A plant power load monitoring and early warning method based on digital twinning

Abstract

This invention provides a digital twin-based method for monitoring and early warning of factory power load, belonging to the fields of digital twin and power load forecasting technology. This invention combines WTConv with iTransformer for power load forecasting, solving the challenge of traditional power load forecasting methods struggling in complex and dynamic factory environments. In actual factory operation, power load is affected by multiple factors such as production rhythm, equipment operating status, and external environment, resulting in load data exhibiting strong time-varying and nonlinear characteristics. By combining WTConv with iTransformer, this invention can better capture complex spatiotemporal patterns and multi-scale features, thereby significantly improving the accuracy and robustness of power load forecasting, providing more reliable load forecasting support for power systems, helping enterprises optimize production scheduling, reduce energy losses, and effectively prevent power system load risks.

Description

Technical Field

[0001] This invention relates to the fields of digital twin and power load forecasting technology, and in particular to a method for monitoring and early warning of factory power load based on digital twin. Background Technology

[0002] In modern industrial production, electricity, as one of the core energy sources, directly impacts the stability of production processes and energy efficiency due to load variations. To ensure the stability, safety, and energy efficiency of the power system during factory operation, load monitoring and early warning have become crucial aspects of smart factory construction. Currently, mainstream power load monitoring methods are mostly based on SCADA systems, edge sensing devices, and rule-based early warning mechanisms. These methods monitor the system's operating status by collecting parameters such as current, voltage, and power factor in real time, and issue early warnings for anomalies using manually set thresholds. Furthermore, to improve prediction accuracy, some studies have attempted to introduce time series models (such as ARIMA and SARIMA), machine learning algorithms (such as SVR and random forests), or deep learning networks (such as LSTM and Transformer) to model and predict load trends. While these methods have improved data processing capabilities and prediction performance to some extent, they generally suffer from insufficient generalization ability, poor model interpretability, and slow prediction speed, making it difficult to meet the current monitoring and early warning needs of factory power loads in dynamic, complex, and heterogeneous environments.

[0003] Currently, there is a lack of a power load management and early warning framework that can integrate multi-source heterogeneous data, has real-time feedback capabilities, and can dynamically evolve. Therefore, this invention designs a method that combines digital twin technology to achieve synchronous mapping between the virtual and real worlds of a factory power system and full lifecycle load modeling, in order to improve monitoring accuracy, early warning foresight, and system operability, and to facilitate the intelligent and green transformation of industrial energy systems. Summary of the Invention

[0004] This invention proposes a method for monitoring and early warning of factory power load based on digital twins. By using digital twin technology to virtualize the real factory and combining the actual power system of the factory with the virtual model, the method can achieve real-time monitoring, prediction and early warning of power load, thereby improving the operating efficiency and safety of the factory power system.

[0005] This invention is implemented as follows:

[0006] A method for monitoring and early warning of factory power load based on digital twins, the method comprising the following steps:

[0007] S1. Provides stable power support for the entire power load monitoring and early warning system through the power supply module;

[0008] S2. Collect real-time data from sensors and actuators through industrial protocols to complete equipment load status monitoring;

[0009] S3. Based on the collected real-time data, a three-dimensional virtual factory model is constructed using digital twin technology and factory data information. The real-time data is then mapped onto the virtual model through a data interface to achieve dynamic simulation and interactive visualization of the model.

[0010] S4. The differential autoregressive moving average model (ARIMA) is used to analyze the collected real-time data and uncover potential power load change trends for normal equipment operation.

[0011] S5. Based on the potential power load change trend of normal equipment operation, apply WTConv wavelet transform to iTransformer model to predict power load data for the next hour.

[0012] S6. Output and visualize the prediction results;

[0013] S7. Identify overload events by monitoring and predicting data, trigger early warnings, and perform fault diagnosis and analysis.

[0014] Furthermore, the specific method for collecting real-time data from sensors and actuators through industrial protocols to complete equipment load status monitoring in step S2 includes:

[0015] The industrial data acquisition terminal of the edge gateway periodically acquires real-time data from factory equipment using the OPC UA industrial communication protocol and performs preliminary verification of the real-time data.

[0016] The collected real-time data specifically includes equipment operation data and environmental data;

[0017] The equipment operation data includes power parameters and status parameters. The power parameters include: voltage, current, active power, reactive power, and power load. The status parameters include: equipment on / off status, operating mode, and fault status.

[0018] The environmental data includes environmental factors, such as indoor temperature, humidity, and air pressure in the factory.

[0019] Furthermore, the specific methods described in step S3 for constructing a three-dimensional virtual model based on the collected real-time data using digital twin technology and factory data information, and mapping the real-time data onto the virtual model through a data interface to achieve dynamic simulation and interactive visualization of the model include:

[0020] S31. Using digital twin technology, collect the factory's geographic information, production unit layout data, and equipment locations through sensors and engineering design documents;

[0021] The geographic information, the production unit layout data, and real-time equipment operation data are integrated to generate a three-dimensional virtual factory model.

