Multi-main variable coordination dynamic capacity adjustment multi-target precise evaluation and closed-loop regulation method

Abstract

The application relates to a multi-main transformer coordination dynamic capacity regulation multi-target precise evaluation and closed-loop regulation method, relates to the transformer regulation technical field, breaks through the traditional single-main transformer rated capacity constraint, constructs a trinity regulation system of'multi-main transformer global coordination + hierarchical precise evaluation + edge-main station closed loop', collects the insulation aggregation degree (DP) and oil furfural content of the main transformer in real time through the deployment of an edge terminal, dynamically back calculates a thermal aging state, calculates individualized available margin, quickly screens a safe and feasible scheme based on a voltage / frequency sensitivity model, removes a high-risk capacity regulation strategy in combination with the thermal aging kinetic quantization life loss cost, adopts a deep reinforcement learning intelligent agent, cooperatively optimizes the capacity regulation amplitude, time sequence and power flow distribution of each main transformer, and realizes millisecond-level risk suppression through 100ms-level edge early warning and step-by-step back-off mechanism, and forms a 'perception-decision-execution-learning-evolution' closed loop.

Description

Technical Field

[0001] This application relates to the field of transformer control technology, specifically to a method for multi-transformer coordinated dynamic capacity adjustment with multi-objective precise evaluation and closed-loop control. Background Technology

[0002] As is well known, current dynamic capacity adjustment technologies for main transformers in power systems are mostly based on independent assessment and control of a single main transformer, using rated capacity as a hard constraint or simply switching between start-up and shutdown based on load rate thresholds. They do not consider the potential for power flow coupling and capacity coordination among multiple main transformers within a substation cluster. Traditional methods typically use fixed thresholds to preset capacity adjustment conditions, lacking the ability to dynamically perceive and quickly predict the real-time power flow distribution of the grid, voltage-sensitive nodes, and system frequency response characteristics. This results in delayed capacity adjustment decisions and slow response, making it difficult to adapt to the drastic load fluctuations under the high penetration of new energy sources. At the same time, existing technologies generally ignore the differentiated impact of the aging state of main transformer insulation and have not established a quantitative model of life loss based on thermal aging dynamics. In pursuing capacity release, this can easily lead to accelerated insulation degradation, creating hidden dangers for equipment reliability. In addition, the control strategies are mostly static open-loop designs, lacking real-time risk monitoring and closed-loop feedback mechanisms during execution. Once abnormal conditions such as sudden temperature rise or voltage exceeding limits occur, timely intervention is not possible, posing a risk of safety risk spillover. At the system level, the communication architecture mostly adopts the full data upload mode, which leads to high pressure on the communication bandwidth of the edge side and heavy computing load of the main station, making it difficult to meet the millisecond-level response requirements. Moreover, once the control strategy is deployed, it is fixed for a long time and lacks the ability to learn and optimize based on operating experience, making it difficult to adapt to the long-term evolution of the power grid structure and operating conditions.

[0003] In the aforementioned technologies, the isolated evaluation mode of a single main transformer cannot coordinate the distribution of cluster capacity resources, resulting in structural waste of "unused surplus capacity and uneven distribution" in the region; the lack of a collaborative optimization mechanism among multiple objectives (system stability, equipment lifespan, and economic benefits) often leads to unbalanced decisions that sacrifice economy for safety or lifespan for increased revenue; risk management relies on manual inspections or post-event alarms, with responses lagging behind dynamic changes and unable to achieve millisecond-level closed-loop suppression; and due to the lack of a data-driven model update mechanism, the control strategy remains fixed for a long time, with poor generalization ability, making it difficult to adapt to the high uncertainty and strong volatility operating environment of new power systems, ultimately resulting in problems such as low regional capacity utilization, uncontrolled main transformer lifespan loss, and frequent voltage and frequency exceedances, which seriously restrict the scientific release of main transformer capacity potential and the improvement of the safety and economy of power grid operation. Therefore, we propose a multi-main transformer collaborative dynamic capacity adjustment multi-objective accurate evaluation and closed-loop control method. Summary of the Invention

