A device power recovery method for discrete manufacturing one table multiple machines metering

Abstract

This invention relates to the field of industrial energy data processing technology and discloses a method for restoring equipment power in discrete manufacturing with a single meter for multi-machine metering. First, it acquires the aggregated power sequence, equipment state sequence, and equipment state power band collected by a shared meter, and calibrates the equipment state switching boundaries. Second, it constructs an anchoring window and a reference window, and builds a weak supervision signal based on cross-window aggregated power differences, equipment state change relationships, and compatibility of state power band changes. Then, it uses the weak supervision signal to train an equipment power restoration model. Finally, it inputs the target time window into the trained model to generate the estimated equipment power, and uses feasibility projection to ensure it meets the overall meter consistency and state power band constraints, thus obtaining the equipment power restoration result. This invention can restore equipment power without the need for large-scale installation of equipment-level meters, providing a data foundation for equipment energy consumption attribution, product energy consumption collection, and carbon emission accounting during the manufacturing stage.

Description

Technical Field

[0001] This invention relates to the field of industrial energy data processing technology, and in particular to a method for restoring the power of equipment for discrete manufacturing with one meter for multiple machines. Background Technology

[0002] Energy consumption and carbon emission accounting during the product manufacturing stage requires fine-grained, traceable equipment energy consumption data. Equipment energy consumption is usually obtained by integrating equipment power over time. Therefore, obtaining continuous equipment power curves is an important data foundation for attributing equipment energy consumption, collecting product energy consumption data, and calculating carbon emissions during the manufacturing stage.

[0003] In discrete manufacturing, acquiring equipment power data is often challenging. Multiple machines in the same area or on the same production line often operate simultaneously, serving different orders or product components. Existing discrete manufacturing energy monitoring systems primarily focus on operational monitoring, cost accounting, and regional energy consumption management, typically remaining at the workshop, area, or production line level for aggregated metering. While a few high-power devices may have independent meters, a large number of medium- and low-power devices still rely on shared meters. In this shared metering model, the shared meter collects the aggregated power from multiple devices operating under different conditions. The power contribution of individual devices changes over time and masks each other, making it difficult to directly determine the continuous power curve of each device from the aggregated power.

[0004] Existing methods for acquiring equipment energy consumption mainly include direct metering and indirect allocation. Direct metering typically requires the installation of electricity meters or data acquisition devices on each piece of equipment, resulting in high deployment, maintenance, and data management costs, making it difficult to implement on a large scale for a large number of low- to medium-power devices. Indirect allocation usually allocates energy consumption based on rated power, operating time, output, or empirical rules. Relying on preset allocation relationships, it is difficult to depict the dynamic changes in the power contribution of each device under conditions of concurrent operation of multiple devices, which can easily lead to deviations in equipment energy consumption attribution.

[0005] Non-intrusive load monitoring can recover the power curves of individual energy-consuming objects from aggregated power, providing a potential solution for scenarios with multiple devices under a single meter. However, existing non-intrusive load monitoring methods are mostly geared towards residential or building scenarios, typically relying on relatively clear appliance start / stop events, equipment power tags, or high-quality status tags. For discrete manufacturing scenarios, true equipment power values ​​and high-quality event tags are usually difficult to obtain; what can be utilized is often the coarse-grained status range provided by manufacturing execution system (MES) reporting records. These states have production management semantics and cannot be directly equated to precise power load tags, and may suffer from issues such as time-series offsets, unclear boundaries, and power fluctuations within the state.

[0006] Therefore, there is an urgent need to provide a method for restoring equipment power for multi-machine metering in discrete manufacturing to solve the above technical problems. Summary of the Invention

[0007] This invention provides a method for restoring equipment power through a single meter for multi-machine metering in discrete manufacturing. It utilizes aggregated power sequences, coarse-grained equipment state sequences, and equipment state power bands collected by shared meters to construct a weak supervision signal with an anchored-reference window relationship. Combined with a feasibility projection mechanism, it restores equipment power, thereby providing a data foundation for attributing equipment energy consumption and collecting product energy consumption without installing additional equipment meters.

[0008] This invention provides a method for restoring equipment power for multi-machine metering in discrete manufacturing, comprising:

[0009] S1. Obtain the aggregated power sequence collected by the shared electricity meter, and the coarse-grained state sequence of each of the multiple devices jointly measured by the shared electricity meter. Perform state boundary calibration on the coarse-grained state sequence to obtain the calibrated device state sequence, and obtain the state power band corresponding to each device in the calibrated device state.

[0010] S2. Based on the aggregated power sequence and the calibrated device state sequence, construct an anchoring window and a candidate reference window, and select a reference window from the candidate reference window according to the aggregated power difference, device state change relationship and state power band change compatibility between the anchoring window and the candidate reference window;

[0011] S3. Based on the anchor window and the reference window, construct a weak supervision signal including a time-varying weak supervision signal and a device responsibility weak supervision signal, and train the pre-built device power recovery model by combining the aggregated power sequence, the calibrated device state sequence and the state power band to obtain the target power recovery model;

[0012] S4. Obtain the target time window to be recovered, and input the aggregated power sequence, calibrated device status sequence and status power band within the target time window into the target power recovery model to generate the estimated power of each device within the target time window.

[0013] S5. Adjust the estimated power of the equipment to the feasible region defined by the consistency constraint of the total table and the state power band constraint through feasibility projection to obtain the equipment power recovery result of each equipment; wherein, the consistency constraint of the total table is determined according to the aggregated power sequence within the target time window, and the state power band constraint is determined according to the state power band within the target time window.

[0014] Furthermore, S2 specifically includes:

[0015] The aggregated power sequence and the calibrated device status sequence are time-aligned according to timestamps. The time-aligned sequence data is continuously divided into multiple time windows according to a preset window length and a preset sliding step size. Any time window is selected from the multiple time windows as the anchor window.

[0016] In other time intervals outside the anchor window, several time windows are extracted as candidate reference windows; wherein, the candidate reference windows and the anchor window have the same time length in terms of time span;

[0017] For each candidate reference window, calculate the aggregated power difference, device state change relationship, and state power band change compatibility between it and the anchor window, and obtain the corresponding total table difference strength, device state relationship weight, and compatibility parameters respectively.

