Multi-level energy abnormity early warning method and system and storage medium

Abstract

The invention relates to the field of digital energy management and control, in particular to a multi-level energy abnormity early warning method and system and a storage medium, and the method comprises the steps: obtaining multi-source data of an energy system; performing multi-dimensional feature extraction on the multi-source data and obtaining fusion features; based on the fusion features, constructing a multi-scale sliding window to calculate a dynamic threshold; performing multi-dimensional anomaly judgment on the current energy consumption data according to the dynamic threshold value to obtain a comprehensive anomaly score and a persistent score; and performing multi-level early warning aggregation based on the comprehensive abnormal score and the persistence score to obtain an early warning scheme. According to the embodiment of the invention, by constructing the dynamic adaptive threshold system, the multi-dimensional anomaly judgment mechanism and the multi-level early warning aggregation framework, accurate anomaly detection, timely early warning response and global energy efficiency management of the industrial energy system are realized.

Description

Technical Field

[0001] This invention relates to the field of digital energy management, specifically to a multi-level energy anomaly early warning method, system, and storage medium. Background Technology

[0002] In industrial production, efficient management of energy systems relies on real-time and accurate anomaly detection of massive amounts of multivariate time-series data. While existing technologies have made some progress, limitations remain in practical applications. Existing patent CN202411688863 proposes an unsupervised detection method for edge devices, which reduces computational cost while maintaining accuracy through data similarity input and a lightweight structure. However, this method's strategy is relatively fixed, making it difficult to adapt to dynamic characteristics such as production load fluctuations and seasonal changes. Furthermore, it lacks differentiated analysis of the importance and stability of different indicators, limiting its effectiveness in multi-dimensional energy systems. Another patent, CN202511304396, dynamically determines the mask ratio by analyzing data correlation and stability, solving the problems caused by fixed masks. However, this method is limited to single-time-window detection, unable to effectively capture complex correlation anomalies across windows, multiple devices, and production lines, and lacks the ability to predict future trends, thus hindering preventative management.

[0003] The following technical problems exist in existing industrial energy management systems: First, low anomaly detection accuracy and high false alarm / missed alarm rates. Existing technologies use fixed thresholds or simple statistical methods for anomaly judgment, failing to dynamically adjust detection strategies based on real-time operating condition changes, production load fluctuations, seasonal patterns, and other factors. When production load suddenly increases, normal energy consumption surges are mistakenly identified as anomalies; conversely, when equipment gradually deteriorates leading to a slow decline in energy efficiency, anomalies are difficult to detect in a timely manner, resulting in persistently high false alarm and missed alarm rates. Second, limited early warning levels and lack of a global perspective. Traditional systems can only monitor anomalies for individual devices, unable to aggregate and correlate anomalies across multiple levels such as production lines, workshops, and factory areas. This affects the efficiency of root cause localization and fails to provide comprehensive guidance for enterprise energy management decisions. When multiple devices simultaneously exhibit minor anomalies, single-device early warning mechanisms struggle to identify overall energy waste. Third, insufficient predictive capabilities and a passive response approach. Existing methods lack predictive models based on historical data and time-series patterns, failing to anticipate energy consumption anomaly trends and only responding passively after anomalies occur, missing optimal handling opportunities and increasing energy waste and production losses.

[0004] In summary, existing industrial time-series data anomaly detection technologies still face numerous challenges in practical applications. These challenges primarily manifest as insufficient dynamic adaptability, making it difficult to effectively cope with complex and ever-changing operating conditions. Furthermore, the lack of multi-level global correlation analysis capabilities results in a missing comprehensive anomaly diagnosis mechanism covering equipment, production lines, and workshop dimensions. In addition, due to insufficient predictive capabilities, existing technologies cannot anticipate anomaly development trends to aid preventative decision-making. Moreover, the balance between accuracy, real-time performance, and resource consumption when processing large-scale data requires further optimization. These issues collectively limit the level of intelligent management in industrial energy systems. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-level energy anomaly early warning method, system, and storage medium. By constructing a dynamic adaptive threshold system, a multi-dimensional anomaly judgment mechanism, and a multi-level early warning aggregation framework, it enables accurate anomaly detection, timely early warning response, and global energy efficiency management of industrial energy systems, thereby solving the technical problems of low anomaly detection accuracy, single early warning level, and insufficient predictive ability in the prior art.

[0006] To achieve the above objectives, embodiments of the present invention provide a multi-level energy anomaly early warning method, comprising: Acquire multi-source data from the energy system; Multidimensional feature extraction is performed on the multi-source data to obtain fused features; Based on the fusion features, a multi-scale sliding window is constructed to calculate the dynamic threshold; The current energy consumption data is subjected to multi-dimensional anomaly determination based on the dynamic threshold to obtain a comprehensive anomaly score and a persistence score. Multi-level early warning aggregation is performed based on the comprehensive anomaly score and the persistence score to obtain an early warning scheme.

[0007] Optionally, acquiring multi-source data from the energy system includes: The system acquires real-time operating data, historical forecast data, baseline energy consumption data, and operating condition parameter information of the energy system. The historical forecast data is the system energy consumption forecast data based on historical data using an LSTM time series forecast model. The baseline energy consumption data is obtained by hierarchically statistically analyzing historical data according to time period, load, season, and operating condition.

[0008] Optionally, performing multidimensional feature extraction and obtaining fused features from the multi-source data includes: Statistical features, time-series features, load features, and comparative features of the multi-source data are extracted, wherein the time-series features are used to extract the first-order autocorrelation coefficient, the load features are used to extract the correlation coefficient between energy consumption and load, and the comparative features are used to extract the year-on-year change rate. Dynamic weighting is applied to real-time energy consumption data, energy consumption prediction data, and baseline energy consumption data to obtain fusion characteristics.

[0009] Optionally, based on the fusion features, constructing a multi-scale sliding window to calculate the dynamic threshold includes: Calculate the statistics for short window, medium window, and long window respectively; Based on the short window statistics, medium window statistics, and long window statistics, multi-scale weighted statistics are obtained; Based on the short window statistics, medium window statistics, and long window statistics, adaptive coefficients are obtained; Based on the multi-scale weighted statistics and adaptive coefficients, dynamic upper and lower thresholds are obtained.

[0010] Optionally, obtaining multi-scale weighted statistics based on the short-window, medium-window, and long-window statistics includes: Obtain the multiscale weighted statistics according to formulas (1) and (2): (1) (2) in, The average value is the comprehensive average. For the composite standard deviation, This is the short window mean. The mean of the middle window. The average value over a long window. , and These are the weighting coefficients. For short window standard deviation, The standard deviation of the middle window. The standard deviation is for the long window.

[0011] Optionally, obtaining the adaptive coefficients based on the short window statistics, medium window statistics, and long window statistics includes: Calculate the adaptive coefficients according to formulas (3) to (7). (3) (4) (5) (6) (7) in, For adaptive coefficients, Based on the coefficient, This is the volatility adjustment factor. This is a trend adjustment coefficient. This is a seasonal adjustment factor. This is the load-related adjustment factor. The coefficient of variation for the short window. For the medium window trend, For seasonal fluctuations, For standard time functions, For phase deviation, This represents the real-time load rate.

