the process adjustment priority, adjust the execution order of the highest priority process node to be optimized, move it from the high energy consumption period to the low energy consumption period, and recalculate the adjusted energy consumption prediction subsequence.

[0134] High energy consumption periods are those when the predicted energy consumption value in the energy consumption prediction sequence is higher than the preset high energy consumption threshold, and low energy consumption periods are those when the predicted energy consumption value in the energy consumption prediction sequence is lower than the preset low energy consumption threshold.

[0135] Based on process adjustment priorities, the highest priority process nodes to be optimized are processed first. Moving the process from high-energy-consumption periods to low-energy-consumption periods can effectively reduce overall energy consumption. High-energy-consumption periods are determined based on energy consumption prediction sequences; during these periods, the predicted energy consumption value is higher than a preset high-energy-consumption threshold, indicating higher energy costs. Low-energy-consumption periods are those where the predicted energy consumption value is lower than a preset low-energy-consumption threshold; performing processes during these periods can reduce energy consumption.

[0136] In one implementation, step S430 may specifically include: Step S431: Determine the distribution of energy consumption prediction values ​​for each time period within the current manufacturing cycle, mark the time periods with energy consumption prediction values ​​lower than the preset low energy consumption threshold as low energy consumption time periods, and generate a list of low energy consumption time periods.

[0137] Determining the distribution of predicted energy consumption across different time periods within the current manufacturing cycle requires a detailed analysis of the energy consumption prediction sequence. By comparing the predicted energy consumption for each time period with a preset low energy consumption threshold, periods with lower energy consumption can be identified.

[0138] For example, the energy consumption prediction sequence provides predicted energy consumption values ​​for each time unit within the future manufacturing cycle. These predicted values ​​are compared one by one with a preset low-energy consumption threshold. When the predicted energy consumption value for a certain period is lower than the threshold, that period is marked as a low-energy consumption period. All marked low-energy consumption periods are compiled to generate a low-energy consumption period list. This list provides a range of options for adjusting the execution time of processes, allowing the factory to select appropriate low-energy consumption periods to execute processes based on process requirements and actual conditions.

[0139] Step S432: Select target low-energy periods from the list of low-energy periods that match the execution duration of the process node to be optimized, so that the duration of the target low-energy periods is not less than the execution duration of the process node to be optimized.

[0140] Selecting suitable target low-energy periods from the list of low-energy periods requires considering the execution duration of the process node to be optimized. The duration of the target low-energy period must be no less than the execution duration of the process node to be optimized to ensure that the process can be fully executed within this period.

[0141] Step S433: Calculate the energy consumption change after the process node to be optimized is migrated to different target low energy consumption periods, and select the target low energy consumption period with the smallest energy consumption change as the optimal migration period.

[0142] Calculating the energy consumption change after migrating the process node to be optimized to different target low-energy consumption periods is to evaluate the effectiveness of each migration scheme. The energy consumption change reflects the degree of energy reduction after the process migration. Selecting the target low-energy consumption period with the smallest energy consumption change as the optimal migration period can minimize energy consumption.

[0143] For example, for each target low-energy consumption period, the energy consumption of the process node to be optimized after migrating to that period is simulated. By comparing the predicted energy consumption before and after the migration, the change in energy consumption is calculated. For instance, by migrating a welding process to different target low-energy consumption periods, the difference between the predicted energy consumption after migration and the original planned energy consumption is calculated, resulting in different changes in energy consumption. The target low-energy consumption period with the smallest change in energy consumption is selected as the optimal migration period. This minimizes energy consumption while ensuring the normal execution of the process.

[0144] Step S434: Adjust the execution time interval of the process node to be optimized to the optimal migration time period, update the execution sequence table of the manufacturing process, and record the time interval change information before and after the process adjustment.

[0145] Adjusting the execution time interval of the process node to be optimized to the optimal migration period is a step in putting the previous analysis and decisions into practice. Updating the execution sequence table of the manufacturing processes ensures that the execution order and timing of all processes are up-to-date. Recording the changes in time intervals before and after the process adjustment helps with subsequent tracking and evaluation.

