Machine learning based device energy consumption collaborative intelligent optimization method and system

Abstract

The application provides a device energy consumption collaborative intelligent optimization method and system based on machine learning. The method comprises: acquiring multi-source data in the production, supply and use processes of an energy system, and writing fact time parameters and storage time parameters for the multi-source data; performing windowed time alignment on the multi-source data based on the fact time parameters, and performing backfilling and merging on late data to form a unified time series data set; constructing an energy unit graph model based on the unified time series data set; training an energy consumption prediction model and outputting multi-step energy consumption prediction results; performing hierarchical classification processing on alarm event objects to generate alarm event structured vectors; within a rolling scheduling window, solving a collaborative optimization scheduling result, and converting the collaborative optimization scheduling result into execution instructions for device objects for updating device running states or energy supply running parameters. The application can improve the data fusion efficiency of production, supply and use, improve the energy consumption prediction accuracy, and enhance the stability of rolling scheduling issuance.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and industrial internet technology, and in particular to a method and system for collaborative intelligent optimization of equipment energy consumption based on machine learning. Background Technology

[0002] In the context of the Industrial Internet and smart manufacturing, energy systems typically encompass multiple energy sources, including electricity, steam, compressed air, and refrigerants / heaters, involving the coordinated operation of production processes, energy supply facilities, and energy-consuming equipment. To reduce energy consumption and support production organization, enterprises generally deploy energy management systems and energy consumption monitoring platforms. These systems achieve energy management by collecting metering data, conducting statistical analysis, and generating reports for performance evaluation. Furthermore, predictive models or scheduling strategies are introduced to optimize energy supply and consumption.

[0003] In existing technologies, multi-source energy consumption data is typically scattered across control systems, production execution systems, and energy management platforms. Equipment, processes, media, metering points, and alarm events lack unified object identification and a unified time reference, making data alignment and integration difficult. Furthermore, historical window data is prone to repeated corrections during network jitter, delayed reporting, and supplementary data collection, affecting the consistency of prediction and scheduling. On the other hand, alarm management is mostly event-centric, allowing for hierarchical classification and suppression, but alarm results are difficult to translate into constraints for energy consumption optimization, leading to a disconnect between faults, alarms, and energy balance scheduling. Moreover, prediction and scheduling often employ centralized computation, which, constrained by real-time data throughput, latency, and edge execution feasibility, struggles to stably output directly executable device-level commands within a rolling window, making it difficult for real-time data to support closed-loop applications for collaborative energy consumption optimization. Summary of the Invention

[0004] In view of this, embodiments of this application provide a machine learning-based collaborative intelligent optimization method and system for equipment energy consumption, in order to solve the problems of difficulty in unifying and aligning multi-source data of production, supply and consumption, difficulty in constraining alarms to participate in scheduling, and difficulty in stably issuing rolling scheduling instructions in the existing technology.

[0005] A first aspect of this application provides a machine learning-based collaborative intelligent optimization method for equipment energy consumption, comprising: acquiring multi-source data from the production, supply, and consumption processes of an energy system; establishing an association mapping based on a unified object identifier for equipment objects, process objects, energy medium objects, metering point objects, and alarm event objects; and writing factual time parameters and storage time parameters for the multi-source data; performing windowed time alignment on the multi-source data based on the factual time parameters; and performing backfilling and merging on late data based on the storage time parameters to form a unified time-series dataset; constructing an energy unit graph model based on the unified time-series dataset, and configuring energy balance constraints and schedulable parameters, wherein the energy unit... The metagraph model is used to represent the energy flow relationship between equipment objects, process objects, and energy supply objects; an energy consumption prediction model is trained based on the energy unit graph model and a unified time series dataset, and multi-step energy consumption prediction results are output; alarm event objects are subjected to hierarchical classification processing to generate alarm event structured vectors carrying priority and associated object identifier sets, and the alarm event structured vectors are converted into collaborative scheduling constraints; within the rolling scheduling window, the collaborative optimization scheduling results are solved based on the multi-step energy consumption prediction results, energy balance constraints, and collaborative scheduling constraints, and the collaborative optimization scheduling results are converted into execution instructions for equipment objects to update equipment operating status or energy supply operating parameters.

[0006] A second aspect of this application provides a machine learning-based equipment energy consumption collaborative intelligent optimization system, comprising: an acquisition module for acquiring multi-source data from the production, supply, and consumption processes of an energy system, establishing association mappings based on unified object identifiers for equipment objects, process objects, energy medium objects, metering point objects, and alarm event objects, and writing factual time parameters and storage time parameters for the multi-source data; a merging module for performing windowed time alignment on the multi-source data based on the factual time parameters, and performing backfilling and merging on late data based on the storage time parameters to form a unified time-series dataset; and a construction module for constructing an energy unit graph model based on the unified time-series dataset, and configuring energy balance constraints and schedulable parameters, wherein the energy unit... The metagraph model is used to represent the energy flow relationship between equipment objects, process objects, and energy supply objects; the training module is used to train the energy consumption prediction model based on the energy unit graph model and the unified time series dataset and output multi-step energy consumption prediction results; the generation module is used to perform hierarchical classification processing on alarm event objects to generate alarm event structured vectors carrying priority and associated object identifier sets, and convert the alarm event structured vectors into collaborative scheduling constraints; the conversion module is used to solve the collaborative optimization scheduling results based on the multi-step energy consumption prediction results, energy balance constraints, and collaborative scheduling constraints within the rolling scheduling window, and convert the collaborative optimization scheduling results into execution instructions for equipment objects to update the equipment operating status or energy supply operating parameters.

[0007] A third aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0008] The above-described technical solutions adopted in the embodiments of this application can achieve the following beneficial effects: By acquiring multi-source data from the energy system's production, supply, and consumption processes, a correlation mapping is established based on unified object identifiers for equipment objects, process objects, energy medium objects, metering point objects, and alarm event objects. Real-time parameters and storage entry time parameters are written to the multi-source data. Windowed time alignment is performed on the multi-source data based on the real-time parameters, and late data is backfilled and merged based on the storage entry time parameters to form a unified time-series dataset. An energy unit graph model is constructed based on the unified time-series dataset, and energy balance constraints and schedulable parameters are configured. The energy unit graph model is used to represent equipment objects, process objects, and energy medium objects. This application establishes the energy flow relationship between energy objects; trains an energy consumption prediction model based on an energy unit graph model and a unified time-series dataset, and outputs multi-step energy consumption prediction results; performs hierarchical classification processing on alarm event objects to generate structured vectors of alarm events carrying priority and associated object identifier sets, and converts the structured vectors of alarm events into collaborative scheduling constraints; within the rolling scheduling window, solves the collaborative optimization scheduling results based on the multi-step energy consumption prediction results, energy balance constraints, and collaborative scheduling constraints, and converts the collaborative optimization scheduling results into execution instructions for equipment objects to update equipment operating status or energy supply operating parameters. This application can improve the efficiency of production, supply, and consumption data fusion, improve the accuracy of energy consumption prediction, and enhance the stability of rolling scheduling. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating the machine learning-based collaborative intelligent optimization method for device energy consumption provided in an embodiment of this application. Figure 2 This is a schematic diagram of the structure of the machine learning-based device energy consumption collaborative intelligent optimization system provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0011] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0012] In existing technologies, enterprises mostly collect metering data through energy management systems and energy consumption monitoring platforms and perform statistical analysis. In some scenarios, predictive models or scheduling strategies are introduced to optimize energy supply and consumption. However, production, supply, and consumption-related data are usually scattered among control systems, production execution systems, and energy management platforms. Equipment, processes, energy media, metering points, and alarm events lack unified object identification and a unified time reference. Late reporting and supplementary data collection can easily lead to repeated corrections of window data, resulting in unstable predictive inputs and scheduling basis. At the same time, alarm classification and grading are mostly limited to the event management level and are difficult to transform into constraints that can be directly used for energy consumption optimization. Faults, alarms, and energy balance scheduling are disconnected. Furthermore, due to the constraints of real-time data throughput and latency, rolling scheduling results are difficult to reliably convert into equipment-level instructions that can be issued.

[0013] Based on this, the technical problem to be solved by this application is that it is difficult to unify and align multi-source data of production, supply and use, it is difficult to constrain alarm information to participate in collaborative scheduling, and it is difficult to reliably issue rolling scheduling instructions.

[0014] To address the aforementioned technical issues, this application provides a machine learning-based collaborative intelligent optimization method for equipment energy consumption. This method establishes a unified object identifier association mapping for equipment objects, process objects, energy medium objects, metering point objects, and alarm event objects, and writes factual time parameters and storage time parameters into multi-source data. Based on the factual time parameters, windowed time alignment is performed, and late data is supplemented and merged based on the storage time parameters to form a unified time-series dataset. Based on the unified time-series dataset, an energy unit graph model representing energy flow relationships is constructed, and energy balance constraints and schedulable parameters are configured. Based on the energy unit graph model and the unified time-series dataset, an energy consumption prediction model is trained, outputting multi-step energy consumption prediction results. Alarm event objects are classified and categorized to generate structured vectors of alarm events carrying priority and associated object identifier sets, which are then converted into collaborative scheduling constraints. Within a rolling scheduling window, collaborative optimization scheduling results are solved based on multi-step energy consumption prediction results, energy balance constraints, and collaborative scheduling constraints, and the scheduling results are converted into execution instructions for equipment objects to update equipment operating status or energy supply operating parameters.

[0015] Through the above-described technical solution of this application, this application can improve the efficiency of data fusion between production, supply and consumption, improve the accuracy of energy consumption prediction, and enhance the stability of rolling scheduling instructions, thereby supporting the collaborative optimization and continuous operation management of equipment energy consumption.

[0016] The technical solution of this application will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0017] Figure 1 This is a flowchart illustrating the machine learning-based collaborative intelligent optimization method for device energy consumption provided in an embodiment of this application. Figure 1 As shown, the method may specifically include: S101: Obtain multi-source data from the production, supply, and consumption processes of the energy system; establish an association mapping based on a unified object identifier for equipment objects, process objects, energy medium objects, metering point objects, and alarm event objects; and write fact time parameters and storage time parameters for the multi-source data. S102, perform windowed time alignment on multi-source data based on the fact time parameter, and perform backfilling and merging on late data based on the inbound time parameter to form a unified time series dataset; S103, construct an energy unit graph model based on a unified time series dataset, and configure energy balance constraints and schedulable parameters. The energy unit graph model is used to characterize the energy flow relationship between equipment objects, process objects and energy supply objects. S104, based on the energy unit graph model and the unified time series dataset, trains the energy consumption prediction model and outputs multi-step energy consumption prediction results; S105, perform hierarchical classification processing on alarm event objects to generate alarm event structured vectors carrying priority and associated object identifier sets, and convert alarm event structured vectors into collaborative scheduling constraint terms; S106, within the rolling scheduling window, solve the collaborative optimization scheduling result based on the multi-step energy consumption prediction result, energy balance constraint and collaborative scheduling constraint, and convert the collaborative optimization scheduling result into execution instructions for equipment objects to update the equipment operating status or energy supply operating parameters.