[0012] Optionally, multi-dimensional anomaly determination is performed on the current energy consumption data based on the dynamic threshold to obtain a comprehensive anomaly score and a persistence score, including: Scores were calculated for five dimensions, including absolute threshold deviation index, prediction deviation rate, baseline deviation rate, confidence interval deviation index, and time change rate. Calculate the comprehensive anomaly score according to formula (8). (8) in, For comprehensive anomaly scoring, For the first Abnormal indicators in each dimension For the first Scoring across multiple dimensions For the first Weighting coefficients for each dimension of the score; Calculate the persistence score according to formula (9). (9) in, For continuous scoring, Indicates a continuous evaluation window. For indicator functions, For a historic moment The overall abnormality score.

[0013] Optionally, a multi-level early warning aggregation is performed based on the comprehensive anomaly score and the persistence score to obtain an early warning scheme, including: Perform device-level early warning judgment, including: triggering an early warning when the overall anomaly score of the current device reaches a threshold; Perform production line-level aggregation analysis, including: calculating the proportion of abnormal equipment on the production line, the total abnormal score of the production line, and the total energy consumption deviation of the production line respectively; and triggering an early warning if at least one of the proportion of abnormal equipment on the production line, the total abnormal score of the production line, and the total energy consumption deviation of the production line reaches a threshold. Perform workshop-level aggregate analysis, including: calculating the proportion of abnormal production lines and the overall abnormality of the workshop; Perform plant-level aggregate analysis, including: calculating multi-level anomalies and overall plant performance; Based on the equipment-level early warning judgment, production line-level aggregate analysis, workshop-level aggregate analysis, and factory-level aggregate analysis, a multi-level early warning output scheme is generated.

[0014] On the other hand, the present invention also provides a multi-level energy anomaly early warning system, the system comprising: The data acquisition module is used to acquire multi-source data from the energy system; The data processing module is used to preprocess and extract features from the multi-source data to generate fused data; The dynamic anomaly detection module is used to construct a multi-scale sliding window on the fused data to generate a dynamic threshold and obtain a comprehensive anomaly score and a persistence score. The hierarchical early warning module is used to aggregate and analyze the comprehensive anomaly score and the persistence score layer by layer according to the equipment level, production line level, workshop level and factory area level, and generate a multi-level early warning output scheme. A processor is used to connect the data acquisition module, the data processing module, the dynamic anomaly determination module, and the hierarchical early warning module, and the processor is configured to execute any of the methods described above.

[0015] In another aspect, the present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement any of the methods described above.

[0016] Compared with the prior art, the present invention has the following advantages: (1) Significantly improves anomaly detection accuracy. Through a three-layer sliding window system and adaptive adjustment coefficient, intelligent calculation of dynamic thresholds is achieved, overcoming the shortcomings of traditional fixed thresholds that cannot adapt to changes in working conditions. Combined with a five-dimensional anomaly judgment system, the anomaly detection recall rate is ≥95% and the precision rate is ≥90% based on five dimensions: absolute threshold, prediction deviation, baseline deviation, confidence interval, and rate of change. Compared with traditional methods, the false alarm rate is reduced by more than 40%, and the false negative rate is reduced by more than 35%.

[0017] (2) Achieve multi-level global early warning. Construct a four-level hierarchical early warning system at the equipment level, production line level, workshop level, and plant level. Through a three-dimensional judgment mechanism of abnormal equipment ratio, total abnormal score, and total energy consumption deviation rate, achieve comprehensive monitoring from micro to macro. It can accurately locate single equipment problems and identify overall energy waste, providing hierarchical decision support for managers, with an early warning response time of ≤5 minutes.

[0018] (3) Enhance system adaptability. Generate prediction curves for the next 24 hours using the LSTM time series prediction model to achieve predictive anomaly early warning, predict abnormal energy consumption trends in advance, and transform from passive response to proactive prevention. The gradient descent method is used to continuously optimize the adaptive coefficients, and the baseline model is updated daily and monthly to enable the system to gradually adapt to changes in production mode and maintain stable accuracy in long-term operation.

[0019] (4) Provide comprehensive performance evaluation. Quantitatively evaluate the anomaly detection effect through multi-dimensional performance indicators such as precision, recall, and F1 score. Combine comprehensive energy efficiency calculation and energy-saving potential quantitative evaluation of the plant area to help enterprises identify the main sources of energy waste, quantify energy-saving improvement space, and achieve refined energy management.

[0020] (5) Highly efficient and reliable technology. Real-time data preprocessing is performed using edge computing nodes, reducing the computing pressure on the central server and reducing data processing latency to <3 seconds. A multi-terminal visualization output mechanism synchronously pushes the evaluation results to the energy management platform, production monitoring center, and enterprise energy management system, supporting multi-role collaborative decision-making. The system supports 100% device coverage, and early warning signals are pushed to the corresponding responsible persons in a tiered manner to avoid information overload or missed reports.

[0021] (6) Significant economic benefits. Through precise anomaly detection and multi-level early warning, it helps enterprises to promptly identify and handle energy anomalies, avoiding energy waste and production losses. In practical applications, it can improve enterprises' energy utilization efficiency by 5-15%, reduce annual energy costs by -10%, and reduce equipment failure rates by more than 25%, resulting in good economic and social benefits.

[0022] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0023] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 A flowchart of a multi-level energy anomaly early warning method according to an embodiment of the present invention; Figure 2 A flowchart illustrating a method for constructing a multi-scale sliding window to calculate a dynamic threshold according to an embodiment of the present invention; Figure 3 This is an architecture diagram of a multi-level energy anomaly early warning method according to an embodiment of the present invention. Detailed Implementation

[0024] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0025] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with relevant laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0026] like Figure 1 The diagram shows a flowchart of a multi-level energy anomaly early warning method according to an embodiment of the present invention. Figure 1 In this context, the early warning method may include the following steps: In step S10, multi-source data of the energy system is acquired; In step S11, multidimensional feature extraction is performed on the multi-source data and fusion features are obtained; In step S12, a multi-scale sliding window is constructed to calculate the dynamic threshold based on the fusion features; In step S13, the current energy consumption data is subjected to multi-dimensional anomaly determination based on the dynamic threshold to obtain a comprehensive anomaly score and a persistence score. In step S14, multi-level early warning aggregation is performed based on the comprehensive anomaly score and the persistence score to obtain an early warning scheme.