[0146] For example, based on the previously determined optimal migration period, the execution time interval of the process node to be optimized is adjusted to that period. For instance, the execution time of a painting process is adjusted from its original high-energy-consumption period to the optimal migration period. Simultaneously, the execution sequence table of the manufacturing processes is updated, modifying the time information for that process in the table. The changes in time intervals before and after the process adjustment are recorded, such as the original execution time interval and the new execution time interval. This information can be used to analyze the impact of the process adjustment on overall production and energy consumption.

[0147] Step S435: Based on the adjusted execution order table, the graph neural network model is called again to perform time-series evolution prediction processing on the knowledge graph of manufacturing process-energy consumption association, generating an adjusted energy consumption prediction subsequence, and verifying whether the adjusted energy consumption prediction value is within the preset threshold range.

[0148] Based on the adjusted execution sequence list, the graph neural network model is re-invoked for time-series evolution prediction processing to obtain accurate energy consumption prediction results for the adjusted process. Generating the adjusted energy consumption prediction subsequence allows us to understand the future energy consumption of the adjusted process. Verifying whether the adjusted energy consumption prediction value is within the preset threshold range checks whether the process adjustment has achieved the expected energy consumption optimization effect.

[0149] For example, the adjusted execution sequence list is input into a graph neural network model. Based on the new execution order and related information, the model performs temporal evolution prediction on the manufacturing process-energy consumption association knowledge graph. The generated adjusted energy consumption prediction subsequence reflects the energy consumption prediction of the adjusted process. The adjusted energy consumption prediction value is compared with a preset threshold range. If the prediction value is within the preset threshold range, it indicates that the process adjustment has achieved a certain energy consumption optimization effect; if it still exceeds the threshold range, further analysis and adjustment are needed.

[0150] Step S440: Based on the adjusted energy consumption prediction subsequence, optimize the allocation scheme of manufacturing resources, and prioritize the allocation of resources to process execution nodes with lower energy consumption prediction values ​​and higher process importance coefficients. Wherein, lower energy consumption prediction value means that the energy consumption prediction value of the process execution node is lower than the preset energy consumption threshold, and higher process importance coefficient means that the process importance coefficient of the process execution node is higher than the preset importance coefficient threshold.

[0151] Based on the adjusted energy consumption prediction subsequence, we can understand the adjusted energy consumption of each process execution node. Optimizing the allocation scheme of manufacturing resources aims to improve resource utilization and reduce overall energy consumption. Prioritizing the allocation of resources to process execution nodes with lower predicted energy consumption values ​​and higher process importance coefficients ensures that important processes receive sufficient resource support while reducing unnecessary energy consumption.

[0152] As one implementation method, step S440 specifically includes: Step S441: Collect information on the availability of current manufacturing resources.

[0153] For example, a resource availability status matrix can be generated, including equipment operating status, personnel scheduling, and raw material inventory levels. Various monitoring systems and management records can be used to collect equipment operating status information, such as equipment start-up time and fault conditions; understand personnel scheduling, including each employee's working hours and task assignments; and calculate raw material inventory levels, such as steel and plastics. This information is then organized into a resource availability status matrix, where rows can represent different resource types and columns can represent different times or processes.

[0154] Step S442: For each process execution node in the adjusted energy consumption prediction subsequence, calculate the comprehensive score of the energy consumption prediction value and the process importance coefficient, and generate the process resource demand priority.

[0155] To comprehensively score each execution node in the adjusted energy consumption prediction subsequence, two factors need to be considered: the predicted energy consumption value and the process importance coefficient. Calculating the comprehensive score of these two factors yields the resource requirement priority for each execution node.

[0156] For example, for each process execution node, a comprehensive calculation is performed based on its predicted energy consumption and process importance coefficient. For instance, a weighted average method can be used, assigning different weights to the predicted energy consumption and process importance coefficient, and then calculating the weighted comprehensive score. Process execution nodes with higher comprehensive scores have higher priority in resource requirements.

[0157] Step S443: Based on the priority of process resource requirements and the resource availability status matrix, establish a resource allocation optimization model with the optimization objectives of minimizing total energy consumption and maximizing resource utilization.