[0022] Static metadata of the equipment is embedded in the three-dimensional virtual factory model. The static metadata includes equipment type, specifications, unique identifier, and maintenance cycle. A mapping relationship between the equipment location and virtual nodes in the three-dimensional virtual factory model is established based on the unique identifier.

[0023] S32. Connect the real-time data to the virtual model through an interface, and display the key parameter information on the three-dimensional model in real time to form a dynamic simulation effect.

[0024] Furthermore, the specific method described in step S4 for using the differential autoregressive moving average (ARIMA) model to uncover potential power load variation trends for normal equipment operation is as follows:

[0025] The ARIMA model combines autoregressive (AR), differencing (I), and moving average (MA), and the formula is as follows:

[0026]

[0027] The autoregression section describes the current observed value X. t How to use the past p observations and the random error term ∈ t Decision, φ i Autoregressive coefficients, where c is a constant term;

[0028] Δ d X t =(1-M) n X t

[0029] Difference operations are used to transform non-stationary time series into stationary series; n represents the order of the difference, Δ n X t It is an n-order difference time series, where M is the lag operator, (1-M). n Indicates the difference operation;

[0030]

[0031] The moving average represents the current observation value X. t From past error terms ∈ t The weighted sum of the lag error terms; μ represents the mean of the model, θ i It is the moving average coefficient, representing the impact of the past i error terms on the current value; ∈ t-i It is the error term from the past i time steps.

[0032] The ARIMA model is used to analyze whether the time series exhibits a monotonically rising or falling trend and to identify seasonal fluctuations in the time series. The autoregressive part is used to determine the long-term dependencies in the data, and the moving average part is used to determine whether noise in the data has a significant impact on the current observation.

[0033] Furthermore, the specific method described in step S5 for applying the WTConv wavelet transform to the iTransformer model based on the potential power load change trend of normal equipment operation to predict the power load data for the next hour is as follows:

[0034] S51. The input is historical time-series data of power load, usually expressed as:

[0035] X = [x1, x2, x3, ..., x T ]∈R T×d

[0036] Where T is the number of time steps and d is the feature dimension;

[0037] The power load time-series data is normalized using the following formula:

[0038]

[0039] Where x is the input power time series data, μ(x) is the mean of the input data, and σ(x) is the standard deviation of the input data;

[0040] S52. The wavelet transform as convolutions is used to extract low-frequency and high-frequency information from the power load time series data; based on a time series x of length L, the formula for this process is as follows:

[0041]

[0042] Where WT(x) is the wavelet transform operation, and Convld(x, filters) is the one-dimensional convolution operation, where x is the input signal and filters are the convolution kernels;

[0043] After wavelet transform, the sequence is decomposed into multiple wavelet sub-bands, representing information from different frequency bands: x L ,x H , where x L For low-frequency components, x H High-frequency components; then for x L ,x H F is obtained by extracting local features using 1D convolution. L =Conv1D(X L ), FH =Conv1D(X H );

[0044] S53. Through a learnable linear transformation, multi-scale features are projected onto a feature space of the same dimension, where W L W H ∈R d'×d It is the projection matrix:

[0045] E L =F L ·W L +b L E H =F H ·W H +b H

[0046] Among them, E L E H The embedding result is the low-frequency and high-frequency components after linear transformation; b L and b H These are bias terms for low-frequency and high-frequency characteristics;

[0047] Then, the two are concatenated to form the final frequency component E = Concat(E L E H The low-frequency and high-frequency features of different frequency components after processing by WTConv are embedded into the variable tokens of the iTransformer prediction model. In the embedding stage, each frequency band component is embedded into an independent variable token to maintain the independence of information in each frequency band.

[0048] The iTransformer basic model proposes a reverse innovation approach through a redesigned architecture, the core of which lies in redefining the sequence embedding method with variables at its center; in iTransformer, based on the input time series X... ：,n Information is passed layer by layer through the Transformer module TrmBlock, and the final prediction result is... The process is as follows (using the formula):

[0049]

[0050] H l+1 =TrmBlock(H l ), l=0,...,l-1,

[0051]

[0052] It is the initial representation of the nth time series after passing through the embedding layer; embedding represents the embedding operation; H l+1 It is the output of the Transformer block at layer (l+1), (H l ) is the output of the previous layer. The final output contains the context vector associated with the nth variable, where C represents the module layer number. The predicted feature representation is input into the Projection layer.

[0053] It contains N embedding tokens with dimension D; iTransformer embeds the input sequence into variable tokens and calculates the correlation between variables through a self-attention mechanism; a feedforward network is used on each variable token to focus on extracting non-linear features of the time series;

[0054] S54. The main stage of power load forecasting adopts the iTransformer model, and the multi-head self-attention mechanism is used to capture the dependencies between multiple variables in each time series.