[0027] In such Figure 1 In the multi-level energy anomaly early warning method shown, step S10 is used to acquire multi-source data of the energy system, including real-time operating data, historical prediction data, benchmark energy consumption data, and operating condition parameter information. In this embodiment, the specific method for acquiring multi-source data of the energy system in step S10 can be of various forms known to those skilled in the art. In one example of this invention, step S10 can be achieved through a three-level architecture of "distributed intelligent metering terminal - multi-protocol energy data acquisition - edge node real-time processing," realizing real-time and accurate acquisition and standard coal equivalent conversion of various energy media such as electricity, water, natural gas, and compressed air in the industrial energy system. This covers four core data dimensions: real-time operating data, historical prediction data, benchmark energy consumption data, and operating condition parameter information, providing a complete data foundation for subsequent multi-scale anomaly detection and dynamic threshold calculation. The data acquisition frequency is uniformly set to 60 seconds to ensure data timeliness and consistency.

[0028] Specifically, in this example, the acquisition of electrical energy data is centered on four-dimensional monitoring of "power-current-voltage-cumulative energy". Smart meters and current transformers are installed in the power distribution circuit of each monitoring device to collect three-phase power P (kW), three-phase current I (A), three-phase voltage U (V), and cumulative energy E (kWh) in real time. The acquisition frequency is adjustable from 1 to 60 seconds per acquisition. The acquisition devices are multi-functional smart meters (model: DTZY341-Z) and Hall effect current transformers (model: ACS758). The meters are deployed at the inlet of the power distribution box of each monitoring device, and the transformers are installed through the three-phase cables. The meters collect instantaneous power, current, and voltage values ​​once per second and accumulate energy data once per minute. When the power fluctuation exceeds 20% of the rated value, a 0.1-second high-frequency acquisition mode is automatically triggered. Real-time energy consumption is obtained through integral calculation, using the following formula: (10) in, This indicates the sampling interval in minutes. If the voltage deviation exceeds ±10% of the rated value (e.g., 418V measured when the rated value is 380V), a voltage anomaly flag is triggered and pushed to the monitoring terminal simultaneously to avoid voltage fluctuations causing distortion of energy consumption data.

[0029] In this example, the acquisition of tap water consumption data is based on three-dimensional monitoring of "flow rate-pressure-temperature". Electromagnetic flow meters and pressure sensors are installed on the main industrial water supply line and the branch lines of each production process water supply line to collect instantaneous flow rate in real time. Pipeline pressure Water temperature The data acquisition frequency is set to 60 seconds / time. The acquisition devices are an electromagnetic flowmeter (model: LDCK-80) and a pressure transmitter (model: MPM480). The flowmeter is installed on the pipeline via a flange connection, and the pressure transmitter is threaded and installed 50cm downstream of the flowmeter. The flowmeter outputs instantaneous and cumulative flow values ​​every 5 seconds, and the pressure transmitter synchronously acquires the pipeline pressure. When the flow rate suddenly changes by more than 20% (e.g., from 10 m³ / h to 7 m³ / h) and the pressure drops by more than 0.1 MPa, it is considered a "pipeline leakage warning" and an anomaly flag is triggered. Water energy consumption calculations based on standard coal equivalent must consider water temperature. When the water temperature is below 5℃ or above 60℃, water energy consumption is calculated as "basic water consumption × temperature correction factor," using the following formula: (11) in, Instantaneous flow rate ( ), integral term Indicates the sampling interval Basic water consumption within the facility; The temperature correction factor is obtained by looking up a table (e.g., 1.15 for 5℃ and 0.92 for 60℃).

[0030] In this example, natural gas data acquisition focuses on three-dimensional monitoring of "flow rate-pressure-temperature". Gas turbine flow meters and temperature and pressure compensators are installed in the natural gas intake main and each combustion equipment branch to collect instantaneous flow rate under operating conditions in real time. Pipeline pressure Gas temperature The data acquisition frequency was set to 15 seconds per acquisition. The acquisition devices were a gas turbine flow meter (model: LWQ-100) and an intelligent flow totalizer (model: XSR-21R). The flow meter was installed on a DN100 pipeline using a flange connection. The totalizer has a built-in temperature and pressure compensation algorithm to convert the operating flow rate into the instantaneous flow rate under standard conditions (20℃, 101.325kPa) in real time. The formula for calculating real-time energy consumption (standard volume) of natural gas is: (12) in, Indicates the sampling interval Standard volume consumption of natural gas within (hours) ); The local atmospheric pressure is taken as 101.325 kPa. Standard atmospheric pressure; The standard temperature is 293.15K; this calculated value is directly used in subsequent calculations for standard coal equivalent. Variable input.

[0031] In this example, compressed air data acquisition is centered on three-dimensional monitoring of "flow rate-pressure-dew point". Thermal mass flow meters and dew point meters are installed on the main pipe of the air compressor station and on the branch lines of each air-consuming equipment to collect instantaneous flow rate data in real time. Pipeline pressure Dew point temperature The sampling frequency is set to 5 seconds / time. The sampling devices are a thermal mass flow meter (model: MT100) and a dew point transmitter (model: DMT340). The flow meter is inserted into the pipeline, and the dew point transmitter is installed 1m downstream of the flow meter. The flow meter outputs instantaneous flow rate once every 5 seconds (automatically compensated for temperature and pressure to standard conditions of 0℃ and 0.101325MPa). The dew point transmitter synchronously collects the dew point temperature. When the dew point is higher than -20℃, a warning for excessive moisture content in compressed air is triggered (standard requirement ≤-20℃), which affects the stability of the air-using equipment. The compressed air energy consumption calculation needs to consider the compression ratio and motor efficiency. The calculation formula is: (13) in, The compression ratio is 1.4 (typical value). If the pressure is lower than 80% of the rated value (e.g., 0.5 MPa when the rated value is 0.7 MPa), it is determined to be insufficient pressure, and the air compressor operating data needs to be correlated to analyze the leak point.

[0032] The unified conversion of multi-energy data to standard coal equivalent adopts a mechanism of "conversion by energy type + weighted aggregation" to achieve comparability assessment of different energy media. The standard coal equivalent coefficient for electricity is set at 0.1229 kgce / kWh (national standard GB / T2589-2020), hydropower is converted according to the "energy consumption for tap water preparation" coefficient of 0.0857 kgce / t, natural gas is converted according to real-time calorific value (typical coefficient 1.33 kgce / Nm³, corresponding to 38 MJ / Nm³ calorific value), and compressed air is converted according to the "unit energy consumption of air compressor" coefficient of 0.132 kgce / Nm³ (corresponding to 0.7 MPa pressure level).

[0033] The real-time standard coal equivalent calculation formula is as follows: ; in Electricity (kWh) Water volume (t) Gas volume (Nm³) Compressed air volume (Nm³) The data is presented in real-time calorific value (MJ / Nm³). The cumulative values ​​of each energy type are summarized every 15 minutes, and the total equivalent standard coal amount is calculated. When a sensor malfunctions and data is missing for a certain energy type, the missing data is supplemented by "historical average value × current operating condition correction coefficient". The correction coefficient is calculated by associating with the load rate (e.g., if the current load rate is 0.8 and the historical average is 0.7, the correction coefficient = 0.8 / 0.7 ≈ 1.14). This avoids the impact of missing data for a single energy source on the accuracy of total energy consumption statistics.

