Electrical equipment energy efficiency evaluation method and system based on neural network model

Abstract

This invention discloses a method and system for evaluating the energy efficiency of electrical equipment based on a neural network model, belonging to the field of energy efficiency data management. The method includes the following steps: acquiring model training data; extracting the recording time points of several cumulative metering data of the same type from the model training data; sorting and managing the cumulative metering data according to chronological order to obtain a cumulative metering sequence arranged chronologically; performing metering trajectory break detection on the cumulative metering sequence; extracting each break interval and anomaly type after detection; segmenting the cumulative metering sequence using each break interval as a boundary and correcting the cumulative trajectory; generating candidate supplementary sequences for each break interval; and performing data consistency verification on the candidate supplementary sequences. This solves the problem of overfitting to anomalies and weakening the ability to characterize normal operating conditions in existing technologies, ultimately improving the stability, accuracy, and engineering usability of energy efficiency evaluation results.

Description

Technical Field

[0001] This invention relates to the field of energy efficiency data management technology, and in particular to a method and system for evaluating the energy efficiency of electrical equipment based on a neural network model. Background Technology

[0002] The increasing adoption of neural network models in electrical equipment energy efficiency assessments stems from the fact that energy consumption is often influenced by equipment operating conditions, control strategies, environmental conditions, and user behavior, exhibiting significant nonlinearity, time-varying characteristics, and coupling. Traditional assessment methods based on fixed thresholds or a limited set of rules are prone to increasing misjudgments; while neural networks are better suited to learning normal patterns and efficiency boundaries under similar operating conditions from historical data. They maintain good fitting and generalization capabilities even when external conditions change, and can be more naturally integrated with the data acquisition, monitoring, and alarm closed loop of energy management systems.

[0003] Current energy efficiency assessments of electrical equipment based on neural network models typically involve multi-source data collection and processing (such as metering data, equipment status, environmental and occupancy information, etc.), followed by data cleaning and consistency verification, and the construction of samples organized by equipment-circuit-region-time window. Subsequently, energy efficiency labels or benchmarks (which may come from benchmarking, statistical stratification, or mechanistic constraints) are formed under a predetermined assessment caliber, and the neural network model is trained to output energy efficiency scores, deviations, or compliance judgments. After the model goes online, the system performs the same feature construction and inference output on real-time data, and triggers graded alarms in conjunction with confidence / data quality gating strategies. At the same time, operational statistics are accumulated for periodic retraining and adaptive parameter updates, forming a continuously iterative closed loop for energy efficiency management.

[0004] For example, Chinese invention patent CN116663996B discloses a method and apparatus for evaluating the energy efficiency level of wind turbine generator sets. The evaluation method is based on neural networks and adaptive thresholds, and includes the following steps: S1, establishing a neural network model based on historical data and training it to obtain predicted values ​​of core parameters; S2, calculating the reliability loss of core parameters based on the predicted and actual values ​​of core parameters; S3, setting the weights of core parameters; S4, calculating the energy efficiency level of the subsystem corresponding to the core parameters based on the weights and reliability loss; S5, based on extreme value theory, directly fitting the historical data of reliability loss using the exponential probability density form to obtain the initial value of the adaptive threshold for reliability loss; S6, establishing a sliding window to calculate the adaptive threshold at each time point; S7, judging whether the energy efficiency level of the wind turbine generator set meets the standard based on the adaptive threshold.

[0005] For example, Chinese invention patent CN115330162A discloses an industrial energy efficiency analysis system based on big data, which includes a data acquisition module, an index data storage module, and an energy efficiency assessment model building module. The data acquisition module is used to collect energy consumption data using the Hall current method, the index data storage module is used to process the data in blocks according to data attributes, and the energy efficiency assessment model building module is used to analyze and process the collected energy consumption. The system uses a neural network model. First, in order to obtain the number of neuron vectors in the hidden layer of the model, specific index factors are selected, and the current and voltage in the Hall device circuit are used for detection.

[0006] The above-mentioned technology has at least the following technical problems:

[0007] In existing technologies, the metering data sequence, which is continuously accumulated over time, may experience resets, rollbacks, retransmissions, and duplicate overwriting at the engineering site. This causes the incremental data within the window time interval obtained by the difference processing between adjacent data to produce negative values ​​or abnormal jumps, thereby contaminating the training labels and evaluation benchmarks with outliers. This manifests as drifting score outputs, unstable rankings, and anomaly location bias. The judgment path based on the prediction deviation to construct the threshold is highly sensitive to the stability of the deviation distribution. When data quality fluctuations, equipment replacements, or control strategy adjustments cause the deviation distribution to shift, the threshold is prone to drift during the time update process, resulting in inconsistent judgment boundaries and alternating false alarms and false negatives. Data should be processed according to the consistency characteristics of equipment operating conditions. When grouping and classifying data, if the stratification dimension is limited or the stratification granularity is too coarse, different usage behaviors and different service intensities are easily mixed into the same layer, causing the reference data to shift and leading to a decrease in the consistency and interpretability of the model's learning data. During the training phase, when samples with different confidence levels are uniformly included in parameter updates, the extreme errors generated by the low confidence window will form disproportionate perturbations in the gradient, causing the model to overfit to outliers and weaken its ability to characterize normal operating conditions. In the online phase, when the scoring output directly drives alarm triggering, situations such as the fracture recovery period, aggravated power loss, or inconsistent candidate evidence can cause short-term spikes or drops in the score, resulting in frequent alarm fluctuations, ultimately affecting the stability, accuracy, and usability of the energy efficiency assessment results. Summary of the Invention

[0008] To address the technical problem of existing models overfitting to outliers and weakening their ability to characterize normal operating conditions, this invention provides a method and system for evaluating the energy efficiency of electrical equipment based on a neural network model. The technical solution is as follows:

[0009] On the one hand, a method for constructing a neural network model for energy efficiency assessment of electrical equipment is provided. This method includes: Step 1: Obtaining model training data, extracting the recording time points of several cumulative metering data of the same type from the model training data, sorting and managing the cumulative metering data according to chronological order to obtain a cumulative metering sequence arranged chronologically; Step 2: Performing metering trajectory break detection on the cumulative metering sequence, extracting each break interval and anomaly type after detection, segmenting the cumulative metering sequence using each break interval as a boundary, and correcting the cumulative trajectory; Step 3: Generating candidate supplementary sequences for each break interval, performing data consistency verification on the candidate supplementary sequences, updating the cumulative metering sequence after verification, and associating scoring labels with the updated cumulative metering sequence; Step 4: Introducing scoring labels into the total loss function, iteratively updating the neural network model until the preset model training completion conditions are met, thus obtaining a neural network model for predicting the energy efficiency of electrical equipment.

[0010] On the other hand, a method for evaluating the energy efficiency of electrical equipment based on a neural network model constructed using a specific method is provided. This method includes: Step 1: Collecting cumulative metering data and operational status data of electrical equipment in the target area; sorting the cumulative metering data by event occurrence and arrival times to generate several cumulative metering sequences; Step 2: Performing metering trajectory break detection on each cumulative metering sequence; extracting each break interval and anomaly type corresponding to each cumulative metering sequence; generating corresponding candidate supplementary sequences for different break intervals and performing data consistency verification to update each cumulative metering sequence; Step 3: Based on the updated cumulative metering sequences, generating cumulative metering sample values ​​according to a preset evaluation window; associating these with operational status data to form evaluation window samples; inputting the evaluation window samples into the trained neural network model; and outputting the energy efficiency evaluation results of the electrical equipment in the target area.

[0011] On the other hand, a system for constructing a neural network model for energy efficiency assessment of electrical equipment is provided. This system includes: a metering sequence generation module for collecting cumulative metering data and operating status data of electrical equipment in a target area, sorting the cumulative metering data by event occurrence and arrival times, and generating several cumulative metering sequences; a data consistency verification module for performing metering trajectory break detection on each cumulative metering sequence, extracting each break interval and anomaly type corresponding to each cumulative metering sequence, generating corresponding candidate supplementary sequences for different break intervals and performing data consistency verification, thereby updating each cumulative metering sequence; and an assessment result output module for generating cumulative metering sample values ​​according to a preset assessment window based on the updated cumulative metering sequences, associating them with operating status data to form assessment window samples, inputting the assessment window samples into the trained neural network model, and outputting the energy efficiency assessment results of the electrical equipment in the target area.