[0018] It should be noted that the following embodiments of this application will use the industrial internet platform application of a pharmaceutical factory as a case background for description. However, it should be understood that the application scenarios of the technical solution of this application are not limited to the scenario of a pharmaceutical factory. This solution is applicable to energy management systems and energy consumption monitoring platforms in any intelligent manufacturing scenario. The following scenarios are only examples and do not constitute a limitation on the applicable scenarios of this solution.

[0019] In some embodiments, a correlation mapping is established based on a unified object identifier for equipment objects, process objects, energy medium objects, metering point objects, and alarm event objects, and fact time parameters and data entry time parameters are written for multi-source data, including: Global object identifiers are assigned to various types of objects, and object association entries are generated based on preset object relationship rules. Object association entries are used to indicate the association relationship between equipment objects and process objects, energy medium objects, metering point objects, and alarm event objects. When receiving multi-source data associated with each object, extract the collection time from the multi-source data and write it into the fact time parameter, and obtain the access time of the multi-source data entering the data processing link and write it into the database entry time parameter. Perform consistency checks on multi-source data to generate check tags. When the check passes, bind and encapsulate the multi-source data with the corresponding global object identifier, object association entry, fact time parameter, entry time parameter, and check tag to generate a unified data record that can be used for subsequent windowed time alignment.

[0020] Specifically, in a scenario involving the collaboration of energy and production systems in a pharmaceutical factory, a data foundation for intelligent optimization of energy consumption is constructed. Pharmaceutical factories typically include production, supply, and consumption stages such as purified water preparation, water for injection preparation, cleanroom air conditioning, compressed air, steam supply, solution preparation, filling, freeze-drying, and packaging. Multiple energy media, including electricity, steam, chilled water, hot water, and compressed air, are simultaneously present on-site. Since energy consumption data, process data, and alarm data originate from different systems and acquisition links, this embodiment establishes a mapping relationship between equipment objects, process objects, energy medium objects, metering point objects, and alarm event objects through a unified object identifier and a dual-time parameter mechanism, forming a unified data record that can be used for subsequent windowed time alignment.

[0021] Global object identifiers are coded carriers that uniquely identify equipment objects, process objects, energy medium objects, metering point objects, and alarm event objects within a pharmaceutical factory. They are used to maintain consistent referencing of the same object during cross-system and cross-link data processing, avoiding object confusion caused by inconsistent names in different systems.

[0022] Object association entries are relational description units constructed using global object identifiers as indexes. They are used to express the association relationships between equipment objects and process objects, energy medium objects, metering point objects, and alarm event objects. Object association entries include association direction, association type, and association constraint information. The association type is used to indicate relationship categories such as "equipment belongs to process", "equipment consumes medium", "metering point metering object", and "alarm event associated object".

[0023] The fact time parameter is the moment when the physical facts reflected by the multi-source data occur or are collected, used to characterize the actual production and energy consumption status corresponding to the data; the storage time parameter is the moment when the multi-source data enters the data processing link and is received by the access side, used to characterize the order and delay of the data arriving in the system, thereby supporting the identification, replenishment and traceability of late data.

[0024] In its implementation, this embodiment first assigns global object identifiers to various types of objects and generates object association entries based on preset object relationship rules. Taking a pharmaceutical factory as an example, global object identifiers can be generated by combining object category, factory area, production line, level, and instance number. For example, equipment objects can be identified by a hierarchical structure of "factory area-workshop-system-equipment type-instance number," process objects can be identified by a hierarchical structure of "production line-section-process section-batch unit," energy medium objects can be identified by "medium type-power supply station-pipeline area-branch number," metering point objects can be identified by "metering medium type-installation location-loop number," and alarm event objects can be identified by "alarm source object identifier-alarm type-trigger sequence number."

[0025] In some examples, pre-defined object relationship rules are used to constrain which objects need to be associated and the direction and type of such associations. These include process affiliation rules, energy medium consumption rules, metering mapping rules, and alarm association rules. Process affiliation rules determine which process an object belongs to; for example, a liquid preparation tank, circulating pump, and temperature control unit belong to the "liquid preparation process." Energy medium consumption rules determine which energy medium objects an object consumes; for example, a cleanroom air conditioning unit consumes electricity and chilled water, while a purified water preparation system consumes electricity and steam. Metering mapping rules determine which object's energy consumption or flow rate a metering point measures; for example, an electricity meter measures the electricity consumption of a freeze dryer circuit, while a steam flow meter measures the steam flow rate of a heat exchange station branch. Alarm association rules determine which objects an alarm event object is associated with; for example, a "freeze dryer vacuum pump fault alarm" is associated with both the freeze dryer object and the freeze drying process object, and can also be associated with its corresponding electricity metering point object to support subsequent energy consumption anomaly analysis. After object association entries are generated, they can be indexed and stored using global object identifiers, allowing any subsequent data record to quickly locate associated processes, media, metering points, and alarm events through its object identifier.

[0026] Furthermore, when receiving multi-source data associated with each object, the acquisition time is extracted from the multi-source data and written into the factual time parameter, and the access time of the multi-source data entering the data processing link is obtained and written into the storage time parameter. Data sources in pharmaceutical factories typically include equipment operation data acquired by the control system, electricity and flow data acquired by the energy metering system, batch and process status data generated by the production execution system, and alarm event data output by the alarm system. For control system data, the acquisition time is usually written by the control system during sampling; for metering system data, the acquisition time is written by the metering equipment or acquisition gateway; for production execution system data, the acquisition time can be the time when the process status change occurs; for alarm event data, the acquisition time can be the alarm trigger time.

[0027] The data entry time parameter is written by the data access side in this embodiment when it receives the data. This parameter is independent of whether the data source provides a timestamp, thus allowing it to be used to identify delayed data caused by network jitter, link congestion, or offline cache reporting. In some scenarios, the metering gateway may report the power increment every 60 seconds, but due to network jitter, a report may be delayed. In this case, the factual time parameter of the data still corresponds to the original collection time, while the data entry time parameter is significantly later than the data entry time parameters of other data in the same window. This difference will serve as the basis for subsequent lateness determination and compensation processing.

[0028] Furthermore, consistency checks are performed on the multi-source data to generate check tags. Upon successful check, the multi-source data is bound and encapsulated with the corresponding global object identifier, object association entry, fact time parameter, database entry time parameter, and check tag, generating a unified data record that can be used for subsequent windowed time alignment. Consistency checks include object consistency checks, time consistency checks, structural consistency checks, and value consistency checks.

[0029] Object consistency verification checks whether the object identifier carried in a data record can be resolved in the global object identifier system and whether the corresponding object association entry can be located. If it cannot be resolved, a verification flag indicating a missing object is generated and the data record is set to a pending completion state. Time consistency verification checks whether the actual time parameter falls within the allowed time range and whether the time interval with adjacent sampled records meets the preset sampling rules. If reverse order or excessively large intervals occur, a verification flag indicating time anomalies is generated.

[0030] The structural consistency check is used to verify whether the field set, field type, and unit of measurement label meet the preset structural constraints. For example, the electricity field must carry a unit label and be convertible to a uniform unit. If the unit is missing or the field type does not match, a structural anomaly label is generated. The value consistency check is used to verify whether key fields are within a reasonable range. For example, steam flow rate should not be negative, and the return water temperature of chilled water in clean air conditioning should not exceed the preset upper limit. If the limits are exceeded, a value anomaly label is generated.

[0031] Upon successful verification, the global object identifier of the data record is bound to its corresponding object association entry. Simultaneously, the fact time parameter and the entry time parameter are written to the record, along with a verification flag, forming a unified data record. This unified data record logically becomes the smallest data unit for subsequent windowed time alignment, enabling subsequent processing to complete window-based aggregation, object-based alignment, and relationship-based fusion without needing to query object relationships across systems.

[0032] The above process is further illustrated below with a specific example from a pharmaceutical factory. Taking an injectable drug production line as an example, the freeze-drying process includes equipment objects such as freeze dryers, vacuum pump sets, and condensation systems. The energy supply side includes energy supply objects such as steam heat exchange stations and refrigeration rooms. Energy medium objects include steam, chilled water, and electricity. Metering point objects include freeze dryer meters, steam branch flow meters, and chilled water branch flow meters. First, global object identifiers are assigned to freeze dryers, vacuum pump sets, steam heat exchange stations, refrigeration rooms, steam media, chilled water media, and freeze dryer meters, and object association entries are generated. For example, "freeze dryer equipment object identifier associated with freeze-drying process object identifier", "freeze dryer equipment object identifier associated with electrical medium object identifier and freeze dryer meter metering point object identifier", "steam heat exchange station energy supply object identifier associated with steam media object identifier and steam flow meter metering point object identifier", and "refrigeration room energy supply object identifier associated with chilled water media object identifier and chilled water flow meter metering point object identifier".

[0033] For example, when the freeze dryer's electrical meter collects the electricity increment at 10:00:00 and is received by the data access side at 10:00:02, the fact time parameter is written as 10:00:00 and the storage time parameter is written as 10:00:02. When the steam flow meter collects the steam flow at 10:00:00 but arrives at the access side at 10:01:10 due to network jitter, the fact time parameter is written as 10:00:00 and the storage time parameter is written as 10:01:10. The system performs a consistency check on the above data. After confirming that the object identifier can be parsed, the field unit can be converted, and the value is within the range, a check mark is generated as passed, and the multi-source data is bound and encapsulated with the corresponding object-related entries to form a unified data record. During subsequent windowed time alignment, the two records can be directly grouped into the same window based on the fact time parameter, and the steam data can be identified as late data based on the storage time parameter, providing a clear basis for subsequent backfilling and merging.

[0034] Through the implementation method of this embodiment, multi-source data in the production, supply and use processes obtain stable association mapping under a unified object identification system, and form a traceable dual time description through fact time parameters and warehousing time parameters. Combined with consistency verification and binding encapsulation, a unified data record is obtained, thereby providing a consistent data entry and reference basis for subsequent windowed time alignment, late data replenishment and merging, and energy unit graph model construction.

[0035] In some embodiments, windowed time alignment is performed on multi-source data based on a fact time parameter, and late data is backfilled and merged based on an inbound time parameter to form a unified time-series dataset, including: Window identifiers are assigned to multi-source data based on fact time parameters. The window identifiers are used to indicate the windowed time range to which the multi-source data belongs. Under the same window identifier, multi-source data are aggregated according to the global object identifier to generate object window data. For object window data, an alignment key is generated based on the object association entries. The object window data of equipment objects, process objects, energy medium objects, metering point objects, and alarm event objects are time-aligned according to the alignment key to form aligned window data. When the inbound time parameter and the deadline corresponding to the window identifier meet the preset late judgment conditions, the corresponding multi-source data will be identified as late data and a replenishment mark will be generated for the late data. Perform a backfill merging process on late data carrying a backfill flag. The backfill merging process includes writing the late data into the aligned window data corresponding to the window identifier and recording the backfill version flag. A traceable unified time-series dataset is generated based on the back-version tag. The unified time-series dataset includes aligned window data and back-version tags associated with the aligned window data.

