Factory load dynamic prediction method and system

Abstract

The invention relates to the technical field of industrial big data analysis and energy management, and provides a factory load dynamic prediction method and system, and the method comprises the following steps: collecting the operation state data of key equipment in a factory in real time, predicting the remaining service life of the equipment, and calculating the health index of a current time point; dynamically correcting the reference load predicted value of the equipment to obtain a real load predicted value; knowledge migration is carried out on the target plant area to construct a load prediction model of the target plant area; performing real-time compensation on a load prediction result based on random event simulation; and when an event is detected, triggering a Gaussian process regression model to calculate a load deviation delta P, and dynamically adjusting an event sensitivity coefficient by utilizing reinforcement learning. According to the method, the accuracy problem of long-term prediction is solved through equipment recession modeling, the problem of data scarcity of a new plant area is solved through transfer learning, and the problem of short-term sudden disturbance is solved through random event simulation.

Description

Technical Field

[0001] This invention belongs to the field of industrial big data analysis and energy management technology, and specifically relates to a method and system for dynamic prediction of plant load based on dynamic assessment of equipment health status, data migration learning and random event simulation. It is applicable to accurate load prediction and energy dispatching in process industries and discrete manufacturing industries. Background Technology

[0002] Plant load forecasting is a key technology for industrial energy management, production scheduling, and power optimization, and its accuracy directly impacts enterprise operating costs, equipment reliability, and grid stability. However, existing technologies have significant shortcomings:

[0003] Static models cannot adapt to the dynamic degradation of equipment: During long-term operation, industrial equipment (such as motors and compressors) experiences a non-linear decline in its actual load capacity due to mechanical wear, insulation aging, and efficiency reduction. Existing predictive models (such as neural networks or regression models trained on historical data, like the elastic assessment model described in patent application CN117272850A) are mostly static or use only a fixed linear decay coefficient for correction, failing to capture and adapt in real time to this load characteristic drift caused by changes in equipment health. This leads to the predictive model gradually becoming inaccurate as the equipment ages, with long-term prediction errors continuously increasing. For example, a compressor with a rated power of 250kW may have its actual maximum load capacity reduced to below 200kW at the end of its lifespan, but a static model still predicts 250kW, resulting in a continuous overestimation.

[0004] Data-driven models face the cold-start challenge: for newly built factories or new production lines, there is a lack of sufficient historical data for training data-driven models such as deep learning. Although transfer learning has been used to address this issue, existing methods, when applied across different plants, suffer from negative transfer due to significant differences in equipment models, process parameters, production rhythms, and other factors, leading to a decrease in prediction accuracy. Current technologies lack a fine-grained transfer mechanism capable of evaluating and matching feature similarity at different levels (equipment, process, plant-wide), thus hindering accurate knowledge reuse.

[0005] The impact of random events lacks quantification and adaptive mechanisms: Random events such as urgent orders, sudden equipment failures, and changes in raw material properties are common in production processes, and these events can instantly disrupt the stable load patterns. Existing research (such as patent application CN118432114A) focuses on random events, but primarily emphasizes macro-level load component identification or post-event analysis, lacking an automated simulation and compensation mechanism for plant production scheduling that can capture events in real time and quickly quantify their load impact. Existing methods often rely on expert experience to manually set disturbance ranges, resulting in slow response times (typically several hours) and an inability to self-learn and optimize based on feedback, leading to rigid performance of the prediction system in the face of sudden events.

[0006] In summary, existing technical solutions often address the three types of problems—device degradation, data scarcity, and random events—in a fragmented manner, lacking an integrated solution. Summary of the Invention

[0007] To address the shortcomings of existing technologies, a method and system for dynamic forecasting of plant load is provided, which can adapt to equipment degradation, efficiently utilize heterogeneous data, and respond to random events in real time.

[0008] To achieve the above objectives, in a first aspect, the present invention proposes a method for dynamic prediction of plant load, comprising the following steps:

[0009] S1: Real-time acquisition of operating status data of key equipment in the plant area, including current harmonic distortion rate (THD), vibration value (Vibration), and winding temperature, and extraction of characteristic frequency band energy of vibration signal using wavelet packet decomposition (WPD).