[0055] First, based on the input E = Concat(E) L E H Generate query, key, and value vectors:

[0056] Q = E·W Q K = E·W K V = E·W V

[0057] Where Q is the query vector, K is the key vector, V is the value vector, and W is the value vector. Q W K W V These are learnable weight matrices, used to map the input features E to the query, key, and value spaces, respectively.

[0058] In the attention mechanism, low-frequency variable tokens are assigned higher weights, while high-frequency variable tokens are assigned lower weights; the attention calculation method is as follows:

[0059]

[0060] Among them, Q w ,K w V w These represent the query, key, and value, respectively, and are all vectors within a local window w. k It is the vector dimension;

[0061] Multi-head concatenation and linear mapping: MHSA(E) = Concat(head1,...,head) H )·W O ;

[0062] S55. Using residual connectivity and layer normalization, it is expressed as follows:

[0063] Z = LayerNorm(E + MHSA(E))

[0064] The features of each frequency band are recombine into a whole feature using the Inverse Discrete Wavelet Transform (IDWT): Z agg =IDWT(Z);

[0065] S56, The above overall features Z agg The input is a feedforward network, with a two-layer MLP structure, as shown in the following formula:

[0066] h1 = ReLU(z) agg ·W1+b1)

[0067] h2 = h1·W2 + b2

[0068] Where ReLU is the activation function, W1 and W2 are weight matrices, and b1 and b2 are bias terms;

[0069] S57, Output is the predicted load value for the next time step. The system predicts the power load data for the next hour and stores the prediction results in the system.

[0070] Furthermore, the specific method for identifying overload events by monitoring and predicting data, triggering early warnings, and performing fault diagnosis and analysis as described in step S7 includes:

[0071] S71. Based on the predicted data, issue an early warning when the load is about to be overloaded to remind the equipment administrator to handle it in time;

[0072] S72. By analyzing data and historical fault data, determine the cause of the fault and provide steps and suggestions for repairing the fault.

[0073] Furthermore, in step S71, based on the predicted data, a warning is issued when the load is about to be overloaded, reminding the equipment administrator to take timely action. The specific method is as follows:

[0074] The load change rate is used to determine the rate of load growth by calculating the ratio of the current load to the load at the previous moment. The formula is as follows:

[0075] ΔL t =L t -Lt-1

[0076] L t For the load data at the current moment, L t-1 The load data from the previous moment, ΔL t This refers to the change in load.

[0077] The average load calculated based on historical data is μ. L The standard deviation is σ L The safe load threshold can be set as follows:

[0078] L th =μ L +α·σ L

[0079] Where α is the safety factor, which can be adjusted by managers. It is usually set to 2, which represents the situation where the load exceeds the average load by a certain standard deviation.

[0080] The system compares the predicted load value with the set threshold. If the predicted value exceeds the safe load threshold, the system will issue an overload warning signal.

[0081] Furthermore, in step S72, the cause of the fault is determined through data analysis and historical fault data, and specific methods and suggestions for repairing the fault are provided:

[0082] Using Bayesian networks, we can determine the conditional probability P(Cause) i The Fault infers the optimal root cause and provides steps and suggestions for fixing the fault, as shown in the following formula:

[0083]

[0084] Wherein, P(Cause) i |Fault) represents the probability of a certain cause of failure occurring given that a failure has occurred; P(Fault|Cause) i P(Cause) represents the probability of a failure occurring given that a certain cause of failure has occurred; i P(Fault) represents the prior probability of a fault occurring; P(Fault) represents the total probability of a fault occurring.

[0085] Compared with the prior art, the present invention has the following advantages:

[0086] 1. The digital twin factory power load monitoring method of this invention effectively solves the shortcomings of traditional factory power load monitoring systems in practical applications, especially the inability to reflect load fluctuations, equipment operating status, and process parameter changes during factory production in real time. By introducing digital twin technology, a virtual factory model can be established and synchronized with real-time data from the actual factory, thereby accurately simulating and monitoring real-time changes in the factory's power load. This method provides factories with higher-precision load monitoring capabilities and can detect potential overload or equipment failure risks in advance, improving the stability and security of the power system.

[0087] 2. This invention combines WTConv with iTransformer for power load forecasting, addressing the challenge of traditional power load forecasting methods struggling in complex and dynamic factory environments. In actual factory operation, power load is influenced by multiple factors, including production rhythm, equipment operating status, and the external environment, resulting in load data exhibiting strong time-varying and nonlinear characteristics. By combining WTConv and iTransformer, this invention can better capture complex spatiotemporal patterns and multi-scale features, thereby significantly improving the accuracy and robustness of power load forecasting. This provides more reliable load forecasting support for power systems, helping enterprises optimize production scheduling, reduce energy losses, and effectively prevent load risks in power systems.

[0088] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. Attached Figure Description

[0089] 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.

