Industrial load regulation method and system based on digital twinning

Abstract

The application discloses an industrial load regulation method and system based on digital twinning, and belongs to the technical field of power system automation. The method comprises the following steps: collecting real-time and historical data of industrial load, energy storage and new energy; establishing and dynamically updating an industrial load digital twinning model; calculating the adjustable capacity vector of each load unit based on the model; predicting the future power sequence of new energy by using a method based on a wind direction adaptive attention mechanism; calculating the new energy consumption capacity by a dynamic evaluation method considering voltage flicker risk propagation; and generating a load regulation scheme according to the calculation result. The application accurately quantifies the load regulation capacity, improves the power prediction accuracy, and introduces voltage flicker risk evaluation, ensuring the power quality safety of the regulation process, realizing efficient and safe collaborative scheduling of industrial load, and effectively improving the new energy consumption level.

Description

Technical Field

[0001] This application relates to the field of power system automation technology, and in particular to an industrial load regulation method and system based on digital twins. Background Technology

[0002] With the advancement of renewable energy policies, the penetration rate of new energy sources such as wind power and photovoltaics in the power system continues to increase. However, the output of new energy sources is highly volatile and intermittent, and the accuracy of their power prediction is greatly affected by meteorological factors. Especially under complex conditions such as sudden changes in wind direction, traditional prediction models struggle to accurately capture power output dynamics, affecting the reliability of grid dispatch decisions.

[0003] Meanwhile, industrial loads, as an important adjustable resource, operate under complex conditions and involve diverse equipment types. Traditional methods often rely on static parameters or empirical estimations to determine their adjustment capacity, lacking real-time and precise quantification of load adjustment potential. This results in their flexibility being "adjustable but invisible," hindering their effective application in mitigating renewable energy fluctuations and supporting renewable energy consumption.

[0004] Furthermore, existing assessment methods for renewable energy absorption capacity often focus on power balance and capacity matching, neglecting power quality issues. When renewable energy output fluctuates drastically or industrial loads adjust rapidly, power quality disturbances such as voltage flicker can easily occur, potentially propagating and accumulating in the grid and threatening the safe operation of sensitive equipment. Traditional assessment methods fail to model the propagation path and dynamic risks of voltage flicker, leading to overly optimistic assessment results and potential safety hazards.

[0005] Therefore, existing technologies suffer from defects such as low accuracy in predicting new energy output, difficulty in quantifying industrial load regulation capabilities in real time, and insufficient consideration of voltage flicker risks. These defects lead to inaccurate assessment of new energy absorption capacity, unsafe and inefficient control schemes, and an urgent need for a new technical solution to address these problems. Summary of the Invention

[0006] The purpose of this application is to provide an industrial load regulation method and system based on digital twins, aiming to solve the technical problems in the background art, such as low accuracy of new energy output prediction, difficulty in real-time quantification of industrial load regulation capacity, and insufficient consideration of voltage flicker risk, which lead to inaccurate assessment of new energy absorption capacity, unsafe regulation and low efficiency.

[0007] To achieve the above objectives, this application provides an industrial load regulation method based on digital twins, comprising the following steps: Real-time and historical operating data of industrial load, energy storage device and new energy output are collected and preprocessed and synchronized with time; wherein, the real-time data includes at least the active power of industrial load, voltage of energy storage device and power output of new energy. A digital twin model of the industrial load is established based on the historical operating data of the industrial load, and the real-time data is mapped to the digital twin model of the industrial load for dynamic updates; Based on the updated digital twin model of industrial load, the volatility, flexibility and responsiveness of each load unit are calculated and integrated to form an adjustable capability vector of each load unit. Based on real-time data of renewable energy output, a renewable energy power prediction method based on wind direction adaptive attention mechanism is used to perform short-term prediction and generate future power sequences. Based on the adjustable capacity vector and the future power sequence, the industrial load's ability to support new energy absorption and the absorption rate are calculated in real time using a dynamic assessment method based on voltage flicker risk propagation. Based on the calculation results of the absorption capacity and absorption rate, a corresponding load adjustment scheme is generated.

[0008] Optionally, establishing a digital twin model of the industrial load based on the historical operating data of the industrial load includes: Extract load dynamic characteristics from the historical operational data; A basic industrial load model is constructed based on long short-term memory networks to describe the dynamic response relationship of load under different operating conditions. The topology between load units is constructed based on the physical connections and power transmission relationships of the industrial system, and the topology is embedded into the basic industrial load model. At the same time, the characteristics of energy storage units and new energy equipment are incorporated into the model to generate the digital twin model of the industrial load.

[0009] Optionally, the calculation of the volatility, flexibility, and responsiveness of each load unit based on the updated industrial load digital twin model includes: For each load unit, the standard deviation of its active power is calculated within a specified time window to obtain the load fluctuation rate; Based on the aforementioned digital twin model, the adjustable range of the load unit under different adjustment commands is simulated to obtain the flexibility. The response of a simulated load unit to a regulation command is used to obtain a response capability index, which describes the load's ability to quickly adjust power after receiving a regulation command. The load volatility, flexibility, and responsiveness indicators are integrated into the adjustable capability vector.

[0010] Optionally, the renewable energy power prediction method based on wind direction adaptive attention mechanism includes: Real-time wind direction data, including wind direction angle and wind direction change rate, is collected and synchronized with the power output data of new energy sources to construct a power time series feature vector. The wind direction angle is divided into multiple azimuth intervals, and the initial power output correction coefficient for each azimuth interval is calculated based on historical data. Calculate the wind direction fluctuation index based on the time series of wind direction angle; Construct a mapping relationship between wind direction zones and power output to form a wind direction feature vector; A wind direction attention layer is introduced into the long short-term memory network model to jointly model the wind direction feature vector and the power time series feature vector to generate a comprehensive feature sequence. The prediction time window is dynamically adjusted based on the wind direction change rate. The corresponding correction coefficient is obtained based on the current wind direction and its azimuth range, and the correction coefficient is updated on a rolling basis in combination with recent actual power output data; The updated comprehensive feature sequence is input into the trained long short-term memory network model, and the future power sequence is output.

[0011] In one possible implementation, the introduction of a wind direction attention layer into the long short-term memory network model to jointly model the wind direction feature vector and the power time series feature vector includes: The wind direction feature vector and the power time series feature vector are aligned and fused in the time dimension to form a multimodal input feature matrix; The correlation weights of upwind characteristics to power changes at each time step are calculated using the wind direction attention layer. Based on the soft attention mechanism, the correlation weights are used to weight the wind direction feature vectors at different time steps in the multimodal input feature matrix; The weighted wind direction feature vector is fused with the power time series feature to generate the comprehensive feature sequence, which is then input into the hidden state update unit of the long short-term memory network.