[0012] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0013] 1. The electrical equipment energy efficiency assessment method and system based on a neural network model provided by this invention sorts and manages the occurrence and arrival times of events in the cumulative metering sequence, and performs metering trajectory break detection on the cumulative metering sequence to extract break intervals and anomaly types; segments the cumulative metering sequence with the break intervals as boundaries and corrects the cumulative trajectory, further generates candidate supplementary sequences for the break intervals and performs data consistency verification, determines the supplementary results and confidence labels and backfills them into the cumulative metering sequence, thereby obtaining a continuous and reliable cumulative metering sequence and associating it with stable scoring labels; during the training phase, robust regression terms and label loss terms based on scoring labels are introduced into the total loss function to suppress the disturbance of low-confidence samples and extreme errors on parameter updates and improve the model's ability to characterize normal operating conditions; in the online phase, weight reduction and gating are performed based on confidence labels to reduce scoring spikes and alarm jitter caused by break recovery periods and data missing data. This achieves the technical effects of stable score output, improved ranking consistency, and reduced anomaly localization bias, effectively solving the problems in existing technologies such as abnormal window increment caused by reset, callback, resend and repeated coverage, contamination of training labels and evaluation benchmarks, alternating false positives and false negatives caused by threshold drift, and decreased interpretability caused by layered reference line offset.

[0014] 2. This invention applies a segmentation strategy to the cumulative measurement sequence sequentially, managing the cumulative measurement sequence in segments with the break interval as the boundary. It calculates the connection offset for the segments after the break interval, ensuring a continuous connection between the corrected cumulative trajectory and the cumulative trajectory before the break interval, thus providing a stable trajectory base for the supplementary reconstruction of the break interval. Based on this, the invention generates a candidate supplementary sequence one for any break interval based on linear difference, and performs a data monotonic consistency check on the candidate supplementary sequence one. If the check passes, the candidate supplementary sequence one is filled into the corresponding break interval to update the cumulative measurement sequence. Simultaneously, the invention also generates a candidate supplementary sequence two and a candidate supplementary sequence three for any break interval. The candidate supplementary sequence two is obtained by integrating the instantaneous measurement intensity, and the candidate supplementary sequence three is inferred from the statistical characteristics of the segments before and after the break interval. The consistency residual rate of the candidate supplementary sequences is calculated based on the candidate supplementary sequences two and three. Different interpolation methods are selected based on the consistency residual rate, thereby achieving controllable continuity and reliability of the incremental reconstruction of the break interval window, reducing the impact of abnormal jumps caused by reset, callback, and re-issuance on subsequent energy efficiency assessments.

[0015] 3. This invention groups the model training data based on a preset working condition classification rule, and within each group, uses cumulative measurement sample sequences to form a statistical distribution, extracting efficient and inefficient reference lines. This maps the updated cumulative measurement sequences to rating labels ranging from zero to one hundred points, ensuring the rating labels are comparable for similar working conditions and reducing reference line shifts caused by the mixing of different usage behaviors. Based on this, the invention constructs window samples based on the cumulative measurement sequences, obtains predicted scores, and calculates the residual between the predicted scores and the rating labels. A robust regression approach is then used to construct the rating regression loss term, and a Hubble-type piecewise function is employed to suppress the disturbance of extreme errors to parameter updates. Simultaneously, the invention statistically analyzes the anomaly types and anomaly markers in the cumulative measurement sequences, matches corresponding label loss values ​​from the database, and accumulates them to form a label loss term. This explicitly incorporates data quality information such as duplicate coverage and late arrival markers into the optimization objective, thereby reducing the interference of low-confidence samples on the training direction. Finally, in each training round, the total loss value is calculated, the gradient vector of the total loss value with respect to the neural network model parameters is obtained, and the model parameters are updated in the opposite direction of the gradient vector with a step size determined by the learning rate. This process is repeated iteratively until a neural network model for predicting the energy efficiency of electrical equipment is obtained, thereby improving the stability, accuracy, and engineering usability of the energy efficiency score output. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating a method for constructing a neural network model for energy efficiency assessment of electrical equipment, provided in an embodiment of this application;

[0018] Figure 2 A flowchart illustrating an electrical equipment energy efficiency assessment method based on a neural network model constructed using a construction method, provided as an embodiment of this application;

[0019] Figure 3 A schematic diagram of a system structure for constructing a neural network model for energy efficiency assessment of electrical equipment, provided in an embodiment of this application;

[0020] Figure 4 This is a schematic diagram of the model mean square error comparison provided in the embodiments of this application;

[0021] Figure 5 This is a schematic diagram of model determination coefficient comparison provided in an embodiment of this application;

[0022] Figure 6 This is a schematic diagram showing the comparison of model prediction results provided in the embodiments of this application. Detailed Implementation

[0023] Embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present disclosure are shown in the drawings, it should be understood that embodiments of the present disclosure may be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present disclosure.

[0024] It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure. In the description of the embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "this embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects.

[0025] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0026] Example 1: As Figure 1 The diagram shown is a flowchart of a method for constructing a neural network model for energy efficiency assessment of electrical equipment, provided in an embodiment of this application. The method includes the following steps:

[0027] In school lighting energy efficiency assessment scenarios, model training data typically originates from rail-mounted energy meters or smart circuit breakers within building distribution boxes, which periodically report cumulative electricity consumption readings for lighting circuits in corridors, classrooms, dormitories, etc. Cumulative electricity consumption records from the same circuit and using the same metering caliber are grouped into a single cumulative metering data set, which is then included in the model training data management scope. To ensure the reliability of subsequent window incremental calculations and energy efficiency scoring training, during the data entry phase, both the event occurrence time and arrival time are retained for each cumulative metering record. The event occurrence time represents the sampling moment when the metering device generates the reading, while the arrival time represents the moment when the reading is received by the platform and written into storage after being uploaded via the communication link. In addition to cumulative metering data, model training data also includes operational status data (e.g., switch status, dimming level, occupancy trigger, and time period labels).

[0028] In this context, operational status data can be associated with windows by aligning the event occurrence timestamps. However, cumulative metering data is subject to various issues such as resets, rollbacks, reissues, and duplicate overwriting at the engineering site, which can undermine the premise of differential cumulative readings. This can contaminate window increments, scoring labels, and evaluation benchmarks. Timestamp alignment can only ensure that records from the same moment are retrieved within the same evaluation window, and cannot solve the non-monotonicity and version uncertainty problems of cumulative metering sequences. Specifically, even if the event occurrence timestamps are perfectly aligned, two adjacent cumulative readings may still decrease due to equipment resets or rollbacks, resulting in negative window differences. Furthermore, reissues and duplicate overwritings may lead to multiple records at the same event occurrence time, making it impossible to determine which record is the final valid record after alignment. This can cause extreme errors in window increments during the training phase, manifesting as score output drift, unstable rankings, and abnormal positioning deviations. Therefore, this application only introduces an enhanced governance process for cumulative metering sequences to ensure the stability of energy efficiency scores and engineering usability.

[0029] Taking the lighting circuit of a teaching building corridor as an example, the galvanic rail meter reports the cumulative electricity consumption every 15 minutes, and the evaluation window is 15 minutes. The cumulative electricity consumption records from 20:00 to 20:45 on a certain day (all event timestamps are correct) are as follows: 20:00 cumulative electricity consumption = 1250.40; 20:15 cumulative electricity consumption = 1250.55; 20:30 cumulative electricity consumption = 0.12 (the meter restarted and reset, recounting); 20:45 cumulative electricity consumption = 0.28. If only the event timestamps are aligned and sorted, and then the difference between adjacent readings is calculated, the window increment from 20:15 to 20:30 = 0.12 - 1250.55 = -1250.43, resulting in a significant negative value. This negative value directly pollutes the training labels and the scoring mapping benchmark, causing the model to be pulled by outliers during training, and online scoring is prone to short-term drops and alarm jitter. Therefore, aligning timestamps cannot eliminate the differential distortion caused by reset / rollback.