[0010] S2: Input the data processed in step S1 into a Long Short-Term Memory (LSTM) network to predict the remaining useful life (RUL) of the device and calculate the health index HI(t) at the current time point.

[0011] S3: Based on the health index HI(t), predict the baseline load value P of the equipment. base (t) is dynamically corrected to obtain the true load forecast value P that takes into account equipment degradation. real (t);

[0012] S4: Based on a three-level transfer learning architecture, using source plant data, knowledge is transferred to the target plant through three levels: equipment level, process level, and plant area level, in order to build a load prediction model for the target plant.

[0013] S5: Based on random event simulation, real-time compensation is performed on load prediction results; a three-dimensional classification knowledge base (EID) of event-impact-duration is constructed; events in the manufacturing execution system or data acquisition and monitoring control system (SCADA) are monitored through the OPC UA interface; when an event is detected, the Gaussian process regression (GPR) model is triggered to calculate the load deviation ΔP, and the event sensitivity coefficient is dynamically adjusted using reinforcement learning (RL).

[0014] Furthermore, the specific formula for the dynamic correction described in step S3 is as follows:

[0015]

[0016] in, The environmental degradation factor is calculated using the following formula:

[0017]

[0018] In the formula, The current harmonic distortion influence coefficient is obtained by regression analysis of the correlation between THD and load capacity in historical data. The vibration influence coefficient is obtained by analyzing the correlation between vibration values ​​and load capacity in historical data through regression analysis; THD(t) is the measured value of total harmonic distortion of current at time t; Vibration(t) is the measured value of effective vibration velocity at time t.

[0019] The health index HI(t) reflects the trend of permanent performance degradation caused by long-term aging of equipment, while This reflects the instantaneous performance fluctuations caused by the current operating conditions (power quality, mechanical condition). Multiplying these two values ​​by the baseline prediction achieves a dual dynamic correction to the equipment load capacity. The max function ensures that the predicted value will not be a non-physical negative value, enhancing the model's robustness. This formula transforms the abstract health state into a concrete, calculable power attenuation, enabling the prediction model to adaptively track changes in equipment performance.

[0020] Furthermore, the dynamic weighted domain adaptation algorithm used in step S4 for plant-level migration has the following formula for calculating the weight ω:

[0021]

[0022] in, is the cosine similarity between source plant feature A and target plant feature B, and 3.0 is the temperature coefficient used to control the steepness of the weight distribution.

[0023] This formula is essentially a variant of the Softmax function, used to map cosine similarity to weights between (0, 1). The temperature coefficient T = 3.0 is a key hyperparameter that controls the sensitivity of the weights to changes in similarity. A larger T value results in a steeper weight distribution, amplifying small differences in similarity and making the transfer strategy more selective; a smaller T value results in a flatter distribution and a more lenient transfer strategy. By adjusting the T value, the transfer effect can be optimized, avoiding negative transfer. This method achieves automated, non-linear allocation of transfer weights, offering better adaptability than simple linear interpolation or fixed threshold methods.

[0024] Furthermore, the formula for calculating the load deviation ΔP in step S5 is as follows:

[0025]

[0026] in, Total load deviation, in kilowatts (kW); The weight coefficient for the k-th type of event is obtained through statistical analysis of historical data. For the k-th type of event The corresponding load change function is calibrated through regression analysis.

[0027] This formula provides a modular, interpretable framework for quantifying event impact. It decomposes the impact of a complex event into two parts: and . Regression analysis is used to learn the nonlinear relationship between events (such as duration and intensity) and load changes. The overall impact of this type of event is obtained through historical statistics. This decomposition reduces the difficulty of model learning and improves the generalization ability under small sample events. At the same time, the formula allows for the linear superposition of the combined effects of multiple concurrent events, which is consistent with the assumption of the independence of disturbance sources in most industrial scenarios.