[0158] Based on the priority of process resource requirements and the resource availability matrix, a resource allocation optimization model is established to find the optimal resource allocation scheme. The optimization objective is to minimize total energy consumption and maximize resource utilization, that is, to reduce energy consumption and improve resource utilization efficiency as much as possible while meeting the production needs of each process.

[0159] In one implementation, step S443 may specifically include: Step S4431: Convert the priority of process resource demand into the objective function coefficient of resource allocation. The objective function coefficient is positively correlated with the value of the priority of process resource demand.

[0160] Transforming the priority of process resource requirements into objective function coefficients for resource allocation is to reflect the degree of resource demand of each process in the optimization model. The objective function coefficients are positively correlated with the values ​​of the process resource requirement priorities; that is, the higher the priority of a process resource requirement, the larger its corresponding objective function coefficient.

[0161] For example, for each process execution node, the corresponding objective function coefficients are determined based on the priority of its process resource requirements. For instance, if a process has a higher resource requirement priority, then the coefficient corresponding to that process will be larger in the objective function of resource allocation.

[0162] Step S4432: Construct a set of resource constraints.

[0163] The set of resource constraints includes equipment load rate constraints, personnel skill matching constraints, and raw material supply continuity constraints. The equipment load rate constraint limits the number of processes that a single piece of equipment can process simultaneously to no more than the equipment's maximum parallel processing capacity. The personnel skill matching constraint limits the degree of matching between the skill level of the operator and the skill level required by the process to be no less than a preset threshold. The raw material supply continuity constraint limits the difference between the replenishment rate and the consumption rate of raw materials to be no less than zero.

[0164] Step S4433: The weighted summation method is used to merge the goal of minimizing total energy consumption and the goal of maximizing resource utilization into a single objective optimization function. The priority of the two objectives is adjusted by a dynamic weight factor, which is automatically updated according to the proportion of energy consumption exceeding the threshold in the current manufacturing cycle.

[0165] By employing a weighted summation method to merge the objectives of minimizing total energy consumption and maximizing resource utilization into a single objective function, the aim is to consider both objectives within a single objective function. A dynamic weighting factor is used to adjust the priority of the two objectives, automatically updated based on the predicted energy consumption exceeding the threshold during the current manufacturing cycle.

[0166] For example, if the predicted energy consumption exceeding the threshold in the current manufacturing cycle is high, it indicates a serious energy consumption problem. In this case, the dynamic weighting factor will increase the weight of minimizing the total energy consumption objective, making the optimization model more focused on reducing energy consumption. Conversely, if the predicted energy consumption exceeding the threshold is low, the dynamic weighting factor will increase the weight of maximizing resource utilization, making the model more focused on improving resource efficiency. In this way, the priority of the two objectives is dynamically adjusted according to the actual situation, achieving a more reasonable allocation of resources.

[0167] Step S4434: Transform the resource availability state matrix into the decision variable boundary conditions of the optimization model. The decision variable boundary conditions limit the upper and lower limits of the allocation of each resource. The upper limit is the maximum available quantity in the resource availability state matrix, and the lower limit is the minimum resource quota required for the process.

[0168] For example, regarding equipment resources, if the resource availability state matrix shows that the maximum available time for a certain piece of equipment is a certain duration, then in the optimization model, the time allocated to that equipment for each process cannot exceed this maximum available time. For personnel resources, the upper and lower limits of the number of personnel allocated to each process are determined based on personnel scheduling and skill levels. For raw material resources, the upper and lower limits of the quantity of raw materials allocated to each process are determined based on inventory levels and replenishment plans. By setting boundary conditions for decision variables, the feasibility of the resource allocation scheme can be ensured.

[0169] Step S4435: Introduce a resource conflict resolution penalty term. When multiple processes have competing demands for the same resource, calculate the conflict penalty coefficient based on the difference in process adjustment priorities. Add the conflict penalty coefficient to the optimization objective function. The larger the difference in process adjustment priorities, the larger the value of the conflict penalty coefficient.