[0018] Based on the difference strength of the summary table, the weight of the device status relationship, and the compatibility parameter, a comprehensive relationship quality score is calculated for each candidate reference window. Based on the comprehensive relationship quality score, several candidate reference windows are selected from the candidate reference windows to form a reference window set.

[0019] Furthermore, the comprehensive relationship quality score The calculation formula is:

[0020]

[0021] in,

[0022]

[0023]

[0024]

[0025] In the formula, The strength of the differences in the overall table; Weights for device status relationships; For compatibility parameters; is the preset power normalization function; T is the length of the time window; and These represent the aggregate power of anchor window a and candidate reference window r at time t, respectively. The function is an indicator function; N is the total number of devices that are jointly metered by a shared meter. The maximum number of devices allowed to undergo state changes; The number of devices whose status changes according to the set preferences; This is the state change indicator for device i. It takes the value of 1 when device i changes state between the anchor window and the candidate reference window, and takes the value of 0 otherwise. and These represent the center values ​​of the state power bands of device i in anchoring window a and candidate reference window r, respectively. This is a smooth minimum term.

[0026] Furthermore, S3 specifically includes:

[0027] Soft event weights are constructed by utilizing the local variation differences in the total power, and the soft event weights are fused with the amplitude of the total power variation to obtain a time-varying weak supervision signal.

[0028] The equipment explanation responsibility distribution caused by explicit state changes and the residual candidate responsibility distribution caused by power changes within the operating state are constructed respectively. Based on the degree of explanation of the aggregate power difference by the explicit state changes, the equipment explanation responsibility distribution and the residual candidate responsibility distribution are weighted and fused to obtain the weak supervision signal of equipment responsibility.

[0029] A device power recovery model is pre-constructed, comprising a time-series feature extraction module, a state-aware power decomposition module, and a feasibility projection module. The time-series feature extraction module extracts time features from the aggregated power sequence. The state-aware power decomposition module fuses the time features with the calibrated device state sequence and state power band, outputs the in-band position parameters of the device within the corresponding state power band, and generates the corresponding estimated device power.

[0030] The target power recovery model is obtained by training a pre-built device power recovery model using a joint loss function based on single-window physical consistency loss and anchor-reference window relationship loss.

[0031] Furthermore, the time-varying weak supervision signal The calculation formula is:

[0032]

[0033]

[0034] In the formula, For soft events; For time neighborhood expansion operators; For normalized time smoothing kernel; , These are the local change thresholds within the anchor window and the reference window, respectively; and These represent the aggregated power of the anchor window and the reference window at time t, respectively. For indicator functions; The mixing coefficient; This is the time normalization function; This represents the magnitude of the differences in the summary table; For smooth minimum terms;

[0035] The weak monitoring signal of equipment responsibility The calculation formula is:

[0036]

[0037]

[0038]

[0039]

[0040] In the formula, Explain the distribution of responsibility for the equipment; The residual candidate responsibility distribution; This is an indication of the state change of device i; and These represent the center values ​​of the state power bands of device i in the anchor window and the reference window, respectively; For compatibility parameters; This is the indicator parameter when the equipment is in processing mode in both windows; it is set to 1 when both windows are in processing mode. and These are the upper and lower bounds of the power for the corresponding device state, respectively; The strength of the differences in the overall table; These are weighting coefficients; This is a smooth minimum term.

[0041] Furthermore, the device's estimated power

[0042]

[0043] In the formula, and These represent the lower and upper bounds of the state power band of device i at time t, respectively. This refers to the in-band position parameters of the device within the corresponding power band.

[0044] The single-window physical consistency loss includes the total table consistency loss used to constrain the sum of the estimated power of each device at the same time to be consistent with the aggregated power, and the state power band feasibility loss used to constrain the estimated power of each device to be within the range of its corresponding state power band.

[0045] The anchor-reference window relationship loss includes a time variation distribution loss for constraining the distribution of cross-window device power variation in the time dimension to be consistent with the weak time variation supervision signal, and a device responsibility distribution loss for constraining the device responsibility distribution to be consistent with the weak device responsibility supervision signal.

[0046] Furthermore, S5 specifically includes:

[0047] For any moment within the target time window, a feasible region is constructed, which is jointly defined by the consistency constraint of the total table and the state power band constraint. Based on the device state characteristics, the projection adjustment cost of different devices at different times is dynamically set.

[0048] Within the feasible region, a weighted feasibility projection problem is constructed. The weighted feasibility projection problem is solved by minimizing the weighted deviation between the estimated power of the device and the projected power of the device, thereby obtaining the continuous device power recovery results for each device.

[0049] Furthermore, the expression for the feasible region is:

[0050]

[0051] In the formula, This represents the feasible region at time t within the target time window; This represents the power vector of each device at time t; Let N represent an N-dimensional real space, where N is the total number of devices sharing the same electricity meter; This represents the power of device i at time t; This represents the aggregated power collected by the shared meter at time t; and These represent the lower and upper bounds of the state power of device i at time t, respectively.

[0052] The projected device power of the device power recovery result Its formula is:

[0053]

[0054] In the formula, Estimated device power generated by model forward inference; The cost of adjusting the projection as set.

[0055] The present invention also provides a power recovery device for equipment with one meter for multiple machines in discrete manufacturing, based on the power recovery method for equipment with one meter for multiple machines in discrete manufacturing as described above, the device comprising:

[0056] The data acquisition module is used to acquire the aggregated power sequence collected by the shared electricity meter, and the coarse-grained state sequence of each of the multiple devices jointly measured by the shared electricity meter, to perform state boundary calibration on the coarse-grained state sequence to obtain the calibrated device state sequence, and to acquire the state power band corresponding to each device in the calibrated device state.

[0057] The window construction module is used to construct an anchor window and a candidate reference window based on the aggregated power sequence and the calibrated device state sequence, and to select a reference window from the candidate reference window according to the aggregated power difference, device state change relationship and state power band change compatibility between the anchor window and the candidate reference window;

[0058] The weak supervision signal construction module is used to construct a weak supervision signal including a time-varying weak supervision signal and a device responsibility weak supervision signal based on the anchoring window and the reference window, and to train the pre-built device power recovery model by combining the aggregated power sequence, the calibrated device state sequence and the state power band to obtain the target power recovery model.