[0012] Optionally, the dynamic assessment method for absorption capacity based on voltage flicker risk propagation includes: Based on the power grid topology, a path network model for voltage flicker propagation is established, and the flicker influence factor matrix between nodes is calculated. Real-time monitoring of short-term flicker values ​​at each node, and calculation of cumulative voltage flicker values ​​using a moving average; Based on the real-time monitored voltage flicker value, a nonlinear function is used to calculate the voltage flicker risk index of each node, and it is mapped to the constraint factor of the adjustable power of the node to form the adjustable power range under risk constraints. Based on the risk-constrained adjustable power range and the future power sequence, combined with the length of the voltage flicker propagation path and the attenuation coefficient, the total adjustable power of the system is calculated. With the objectives of maximizing the utilization rate of new energy output and minimizing the node risk power deviation, a multi-objective optimization function considering flicker constraints is constructed. By solving this function, the absorption capacity and absorption rate under risk constraints are obtained.

[0013] In one possible implementation, the voltage flicker risk index of each node is calculated using a nonlinear function based on the real-time monitored voltage flicker value, and then mapped to a constraint factor for the adjustable power of the node, forming an adjustable power range under risk constraints, including: Based on short-term flicker values ​​and cumulative flicker values, calculate the flicker risk index for each node; Map the flash risk index to a constraint factor; Multiply the constraint factor by the node's own maximum physical adjustable power to obtain the upper limit of adjustable power under risk constraints; and establish an adjustable power model under risk constraints based on the upper limit of adjustable power. The adjustable power of the risk constraint is obtained by solving the adjustable power model of the risk constraint.

[0014] In one possible implementation, the calculation of the total adjustable power of the system based on the risk-constrained adjustable power range and the future power sequence, combined with the length and attenuation coefficient of the voltage flicker propagation path, includes: Based on risk-constrained adjustable power and future power sequences, the length and attenuation coefficient of the voltage flicker propagation path are obtained. Construct a path attenuation matrix based on the length and attenuation coefficient of the voltage flicker propagation path; By combining the flicker influence factor matrix, the effective contribution of each node to the sensitive node is calculated; The candidate value of the total adjustable power of the system is obtained by summing the effective contributions of all nodes. By adjusting the candidate values ​​using a linear programming algorithm, the final total adjustable power of the system is obtained while satisfying the risk constraints and power balance of each node.

[0015] Optionally, generating a corresponding load adjustment scheme based on the calculation results of the absorption capacity and absorption rate includes: Based on the adjustable capacity vector of each load unit, a candidate operation list for load start-up and power regulation is generated; Based on the state of the energy storage unit and the future power sequence, candidate operations for energy storage charging and discharging are generated; The load and energy storage candidate operations are combined to form a set of system-level regulation schemes and their feasibility is verified. The verified solution is then translated into actual control commands.

[0016] This application also provides an industrial load regulation system based on digital twins, comprising: The data acquisition and processing module is used to collect real-time data and historical operating data of industrial load, energy storage device and new energy output, and to perform preprocessing and time synchronization; wherein, the real-time data includes at least the active power of industrial load, voltage of energy storage device and power output of new energy. The industrial load digital twin modeling and updating module is used to establish an industrial load digital twin model based on the historical operating data of the industrial load, and to map the real-time data to the industrial load digital twin model for dynamic updating; The load adjustability assessment module is used to calculate the volatility, flexibility and response capability of each load unit based on the updated industrial load digital twin model, and integrate them to form the adjustability vector of each load unit. The new energy power prediction module is used to make short-term predictions based on real-time data of new energy output and a new energy power prediction method based on wind direction adaptive attention mechanism, and generate future power sequences. The dynamic evaluation module for absorption capacity is used to calculate in real time the industrial load's ability to support new energy absorption capacity and absorption rate based on the adjustable capacity vector and the future power sequence, using a dynamic evaluation method for absorption capacity based on voltage flicker risk propagation. The load regulation scheme generation module is used to generate a corresponding load regulation scheme based on the calculation results of the absorption capacity and absorption rate.

[0017] Compared with the prior art, the beneficial effects of this application are as follows: 1. Improved accuracy of renewable energy power forecasting: By constructing a forecasting method based on wind direction adaptive attention mechanism, which integrates multiple wind direction features such as wind direction angle and rate of change, and introduces dynamic correction and adaptive forecasting window, it can effectively capture the impact of sudden wind direction changes on power output, significantly improve the accuracy and stability of short-term power forecasting, and provide a reliable data foundation for subsequent absorption capacity assessment and load regulation.

[0018] 2. Achieved a safe assessment of absorption capacity while considering power quality: By introducing a dynamic assessment method based on voltage flicker risk propagation, power quality risks are quantified as dynamic constraints on the adjustable power of the load, fully considering the safe and stable operation boundary of the power grid when assessing absorption capacity. This breaks through the limitation of traditional assessment methods that only focus on power balance, avoids the problem of power quality deterioration caused by load adjustment, and makes the assessment results safer and more reliable.

[0019] 3. Precise quantification and efficient utilization of load regulation resources: By establishing a digital twin model of industrial loads and calculating a multi-dimensional adjustable capacity vector including volatility, flexibility, and responsiveness, the previously ambiguous load regulation potential becomes "visible and quantifiable." Combined with the results of safe absorption capacity assessment, load resources with different characteristics can be optimally combined and coordinated for scheduling. Under the premise of ensuring safety, the flexibility of industrial loads is maximized, thereby effectively improving the local absorption level of new energy. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. Furthermore, these drawings and textual descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this application to those skilled in the art through reference to specific embodiments.

[0021] Figure 1 This is a schematic flowchart of an industrial load regulation method based on digital twins, provided as an embodiment of this application.

[0022] Figure 2 This is a schematic diagram of the structure of an industrial load regulation system based on digital twins, provided as an embodiment of this application. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0024] The core idea of ​​this application is to accurately characterize and map the energy consumption characteristics of industrial systems in real time by constructing a high-fidelity digital twin model of industrial load. Based on this, an innovative method for predicting renewable energy power that considers the dynamic impact of wind direction, and a capacity assessment model that takes into account the propagation risk of power quality (especially voltage flicker), are combined to ultimately form a closed-loop load regulation strategy that balances safety and economy. This method transforms the abstract "adjustable potential" into concrete, quantifiable, and risk-controllable dispatch instructions, thereby maximizing the absorption of fluctuating renewable energy sources while ensuring the safe and stable operation of the industrial power grid.