[0036] Specifically, object window data is a data set or aggregation result obtained by aggregating multi-source data corresponding to the same global object identifier under the same window identifier. It is used to characterize information such as the object's operating status, energy consumption metering, media flow and alarm status within the window.

[0037] Alignment key is the association index information generated based on the object association entries. It is used to align and match data of different categories of objects within the same window identifier according to a predetermined association relationship. Alignment key includes target global object identifier, associated object identifier, and association type marker.

[0038] Late data refers to data records whose entry time parameter meets the preset late judgment conditions relative to the window deadline. The late judgment conditions include exceeding the preset late threshold or crossing the preset replenishment boundary.

[0039] The backfill flag is a processing identifier for late data, used to indicate that the data needs to be written into the already formed alignment window data and trigger backfill merging.

[0040] A unified time-series dataset is a collection of data organized in window order by multiple window identifiers corresponding to aligned window data, and carries a backfill version tag associated with each aligned window data to support subsequent window-by-window playback, tracing and incremental updates.

[0041] In practical implementation, firstly, window identifiers are assigned to multi-source data based on factual time parameters. Then, under the same window identifier, multi-source data is aggregated according to global object identifiers to generate object window data. In this embodiment, the preset window granularity can be 60s or 300s to balance computational load and scheduling accuracy. After obtaining unified data records, the data access side locates the start and end times of the window to which it belongs based on the factual time parameters in the unified data records, and generates corresponding window identifiers. For multiple records within the same window identifier, the records are divided into aggregation slots of the corresponding objects according to the global object identifiers, thereby forming object window data.

[0042] In some examples, the aggregation method for object window data can be customized based on the object type. For instance, equipment object window data can aggregate operating status, load, start / stop events, and power statistics; energy medium object window data can aggregate supply-side flow, pressure, and temperature statistics; metering point object window data can aggregate power increments, flow integrals, or instantaneous value sequences; and alarm event object window data can aggregate newly added alarms, confirmed alarms, and alarm clearing events within the window, while retaining priority and impact domain labels. To ensure subsequent traceability, the aggregation process can also retain a summary of the entry time parameters for each record within the window to pinpoint late data insertion points.

[0043] Furthermore, for the object window data, alignment keys are generated based on object association entries. The object window data of equipment objects, process objects, energy medium objects, metering point objects, and alarm event objects are then time-aligned according to these alignment keys to form aligned window data. In this embodiment, the alignment key is generated driven by object association entries, typically using a process object or equipment object as the alignment center object. Taking a pharmaceutical factory as an example, the process object can be used as the upper-level alignment center object, and the equipment objects, medium objects, metering point objects, and alarm event objects associated with this process object can be attached to the same alignment structure via alignment keys.

[0044] Specifically, the system reads the object association entries of a certain process object, extracts its associated equipment object identifier list, energy medium object identifier list, metering point object identifier list, and alarm event object identifier list, and generates an alignment key set for each type of association, containing "central object identifier - associated object identifier - association type tag". Within the same window identifier, the system extracts the data segment corresponding to the central object from the object window data of each associated object according to the alignment key set, and forms the alignment window data. Structurally, the alignment window data can contain the central object window data and the associated object window data linked by the alignment key index, so that the production, supply, and usage status and alarm status within the window are aligned in the same time slice.

[0045] Furthermore, when the entry time parameter and the deadline corresponding to the window identifier meet the preset lateness determination conditions, the corresponding multi-source data is identified as late data, and a replenishment mark is generated for the late data. In this embodiment, the lateness determination conditions include at least two types: one is timeout threshold determination, that is, the entry time parameter is later than the window deadline and exceeds the preset lateness threshold; the other is cross-window boundary determination, that is, the entry time parameter enters a subsequent window or a later window, and the aligned window data of the window corresponding to the data has been formed and read by the downstream module. The preset lateness threshold can be configured differently according to the data source category. For example, the threshold for data collected by the control system is relatively small, while the threshold for metering system and alarm system can be appropriately relaxed. A replenishment mark is written to the data record determined to be late, and the replenishment mark is bound to its window identifier and global object identifier to ensure accurate positioning of subsequent replenishment writing.

[0046] Furthermore, a backfilling and merging process is performed on the late data carrying the backfilling mark. The late data is written into the aligned window data corresponding to the window identifier and the backfilling version mark is recorded. In this embodiment, the backfilling and merging process is performed based on the principle of "partial update within the window". The system locates the target aligned window data according to the window identifier of the late data, determines the object window data slot to be written according to the global object identifier of the late data, and determines its attachment position in the aligned window data according to the alignment key.

[0047] In some examples, the write operation can employ overwrite, append, or aggregate recalculation methods: for metering incremental data, aggregate recalculation can be performed to update the window's cumulative value; for alarm event data, appending can be performed and the alarm status summary within the window can be refreshed; for device status data, the window statistics can be inserted and recalculated according to the fact-time parameter order. After the write operation is complete, a new backfill version tag is generated and bound to the aligned window data, while retaining the version tag and differential summary of the previous version for subsequent traceability. To prevent downstream modules from reading inconsistent data, this embodiment can set a version lock or read consistency tag on the aligned window data, ensuring that downstream modules use the specified version tag when reading.

[0048] Furthermore, a traceable unified time-series dataset is generated based on the backdated version markers. This unified time-series dataset includes aligned window data and backdated version markers associated with the aligned window data. In this embodiment, the unified time-series dataset is organized according to window identifier order, and each window entry contains at least aligned window data and a corresponding backdated version marker. For windows that have undergone backdating, the unified time-series dataset can simultaneously retain the latest version marker and the reference information of the previous version marker, enabling subsequent energy consumption prediction model training and scheduling to select either "fixed version input" or "latest version input." In some implementations, when the backdated version marker changes and the window is within the model training sample coverage area, the version change information can be further output to the model training module to trigger a local incremental update of the affected sample window range, thereby maintaining data and model consistency.

[0049] The above process is illustrated below using a pharmaceutical factory example. Taking the freeze-drying process as an example, the preset window granularity is 300 seconds, and the corresponding window identifier can be represented as 10:00:00 to 10:04:59. The freeze-dryer's electricity meter collects the electricity increment at 10:02:00 and enters the data into the database at 10:02:02. This record is assigned to the aforementioned window identifier and aggregated into the object window data of the freeze-dryer equipment object and the electricity metering point object. The steam flow meter collects the flow rate at 10:03:30, but due to network jitter, it is not entered into the database until 10:06:10. Its actual time parameter still falls within the 10:00:00 to 10:04:59 window, but the entry time parameter has exceeded the window deadline and meets the late threshold. Therefore, it is judged as late data and a replenishment mark is generated. The system locates the aligned window data for the period from 10:00:00 to 10:04:59, writes the delayed steam flow record into the object window data of the steam medium object and the steam flow meter metering point object, and attaches it to the freeze-drying process alignment structure using the alignment key. Simultaneously, it recalculates the cumulative steam volume for this window and generates a new backfill version marker, for example, updating from version 1 to version 2. The unified time-series dataset records the version 2 corresponding to this window and retains the reference information for version 1. This allows the subsequent energy unit diagram model to directly retrieve version 2 data when extracting steam supply parameters for this window, and also enables the identification of the reason for and content of version changes when traceability is required.

[0050] Through the implementation method of this embodiment, multi-source data can be aggregated and time-aligned under a unified window identifier based on fact time parameters. It can also use the entry time parameter to deterministically identify and replenish late data, and combine the replenished version tag to form a traceable unified time series dataset. This improves the stability and consistency of cross-system data alignment, reduces the disturbance of late data to subsequent model training and rolling scheduling input, and provides a reusable time series foundation for subsequent energy unit graph model construction and collaborative optimization solution.

[0051] In some embodiments, an energy unit graph model is constructed based on a unified time-series dataset, and energy balance constraints and schedulable parameters are configured, including: Based on global object identifiers and object association entries, the set of nodes and the set of edges of the energy unit graph model are determined; Configure medium type labels and metering point object identifiers for each edge in the edge set, and extract the medium flow parameters or power parameters corresponding to the metering point object identifiers based on the unified time series dataset to determine the time-varying energy flow parameters of each edge; Configure a set of schedulable parameters for each node in the node set, and determine the range of parameter values ​​corresponding to the set of schedulable parameters based on a unified time series dataset; Based on the time-varying energy flow parameters of edges and the set of schedulable parameters of nodes, energy balance constraints are constructed within a preset scheduling window. The energy balance constraints are used to limit the constraint relationship between the supply, consumption and storage changes within the same scheduling window. The set of nodes, the set of edges, the medium type label, the time-varying energy flow parameters, the set of schedulable parameters, and the energy balance constraints are encapsulated into an energy unit graph model for subsequent energy consumption prediction model training and collaborative optimization scheduling result solving.

[0052] Specifically, the energy unit graph model is a model description data used to represent energy system objects and their energy flow relationships in a graph structure, where nodes are used to represent energy-related objects and edges are used to represent the energy medium flow or energy conversion relationship between nodes.

[0053] Time-varying energy flow parameters are a set of parameters associated with edges that change over a time window. They are used to characterize the intensity of energy supply, transmission, or consumption within different windows. Time-varying energy flow parameters can be determined by the medium flow rate parameters, power parameters, or their integral quantities corresponding to the metering point object identifier.

[0054] The schedulable parameter set is a set of parameters associated with a node that can be adjusted by the schedulable solution process. It describes the operating parameters of a device or power supply object that can be set or constrained within the scheduling window, including one or more of the following: start / stop status parameters, load level parameters, ramp constraint parameters, and minimum duration parameters.

[0055] Energy balance constraints are constraint descriptions established for a preset scheduling window. They are used to limit the constraint relationship between supply, consumption and storage changes within the same window. Energy balance constraints can further include constraint terms for metering error relaxation and missing data compensation to adapt to actual metering deviations and incomplete data scenarios.

[0056] In practical implementation, the node set and edge set of the energy unit graph model are first determined based on global object identifiers and object association entries. In this embodiment, the node set includes equipment object nodes, process object nodes, and energy supply object nodes, and energy storage object nodes can be introduced according to the characteristics of the energy system of the pharmaceutical factory, such as chilled water buffer tanks, hot water insulation tanks, or compressed air storage tanks. Object association entries provide a structural basis for the generation of nodes and edges. For example, "equipment object belongs to process object" is used to establish the association between equipment object nodes and process object nodes; "equipment object consumes energy medium object" is used to establish the edge related to the medium channel between the equipment object node and the medium channel; "energy supply object generates or converts energy medium object" is used to establish the edge related to the medium channel between the energy supply object node and the medium channel; and "metering point object metering object" is used to bind the metering point object identifier to the edge or node to support parameter extraction.

[0057] Edge sets are used to characterize the energy medium flow relationships between node sets, including supply edges from energy supply objects to energy medium channels, consumption edges from energy medium channels to energy-consuming equipment, process energy consumption mapping edges from energy-consuming equipment to process objects, and charging and discharging edges of energy storage objects. In some implementations, process objects can be regarded as upper-level aggregation nodes to carry out energy consumption accounting and scheduling constraints at the production line level, so that equipment-level decisions can be mapped to process-level goals.