[0004] This application provides a method for multi-transformer coordinated dynamic capacity adjustment with multi-objective precise assessment and closed-loop control, including the following steps: Deploying edge sensing terminals on each main transformer in the substation cluster to collect real-time main transformer operating parameters, grid operating parameters, and environmental parameters, using a hierarchical transmission mechanism: the edge terminals locally store fine-grained data for lifetime loss calculation, and only upload a global state summary to the regional control master station; based on real-time power flow and voltage distribution of the grid, constructing a voltage fluctuation sensitivity model and a frequency deviation response model to quickly predict the transient stability of candidate capacity adjustment schemes, and selecting a set of feasible schemes that meet the constraints of the "Guidelines for the Safety and Stability of Power Systems"; within the set of feasible schemes, constructing a multi-transformer global capacity margin collaborative model, comprehensively considering the insulation aging state of each main transformer and grid power flow constraints, calculating the available dynamic margin of each main transformer, and obtaining the total regional dynamic capacity margin through weighted summation using margin adjustable coefficients; simultaneously, based on the thermal aging dynamic model... The insulation lifetime loss cost of each main transformer during capacity adjustment is calculated, and schemes with lifetime loss exceeding a preset threshold are eliminated. The remaining schemes are modeled as Markov decision processes, and a deep reinforcement learning agent is used to maximize the net benefit of the region, outputting the target load rate, capacity adjustment timing, and power flow allocation ratio of each main transformer. The net benefit is the capacity adjustment gain minus the lifetime loss conversion cost and operating cost. During capacity adjustment, the edge terminal monitors the temperature change rate, voltage deviation rate, and frequency deviation rate every 100ms. If any indicator exceeds the limit, an early warning is triggered, and the main station reconstructs the collaborative strategy in real time to suppress the risk. If an extreme risk occurs, a step-by-step rollback mechanism is triggered to restore the original state. After each adjustment, the full-cycle operation data, risk events, and actual benefits are stored in the experience database for offline self-updating of the deep reinforcement learning agent. The cumulative lifetime loss of the main transformer is checked monthly, and the aging model and benefit coefficient are updated quarterly to achieve continuous optimization of the control strategy.

[0005] The aforementioned method for multi-transformer coordinated dynamic capacity adjustment with multi-objective precise assessment and closed-loop control can be used to calculate the dynamic margin using the following formula: in, For the first The maximum allowable overload capacity of the main transformer under the condition of not exceeding the preset life loss threshold is dynamically determined by back-calculating the thermal aging state based on the current degree of polymerization of the insulation paper and the furfural content.

[0006] As described above, the multi-objective precise assessment and closed-loop control method for multi-master dynamic capacity adjustment, the formula for calculating the regional total dynamic capacity margin is: in, For the first The margin adjustment coefficient of the main transformer is calculated by comprehensively considering its voltage sensitivity to key nodes and power flow transmission capacity, reflecting its contribution weight in the coordinated capacity adjustment of the power grid topology.

[0007] As mentioned above, the multi-transformer coordinated dynamic capacity adjustment multi-objective precise assessment and closed-loop control method, insulation life loss cost The calculations are based on a thermal aging kinetic model: in, Temperature of the main transformer winding hot spot (°C). The relative aging rate is used to calculate the cumulative aging amount, which is calculated in conjunction with the duration of capacity adjustment, T (in hours). And it is calculated as an equivalent economic loss, denoted as If the cumulative life loss in a single cycle exceeds 0.01% of the total life of the main transformer, the capacity adjustment scheme will be eliminated.

[0008] As described above, the multi-master variable collaborative dynamic capacity adjustment multi-objective precise assessment and closed-loop control method, regional net income The calculation formula is: in,

[0009] The new transmittable power is represented by T, where T is the duration of capacity adjustment and P is the real-time electricity price. To reduce the operating and communication costs of the capacity adjustment switch, Costs are calculated based on insulation aging.

[0010] As described above, the multi-transformer collaborative dynamic capacity adjustment multi-objective precise evaluation and closed-loop control method utilizes a deep reinforcement learning agent whose state space includes: current load rate of each transformer, hotspot temperature, remaining insulation aging rate, grid node voltage deviation, system frequency, and electricity price signal; its action space comprises the target capacity adjustment amplitude and timing sequence for each transformer; and its reward function is defined as: in, The newly added transmittable power of the system after capacity adjustment, in MW;

[0011] p represents the real-time electricity price, expressed in yuan / (kW·h);

[0012] The cost is calculated based on the insulation life loss as defined in claim 4, in yuan.