[0170] The resource conflict resolution penalty term is introduced to address the problem of multiple processes competing for the same resource. When resource competition occurs, a conflict penalty coefficient is calculated based on the difference in process adjustment priorities and added to the optimization objective function. The larger the difference in process adjustment priorities, the greater the priority difference between the two processes in energy consumption optimization, and the larger the value of the conflict penalty coefficient.

[0171] Step S444: Solve the resource allocation optimization model using an integer programming algorithm to determine the resource allocation quantity and allocation time point for each process execution node within the optimal migration period.

[0172] In one implementation, step S444 may specifically include: Step S4441: Use a greedy algorithm to generate an initial feasible solution for the resource allocation optimization model. Allocate available resources to each process in descending order of process resource requirements until all resources are allocated or all processes have obtained a basic resource quota. The basic resource quota is the minimum amount of resources required to maintain the process.

[0173] For example, based on the previously generated process resource requirement priorities, resources are allocated to each process sequentially, starting with the highest priority process. During the allocation process, the needs of high-priority processes are met first, until all resources are allocated or all processes have obtained their basic resource quotas. The basic resource quota is the minimum amount of resources required to maintain a process. For example, a welding process requires at least one welding machine and one operator with the corresponding skills to perform; these constitute the basic resource quota for that process.

[0174] Step S4442: Calculate the objective function value corresponding to the initial feasible solution, identify the key decision variables that have the greatest impact on the objective function value through sensitivity analysis, and take the key decision variables as the focus of optimization of the integer programming algorithm. Sensitivity analysis calculates the degree of influence of each decision variable on the objective function value through partial derivatives.

[0175] Calculating the objective function value corresponding to the initial feasible solution is to evaluate the quality of that solution. Sensitivity analysis is used to identify the key decision variables that have the greatest impact on the objective function value, and these variables are then the focus of optimization for the integer programming algorithm.

[0176] For example, based on the initial feasible solution, the objective function value of the resource allocation optimization model is calculated. This value reflects the total energy consumption and resource utilization rate of the current resource allocation scheme. Sensitivity analysis is then used to calculate the degree of influence of each decision variable on the objective function value. For instance, analyzing the partial derivatives of decision variables such as the number of equipment allocated, the number of personnel allocated, and the number of raw materials allocated with respect to the objective function value identifies the decision variables with larger absolute values ​​of their partial derivatives; these variables are the key decision variables.

[0177] Step S4443: Construct a branch and bound tree to spatially partition the decision variables. Each branch node corresponds to a set of decision variable value constraints. Delete branches whose objective function value is inferior to the current optimal solution through pruning rules. The pruning rules include two types: upper bound pruning and lower bound pruning.

[0178] For example, decision variables include the number of equipment allocated, the number of personnel allocated, and the number of raw materials allocated. When constructing a branch and bound tree, starting from the root node, the decision variables are divided with value constraints. For example, for the allocation quantity of a certain piece of equipment, its value range can be set to a large interval at the root node, and then this interval can be gradually narrowed through branches. Branches with objective function values ​​worse than the current optimal solution are deleted using pruning rules. Upper bound pruning means that when the upper bound of the objective function of a branch is less than the current optimal solution, the branch is deleted; lower bound pruning means that when the lower bound of the objective function of a branch is greater than the current optimal solution, the branch is deleted. Through pruning operations, unnecessary calculations can be reduced, improving the efficiency of the algorithm.

[0179] Step S4444: Solve the relaxation problem using the linear relaxation method in each branch node. If the optimal solution of the relaxation problem is an integer solution and the objective function value is better than the current optimal solution, then update the current optimal solution; if the optimal solution of the relaxation problem is a non-integer solution, then select the decision variable with the largest fraction for branch processing and generate two new sub-branch nodes.

[0180] Solving the relaxation problem at each branch node using the linear relaxation method transforms the integer programming problem into a linear programming problem. The linear relaxation method ignores the integer constraints of the decision variables, making the solution process relatively simple.