[0090] Figure 1 This is a flowchart of a factory power load monitoring and early warning method based on digital twins according to the present invention;

[0091] Figure 2 This is a schematic diagram illustrating data visualization and interaction in this invention;

[0092] Figure 3 This is a schematic diagram of the power load forecasting process in this invention;

[0093] Figure 4 This is a schematic diagram of the data transmitter module in this invention. Detailed Implementation

[0094] To enable those skilled in the art to better understand the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art. Based on the implementation methods provided in the above aspects, this application can also be further combined to provide more implementation methods.

[0095] This invention provides a method for monitoring and early warning of factory power load based on digital twins. See also... Figure 1 , Figure 1 A flowchart illustrating a factory power load monitoring and early warning method based on digital twins, provided for embodiments of the present invention. The load monitoring and early warning method includes the following steps:

[0096] S1. Provides stable power support for the entire power load monitoring and early warning system through the power supply module.

[0097] S2. Collects real-time data from sensors and actuators through industrial protocols to complete equipment load status monitoring.

[0098] S3. Based on the collected real-time data, a three-dimensional virtual factory model is constructed using digital twin technology and factory data information. The real-time data is then mapped onto the virtual model through a data interface to achieve dynamic simulation and interactive visualization of the model.

[0099] S4. The Autoregressive Moving Average (ARIMA) model is used to analyze the collected real-time data and uncover potential power load change trends for normal equipment operation.

[0100] S5. Based on the potential power load change trend of normal equipment operation, the WTConv wavelet transform is applied to the iTransformer model to predict the power load data for the next hour.

[0101] S6. Output and visualize the prediction results.

[0102] S7. Identify overload events by monitoring and predicting data, trigger early warnings, and perform fault diagnosis and analysis.

[0103] Specifically, the power supply module provides stable power support to the system, while the equipment load status monitoring module monitors the power load status of each device in real time. Simultaneously, the data visualization and interaction module, through digital twin technology, constructs a 3D virtual model of the factory's actual layout, production lines, and equipment. Equipment power data visualization and environmental data visualization intuitively display the power load status and its changing trends, providing comprehensive data support for management personnel. In the data analysis phase, through in-depth analysis of the collected power load data, the system can detect potential abnormal fluctuations, further providing a basis for the early warning system. The power load prediction module uses time series prediction algorithms to predict future load changes based on historical data trends, helping factory managers to provide early warnings of abnormal power loads and avoid the risks of power overload. The data output module outputs the analysis and prediction results in a visualized form for relevant personnel to process and make decisions. The data publishing module implements power load overload early warning and fault diagnosis and analysis. It promptly issues alarms when abnormal power loads are detected, notifying relevant personnel to take appropriate measures to reduce the probability of equipment failure and improve system stability.

[0104] This invention improves the real-time performance, accuracy, and intelligence of power load monitoring through the organic collaboration of multiple modules, effectively reduces the risk of power system failures, and ensures the efficient and safe operation of the factory's power system, thus having significant practical application value.

[0105] Optionally, the specific method for monitoring the equipment load status in step S2 by collecting real-time data from sensors and actuators of factory equipment using industrial protocols includes:

[0106] Through an edge gateway industrial data acquisition terminal, real-time data is periodically acquired from factory equipment using the OPC UA industrial communication protocol, and the data undergoes preliminary verification. This process includes data format parsing and integrity checks to ensure that the data is not damaged during transmission and to meet the system's real-time requirements, ensuring that the data is delivered to subsequent processing units in a timely and reliable manner. The acquired data specifically includes equipment operation data and environmental data. The equipment operation data includes power parameters and status parameters. The power parameters include voltage, current, active power, reactive power, and power load. The status parameters include equipment on / off status, operating mode, and fault status. The environmental data includes environmental factors, such as indoor temperature, humidity, and air pressure in the factory.

[0107] Optionally, such as Figure 2As shown, step S3 involves constructing a 3D virtual model of the factory's actual layout, production lines, and equipment using digital twin technology based on the collected real-time data. The specific methods for mapping the real-time data onto the virtual model through a data interface to achieve dynamic simulation and interactive visualization of the model include:

[0108] S31. Using digital twin technology, a realistic 3D factory model is constructed using real-time data. Geographic information of the factory and layout data of production units are collected to generate a 3D spatial model of the factory. This model not only covers the actual layout of the production line but also the location and dimensions of each piece of equipment. By combining CAD drawings and other engineering design documents, detailed information about each piece of equipment is also incorporated into the virtual model, including the type, specifications, and functions of the equipment. It not only displays the structure and supports embedded attribute definitions but also includes a unique identifier, equipment model, and static metadata of the operation and maintenance cycle for each piece of electrical equipment. The key to this process is to map the real-time changes of the physical world to the virtual model using data collected by sensors, ensuring that the 3D model of the factory can dynamically reflect the operating status of the equipment and changes in the production environment.