[0028] Furthermore, step S5 includes dynamically adjusting the weight coefficients using reinforcement learning. Its update formula is:

[0029]

[0030] in, The learning rate; Let be the reward signal obtained at time t, defined as the negative value of the mean squared error (MSE) within the event window, i.e. ; For state Take action below (i.e., adjustment) The state-action value function.

[0031] Introducing reinforcement learning (RL) transforms the event compensation system from a static system into a dynamic, self-learning system. The RL agent (this system) continuously interacts with the environment (the real factory), trying different... (Action), and receive changes in prediction accuracy as a reward signal. Through the above update formula (a form of policy gradient method), the agent gradually learns what... The system is configured to maximize long-term rewards (i.e., highest prediction accuracy) under specific conditions. This allows the system to adaptively optimize compensation strategies to address the impact of new patterns or changing events that have not appeared in historical data.

[0032] Secondly, the present invention also proposes a dynamic load prediction system for a plant area to implement the above-mentioned method, the system comprising:

[0033] The data acquisition and preprocessing module is used to acquire the current, vibration, and temperature data of the equipment in real time, and to perform wavelet packet decomposition processing on the vibration signal.

[0034] The equipment health status assessment module, with a built-in LSTM network, is used to predict the remaining useful life RUL(t) and calculate the health index HI(t).

[0035] The load forecasting dynamic correction module receives the health index and baseline load forecast values, and calculates the actual load forecast value P. real (t) or load deviation ΔP;

[0036] The three-level transfer learning module is used to perform equipment-level, process-level, and plant-level transfer learning. Its plant-level transfer learning adopts a dynamic weighted domain adaptation algorithm.

[0037] The random event simulation and compensation module is used to compensate for load forecast results in real time; construct a three-dimensional classification knowledge base of event-impact-duration; monitor events in manufacturing execution systems or data acquisition and monitoring control systems; trigger Gaussian process regression models to calculate load deviation ΔP, and dynamically adjust the event sensitivity coefficient using reinforcement learning; and

[0038] The central processing and control module is used to coordinate and control the operation of each module and output the final load dynamic prediction results.

[0039] This system materializes each step of the methodology into functionally decoupled but data-coupled software modules, forming a complete end-to-end solution. The modular design facilitates system development, deployment, and maintenance. The modules are coordinated through a central processing and control module, realizing an automated pipeline from data acquisition to predictive output. It possesses excellent engineering feasibility and system integration capabilities, laying the foundation for the industrial application of the technology.

[0040] Furthermore, the data acquisition and preprocessing module is deployed on the edge computing node of the factory area, while the remaining modules are deployed on the cloud server.

[0041] A heterogeneous cloud-edge collaborative architecture is adopted. Tasks are rationally allocated based on the real-time nature of data processing and computational load. The edge, located close to the data source, is responsible for real-time acquisition of high-frequency data and lightweight, low-latency preprocessing, effectively reducing network bandwidth pressure and data transmission latency. The cloud handles complex model training, large-scale historical data storage, and global optimization calculations, fully utilizing the powerful computing and storage resources of the cloud. This architecture ensures real-time performance while also meeting the computational resource demands of complex algorithms, making it a preferred architecture for industrial internet applications.

[0042] Furthermore, the random event simulation and compensation module integrates an FPGA-based hardware accelerator to accelerate the inference calculation of the Gaussian process regression model.

[0043] The computational complexity of Gaussian process regression (GPR) models increases cubically with the amount of training data, which may not meet high real-time requirements (such as millisecond-level response). Using FPGAs to accelerate large-scale matrix operations in the GPR inference process can significantly improve computation speed, ensuring that the system can quickly output compensation results even when processing high-frequency, complex events, thus meeting the needs of high real-time industrial applications.

[0044] Furthermore, the process-level transfer in the three-level transfer learning module is implemented using a graph neural network (GNN), whose node feature vectors include equipment power, energy consumption, and runtime, and the edge weights represent the intensity of material flow or energy flow.

[0045] Abstracting the production process into a graph structure, where nodes represent equipment and edges represent process relationships, closely reflects the actual conditions of industrial sites. Generative Neural Networks (GNNs) can efficiently process this non-Euclidean data, capturing the interactions between equipment through message passing mechanisms, thereby learning deep, process-logic-based load correlation patterns—something traditional machine learning models struggle to achieve.