[0058] Furthermore, each edge in the edge set is configured with a medium type tag and a metering point object identifier, and the medium flow rate parameter or power parameter corresponding to the metering point object identifier is extracted based on the unified time series dataset to determine the time-varying energy flow parameters of each edge. In this embodiment, each edge is bound to at least one medium type tag to distinguish the dimensions and metering methods of different media. For example, a dielectric edge can be bound to a power metering point object identifier or an electric power metering point object identifier, a steam medium edge can be bound to a steam flow meter object identifier and combined with enthalpy or temperature and pressure parameters to form an equivalent energy quantity, and a chilled water medium edge can be bound to a flow meter object identifier and combined with the supply and return water temperature difference to form a cooling quantity parameter. The unified time series dataset provides metering point object window data aligned to a window. The system reads the data of the corresponding metering point object identifier in each preset scheduling window to obtain the power parameter, flow rate parameter or its cumulative amount for that window, and writes it into the time-varying energy flow parameters of that edge. If the unified time series dataset carries a backfill version tag, this embodiment can preferentially select the aligned window data corresponding to the latest version tag to update the time-varying energy flow parameters and record the parameter source version for traceability.

[0059] Furthermore, a set of schedulable parameters is configured for each node in the node set, and the parameter value range corresponding to the schedulable parameter set is determined based on a unified time-series dataset. In this embodiment, the configuration of the schedulable parameter set is related to the object type. For energy supply object nodes, such as boilers, steam heat exchange stations, refrigeration units, and air compressor units, the schedulable parameter set may include start-stop state parameters, output upper limit parameters, ramp constraint parameters, and minimum duration parameters; for energy consumption equipment object nodes, such as liquid mixing tanks, freeze dryers, filling lines, and cleanroom air conditioning units, the schedulable parameter set may include start-stop state parameters, load level parameters, and cycle time correlation parameters; for energy storage object nodes, the schedulable parameter set may include charge / discharge power parameters, storage capacity parameters, and charge / discharge conversion constraint parameters. The parameter value range can be determined comprehensively from multi-source data and maintenance information.

[0060] For example, if a device's shutdown status and alarm event structured vector indicate that the fault has not been eliminated within the maintenance window, the feasible range of the device's corresponding start / stop status parameters within the scheduling window can be narrowed. If the fan frequency and chilled water valve position are within a certain range in the cleanroom air conditioning unit's historical window, the initial range of the load level parameters can be determined from the higher-order range of that range and expanded to a wider candidate range based on process requirements to support scheduling solution exploration. To improve traceability, this embodiment can bind and store the parameter value range with its source window identifier and backfill version tag.

[0061] Subsequently, based on the time-varying energy flow parameters of the edges and the set of schedulable parameters of the nodes, energy balance constraints are constructed within a preset scheduling window to limit the constraint relationship between the supply, consumption, and storage changes within the same scheduling window. In this embodiment, energy balance constraints can be established at different levels.

[0062] On the one hand, a media conservation constraint is established in the energy medium layer. For example, in the steam medium channel, the cumulative supply of steam supply side and the cumulative consumption of each steam consumption side in the same scheduling window satisfy the constraint relationship, and a metering error relaxation amount can be introduced to cover metering inconsistencies; in the chilled water channel, the cold energy supply and cold energy consumption satisfy the constraint relationship, and the storage change amount of the storage object can be introduced to characterize the energy access of the buffer tank in the window.

[0063] On the other hand, process energy consumption allocation constraints are established at the process level. For example, within the freeze-drying process window, the power consumption and steam consumption of the freeze dryer and its auxiliary equipment can be mapped to process energy consumption items and associated with the process cycle time or batch status, so that the scheduling solution can simultaneously meet the process capacity boundary when adjusting equipment load. To cope with missing or delayed data replenishment, this embodiment can introduce a missing data compensation amount into the energy balance constraint and associate its upper limit with the replenishment version flag to avoid constraint failure or excessive shrinkage of the feasible region when data quality is insufficient.

[0064] Finally, the node set, edge set, medium type label, time-varying energy flow parameters, schedulable parameter set, and energy balance constraints are encapsulated into an energy unit graph model for subsequent energy consumption prediction model training and collaborative optimization scheduling result solving. In this embodiment, the encapsulation result includes, but is not limited to: graph structure description, parameter dictionary for each node and edge, time-varying parameter index for each window, and energy balance constraint set, and a model version tag is written to this encapsulation. The model version tag can increment when object association entries change, medium channel topology changes, or parameter value ranges are updated, so that the energy consumption prediction model training and scheduling solution can explicitly refer to a consistent version of the energy unit graph model when called.

[0065] The above process is illustrated below using a pharmaceutical factory scenario as an example. Taking an injection workshop as an example, the energy supply side includes a refrigeration room and a steam heat exchange station, while the energy consumption side includes freeze-drying equipment and a cleanroom air conditioning system. The system generates a set of nodes based on global object identifiers, including refrigeration room energy supply object nodes, steam heat exchange station energy supply object nodes, freeze-drying equipment object nodes, vacuum pump group equipment object nodes, air conditioning unit equipment object nodes, freeze-drying process object nodes, and chilled water buffer tank energy storage object nodes.

[0066] The system further generates an edge set, including supply edges from the chiller room to the chilled water channel, consumption edges from the chilled water channel to the air conditioning unit, supply edges from the steam heat exchange station to the steam channel, consumption edges from the steam channel to the freeze-drying process, consumption edges from the dielectric channel to each piece of equipment, and buffer tank charging / discharging edges. The chilled water channel edges are configured with the medium type labeled as chilled water and bound to a chilled water flow meter metering point object identifier; the steam channel edges are configured with the medium type labeled as steam and bound to a steam flow meter metering point object identifier; the dielectric edges are configured with the medium type labeled as electricity and bound to the electricity meter metering point object identifiers of each piece of equipment. The system reads the flow and power parameters of the corresponding window from a unified time-series dataset to form the time-varying energy flow parameters for each edge.

[0067] Subsequently, start / stop status parameters, output limit parameters, and ramp constraint parameters are configured for the chiller room nodes; load level parameters and start / stop status parameters are configured for the air conditioning unit nodes; start / stop status parameters and load level parameters are configured for the freeze dryer nodes; and charging / discharging power parameters and storage capacity parameters are configured for the buffer tank nodes. The value ranges for each parameter are determined based on historical window statistics and maintenance status from a unified time-series dataset. Finally, energy balance constraints are constructed within each preset scheduling window, ensuring that changes in chilled water supply, consumption, and storage satisfy constraints within the window; steam supply and consumption satisfy constraints; and dielectric consumption and the load settings of each device satisfy correlation constraints. This information is then encapsulated into an energy unit graph model for subsequent module calls.

[0068] The implementation method of this embodiment can form an energy unit diagram model consistent with the production, supply and consumption structure of pharmaceutical factories on the basis of a unified time series dataset. It can also encapsulate multi-media energy flow parameters, equipment schedulable parameters and window-level energy balance constraints in a unified manner, so that the subsequent energy consumption prediction model training and rolling collaborative scheduling solution can obtain consistent structured input and constrained expression, improve the consistency of energy relationship modeling across processes, equipment and media and parameter reference stability, and provide a traceable model foundation for the subsequent collaborative optimization process.

[0069] In some embodiments, an energy consumption prediction model is trained based on an energy unit graph model and a unified time-series dataset, and multi-step energy consumption prediction results are output, including: The structural prior features are generated based on the energy unit graph model. The structural prior features include node adjacency features, medium type labeling features, and schedulable parameter features of nodes and edges. Based on the unified time series dataset, sample window features are constructed for each node and edge of the energy unit graph model; The structural prior features and sample window features are input into the energy consumption prediction model to generate multi-step prediction output. The multi-step prediction output includes the energy consumption prediction value corresponding to each prediction step and the uncertainty parameter associated with the energy consumption prediction value. When receiving an update of the unified time series dataset carrying a backdated version tag, the affected sample window range is determined based on the backdated version tag, and a local incremental update is performed on the energy consumption prediction model to generate an updated energy consumption prediction model. The updated energy consumption prediction model is invoked, and the multi-step energy consumption prediction results corresponding to the preset prediction step size set are output based on the input features of the current scheduling window.

[0070] Specifically, structural prior features are a set of structured features directly derived from the energy unit graph model. They are used to provide the energy consumption prediction model with information on the topological relationships and media attributes between production, supply, and consumption objects. Structural prior features include, but are not limited to: node adjacency relationship features, media type labeling features, and schedulable parameter features of nodes and edges.

[0071] Node adjacency features represent the connection relationships and directions between nodes, indicating the strength and path structure between energy-supplying objects and energy-consuming equipment, between equipment and processes, and between energy storage objects and media channels. Media type labeling features distinguish the types of media carried by different edges, enabling the model to differentiate the dimensional differences and temporal characteristics of media such as electricity, steam, and chilled water. Sample window features are a set of input features constructed within a preset time window based on a unified time-series dataset, used to characterize dynamic information such as energy consumption, load, power supply status, metering data, and alarm status of nodes and edges within that window.

[0072] In some examples, multi-step prediction output is a set of prediction results generated for a preset prediction step size set. This preset step size set can include step size indices such as 1, 2, 3, 6, and 12, or corresponding future minutes, to cover short-term and medium-term predictions. Uncertainty parameters are prediction confidence description parameters corresponding to each prediction step size. They can be used to characterize the dispersion of prediction results or the width of the confidence interval, facilitating subsequent scheduling and solving for stricter constraints or redundant reserves for high-uncertainty predictions. Local incremental updates refer to an update method where, when a backtest version update occurs on a unified time-series dataset, only the data within the affected sample window is retrained or parameters are updated, without retraining all historical samples. This reduces computational overhead and shortens the model update chain. The affected sample window range is the set of window identifiers located based on the backtest version marker, along with the set of preceding context windows involved in sample construction.

[0073] In its implementation, the system first generates structural prior features based on the energy unit graph model. In this embodiment, the system reads the node and edge sets from the energy unit graph model to generate node adjacency features that describe the incoming and outgoing edges, adjacent node categories, and connection directions of each node. Simultaneously, it reads the medium type markers of the edges to generate medium type marker features, enabling the model to distinguish different medium channels and learn the unique dynamic responses of the medium. Furthermore, it reads the schedulable parameter sets of nodes and edges and their value ranges to generate schedulable parameter features for nodes and edges, which express the structural information of controllable schedulable variables such as equipment start-up / shutdown, load level, ramp constraints, and minimum duration. To avoid directly introducing overly detailed equipment model information, the structural prior features in this embodiment can be expressed using a higher-level encoding method, such as using node category, medium category, schedulable parameter category, and constraint category as encoding dimensions, thereby balancing model generalization and protection scope.