[0059] The equipment power estimation module is used to obtain the target time window to be restored, input the aggregated power sequence, the calibrated equipment state sequence and the state power band within the target time window into the target power recovery model, and generate the equipment estimated power of each device within the target time window.

[0060] The result output module is used to adjust the estimated power of the equipment to a feasible region defined by the overall table consistency constraint and the state power band constraint through feasibility projection, so as to obtain the equipment power recovery result of each equipment; wherein, the overall table consistency constraint is determined according to the aggregated power sequence within the target time window, and the state power band constraint is determined according to the state power band within the target time window.

[0061] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0062] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0063] The beneficial effects of this invention are as follows:

[0064] This invention acquires aggregated power sequences and calibrated device state sequences, constructs anchoring and reference windows in the time dimension, and extracts cross-window aggregated power differences and state change mapping relationships to construct a weak supervision signal containing time variations and device responsibility. This achieves efficient training of the device power recovery model without the need for additional invasive sub-metering sensor hardware. Furthermore, the device estimated power generated by the model inference is introduced into the feasible region jointly defined by the physical consistency of the master table and the device state power band through weighted feasibility projection. Dynamic adjustment costs are used to correct deviations in physical constraints caused by state recording errors or model inference biases, ensuring that the final output device power not only maps the underlying data change logic in terms of time series characteristics but also follows the physical reality in terms of numerical boundaries. Ultimately, this provides highly reliable data support for accurate energy consumption aggregation of multi-device products and closed-loop optimization of production energy efficiency in discrete manufacturing scenarios. Attached Figure Description

[0065] Figure 1 This is a schematic diagram of a method flow according to an embodiment of the present invention.

[0066] Figure 2 This is a schematic diagram of the device structure according to an embodiment of the present invention.

[0067] Figure 3 This is a schematic diagram of the internal structure of a computer device according to an embodiment of the present invention.

[0068] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0069] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0070] like Figure 1 As shown, this invention provides a method for restoring equipment power through a single meter for multi-machine metering in discrete manufacturing. By utilizing the aggregated power of a shared meter, the coarse-grained status of the equipment, and the power range of the equipment status, it restores the continuous power of the equipment without relying on the true value of the equipment power. This provides a data foundation for equipment energy consumption attribution, product energy consumption aggregation, and carbon emission accounting during the manufacturing stage. Specifically, the method includes the following steps:

[0071] S1. Obtain the aggregated power sequence collected by the shared electricity meter, and the coarse-grained state sequence of each of the multiple devices jointly measured by the shared electricity meter. Perform state boundary calibration on the coarse-grained state sequence to obtain the calibrated device state sequence, and obtain the state power band corresponding to each device in the calibrated device state.

[0072] In one specific embodiment, step S1 includes the following sub-steps:

[0073] S101. Obtain the aggregated power sequence collected by the shared electricity meter, and obtain the coarse-grained state sequence of each of the multiple devices jointly measured by the shared electricity meter.

[0074] In discrete manufacturing environments, a large number of low- to medium-power devices are typically metered using shared electricity meters. First, aggregated power data is acquired via a data interface or acquisition terminal from shared meters used to monitor multiple devices in the same area or production line (e.g., one meter monitors six devices). The aggregated power sequence is a continuous time series formed by superimposing data from multiple devices under different operating conditions within the same shared meter. In a preferred embodiment, the sampling interval of this sequence can be set to 1 second, and the power measurement unit is kilowatts (kW). The acquired aggregated power sequence data is then stored in a database, providing fundamental observational data in the time series dimension for subsequent model training.

[0075] Simultaneously acquire production records, such as manufacturing execution system (MES) reports, from multiple devices that are strictly aligned with the aggregated power sequence in timestamps. Parse and convert these reports into a coarse-grained state sequence at the device level. This coarse-grained state sequence includes at least one or more of the following physical operating states: shutdown, standby, processing, and uncertain. Specifically, shutdown indicates the device is completely shut down or not running; standby indicates the device is awaiting processing or in an auxiliary system operating state; processing indicates the device is in a substantive production / processing state; and uncertain states represent time intervals where device state records are missing, state boundaries are unreliable, or the actual state cannot be determined.

[0076] S102. Perform state boundary calibration on the coarse-grained state sequence to obtain the calibrated device state sequence, and obtain the state power band corresponding to each device in the calibrated device state.

[0077] Due to issues such as time skew, unclear boundaries, and internal state fluctuations in work reporting records in discrete manufacturing scenarios, directly using coarse-grained states with production management semantics as power load labels would cause severe data alignment errors. Therefore, state boundary calibration is necessary. Specifically, time series analysis is used to identify state transition points between states in the coarse-grained state sequence; subsequently, a preset time interval near the detected state transition point is forcibly marked as an uncertain state. In one specific embodiment, the preset time interval can be set to a 2-minute period before and after the state transition point. This calibration operation effectively shields time series skew errors caused by manual work reporting or system recording, resulting in a calibrated equipment state sequence with clearer physical boundaries.

[0078] A rigorous physical power boundary is assigned to each moment in the above-mentioned calibrated equipment state sequence. Specifically, the state power band includes the lower and upper power boundaries corresponding to the equipment in different coarse-grained states. The upper and lower power boundaries are determined by one or more methods, including equipment nameplate parameters, equipment operating instructions, historical power statistics, and field engineering experience.

[0079] Specifically, for uncertain states newly generated due to boundary calibration, since they are in the transition period between two defined operating states, the state power band corresponding to the uncertain state is set as the union of the state power bands of the two states before and after the state switching action (e.g., switching from standby state to processing state). This ensures that the actual power value of the device at any given time strictly falls within the envelope of its corresponding state power band.

[0080] S2. Based on the aggregated power sequence and the calibrated device state sequence, construct an anchor window and a candidate reference window, and select a reference window from the candidate reference window according to the aggregated power difference, device state change relationship and state power band change compatibility between the anchor window and the candidate reference window.