[0046] Furthermore, the system also includes a visual human-computer interaction interface for displaying equipment health status trends, load forecast curves, event alarm logs, and transfer learning effect evaluation dashboards.

[0047] Visualized human-computer interfaces provide a good interface for human-computer interaction, presenting complex algorithm results to users (such as factory energy managers) in a graphical and intuitive way. This not only enhances the usability and interpretability of the system, but also makes it easier for users to monitor the system status, trust the prediction results, and make necessary manual interventions, thereby better integrating the system into actual industrial management processes.

[0048] The beneficial effects of this invention are: through the closed-loop execution of the above five steps, this invention comprehensively solves the three major problems raised in the background art. Its core principle lies in constructing a dynamic feedback system of perception, evaluation, prediction, transfer, compensation, and learning. Online monitoring data (S1) provides the system with sensing capabilities; equipment health status assessment (S2), based on the powerful processing capabilities of LSTM networks for time-series data, quantifies the abstract equipment status into a health index HI(t), providing a scientific basis for subsequent corrections; dynamic load forecast correction (S3) combines HI(t) with real-time operating parameters, and through a mathematical model with clear physical meaning, achieves dual dynamic decay of the baseline forecast value, fundamentally solving the model inaccuracy problem caused by equipment aging; three-level transfer learning (S4), through feature alignment and knowledge reuse at the equipment, process, and plant levels, effectively solves the problem of data scarcity in new plant areas, and its dynamic weighting mechanism avoids negative transfer; random event simulation (S5), through the EID knowledge base and GPR model, transforms unpredictable discrete events into calculable load deviations, and introduces reinforcement learning for online policy optimization, enabling the system to have adaptive capabilities to cope with uncertainty. These three elements are tightly coupled through data flow, forming a highly adaptable and robust intelligent forecasting system. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the dynamic load prediction method for the plant area according to the present invention. Detailed Implementation

[0050] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] Name Explanation

[0052] HI(t): Health index function, representing the health status of the device at a specific time t. It is a variable that changes with time t, and its range is [0, 1], where 0 represents complete failure and 1 represents an ideal health state. The calculation formula is as follows: .

[0053] HI: Health Index, is the core concept and name used in this method to quantify the health status of equipment. In the formulas and specific calculations, the specific time point must be specified, i.e., HI(t) must be used.

[0054] RUL(t): Remaining Useful Life function, representing the predicted remaining lifespan of the equipment at a specific time t. It is a variable, and the unit is hours (h). Remaining Useful Life (RUL): refers to the expected length of time from the current moment when the equipment is in operation until it fails.

[0055] Rated service life of equipment is a constant for the equipment, expressed in hours (h), and remains unchanged throughout the equipment's lifespan.

[0056] Pbase(t): Baseline load forecast function, representing the load power value calculated by the forecast model at a specific time t, assuming the equipment is in ideal health condition, in kilowatts (kW). It is a variable.

[0057] Preal(t): The actual load prediction function, representing the final output load power prediction value after correction by the method of this invention at a specific time t, in kilowatts (kW). It is a variable.

[0058] THD(t): The total harmonic distortion rate of the current at time t is a measured value. It is an instantaneous variable, expressed as a percentage (%), and is used to characterize the power quality at the current moment.

[0059] The harmonic distortion (THD) factor is a dimensionless constant obtained through regression analysis of historical data. It represents the average attenuation of equipment load capacity caused by a 1% increase in THD.

[0060] Vibration(t): The effective value of vibration velocity at time t, which is an instantaneous variable with the unit of millimeters per second (mm / s), used to characterize the mechanical vibration state of the equipment at the current moment.

[0061] The vibration influence coefficient is a dimensionless constant obtained through regression analysis of historical data. It represents the average attenuation of the equipment's load capacity caused by each 1 mm / s increase in vibration value.