[0025] Please see Figure 1This document illustrates the overall process of an industrial load regulation method based on digital twins, as provided in an embodiment of this application. This method can be executed by an industrial load regulation system deployed on a cloud server or a local edge computing device in an industrial park. The system may include a processor, memory, communication interface, etc., in hardware. The memory stores a computer program, which, when executed by the processor, implements… Figure 1 The steps are shown. In a specific system architecture, this system can be embodied as an industrial load regulation system based on digital twins, such as... Figure 2 As shown, it includes a data acquisition and processing module 100, an industrial load digital twin modeling and updating module 200, a load adjustability assessment module 300, a new energy power prediction module 400, a dynamic assessment module for absorption capacity 500, and a load adjustment scheme generation module 600. These six modules are used sequentially to execute steps S1 to S6 in Example 1. The method flow will be described in detail below in conjunction with this system architecture.

[0026] Example 1 This application provides an industrial load regulation method based on digital twins, such as... Figure 1 As shown, the method first executes step S1: collecting real-time and historical operating data of industrial load, energy storage devices, and new energy output, and performing preprocessing and time synchronization. This step is the data foundation of the entire method, and its accuracy and completeness directly affect the quality of all subsequent analyses and decisions. This step is executed by the data acquisition and processing module.

[0027] Specifically, the sources of data collection are diverse. For industrial loads, data typically comes from smart meters and power quality analyzers installed on various production lines and large equipment (such as electric arc furnaces, rolling mills, and electrolytic cells), or is obtained through the factory's Manufacturing Execution System (MES) and Distributed Control System (DCS). For energy storage devices, data comes from the Battery Management System (BMS), including voltage, current, state of charge (SOC), and temperature. For renewable energy output, data mainly comes from the Supervisory Control and Data Acquisition (SCADA) systems of renewable energy power plants, such as wind turbine controllers and photovoltaic inverters. In addition, relevant meteorological data, such as wind speed, wind direction, solar irradiance, temperature, and humidity, are also collected; these data are crucial for renewable energy power forecasting.

[0028] Real-time data is crucial for reflecting the current state of the system. In this embodiment, real-time data includes at least the active power of the industrial load, the terminal voltage of the energy storage device, and the real-time output power of the new energy source. In better practices, the collected data dimensions will be richer, such as the reactive power, power factor, phase current and voltage of the industrial load, the charging and discharging current and cycle number of the energy storage device, and the voltage, frequency, and short-term flicker (Pst) value of key nodes in the power grid. This multi-dimensional data provides a solid foundation for building a high-precision model. The sampling frequency of real-time data depends on the dynamic nature of the application scenario. For example, for a rapidly fluctuating electric arc furnace load, the sampling frequency may reach the kHz level, while for a relatively stable continuous production line load, sampling at the second or minute level is sufficient.

[0029] Historical operational data forms the foundation for model training and pattern discovery. The system stores all types of data over a long period (e.g., more than a year) to form a high-resolution data warehouse. This historical data not only includes operating records of equipment under various conditions but also contains the combined impact of factors such as seasons, weather, and production plans on the power system.

[0030] Data preprocessing is a crucial step in ensuring data quality. Raw data often contains noise, missing data, outliers, and misaligned timestamps. Therefore, preprocessing steps include: 1. Data Cleaning: Identifying and removing obvious outliers or bad data by setting thresholds or using statistical algorithms (such as the 3σ criterion). For example, a sudden power spike to zero or a maximum value caused by a momentary sensor malfunction.

[0031] 2. Data imputation: For missing data points, appropriate interpolation methods are used to imput them, such as linear interpolation, spline interpolation, or imputation based on the mean of data from adjacent time points, in order to ensure the continuity of the data sequence.

[0032] 3. Data Denoising: Filtering algorithms, such as moving average filtering, Gaussian filtering, or wavelet transform, are used to remove high-frequency noise from the data and extract the effective signal. For example, the output power of wind turbines is smoothed to reduce instantaneous disturbances caused by gusts.

[0033] 4. Time Synchronization: Due to slight timestamp discrepancies between data from different sources, it is essential to align all data streams to a unified time base using methods such as Network Time Protocol (NTP) or GPS clock synchronization. This ensures that at any given point in time, all data reflects the system state at the same instant. This is a prerequisite for conducting multivariate joint analysis.

[0034] After data acquisition and preprocessing are completed, the method proceeds to step S2: A digital twin model of the industrial load is established based on historical operational data, and real-time data is mapped to the digital twin model for dynamic updates. This step, completed by the industrial load digital twin modeling and updating module, is one of the core components of this application, aiming to create a highly realistic virtual copy that is synchronized in real time with the physical industrial system.

[0035] A digital twin model is not a single model, but a collection of multi-dimensional, multi-scale models. It first needs to depict the physical entities and topology of an industrial system. For example, within an industrial park, the model would precisely represent the location and electrical connections of entities such as substations, feeders, factory workshops, large electrical equipment, energy storage stations, and photovoltaic arrays, using nodes and edges. This topology defines the flow paths of electrical energy and the interactions between devices.

[0036] Secondly, the core of the digital twin model is to establish a dynamic behavior model for each key load unit. Unlike traditional static ledger parameters, this model aims to capture the dynamic response characteristics of the load under different operating commands and conditions. For example, for a rolling mill, the model not only records its rated power but also simulates the rapid changes in power demand throughout the entire process of billet entry, rolling, and exit. This dynamic behavior model is typically built based on historical data-driven methods, such as using neural networks, support vector machines, or time series analysis models.

[0037] After establishing the digital twin model, the most crucial step is to achieve "twin"—that is, real-time synchronization with the physical world. This is accomplished by continuously "feeding" the real-time data collected and preprocessed in step S1 into the digital twin model. This process is called data mapping and model updating. For example, when the actual operating power of a motor in the physical world changes, this real-time power value will be immediately updated in the corresponding virtual object of the motor in the digital twin model. This dynamic updating allows the digital twin model to accurately reflect the real state of the physical system at any given time, including the operating parameters of each device and the power flow distribution of the power grid.

[0038] The advantage of this dynamic updating is that it provides the most accurate initial conditions available for various simulations and extrapolations. We can perform "what-if" analyses in this virtual world; for example, simulating how the system power flow will change after executing a certain regulation command, or whether the voltage at a certain node will exceed its limit—all without causing any interference or risk to the actual physical system. This provides a powerful tool for subsequent load capacity assessment and regulation scheme development.