[0109] S32. Real-time data is connected to the virtual model through an interface, displaying key parameters such as temperature, pressure, and power on the 3D model in real time to create a dynamic simulation effect. Users can operate the virtual model through the interactive interface, view detailed data of different equipment and production lines, and present data analysis results in various forms such as charts and heat maps. Administrators can view the operating status of each piece of equipment in real time.

[0110] Optionally, the specific method for using the differential autoregressive moving average (ARIMA) model to analyze and uncover potential power load variation trends for normal equipment operation described in step S4 is as follows:

[0111] The ARIMA model combines autoregressive (AR), differencing (I), and moving average (MA), and the formula is as follows:

[0112]

[0113] The autoregression section describes the current observed value X. t How to use the past p observations and the random error term ∈ t Decision, φ i Autoregressive coefficients, where c is a constant term;

[0114] Δ d X t =(1-M) n X t

[0115] Difference operations are used to transform non-stationary time series into stationary series; n represents the order of the difference, Δn X t It is an n-order difference time series, where M is the lag operator, (1-M). n This indicates a difference operation.

[0116]

[0117] The moving average represents the current observation value X. t From past error terms ∈ t The weighted sum of the lag error terms; μ represents the model mean, θ i It is the moving average coefficient, representing the impact of the past i error terms on the current value; ∈ t-i It is the error term from the past i time steps.

[0118] The ARIMA model is used to analyze whether the time series exhibits a monotonically rising or falling trend and to identify seasonal fluctuations in the time series. The autoregressive part is used to determine the long-term dependencies in the data, and the moving average part is used to determine whether noise in the data has a significant impact on the current observation.

[0119] Optionally, such as Figure 3 As shown, the specific method described in step S5, which involves applying the WTConv wavelet transform to the iTransformer model based on the power load change trend during normal equipment operation, to predict the power load data for the next hour, is as follows:

[0120] S51. Input Time Series Data: The input is historical time series data of the power load, usually expressed as:

[0121] X = [x1, x2, x3, ..., x T ]∈R T×d

[0122] Where T is the number of time steps and d is the feature dimension.

[0123] The power load time-series data is normalized using the following formula:

[0124]

[0125] Where x is the input power time series data, μ(x) is the mean of the input data, and σ(x) is the standard deviation of the input data.

[0126] S52, WTConv Frequency Band Decomposition: Power load time series data typically exhibits characteristics such as periodicity, trend, and long-range dependence. The Wavelet Transform as Convolutions is used to efficiently extract these low-frequency and high-frequency information, thereby improving the performance of the load forecasting model. Based on a time series x of length L, the formula for this process is as follows:

[0127]

[0128] Here, WT(x) is the wavelet transform operation, and Convld(x, filters) is the one-dimensional convolution operation, where x is the input signal and filters are the convolution kernels.

[0129] After wavelet transform, the sequence is decomposed into multiple wavelet sub-bands, representing information from different frequency bands: x L ,x H , where x L For low-frequency components, x H High-frequency components; then for x L ,x H F is obtained by extracting local features using 1D convolution. L =Conv1D(X L ), F H =Conv1D(X H ).

[0130] S53, Low-frequency and high-frequency embeddings: Through learnable linear transformations, multi-scale features are projected onto a feature space of the same dimension, facilitating subsequent fusion and modeling, where W... L W H ∈R d'×d It is the projection matrix:

[0131] E L =F L ·W L +b L E H =F H ·W H +b H

[0132] Among them, E L E H The embedding result is the low-frequency and high-frequency components after linear transformation; b L and b H This is the bias term for low-frequency and high-frequency characteristics.

[0133] Then, the two are concatenated to form the final frequency component E = Concat(E L E H The low-frequency and high-frequency features of different frequency components after WTConv processing are embedded as variable tokens of the iTransformer prediction model. In the embedding stage, each frequency band component is embedded into an independent token to maintain the independence of each frequency band information.

[0134] The iTransformer basic model proposes a reverse innovation approach through a redesigned architecture, the core of which lies in redefining the sequence embedding method with variables at its center; in iTransformer, based on the input time series X... ：,n Information is passed layer by layer through the Transformer module TrmBlock, and the final prediction result is... The process is as follows (using the formula):

[0135]

[0136] H l+1 =TrmBlock(H l ), l=0,...,l-1,

[0137]

[0138] It is the initial representation of the nth time series after passing through the embedding layer; embedding represents the embedding operation; H l+1 It is the output of the Transformer block at layer (l+1), (H l ) is the output of the previous layer. The final output contains the context vector associated with the nth variable, where C represents the module layer number. The predicted feature representation is input into the Projection layer.

[0139] It contains N embedding tokens with dimension D; iTransformer embeds the input sequence into variable tokens and calculates the correlation between variables through a self-attention mechanism; a feedforward network is used on each variable token to focus on extracting nonlinear features of the time series.