[0081] In one specific embodiment, step S2 includes the following sub-steps:

[0082] S201, data alignment and time window division, and construction of anchor windows and candidate reference windows.

[0083] First, the acquired aggregated power sequence and the calibrated device status sequences of each device are strictly time-aligned according to timestamps. Based on this, the time-aligned long sequence data is continuously divided into multiple time windows according to a preset window length and a preset sliding step size. After this division process, each time window completely contains the aggregated power sequence within that specific time segment, the device status segments of each device, and the corresponding state power band data matched based on the status of each device. The aforementioned divided window data serves as the base sample pool for subsequent window pairing sampling.

[0084] From the multiple time windows generated by the above division, any one time window is selected as the current benchmark for investigation and denoted as anchor window a. Subsequently, several time windows outside anchor window a are extracted from other time intervals on the time axis as candidate reference windows, denoted as candidate reference windows r. In order to accurately perform subsequent feature comparison measurements on the same dimension, the candidate reference windows and anchor windows must have the same time span to allow for equal-scale comparative analysis of the differences in aggregate power, equipment state change relationships, and state power band change relationships between them.

[0085] S202. Calculate the difference in aggregated power between the anchor window and the candidate reference window, the relationship between device state changes, and the compatibility of state power band changes.

[0086] In order to select the most instructive window pairs from a large number of candidate samples, the candidate reference windows must meet the following conditions: there is a sufficient difference in total power between them and the anchor window, the number of devices whose states change is within a reasonable range, and the amplitude of the change in state power band is physically and logically compatible with the difference in total power.

[0087] Therefore, for each candidate reference window, a relational metric representing the above conditions is calculated, specifically including:

[0088] (1) Calculate the polymerization power difference. In this embodiment, the polymerization power difference is specifically calculated using the difference intensity in the summary table. To measure and characterize. Its calculation formula is:

[0089]

[0090] In the formula, T is the length of the time window. and These represent the aggregated power sequence values ​​of the anchor window and the candidate reference window at time t, respectively.

[0091] (2) Calculate the equipment state change relationship. In this embodiment, the equipment state change relationship is specifically calculated using the equipment state relationship weight. To conduct a quantitative assessment. The calculation formula is:

[0092]

[0093] In the formula, For indicator functions; The maximum number of devices allowed to undergo state changes; The number of devices whose status changes according to the set preferences; This is a state change indication value for device i. The value is 1 when device i changes state between the anchor window and the candidate reference window, and 0 otherwise.

[0094] (3) Calculate the compatibility of state power band changes. In this embodiment, the compatibility of state power band changes is specifically determined by the compatibility degree parameter. The calculation is as follows:

[0095]

[0096] In the formula, and These represent the center values ​​of the state power bands corresponding to the states of device i in the anchoring window a and the candidate reference window r, respectively. The introduced smooth miniterms are used to ensure the stability of the division operation.

[0097] S203. Calculate the relationship quality score and complete the reference window filtering.

[0098] Based on the three relationship metrics obtained in step S202, a comprehensive relationship quality score between each candidate reference window and the anchor window is calculated. The formula for calculating the relationship quality score is as follows:

[0099]

[0100] Among them, relationship quality score The value range of is within [0,1]; This is a preset power normalization function, specifically used to perform nonlinear mapping adjustment on the difference intensity distribution of the candidate reference window, in order to facilitate subsequent sorting and filtering operations.

[0101] The final relationship quality score is calculated. As a result, for each set anchor window 'a', several candidate reference windows with different relation patterns and higher relation quality scores are selected, thus forming a high-confidence reference window set R(a). This screening action eliminates irrelevant or noisy data pairs, providing high-quality data source support for constructing accurate weak supervision signals.

[0102] S3. Based on the anchor window and the reference window, construct a weak supervision signal including a time-varying weak supervision signal and a device responsibility weak supervision signal, and train the pre-built device power recovery model by combining the aggregated power sequence, the calibrated device state sequence and the state power band to obtain the target power recovery model.

[0103] For any anchor window 'a' and its corresponding reference window 'r' obtained in step S2, the aggregated power segment, device state, and corresponding state power band of each device are obtained in the anchor window and reference window, respectively. Based on the aggregated power difference, device state change relationship, and state power band change relationship between the anchor window and the reference window, a cross-window weak supervision signal is constructed for training the device-level power recovery model. In a specific embodiment, step S3 includes the following sub-steps:

[0104] S301, constructing a weak monitoring signal with time variation.

[0105] The time-varying weak supervision signal represents the relative weight of the difference in total power between the anchor window and the reference window on the time axis. For any pair of anchor windows and reference windows, although... It can reflect the magnitude of the time-by-time summary power difference, but its direct use is easily affected by window baseline differences. To highlight areas more likely to contain event variations, soft event weights are constructed using the local variation differences in summary power. :

[0106]

[0107] In the formula, This is a temporal neighborhood expansion operator used to extend the detected event location to its neighborhood range; It is a normalized time smoothing kernel used to reduce instability caused by event boundaries; These are the local change thresholds within the anchor window and the reference window, respectively; and These represent the aggregated power of the anchor window and the reference window at time t, respectively.

[0108] Based on this, the soft event weights are fused with the difference magnitudes in the summary table to obtain the time-varying weak supervision signal. :

[0109]

[0110] In the formula, The mixing coefficient; This is the time normalization function. Through this normalization process, a time-varying weight distribution that is non-negative in the time dimension and sums to 1 can be obtained, eliminating the tendency for the model to uniformly shift across the entire window and helping to reduce the tendency of the model to interpret local power changes as uniform shifts across the entire window.

[0111] S302, Weak supervision signal for construction equipment responsibility.

[0112] Weakly supervised equipment liability signals are used to characterize the degree to which different equipment interprets the power differences in the aggregate table between the anchor window and the reference window. For any pair of anchor windows and reference windows, since the cross-window aggregated power difference may be caused by explicit changes in equipment state or by power fluctuations within the same processing state of the equipment, equipment liability includes the distribution of equipment interpretation liability caused by explicit state changes. and the distribution of residual candidate responsibilities caused by power changes during operation. .