[0074] Furthermore, sample window features are constructed for each node and edge of the energy unit graph model based on a unified time-series dataset. In this embodiment, the sample window features are constructed based on aligned window data. For equipment object nodes, the sample window features may include power statistics within the window, start / stop status change markers, load level statistics, and structured vector summaries of alarm events associated with the equipment; for process object nodes, the sample window features may include process status markers, batch stage markers, cycle load statistics, and process aggregated energy consumption statistics; for energy supply object nodes, the sample window features may include energy supply output, energy supply side pressure and temperature statistics, and start / stop status; for edges, the sample window features may include medium flow parameters, power parameters, cumulative parameters, and metering anomaly verification markers.

[0075] In some examples, to form trainable samples, this embodiment concatenates the features of multiple consecutive window samples in chronological order to form an input sequence, and uses the target energy consumption corresponding to a preset prediction step size as a supervision label. In some implementations, the target energy consumption can be selected as the future power consumption of a device node, the future comprehensive energy consumption of a process node, or the future supply of a medium edge, and different output heads can be trained according to different prediction tasks to adapt to multi-target prediction.

[0076] Furthermore, the structural prior features and sample window features are input into the energy consumption prediction model to generate multi-step prediction outputs. In this embodiment, the energy consumption prediction model adopts a modeling approach that integrates "graph structure information and time series information," enabling the model to utilize both the topological structure of the energy unit graph model and learn the temporal dynamics in the unified time series dataset. Specifically, the structural prior features are used to determine the information propagation paths and medium channel differences between nodes, while the sample window features are used to provide dynamic observations within each window.

[0077] In some examples, the model outputs multi-step predictions, which include the predicted energy consumption value for each prediction step and the uncertainty parameter associated with that value. The uncertainty parameter can be generated synchronously by the model at the output and corresponds one-to-one with each prediction step, characterizing the reliability of the prediction results. In subsequent collaborative scheduling constraint construction, the uncertainty parameter can be used as input to set more conservative feasible region boundaries or retention parameters for predictions with high uncertainty. However, in this embodiment, it is not described as an effect, but only used to illustrate its structured output that can be referenced in subsequent steps.

[0078] Furthermore, upon receiving an update to the unified time-series dataset carrying a backdated version marker, the affected sample window range is determined based on the backdated version marker, and a local incremental update is performed on the energy consumption prediction model to generate an updated energy consumption prediction model. In this embodiment, when late data triggers backdated merging and generates a new backdated version marker, the data processing link outputs an update notification of "window identifier - backdated version marker - set of changed object identifiers". The model training module reads this update notification, determines which window identifiers have changed data versions, and further determines the input sequence window segment and label window segment corresponding to these window identifiers in the sample construction, thereby forming the affected sample window range.

[0079] In some examples, local incremental updates can be implemented as follows: reconstruct the sample window features and supervision labels for the affected sample window range, keep the structural prior features unchanged, perform a limited number of rounds of parameter updates on the model or perform updates on a subset of the model's parameters, and write the updated model to the model version repository. At the same time, record the model update version tag associated with the backfill version tag so that the data version on which the model is based can be traced later.

[0080] Furthermore, the updated energy consumption prediction model is invoked, and the multi-step energy consumption prediction results corresponding to the preset prediction step size set are output based on the input features of the current scheduling window. In this embodiment, at the beginning of each rolling scheduling window, the system reads the aligned window data of the current window and several preceding windows from a unified time-series dataset to construct sample window features consistent with the model input; at the same time, it reads the structural prior features from the energy unit graph model.

[0081] In some implementations, the two types of features are input into the updated energy consumption prediction model to obtain multi-step energy consumption prediction results covering a preset prediction step size set. For example, energy consumption prediction values ​​and uncertainty parameters are output for one, two, three, and six future windows. The multi-step energy consumption prediction results are then used as prediction inputs for the cooperative optimization scheduling solution, and together with energy balance constraints and cooperative scheduling constraints, they constitute the solution conditions within the rolling window.

[0082] The above process is illustrated below using a pharmaceutical factory scenario as an example. Taking the freeze-drying process in an injection workshop as an example, the energy unit graph model includes nodes for energy supply to the refrigeration room, the steam heat exchange station, the freeze dryer, and the cleanroom air conditioning unit, as well as media edges such as steam, chilled water, and electricity. In the structural prior features, the node adjacency relationship feature indicates that the steam heat exchange station supplies energy to the freeze-drying process through the steam media edge, the refrigeration room supplies energy to the air conditioning unit through the chilled water media edge, and the electricity media edge supplies energy to each piece of equipment; the media type label feature distinguishes between steam, chilled water, and electricity; the schedulable parameter feature indicates that the refrigeration room and air conditioning unit have start-stop and load parameters, and the freeze dryer has phased load parameters.

[0083] The sample window features are provided by a unified time-series dataset, such as the freeze dryer power statistics, cumulative steam flow, chilled water flow and fan frequency statistics of the air conditioning unit, and alarm summary vectors within each 300-second window during the freezing phase. After training, the model outputs prediction results for multiple future windows, such as the comprehensive energy consumption prediction values ​​and corresponding uncertainty parameters for the freeze-drying process in windows 1, 2, 3, and 6. If a delayed report from the steam flow meter causes the cumulative steam volume of a certain historical window to be replenished and a new replenished version marker to be generated, the system locates the affected sample window range and performs a local incremental update, allowing the model parameters to absorb the replenished data without full retraining. Subsequently, at the start of the next rolling scheduling window, the updated model is called to output new multi-step energy consumption prediction results for scheduling solutions.

[0084] The implementation method of this embodiment can combine the structural information provided by the energy unit graph model with the dynamic observation information provided by the unified time series dataset for training the energy consumption prediction model, outputting energy consumption prediction values ​​and their uncertainty parameters covering multiple prediction steps, and realizing local incremental updates of affected samples based on the replenishment version mark when the data version changes due to late replenishment, thereby improving the consistency and traceability of prediction input and data version, and supporting the collaborative optimization solution within the rolling scheduling window to obtain stable multi-step prediction basis.

[0085] In some embodiments, hierarchical classification processing is performed on alarm event objects to generate alarm event structured vectors carrying priority and associated object identifier sets, and the alarm event structured vectors are converted into cooperative scheduling constraints, including: Based on the event type information, trigger condition information and confirmation status information of the alarm event object, generate an alarm tag set corresponding to the alarm event object. The alarm tag set includes one or more of the severity level tag, impact domain tag and handling status tag. The priority of alarm event objects is determined based on the alarm tag set, and the equipment object identifier, process object identifier, or power supply object identifier associated with the alarm event object is extracted based on the object association entries to generate an associated object identifier set. The priority, associated object identifier set, and alarm tag set are encapsulated into an alarm event structured vector, and an effective window flag is written into the alarm event structured vector. The effective window flag is used to indicate the scheduling window range to which the alarm event structured vector is applicable. Based on the influence domain label and priority, the structured vector of alarm events is converted into a collaborative scheduling constraint. The collaborative scheduling constraint includes one or more of the following: boundary constraints on the set of schedulable parameters corresponding to the set of associated object identifiers, switching constraints on start-stop switching within the scheduling window, and relaxed upper limit constraints on energy balance constraints.

[0086] Specifically, priority is a sortable parameter obtained by mapping the alarm tag set, used to characterize the constraint strength or constraint effectiveness order of alarm events in the scheduling constraint generation process.

[0087] The associated object identifier set is a set of object identifiers extracted from the object association entries that are associated with the alarm event object. It includes one or more of the following: equipment object identifier, process object identifier, or power supply object identifier, and is used to limit the target of the alarm constraint.

[0088] An alarm event structured vector is a structured data carrier obtained by encapsulating an alarm event object. It includes, but is not limited to, priority, a set of associated object identifiers and a set of alarm tags, and may also include a trigger time summary and an acknowledgment status summary.

[0089] Cooperative scheduling constraints are constraint descriptions obtained by transforming alarm events into structured vectors. They are used to limit the feasible domain of the schedulable parameter set or energy balance constraints when solving cooperative optimization scheduling problems. Cooperative scheduling constraints include one or more of the following: boundary constraints, switching constraints, and relaxation upper limit constraints.

[0090] In practical implementation, firstly, based on the event type information, trigger condition information, and confirmation status information of the alarm event object, a set of alarm tags corresponding to the alarm event object is generated. In this embodiment, the alarm event object is output by the alarm system, control system, or equipment diagnostic module, and carries at least event type information, trigger condition information, and confirmation status information. The event type information is used to indicate the category to which the alarm belongs, such as equipment failure, process deviation, energy supply abnormality, metering abnormality, or communication abnormality; the trigger condition information is used to indicate the threshold conditions or logical conditions for alarm triggering, such as temperature exceeding the limit, pressure falling below the limit, vibration exceeding the limit, abnormal difference between supply and return water temperatures, steam pressure fluctuation, or metering point offline; the confirmation status information is used to indicate whether the alarm has been manually confirmed or confirmed by the system.

[0091] Furthermore, the system generates an alarm tag set based on preset tag mapping rules. Severity level tags can be mapped by a combination of event type and triggering conditions; for example, alarms involving aseptic environments and deviations in critical process parameters are mapped to high severity, while short-term packet loss at metering points is mapped to medium-low severity. Impact domain tags can be derived from event type and associated object entries; for example, the impact domain of a chiller room power supply anomaly includes chilled water media channels and cleanroom air conditioning units that rely on chilled water, while the impact domain of a freeze dryer vacuum system anomaly includes the freeze dryer equipment and the freeze drying process objects. Disposal status tags are directly mapped from confirmation status information, such as unconfirmed, confirmed, isolated, and released.

[0092] Furthermore, the priority of alarm event objects is determined based on the alarm tag set, and the equipment object identifier, process object identifier, or power supply object identifier associated with the alarm event object is extracted based on the object association entries to generate an associated object identifier set. In this embodiment, the priority generation rule considers at least the severity level tag and the handling status tag. Generally, the higher the severity level of the alarm and the handling status is not cleared, the higher its priority; for alarms that have been cleared or isolated, the priority can be reduced or they can be marked as being used only for traceability and not for constraint generation.

[0093] Subsequently, the system uses the global object identifier of the alarm event object itself as an index to read the object association entries, extract the equipment object identifier, process object identifier, and power supply object identifier associated with the alarm, and form a set of associated object identifiers. In some implementations, if the alarm impact domain label indicates the impact of the media channel, the set of associated object identifiers can be further expanded to include the power supply object identifier upstream of the media channel and the energy-consuming equipment object identifier downstream, to ensure that the constraint scope covers the actual affected objects.

[0094] Furthermore, the priority, the set of associated object identifiers, and the set of alarm tags are encapsulated into an alarm event structured vector, and an effective window marker is written into the alarm event structured vector. In this embodiment, the alarm event structured vector serves as an intermediate carrier for generating scheduling constraints. Its content includes, but is not limited to: priority, set of associated object identifiers, severity level label, impact domain label, and handling status label, and may also include a summary of trigger fact time parameters and entry time parameters for alignment with a unified time series dataset.