[0140] S54. Multi-head Self-Attention Mechanism: The main stage of power load forecasting adopts the iTransformer model, and the multi-head self-attention mechanism is used to capture the dependencies between multiple variables in each time series; iTransformer can establish the correlation between variables in different frequency bands, improving the accuracy of the model in time series forecasting; firstly, according to the input E = Concat(E L E H Generate query, key, and value vectors:

[0141] Q = E·W Q K = E·W K V = E·W V ;

[0142] Where Q is the query vector, K is the key vector, V is the value vector, and W is the value vector. Q W K W V These are learnable weight matrices, used to map the input features E to the query, key, and value spaces, respectively.

[0143] In the attention mechanism, low-frequency tokens are given higher weights to emphasize long-term trends, while high-frequency tokens are given lower weights to preserve detailed features. The attention calculation method is as follows:

[0144]

[0145] Among them, Q w ,K w V w These represent the query, key, and value, respectively, and are all vectors within a local window w. k It is the vector dimension;

[0146] Multi-head concatenation and linear mapping: MHSA(E) = Concat(head1,...,head) H )·W O .

[0147] S55. Feature Aggregation: To enhance stability and generalization ability, residual connections and layer normalization are used, expressed as:

[0148] Z = LayerNorm(E + MHSA(E))

[0149] The features of each frequency band are recombine into a whole feature using the Inverse Discrete Wavelet Transform (IDWT): Z agg =IDWT(Z).

[0150] S56. Feedforward Network: The ReLU activation function in the feedforward network introduces non-linearity into the model. The addition of the activation function enables the feedforward network to learn more complex and expressive features; the overall features Z... agg The input is a feedforward network, with a two-layer MLP structure, as shown in the following formula:

[0151] h1 = ReLU(z) agg ·W1+b1)

[0152] h2 = h1·W2 + b2

[0153] Where W1 and W2 are weight matrices, and b1 and b2 are bias terms. W1 and b1 are used to transform the input to a higher-dimensional space and then perform a non-linear transformation through an activation function; W2 and b2 are used to map the expanded features back to the original dimension to ensure consistency with the input dimension of the model.

[0154] S57. Generate load forecast results: The output is the predicted load value for the next time step. The system predicts the power load data for the next hour and stores the prediction results in the system.

[0155] Optionally, the specific method described in step S5 for applying WTConv wavelet transform to the iTransformer model based on the power load change trend during normal equipment operation to predict power load data for the next hour is as follows:

[0156] Optionally, such as Figure 4 As shown, the specific method for assessing the load status based on the prediction results in step S7, triggering an overload warning when the predicted value exceeds a preset safety threshold, and simultaneously performing diagnostic analysis of potential faults by combining historical fault data includes:

[0157] S71, Power Load Overload Warning: Based on predicted data, an early warning is issued when the load is about to become overloaded, reminding the equipment administrator to handle the situation in a timely manner.

[0158] S72. Fault Diagnosis and Analysis: Through data analysis and historical fault data, determine the cause of the fault and provide steps and suggestions for repairing the fault.

[0159] Optionally, the specific method for issuing an early warning based on the predicted data when the load is about to be overloaded, reminding the equipment administrator to handle the situation in a timely manner, in step S71 is as follows:

[0160] By combining the load data predicted in previous steps for the next hour with historical data and load change trends, abnormal patterns of load growth are identified. Load change rate: The rate of load growth is determined by calculating the ratio of the current load to the load at the previous moment, as shown in the following formula:

[0161] ΔL t =L t -L t-1

[0162] L t For the load data at the current moment, L t-1 The load data from the previous moment, ΔL t This represents the change in load.

[0163] The average load calculated based on historical data is μ. L The standard deviation is σ L The safe load threshold can be set as follows:

[0164] L th =μ L +α·σ L

[0165] Here, α is the safety factor, which can be adjusted by managers. It is usually set to 2, which indicates the situation where the load exceeds the average load by a certain standard deviation. Based on the comparison between the predicted load value and the set threshold, if the predicted value exceeds the safe load threshold, the system will issue an overload warning signal.

[0166] Optionally, in step S72, the cause of the fault is determined through data analysis and historical fault data, and specific methods for fault repair steps and suggestions are provided:

[0167] By analyzing data and historical fault data, a trained classification model is used to identify different fault types; a Bayesian network is used to determine the conditional probability P(Cause) based on the fault. i The Fault (or similar function) infers the most likely root cause and provides steps and suggestions for fixing the fault, using the following formula:

[0168]

[0169] Wherein, P(Cause) i |Fault) represents the probability of a certain cause of failure occurring given that a failure has occurred; P(Fault|Cause) i P(Cause) represents the probability of a failure occurring given that a certain cause of failure has occurred; i P(Fault) represents the prior probability of a fault occurring; P(Fault) represents the total probability of a fault occurring.

[0170] With the support of historical data and failure cases, the system can also generate experience-based repair suggestions.

[0171] The power load forecasting method in step S5 was tuned and tested using the same dataset. Multiple evaluation metrics were used to assess the performance of the forecasting model, such as mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (RMAE), and coefficient of determination (R²). 2 ).