[0113] First, construct the explicit state change candidate responsibility distribution. For device i, determine whether a state change has occurred between the anchor window and the reference window. If a state change has occurred, then its state change indication... Set to 1 if the value is less than 0 otherwise. Further, based on the magnitude of the change in the state power band center value of device i within the anchoring window and the reference window, and the compatibility between this change and the aggregated power difference, determine the explicit state change candidate responsibility for device i:

[0114]

[0115] Secondly, the residual candidate responsibility distribution is constructed. For equipment that is in processing mode in both the anchoring window and the reference window, even if its state does not change, power differences may occur due to variations in processing load, process stage, or short pauses. Therefore, residual candidate responsibilities are preferentially assigned to equipment that is in processing mode in both windows and has a wider state power band:

[0116]

[0117]

[0118] In the formula, when the device is in processing mode (Proc) in both windows, , and These are the upper and lower bounds of the power for the corresponding state of the device.

[0119] Based on the degree to which explicit state changes explain the differences in polymerization power, and by combining the two types of responsibility distributions, the final weak supervision signal of equipment responsibility is obtained. :

[0120]

[0121]

[0122] Through the above steps, the time-varying weak supervision signals and device responsibility weak supervision signals corresponding to the anchor window a and the reference window r can be obtained. These weak supervision signals are not actual device power labels, but rather relational constraints constructed from aggregated power, device state, and state power bands. They are used to guide the device-level power recovery model in learning the correspondence between cross-window power differences and device state changes during subsequent model training.

[0123] S303. Construct the network architecture for the device power recovery model.

[0124] A device-level power recovery model is constructed, which is used to generate power estimation results for each device within a time window based on the aggregated power of shared meters, device status, and status power bands. In one embodiment, the device-level power recovery model includes a time-series feature extraction module, a status-aware power decomposition module, and a feasibility projection module.

[0125] The temporal feature extraction module is used to extract temporal features from the aggregated power sequence. Specifically, the aggregated power sequence within a time window is input into the temporal feature extraction module to obtain the corresponding temporal feature representation. The temporal feature extraction module can employ convolutional networks, recurrent neural networks, attention networks, Transformer networks, temporal convolutional networks, or combinations thereof. In one embodiment, the temporal feature extraction module includes a multi-scale one-dimensional convolutional module and a self-attention encoding module; wherein, the multi-scale one-dimensional convolutional module is used to extract local power change features at different time scales, and the self-attention encoding module is used to extract long-range dependency features within the time window.

[0126] The state-aware power decomposition module fuses the aforementioned time features with the calibrated device state sequence and state power band. This module includes a non-processing state power estimation path and a processing state power estimation path. The non-processing state power estimation path estimates the power of the device in a stopped, standby, or uncertain state, while the processing state power estimation path estimates the power of the device in a processing state.

[0127] For the state power band of device i at time t, let its lower power bound and upper power bound be respectively... and The state-aware power decomposition module outputs the in-band position parameters of the device at time t. ,in The value range is [0,1], and the device-level estimated power is generated according to the following formula. :

[0128]

[0129] in, This represents the device-level estimated power of device i at time t. In this way, the device-level estimated power is limited to the corresponding state power band.

[0130] The feasibility projection module is used to adjust the estimated power of the aforementioned devices to a physically feasible region defined by the overall meter consistency and the state power band during model training and application. Specifically, the feasibility projection module takes the estimated power of the devices output by the state-aware power decomposition module as input and projects it onto the feasible region defined by the aggregated power of the shared meter and the state power band of the devices to obtain the projected power. The specific implementation of the feasibility projection module can be found in step S5.

[0131] S304. Model training based on joint loss function.

[0132] The device-level power recovery model is trained using aggregated power sequences, calibrated device state sequences, state power bands, and time-varying weak supervision signals and device responsibility weak supervision signals. In one embodiment, the training loss includes one or more of single-window physical consistency loss and anchor-reference window relationship loss.

[0133] Single-window physical consistency Losses are used to constrain the consistency between the device's estimated power and the aggregated power and state power bands of the shared meters, including total meter consistency losses. and state power band feasibility loss Among them, the overall meter consistency loss is used to constrain the sum of the estimated power of each device at the same time to be consistent with or nearly consistent with the aggregated power of the shared meter; the state power band feasibility loss is used to constrain the estimated power of each device to be within its corresponding state power band range. Specifically:

[0134]

[0135]

[0136]

[0137] In the formula, This results in a loss of consistency in the overall table. For state power band feasibility loss; and These represent the lower and upper bounds of the state power band of device i at time t, respectively. and These are the corresponding weighting coefficients.

[0138] The anchor-reference window relationship loss is used to constrain the device power variation output by the model to be consistent with the weak supervision signal constructed in steps S301 and S302. The anchor-reference window relationship loss includes one or more of time-varying distribution loss and device responsibility distribution loss. Specifically, the time-varying distribution loss constrains the distribution of cross-window device power variation predicted by the model to be consistent with the time-varying weak supervision signal. The device responsibility distribution loss constrains the distribution of device responsibility predicted by the model to be consistent with the device responsibility weak supervision signal. It is worth noting that since the weak supervision signal is not a true label, it is not used as a deterministic supervision signal, but rather determined through distribution alignment and window relationship quality. Weighted constraints are applied to reduce the impact of low-confidence reference windows on training. Specifically, the formula for the anchor-reference window relationship loss is:

[0139]

[0140]

[0141]

[0142] In the formula, Loss due to time variation; For equipment liability losses; and These are the corresponding weighting coefficients.

[0143] Through the above training process, the device power recovery model can learn the correspondence between the aggregated power of shared meters, device status, state power band, and weakly supervised signals across windows without using the actual power labels of the devices. After multiple iterations to reach the preset convergence condition, the trained network topology and its corresponding weight parameters are stored together. The resulting solidified model is the target power recovery model, which can be used for forward inference prediction of subsequent application data.

[0144] S4. Obtain the target time window to be restored, and input the aggregated power sequence, calibrated device state sequence and state power band within the target time window into the target power recovery model to generate the estimated power of each device within the target time window.