[0181] For example, for the relaxation problem at each branch node, the optimal solution is obtained. If the optimal solution is an integer solution and its objective function value is better than the current optimal solution, it indicates that a better resource allocation scheme has been found, and the current optimal solution is updated. If the optimal solution is a non-integer solution, the decision variable with the largest fractional value is selected for branching. For example, if a decision variable has a non-integer value and a large fractional value, the range of values ​​for this variable is further divided, generating two new sub-branch nodes, one for values ​​greater than and the other for values ​​less than the non-integer value.

[0182] Step S4445: Repeat the branch and bound process until all valid branch nodes are traversed, output the final integer optimal solution, map the decision variable values ​​in the optimal solution to the resource allocation quantity and allocation time point of each process execution node, and generate a detailed execution plan table of the resource allocation scheme. The detailed execution plan table includes the resource allocation quantity, resource start time point and resource release time point of each process within the optimal migration period. The time difference between the resource start time point and the process execution start time point does not exceed the preset resource preparation threshold.

[0183] Step S445: Integrate the resource allocation quantity, allocation time point and process execution sequence adjustment results into a resource allocation scheme to make the resource allocation scheme consistent with the energy consumption optimization target, and finally generate an energy consumption optimization-oriented manufacturing execution instruction that includes process adjustment priority, new execution time period and resource allocation list.

[0184] Integrating the resource allocation quantity, allocation time point, and process execution sequence adjustment results into a resource allocation plan is a comprehensive step in synthesizing the preceding analysis and decisions. The resource allocation plan needs to align with the energy consumption optimization objective, namely, reducing overall energy consumption through reasonable resource allocation and process adjustments. For example, it integrates the resource allocation quantity, allocation time point, and process execution sequence adjustment results for each process. For instance, it combines the resource allocation quantity, resource start time point, and resource release time point for a specific process with the adjusted execution period for that process to form a complete resource allocation plan. The final energy-optimized manufacturing execution instruction includes process adjustment priorities, new execution periods, and a resource allocation list.

[0185] Step S450: Integrate the results of process execution sequence adjustment and resource allocation optimization to generate an energy-optimized manufacturing execution instruction that includes process adjustment priority ranking, new execution time period, and resource allocation list.

[0186] Integrating the results of process execution sequence adjustments with resource allocation optimization results serves as a summary and output of the entire optimization process. The generated energy-optimized manufacturing execution instructions include process adjustment priority ranking, new execution periods, and resource allocation lists, providing comprehensive guidance for factory production execution.

[0187] For example, the previously determined process adjustment priority ranking, the new execution time slots for each process, and the resource allocation list are integrated. The process adjustment priority ranking clarifies the importance and adjustment order of each process in energy consumption optimization; the new execution time slots specify the exact execution time for each process; and the resource allocation list details the allocation of resources such as equipment, personnel, and raw materials required for each process. Through this energy consumption optimization-guided manufacturing execution instruction, the factory can produce according to the optimized plan, effectively reducing energy consumption, improving resource utilization, and achieving sustainable production operations.

[0188] It is understood that the various algorithms involved in the above descriptions of the embodiments of the present invention can be obtained from relevant content in the prior art. To save space, they will not be elaborated on in the embodiments of the present invention. In addition, those skilled in the art can supplement the details based on common knowledge in the art when implementing the solutions of the present invention. For example, they can use normalization to eliminate dimensional conflicts before feature fusion, use interpolation to eliminate dimensional differences, reasonably set thresholds based on historical data, experience or business scenario requirements, train the model based on a general model training method, set the number of layers in the model structure based on actual needs, select activation functions, etc. The present invention will not provide redundant descriptions of overly detailed implementation processes here.

[0189] Please see Figure 2 , Figure 2 This is a schematic diagram of a computer system provided in an embodiment of the present invention. The computer system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the computer system, capable of parsing various instructions and processing various data within the computer system. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, mobile communication interface, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the computer system. The memory 103 is a storage device in the computer system used to store programs and data. It is understood that the memory 103 here can include the computer system's built-in memory, or it can include extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system; this invention does not limit this storage space.

[0190] In one embodiment, the processor 101 executes the smart factory manufacturing optimization method based on energy consumption prediction provided above in the embodiments of the present invention by running a computer program in the memory 103.