[0034] For acquiring historical prediction data, an LSTM time-series prediction model is used, employing a three-level network structure: "multi-step temporal feature extraction in the input layer - long short-term memory learning in the hidden layer - prediction value generation in the output layer." The computation process of the LSTM model can be represented as follows: (14) (15) (16) (17) in, This is the predicted energy consumption value at the current moment. Indicates the Gate of Oblivion; Indicates the input gate; Indicates the cell state; Indicates a hidden state. For the current time step, control the memory unit of the previous time step. How much information is retained; Forget gate weight matrix This is the hidden state vector from the previous time step; The current input vector (which is the preprocessed fused data, containing features such as energy consumption and load factor at the current moment) includes: , Parameters such as ambient temperature and time (system clock, calendar). This is the forget gate bias vector; The current time-to-memory unit vector, The candidate memory unit vector, calculated using the tanh function, represents the new memory that may be formed from the current input.

[0035] In addition, the prediction confidence interval needs to be calculated. The 95% confidence interval is calculated based on the prediction standard error: (18) (19) (20) in, For a historic moment, For the current moment, This represents the historical actual energy consumption value. Historical predicted energy consumption values.

[0036] For obtaining baseline energy consumption data, a "multi-dimensional hierarchical statistics + dynamic correction" mechanism is adopted, which statistically analyzes historical data in multiple dimensions such as time period, load, season, and operating condition. ,(twenty one) in, Indicates hours (0-23); Indicates the load factor (0-1); Indicates the season; Indicates ambient temperature; Indicates date type; This represents the temperature correction factor; This indicates a correction factor for special dates.

[0037] For operating parameter data, including production parameters, environmental parameters, time parameters, and equipment status parameters, the load factor calculation formula is: ,(twenty two) in, Indicates actual power; Indicates the rated power.

[0038] The acquisition of operating parameter data is based on a four-dimensional correlation of "production parameters - environmental parameters - time parameters - equipment status". Through data interaction with MES (Manufacturing Execution System), EMS (Energy Management System), and DCS (Distributed Control System), real-time production load, environmental conditions, time characteristics, and equipment operating status are obtained. Production parameters include actual power. With rated power Through the load factor calculation formula Quantify production intensity as operating condition parameter data.

[0039] After acquiring multi-source data from the energy system, the data needs to be preprocessed. Then, step S11 involves multi-dimensional feature extraction and fusion feature acquisition from the preprocessed multi-source data. In this embodiment, the specific method for preprocessing the multi-source data can be of various forms known to those skilled in the art. In one example of this invention, the preprocessing step can involve constructing a distributed data cleaning system using an IoT sensor network and edge computing nodes. This includes outlier detection and removal, intelligent completion of missing values, data denoising and smoothing, and sliding window standardization. This process eliminates data noise, fills in missing data, further extracts key features, and fuses multi-source information to generate a high-quality standardized energy consumption dataset. This provides reliable input for subsequent multi-scale sliding window anomaly detection, with the overall data preprocessing latency controlled within 2 seconds.

[0040] Specifically, in this example, outlier detection and removal employ... The criteria are used for outlier detection: ,(twenty three) ,(twenty four) in, This represents the mean value of the data within the detection window; Indicates the first in the detection window Each energy consumption data point ; It represents the standard deviation.

[0041] Outlier removal uses box plots for auxiliary identification, calculating the quartiles (Q1, Q3) and interquartile range (IQR = Q3 - Q1) of the data, and removing outliers less than Q1 - 1.5. IQR or greater than Q3 + 1.5 IQR data points are identified as outliers. This method complements the 3σ criterion to identify outliers that are not normally distributed. If the proportion of outliers is less than 5%, they are directly deleted; if the proportion of outliers is high, median replacement or linear interpolation based on preceding and following data is used to fill in the gaps.

[0042] In this example, intelligent missing value completion uses a hierarchical mechanism of "short-term missing linear interpolation + long-term missing ARIMA model prediction". When the missing duration is ≤5 minutes (≤5 data points), linear interpolation is used. (25) in, Indicates the last valid moment before the start of the missing period; This represents the first valid moment after the end of the missing period. For long-term missing data, an ARIMA model is used for prediction and imputation. Model parameters are automatically selected based on the AIC criterion. For example, if the historical data of a production line exhibits periodic fluctuations (daily cycle), an ARIMA(2,1,2) model (p=2nd order autoregression, d=1st order differencing, q=2nd order moving average) is selected. The model is trained based on the data from the 30 minutes prior to the missing period to predict the missing value. For instance, if a device has a 10-minute data gap from 10:10 to 10:20 (sensor disconnection), an ARIMA model is trained based on the data from 10:00 to 10:10, predicting energy consumption values ​​of 58.5-61.8 kgce / h for each moment from 10:10 to 10:20. After imputation, the data smoothly connects with the actual recovered value of 62 kgce / h at 10:20, avoiding data gaps. The imputed data is labeled "predicted value" to facilitate differentiation between measured and predicted data in subsequent analysis. This process aims to obtain complete time-series data without missing data. .

[0043] In this example, data denoising and smoothing employs the Exponentially Weighted Moving Average (EWMA) algorithm, which smooths current noise by weighting historical data for the initial time period. Set the initial smoothing value to equal the first measured value, i.e. ;for The time is calculated using the following formula: (26) Among them, the smoothing coefficient . for Energy consumption value after time smoothing for Energy consumption value after filling .

[0044] Sliding window standardization employs the sliding window Z-Score standardization method to eliminate differences in energy consumption levels across different time periods and devices. The sliding window statistic is calculated based on data smoothed by EWMA. (27) (28) (29) in, For the first in the window Energy consumption value after EWMA smoothing at each moment; This indicates the window size, typically ranging from 60 to 1440 minutes. This is the standardized energy consumption value.

[0045] After preprocessing the multi-source data, multi-dimensional feature extraction is performed in step S11. In this example, multi-dimensional feature extraction adopts a four-dimensional parallel extraction mechanism of "statistical features - time-series features - load features - comparative features" to mine key features for anomaly detection from the original energy consumption data. The core purpose of multi-dimensional feature extraction is to provide accurate guidance for the subsequent five-dimensional anomaly judgment system and multi-level early warning aggregation, thereby solving the three major pain points of the background technology: high false positives and false negatives, lack of global perspective, and only passive response. In adaptive optimization, the system needs to evaluate the judgment performance. Multi-dimensional features (such as prediction bias rate and baseline bias rate) are themselves inputs for quantifying the anomaly dimension. By analyzing which features contribute significantly to false positives / false negatives, the system can back-optimize the feature extraction window or threshold calculation model to achieve closed-loop evolution.