[0039] Next, the method executes step S3: based on the updated industrial load digital twin model, the volatility, flexibility, and responsiveness of each load unit are calculated, and integrated to form an adjustable capacity vector for each load unit. This step is performed by the load adjustable capacity assessment module, and its purpose is to quantify and profile the "adjustable potential" of the industrial load.

[0040] In traditional power grid dispatching, industrial loads are typically treated as a fixed, unadjustable whole. However, many industrial production processes actually possess a degree of flexibility. This step aims to uncover and quantify this flexibility. Based on a real-time updated digital twin model, the system can perform in-depth analysis of each load unit (which could be a piece of equipment, a production line, or a workshop), specifically including the following aspects: Volatility: The standard deviation or coefficient of variation of power is calculated by analyzing its historical and real-time power curves over a certain time window. High-volatility loads (such as electric arc furnaces) are themselves sources of disturbance, but sometimes they also imply the possibility of adjustment within the fluctuation cycle.

[0041] Flexibility: This is an indicator that measures the adjustable range of the load. Through simulation in a digital twin model, the system can determine the maximum adjustable power and the maximum adjustable power of the load without affecting core production processes and product quality. For example, for temperature-controlled loads (such as industrial freezers), their start and stop can be adjusted within a certain temperature range, which reflects their flexibility.

[0042] Response time: This measures how quickly and accurately a load executes regulation commands. Also simulated using a digital twin model, the system can assess the time a load takes from receiving a command to completing power regulation (response time) and the minimum regulation step size it can achieve (regulation accuracy). For example, inverter-controlled motors respond quickly, while equipment requiring preheating responds slowly.

[0043] By calculating these three key indicators, the system generates an adjustable capacity vector for each load unit, for example... This vector acts like a "capability business card" for each load, clearly demonstrating its strengths and limitations in participating in grid regulation. For example, one vector might indicate that a load has a "large regulation range but slow response," while another might indicate a "small regulation range but fast response." This provides a basis for subsequently developing refined regulation strategies.

[0044] Having gained a clear understanding of the load side, the method shifts its focus to the supply side, namely the output of new energy sources. This leads to step S4: based on real-time data of new energy output, a new energy power prediction method based on a wind direction adaptive attention mechanism is used to perform short-term predictions and generate future power sequences. This step, executed by the new energy power prediction module, is another major core innovation of this application.

[0045] The output of new energy sources (especially wind and solar power) is highly volatile and uncertain, making accurate short-term power forecasting a prerequisite for effective grid integration. Traditional forecasting methods often focus only on key factors such as wind speed and solar irradiance. However, in real-world scenarios, especially for wind power, sudden changes in wind direction are a major cause of dramatically increased forecasting errors. For example, rapid wind deflection can cause wind turbines to yaw frequently or cause downstream turbines to enter the wake region of upstream turbines, resulting in actual output far lower than the wind speed-based forecast.

[0046] The prediction method proposed in this application, based on a "wind direction adaptive attention mechanism," aims to address this pain point. It doesn't simply treat wind direction as a common input feature, but rather delves into its dynamic characteristics. This method collects data such as wind direction angle and rate of change in real time and constructs a special attention layer. This attention layer dynamically analyzes which moments in the current and historical time series have the greatest impact on future power changes, assigning higher weights to these key pieces of information.

[0047] For example, when the system detects a sudden increase in the rate of wind direction change, the attention mechanism automatically "focuses" on this change and may trigger the model to dynamically adjust its forecasting strategy, such as shortening the forecast time window, to cope with impending uncertainty. This mechanism allows the model to act like an experienced meteorologist, remaining highly vigilant about subtle changes in wind direction and understanding their potential impact on wind turbine output. In this way, the model can generate a more accurate and reliable power forecast sequence for a future period (such as the next 15 minutes to 4 hours).

[0048] With a precise grasp of load regulation capacity and future renewable energy output, the method enters the crucial decision-making and evaluation stage, namely step S5: based on the adjustable capacity vector and future power sequence, the industrial load's ability to support renewable energy absorption and the absorption rate are calculated in real time using a dynamic assessment method based on voltage flicker risk propagation. This step is executed by the dynamic assessment module for absorption capacity.

[0049] Absorption capacity is not simply a matter of adding or subtracting the predicted output of new energy sources from the adjustable load range. A key, but often overlooked, constraint is power quality, particularly voltage flicker. Voltage flicker is a voltage fluctuation caused by rapid, large-scale load changes (such as the start-up of electric arc furnaces or the switching of large motors). It propagates through the power grid like ripples, causing interference or even damage to sensitive equipment (such as precision instruments and servers). Similarly, large-scale, rapid adjustments to industrial loads to absorb new energy sources can also trigger severe voltage flicker problems.

[0050] The "Dynamic Assessment Method for Absorption Capacity Based on Voltage Flicker Risk Propagation" proposed in this application innovatively incorporates this risk into the assessment model. This method first establishes a network model of voltage flicker propagation paths based on the power grid topology. This model can calculate the flicker impact of power fluctuations at any node on other nodes in the power grid, forming a "flicker impact factor matrix."

[0051] The system then monitors the voltage flicker values ​​of each key node in real time and, combined with historical data, calculates a "voltage flicker risk index." A higher index indicates a node is more sensitive to voltage fluctuations or is already at a high flicker level. This risk index is then used as a dynamic "constraint factor" to limit the adjustable power range of the load on that node. For example, a node adjacent to a data center will have a high risk index, meaning that even if the adjustable load connected to that node can be significantly adjusted physically, its allowable adjustment range will be strictly limited to protect the data center's security.

[0052] Ultimately, within a multi-objective optimization framework, the system comprehensively considers the future power sequence of new energy sources, the adjustability vector of each load, and the dynamic security constraints defined by voltage flicker risk. It calculates the maximum new energy power that the entire industrial system can absorb, i.e., the "absorption capacity," and the corresponding absorption rate, under the premise of ensuring grid safety (especially preventing flicker from exceeding limits). This result is dynamic, recalculated every few minutes based on the latest system state, achieving a real-time, accurate, and safe assessment of absorption potential.

[0053] Finally, the method executes step S6: based on the calculated absorption capacity and absorption rate, a corresponding load adjustment scheme is generated. This step is executed by the load adjustment scheme generation module and is the final step in translating the assessment results into practical action.