[0172] The test results are as follows:

[0173]

[0174] A comparative experiment was conducted on the ETT (Electricity Transformer Temperature) open-source electricity load dataset. The experimental results are shown in the table below:

[0175]

[0176] In this power load forecasting experiment, the method in step S5 of the invention outperformed existing mainstream models in several key performance indicators, demonstrating superior forecasting ability and stability; specifically, the model of the present invention outperforms existing mainstream models in terms of the coefficient of determination (R²). 2 The accuracy of the proposed method is slightly higher than that of iTransformer, indicating that it fits the actual load change trend better. In terms of error indicators, the mean square error, root mean square error, and mean absolute error of the proposed method are all better than other comparative models, showing that it has higher prediction accuracy. The mean absolute percentage error (MAPE) is 0.970%, which fully demonstrates that the method can still maintain high relative accuracy in multi-period and multi-fluctuation scenarios, and has good generalization ability and robustness. The proposed step S5 method has stronger fitting ability, lower prediction error and higher relative accuracy in power load prediction tasks, and can effectively cope with complex load change trends, and has broad engineering application prospects.

[0177] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for monitoring and early warning of factory power load based on digital twins, characterized in that, The load monitoring and early warning method includes the following steps: S1. Provides stable power support for the entire power load monitoring and early warning system through the power supply module; S2. Collect real-time data from sensors and actuators through industrial protocols to complete equipment load status monitoring; S3. Based on the collected real-time data, a three-dimensional virtual factory model is constructed using digital twin technology and factory data information. The real-time data is then mapped onto the virtual model through a data interface to achieve dynamic simulation and interactive visualization of the model. S4. The differential autoregressive moving average model (ARIMA) is used to analyze the collected real-time data and uncover potential power load change trends for normal equipment operation. S5. Based on the potential power load change trend of normal equipment operation, apply WTConv wavelet transform to iTransformer model to predict power load data for the next hour. S6. Output and visualize the prediction results; S7. Identify overload events by monitoring and predicting data, trigger early warnings, and perform fault diagnosis and analysis; The specific method described in step S5, which involves applying the WTConv wavelet transform to the iTransformer model based on the potential power load change trend of normal equipment operation, to predict the power load data for the next hour, is as follows: S51. The input is historical time-series data of power load, represented as follows: ； Where T is the number of time steps. It is the feature dimension; The power load time-series data is normalized using the following formula: ； Where x is the input power time series data, It is the mean of the input data. It is the standard deviation of the input data; S52. The wavelet transform as convolutions is used to extract low-frequency and high-frequency information from the power load time series data; based on a time series x of length L, the formula for this process is as follows: ； Where WT(x) is the wavelet transform operation, and Convld(x, filters) is the one-dimensional convolution operation, where x is the input signal and filters are the convolution kernels; After wavelet transform, the sequence is decomposed into multiple wavelet sub-bands, representing information from different frequency bands: ,in Low-frequency components, High-frequency components; then for Local features were extracted using 1D convolution respectively. ; S53. Through learnable linear transformations, multi-scale features are projected onto a feature space of the same dimension, where... It is the projection matrix: ； Among them, E L E H The embedding result is the low-frequency and high-frequency components after linear transformation; b L and b H These are bias terms for low-frequency and high-frequency characteristics; Then, the two are spliced ​​together to form the final frequency components. The low-frequency and high-frequency features of different frequency components after processing by WTConv are embedded into the variable tokens of the iTransformer prediction model. In the embedding stage, each frequency band component is embedded into an independent variable token to maintain the independence of information in each frequency band. The iTransformer basic model proposes a reverse innovation approach through a redesigned architecture, the core of which lies in redefining sequence embedding around variables; in iTransformer, based on the input time series... Through the Transformer module Information is passed layer by layer, and the final prediction result is... The process is as follows: ； It is the initial representation of the nth time series after the embedding layer; Indicates an embedding operation; It is the output of the Transformer block at layer l+1. It is the output of the previous layer; This is the context vector related to the nth variable in the final output, where C represents the module level. As the feature representation for prediction, input into layer; It contains N embedding tokens with dimension D; iTransformer embeds the input sequence into variable tokens and calculates the correlation between variables through a self-attention mechanism; a feedforward network is used on each variable token to focus on extracting non-linear features of the time series; S54. The main stage of power load forecasting adopts the iTransformer model, and the multi-head self-attention mechanism is used to capture the dependencies between multiple variables in each time series. First, based on the input Generate query, key, and value vectors: ； Where Q is the query vector, K is the key vector, V is the value vector, and W is the value vector. Q W K W V These are learnable weight matrices, used to map the input features E to the query, key, and value spaces, respectively. In the attention mechanism, low-frequency variable tokens are assigned higher weights, while high-frequency variable tokens are assigned lower weights; the attention calculation method is as follows: ； in, These represent the query, key, and value, respectively, and are all local windows. Vectors within, It is the vector dimension; Multi-head splicing and linear mapping: ; S55. Using residual connectivity and layer normalization, it is expressed as follows: ； The features of each frequency band are recombined into a whole feature by using the Inverse Discrete Wavelet Transform (IDWT): ; S56. The above overall features The input is a feedforward network, which has a two-layer MLP structure, as shown in the following formula: ； Where ReLU is the activation function. , It is a weight matrix. , It is a bias term; S57, Output is the predicted load value for the next time step. It predicts the power load data for the next hour and stores the prediction results in the system.