[0046] Specifically, in this example, the statistical characteristic includes the mean. ,variance Standard deviation skewness (Measures distribution symmetry; positive skewness indicates that high values ​​have longer tails), kurtosis (To measure the kurtosis of the distribution, positive kurtosis indicates that the distribution is sharper than a normal distribution), statistical characteristics of the past hour are calculated using a sliding window.

[0047] In this example, the first-order autocorrelation coefficient is extracted from the time-series features. To measure the correlation between data and its own data lagged by one period: (30) in, For the temporary pair window [1, n] containing the first... Standardized energy consumption value at a given time Calculate its mean. 'n' represents the window size, the window length used to calculate the autocorrelation coefficient. By extracting the autocorrelation coefficient, it's possible to determine whether the anomaly is a sudden shock or a continuation of a trend.

[0048] In this example, the correlation coefficient between load feature extraction energy consumption and load is used. Quantifying the synchronicity of energy consumption changes with production load: (31) By extracting the load correlation coefficient, it is immediately possible to determine whether high energy consumption is due to production saturation (high load rate) or equipment idling (low load rate).

[0049] In this example, the year-on-year change rate of the comparative feature extraction is compared. Capture periodic anomalies: (32) in, This represents the energy consumption value 24 hours ago (at the same time the previous day). By extracting the year-on-year change rate, it's possible to quickly determine whether the problem is a sudden occurrence today or a historical accumulation.

[0050] The multi-source data weighted fusion adopts a "dynamic weight allocation + confidence correction" mechanism, integrating real-time data, predicted data, and benchmark data to generate a comprehensive energy consumption reference value. The fusion formula is as follows: (33) in, As a feature of fusion, This is the real-time energy consumption value. This is the predicted energy consumption value. As the baseline energy consumption value, , and These are the weighting coefficients. From formula (17), It is obtained from formula (21); .

[0051] In this example, for , , The value can be: real-time data weight The typical value is 0.5. It remains unchanged when the real-time data quality is marked as Good (data meets physical constraints and statistical characteristics are within the normal range), and remains unchanged when it is Uncertain (continuous...). The values ​​are exactly the same in each sampling period. or short-time coefficient of variation When the value is reduced by 20% (e.g., from 0.5 to 0.4), it is considered a Bad ( This violates common sense in physics, such as negative energy consumption or Severe over-range or data timestamp delay The weight of the forecast data is adjusted to 0 (fully dependent on the forecast and the baseline). Based on dynamic adjustment of prediction accuracy, the prediction error over the past 24 hours is calculated. (Mean absolute error), when hour (Prediction reliable), when hour, ,when hour (Reducing weight due to large prediction bias). Baseline data weight. Based on dynamic adjustment of operating condition similarity, the current operating condition parameters (load rate) are calculated. ,temperature Euclidean distance from historical benchmark conditions ,when (When operating conditions are highly similar) ,when (When operating conditions vary greatly) For example: Real-time data at a certain moment: 65 kgce / h (quality Good). ), Forecast data 68 kgce / h (recent) ), baseline data 62 kgce / h (working distance) After normalizing the weights) , , fusion value The data fusion approach balances real-time performance (real-time data has the highest weight), predictability (predictive data provides trends), and rationality (benchmark data constrains anomalies), providing a robust reference for anomaly detection.

[0052] Step S12 is used to construct a multi-scale sliding window to calculate the dynamic threshold based on the fusion features. In this embodiment, the specific method for constructing the multi-scale sliding window to calculate the dynamic threshold in step S12 can be of various forms known to those skilled in the art. In one example of the present invention, step S12 may include, for example... Figure 2 The steps shown are described. Figure 2 In this context, step S12 may include: In step S20, the short window statistic, medium window statistic, and long window statistic are calculated respectively; In step S21, a multi-scale weighted statistic is obtained based on the short window statistic, medium window statistic, and long window statistic. In step S22, adaptive coefficients are obtained based on the short window statistics, medium window statistics, and long window statistics; In step S23, dynamic upper and lower thresholds are obtained based on the multi-scale weighted statistics and adaptive coefficients.

[0053] In such Figure 2 In the method shown, in step S20, the short window statistic, the medium window statistic, and the long window statistic are calculated respectively. Specifically, in this embodiment, the short window statistic is used to measure instantaneous volatility. The calculation method can be, for example, using formula (34) or formula (36): (34) (35) (36) in, Represents the coefficient of variation for the short window. This indicates the size of the short window; it's a system-preset parameter.

[0054] In the calculation, window statistics and trend terms are used to measure intraday trends, and the calculation formulas are as follows: (37) (38) (39) in, Indicates the trend in the middle window. Medium window size.

[0055] The formula for calculating long window statistics is as follows: (40) (41) After calculating the short-window, medium-window, and long-window statistics, a multi-scale weighted statistic is calculated in step S21. The three-layer window statistics are then fused using dynamic weights to generate a comprehensive mean and standard deviation. Specifically, in this example, the multi-scale weighted statistic can be obtained using formulas (1) and (2): (1) (2) in, The average value is the comprehensive average. For the composite standard deviation, This is the short window mean. The mean of the middle window. The average value over a long window. , and These are the weighting coefficients. For short window standard deviation, The standard deviation of the middle window. This represents the standard deviation for a long window. In this example, a timeliness weight is set. (Shortest window has the highest weight and responds first to immediate changes.) (The middle window is secondary, reflecting the intraday trend.) (Long window minimum, providing baseline reference), satisfies normalization constraints. .

[0056] Step S22 is used to obtain the adaptive coefficient. The adaptive coefficient is calculated using a "base coefficient × multidimensional adjustment coefficient" multiplication mechanism, which comprehensively considers the impact of four dimensions—volatility, trend, seasonality, and load—on the threshold. The calculation formula is as follows: (3) in, For adaptive coefficients, Based on the coefficient, This is the volatility adjustment factor. This is a trend adjustment coefficient. This is a seasonal adjustment factor. This is a load-related adjustment factor. In this example, the base factor... It could be version 3.0.

[0057] Volatility Adjustment Factor Calculation using the continuous formula: (4) Among them, the range is . It is obtained from formula (36).

[0058] Trend adjustment coefficient Based on the trend item in the middle window The calculation formula is as follows: (5) The range is . This is obtained from formula (39). When (When there is no trend) (No adjustment); when During an upward trend, (Relax the upper limit threshold to avoid misjudging upward trends as abnormal); when (During a downward trend) (Tighten the lower threshold to enhance anomaly detection capabilities); an amplification factor of 20 ensures significant adjustments occur when the trend term is within ±5%. For example: a production line experiences a continuous increase in energy consumption due to capacity expansion. (Daily increase of 6%), calculate: ; The threshold is raised by 8% to adapt to the upward trend and avoid normal capacity increases being misjudged as abnormal energy consumption; conversely, if... (Daily energy saving 8% after energy-saving renovation) The threshold is lowered by 8% to quickly identify whether the energy-saving effect meets the standard.