[0095] The generation of the effective window marker is based on the alarm trigger time and handling status: when the alarm is in an unresolved state, the starting window of the effective window marker can be set to the identifier of the window where the alarm was triggered, and the ending window can be set to the preset number of continuation windows after the current rolling scheduling window; when the alarm is in a resolved state, the ending window of the effective window marker can be set to the identifier of the window where the alarm was resolved, so that the alarm constraint only participates in the solution during its effective period. To deal with late-reported alarm events, this embodiment can use the database entry time parameter and the version update marker mechanism to write the late alarm into the corresponding historical window and generate a new effective window marker, while recording version changes for traceability.

[0096] Finally, the structured vector of alarm events is converted into collaborative scheduling constraints based on the influence domain label and priority. In this embodiment, the collaborative scheduling constraints include one or more of the following three categories. Boundary constraints are used to impose upper and lower bounds or fixed value restrictions on the set of schedulable parameters corresponding to the set of associated object identifiers. For example, when a high-priority fault alarm occurs in the chiller room and the influence domain label points to the chilled water channel, a lower upper bound can be imposed on the upper limit parameter of the chiller room's output, and an upper bound restriction can be imposed on the load level parameter of the clean air conditioning unit that depends on chilled water; when a motor overload alarm occurs in a critical piece of equipment, an upper bound restriction can be imposed on the load level parameter of the equipment or the start / stop status parameter can be limited to the shutdown state.

[0097] In some examples, switching constraints are used to limit start-stop switching behavior within a scheduling window, preventing frequent start-stops or disallowed switching while an alarm is still active. For instance, when an air compressor unit experiences a lubrication system alarm and the status is not yet cleared, a switching constraint prohibiting switching from stop to start can be applied to the air compressor unit's start-stop status parameters; when a valve jamming alarm occurs at a steam heat exchange station, a minimum duration constraint can be set for the start-stop switching of the relevant equipment to avoid short-cycle switching.

[0098] The relaxation upper limit constraint is used to limit the upper limit of the relaxation amount when introducing metering error relaxation or missing compensation amount into the energy balance constraint, and adjusts the value range of the relaxation upper limit according to the alarm priority. For example, when a metering point object triggers a metering anomaly alarm and the influence domain label points to a certain medium channel, the system can set the relaxation upper limit of the energy balance constraint of that channel to a stricter range, or mark the time-varying energy flow parameters of the relevant edges of that channel as low-confidence inputs, thereby enabling the scheduling solution to make computable boundary limits on uncertain data at the constraint construction level. The generated collaborative scheduling constraint items are bound to the effective window label and written to the constraint index according to the set of associated object identifiers, so that the effective constraints can be quickly loaded according to the current window identifier when solving the rolling scheduling window.

[0099] The following example, using a pharmaceutical factory scenario, illustrates the above process. Taking the cleanroom air conditioning system in an injection preparation workshop as an example, at a certain moment, the chiller room outputs a "chiller unit high-pressure protection alarm." The event type information indicates a power supply equipment failure, the trigger condition information indicates that the condensing pressure exceeds the threshold, and the confirmation status information indicates that it has been confirmed but not yet deactivated. The system generates a set of alarm tags, where the severity level tag is "high," the impact domain tag points to the chilled water channel and the cleanroom air conditioning unit, and the handling status tag is "not deactivated." Based on the tag mapping rules, the priority is determined to be high, and a set of associated object identifiers is extracted from the object association entries, including the chiller room power supply object identifier, the chilled water medium object identifier, and the air conditioning unit equipment object identifier that depends on chilled water.

[0100] Furthermore, the system encapsulates the above information into a structured vector of alarm events and writes it into the effective window marker, so that it covers several subsequent rolling scheduling windows starting from the alarm trigger window. Then, based on the influence domain label and priority, it generates collaborative scheduling constraints: boundary constraints are generated for the chiller room output limit parameter, boundary constraints are generated for the air conditioning unit load level parameter, and switching constraints are generated for the air conditioning unit start-stop switching, to limit the operating strategy during the period when the alarm is not cleared. This collaborative scheduling constraint takes effect together with the energy balance constraint during the rolling scheduling solution, thus ensuring that the schedulable parameter trajectory of the scheduling result remains consistent with the alarm state in a computational sense.

[0101] The implementation method of this embodiment can convert alarm events from event text or discrete alarm forms into a computable set of alarm tags and a structured vector of alarm events. Based on object association entries, the alarm influence domain is mapped to the set of associated object identifiers. Furthermore, under the effective window marking constraint, collaborative scheduling constraint items such as boundary constraints, switching constraints, and relaxation upper limit constraints are generated. This achieves structured coupling between alarm classification results and the rolling collaborative scheduling solution process, improves the availability and consistency of alarm information on the scheduling input side, and supports subsequent device-level execution instruction generation processes to be updated under constraints.

[0102] In some embodiments, within a rolling scheduling window, the collaborative optimization scheduling result is solved based on multi-step energy consumption prediction results, energy balance constraints, and collaborative scheduling constraints, including: At the beginning of each rolling scheduling window, the collaborative scheduling problem is decomposed into multiple local sub-problems corresponding to equipment object nodes, process object nodes and energy supply object nodes based on the energy unit graph model, and a set of consistency variables is assigned to each local sub-problem. For each local subproblem, the candidate parameter trajectory of the schedulable parameter set within the rolling scheduling window is determined based on the multi-step energy consumption prediction results, and the local scheduling candidate solutions are solved under the constraints of energy balance constraints and cooperative scheduling constraints. Based on the set of consistency variables, a consistency coordination update is performed on each local scheduling candidate solution. The consistency coordination update includes iteratively updating the multiplier parameters of the set of consistency variables and performing projection correction on the local scheduling candidate solutions. When the preset convergence criterion is met, the candidate solutions of each local scheduling are merged to generate a collaborative optimization scheduling result. The collaborative optimization scheduling result includes the schedulable parameter trajectory within the rolling scheduling window corresponding to each global object identifier.

[0103] Specifically, the consistency variable set is a set of variables used to pass shared boundary conditions among multiple local solution subproblems, expressing window-level energy boundary parameters or shared resource capacity parameters shared across objects. A local solution subproblem is a subproblem constructed around a device object node, process object node, or energy supply object node. Its decision variables are the set of schedulable parameters corresponding to the central object, and the constraints include local terms of energy balance constraints and coordinated scheduling constraints associated with that object. The multiplier parameter is an update parameter used to measure the degree of deviation of the local solution subproblem from the consistency variables, driving each local solution to converge towards the consistency constraints. Projection correction is a process of mapping the feasible region of local scheduling candidate solutions, used to correct the local solutions to the feasible set under the constraints of boundary constraints, switching constraints, and parameter value ranges. The preset convergence criterion is a set of conditions used to determine whether the consistency coordination update has ended, including one or more of the following: consistency variable residual threshold, objective function change threshold, and maximum iteration number threshold.

[0104] In its implementation, at the start of each rolling scheduling window, the system decomposes the collaborative scheduling problem based on the energy unit graph model. Specifically, it determines the boundaries of sub-problems for energy supply objects, energy-consuming equipment, and process objects based on the set of nodes and edges. It also abstracts the window-level supply boundary, medium channel capacity boundary, or energy storage object charging / discharging boundary on the medium channel edge into a set of consistency variables. For a pharmaceutical factory's injection workshop, the energy supply side includes refrigeration rooms and steam heat exchange stations, the energy consumption side includes freeze dryers, cleanroom air conditioning units, and liquid preparation systems, and the process side includes freeze drying and cleanroom protection processes. The set of consistency variables can include variables such as the upper limit of chilled water window supply, the upper limit of steam window supply, and the allowable relaxation amount of critical medium channels, ensuring that each sub-problem shares consistent energy boundary conditions within the same window.

[0105] Furthermore, for each local subproblem, the candidate parameter trajectory of the schedulable parameter set within the rolling scheduling window is determined based on the multi-step energy consumption prediction results, and the local scheduling candidate solutions are solved under the constraints of energy balance constraints and cooperative scheduling constraints. Specifically, the system reads the predicted values ​​corresponding to the current window and several future windows from the multi-step energy consumption prediction results, and generates the initial range of the candidate parameter trajectory by combining it with the uncertainty parameter.

[0106] For example, for the chiller house problem, candidate trajectories for chiller output and start / stop are generated based on future cooling demand forecasts and uncertainty parameters over multiple steps. For the cleanroom air conditioning unit sub-problem, candidate trajectories for fan frequency and valve position are generated based on future chilled water supply forecasts and the load range corresponding to the cleanroom cleanliness level constraints. For the freeze dryer sub-problem, candidate trajectories for equipment load are generated based on process stage labels and future power consumption forecasts. If the collaborative scheduling constraint indicates that there is a high-priority alarm in the chiller house, the output upper limit range is narrowed or the number of start / stop switching is limited during the candidate trajectory generation stage, so that local solution candidates are generated within the constraints.

[0107] Furthermore, after obtaining each local scheduling candidate solution, a consistency coordination update is performed based on the consistency variable set. Specifically, the system calculates the consistency residual of each local candidate solution on the shared boundary variables, such as the deviation between the window supply output of the chilled water channel supply subproblem and the window consumption output of the air conditioning unit consumption subproblem on the consistency variables; then, the multiplier parameters are updated based on the deviation, so that the direction with larger deviations receives stronger coordination penalties in the next round of local solution. Subsequently, a projection correction is performed on the local scheduling candidate solutions, mapping the candidate solutions to a feasible region that includes parameter value ranges, start-stop switching constraints, and boundary constraints. For example, when a round of coordination causes the air conditioning unit load candidate trajectory to exceed the upper limit of the boundary constraint, it is limited to the allowable range through projection correction, while maintaining consistency with the minimum duration constraint; when a round of coordination causes the steam supply candidate trajectory to touch the relaxation upper limit constraint, the relaxation amount is limited to within the upper limit and the relaxation occupancy summary of the window is recorded.

[0108] Furthermore, when a preset convergence criterion is met, the various local scheduling candidate solutions are merged to generate a collaborative optimization scheduling result. For example, the system uses a set of consistent variables as a hub to align the schedulable parameter trajectories of energy supply objects, energy-consuming equipment objects, and process objects along the window dimension, and outputs the trajectory set according to the global object identifier to form a collaborative optimization scheduling result. For the pharmaceutical factory example, the collaborative optimization scheduling result may include the output trajectory and start / stop trajectory of the chiller room in the current window and several future windows, the load trajectory and switching trajectory of the clean air conditioning unit, the load trajectory of the freeze dryer at different stages, and the supply trajectory of the steam heat exchange station, and may carry window activation flags related to alarm constraints for reference when generating subsequent instructions.

[0109] The implementation method of this embodiment can integrate multi-step energy consumption prediction results, energy balance constraints, and collaborative scheduling constraints into the solution process within a rolling scheduling window. Through problem decomposition and consistency coordination, collaborative parameter trajectories across the energy supply side, energy consumption side, and process side are generated. Under the constraints of convergence criteria, traceable collaborative optimization scheduling results are output, thereby improving the scalability and consistency of scheduling solutions in multi-object and multi-media scenarios, and enhancing the adaptability of scheduling results to alarm constraints and energy boundary constraints.