[0030] Taking resending and duplicate overwriting as an example, the same event occurrence time of 21:00 on the same night may correspond to two records: Record A (event occurrence time = 21:00, cumulative battery power = 0.60, arrival time = 21:01); Record B (event occurrence time = 21:00, cumulative battery power = 0.72, arrival time = 21:10, which is the corrected value after resending). If only the event occurrence time is aligned, both records will fall at 21:00, and the system cannot determine which one to use based solely on alignment; if Record A is mistakenly used, the subsequent window difference will be too small, the score will be raised, and inefficient loops will be missed. Therefore, it is necessary to retain the arrival time as well, using the record with the latest arrival time as the valid record, and generating a duplicate overwriting mark, which is marked on the data in the current window for subsequent reliability control.

[0031] In the current scenario, technicians set the interval threshold to 30 minutes (0.5 hours) based on communication status. This represents the maximum allowable interval between the occurrence and arrival times of an event. When the occurrence time of a cumulative measurement data is 21:00 and the arrival time is 21:10, the interval is 10 minutes, and no late marker is generated. When the occurrence time of another cumulative measurement data is 18:00 and the arrival time is 20:10, the interval is 130 minutes, exceeding the 30-minute threshold. The system generates a late marker for this data and reduces the confidence weight of the evaluation window covered by this data in subsequent window incremental reconstruction and scoring label construction to suppress the disturbance of the updated statistical reference line and model parameters caused by the resent data.

[0032] To address the aforementioned issues, this application implements metering trajectory breakage detection for cumulative metering sequences: the system sorts the reported data according to the event occurrence time (i.e., the sampling time when the electricity meter generates the reading), obtaining a cumulative metering sequence of cumulative electricity consumption (i.e., a cumulative reading sequence arranged chronologically). During the period from 20:00 to 20:45 on a certain day, the four records in the cumulative metering sequence are as follows: cumulative electricity consumption of 1250.40 at 20:00; cumulative electricity consumption of 1250.55 at 20:15; cumulative electricity consumption of 0.12 at 20:30; and cumulative electricity consumption of 0.28 at 20:45. The system reads adjacent data sequentially according to the time of the event, taking the record at 20:15 as the previous record and the record at 20:30 as the next record, and calculates the cumulative difference between the next record and the previous record (i.e., the difference between 1250.55 and 0.12) and the interval between the two records (i.e., 20:30 minus 15 minutes of 20:15, converted to 0.25 hours).

[0033] In the rollback detection, it is determined whether the cumulative electricity consumption has decreased: when the cumulative difference is negative and the absolute value of the cumulative difference exceeds the preset tolerance threshold in the database (representing the maximum allowed absolute value of the cumulative difference, for example, the tolerance threshold is set to 0.01), it is determined that there is a rollback anomaly in the adjacent interval. In this example, the cumulative difference is 0.12-1250.55=-1250.43, which is significantly negative and exceeds the tolerance threshold. Therefore, the system marks 20:15-20:30 as the rollback anomaly interval. Subsequently, the system performs a jump detection: the system converts the cumulative difference into average power according to the interval duration (i.e., the average power converted from the difference, calculated by dividing the window electricity increment by the window hours), and compares this average power with the installed power limit or the historical reliable limit; where the installed power limit is the sum of the rated power of the lighting fixtures in this circuit (for example, the installed power limit of this corridor circuit is 2.0kW), and the historical reliable limit is the high quantile reference limit of the average power distribution of this circuit during the historical stable period (for example, taking the 90th quantile gives 1.6kW). In this example, since the backtracking anomaly has been established, the system still calculates the average power for cross-validation: average power = (-1250.43) / 0.25 = -5001.72kW, which is a result that is clearly inconsistent with physical meaning. Based on this, the system retains the anomaly type of this interval as the backtracking dominant type and uses this result as the basis for subsequent fracture interval location.

[0034] The system continues to perform consistency checks on the next pair of adjacent records. The system uses 20:30 as the previous record and 20:45 as the next record, calculating the cumulative difference as 0.28 - 0.12 = 0.16. With an interval of 0.25 hours, the average power is calculated to be 0.16 / 0.25 = 0.64 kW. Simultaneously, the system reads the instantaneous power sequence (instantaneous power refers to the real-time power measurement value of the lighting circuit within the sampling period) from the same evaluation window, and integrates this instantaneous power sequence over time to obtain the integrated power consumption (integrated power consumption refers to the window power estimate obtained by accumulating instantaneous power over time). For example, if the average instantaneous power within the 20:30-20:45 window is approximately 0.62 kW, then the integrated power consumption is approximately 0.62 × 0.25 = 0.155 kWh. The system compares the differential energy (i.e., the cumulative difference of 0.16 kWh) with the power integral energy (0.155 kWh). If the deviation between the two exceeds a preset consistency threshold, an inconsistency is determined. This consistency threshold can be defined by a residual rate constraint. For example, the residual rate is defined as the absolute value of the difference between the differential energy and the power integral energy divided by the power integral energy. When the residual rate is greater than 0.20, the deviation is considered too large. In this example, the residual rate = |0.16 - 0.155| / 0.155 ≈ 0.032, which is less than 0.20. Therefore, the interval from 20:30 to 20:45 does not constitute an inconsistency anomaly.

[0035] After summarizing the detection results of rollback detection, jump detection, and consistency detection, the system extracts the break intervals from the detection results and outputs the anomaly type of the break intervals. Since 20:15-20:30 is marked as a rollback anomaly interval, the system uses it as the break interval, with the start and end times of the break interval being 20:15 and 20:30 respectively, and the anomaly type output is "rollback anomaly". Subsequently, the system segments the cumulative measurement sequence using this break interval as the boundary: 20:00 and 20:15 before the break interval belong to the first segment, and 20:30 and 20:45 after the break interval belong to the second segment. To correct the cumulative trajectory (the cumulative trajectory refers to the curve formed by the cumulative reading changing over time), the system calculates a connection offset for the second segment after the break interval, ensuring a continuous connection between the corrected cumulative reading of the second segment and the final value of the first segment: The offset is calculated as 1250.55 - 0.12 = 1250.43, using the last valid cumulative reading of 1250.55 before the break interval and the first cumulative reading of 0.12 in the second segment. This offset is then added to the entire cumulative reading of the second segment, resulting in a corrected cumulative energy consumption of 1250.55 at 20:30 and 1250.71 at 20:45. Through this segmentation and offset correction, the system eliminates the cumulative trajectory break caused by reset / rewind, restoring the monotonic accumulation characteristic of the corrected cumulative trajectory. This ensures that subsequent differential calculations based on the evaluation window will not produce negative values ​​or abnormal jumps, supporting stable output and interpretable benchmarking of the school lighting energy efficiency score.

[0036] Taking the lighting circuit of the corridor on the 3rd floor of Building A in the teaching building as an example, the rail-mounted energy meter in the distribution box reports one cumulative energy reading every 15 minutes. The system sorts the historical training data of the same circuit according to the time point of the event (the sampling time when the energy meter generates the reading), and obtains a cumulative metering sequence (a cumulative reading sequence arranged in chronological order). The cumulative metering sequence segment from 19:00 to 21:00 on a certain workday is as follows (unit: kWh): 19:00 is 1250.10, 19:15 is 1250.18, 19:30 is 1250.26, 19:45 is 1250.34, 20:00 is 1250.41, 20:15 is 1250.49, 20:30 is 0.06, 20:45 is 0.14, and 21:00 is 0.22. The system sequentially takes two adjacent records as the previous and next records according to the order of the event occurrence time, and calculates the cumulative difference (cumulative amount of the next record minus the cumulative amount of the previous record) and the interval duration (the time difference between the occurrence times of two adjacent events, which is 15 minutes or 0.25 hours in this example). Within each adjacent interval, it sequentially performs rollback detection, jump detection, and consistency detection: Rollback detection is used to determine whether the cumulative power consumption has decreased, and the judgment rule is that the cumulative difference is negative and exceeds the tolerance threshold; Jump detection is used to determine whether the average power converted from the difference exceeds the upper limit of installed power or the upper limit of historical reliable upper quantile, where the average power is obtained by "cumulative difference ÷ interval duration"; Consistency detection is used to determine whether the deviation between the differential power consumption and the power integral power consumption is too large, where the power integral power consumption is obtained by integrating the instantaneous power sequence within the window over time.