[0059] Seasonal adjustment coefficient Based on the annual seasonal cycle, the calculation formula is as follows: (6) in, The seasonal fluctuation range (±15%) This is a phase shift (to make the winter value correspond to the maximum value and the summer value correspond to the minimum value, which is consistent with the pattern that the energy consumption for heating in the manufacturing industry is high in winter and relatively low in summer). This is a standard time function, representing the day of the year.

[0060] Load-related adjustment factor Based on real-time load rate The calculation formula is as follows: (7) The calculation range is .when (Unloaded) (Tighten the threshold; no-load energy consumption should be extremely low); when (When fully loaded) (Relaxed threshold, energy consumption fluctuates greatly under full load); linear relationship ensures that the threshold changes synchronously with the load. It is obtained from formula (22).

[0061] Step S23 is used to calculate the dynamic upper and lower thresholds based on the multi-scale weighted statistics and adaptive coefficients: (42) (43) Next, the threshold is smoothed using exponential smoothing to avoid abrupt changes, for the initial time step. ,set up ;for The time is calculated using the following formula: (44) Among them, the smoothing coefficient The typical value is 0.2.

[0062] By employing a three-layer window system—short-window instantaneous fluctuation capture, medium-window intraday pattern recognition, and long-window periodic pattern extraction—energy consumption characteristics monitoring across time scales is achieved. Combined with a four-dimensional adaptive coefficient of "volatility, trend, seasonality, and load," the anomaly detection threshold is dynamically adjusted to generate an upper limit threshold. With lower threshold This ensures that the threshold can adapt to energy consumption fluctuation characteristics, trend changes, seasonal patterns, and load status, with an anomaly detection accuracy of ≥90%.

[0063] Step S13 is used to perform multi-dimensional anomaly determination on the current energy consumption data based on the dynamic threshold to obtain a comprehensive anomaly score and a persistence score. In this embodiment, the specific method for multi-dimensional anomaly determination can be a five-dimensional anomaly determination system consisting of "absolute threshold deviation determination - prediction deviation determination - baseline deviation determination - confidence interval deviation determination - rate of change determination," which quantifies the degree of energy consumption anomaly from different perspectives. A "non-linear scoring activation function + weighted comprehensive scoring" mechanism is used to integrate the five-dimensional indicators into a comprehensive anomaly score. Based on the scoring to classify the anomaly level (P1 severe / P2 moderate / P3 slight / P4 normal), combined with continuous evaluation to avoid false alarms due to transient noise, the system generates equipment-level anomaly judgment results with an anomaly identification recall rate ≥95% and precision rate ≥90%.

[0064] Specifically, in this example, dimension 1 is the absolute threshold deviation determination, and the absolute threshold deviation index is calculated: (45) For the current moment The real-time energy consumption value is the raw energy consumption value that is directly collected from the sensor and preliminarily calculated, but has not undergone in-depth preprocessing (such as smoothing or fusion).

[0065] Dimension 2 is for determining prediction bias, calculating the prediction bias rate: (46) Dimension 3 is for determining the benchmark deviation and calculating the benchmark deviation rate: (47) Dimension 4 is the confidence interval deviation determination: (48) , It is obtained from formulas (19) and (20).

[0066] Dimension 5 is the rate of change determination, which calculates the rate of change per unit time: (49) in, This indicates a time interval, typically 5 minutes. Real-time energy consumption value before time Δt.

[0067] Define a scoring activation function to convert the five-dimensional anomaly index into a non-linear score, amplifying significant anomalies and suppressing minor biases: (50) Next, a comprehensive anomaly score is calculated, and a comprehensive score is generated by weighted summation and fusion of the five-dimensional scores. Specifically, the comprehensive anomaly score can be calculated, for example, using formula (8): (8) in, For comprehensive anomaly scoring, For the first Abnormal indicators in each dimension For the first Scoring across multiple dimensions For the first The weighting coefficients for each dimension's score. In this example, the weights are configured as follows: (Absolute threshold deviation has the highest weight and most directly reflects anomalies.) (Prediction bias is secondary) (Benchmark deviation is secondary) (Confidence intervals have lower weights and are used to assist in the determination) (The rate of change has a higher weighting, capturing mutations).

[0068] Based on the scoring system, anomalies are classified into four levels: P1 Severe Anomalies (…) P2 is moderately abnormal. P3 slightly abnormal ( P4 is normal. Example: The aforementioned overall score is 8.1 points, because... The anomalies are categorized into several levels: P1 (Severe Anomaly), triggering a high-priority alert (e.g., sending SMS / email notifications to operations personnel and initiating emergency response procedures); a score of 6.5 at any given time indicates a P2 (Moderate Anomaly), which records the anomaly log and pushes it to the monitoring platform, but does not trigger an immediate notification; a score of 3.2 indicates a P3 (Minor Anomaly), which only records the log for subsequent analysis; and a score of 1.5 indicates a P4 (Normal), requiring no action. This tiered mechanism ensures that operations personnel prioritize high-risk anomalies, preventing low-level anomalies from overwhelming critical alerts.

[0069] The persistence of abnormalities is assessed by calculating a persistence score according to formula (9). (9) in, For continuous scoring, This indicates a continuous evaluation window, typically 30 minutes. For indicator functions, For a historic moment The comprehensive anomaly score. Anomaly persistence assessment can suppress transient disturbances, reduce false alarms, and quantify auxiliary dimensions of fault severity and urgency.

[0070] In step S14, multi-level early warning aggregation is performed based on the comprehensive anomaly score and the persistence score to obtain an early warning plan. In this embodiment, a four-level hierarchical early warning system of "equipment-level early warning judgment - production line-level early warning aggregation - workshop-level early warning aggregation - plant-wide early warning aggregation" is used to achieve hierarchical aggregation and risk assessment of abnormal information from single equipment to the entire plant. A three-dimensional judgment mechanism of "abnormal equipment ratio + comprehensive score sum + energy consumption deviation rate" is adopted to generate early warning trigger signals and risk levels for each level. Combined with workshop energy efficiency calculation and multi-level anomaly assessment of the plant area, hierarchical decision support is provided for managers, and the early warning response time is ≤5 minutes.

[0071] In this example, device-level early warning determination can be based on device anomaly scoring. An alert is triggered when the threshold is reached: (51) in, Indicates the equipment number; This indicates the warning threshold, with a typical value of 2.

[0072] Production line-level early warning aggregation calculates the proportion of abnormal equipment, the total abnormal score of the production line, and the total energy consumption deviation of the production line. The proportion of abnormal equipment is calculated according to formula (52): (52) in, Indicates production line The set of devices included.

[0073] The total production line anomaly score is calculated according to formula (53): (53) The total energy consumption deviation of the production line is calculated according to formula (54): (54) in, Indicates the actual total energy consumption of the production line ; It represents the baseline total energy consumption of the production line, which is the sum of the baseline energy consumption of all equipment in the production line.

[0074] The production line-level early warning trigger criteria are as follows: (55) The threshold conditions for judgment are the proportion of abnormal devices, the total abnormal score, and the trigger threshold of energy consumption deviation rate, all of which are preset parameters of the system.