[0110] In some embodiments, the collaborative optimization scheduling result is converted into execution instructions for device objects to update the device operating status or power supply operating parameters, including: Based on the collaborative optimization scheduling results, the schedulable parameter trajectories corresponding to the object identifiers of each device are extracted, and the schedulable parameter trajectories are converted into a set of instruction parameters according to the preset instruction mapping rules; Generate an instruction encapsulation body for each device object. The instruction encapsulation body includes a device object identifier, a target scheduling window identifier, an instruction parameter set, an instruction version flag, and a constraint reference flag associated with the instruction version flag. The constraint reference flag is used to indicate the energy balance constraints and cooperative scheduling constraints used when generating the instruction parameter set. Before the instruction package is sent to the execution side, a safe and feasible domain check is performed on the instruction parameter set. The safe and feasible domain check generates a feasible domain determination result based on the status feedback parameters of the device object, the parameter value range of the schedulable parameter set, and the boundary constraints in the collaborative scheduling constraint item. When the feasible region determination result indicates that it is infeasible, a feasible modification is performed on the instruction parameter set to generate a modified instruction parameter set. The feasible modification includes limiting or backing up the instruction parameter set based on the projection operator. Write the instruction package or the instruction package carrying the set of modified instruction parameters into the instruction queue and trigger execution to update the device operating status or power supply operating parameters.

[0111] Specifically, the instruction mapping rules are a set of rules used to convert schedulable parameter trajectories into a set of instruction parameters that can be recognized by the execution side. These rules include, but are not limited to, parameter name mapping, dimension conversion, discretization rules, and timing sampling rules. The instruction parameter set is a set of parameters generated according to the instruction mapping rules. It describes the control setpoints or set ranges of the device object within the target scheduling window. The instruction parameter set includes one or more of the following: start / stop control parameters, load setting parameters, valve position setting parameters, and frequency setting parameters.

[0112] The instruction encapsulation body is an encapsulated data structure used to carry the metadata and instruction parameter set required for instruction issuance. It includes, but is not limited to: device object identifier, target scheduling window identifier, instruction parameter set, instruction version flag, and constraint reference flag.

[0113] The safe and feasible domain verification is the process of determining the executability of the instruction parameter set. It is used to form a feasible domain determination result before issuance, combining equipment status feedback, parameter value range, and boundary constraints. Feasibility correction is the process of correcting the instruction parameter set when it is determined to be infeasible. Feasibility correction includes two types of operations: limiting and rollback. Limiting is used to project parameter values ​​into the allowable range, and rollback is used to revert the parameter trajectory to the previous version of the instruction or the stable parameters of the previous window.

[0114] In practical implementation, the system first extracts schedulable parameter trajectories corresponding to each device object identifier based on the collaborative optimization scheduling results, and then converts these schedulable parameter trajectories into a set of instruction parameters according to a preset instruction mapping rule. In this embodiment, the collaborative optimization scheduling results include schedulable parameter trajectories within a window corresponding to a global object identifier. The system then filters the trajectory entries corresponding to the device object identifiers and selects a mapping rule based on the device type.

[0115] For example, for chiller units in a chiller room, the schedulable parameter trajectory includes start / stop status trajectory and output setting trajectory, mapped to start / stop command parameters and target output command parameters; for cleanroom air conditioning units, the schedulable parameter trajectory includes fan frequency trajectory and chilled water valve position trajectory, mapped to fan frequency setting parameters and valve position setting parameters; for steam heat exchange stations, the schedulable parameter trajectory includes valve opening trajectory or steam supply pressure setting trajectory, mapped to valve opening setting parameters or pressure setting parameters. The command mapping rules simultaneously perform dimension conversion and discretization. For example, the continuous load setpoint is discretized according to the equipment control resolution, the window-level trajectory is extracted according to the execution side sampling period to generate a time-series setpoint sequence, and a target scheduling window identifier is bound to each setpoint.

[0116] Furthermore, an instruction encapsulation body is generated for each device object, and a constraint reference tag is written to it. In this embodiment, the system generates one or more instruction encapsulation bodies for each device object. The instruction encapsulation body includes a device object identifier, a target scheduling window identifier, and a set of instruction parameters. The instruction version tag is generated by combining the current rolling scheduling round identifier and the data version identifier, and is used to distinguish different instruction versions within the same window. The constraint reference tag is used to record the constraint source on which the instruction is based, including the energy balance constraint version, the cooperative scheduling constraint item version, and the alarm constraint effective window tag summary, so that the subsequent execution side or audit side can trace the constraint set on which the instruction was relied upon when it was generated. In some implementations, the constraint reference tag may also include a backed-up version tag summary, which is used to indicate whether the instruction is generated based on the backed-up unified time-series dataset.

[0117] Furthermore, before sending the instruction package to the execution side, a safety and feasibility domain check is performed on the instruction parameter set. In this embodiment, the safety and feasibility domain check generates a feasibility domain determination result based on at least three types of inputs. The first type is the status feedback parameters of the device object, such as the current start / stop status, current load, current valve position, temperature and pressure status, and interlock status; the second type is the parameter value range of the schedulable parameter set, such as the set of allowed start / stop states, upper and lower load limits, ramp rate upper limit, and minimum duration; the third type is the boundary constraints and switching constraints in the collaborative scheduling constraints, such as the output upper limit reduction or prohibition of start / stop switching caused by high-priority alarms. The check process can determine each parameter item by item and form a summary determination result: if any parameter setting value exceeds the value range, or the setting value change violates the ramp constraint, or the start / stop switching violates the switching constraint, it is determined to be infeasible and an infeasibility reason flag is output; if all are satisfied, it is determined to be feasible and a feasible flag is output.

[0118] Furthermore, when the feasible region determination result indicates infeasibility, a feasibility correction is performed on the instruction parameter set to generate a corrected instruction parameter set. In this embodiment, the feasibility correction is performed with the projection operator as the core. For continuous parameters that exceed the boundary, a limiting operation is performed to project them into the allowable range, and the feasible value closest to the original setting value is selected first while maintaining the consistency of window-level constraints; for start-stop parameters that violate switching constraints or minimum duration constraints, a rollback operation is performed to roll back the start-stop settings to the stable state executed in the previous window or the start-stop settings of the previous version of the instruction, and subsequent setting points are adjusted synchronously to avoid short-cycle repeated switching.

[0119] For example, in a pharmaceutical factory scenario, if a cleanroom air conditioning unit is limited to its load limit under a high-priority alarm constraint, but the scheduling result gives a higher frequency setting, then the frequency setting is projected to the upper limit through amplitude limiting; if a refrigeration unit is in a protection lockout state but the scheduling result requires it to start, then the start / stop command is rolled back to the stop state through rollback, and the rollback reason is recorded in the constraint reference mark.

[0120] Finally, the instruction encapsulation body, or the instruction encapsulation body carrying the set of modified instruction parameters, is written into the instruction queue and execution is triggered for updating the device operating status or power supply operating parameters. In this embodiment, the instruction queue is managed by partition according to the device object identifier and sorted according to the target scheduling window identifier. When triggering execution, the system sends the corresponding instruction encapsulation body to the execution side control interface or gateway interface according to the target scheduling window identifier and the execution side communication cycle. After receiving the instruction encapsulation body, the execution side can perform deduplication and replacement according to the instruction version tag, and record the instruction execution log according to the constraint reference tag for subsequent anomaly tracing and auditing. If a new cooperative scheduling constraint item or energy balance constraint version update is received during the execution, this embodiment can generate a new instruction version tag and insert the new instruction encapsulation body into the instruction queue to cover the window setpoint that has not yet been executed.

[0121] The following example, using a pharmaceutical factory scenario, illustrates the above process. Taking an injection workshop as an example, the collaborative optimization scheduling result provides the output trajectory and start / stop trajectory of the refrigeration room in the next scheduling window, as well as the fan frequency trajectory and valve position trajectory of the cleanroom air conditioning unit. The system converts the output trajectory into output setting parameters and the frequency trajectory into frequency setting parameters according to the instruction mapping rules, and generates an instruction encapsulation containing the equipment object identifier and target window identifier. If, at this time, the refrigeration room has a high-pressure protection alarm and the collaborative scheduling constraint limits the output upper limit, and the safety feasible domain check finds that the output setting exceeds the upper limit, then a limit correction is performed, and a set of correction instruction parameters is generated. Subsequently, the instruction encapsulation carrying the correction parameters is written into the instruction queue and triggered for issuance; simultaneously, the energy balance constraint version and alarm constraint version used by the instruction are recorded in the constraint reference flag, so that the basis for this limit correction can be located subsequently.

[0122] The implementation method of this embodiment can stably convert the collaborative optimization scheduling results into a set of instruction parameters and instruction encapsulation that can be identified by the device object. Before issuing the instruction, the safety and feasibility domain verification and projection correction are performed in combination with the device status feedback, parameter value range and collaborative scheduling constraints, thereby improving the executability and consistency of instruction issuance, enhancing the traceability of the source of instructions and constraints, and supporting the controllable updating of device operating status and power supply operating parameters by version within the rolling scheduling window.

[0123] The following are system embodiments of this application, which can be used to execute the method embodiments of this application. For details not disclosed in the system embodiments of this application, please refer to the method embodiments of this application.

[0124] Figure 2 This is a schematic diagram of the structure of the machine learning-based device energy consumption collaborative intelligent optimization system provided in an embodiment of this application. Figure 2 As shown, the system includes: The acquisition module 201 is used to acquire multi-source data in the production, supply and consumption processes of the energy system, establish an association mapping based on a unified object identifier as equipment object, process object, energy medium object, metering point object and alarm event object, and write fact time parameters and storage time parameters for the multi-source data; The merging module 202 is used to perform windowed time alignment on multi-source data based on the fact time parameter, and to perform backfill merging on late data based on the inbound time parameter to form a unified time series dataset; Module 203 is used to build an energy unit graph model based on a unified time series dataset and configure energy balance constraints and schedulable parameters. The energy unit graph model is used to represent the energy flow relationship between equipment objects, process objects and energy supply objects. Training module 204 is used to train an energy consumption prediction model based on the energy unit graph model and a unified time series dataset and output multi-step energy consumption prediction results; The generation module 205 is used to perform hierarchical classification processing on alarm event objects to generate alarm event structured vectors carrying priority and associated object identifier sets, and to convert alarm event structured vectors into collaborative scheduling constraint terms; The conversion module 206 is used to solve the collaborative optimization scheduling result based on the multi-step energy consumption prediction result, energy balance constraint and collaborative scheduling constraint term within the rolling scheduling window, and convert the collaborative optimization scheduling result into execution instructions for equipment objects to update the equipment operating status or energy supply operating parameters.

[0125] Figure 3 This is a schematic diagram of the electronic device 3 provided in an embodiment of this application. Figure 3 As shown, the electronic device 3 of this embodiment includes a processor 301, a memory 302, and a computer program 303 stored in the memory 302 and executable on the processor 301. When the processor 301 executes the computer program 303, it implements the steps in the various method embodiments described above. Alternatively, when the processor 301 executes the computer program 303, it implements the functions of each module / unit in the system embodiments described above.