[0062] Cosine similarity: A metric that measures the similarity in direction between two vectors. Its value ranges from -1 to 1; the closer the value is to 1, the more similar the two vectors are in direction. The calculation formula is as follows: .

[0063] Domain adaptation: A type of transfer learning that aims to apply knowledge learned from one domain (source domain) to another different but related domain (target domain).

[0064] Gaussian process regression (GPR): a nonparametric probabilistic model suitable for handling small sample sizes and nonlinear regression problems, and can provide estimates of the uncertainty of predictions.

[0065] Reinforcement Learning (RL): A machine learning paradigm in which an agent learns the optimal strategy to achieve a goal by interacting with the environment and receiving rewards.

[0066] This embodiment takes the load prediction of an air compressor station in a chemical plant as an example.

[0067] I. Preliminary Preparations

[0068] On the cloud server, we deployed the core brain program. This includes:

[0069] A pre-trained LSTM network model. This model was trained using equipment data (including current, vibration, and temperature) from a factory area over three years, specifically designed to predict the remaining lifespan of equipment.

[0070] A Gaussian process regression (GPR) model. This model was trained using historical event records, such as how the load changed during the past 100 "feed change" events.

[0071] A transfer learning model includes programs for comparing equipment similarity and graph neural networks for analyzing the similarity of process flow diagrams.

[0072] In the edge computing box at the factory site, a data acquisition program was deployed. It was set to collect current harmonics, vibration (high-speed acquisition, 10,000 points per second), and temperature data once per second. The box also has a built-in wavelet packet decomposition algorithm that can analyze vibration signals in real time.

[0073] We've pre-built an "Emergency Event Handbook" (EID Knowledge Base) in the database. For example, one entry is recorded like this:

[0074] Event ID: GAS001

[0075] Event type: Raw material batch change

[0076] Typical impact: It will cause a decrease in the total load of the entire plant, with an average decrease of about 15%, and the fluctuation is not significant. It usually lasts for 4 to 12 hours.

[0077] Triggering condition: This event is considered to have occurred when the batch number in the MES system changes and the opening degree of a critical valve on the production line exceeds 50%.

[0078] Configure data monitoring: Enable our system to listen for OPC UA messages published by the MES system in real time.

[0079] II. A Day in the Life of the System

[0080] 9:00 AM (S1, Data Acquisition and Preprocessing):

[0081] (Data Acquisition): The edge box on-site acquired the following data from air compressor #1 at this moment: current harmonic distortion (THD) = 4.5%, original vibration signal, temperature 75.2℃. The box immediately performed wavelet packet analysis on the vibration signal and calculated the energy value of the frequency band (1250-2500Hz) to be 0.015 joules (J). Then it sent the data packet {Time: 09:00, THD: 4.5, Vibration Energy: 0.015, Temperature: 75.2} to the cloud.

[0082] 9:00 AM + 0.5 seconds (S2, Health Status Assessment):

[0083] After receiving the data, the cloud-based equipment health assessment module retrieves all data from the air compressor between 8:00 and 9:00 AM, compiles it into a 12-item sequence, and feeds it into the LSTM model. The model outputs a RUL prediction value: RUL(t) = 51800 hours, meaning the equipment can likely operate healthily for another 51800 hours. Its design life is known. =60,000 hours, then its current Health Index (HI) = 51800 / 60000 = 0.863. This means that its overall capability is 86.3% of its original state.

[0084] 9:00:01 AM (S3, Load Forecast Dynamic Correction):

[0085] The baseline load forecasting module (ARIMA algorithm) outputs the baseline load value P under ideal healthy conditions based on historical load curves and current operating conditions. base (t)=201.5kW. The correction module reads HI(t)=0.863 and real-time data THD=4.5 and Vibration=0.07 from the database (the vibration velocity value is uploaded by the edge module).

[0086] Calculate the environmental degradation factor: known coefficients =0.02, =0.05, calculate the environmental impact factor. =0.9065.

[0087] Then calculate the final corrected load value P. real (t) = 201.5kW × 0.863 × 0.9065 ≈ 157.7kW. This 157.7kW is the predicted final output value of air compressor #1 at the current moment.