[0145] In one specific embodiment, step S4 includes the following sub-steps:

[0146] S401. Obtain the target time window to be restored and the corresponding application period sequence data, and use the target power recovery model to perform forward inference to generate the estimated power of the equipment.

[0147] After model training is complete, the actual power decomposition task begins. First, the target time window for restoring the power of the devices to be recovered is determined. Then, within this time window, the aggregated power sequence of shared meters, the calibrated device status, and the corresponding state power band for each device are acquired in real-time or offline. This step repeats the basic data acquisition and state boundary calibration actions on the application side, ensuring that the application-period data fed into the model is completely consistent with the training-period data in terms of format and physical meaning.

[0148] The prepared application-period data is input into the trained device power recovery model (i.e., the target power recovery model). The model performs forward inference calculations based on the learned cross-window association mapping relationship and outputs the device-level estimated power of each device within the target time window. The generated device estimated power data is then stored in a database or memory, forming the preliminary estimate of device power decomposition.

[0149] S402. Analyze and evaluate the potential deviations from physical constraints in the initial output results of the model.

[0150] In practical applications of discrete manufacturing, results obtained directly from neural network models are not always perfect. Specifically, due to the potential impact of state recording errors, concurrent equipment operation, and model estimation errors on the model output, the device-level estimated power may not strictly meet the overall table consistency constraint or the state power band constraint. This objective existence of physical deviation clearly explains why the estimated power directly output by the neural network cannot be used as the final result, and constitutes the technical motivation for subsequent post-processing correction of this preliminary result. This provides causal support for executing step S5 (i.e., feasibility projection processing).

[0151] S5. Adjust the estimated power of the equipment to the feasible region defined by the consistency constraint of the total table and the state power band constraint through feasibility projection to obtain the equipment power recovery result of each equipment; wherein, the consistency constraint of the total table is determined according to the aggregated power sequence within the target time window, and the state power band constraint is determined according to the state power band within the target time window.

[0152] In one specific embodiment, step S5 includes the following sub-steps:

[0153] S501. Construct the feasible region under the dual constraints of overall table consistency and state power band.

[0154] For any time t within the target time window, construct a feasible region jointly defined by the overall table consistency constraint and the state power band constraint. The overall meter consistency constraint is used to ensure that the sum of the power of each device at the same time is equal to or approximately equal to the aggregated power collected by the shared meter, so as to ensure the principle of energy conservation; the state power band constraint is used to ensure that the power of each device is strictly within its corresponding state power band range, so as to ensure that the device-level power does not violate the objective nameplate parameters or operating procedures.

[0155] The feasible region is specifically expressed as follows:

[0156]

[0157] In the formula, This represents the feasible region at time t within the target time window; This represents the power vector of each device at time t; Let N represent an N-dimensional real space, where N is the total number of devices sharing the same electricity meter; This represents the power of device i at time t; This represents the aggregated power collected by the shared meter at time t; and Let represent the lower and upper bounds of the state power of device i at time t, respectively.

[0158] S502, dynamically set the projection adjustment cost based on equipment status characteristics.

[0159] To ensure that power corrections are reasonably allocated to equipment capable of handling fluctuations, projection adjustment costs (i.e., movement costs) for different equipment at different times are set based on equipment status, state power bandwidth, and state boundary uncertainties. ).

[0160] The specific hierarchical setting logic is as follows: for equipment in a shutdown state with a narrow power band, a higher projection adjustment cost can be set to avoid such equipment absorbing excessive power correction; for equipment in a processing state, an uncertain state, or a wide power band, a lower projection adjustment cost can be set to allow such equipment to undertake more reasonable power correction. The design of this cost function fully incorporates engineering experience from discrete manufacturing sites, making the projection process not only a purely mathematical optimization but also an adjustment that conforms to physical laws.

[0161] S503. Construct and solve the weighted feasibility projection problem, obtain the power recovery results, and store them in the database.

[0162] Within the feasible region defined in step S501, the weighted deviation between the estimated power of the device and the projected power of the device is minimized by solving the weighted projection problem, thereby obtaining the final device power recovery result.

[0163] In one implementation, the projected device power is obtained by solving the following mathematical optimization problem. :

[0164]

[0165] In the formula, Estimate the device power generated by the model forward inference in step S4; The movement cost set in step S502 is used to control the correction magnitude of the projection layer for different devices and states. This parameter is determined by the device state and power bandwidth.

[0166] Through the above optimization solution, the uncontrolled predicted power output by the simple neural network is mapped back to a space that conforms to the actual physical operating constraints on site. The final solution yields the continuous device power recovery results for each device at each time step. The data is then stored in the database for use in underlying energy consumption analysis and high-dimensional report display.

[0167] S504, output the power recovery results of the device and collect energy consumption and carbon emissions.

[0168] After obtaining the power recovery results of each of the aforementioned devices, the continuous power recovery results of each device within the target time window are output and presented. The power recovery results can be stored in a database according to the device number, timestamp, and recovered power value, or they can be associated with device status, production task, order number, process number, or product number to support subsequent in-depth device energy consumption analysis and product energy consumption aggregation.

[0169] In one specific embodiment, based on the device power recovery results, the energy consumption of the device within a preset time interval is further statistically analyzed to obtain the device energy consumption attribution result. For device i within the time interval... The energy consumption within the space is calculated using the following formula:

[0170]

[0171] In the formula, This represents the energy consumption of product k on device i; This represents the processing time interval of product k on equipment i; This represents the power recovery result of device i at time t; Indicates the time step or sampling interval.

[0172] Furthermore, given the availability of the product processing time interval, the processing time interval of the product on the corresponding equipment is time-matched with the continuous power recovery results of the equipment. The power of the corresponding equipment is then integrated over time within the processing time interval to obtain the energy consumption of the product on that equipment. Subsequently, the energy consumption of the product on multiple related equipment is summarized to obtain the energy consumption aggregation results for the product manufacturing stage. :

[0173]

[0174] In the formula, This represents the total energy consumption of product k during its manufacturing stage; N represents the total number of equipment related to the processing of this product.