Claims

1. A smart factory manufacturing optimization method based on energy consumption prediction, characterized in that, The method includes: The historical energy consumption records continuously collected during the factory manufacturing process are processed by time series structuring. Energy consumption sequence units with continuous time markers are generated by dividing time windows and verifying data correlation. The energy consumption sequence units include the energy consumption fluctuation characteristics of the manufacturing process execution period and the corresponding period. Based on the energy consumption sequence unit, a manufacturing process-energy consumption association knowledge graph is constructed. By identifying the causal relationship between process execution nodes and energy consumption fluctuation characteristics, a structured knowledge representation containing node attributes and edge relationship weights is established. The pre-trained graph neural network model is invoked to perform time-series evolution prediction processing on the knowledge graph of the manufacturing process-energy consumption association, and the energy consumption prediction sequence for the future preset manufacturing cycle is generated by combining the energy consumption sequence unit change trend of historical periods. Based on the energy consumption prediction sequence, the execution order and resource allocation scheme of the current manufacturing process are dynamically scheduled and optimized to generate energy consumption optimization-oriented manufacturing execution instructions that include process adjustment priorities and resource allocation schemes.

2. The method according to claim 1, characterized in that, The process-energy consumption association knowledge graph constructed based on the energy consumption sequence units, establishes a structured knowledge representation including node attributes and edge relationship weights by identifying the causal relationship between process execution nodes and energy consumption fluctuation characteristics. The energy consumption sequence unit is processed by extracting process execution nodes. Independent process execution nodes are divided according to the start and end times of the manufacturing process. Each process execution node corresponds to a unique process identifier and time interval. The energy consumption fluctuation characteristics in the energy consumption sequence unit are classified by fluctuation pattern processing to identify the energy consumption rising segment, stable segment and falling segment corresponding to different process execution nodes, and to generate the time distribution curve of energy consumption fluctuation characteristics. Calculate the temporal adjacency relationship between execution nodes of different processes, determine the sequential order and time interval of consecutively executed processes, and establish temporal association edges between process execution nodes; Analyze the differences in energy consumption fluctuation characteristics of the same process execution node in different historical periods, and extract the mean change rate and peak occurrence time of energy consumption fluctuation characteristics as node attributes; Based on the time distribution curve and node attributes of the energy consumption fluctuation characteristics, the correlation strength value between the process execution node and the energy consumption fluctuation characteristics is calculated by the causal relationship identification algorithm. The correlation strength value is then used as the edge relation weight to fill the edge attributes of the knowledge graph, resulting in a manufacturing process-energy consumption correlation knowledge graph containing node attributes and edge relation weights.

3. The method according to claim 1, characterized in that, The process-energy consumption association knowledge graph constructed based on the energy consumption sequence units, establishes a structured knowledge representation including node attributes and edge relationship weights by identifying the causal relationship between process execution nodes and energy consumption fluctuation characteristics. The energy consumption sequence unit is subjected to outlier filtering processing to delete abnormal data points that exceed the preset energy consumption fluctuation range, and retain the effective energy consumption sequence unit that conforms to the normal energy consumption characteristics of the process. The time intervals of the process execution in the effective energy consumption sequence unit are aligned with the energy consumption values ​​on the time axis so that the time interval of each process execution node is completely matched with the collection time of the corresponding energy consumption fluctuation characteristics. Define the process execution nodes as entities, and use process identifiers, time intervals, and node attributes as the basic descriptive information of the entities to generate a set of nodes for a knowledge graph; The energy consumption fluctuation characteristics of different process execution nodes are similar. Energy consumption fluctuation characteristics with similarity higher than a preset threshold are classified into the same feature category, and the association edge between the feature category and the process execution node is established. The initial value of the edge relationship weight is set according to the type of the associated edge, and the edge relationship weight is adjusted by an iterative optimization algorithm so that the association relationship between nodes and edges in the knowledge graph conforms to the actual fluctuation pattern of historical energy consumption data.