[0054] Based on the optimal load reduction strategy calculated in step S5, this module generates a set of specific, executable control instructions. This is no longer a vague, macroscopic instruction to "increase / decrease the load," but a detailed sequence of operations. For example, the plan might include: "Within the next 5 minutes, reduce the power of rolling mill No. 3 in workshop A by 1.5MW; simultaneously, activate the energy storage system to charge at a power of 0.5MW; and notify the temperature control system in workshop B that it can enter energy-saving operation mode in 10 minutes." The generation of this scheme is a complex combinatorial optimization process. The system iterates through all available regulation operations of all load units and energy storage units (based on their adjustable capability vectors), combining them into thousands or even tens of thousands of possible system-level regulation schemes. Then, it uses a digital twin model to quickly verify the feasibility of each candidate scheme, checking whether it meets power balance, whether it violates voltage or flicker constraints, and whether it meets the process requirements of each device. Finally, the system selects one or a set of optimal schemes and translates them into actual control commands conforming to industrial control protocols (such as Modbus, PROFIBUS), which are then issued to the controllers of each device for execution.

[0055] In summary, this application constructs a complete closed loop through these six interconnected steps, from data perception and model cognition to accurate prediction, safety assessment, and optimized control. It not only solves the problem of "how much can be adjusted," but more importantly, it addresses the core challenges of "when to adjust, how to adjust, and what risks are involved after adjustment," providing strong technical support for the orderly, efficient, and safe participation of industrial loads in renewable energy consumption.

[0056] Example 2 Based on the above embodiments, in order to further improve the performance and accuracy of the method, this application also provides a series of preferred implementation methods.

[0057] In a preferred embodiment, the process of establishing a digital twin model of the industrial load in step S2 is refined. First, when extracting dynamic characteristics of the load from historical data, it is not only statistical features, but also includes the extraction of periodic features and harmonic features of the load using signal processing techniques (such as Fourier transform and wavelet transform). These deeper features help to more accurately depict the power quality profile of the load.

[0058] Next, when constructing the basic industrial load model based on Long Short-Term Memory (LSTM) networks, the model is not a simple single-layer LSTM. Preferably, a stacked LSTM or bidirectional LSTM architecture can be used. Stacked LSTM increases the network depth, enabling it to learn higher-level temporal abstractions from the data. Bidirectional LSTM can learn from both past and future information simultaneously, which is very helpful for understanding certain industrial processes with sequential dependencies (e.g., the end of one operation step foreshadows the start of the next high-power step). The model's input layer receives not only historical power sequences but also diverse auxiliary features such as equipment switching status, production order numbers, and ambient temperature. The model's output layer uses a fully connected network to map the hidden states of the LSTM to power predictions for multiple future time steps, achieving multi-step prediction.

[0059] Furthermore, embedding the physical topology of the industrial system into the model is crucial for building a high-fidelity digital twin. This involves more than simply drawing a connection graph; it involves mathematically representing the topological relationships, for example, using adjacency matrices or graph Laplacian matrices. This graph structure information is then fused with the LSTM model. An advanced implementation employs graph neural networks (GNNs), such as GCNs (Graph Convolutional Networks) or GATs (Graph Attention Networks). At each time step, a GNN can aggregate information about a load node and its neighboring nodes, allowing the model to understand the transmission effects of power regulation in the grid. For example, when the power of load A decreases, the voltage of its directly connected load node B will rise slightly; the GNN can learn this local influence. Simultaneously, the characteristics of energy storage units and new energy devices, such as the SOC and charge / discharge efficiency curves of energy storage, and the MPPT (Maximum Power Point Tracking) algorithm characteristics of new energy sources, are also integrated as node attributes or dedicated model modules, ultimately generating a highly complex industrial load digital twin model capable of simulating the dynamic interactions of the entire system.

[0060] In another preferred embodiment, the process of calculating the adjustable capacity vector in step S3 is described in detail. For each load unit b, its volatility is calculated. This is performed within a dynamically identified analysis time window N. For example, for periodic loads, window N should cover at least one complete production cycle. The calculation formula is as follows: ,in It is the active power at time t. It represents the average power within the window. This indicator reflects the inherent instability of the load.

[0061] Calculate flexibility In this process, the digital twin model plays a crucial role. The system issues a series of increasing or decreasing power regulation commands to load b within the virtual model, while simultaneously monitoring its associated key process parameters (such as temperature, pressure, and product quality indicators). Flexibility Defined as the maximum power at which the load can operate stably under the premise that all key process parameters do not exceed the allowable range. With minimum power The difference, that is This is a simulation-based, non-invasive approach to potential discovery.

[0062] Calculate response capability indicators Similarly, this is based on digital twin simulation. The system simulates sending a step power regulation command to load b (e.g., adjusting from the current power P_current to the target power P_target) and records its power response curve. Response capability indicators It can be defined as the ratio of power change to response time, i.e. ,in This refers to the actual achievable power variation range. This is the time required for power to reach 90% of the target value. This metric describes the load's "agility." Ultimately, these three quantifiable metrics... , and Integrated into a three-dimensional vector A precise profile was created for each load unit.

[0063] One of the core innovations of this application, namely the new energy power prediction method in step S4, is described in great detail in a preferred embodiment. This method is called the prediction method based on the "wind direction adaptive attention mechanism". First, in step S4.1, the system collects the wind direction angle and wind direction angle change rate reported by the wind turbine yaw system in real time, and performs strict time synchronization with the power data to construct a multi-dimensional power time series feature vector.

[0064] In step S4.2, the 360-degree wind direction angle is divided into multiple azimuth intervals, such as eight 45-degree intervals (due north, northeast, due east, etc.). Based on long-term historical data, the average power output level of the wind turbines within each wind direction interval is statistically analyzed. Then, the initial power output correction coefficient for each wind direction zone is calculated. Here It could be the total average power across all wind directions, and It represents the average power within a specific wind direction zone g. This coefficient reflects the systematic deviation in power output under different wind directions caused by fixed factors such as topography and wake.

[0065] In step S4.3, a "wind direction fluctuation index" is introduced to quantify the dynamic instability of wind direction. This index can be the variance or standard deviation of the wind direction angle change per unit time. A high wind direction fluctuation index indicates unstable wind conditions and increased difficulty in prediction.

[0066] In step S4.4, information such as wind direction angle, wind direction change rate, wind direction fluctuation index, and correction coefficients for the corresponding zones are integrated into a "wind direction feature vector". This vector comprehensively describes the static and dynamic characteristics of the current wind direction.