[0037] Within the adjacent interval from 20:15 to 20:30, the cumulative power consumption of the previous record was 1250.49, and the cumulative power consumption of the next record was 0.06. The difference in cumulative power consumption is 0.06 - 1250.49 = -1250.43, which is a significantly negative value and exceeds the tolerance threshold. Therefore, the rollback detection is valid. At the same time, the average power of this interval is -1250.43 ÷ 0.25 = -5001.72 kW, which clearly does not conform to the physical boundary of the lighting circuit. The sudden jump detection can also serve as evidence of crossover anomalies. Based on this, the system marks 20:15-20:30 as a broken interval and the anomaly type as a rollback anomaly (this anomaly type is used to characterize the cumulative trajectory break caused by reset / rewind). The system merges adjacent intervals with consecutive anomalies. In this example, the broken interval is a single segment, 20:15-20:30. The system then segments the cumulative measurement sequence using the break interval as the boundary: the first segment is before the break interval (19:00 to 20:15), and the second segment is after the break interval (20:30 to 21:00). A connection offset is calculated for the second segment to ensure continuity between the corrected cumulative trajectory of the second segment and the final value of the first segment. The connection offset is calculated as "the last valid cumulative value before the break interval - the first cumulative value after the break interval," which in this example is 1250.49 - 0.06 = 1250.43. The system adds 1250.43 to each cumulative reading in the second segment to obtain the correction value: 20:30 is corrected to 1250.49, 20:45 to 1250.57, and 21:00 to 1250.65, thus eliminating the impact of the break on subsequent differences and restoring the continuity of the cumulative trajectory after the break interval.

[0038] Next, a candidate supplementary sequence 1 is generated for the broken interval, and after performing a data monotonic consistency check on the candidate supplementary sequence 1, the cumulative metering sequence is backfilled and updated. The candidate supplementary sequence 1 refers to the sequence generated by linearly interpolating the cumulative reading points that need to be filled within the broken interval; linear interpolation refers to allocating the cumulative values ​​of intermediate points according to the time ratio, using the front boundary point and the back boundary point of the broken interval as endpoints. In this example, the front boundary point of the broken interval is the cumulative electricity of 1250.49 at 20:15, and the back boundary point of the broken interval is the cumulative electricity of 1250.49 at 20:30 after the connection correction. If the system has missing intermediate sampling points within the broken interval (for example, the electricity meter should have reported two records at 20:20 and 20:25 but failed to do so), the system will use 20:20 and 20:25 as the time points to be filled and generate candidate supplementary sequence one. Since the endpoints of 20:15 and 20:30 are the same, the linear difference yields a supplementary value of 1250.49 for 20:20 and 1250.49 for 20:25. The system then performs a data monotonicity consistency check on candidate supplementary sequence one. This check ensures that the cumulative trajectory after filling does not decrease over time and does not produce negative differences. In this example, the filled sequence is 1250.49 at 20:15, 20:20, 20:25, and 20:30, satisfying the non-decreasing constraint, and the check passes. Based on this, the system fills the missing points within the broken interval with candidate supplementary sequence one and updates the cumulative metering sequence, ensuring that the broken interval also has a continuous reading basis suitable for window differencing.

[0039] To associate the updated cumulative measurement sequence with the scoring labels, the model training data is first grouped according to the preset operating condition classification rules. Operating condition refers to a statistically comparable combination of operating states of lighting circuits, such as circuit type (corridor / classroom / dormitory), date type (weekday / weekend / holiday), day / night time period, light-on percentage range, dimming range, etc. Grouping means classifying samples that meet the same set of operating condition conditions into the same set, avoiding comparing different usage behaviors and windows with different service intensities together.

[0040] For example, the cumulative electricity consumption readings at 20:00, 20:15, 20:30, 20:45, and 21:00 are 1250.41, 1250.49, 1250.58, 1250.66, and 1250.74 (in kWh), respectively, with adjacent differences corresponding to the electricity consumption increment within the evaluation window. During the training phase, the system needs to associate scoring labels with each cumulative metering sequence. Therefore, the model training data is first grouped according to preset operating condition classification rules, such as limiting it to corridor lighting circuits, weekdays, 20:00-22:00, a lighting on-duty ratio of 0.8-1.0, and a dimming range of 0.9-1.0, thereby avoiding baseline drift caused by the mixing of different usage behaviors.

[0041] Within each group, instead of directly scoring individual windows, a mean index for mapping is calculated for each cumulative metering sequence. Specifically, the cumulative metering sequence is divided by adjacent differences within an evaluation window (e.g., 15 minutes) to obtain the window's electricity consumption sequence. The mean of this window's electricity consumption sequence is then calculated as the sequence mean of the cumulative metering sequence. Here, the sequence mean refers to the average electricity consumption level of the sequence at the evaluation window granularity, which can be obtained by dividing the difference in cumulative readings at the beginning and end of the sequence by the number of windows. For example, the above sequence has four 15-minute windows from 20:00 to 21:00, and its sequence mean is (1250.74-1250.41) / 4=0.0825 (unit: kWh / 15min). The system calculates the sequence mean for all cumulative metering sequences within a group, obtaining a set of sequence mean sample sets, and sorts and statistically analyzes this set to form a statistical distribution. This statistical distribution is used to characterize the differences in average electricity consumption levels among different sequences within the same operating condition group.

[0042] Based on this statistical distribution, efficient and inefficient reference lines are extracted and used as mapping benchmarks from sequence means to scoring labels. Within each group, the system first aggregates the sequence means corresponding to all cumulative measurement sequences within that group into a sample set, and then sorts the sample set according to numerical value. Sorting means arranging the sequence means from smallest to largest, so that the earlier positions correspond to smaller average energy consumption levels, and the later positions correspond to larger average energy consumption levels. This sorting method is consistent with the evaluation direction that lower energy consumption is more efficient, which facilitates the subsequent extraction of reference values ​​based on quantile positions.

[0043] After sorting the samples from smallest to largest, the system selects values ​​from the sorted sequence according to preset quantile values. Quantile values ​​are pre-defined proportional positional parameters used to locate representative values ​​within the sorted sample set: for example, the 10th percentile represents the value in the top 10% after sorting from smallest to largest, and the 90th percentile represents the value in the bottom 10%. The purpose of using quantile values ​​is to avoid the influence of extreme samples introduced by directly using the minimum and maximum values, making the reference values ​​more stable within similar operating condition groups.

[0044] Based on the above sorting and quantile values, the system extracts high-efficiency and low-efficiency reference lines. The high-efficiency reference line uses the quantile value on the low-energy-consumption side to represent the typical sequence mean corresponding to the relative energy-saving level within the operating condition group; that is, the representative value at the beginning of the sorted sequence corresponding to the high-efficiency reference line. The low-efficiency reference line uses the quantile value on the high-energy-consumption side to represent the typical sequence mean corresponding to the relative energy consumption level within the operating condition group; that is, the representative value at the end of the sorted sequence corresponding to the low-efficiency reference line. Therefore, the high-efficiency reference line is usually smaller than the low-efficiency reference line, and together they constitute the scoring mapping interval for the operating condition group: the closer the sequence mean is to the high-efficiency reference line, the higher the scoring label; the closer the sequence mean is to the low-efficiency reference line, the lower the scoring label; when the sequence mean is lower than the high-efficiency reference line or higher than the low-efficiency reference line, the scoring label takes the upper or lower limit respectively to ensure that the range of scoring labels is consistent with the evaluation direction.

[0045] For example, within the above working condition group, the efficient reference line can be extracted as 0.07 and the inefficient reference line as 0.16 according to the preset quantile position.

[0046] The system then performs sequence mapping, which maps the mean of each cumulative econometric sequence to a rating label ranging from 0 to 100, thus obtaining sequence-level rating labels. The mapping can take the following form:

[0047] ;

[0048] Among them, clip( , 0, 1) indicates boundary constraints: when the parentheses contain When the result is less than 0, it is taken as 0; when it is greater than 1, it is taken as 1; otherwise, the original value is retained to ensure that the rating label falls within the range of 0 to 100. Following the example above, with an efficient reference line of 0.07 and an inefficient reference line of 0.16, the rating label for a cumulative measurement sequence with a mean of 0.07 is 100; the rating label for a cumulative measurement sequence with a mean of 0.16 is 0; and the rating label for a cumulative measurement sequence with a mean of 0.10 is 66.67. Thus, the system uses the mean of the cumulative measurement sequence as a mapping basis to map each cumulative measurement sequence within the same working condition group to a rating label of 0-100, providing a stable and comparable supervision signal for subsequent neural network model training.