[0088] 9:30 AM (S5, Random Event Handling: Triggered)

[0089] (Unexpected event): The MES system suddenly issued a message: "Batch number changed to 247". The OPC UA client of the cloud event simulation module detected the change in this node value in real time and immediately captured this event message.

[0090] (Event Matching): The module parses the message and matches it against the "Emergency Event Manual" (EID knowledge base), finding a match for the "Raw Material Batch Change" (GAS001) event. Upon further inspection, it discovers that the valve opening on the production line is 65%, meeting the triggering conditions. Thus, the "Raw Material Batch Change" event is officially triggered.

[0091] 9:30 AM + 100 milliseconds (S5, Random Event Handling: Compensation)

[0092] The event simulation module triggers the GPR model for inference. The model inputs are: event type = GAS001, expected duration = 8h (sampled from a uniform distribution in the EID library), and instantaneous total load before the event = 6520kW;

[0093] The GPR model outputs a predicted load deviation of ΔP = -728kW. (Calculation principle: ,initial =1.0, (Calculated from GPR).

[0094] The deviation value of -728kW was immediately sent to the central processing module. The central processing module overlaid it with the current baseline load forecast to generate a corrected forecast curve showing a decrease of 728kW over the next 8 hours.

[0095] The dispatchers could see almost in real time that the forecast curve jumped down due to this sudden event.

[0096] 5:30 PM (S5, Random Time Processing: Learning):

[0097] The event has ended, and the system reviews its predictive performance during the event's duration: it examines the difference between the predicted and actual loads over the 8 hours from 9:30 AM to 5:30 PM. It calculates the mean squared error between the predicted and actual values ​​within this window; if the actual load drop is greater than predicted (e.g., a drop of 750kW), it silently records this: "Next time a similar event occurs, the compensation should be increased." Through an internal self-learning algorithm (such as reinforcement learning), the reinforcement learning module processes the experienced state (event type, duration, initial load), actions (using W...)... k=1.0), and the reward (R(t) =-120.5) are stored in the experience replay buffer. The RL model initiates a learning update, fine-tuning the policy network parameters using the policy gradient algorithm. Final decision: Set the weight coefficient W of event GAS001... k The value was adjusted from 1.00 to 1.03. This way, when a similar event is triggered again, the predicted impact will be 1.03*(-728kW)≈-750kW, thus more accurately approximating the actual impact.

[0098] The next morning at 9:00 AM (Application of S4 transfer learning; continuous optimization):

[0099] (Transfer Learning in Action): The plant-level transfer module calculates the cosine similarity of equipment features between the source and target plants daily. For example, for a new plant, the system also silently compares its similarity with that of the old plant daily. If the calculated similarity is 0.72 (out of 1), then the trust weight ω when learning from the old plant data today is:

[0100]

[0101] This means that today it will be more inclined to draw on the experience of the old factory areas (89.6% weight) rather than relying solely on its own limited data. This weight is used to weight the contribution of the source factory area loss function in today's model fine-tuning training, achieving smooth and stable knowledge fusion.

[0102] This invention addresses the accuracy issue of long-term prediction through equipment degradation modeling, the data scarcity problem in new plant areas through transfer learning, and the short-term sudden disturbance problem through random event simulation. The health index provides a benchmark for state alignment in transfer learning, and event compensation corrects prediction biases based on historical data. Field tests show that this invention reduces the one-year prediction error by more than 12% compared to traditional methods, achieves a prediction accuracy of 9% MAPE for new plant areas with only 500 data points, and reduces the response time to random events to less than 5 minutes. Its overall performance is significantly superior to existing technologies.