[0067] Step S4.5 details the core of this method—the wind direction attention layer. In LSTM models, the traditional approach is to treat all input features (such as historical power, wind speed, and wind direction) equally at each time step. However, the attention mechanism introduced in this application differs. In step S4.51, the wind direction feature vector and the power time series feature vector are aligned and fused into a multimodal input feature matrix. In step S4.52, the attention layer calculates a relevance weight. This weight represents the importance of wind direction characteristics for predicting future power at time t. For example, this weight may be low when the wind direction is stable, but it will increase significantly when the wind direction changes drastically. In step S4.53, this weight is used through a soft attention mechanism. The input wind direction feature vector is weighted. In step S4.54, the weighted wind direction feature is fused with the power feature to generate a comprehensive feature sequence. The data is then fed into the hidden state update unit of the LSTM. This is similar to how the model learns to "highlight" information, intelligently focusing on wind direction changes that have the greatest impact on the prediction results.

[0068] Step S4.6 further demonstrates adaptability. The system dynamically adjusts the length of the prediction time window based on the real-time wind direction change rate. When wind direction changes are gradual, a longer time window (e.g., 1-4 hours) can be used for longer-term planning; when wind direction changes drastically, the prediction window is automatically shortened (e.g., 15-30 minutes), focusing on accurate short-term predictions and avoiding the contamination of long-term predictions by huge uncertainties. Simultaneously, to prevent abrupt changes in the prediction window length from causing jumps in the output power sequence, a smooth transition mechanism is designed to ensure that changes in the window length are gradual.

[0069] Step S4.7 implements the model's self-learning capability. The system will obtain the corresponding correction coefficients based on the azimuth interval of the current wind direction. However, this coefficient is not static. The system updates it by incorporating actual power output data from the same wind direction range over a recent period (e.g., 24 hours). For example, if it is found that the actual power output in the northeast wind direction is generally 5% higher than the historical average in the most recent day, the model will automatically increase the correction coefficient for the northeast wind direction. This rolling update allows the model to adapt to slowly changing factors such as seasonal variations or equipment aging.

[0070] Finally, in step S4.8, the comprehensive feature sequence, which has undergone attention weighting, feature fusion, and incorporates the latest correction coefficients, is input into the already trained LSTM model to output a highly accurate future power sequence.

[0071] Similarly, the dynamic evaluation method for absorption capacity in step S5 is also detailed in the preferred embodiment, corresponding to claims 6, 7, and 8. In step S5.1, a path network model for voltage flicker propagation is established, and the flicker influence factor matrix is ​​calculated. .in, It is the transfer impedance from node i to node j in the power grid topology. It is the short-circuit capacity of node j. This is the system's rated voltage. The physical meaning of this matrix is ​​clear: it quantifies how much voltage flicker will occur at node j due to a unit power fluctuation at node i.

[0072] In step S5.2, the system monitors the short-term flicker value Pst of each node in real time through the power quality monitoring device, and calculates the cumulative voltage flicker value Plt by moving average to capture the long-term cumulative effect of flicker.

[0073] Step S5.3 is the core of risk quantification. The system uses a non-linear function to calculate the "flicker risk index" for each node. For example, a sigmoid function or a hyperbolic tangent function (tanh) can be used. A specific calculation formula could be:

[0074] in, and These are the real-time short-term and cumulative flicker values, respectively. and It is the flicker limit specified by national or industry standards. and This is a weighting coefficient that can be adjusted based on the sensitivity of the devices connected to that node. For example, for a node connecting to precision instruments, The weighting can be set higher because the cumulative flicker effect is more harmful. Then, this risk index... It is mapped to a "constraint factor" between 0 and 1. Finally, this constraint factor is multiplied by the node's own physical maximum adjustable power to obtain the upper limit of adjustable power under the risk constraint. This creates a dynamic "airbag": the higher the risk of flicker in a node, the smaller the allowable adjustment range.

[0075] Step S5.4 calculates the total adjustable power of the entire system based on the adjustable power of each node. This requires considering the attenuation of flicker along the propagation path. In steps S5.41 and S5.42, the path length is used as a basis for calculation. and line attenuation coefficient Construct a path decay matrix In step S5.43, the flicker influence factor matrix is ​​combined. and decay matrix Calculate the adjustable power of each node i. "Effective contribution" to sensitive node j In step S5.44, the effective contributions of all nodes are summed to obtain a candidate value for the total adjustable power of the system. Finally, in step S5.45, this candidate value is adjusted using optimization algorithms such as linear programming, while satisfying the risk constraints of all nodes and the power balance of the entire system, to obtain the final and feasible total adjustable power of the system. .

[0076] Finally, in a preferred embodiment, the load regulation scheme generation process in step S6 is further refined. In step S6.1, the system generates a candidate operation list based on the adjustability vector of each load unit. For example, for a load with a flexibility of [-2MW, +1MW] and a response time of 5 minutes, its candidate operations may include discretized operation items such as "reduce 0.5MW within 5 minutes" or "reduce 1.0MW within 5 minutes". In step S6.2, similarly, candidate charging and discharging operations are generated based on the SOC of energy storage, power limitations, and the peak and valley conditions of future renewable energy sources. In step S6.3, the system combines these candidate operations from different loads and energy storage to form a massive set of system-level regulation schemes. Then, each combined scheme undergoes rapid feasibility verification, eliminating those schemes that would violate grid constraints (voltage, flicker, line capacity). In step S6.4, from all feasible solutions, the optimal solution is selected based on one or more optimization objectives (such as lowest adjustment cost or fastest response speed), and it is translated into a specific, timestamped sequence of control instructions and issued to the on-site actuators.

[0077] Example 3 To more intuitively understand how this application works, this embodiment constructs a specific application scenario. Assume a coastal industrial park includes a steel plant using an electric arc furnace (load A, high power, highly volatile, and a major source of flicker), a data center (load B, stable power, extremely sensitive to power quality), a precision manufacturing workshop equipped with variable frequency fans (load C, medium power, capable of rapid adjustment within a small range), a 50MW wind farm, and a 10MW / 20MWh battery energy storage system (BESS).

[0078] On a certain workday afternoon, the system followed... Figure 1 The process shown continues to run: S10: Data Acquisition and Preprocessing. The system's data acquisition and processing module is gathering data from multiple sources. The SCADA system uploads the total output of the wind farm, the wind speed, wind direction angle, and power of each turbine every second. The steel plant's DCS system uploads the operating status (smelting period, refining period) and instantaneous power of the electric arc furnace (sampling frequency up to 10kHz). The data center's BMS uploads the total power consumption of the server cluster and the voltage flicker value at the PDU (Power Distribution Unit). The smart meters in the precision manufacturing workshop upload the power of the production line. The energy storage system's BMS uploads SOC, voltage, and current. All data is cleaned, timestamped, and then formed into a unified format data stream.