[0175] In another specific embodiment, the product energy consumption aggregation results can be further combined with energy carbon emission factors for carbon emission accounting during the manufacturing stage. For example, when the equipment uses electricity, the indirect carbon emissions during the product manufacturing stage can be calculated based on the product energy consumption aggregation results and the corresponding electricity carbon emission factors. It should be noted that product energy consumption aggregation and carbon emission accounting during the manufacturing stage are downstream applications of equipment power recovery results and do not limit the present invention to product carbon emission accounting scenarios.

[0176] Through the above steps, the present invention can convert the power of equipment that cannot be directly observed under a shared electricity meter into outputtable, storable and traceable equipment power recovery results, and provide a data foundation for equipment-level energy consumption attribution, product energy consumption collection and carbon emission accounting in the manufacturing stage without the installation of equipment-level electricity meters.

[0177] like Figure 2 As shown, the present invention also provides a power recovery device for equipment with one meter for multiple machines in discrete manufacturing, based on the power recovery method for equipment with one meter for multiple machines in discrete manufacturing as described above. The device includes:

[0178] The data acquisition module 1 is used to acquire the aggregated power sequence collected by the shared electricity meter, and the coarse-grained state sequence of each of the multiple devices jointly measured by the shared electricity meter, to perform state boundary calibration on the coarse-grained state sequence to obtain the calibrated device state sequence, and to acquire the state power band corresponding to each device in the calibrated device state.

[0179] Window construction module 2 is used to construct an anchor window and a candidate reference window based on the aggregated power sequence and the calibrated device state sequence, and to select a reference window from the candidate reference window according to the aggregated power difference, device state change relationship and state power band change compatibility between the anchor window and the candidate reference window;

[0180] The weak supervision signal construction module 3 is used to construct a weak supervision signal including a time-varying weak supervision signal and a device responsibility weak supervision signal based on the anchoring window and the reference window, and to train the pre-built device power recovery model by combining the aggregated power sequence, the calibrated device state sequence and the state power band to obtain the target power recovery model.

[0181] The device power estimation module 4 is used to obtain the target time window to be restored, input the aggregated power sequence, the calibrated device state sequence and the state power band within the target time window into the target power recovery model, and generate the device estimated power of each device within the target time window.

[0182] The result output module 5 is used to adjust the estimated power of the equipment to a feasible region defined by the overall table consistency constraint and the state power band constraint through feasibility projection, so as to obtain the equipment power recovery result of each equipment; wherein, the overall table consistency constraint is determined according to the aggregated power sequence within the target time window, and the state power band constraint is determined according to the state power band within the target time window.

[0183] Each of the above modules is used to perform the respective steps in the above-described method for restoring the power of equipment in discrete manufacturing with one meter for multiple machines. The specific implementation methods are as described in the above-described method embodiments, and will not be repeated here.

[0184] like Figure 3 As shown, the present invention also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores all the data required for the process of a multi-machine metering method for discrete manufacturing. The network interface communicates with external terminals via a network connection. The computer program is executed by the processor to implement the multi-machine metering method for discrete manufacturing.

[0185] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.

[0186] An embodiment of this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described equipment power recovery methods for discrete manufacturing multi-machine metering.

[0187] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by hardware related to computer program instructions. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the above method embodiments. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM), such as dynamic RAM (used as main storage) or static RAM (commonly used as cache memory). By way of illustration and not limitation, RAM has various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), and Rambus DRAM (RDRAM).

[0188] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0189] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for restoring equipment power for multi-machine metering in discrete manufacturing, characterized in that, include: S1. Obtain the aggregated power sequence collected by the shared electricity meter, and the coarse-grained state sequence of each of the multiple devices jointly measured by the shared electricity meter. Perform state boundary calibration on the coarse-grained state sequence to obtain the calibrated device state sequence, and obtain the state power band corresponding to each device in the calibrated device state. S2. Based on the aggregated power sequence and the calibrated device state sequence, construct an anchoring window and a candidate reference window, and select a reference window from the candidate reference window according to the aggregated power difference, device state change relationship and state power band change compatibility between the anchoring window and the candidate reference window; S3. Based on the anchor window and the reference window, construct a weak supervision signal including a time-varying weak supervision signal and a device responsibility weak supervision signal, and train the pre-built device power recovery model by combining the aggregated power sequence, the calibrated device state sequence and the state power band to obtain the target power recovery model; S4. Obtain the target time window to be recovered, and input the aggregated power sequence, calibrated device status sequence and status power band within the target time window into the target power recovery model to generate the estimated power of each device within the target time window. S5. Adjust the estimated power of the equipment to the feasible region defined by the consistency constraint of the total table and the state power band constraint through feasibility projection to obtain the equipment power recovery result of each equipment; wherein, the consistency constraint of the total table is determined according to the aggregated power sequence within the target time window, and the state power band constraint is determined according to the state power band within the target time window.

2. The equipment power restoration method for multi-machine metering in discrete manufacturing according to claim 1, characterized in that, S2 specifically includes: The aggregated power sequence and the calibrated device status sequence are time-aligned according to timestamps. The time-aligned sequence data is continuously divided into multiple time windows according to a preset window length and a preset sliding step size. Any time window is selected from the multiple time windows as the anchor window. In other time intervals outside the anchor window, several time windows are extracted as candidate reference windows; wherein, the candidate reference windows and the anchor window have the same time length in terms of time span; For each candidate reference window, calculate the aggregated power difference, device state change relationship, and state power band change compatibility between it and the anchor window, and obtain the corresponding total table difference strength, device state relationship weight, and compatibility parameters respectively. Based on the difference strength of the summary table, the weight of the device status relationship, and the compatibility parameter, a comprehensive relationship quality score is calculated for each candidate reference window. Based on the comprehensive relationship quality score, several candidate reference windows are selected from the candidate reference windows to form a reference window set.

3. The equipment power restoration method for multi-machine metering in discrete manufacturing according to claim 2, characterized in that, The comprehensive relationship quality score The calculation formula is: in, In the formula, The strength of the differences in the overall table; Weights for device status relationships; For compatibility parameters; is the preset power normalization function; T is the length of the time window; and These represent the aggregate power of anchor window a and candidate reference window r at time t, respectively. The function is an indicator function; N is the total number of devices that are jointly metered by a shared meter. The maximum number of devices allowed to undergo state changes; The number of devices whose status changes according to the set preferences; This is the state change indicator for device i. It takes the value of 1 when device i changes state between the anchor window and the candidate reference window, and takes the value of 0 otherwise. and These represent the center values ​​of the state power bands of device i in anchoring window a and candidate reference window r, respectively. This is a smooth minimum term.