[0103] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for dynamic prediction of plant load, characterized in that, Includes the following steps: S1: Real-time acquisition of operating status data of key equipment in the factory area, including current harmonic distortion rate (THD), vibration value (Vibration), and winding temperature, and extraction of characteristic frequency band energy of vibration signal using wavelet packet decomposition; S2: Input the data processed in step S1 into the long short-term memory network to predict the remaining service life of the device and calculate the health index HI(t) at the current time point. S3: Based on the health index HI(t), predict the baseline load value P of the equipment. base (t) is dynamically corrected to obtain the actual load forecast value P of equipment degradation. real (t); S4: Based on a three-level transfer learning architecture, using source plant data, knowledge is transferred to the target plant through three levels: equipment level, process level, and plant area level, in order to build a load prediction model for the target plant. S5: Real-time compensation of load forecast results based on random event simulation; construction of a three-dimensional classification knowledge base of event-impact-duration; Monitor events in the manufacturing execution system or data acquisition and monitoring control system; when an event is detected, trigger the Gaussian process regression model to calculate the load deviation ΔP, and use reinforcement learning to dynamically adjust the event sensitivity coefficient.

2. The method for dynamic prediction of plant load according to claim 1, characterized in that, The specific formula for dynamic correction in step S3 is as follows: ； in, The environmental degradation factor is calculated using the following formula: ； In the formula, This is the current harmonic distortion influence coefficient; is the vibration influence coefficient; THD(t) is the measured value of the total harmonic distortion rate of the current at time t; Vibration(t) is the measured value of the effective value of the vibration velocity at time t.

3. The method for dynamic prediction of plant load according to claim 1, characterized in that, The dynamic weighted domain adaptation algorithm used in step S4 for plant-level migration has the following formula for calculating the weight ω: ； in, 3.0 is the cosine similarity between feature A of the source plant area and feature B of the target plant area, and 3.0 is the temperature coefficient.

4. The method for dynamic prediction of plant load according to claim 1, characterized in that, The formula for calculating the load deviation ΔP in step S5 is as follows: ； in, Total load deviation; is the weight coefficient for the k-th type of event; For the k-th type of event The corresponding load change function.

5. The method for dynamic prediction of plant load according to claim 4, characterized in that, Step S5 includes dynamically adjusting the weight coefficients using reinforcement learning. Its update formula is: ； in, The learning rate; The reward signal obtained at time t; For state Take action below The state-action value function.

6. A dynamic load prediction system for a plant area, used to implement the method described in any one of claims 1-5, characterized in that, The system includes: The data acquisition and preprocessing module is used to acquire the current, vibration, and temperature data of the equipment in real time, and to perform wavelet packet decomposition processing on the vibration signal. The equipment health status assessment module, with a built-in LSTM network, is used to predict the remaining useful life RUL(t) and calculate the health index HI(t). The load forecasting dynamic correction module receives the health index and baseline load forecast values, and calculates the actual load forecast value P. real (t) or load deviation ΔP; The three-level transfer learning module is used to perform equipment-level, process-level, and plant-level transfer learning. Its plant-level transfer learning adopts a dynamic weighted domain adaptation algorithm. The random event simulation and compensation module is used to compensate for load forecast results in real time; construct a three-dimensional classification knowledge base of event-impact-duration; monitor events in manufacturing execution systems or data acquisition and monitoring control systems; trigger Gaussian process regression models to calculate load deviation ΔP, and dynamically adjust the event sensitivity coefficient using reinforcement learning; and The central processing and control module is used to coordinate and control the operation of each module and output the final load dynamic prediction results.

7. The plant area load dynamic prediction system according to claim 6, characterized in that, The data acquisition and preprocessing module is deployed on the edge computing node of the factory area, while the other modules are deployed on the cloud server.

8. A dynamic load prediction system for a factory area according to claim 6, characterized in that, The random event simulation and compensation module integrates an FPGA-based hardware accelerator to accelerate the inference calculation of the Gaussian process regression model.

9. A dynamic load prediction system for a factory area according to claim 6, characterized in that, The process-level transfer learning module in the three-level transfer learning module is implemented using a graph neural network (GNN). The feature vectors of its nodes include equipment power, energy consumption, and runtime, while the edge weights represent the intensity of material flow or energy flow.

10. A dynamic load prediction system for a factory area according to claim 6, characterized in that, The system also includes a visual human-computer interaction interface for displaying equipment health status trends, load forecast curves, event alarm logs, and transfer learning effect evaluation dashboards.

Images