[0079] S20: Digital Twin Model Establishment and Update. At the core of the system, the industrial load digital twin modeling and update module maintains a virtual power grid for a campus. In this model, steel mills, data centers, manufacturing workshops, wind farms, and energy storage stations are nodes in the graph, and the transmission lines between them are edges. The electric arc furnace model is a complex nonlinear time-series model that can predict its next power demand based on its current stage in the smelting cycle. The data center is modeled as a constant load with zero tolerance for voltage disturbances. The manufacturing workshop model is linked to its production schedule, knowing when a short-term power reduction can be implemented. At this moment, real-time data streams are constantly flowing in, dynamically updating the state of each virtual device in the model, and the power flow distribution of the virtual power grid is completely consistent with the real world.

[0080] S30: Adjustable Capability Vector Calculation. Based on this vivid digital twin model, the load adjustability assessment module is creating a "profile" for each load. It calculates that: the steel plant (load A) has enormous flexibility (can reduce power by tens of megawatts during smelting breaks), but poor responsiveness (cannot be started or stopped instantaneously) and extremely high volatility. The data center (load B) has zero flexibility. The precision manufacturing workshop (load C) has relatively low flexibility (±500kW), but excellent responsiveness (second-level response). The adjustable capability vector of the energy storage system includes its current SOC and maximum charge / discharge power. These vectors are updated and stored in real time.

[0081] S40: New Energy Power Forecast. At this time, meteorological data shows a cold wind is approaching, and the wind direction is expected to change from northwest to northeast within 30 minutes. The new energy power forecast module captures this information. Its internal LSTM model, based on a wind direction adaptive attention mechanism, begins to operate. The sharp increase in the wind direction change rate significantly increases the weight of the wind direction attention layer, and the model begins to "focus" on the impact of the wind direction change. Simultaneously, based on the wind direction change rate exceeding a preset threshold, the system automatically shortens the forecast time window from 2 hours to 30 minutes to pursue higher short-term accuracy. The model predicts that within the next 15-25 minutes, due to the wind turbines needing to collectively yaw to adapt to the new wind direction, the total output will experience a brief trough of approximately 15MW.

[0082] S50: Dynamic Assessment of Absorption Capacity. The Dynamic Assessment of Absorption Capacity module receives the predicted sequence of this "output trough" and the adjustable capacity vector of each load. A preliminary calculation shows that to fill this 15MW gap, the energy storage system needs to discharge, and the precision manufacturing plant (load C) may need to reduce its power. However, the core task of this module is to assess the risks of this move. It queries the flicker impact factor matrix and finds that the steel plant (load A) and the data center (load B) are electrically close, resulting in a high impact factor. At this time, the electric arc furnace is in the smelting period of violent power fluctuations, causing the real-time flicker value Pst at the data center access point to be close to the alarm line. The system's flicker risk index calculation module therefore assigns a very high risk index to the data center node and generates an extremely strong constraint factor. This constraint factor acts on the adjustable power range of the steel plant. Although it has the ability to adjust, the model prohibits any large adjustment operations at this time to prevent flicker superposition and data center downtime.

[0083] Next, the system constructs and solves a multi-objective optimization function. The objectives are: during the upcoming 10 minutes (the period of lowest power output), 1) minimize power curtailment (in this scenario, minimizing load reduction); and 2) minimize the risk power deviation at each node. Under the strong flicker risk constraint, the optimization solver concludes that the current safe total adjustable power of the system in the park (adjusted upwards to fill the gap) is 12MW, of which 10MW comes from the full-power discharge of the energy storage system and 2MW comes from the power reduction of the precision manufacturing workshop (load C). This means that there is still a 3MW (15MW-12MW) power gap that cannot be filled by internal resources while ensuring absolute safety, and it may be necessary to purchase electricity from the upper-level grid.

[0084] S60: Load Adjustment Scheme Generation. Finally, the load adjustment scheme generation module translates this optimization result into specific action instructions. It generates the following adjustment schemes: Instruction 1 (for the energy storage system): At T+15 minutes, begin discharging at 10MW power for 10 minutes. Instruction 2 (for the manufacturing workshop): At T+15 minutes, reduce the total power of the production line by 2MW via the frequency converter and maintain this for 10 minutes. This instruction has been verified through a digital twin model and will not affect product accuracy. Instruction 3 (for the dispatch center): Generate an early warning, indicating that there may be a 3MW power shortage in the park between T+15 minutes and T+25 minutes, and suggest preparing to apply for support from the upper-level power grid.

[0085] These instructions were transmitted via industrial Ethernet to the PCS (Power Conversion System) of the energy storage system and the central controller in the manufacturing workshop. The entire process, from prediction to decision-making to execution, was completed within minutes. This successfully anticipated and responded to a potential supply-demand imbalance event caused by a sudden change in wind direction, and was accomplished under the premise of fully considering and avoiding power quality risks, demonstrating the intelligence, precision, and high security of the method proposed in this application.

[0086] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0087] It should be noted that in the description of this application, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this application, unless otherwise stated, "a plurality of" means at least two.

[0088] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.

[0089] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0090] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0091] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc.

[0092] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0093] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. An industrial load regulation method based on digital twins, characterized in that, Includes the following steps: Real-time and historical operating data of industrial load, energy storage device and new energy output are collected and preprocessed and synchronized with time; wherein, the real-time data includes at least the active power of industrial load, voltage of energy storage device and power output of new energy. A digital twin model of the industrial load is established based on the historical operating data of the industrial load, and the real-time data is mapped to the digital twin model of the industrial load for dynamic updates; Based on the updated digital twin model of industrial load, the volatility, flexibility and responsiveness of each load unit are calculated and integrated to form an adjustable capability vector of each load unit. Based on real-time data of renewable energy output, a renewable energy power prediction method based on wind direction adaptive attention mechanism is used to perform short-term prediction and generate future power sequences. Based on the adjustable capacity vector and the future power sequence, the industrial load's ability to support new energy absorption and the absorption rate are calculated in real time using a dynamic assessment method based on voltage flicker risk propagation. Based on the calculation results of the absorption capacity and absorption rate, a corresponding load adjustment scheme is generated.

2. The method according to claim 1, characterized in that, The establishment of the industrial load digital twin model based on the historical operating data of the industrial load includes: Extract load dynamic characteristics from the historical operational data; A basic industrial load model is constructed based on long short-term memory networks to describe the dynamic response relationship of load under different operating conditions. The topology between load units is constructed based on the physical connections and power transmission relationships of the industrial system, and the topology is embedded into the basic industrial load model. At the same time, the characteristics of energy storage units and new energy equipment are incorporated into the model to generate the digital twin model of the industrial load.