[0075] Workshop-level early warning aggregation, calculate the proportion of abnormal production lines in the workshop according to formula (56): (56) in, Workshop The set of production lines included. It refers to the production line At any moment The warning trigger status indicator.

[0076] Calculate the overall anomaly degree of the workshop according to formula (57): (57) Among them, weight , ; The total energy consumption deviation rate of the workshop is obtained as follows: (58) (59) (60) Plant-level early warning aggregation, calculate the multi-level anomaly degree of the plant area according to formula (61): (61) in, Indicates device layer weights; Indicates the weight of the production line layer; Represents the workshop layer weight; satisfies ; and These represent the number of abnormal devices that triggered the warning and the total number of controlled devices across the entire plant at the current moment, respectively. and These represent the number of abnormal production lines that have triggered warnings and the total number of production lines, respectively, across the entire plant at the current moment. and These represent the number of workshops identified as abnormal and the total number of workshops in the entire plant at the current moment, respectively.

[0077] Calculate the overall energy efficiency of the plant area according to formula (62): (62) in, Indicates total output value or total output; This represents the total energy input, which corresponds to the real-time total standard coal equivalent obtained and calculated in step S10. .

[0078] Thus, through the above-mentioned multi-level early warning aggregation, comprehensive monitoring from equipment to the factory area has been completed, realizing the hierarchical aggregation of abnormal information and layered decision support. This helps managers quickly locate the root cause of problems (equipment failure / improper production line configuration / workshop collaboration failure / factory-wide energy management defects) and take targeted measures. The early warning system covers 100% of the four levels of equipment, production line, workshop, and factory area. Early warning signals are pushed to the corresponding responsible persons in a hierarchical manner to avoid information overload or missed reports.

[0079] Furthermore, the present invention provides an adaptive optimization method according to one embodiment. The method includes the following steps: In step S30, periodic performance evaluation is performed. True positives are counted. True negative False positives False negative Four types of sample sizes. In energy anomaly detection scenarios, new devices may be added to the monitoring network, or operating conditions may change. In such cases, it is necessary to evaluate system performance and optimize parameters.

[0080] In step S31, the performance indicators are calculated. The calculation formula is as follows: ; Precision represents precision, recall represents recall, and the F1 score is the harmonic mean of the two.

[0081] In step S32, the optimization objective function is defined: (63) in, Indicates about parameters The objective function value is optimized. This represents the adaptive coefficient parameter variable in the current optimization attempt; This represents the optimal parameter value determined after the previous optimization cycle. This indicates that the parameter takes the value of The false alarm rate at time FP is calculated as: FPR = FP / (FP + TN). This indicates that when the parameter takes the value The false negative rate at that time is FNR = FN / (TP + FN); weight allocation , , While reducing false positives (FPR) and false negatives (FNR), the system avoids drastic changes in parameter k relative to its historical value k_prior, thus maintaining system stability. The adaptive coefficient variable currently being attempted is the total adaptive coefficient in formula (3). .

[0082] In step S33, the adaptive coefficients are optimized using the gradient descent method: (64) in, Indicates the first The parameter update value after the next iteration; Indicates the first The parameter values ​​at the next iteration; This represents the learning rate, typically 0.01. Describe the objective function For parameters The gradient of the objective function. The iterative convergence condition is defined as: when the absolute value of the gradient of the objective function is less than a set threshold (i.e., ... , Typical value ) or the change in the objective function value between two consecutive iterations is less than a set threshold (i.e. When the iteration stops, output the current value. These are the optimal parameters for this round.

[0083] In step S34, the baseline model is updated on a rolling basis, with the baseline values ​​for each time period of the previous day being updated daily: (65) in, Indicates the updated number Baseline energy consumption values ​​for different time periods; This represents the historical energy consumption baseline value before the update; Indicates the number of the previous day Actual energy consumption measurements for the time period; The retention weight (forgetting factor) for historical benchmarks; The updated weights (learning factors) are used for the latest data. The baseline values ​​for each load segment are updated monthly based on the previous month's data.

[0084] Thus, through rolling updates, the baseline model can gradually adapt to changes in production patterns, improving the system's adaptability.

[0085] On the other hand, the present invention also provides a multi-level energy anomaly early warning system, the system comprising a data acquisition module, a data processing module, a dynamic anomaly determination module, a hierarchical early warning module, and a processor. The data acquisition module is used to acquire multi-source data of the energy system, including real-time operating data, historical forecast data, baseline energy consumption data, and operating parameter information of the industrial energy system. The data processing module is used to preprocess and extract features from the multi-source data to generate fused data. The dynamic anomaly determination module is used to construct a multi-scale sliding window on the fused data to generate dynamic thresholds and obtain a comprehensive anomaly score and a persistence score. The hierarchical early warning module is used to aggregate and analyze the comprehensive anomaly score and persistence score layer by layer at the equipment level, production line level, workshop level, and factory area level to generate a multi-level early warning output scheme. The processor is used to connect the data acquisition module, data processing module, dynamic anomaly determination module, and hierarchical early warning module, and the processor is configured to execute any of the methods described in the multi-level energy anomaly early warning method. In one embodiment of the present invention, the architecture diagram of the multi-level energy anomaly early warning method is as follows. Figure 3 As shown.

[0086] In another aspect, the present invention also provides a computer-readable storage medium storing instructions that, when executed by a processor, implement any of the methods described in the multi-level energy anomaly early warning method.

[0087] Compared with the prior art, the present invention has the following advantages: (1) Significantly improves anomaly detection accuracy. Through a three-layer sliding window system and adaptive adjustment coefficient, intelligent calculation of dynamic thresholds is achieved, overcoming the shortcomings of traditional fixed thresholds that cannot adapt to changes in working conditions. Combined with a five-dimensional anomaly judgment system, the anomaly detection recall rate is ≥95% and the precision rate is ≥90% based on five dimensions: absolute threshold, prediction deviation, baseline deviation, confidence interval, and rate of change. Compared with traditional methods, the false alarm rate is reduced by more than 40%, and the false negative rate is reduced by more than 35%.

[0088] (2) Achieve multi-level global early warning. Construct a four-level hierarchical early warning system at the equipment level, production line level, workshop level, and plant level. Through a three-dimensional judgment mechanism of abnormal equipment ratio, total abnormal score, and total energy consumption deviation rate, achieve comprehensive monitoring from micro to macro. It can accurately locate single equipment problems and identify overall energy waste, providing hierarchical decision support for managers, with an early warning response time of ≤5 minutes.

[0089] (3) Enhance system adaptability. Generate prediction curves for the next 24 hours using the LSTM time series prediction model to achieve predictive anomaly early warning, predict abnormal energy consumption trends in advance, and transform from passive response to proactive prevention. The gradient descent method is used to continuously optimize the adaptive coefficients, and the baseline model is updated daily and monthly to enable the system to gradually adapt to changes in production mode and maintain stable accuracy in long-term operation.