4. The equipment power restoration method for multi-machine metering in discrete manufacturing according to claim 1, characterized in that, S3 specifically includes: Soft event weights are constructed by utilizing the local variation differences in the total power, and the soft event weights are fused with the amplitude of the total power variation to obtain a time-varying weak supervision signal. The equipment explanation responsibility distribution caused by explicit state changes and the residual candidate responsibility distribution caused by power changes within the operating state are constructed respectively. Based on the degree of explanation of the aggregate power difference by the explicit state changes, the equipment explanation responsibility distribution and the residual candidate responsibility distribution are weighted and fused to obtain the weak supervision signal of equipment responsibility. A device power recovery model is pre-constructed, comprising a time-series feature extraction module, a state-aware power decomposition module, and a feasibility projection module. The time-series feature extraction module extracts time features from the aggregated power sequence. The state-aware power decomposition module fuses the time features with the calibrated device state sequence and state power band, outputs the in-band position parameters of the device within the corresponding state power band, and generates the corresponding estimated device power. The target power recovery model is obtained by training a pre-built device power recovery model using a joint loss function based on single-window physical consistency loss and anchor-reference window relationship loss.

5. The equipment power restoration method for multi-machine metering in discrete manufacturing according to claim 4, characterized in that, The time-varying weak supervision signal The calculation formula is: In the formula, For soft events; For time neighborhood expansion operators; For normalized time smoothing kernel; , These are the local change thresholds within the anchor window and the reference window, respectively; and These represent the aggregated power of the anchor window and the reference window at time t, respectively. For indicator functions; The mixing coefficient; This is the time normalization function; This represents the magnitude of the differences in the summary table; For smooth minimum terms; The weak monitoring signal of equipment responsibility The calculation formula is: In the formula, Explain the distribution of responsibility for the equipment; The residual candidate responsibility distribution; This is an indication of the state change of device i; and These represent the center values ​​of the state power bands of device i in the anchor window and the reference window, respectively; For compatibility parameters; This is the indicator parameter when the equipment is in processing mode in both windows; it is set to 1 when both windows are in processing mode. and These are the upper and lower bounds of the power for the corresponding device state, respectively; The strength of the differences in the overall table; These are weighting coefficients; This is a smooth minimum term.

6. The equipment power restoration method for multi-machine metering in discrete manufacturing according to claim 4, characterized in that, The estimated power of the equipment In the formula, and These represent the lower and upper bounds of the state power band of device i at time t, respectively. This refers to the in-band position parameters of the device within the corresponding power band. The single-window physical consistency loss includes the total table consistency loss used to constrain the sum of the estimated power of each device at the same time to be consistent with the aggregated power, and the state power band feasibility loss used to constrain the estimated power of each device to be within the range of its corresponding state power band. The anchor-reference window relationship loss includes a time variation distribution loss for constraining the distribution of cross-window device power variation in the time dimension to be consistent with the weak time variation supervision signal, and a device responsibility distribution loss for constraining the device responsibility distribution to be consistent with the weak device responsibility supervision signal.

7. The equipment power restoration method for multi-machine metering in discrete manufacturing according to claim 1, characterized in that, S5 specifically includes: For any moment within the target time window, a feasible region is constructed, which is jointly defined by the consistency constraint of the total table and the state power band constraint. Based on the device state characteristics, the projection adjustment cost of different devices at different times is dynamically set. Within the feasible region, a weighted feasibility projection problem is constructed. The weighted feasibility projection problem is solved by minimizing the weighted deviation between the estimated power of the device and the projected power of the device, thereby obtaining the continuous device power recovery results for each device.

8. The equipment power restoration method for multi-machine metering in discrete manufacturing according to claim 7, characterized in that, The expression for the feasible region is: In the formula, This represents the feasible region at time t within the target time window; This represents the power vector of each device at time t; Let N represent an N-dimensional real space, where N is the total number of devices sharing the same electricity meter; This represents the power of device i at time t; This represents the aggregated power collected by the shared meter at time t; and These represent the lower and upper bounds of the state power of device i at time t, respectively. The projected device power of the device power recovery result Its formula is: In the formula, Estimated device power generated by model forward inference; The cost of adjusting the projection as set.

9. A power restoration device for equipment with one meter for multiple machines in discrete manufacturing, based on the power restoration method for equipment with one meter for multiple machines in discrete manufacturing according to any one of claims 1 to 8, characterized in that, The device includes: The data acquisition module is used to acquire the aggregated power sequence collected by the shared electricity meter, and the coarse-grained state sequence of each of the multiple devices jointly measured by the shared electricity meter, to perform state boundary calibration on the coarse-grained state sequence to obtain the calibrated device state sequence, and to acquire the state power band corresponding to each device in the calibrated device state. The window construction module is used to construct an anchor window and a candidate reference window based on the aggregated power sequence and the calibrated device state sequence, and to select a reference window from the candidate reference window according to the aggregated power difference, device state change relationship and state power band change compatibility between the anchor window and the candidate reference window; The weak supervision signal construction module is used to construct a weak supervision signal including a time-varying weak supervision signal and a device responsibility weak supervision signal based on the anchoring window and the reference window, and to train the pre-built device power recovery model by combining the aggregated power sequence, the calibrated device state sequence and the state power band to obtain the target power recovery model. The equipment power estimation module is used to obtain the target time window to be restored, input the aggregated power sequence, the calibrated equipment state sequence and the state power band within the target time window into the target power recovery model, and generate the equipment estimated power of each device within the target time window. The result output module is used to adjust the estimated power of the equipment to a feasible region defined by the overall table consistency constraint and the state power band constraint through feasibility projection, so as to obtain the equipment power recovery result of each equipment; wherein, the overall table consistency constraint is determined according to the aggregated power sequence within the target time window, and the state power band constraint is determined according to the state power band within the target time window.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

Images