[0049] The rating labels are constructed using a reference line normalization mapping. This is because, within the same operating condition group, the sequence mean can serve as a comparable indicator of energy consumption levels within that group, and lower energy consumption levels generally correspond to higher energy efficiency. To ensure benchmarking across different windows or sequences on the same scale, the system uses an inefficient reference line and an efficient reference line to form a baseline interval for the operating condition group, and normalizes the sequence mean within this interval using a linear ratio. Specifically, the inefficient reference line is subtracted from the sequence mean, so that the smaller the sequence mean, the larger the normalization result, and thus the higher the score. The subtraction of the efficient reference line from the inefficient reference line is used as the denominator, ensuring that the normalization result only reflects the relative position of the sequence mean between the two reference lines, eliminating differences in absolute energy consumption levels between different operating condition groups, and guaranteeing a consistent interpretation of the scores within each group.

[0050] Window samples are constructed based on the updated cumulative econometric sequence, and each window sample is input into the neural network model to output a predicted score. The system uses the score label as a supervision signal to construct a total loss function and use it to iteratively update the model parameters. The total loss function is defined as a weighted sum of the score regression loss term and the label loss term. In each round of training, the total loss is calculated for each window sample as follows:

[0051] ;

[0052] Where L is the total loss value, used to guide the update direction of model parameters in this round of training, H(d) is the rating regression loss, λ1 is the weight index of the rating regression loss term preset in the database, representing the proportion of the rating regression loss term to the total loss value, P is the label loss term, and λ2 is the weight index of the label loss term preset in the database, representing the proportion of the label loss term to the total loss value.

[0053] The specific expression for H(d) is as follows:

[0054] ;

[0055] d = ys - s, where d represents the label difference, ys is the score result of the neural network model for the sample in the window, s is the mean of the score labels of all cumulative measurement sequences in the sample in the window as a whole mapped to 0-100, M is the preset segmentation threshold in the database, representing the maximum value allowed by |d|, and specifying the switching point between the quadratic penalty and the linear penalty.

[0056] The Hubell-type piecewise function design for the rating regression loss aims to balance fitting accuracy and engineering robustness. Specifically, the bias is the difference between the predicted score and the mean of the window-level score labels. When this bias is within a reasonable fluctuation range, a quadratic penalty is used, making the loss more sensitive to small biases. This allows the model to converge quickly and accurately characterize the mapping relationship between scores and features when normal operating conditions predominate. When the bias exceeds the piecewise threshold, an approximately linearly increasing penalty is used. This avoids the gradient corresponding to extreme biases from being amplified exponentially with the bias, thereby suppressing the excessive pull of a small number of anomalous windows on the parameter update direction caused by measurement anomalies such as break recovery periods, re-sending coverage, late arrivals, or duplicate coverage, and reducing the risk of overfitting the model to outliers. The piecewise threshold specifies the switching point between quadratic and linear penalties, and its setting reflects the engineering constraint on the maximum allowable normal bias range: within the threshold, accuracy is emphasized; outside the threshold, robustness and suppression of outlier effects are emphasized. Based on the above mechanism, this design can improve the stability and generalization ability of the scoring output during the training phase, so that the model can still maintain the reliability of the scoring prediction in the field environment where data quality fluctuations and abnormal measurements are inevitable.

[0057] It needs to be explained that constructing window samples refers to using the updated cumulative metering sequence as the primary metering basis, and segmenting the time axis according to the evaluation windows preset by technicians, with each evaluation window serving as the smallest training sample granularity. Specifically, the system reads the cumulative readings of the cumulative metering sequence at the beginning and end boundaries of the window, and uses the difference between the two to represent the cumulative increment within that window, used to characterize the actual energy consumption intensity of that window. Simultaneously, to avoid situations where the cumulative increment alone cannot distinguish between different service intensities with the same energy consumption, the system also incorporates operational status data aligned with the window's time frame into the sample feature set, such as on / off status, dimming level, occupancy trigger, time period type, and outdoor environment. Therefore, each window sample does not only contain cumulative metering data, but is composed of window cumulative increment features, operational status features, and markers. These markers include at least duplicate overlay markers and late arrival markers, used to reflect the reliability of the metering data within that window.

[0058] First, the window increments and their quality characteristics (e.g., number of sequences, missing proportion, duplicate coverage markers, late arrival markers, and backsliding / sudden jump / consistency anomaly statistics) of multiple cumulative measurement sequences within the evaluation window are concatenated with the corresponding operational status characteristics of the window in a preset order to form the feature vector of the window samples. This concatenation is then normalized and standardized. Next, this feature vector is input into a neural network model. During forward inference, the neural network model undergoes several layers of linear transformations and nonlinear activations, each layer assigning a learnable weight G to each input feature and superimposing a learnable bias b. The output of this layer is then obtained through an activation function. The weights and biases are the model parameters to be updated during training. Their role is to allow the model to automatically learn which input features are more important, to what extent, and how the score should change under different feature combinations, thus forming a comprehensive representation of the energy usage behavior of the window. The basic calculation form of each layer can be expressed as Z = σ × (G × c + b), where c is the output (or input feature vector) of the previous layer, and G and b are the trainable parameters of that layer. The activation function is used to express the nonlinear response of the lighting load under switching state, dimming level and changes in environmental conditions. After obtaining the final comprehensive representation, the model maps it to a scalar through a single output node as the score prediction result of the window, which is denoted as the predicted score.

[0059] This embodiment uses energy efficiency scores as the primary output. Therefore, the training objective directly constrains the predicted score to be consistent with the score label. Operating state data is not ignored but is used as input features in model inference, indirectly influencing the formation of the predicted score. In other words, the loss function does not need to set error terms for each type of operating state data; the system uniformly incorporates operating state data into the input side, and the loss function constrains the score deviation at the output. During the back-up process, the model automatically learns the contribution relationship between each operating state feature and the score output. The reason for this design is that the score label itself comprehensively reflects the differences in energy consumption levels under similar operating conditions. During training, only the score output needs to be supervised to achieve unified modeling of multi-source features, avoiding the fragmentation of the training objective caused by directly transferring the complexity of the input features into the loss term.

[0060] Relying solely on the rating regression loss term, the model treats all window samples as equally reliable supervisory signals. However, cumulative measurement sequences in engineering sites may experience issues such as resets, rollbacks, resends, and duplicate overlays, leading to varying degrees of uncertainty in the cumulative increment of the same window even after supplementation or repair. To address this, the system introduces a label loss term, explicitly incorporating the measurement quality and anomaly background of the window into the training objective. The system first statistically analyzes the anomaly types associated with the window (e.g., rollback anomalies, jump anomalies, consistency anomalies) and anomaly labels (at least including duplicate overlay and late arrival labels). Then, it matches the corresponding label loss values ​​from various mapping tables in the database and sums them to obtain the label loss term for the window. The label loss value is set according to a consistent principle: higher penalty values ​​are assigned to anomalies with a greater impact on rating reliability; lower penalty values ​​are assigned to anomalies with a smaller impact or stronger correctability. The role of this label loss term is not to replace rating supervision, but rather to make the training process sensitive to differences in data quality, making model parameter updates more biased towards high-quality window samples.

[0061] The system weights and sums the score regression loss term and the label loss term according to weight coefficients to obtain the total loss function. These weight coefficients balance the accuracy of the fitted score labels with the strength of data quality constraints. The weight coefficients are determined through validation set parameter tuning, aiming to minimize score output fluctuations during scenarios such as the fracture recovery period and the re-coverage period, while ensuring acceptable scoring errors. During training, the system calculates the total loss for each window sample within the current batch and accumulates them to obtain the batch total loss. Then, it calculates the gradient of the batch total loss with respect to the model parameters and updates the model parameters in the opposite direction of the gradient. By repeatedly executing the loop of building window samples—obtaining predicted scores forward—calculating the total loss—updating parameters based on the total loss, until the preset training completion conditions are met, a neural network model for predicting the energy efficiency of electrical equipment is obtained.