2. The factory power load monitoring and early warning method based on digital twins according to claim 1, characterized in that, The specific method for collecting real-time data from sensors and actuators via industrial protocols to complete equipment load status monitoring in step S2 includes: The industrial data acquisition terminal of the edge gateway periodically acquires real-time data from factory equipment using the OPC UA industrial communication protocol and performs preliminary verification of the real-time data. The collected real-time data specifically includes equipment operation data and environmental data; The equipment operation data includes power parameters and status parameters. The power parameters include: voltage, current, active power, reactive power, and power load. The status parameters include: equipment on / off status, operating mode, and fault status. The environmental data includes environmental factors, such as indoor temperature, humidity, and air pressure in the factory.

3. The factory power load monitoring and early warning method based on digital twins according to claim 1, characterized in that, The specific methods described in step S3, which involve constructing a three-dimensional virtual model based on the collected real-time data using digital twin technology and factory data information, and mapping the real-time data onto the virtual model through a data interface to achieve dynamic simulation and interactive visualization of the model, include: S31. Using digital twin technology, collect the factory's geographic information, production unit layout data, and equipment locations through sensors and engineering design documents; The geographic information, the production unit layout data, and real-time equipment operation data are integrated to generate a three-dimensional virtual factory model. Static metadata of the equipment is embedded in the three-dimensional virtual factory model. The static metadata includes equipment type, specifications, unique identifier, and maintenance cycle. A mapping relationship between the equipment location and virtual nodes in the three-dimensional virtual factory model is established based on the unique identifier. S32. Connect the real-time data to the virtual model through an interface, and display the key parameter information on the three-dimensional model in real time to form a dynamic simulation effect.

4. The factory power load monitoring and early warning method based on digital twins according to claim 1, characterized in that, The specific method described in step S4 for using the differential autoregressive moving average (ARIMA) model to uncover potential power load variation trends for normal equipment operation is as follows: The ARIMA model combines autoregressive (AR), differencing (I), and moving average (MA), and the formula is as follows: ； The autoregression section describes the current observations. How from the past Each observation and random error term Decide, Autoregressive coefficient, For constant terms; ； Difference operations are used to convert non-stationary time series into stationary series; n represents the order of the difference. It is an nth-order difference time series. It is a lag operator. Indicates the difference operation; ； The moving average represents the current observation value. From past error terms The weighted sum of the lag error terms constitutes the equation; The mean of the model. It is the moving average coefficient, representing the past... The impact of each error term on the current situation; It's the past Error term at each time point.

5. The factory power load monitoring and early warning method based on digital twins according to claim 1, characterized in that, The specific method for identifying overload events by monitoring and predicting data, triggering early warnings, and performing fault diagnosis and analysis as described in step S7 includes: S71. Based on the predicted data, issue an early warning when the load is about to be overloaded to remind the equipment administrator to handle it in time; S72. By analyzing data and historical fault data, determine the cause of the fault and provide steps and suggestions for repairing the fault.

6. The factory power load monitoring and early warning method based on digital twins according to claim 5, characterized in that, The specific method for issuing an early warning based on predicted data when the load is about to be overloaded, reminding the equipment administrator to take timely action, as described in step S71: The load change rate is used to determine the rate of load growth by calculating the ratio of the current load to the load at the previous moment. The formula is as follows: ； This is the load data at the current moment. This is the load data from the previous moment. This refers to the change in load. The average load calculated based on historical data is The standard deviation is The safe load threshold is set as follows: ； Where α is the safety factor, representing the situation where the load exceeds the average load by a certain standard deviation; The system compares the predicted load value with the set threshold. If the predicted value exceeds the safe load threshold, the system will issue an overload warning signal.

7. The factory power load monitoring and early warning method based on digital twins according to claim 5, characterized in that, In step S72, the cause of the fault is determined through data analysis and historical fault data, and specific methods and suggestions for repairing the fault are provided: Using Bayesian networks, we can apply conditional probabilities The optimal root cause is deduced, and steps and suggestions for fixing the fault are provided, as shown in the following formula: ； in, It represents the probability of a certain cause of failure occurring under the given conditions of failure occurrence; It represents the probability of a failure occurring given that the cause of the failure is known. This represents the prior probability of a particular cause of failure occurring. This represents the total probability of a failure occurring.

Images