[0126] Electronic device 3 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 3 may include, but is not limited to, processor 301 and memory 302. Those skilled in the art will understand that... Figure 3 This is merely an example of electronic device 3 and does not constitute a limitation on electronic device 3. It may include more or fewer components than shown, or different components.

[0127] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0128] The memory 302 can be an internal storage unit of the electronic device 3, such as a hard disk or memory of the electronic device 3. The memory 302 can also be an external storage device of the electronic device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the electronic device 3. The memory 302 can also include both internal and external storage units of the electronic device 3. The memory 302 is used to store computer programs and other programs and data required by the electronic device.

[0129] The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium (e.g., a computer-readable storage medium). Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by a processor, it can implement the steps of the various method embodiments described above. The computer program can include computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0130] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A machine learning-based method for collaborative intelligent optimization of device energy consumption, characterized in that, include: Acquire multi-source data from the production, supply, and consumption processes of the energy system, establish association mappings based on unified object identifiers for equipment objects, process objects, energy medium objects, metering point objects, and alarm event objects, and write fact time parameters and storage time parameters for the multi-source data; Windowed time alignment is performed on multi-source data based on the fact time parameter, and late data is backfilled and merged based on the inbound time parameter to form a unified time series dataset; An energy unit graph model is constructed based on a unified time series dataset, and energy balance constraints and schedulable parameters are configured. The energy unit graph model is used to characterize the energy flow relationship between equipment objects, process objects and energy supply objects. Based on the energy unit graph model and the unified time series dataset, an energy consumption prediction model is trained and multi-step energy consumption prediction results are output. The alarm event objects are subjected to hierarchical classification processing to generate alarm event structured vectors carrying priority and associated object identifier sets, and the alarm event structured vectors are converted into cooperative scheduling constraint terms; Within the rolling scheduling window, the collaborative optimization scheduling result is solved based on the multi-step energy consumption prediction result, energy balance constraint, and collaborative scheduling constraint term. The collaborative optimization scheduling result is then converted into execution instructions for the device object to update the device operating status or energy supply operating parameters.

2. The method according to claim 1, characterized in that, The process involves establishing association mappings based on a unified object identifier for equipment objects, process objects, energy medium objects, metering point objects, and alarm event objects, and writing factual time parameters and data entry time parameters for multi-source data, including: Global object identifiers are assigned to various types of objects, and object association entries are generated based on preset object relationship rules. The object association entries are used to indicate the association relationship between equipment objects and process objects, energy medium objects, metering point objects, and alarm event objects. When receiving multi-source data associated with each object, extract the collection time from the multi-source data and write it into the fact time parameter, and obtain the access time of the multi-source data entering the data processing link and write it into the database entry time parameter. Perform consistency checks on multi-source data to generate check tags. When the check passes, bind and encapsulate the multi-source data with the corresponding global object identifier, object association entry, fact time parameter, entry time parameter, and check tag to generate a unified data record that can be used for subsequent windowed time alignment.

3. The method according to claim 1, characterized in that, The process of performing windowed time alignment on multi-source data based on fact time parameters, and performing backfilling and merging on late data based on inbound time parameters to form a unified time-series dataset, includes: Based on the stated fact time parameter, window identifiers are assigned to multi-source data. The window identifiers are used to indicate the windowed time range to which the multi-source data belongs, and multi-source data are aggregated under the same window identifier according to the global object identifier to generate object window data. For the object window data, an alignment key is generated based on the object association entries. The object window data of the equipment object, process object, energy medium object, metering point object and alarm event object are time aligned according to the alignment key to form aligned window data. When the inbound time parameter and the deadline corresponding to the window identifier meet the preset lateness determination condition, the corresponding multi-source data is determined as late data, and a replenishment mark is generated for the late data. Perform a backfill merging process on the late data carrying the backfill flag. The backfill merging process includes writing the late data into the aligned window data corresponding to the window identifier and recording the backfill version flag. A traceable unified time-series dataset is generated based on the backtracking version tag. The unified time-series dataset includes the alignment window data and the backtracking version tag associated with the alignment window data.

4. The method according to claim 1, characterized in that, The construction of the energy unit graph model based on the unified time-series dataset, and the configuration of energy balance constraints and schedulable parameters, includes: Based on global object identifiers and object association entries, the set of nodes and the set of edges of the energy unit graph model are determined; Configure a medium type label and a metering point object identifier for each edge in the edge set, and extract the medium flow parameters or power parameters corresponding to the metering point object identifier based on the unified time series dataset to determine the time-varying energy flow parameters of each edge; Configure a set of schedulable parameters for each node in the node set, and determine the range of parameter values ​​corresponding to the set of schedulable parameters based on the unified time series dataset; Based on the time-varying energy flow parameters of the edge and the set of schedulable parameters of the node, an energy balance constraint is constructed within a preset scheduling window. The energy balance constraint is used to limit the constraint relationship between the supply, consumption and storage changes within the same scheduling window. The set of nodes, set of edges, medium type markers, time-varying energy flow parameters, set of schedulable parameters, and energy balance constraints are encapsulated into an energy unit graph model for subsequent energy consumption prediction model training and collaborative optimization scheduling result solving.

5. The method according to claim 1, characterized in that, The step of training an energy consumption prediction model based on the energy unit graph model and a unified time-series dataset and outputting multi-step energy consumption prediction results includes: Based on the energy unit graph model, structural prior features are generated, including node adjacency features, medium type labeling features, and schedulable parameter features of nodes and edges. Based on the unified time-series dataset, sample window features are constructed for each node and edge of the energy unit graph model; The structural prior features and the sample window features are input into the energy consumption prediction model to generate a multi-step prediction output, which includes the energy consumption prediction value corresponding to each prediction step and the uncertainty parameter associated with the energy consumption prediction value. When receiving an update of the unified time series dataset carrying a backdated version tag, the affected sample window range is determined based on the backdated version tag, and a local incremental update is performed on the energy consumption prediction model to generate an updated energy consumption prediction model. The updated energy consumption prediction model is invoked, and the multi-step energy consumption prediction results corresponding to the preset prediction step size set are output based on the input features of the current scheduling window.

6. The method according to claim 1, characterized in that, The step of performing hierarchical classification processing on alarm event objects to generate alarm event structured vectors carrying priority and associated object identifier sets, and converting the alarm event structured vectors into cooperative scheduling constraints, includes: Based on the event type information, trigger condition information and confirmation status information of the alarm event object, an alarm tag set corresponding to the alarm event object is generated. The alarm tag set includes one or more of the severity level tag, impact domain tag and handling status tag. The priority of the alarm event object is determined based on the alarm tag set, and the equipment object identifier, process object identifier, or power supply object identifier associated with the alarm event object is extracted based on the object association entries to generate the associated object identifier set. The priority, associated object identifier set, and alarm tag set are encapsulated into the alarm event structured vector, and an effective window flag is written for the alarm event structured vector. The effective window flag is used to indicate the scheduling window range to which the alarm event structured vector is applicable. Based on the influence domain label and the priority, the alarm event structured vector is converted into a collaborative scheduling constraint term. The collaborative scheduling constraint term includes one or more of the following: boundary constraints on the schedulable parameter set corresponding to the associated object identifier set, switching constraints on start / stop switching within the scheduling window, and relaxed upper limit constraints on energy balance constraints.

7. The method according to claim 1, characterized in that, Within the rolling scheduling window, the process of solving for the collaborative optimization scheduling result based on the multi-step energy consumption prediction results, energy balance constraints, and collaborative scheduling constraints includes: At the beginning of each rolling scheduling window, the collaborative scheduling problem is divided into multiple local solution sub-problems corresponding to equipment object nodes, process object nodes and energy supply object nodes based on the energy unit graph model, and a set of consistency variables is assigned to each local solution sub-problem. For each local subproblem, the candidate parameter trajectory of the schedulable parameter set within the rolling scheduling window is determined based on the multi-step energy consumption prediction results, and the local scheduling candidate solutions are solved under the constraints of the energy balance constraint and the cooperative scheduling constraint. Based on the set of consistency variables, a consistency coordination update is performed on each local scheduling candidate solution. The consistency coordination update includes iteratively updating the multiplier parameters of the set of consistency variables and performing projection correction on the local scheduling candidate solutions. When the preset convergence criterion is met, the candidate solutions of each local scheduling are merged to generate the collaborative optimization scheduling result, which includes the schedulable parameter trajectory within the rolling scheduling window corresponding to each global object identifier.

8. The method according to claim 7, characterized in that, The step of converting the collaborative optimization scheduling results into execution instructions for device objects to update the device operating status or power supply operating parameters includes: Based on the collaborative optimization scheduling results, the schedulable parameter trajectories corresponding to the object identifiers of each device are extracted, and the schedulable parameter trajectories are converted into instruction parameter sets according to preset instruction mapping rules; An instruction encapsulation body is generated for each device object. The instruction encapsulation body includes a device object identifier, a target scheduling window identifier, an instruction parameter set, an instruction version flag, and a constraint reference flag associated with the instruction version flag. The constraint reference flag is used to indicate the energy balance constraints and cooperative scheduling constraints used when generating the instruction parameter set. Before the instruction package is sent to the execution side, a safe and feasible domain check is performed on the instruction parameter set. The safe and feasible domain check generates a feasible domain determination result based on the status feedback parameters of the device object, the parameter value range of the schedulable parameter set, and the boundary constraints in the cooperative scheduling constraint item. When the feasible region determination result indicates that it is infeasible, a feasible correction is performed on the instruction parameter set to generate a corrected instruction parameter set. The feasible correction includes performing a limiting or backoff operation on the instruction parameter set based on a projection operator. The instruction package or the instruction package carrying the set of correction instruction parameters is written into the instruction queue and the execution is triggered to update the device operating status or power supply operating parameters.

9. A machine learning-based intelligent optimization system for collaborative energy consumption of devices, characterized in that, include: The acquisition module is used to acquire multi-source data in the production, supply and consumption processes of the energy system. It establishes an association mapping based on a unified object identifier for equipment objects, process objects, energy medium objects, metering point objects and alarm event objects, and writes fact time parameters and storage time parameters for the multi-source data. The merging module is used to perform windowed time alignment on multi-source data based on the fact time parameter, and to perform backfilling and merging on late data based on the inbound time parameter to form a unified time series dataset; The construction module is used to build an energy unit graph model based on a unified time series dataset and configure energy balance constraints and schedulable parameters. The energy unit graph model is used to represent the energy flow relationship between equipment objects, process objects and energy supply objects. The training module is used to train the energy consumption prediction model based on the energy unit graph model and the unified time series dataset and output multi-step energy consumption prediction results. The generation module is used to perform hierarchical classification processing on alarm event objects to generate alarm event structured vectors carrying priority and associated object identifier sets, and to convert the alarm event structured vectors into collaborative scheduling constraint terms; The conversion module is used to solve the collaborative optimization scheduling result based on the multi-step energy consumption prediction result, energy balance constraint and collaborative scheduling constraint term within the rolling scheduling window, and convert the collaborative optimization scheduling result into execution instructions for equipment objects to update the equipment operating status or energy supply operating parameters.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 8.

Images