4. The method according to claim 1, characterized in that, The process involves calling a pre-trained graph neural network model to perform time-series evolution prediction processing on the manufacturing process-energy consumption association knowledge graph, and combining the energy consumption sequence unit change trends of historical periods to generate an energy consumption prediction sequence for the future preset manufacturing cycle, including: The manufacturing process-energy consumption association knowledge graph is input into the graph embedding layer of the graph neural network model, and the node attributes and edge relationship weights are vectorized to generate node embedding vectors and edge embedding vectors. The time-series feature injection module of the graph neural network model transforms the energy consumption sequence units of historical periods into a time-series feature matrix in chronological order. The time-series feature matrix is ​​then associated with the node embedding vector element by element to generate a time-series enhanced node embedding vector that integrates time-series information. The graph convolutional layer of the graph neural network model is used to aggregate the neighbor node information of the temporal enhancement node embedding vector. The contribution of the neighbor nodes is weighted and summed by the edge embedding vector to generate aggregated node features containing local neighborhood association information. The temporal prediction layer of the graph neural network model is invoked to perform long short-term memory modeling on the features of the aggregated nodes, extract the dependence of energy consumption fluctuation features in the time dimension, and generate an energy consumption prediction subsequence for each process execution node in the future preset manufacturing cycle. The energy consumption prediction subsequences of all process execution nodes are spliced ​​together according to the process execution order to generate an energy consumption prediction sequence covering the entire manufacturing cycle. The energy consumption prediction sequence includes the total energy consumption prediction value corresponding to each time unit and the energy consumption ratio of each process.

5. The method according to claim 4, characterized in that, The time-series feature injection module of the graph neural network model transforms the energy consumption sequence units of historical periods into a time-series feature matrix in chronological order, and performs element-wise association processing between the time-series feature matrix and the node embedding vector to generate a time-series enhanced node embedding vector that integrates time-series information, including: The energy consumption sequence units of historical periods are processed with uniform time granularity to convert energy consumption data from different collection frequencies into energy consumption value sequences with the same time interval, thus generating an energy consumption time series with equal intervals. The energy consumption time series is input into the time series feature extractor, and the periodic and trend features of energy consumption fluctuations are extracted by the sliding window Fourier transform. The features are then combined to generate a time series feature matrix containing periodic and trend components. The node embedding vector is subjected to dimensional expansion processing, and the dimension of the node embedding vector is adjusted to be consistent with the column dimension of the temporal feature matrix, thereby generating a dimension-matched expanded node embedding vector. The temporal feature matrix and the extended node embedding vector are associated element by element, and the product value of each element is calculated as the fused feature value to generate a preliminary fused feature matrix. The temporal feature injection module performs feature filtering on the preliminary fused feature matrix, retaining feature components whose contribution to energy consumption prediction is greater than the contribution threshold, and generating the final temporal enhancement node embedding vector.

6. The method according to claim 5, characterized in that, The process involves aggregating neighbor node information on the temporal augmentation node embedding vector using the graph convolutional layer of the graph neural network model. This is achieved by weighting and summing the contributions of neighbor nodes using the edge embedding vectors, generating aggregated node features that include local neighborhood association information. Construct an adjacency matrix for the manufacturing process-energy consumption association knowledge graph, determine the search range of neighbor nodes based on the weight of edge relationships, and define nodes with edge relationship weights greater than a preset threshold as direct neighbor nodes; For each node, the temporal augmentation node embedding vector of the direct neighbor nodes is processed by directional labeling, and directional features are added to the neighbor node embedding vector according to the type of edge relationship; Using the edge embedding vector as a weight coefficient, the neighbor node embedding vectors with directional features are weighted and summed to calculate the contribution of each neighbor node to the center node, generating a neighbor contribution vector. The temporal enhancement node embedding vector of the central node is concatenated with the neighbor contribution vector, and the nonlinear transformation function of the input graph convolutional layer is used for feature mapping to generate preliminary aggregated node features. The initial aggregated node features are subjected to layer normalization to generate aggregated node features with uniform distribution characteristics.