[0062] In each training round, the total loss value is calculated based on the current training samples. The partial derivatives of the total loss value with respect to the neural network model parameters are then calculated and arranged in a predetermined order to obtain gradient vectors corresponding to each neural network model parameter (such as trainable weights and biases). Each dimension of the gradient vector corresponds to a partial derivative of a model parameter, reflecting the direction and sensitivity of the effect of small changes in that parameter on the total loss. The theoretical basis for this step is the chain rule and the backpropagation mechanism: the output of the neural network is composed of multiple differentiable operations, and the derivative of the total loss with respect to the parameters can be efficiently calculated by propagating the error signal layer by layer, thus obtaining the full parameter gradient while keeping the computational complexity under control.

[0063] The parameters are updated in the opposite direction of the gradient vector, with the update step size determined by the learning rate, as follows:

[0064] ;

[0065] For example, where b is the learning rate, θ k G represents the model parameters at the end of the k-th training round. k Indicates at θ k The gradient vector calculated at the current parameter point is used to control the step size of each update. The theoretical basis of this update rule is that, near the current parameter point, the total loss function can be approximated by the first-order Taylor expansion as the current loss value plus the inner product of the gradient and the parameter increment. When the parameter increment is in the negative gradient direction and the learning rate is small enough, the inner product term is negative, thereby reducing the total loss in this approximate sense.

[0066] At the end of the k-th training round, the system has a set of current model parameters (including the weights and biases of each layer). The system first uses these parameters to perform a complete calculation on the current training samples to obtain the predicted score, and then calculates the total loss value based on the predicted score and the score label. Subsequently, the system calculates the partial derivative of the total loss value with respect to each trainable parameter: a positive partial derivative indicates that if the parameter is slightly increased, the total loss will increase; a negative partial derivative indicates that if the parameter is slightly increased, the total loss will decrease; the larger the absolute value of the partial derivative, the more sensitive the parameter is to the total loss. Arranging the partial derivatives of all parameters in a preset order yields the gradient vector, which is equivalent to giving the direction in which changing the parameters at the current parameter position will result in the fastest increase in the total loss.

[0067] To reduce the total loss, the system doesn't adjust the parameters along the direction of fastest gradient, but rather in the opposite direction: each parameter is slightly modified along the negative gradient direction. If a parameter's partial derivative is positive, it's decreased; if it's negative, it's increased. The learning rate controls the magnitude of each adjustment: a larger learning rate results in faster adjustments but is more prone to oscillations or divergence; a smaller learning rate provides more stable adjustments but slower convergence. Through this reverse adjustment, the system obtains the parameter values ​​for the next iteration. The total loss is then recalculated using the new parameters, and the process is repeated. The total loss gradually decreases over multiple iterations until a stopping condition is met, resulting in a trained neural network model.

[0068] The training loop repeatedly executes the steps of calculating the total batch loss, obtaining the gradient vector, and updating the model parameters, stopping when a preset training completion condition is established. The training completion condition can be determined by validation set metrics or loss convergence status, such as when the total loss value drops below the maximum allowed value set in the database, or when the preset maximum number of training epochs is reached. In this case, iteration stops and the current parameters are fixed as the model parameters for training completion.

[0069] Once the neural network model is trained, it can be used for energy efficiency assessment of electrical equipment. Figure 2 A flowchart of an electrical equipment energy efficiency assessment method based on a neural network model constructed using a construction method is provided in this application embodiment. The specific process is as follows:

[0070] Taking the lighting of corridors and classrooms in a university teaching area as an example, the target area includes eight lighting circuits and their respective energy meters under the distribution boxes on two floors. After the model training is completed, the system enters the online energy efficiency assessment stage. The system first collects cumulative metering data and operating status data. The cumulative metering data is the cumulative reading reported by the energy meters of each circuit. The operating status data includes available operating information such as circuit switch status, dimming level, time period type (weekday / weekend / holiday) and building outdoor temperature. Due to network jitter and retransmission coverage issues at the engineering site, the system retains both the event occurrence time and arrival time for each cumulative metering record: the event occurrence time indicates the actual metering time corresponding to the reading, and the arrival time indicates the time when the reading is received and written by the gateway or platform. The system sorts the cumulative metering records using the event occurrence time as the primary key. If there are multiple records for the same event occurrence time, the record with the latest arrival time is selected as the valid record and a duplicate overwrite flag is generated. At the same time, the system calculates the interval between the event occurrence time and the arrival time. If the interval exceeds a preset threshold, a late arrival flag is generated, thus providing a basis for subsequent quality control. Finally, several cumulative metering sequences are formed in ascending order of event occurrence time.

[0071] After obtaining each cumulative metering sequence, the system performs metering trajectory break detection on each sequence to identify discontinuous cumulative trajectories caused by resets, callbacks, or sudden jumps. Specifically, the system reads two adjacent cumulative records sequentially according to the time of the event, calculates the difference in cumulative amount and the corresponding time interval, and determines whether there are backtracking anomalies (the difference is negative and exceeds the tolerance threshold), sudden jump anomalies (the average power converted from the difference exceeds the installed capacity limit or the historical reliable upper limit), and consistency anomalies (the differential increment deviates too much from the instantaneous metering intensity integral increment). The system merges adjacent intervals with consecutive anomalies to form a broken interval and labels the anomaly type for the broken interval. The system then segments the cumulative measurement sequence using the break interval as the boundary. It calculates the connection offset for each segment after the break interval to restore the continuity of the cumulative trajectory before and after the break interval. For each break interval, a candidate supplementary sequence is generated. For example, candidate supplementary sequence two is obtained based on the instantaneous measurement intensity integral, and candidate supplementary sequence three is inferred based on the statistical characteristics of the stable segments before and after the break interval. The system calculates the consistency residual rate of the two candidate supplementary sequences and compares it with the defined consistency residual rate. If the consistency residual rate is lower than a threshold, the candidate supplementary sequences are averaged and backfilled with the break interval. If it is not lower than the threshold, conservative estimation is used for backfilling, and a reduced-weight usage flag is generated, thus obtaining an updated cumulative measurement sequence that provides a continuous reading basis for window differencing and includes confidence information.

[0072] Based on the updated cumulative metering sequence, the system generates cumulative metering sample values ​​according to a preset evaluation window. For example, with a 15-minute evaluation window, the system reads the cumulative readings at the beginning and end of each window boundary and calculates the window increment as the energy used in the window. Simultaneously, the system aggregates the window increments of multiple cumulative metering sequences of the same type within the window to obtain the window-level master metering. The system then aligns and associates this window-level master metering with the corresponding operating status data within the window to form the evaluation window sample. Finally, the system inputs the evaluation window sample into a trained neural network model, and the neural network model outputs the energy efficiency score for that evaluation window as the energy efficiency evaluation result.

[0073] Figure 4 This diagram illustrates the comparison of mean squared errors (MSEs) of the models provided in this embodiment. The horizontal axis represents the model type, namely the traditional model (linear regression model in this embodiment) and the neural network model; the vertical axis represents the MSE value (the smaller the better), i.e., the mean squared error value. As can be seen from the bar height, the neural network model has a smaller mean squared error, indicating that its fit to the energy efficiency score / energy efficiency prediction result is closer to the label or true value, and it can better characterize the nonlinear relationship between input features and output. Figure 5 This is a schematic diagram of the model determination coefficient comparison provided in the embodiments of this application. The horizontal axis represents the model type, namely the traditional model (linear regression model in this embodiment) and the neural network model, respectively. The vertical axis represents R.2 The value (closer to 1 the better), also known as the coefficient of determination, measures the model's ability to explain changes in the target output and its goodness of fit. As can be seen from the bar height, the neural network model has a larger coefficient of determination, indicating that the neural network model can explain more output variance, fits the actual label / true value more fully, and is better able to capture complex nonlinear relationships and interactions under the same input features. This demonstrates its advantages in energy efficiency assessment tasks, such as stronger fitting ability and better predictive consistency.