3. The method according to claim 1, characterized in that, The updated industrial load digital twin model is used to calculate the volatility, flexibility, and responsiveness of each load unit, including: For each load unit, the standard deviation of its active power is calculated within a specified time window to obtain the load fluctuation rate; Based on the aforementioned digital twin model, the adjustable range of the load unit under different adjustment commands is simulated to obtain the flexibility. The response of a simulated load unit to a regulation command is used to obtain a response capability index, which describes the load's ability to quickly adjust power after receiving a regulation command. The load volatility, flexibility, and responsiveness indicators are integrated into the adjustable capability vector.

4. The method according to claim 1, characterized in that, The renewable energy power prediction method based on wind direction adaptive attention mechanism includes: Real-time wind direction data, including wind direction angle and wind direction change rate, is collected and synchronized with the power output data of new energy sources to construct a power time series feature vector. The wind direction angle is divided into multiple azimuth intervals, and the initial power output correction coefficient for each azimuth interval is calculated based on historical data. Calculate the wind direction fluctuation index based on the time series of wind direction angle; Construct a mapping relationship between wind direction zones and power output to form a wind direction feature vector; A wind direction attention layer is introduced into the long short-term memory network model to jointly model the wind direction feature vector and the power time series feature vector to generate a comprehensive feature sequence. The prediction time window is dynamically adjusted based on the wind direction change rate. The corresponding correction coefficient is obtained based on the current wind direction and its azimuth range, and the correction coefficient is updated on a rolling basis in combination with recent actual power output data; The updated comprehensive feature sequence is input into the trained long short-term memory network model, and the future power sequence is output.

5. The method according to claim 4, characterized in that, The method of introducing a wind direction attention layer into the long short-term memory network model to jointly model the wind direction feature vector and the power time series feature vector includes: The wind direction feature vector and the power time series feature vector are aligned and fused in the time dimension to form a multimodal input feature matrix; The correlation weights of upwind characteristics to power changes at each time step are calculated using the wind direction attention layer. Based on the soft attention mechanism, the correlation weights are used to weight the wind direction feature vectors at different time steps in the multimodal input feature matrix; The weighted wind direction feature vector is fused with the power time series feature to generate the comprehensive feature sequence, which is then input into the hidden state update unit of the long short-term memory network.

6. The method according to claim 1, characterized in that, The dynamic assessment method for absorption capacity based on voltage flicker risk propagation includes: Based on the power grid topology, a path network model for voltage flicker propagation is established, and the flicker influence factor matrix between nodes is calculated. Real-time monitoring of short-term flicker values ​​at each node, and calculation of cumulative voltage flicker values ​​using a moving average; Based on the real-time monitored voltage flicker value, a nonlinear function is used to calculate the voltage flicker risk index of each node, and it is mapped to the constraint factor of the adjustable power of the node to form the adjustable power range under risk constraints. Based on the risk-constrained adjustable power range and the future power sequence, combined with the length of the voltage flicker propagation path and the attenuation coefficient, the total adjustable power of the system is calculated. With the objectives of maximizing the utilization rate of new energy output and minimizing the node risk power deviation, a multi-objective optimization function considering flicker constraints is constructed. By solving this function, the absorption capacity and absorption rate under risk constraints are obtained.

7. The method according to claim 6, characterized in that, The voltage flicker value based on real-time monitoring is used to calculate the voltage flicker risk index of each node using a nonlinear function, and then mapped to a constraint factor for the adjustable power of the node, forming an adjustable power range under risk constraints, including: Based on short-term flicker values ​​and cumulative flicker values, calculate the flicker risk index for each node; Map the flash risk index to a constraint factor; Multiply the constraint factor by the node's own maximum physical adjustable power to obtain the upper limit of adjustable power under risk constraints; and establish an adjustable power model under risk constraints based on the upper limit of adjustable power. The adjustable power of the risk constraint is obtained by solving the adjustable power model of the risk constraint.

8. The method according to claim 6, characterized in that, The calculation of the total adjustable power of the system based on the risk-constrained adjustable power range and the future power sequence, combined with the length and attenuation coefficient of the voltage flicker propagation path, includes: Based on risk-constrained adjustable power and future power sequences, the length and attenuation coefficient of the voltage flicker propagation path are obtained. Construct a path attenuation matrix based on the length and attenuation coefficient of the voltage flicker propagation path; By combining the flicker influence factor matrix, the effective contribution of each node to the sensitive node is calculated; The candidate value of the total adjustable power of the system is obtained by summing the effective contributions of all nodes. By adjusting the candidate values ​​using a linear programming algorithm, the final total adjustable power of the system is obtained while satisfying the risk constraints and power balance of each node.

9. The method according to claim 1, characterized in that, The step of generating a corresponding load adjustment scheme based on the calculation results of the absorption capacity and absorption rate includes: Based on the adjustable capacity vector of each load unit, a candidate operation list for load start-up and power regulation is generated; Based on the state of the energy storage unit and the future power sequence, candidate operations for energy storage charging and discharging are generated; The load and energy storage candidate operations are combined to form a set of system-level regulation schemes and their feasibility is verified. The verified solution is then translated into actual control commands.

10. An industrial load regulation system based on digital twins, characterized in that, include: The data acquisition and processing module is used to collect real-time data and historical operating data of industrial load, energy storage device and new energy output, and to perform preprocessing and time synchronization; wherein, the real-time data includes at least the active power of industrial load, voltage of energy storage device and power output of new energy. The industrial load digital twin modeling and updating module is used to establish an industrial load digital twin model based on the historical operating data of the industrial load, and to map the real-time data to the industrial load digital twin model for dynamic updating; The load adjustability assessment module is used to calculate the volatility, flexibility and response capability of each load unit based on the updated industrial load digital twin model, and integrate them to form the adjustability vector of each load unit. The new energy power prediction module is used to make short-term predictions based on real-time data of new energy output and a new energy power prediction method based on wind direction adaptive attention mechanism, and generate future power sequences. The dynamic evaluation module for absorption capacity is used to calculate in real time the industrial load's ability to support new energy absorption capacity and absorption rate based on the adjustable capacity vector and the future power sequence, using a dynamic evaluation method for absorption capacity based on voltage flicker risk propagation. The load regulation scheme generation module is used to generate a corresponding load regulation scheme based on the calculation results of the absorption capacity and absorption rate.

Images