7. The method according to claim 1, characterized in that, The step of dynamically scheduling and optimizing the execution order and resource allocation scheme of the current manufacturing process based on the energy consumption prediction sequence, and generating an energy consumption optimization-oriented manufacturing execution instruction that includes process adjustment priorities and resource allocation schemes, includes: The energy consumption prediction sequence is compared and analyzed with the preset energy consumption threshold range to identify the process execution period and the corresponding process execution node where the energy consumption prediction value exceeds the threshold range, and these are marked as process nodes to be optimized. Energy consumption sensitivity assessment is performed on the process nodes to be optimized. The product of the magnitude of the energy consumption prediction value exceeding the threshold and the process importance coefficient is calculated, and the product result is used as a quantitative indicator of the process adjustment priority. Based on the process adjustment priority, the execution order of the highest priority process node to be optimized is adjusted, and it is moved from the high energy consumption period to the low energy consumption period. The adjusted energy consumption prediction subsequence is recalculated. The high energy consumption period is the period in the energy consumption prediction sequence where the predicted energy consumption value is higher than the preset high energy consumption threshold, and the low energy consumption period is the period in the energy consumption prediction sequence where the predicted energy consumption value is lower than the preset low energy consumption threshold. Based on the adjusted energy consumption prediction subsequence, the allocation scheme of manufacturing resources is optimized, and resources are preferentially allocated to process execution nodes with lower energy consumption prediction values ​​and higher process importance coefficients. The lower energy consumption prediction value means that the energy consumption prediction value of the process execution node is lower than the preset energy consumption threshold, and the higher process importance coefficient means that the process importance coefficient of the process execution node is higher than the preset importance coefficient threshold. Integrate the results of process execution sequence adjustment with resource allocation optimization to generate energy-optimized manufacturing execution instructions that include process adjustment priority ranking, new execution time periods, and resource allocation lists.

8. The method according to claim 7, characterized in that, The step of performing energy consumption sensitivity assessment on the process node to be optimized, calculating the product of the magnitude by which the predicted energy consumption exceeds the threshold and the process importance coefficient, and using the product result as a quantitative indicator for process adjustment priority, includes: Extract the difference between the predicted energy consumption value of the process node to be optimized and the preset energy consumption threshold, and define the ratio of the difference to the preset energy consumption threshold as the energy consumption exceedance range. The importance coefficient of each process execution node is determined by statistical analysis of historical manufacturing data; The energy consumption exceedance and process importance coefficient are standardized and converted into dimensionless values ​​with the same dimension. Calculate the product of the standardized energy consumption excess and the process importance coefficient, and use the product result as the initial quantitative indicator for process adjustment priority; The initial quantitative indicators of all process nodes to be optimized are sorted, and the sorting results are used as the final process adjustment priority.

9. The method according to claim 7, characterized in that, The step of adjusting the execution order of the highest-priority process node to be optimized based on the process adjustment priority, moving it from a high-energy-consumption period to a low-energy-consumption period, and recalculating the adjusted energy consumption prediction subsequence includes: Determine the distribution of energy consumption forecast values ​​for each time period within the current manufacturing cycle, mark the time periods with energy consumption forecast values ​​lower than the preset low energy consumption threshold as low energy consumption time periods, and generate a list of low energy consumption time periods; Select target low-energy consumption periods from the list of low-energy consumption periods that match the execution duration of the process node to be optimized, so that the duration of the target low-energy consumption period is not less than the execution duration of the process node to be optimized. Calculate the energy consumption change after the process node to be optimized is migrated to different target low energy consumption periods, and select the target low energy consumption period with the smallest energy consumption change as the optimal migration period; Adjust the time interval of the process node to be optimized to the optimal migration period, update the execution sequence table of the manufacturing process, and record the time interval change information before and after the process adjustment. Based on the adjusted execution order table, the graph neural network model is called again to perform time-series evolution prediction processing on the knowledge graph of manufacturing process-energy consumption association, generating an adjusted energy consumption prediction subsequence, and verifying whether the adjusted energy consumption prediction value is within the preset threshold range.

10. A computer system, characterized in that, include: A memory, wherein a computer program is stored; A processor for loading the computer program to implement the smart factory manufacturing optimization method based on energy consumption prediction as described in any one of claims 1-9.

Images