[0074] Figure 6 This diagram illustrates the comparison of model prediction results provided in this application embodiment. The horizontal axis represents the true energy efficiency score, and the vertical axis represents the predicted energy efficiency score. Each scatter point in the diagram represents the correspondence between the true score and the predicted score of a sample in an evaluation window under the neural network model. The dashed line is the perfect fit line (the predicted value should fall near this line when it equals the true value). As can be seen from the distribution, the scatter points are clustered in a band along the perfect fit line, and the coefficient of determination given in the diagram is 0.851, indicating that the neural network model can capture the mapping relationship between input features and energy efficiency scores well, making the predicted score and the true score highly consistent.

[0075] Figure 3 This application provides a schematic diagram of a neural network model construction system for energy efficiency assessment of electrical equipment, which includes a metering sequence generation module, a data consistency verification module, an assessment result output module, and a database.

[0076] The database is used to store the parameters involved in the construction system of a neural network model for energy efficiency assessment of electrical equipment. The database is developed based on the parameter requirements of the entire process of energy efficiency assessment of electrical equipment. First, the core categories such as basic equipment attributes, operating conditions, and energy efficiency test data are sorted out. Then, the data format, accuracy standards and storage specifications of various parameters are clarified. Through structured modeling, a data classification and storage architecture is built, and finally a standardized database adapted to energy efficiency assessment calculation and analysis is formed.

[0077] The metering sequence generation module collects cumulative metering data and operating status data of electrical equipment in the target area. It sorts the cumulative metering data by the occurrence and arrival times of events to generate several cumulative metering sequences. The data consistency verification module performs metering trajectory break detection on each cumulative metering sequence, extracts the break intervals and anomaly types corresponding to each cumulative metering sequence, generates corresponding candidate supplementary sequences for different break intervals, and performs data consistency verification to update each cumulative metering sequence. The evaluation result output module generates cumulative metering sample values ​​according to a preset evaluation window based on the updated cumulative metering sequences, associates them with operating status data to form evaluation window samples, inputs the evaluation window samples into the trained neural network model, and outputs the energy efficiency evaluation results of electrical equipment in the target area.

[0078] In Example 1, the cumulative measurement sequence is segmented using the break interval as the boundary, and the continuity of the cumulative trajectory before and after the break interval is achieved by calculating the connection offset. Subsequently, candidate supplementary sequences are generated based solely on the linear difference, and a monotonic consistency check is performed before backfilling and updating. Although this scheme can restore the continuity of the cumulative trajectory in form and eliminate explicit regression, its supplementary basis mainly comes from numerical interpolation at both ends of the break interval, which is difficult to characterize the dynamic process of actual energy consumption changing with operating status within the break interval. When the break interval crosses scenarios such as load abrupt changes, switching strategy adjustments, or occupancy mode switching, the linear difference is prone to producing supplementary results with reasonable shapes but inaccurate energy values, which in turn causes systematic deviations in the window increment statistics within the break interval, resulting in uncertainty in the supervision signals of subsequent scoring label construction and model training samples. In addition, the monotonic consistency check focuses more on constraining the cumulative trajectory from decreasing, and cannot quantitatively compare the consistency between the supplementary results and observable instantaneous measurement intensity evidence, as well as the statistical regularities of the stable segments before and after the break interval. Therefore, supplementary segments that pass the monotonic check but lack credibility may still occur.

[0079] To address the shortcomings of Embodiment 1, Embodiment 2, based on the identification and segmented connection offset correction of the fracture interval from 20:15 to 20:30 in Embodiment 1, further generates candidate supplementary sequence 2 and candidate supplementary sequence 3 for this fracture interval.

[0080] Among them, the second candidate supplementary sequence is obtained by integrating the instantaneous metering intensity. The instantaneous metering intensity is the instantaneous power or equivalent load intensity obtained by the lighting circuit at a high sampling frequency. The system reads this intensity sequence from 20:15 to 20:30 and integrates it at time intervals to obtain the cumulative increment that should be added within the break interval. Then, starting from the cumulative value of 1250.49 at the front boundary of the break interval, it is gradually accumulated to form the second candidate supplementary sequence. For example, if the energy increments converted from the instantaneous metering intensity are 0.02, 0.03, and 0.01 respectively in 5-minute steps from 20:15 to 20:30, the corresponding supplementary cumulative readings of the second candidate supplementary sequence can be 1250.51 at 20:20, 1250.54 at 20:25, and 1250.55 at 20:30.

[0081] The third supplementary sequence is a candidate supplementary sequence inferred from the statistical characteristics of the segmented sequences before and after the break interval. The statistical characteristics refer to the extraction of representative quantities such as the mean, median, or quantile of the window increment or converted power within the stable segments before and after the break interval, which are used to characterize the typical energy consumption level of the circuit under the same time period and similar operating conditions. For example, the system statistically obtained a typical 15-minute increment of 0.06 for the corridor lighting in the adjacent windows of the first segment (19:00–20:15) and the second segment (20:30–21:00). Combined with the break interval length of 15 minutes, this typical increment is used as the inferred increment of the break interval, thus constructing the third candidate supplementary sequence: 1250.51 at 20:20, 1250.53 at 20:25, and 1250.55 at 20:30.

[0082] The system then calculates the consistency residual rate based on candidate supplementary sequence two and candidate supplementary sequence three. The consistency residual rate is used to quantify the relative deviation between the two candidate supplementary sequences, as follows:

[0083] ;

[0084] Where R represents the consistency residual rate, i is the index of the i-th completion time point within the broken interval, i=1,2,3,...,N, and N is the number of time points within the broken interval that need to be completed and compared. A2 i A3 represents the candidate supplementary value of candidate supplementary sequence two at time point i. i A represents the candidate supplementary value of candidate supplementary sequence three at time point i, A0 represents the candidate supplementary value of the starting boundary point of the fracture interval of candidate supplementary sequence two, and a is a preset minimum positive number in the database to avoid the denominator being 0 or too small.

[0085] The construction of this consistency residual rate follows the idea of ​​characterizing the consistency of multi-source supplementary evidence with relative deviation. Candidate supplementary sequence two and candidate supplementary sequence three are derived from different information sources (the former comes from the time integral of instantaneous measurement intensity, and the latter comes from the statistical characteristics inference of the stable segments before and after the fracture interval). The difference between the candidate supplementary values ​​at the same completion time point within the fracture interval directly reflects the degree of divergence between the two pieces of evidence at that moment. Therefore, the numerator uses the summation of the absolute values ​​of the differences at each completion time point to summarize the overall inconsistency magnitude within the fracture interval and avoids the mutual cancellation of positive and negative differences. On the other hand, the magnitude of the absolute difference within the fracture interval varies with the energy level, interval length, and load level. Using only the absolute difference is not conducive to the comparability between different fracture intervals. Therefore, the cumulative increment of the second candidate supplementary sequence relative to the starting boundary point is introduced into the denominator as a normalization scale to characterize the typical change amplitude within the fracture interval, thereby converting the inconsistency amplitude into a dimensionless relative deviation, so that the same threshold can be used for consistency determination in different scale scenarios. When the energy increment within the fracture interval is close to zero or extremely small, causing the denominator to be too small or even zero, a preset minimum positive number is introduced as a lower bound by taking the maximum value, which can avoid numerical instability and unreasonable amplification. Based on the above construction, the consistency residual rate can stably measure the relative consistency of two candidate supplementary sequences within the fracture interval without relying on additional priors: the smaller the residual rate, the closer the inferences of the cumulative trajectory by the two pieces of evidence are, and the more credible the supplementary result is; the larger the residual rate, the more obvious the conflict between the two pieces of evidence is. Based on this, the system can trigger conservative estimation or reduced weighting to reduce the risk of unreliable supplementation polluting subsequent scoring labels and model training.

[0086] After analyzing the consistency residual rate, the consistency residual rate is compared with the preset defined consistency residual rate in the database, which is the maximum allowed consistency residual rate stored in the database. If the consistency residual rate is lower than the defined consistency residual rate, it indicates that the difference between the supplementary trajectories given by the two candidate supplementary sequences in the break interval is within an acceptable range and has mutual verification significance. The fusion is carried out by averaging, that is, the arithmetic mean of the two candidate supplementary values ​​at the same supplementation time point is taken as the final supplementary value.