[0090] (4) Provide comprehensive performance evaluation. Quantitatively evaluate the anomaly detection effect through multi-dimensional performance indicators such as precision, recall, and F1 score. Combine comprehensive energy efficiency calculation and energy-saving potential quantitative evaluation of the plant area to help enterprises identify the main sources of energy waste, quantify energy-saving improvement space, and achieve refined energy management.

[0091] (5) Highly efficient and reliable technology. Real-time data preprocessing is performed using edge computing nodes, reducing the computing pressure on the central server and reducing data processing latency to <3 seconds. A multi-terminal visualization output mechanism synchronously pushes the evaluation results to the energy management platform, production monitoring center, and enterprise energy management system, supporting multi-role collaborative decision-making. The system supports 100% device coverage, and early warning signals are pushed to the corresponding responsible persons in a tiered manner to avoid information overload or missed reports.

[0092] (6) Significant economic benefits. Through precise anomaly detection and multi-level early warning, it helps enterprises to promptly identify and handle energy anomalies, avoiding energy waste and production losses. In practical applications, it can improve enterprises' energy utilization efficiency by 5-15%, reduce annual energy costs by -10%, and reduce equipment failure rates by more than 25%, resulting in good economic and social benefits.

[0093] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0094] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0095] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0096] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0097] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0098] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0099] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0100] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0101] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A multi-level energy anomaly early warning method, characterized in that, The early warning method includes: Acquire multi-source data from the energy system; Multidimensional feature extraction is performed on the multi-source data to obtain fused features; Based on the fusion features, a multi-scale sliding window is constructed to calculate the dynamic threshold; The current energy consumption data is subjected to multi-dimensional anomaly determination based on the dynamic threshold to obtain a comprehensive anomaly score and a persistence score. Multi-level early warning aggregation is performed based on the comprehensive anomaly score and the persistence score to obtain an early warning scheme.

2. The early warning method of claim 1, wherein, Acquiring multi-source data from the energy system includes: The system acquires real-time operating data, historical forecast data, baseline energy consumption data, and operating condition parameter information of the energy system. The historical forecast data is the system energy consumption forecast data based on historical data using an LSTM time series forecast model. The baseline energy consumption data is obtained by hierarchically statistically analyzing historical data according to time period, load, season, and operating condition.

3. The early warning method of claim 1, wherein, The process of extracting multidimensional features from the multi-source data and obtaining fused features includes: Statistical features, time-series features, load features, and comparative features of the multi-source data are extracted, wherein the time-series features are used to extract the first-order autocorrelation coefficient, the load features are used to extract the correlation coefficient between energy consumption and load, and the comparative features are used to extract the year-on-year change rate. Dynamic weighting is applied to real-time energy consumption data, energy consumption prediction data, and baseline energy consumption data to obtain fusion characteristics.

4. The early warning method of claim 1, wherein, Based on the aforementioned fusion features, constructing a multi-scale sliding window to calculate the dynamic threshold includes: Calculate the statistics for short window, medium window, and long window respectively; Based on the short window statistics, medium window statistics, and long window statistics, multi-scale weighted statistics are obtained; Based on the short window statistics, medium window statistics, and long window statistics, adaptive coefficients are obtained; Based on the multi-scale weighted statistics and adaptive coefficients, dynamic upper and lower thresholds are obtained.

5. The early warning method of claim 4, wherein, Based on the short-window, medium-window, and long-window statistics, the multi-scale weighted statistics are obtained, including: Obtain the multiscale weighted statistics according to formulas (1) and (2): ，（1） ，（2） in, The average value is the comprehensive average. For the composite standard deviation, This is the short window mean. The mean of the middle window. The mean of the long window. , and These are the weighting coefficients. For short window standard deviation, The standard deviation of the middle window. The standard deviation is for the long window.

6. The early warning method of claim 4, wherein, Based on the short window statistics, medium window statistics, and long window statistics, the adaptive coefficients are obtained as follows: Calculate the adaptive coefficients according to formulas (3) to (7). ，（3） ，（4） ，（5） ，（6） ，（7） in, For adaptive coefficients, Based on the coefficient, This is the volatility adjustment factor. This is a trend adjustment coefficient. This is a seasonal adjustment factor. This is the load-related adjustment factor. The coefficient of variation for the short window. For the medium window trend, For seasonal fluctuations, For standard time functions, For phase deviation, This represents the real-time load rate.

7. The early warning method of claim 1, wherein, Based on the dynamic threshold, a multi-dimensional anomaly determination is performed on the current energy consumption data to obtain a comprehensive anomaly score and a persistence score, including: Scores were calculated for five dimensions, including absolute threshold deviation index, prediction deviation rate, baseline deviation rate, confidence interval deviation index, and time change rate. Calculate the comprehensive anomaly score according to formula (8). ，（8） in, For comprehensive anomaly scoring, For the first Abnormal indicators in each dimension For the first Scoring across multiple dimensions For the first Weighting coefficients for each dimension of the score; Calculate the persistence score according to formula (9). ，（9） wherein, is a persistence score, represents a persistence evaluation window, is an indicator function, is a historical time of the composite anomaly score.

8. The early warning method of claim 1, wherein, Based on the comprehensive anomaly score and the persistence score, a multi-level early warning aggregation is performed to obtain an early warning scheme, including: Perform device-level early warning judgment, including: triggering an early warning when the overall anomaly score of the current device reaches a threshold; Perform production line-level aggregation analysis, including: calculating the proportion of abnormal equipment on the production line, the total abnormal score of the production line, and the total energy consumption deviation of the production line respectively; and triggering an early warning if at least one of the proportion of abnormal equipment on the production line, the total abnormal score of the production line, and the total energy consumption deviation of the production line reaches a threshold. Perform workshop-level aggregate analysis, including: calculating the proportion of abnormal production lines and the overall abnormality of the workshop; Perform plant-level aggregate analysis, including: calculating multi-level anomalies and overall plant performance; Based on the equipment-level early warning judgment, production line-level aggregate analysis, workshop-level aggregate analysis, and factory-level aggregate analysis, a multi-level early warning output scheme is generated.

9. A multi-level energy anomaly early warning system, characterized in that, The system includes: The data acquisition module is used to acquire multi-source data from the energy system; The data processing module is used to preprocess and extract features from the multi-source data to generate fused data; The dynamic anomaly detection module is used to construct a multi-scale sliding window on the fused data to generate a dynamic threshold and obtain a comprehensive anomaly score and a persistence score. The hierarchical early warning module is used to aggregate and analyze the comprehensive anomaly score and the persistence score layer by layer according to the equipment level, production line level, workshop level and factory area level, and generate a multi-level early warning output scheme. A processor is used to connect the data acquisition module, the data processing module, the dynamic anomaly determination module, and the hierarchical early warning module, and the processor is configured to perform the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a processor, implement the method as described in any one of claims 1 to 8.

Images