[0087] Conversely, if the consistency residual rate is not lower than the defined consistency residual rate, at least one candidate supplementary sequence is deemed unreliable. Unreliability means that the candidate supplementary sequence significantly conflicts with another source of evidence on a relative scale, and direct fusion and backfilling may introduce systematic overestimation or underestimation. In this case, the system no longer performs average fusion, but adopts a conservative estimation strategy, that is, it selects the supplementary path that is less likely to lead to overestimation as the backfilling result. For example, for each completion time point, it takes the smaller of the two candidate supplementary values ​​or takes a smoother incremental path to avoid abnormal integration or statistical drift from filling the broken interval into an abnormal surge. At the same time, the system generates a deweighted usage label for the broken interval, so that the model can identify the deweighted usage label when learning the data, reducing the learning of the data.

[0088] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for constructing a neural network model for energy efficiency evaluation of electrical equipment, characterized in that, Includes the following steps: Step 1: Obtain model training data. Extract the recording time points of several cumulative measurement data of the same type from the model training data. Sort and manage the cumulative measurement data in chronological order to obtain a cumulative measurement sequence arranged in chronological order. Step 2: Perform measurement trajectory break detection on the cumulative measurement sequence. After the detection is completed, extract each break interval and the anomaly type. Divide the cumulative measurement sequence into segments and correct the cumulative trajectory using each break interval as the boundary. Step 3: Generate candidate supplementary sequences for each fracture interval, perform data consistency verification on the candidate supplementary sequences, update the cumulative measurement sequence after verification, and associate scoring labels with the updated cumulative measurement sequence; Step 4: Introduce scoring labels into the total loss function, iteratively update the neural network model until the preset model training completion conditions are met, and obtain the neural network model for predicting the energy efficiency of electrical equipment. The specific process for obtaining the cumulative measurement sequence arranged chronologically is as follows: Extract the recording time points of several cumulative measurement data of the same type from the model training data. The recording time points refer to the time points when the event occurs and the time points when it arrives. Using the time of the event as the primary key, the cumulative measurement data are sorted to form a cumulative measurement sequence arranged in chronological order. When there are several cumulative measurement data at the same time point of occurrence, the cumulative measurement data with the latest arrival time is placed into the cumulative measurement sequence, and a duplicate overwrite mark is generated. Calculate the interval between the occurrence time and arrival time of each cumulative measurement data. If the interval is longer than the interval threshold, generate a late arrival flag. The aforementioned detection of measurement trajectory breaks in the cumulative measurement sequence specifically refers to: Read adjacent data in the cumulative measurement sequence in the order of the time of the event, and use them as the previous record and the next record respectively; Calculate the cumulative difference between the next record and the previous record, and calculate the interval between the two records; Based on the cumulative difference between the next record and the previous record and the interval between the two records, rollback detection, jump detection and consistency detection are performed in sequence. Obtain the detection results, extract each fracture interval from the detection results, and output the anomaly type of each fracture interval; The generation of candidate supplementary sequences for each fracture interval specifically refers to: The segmentation strategy is applied sequentially to the cumulative measurement series; The segmentation strategy refers to segmenting the cumulative measurement sequence with the break interval as the boundary, calculating the connection offset of the segmented sequence after the break interval, so that the cumulative trajectory of the segmented sequence after the break interval is continuously connected with the cumulative trajectory before the break interval. For any break interval, a candidate supplementary sequence 1 is generated based on the linear difference, and a data monotonic consistency check is performed on the candidate supplementary sequence 1. After the check is completed, the candidate supplementary sequence 1 is filled into the corresponding break interval, and the cumulative measurement sequence is updated. The total loss function is constructed as follows: Constructing window samples based on cumulative measurement sequences; Obtain the predicted score and calculate the residual between the predicted score and the score label; The score regression loss term is generated based on the residuals using a robust regression approach, where the score regression loss term employs a Hubell-type piecewise function. The abnormality types and abnormality labels in the cumulative measurement sequence are statistically identified. The corresponding label loss values ​​are matched from the database and accumulated. The accumulated result is labeled as the label loss term. The abnormal markers refer to duplicate overwrite markers and late markers; The weighted summation of the score regression loss term and the label loss term, and the summation result is denoted as the total loss function; The parameters of the neural network model are updated iteratively based on the total loss function.

2. The method for constructing a neural network model for energy efficiency assessment of electrical equipment as described in claim 1, characterized in that: The process of generating candidate supplementary sequences for each fracture interval also includes: For any fracture interval, generate candidate supplementary sequence two and candidate supplementary sequence three; The second candidate supplementary sequence refers to the candidate supplementary sequence obtained by integrating the instantaneous measurement intensity. The third candidate supplementary sequence refers to the candidate supplementary sequence inferred from the statistical characteristics of the segmented sequences before and after the break interval; Based on candidate supplementary sequence two and candidate supplementary sequence three, the consistency residual rate of the candidate supplementary sequences was analyzed. If the consistency residual rate of the candidate supplementary sequence is lower than the defined consistency residual rate, the fusion is performed by averaging. After fusion, the sequence is filled into the corresponding break interval and the cumulative econometric sequence is updated. If the consistency residual rate of the candidate supplementary sequence is not lower than the defined consistency residual rate, it indicates that at least one candidate supplementary sequence is unreliable. In this case, a conservative estimation strategy is adopted and the break interval is used with reduced weight.

3. The method for constructing a neural network model for energy efficiency evaluation of electrical equipment as described in claim 2, characterized in that: The above refers to the updated cumulative quantitative sequence associated scoring label, and the specific association process is as follows: The model training data is grouped based on a preset working condition classification rule; Within each group, a statistical distribution is formed using all cumulative measurement sample sequences, and efficient and inefficient reference lines are extracted: The high-efficiency reference line refers to the low quantile value that defines the statistical distribution of the cumulative measurement sample sequence. The inefficient reference line refers to the high quantile value that defines the statistical distribution of cumulative measurement sample sequences; Based on efficient and inefficient reference lines, cumulative measurement sequences are mapped to rating labels.

4. The method for constructing a neural network model for energy efficiency assessment of electrical equipment as described in claim 1, characterized in that: The iterative update of the neural network model parameters is specifically performed as follows: In each round of training, the total loss value is calculated based on the current training samples; By calculating the partial derivatives of the total loss function with respect to each parameter of the neural network model, the gradient vectors corresponding to the parameters of each neural network model are obtained. The parameters are updated in the opposite direction of the gradient vector, and the update step size is determined by the learning rate. Repeat the process of calculating the total loss value, obtaining the gradient vector, and updating the model parameters until the preset model training completion conditions are met, and obtain a neural network model for predicting the energy efficiency of electrical equipment.

5. A method for evaluating the energy efficiency of electrical equipment based on a neural network model constructed using the construction method described in any one of claims 1-4, characterized in that, Includes the following steps: Collect cumulative metering data and operating status data of electrical equipment in the target area, sort the cumulative metering data by the time of occurrence and arrival of events, and generate several cumulative metering sequences; Perform measurement trajectory break detection on each cumulative measurement sequence, extract each break interval and anomaly type corresponding to each cumulative measurement sequence, generate corresponding candidate supplementary sequences for different break intervals and perform data consistency verification, thereby updating each cumulative measurement sequence; Based on the updated cumulative metering sequence, cumulative metering sample values ​​are generated according to the preset evaluation window, and associated with the operating status data to form evaluation window samples. The evaluation window samples are then input into the trained neural network model to output the energy efficiency evaluation results of electrical equipment in the target area.

6. An electrical equipment energy efficiency evaluation system based on a neural network model constructed using the construction method described in any one of claims 1-4, characterized in that, include: The module includes a measurement sequence generation module, a data consistency verification module, and an evaluation result output module. The metering sequence generation module is used to collect cumulative metering data and operating status data of electrical equipment in the target area, sort the cumulative metering data by the event occurrence time and arrival time, and generate several cumulative metering sequences. The data consistency verification module is used to perform measurement trajectory break detection on each cumulative measurement sequence, extract each break interval and anomaly type corresponding to each cumulative measurement sequence, generate corresponding candidate supplementary sequences for different break intervals and perform data consistency verification, thereby updating each cumulative measurement sequence. The evaluation result output module is used to generate cumulative metering sample values ​​according to a preset evaluation window based on the updated cumulative metering sequence, and associate them with the operating status data to form evaluation window samples. The evaluation window samples are then input into the trained neural network model to output the energy efficiency evaluation results of the electrical equipment in the target area.

Images