[0013] The system stability penalty term is defined as the weighted sum of squares of the voltage deviation rate and the frequency deviation rate, i.e.: in The percentage represents the voltage deviation at critical nodes. Rated voltage, The system frequency deviation is expressed in Hz. The rated frequency is 50 Hz. , These are preset weighting coefficients, with typical values ​​of 0.7 and 0.3 respectively; A positive reward weighting coefficient is used to balance the relative importance of economic gains, lifetime losses, and system stability. Its value is dynamically configured by the power grid operation strategy and is adjusted online based on historical returns and operational risks during the model self-update phase.

[0014] As described above, in the multi-transformer collaborative dynamic capacity adjustment multi-objective precise assessment and closed-loop control method, the edge terminal has local risk identification and millisecond-level early warning capabilities, while the master station has strategy reconstruction and step-by-step backoff control capabilities. When the winding hot spot temperature exceeds 140℃ or the voltage deviation exceeds ±5%, the master station initiates a three-level response: ① adjust the power flow ratio; ② shorten the capacity adjustment time; ③ gradually restore the original operating state step by step.

[0015] As mentioned above, the multi-transformer collaborative dynamic capacity adjustment multi-objective accurate assessment and closed-loop control method includes a model self-updating mechanism that includes: monthly full-cycle lifetime cumulative verification of all participating transformers and updating their remaining lifetime parameters; and quarterly online fine-tuning of thermal aging model parameters, electricity price revenue coefficients, and deep reinforcement learning neural network structure based on historical experience databases to improve the strategy's generalization ability.

[0016] As described above, in the multi-transformer collaborative dynamic capacity adjustment multi-objective accurate assessment and closed-loop control method, in the hierarchical transmission mechanism, the edge terminal only uploads the following summary data to the main station: transformer number, current load rate, hot spot temperature, insulation aging remaining rate, adjustable margin, voltage influence coefficient, and capacity adjustment feasibility indicator. The remaining fine-grained data is retained locally for high-precision life assessment, reducing communication bandwidth by more than 70%.

[0017] The multi-transformer collaborative dynamic capacity adjustment multi-objective precise assessment and closed-loop control method mentioned above is applicable to substation clusters containing 3 or more main transformers. It is suitable for distribution network scenarios with high penetration of new energy and drastic load fluctuations. Compared with the traditional rated capacity constraint method, the regional capacity utilization rate is increased by 15% to 25%, the annual average life loss of main transformers is reduced by more than 40%, and the probability of voltage / frequency exceeding the standard is reduced by more than 80%.

[0018] Applying the technical solution of this application, edge sensing terminals are deployed on each main transformer in the substation cluster. A hierarchical transmission mechanism enables efficient local storage of fine-grained data and uploading of global status summaries, overcoming the bandwidth bottleneck and latency defects of traditional centralized data acquisition. Based on real-time power flow and voltage distribution of the power grid, voltage fluctuation sensitivity and frequency deviation response models are constructed to predict the transient stability of candidate capacity adjustment schemes at the millisecond level, ensuring that only feasible schemes conforming to the "Guidelines for the Safety and Stability of Power Systems" are retained, thus mitigating system instability risks from the source. Furthermore, a multi-main transformer global capacity margin collaborative model is constructed, integrating the insulation aging status of each main transformer with power flow constraints to achieve collaborative quantification and weighted aggregation of dynamic margins for multiple devices, completely changing the situation of low regional capacity utilization caused by isolated assessment of a single main transformer. Simultaneously, a thermal aging kinetic model is introduced to accurately quantify the insulation life loss cost during capacity adjustment, and a rigid threshold is set to eliminate schemes exceeding the limit, achieving for the first time a coupled constraint of the three objectives of safe operation, equipment life, and economic benefits in the scheme generation stage. The remaining schemes are modeled as Markov decision processes. The system, relying on deep reinforcement learning agents, optimizes regional net revenue by dynamically outputting the optimal target load rate, capacity adjustment timing, and power flow allocation ratio, achieving a leap from static rules to intelligent decision-making. During the execution phase, edge terminals monitor temperature change rate, voltage deviation rate, and frequency deviation rate in real time at 100ms intervals. Once these limits are exceeded, an early warning is triggered, driving the main station to reconstruct the strategy in real time, forming a closed loop of "perception-decision-response," completely reversing the passive situation where static prediction lags behind dynamic risks. When encountering extreme risks, a step-by-step rollback mechanism can accurately restore the system to a safe state, ensuring zero system crashes. After each regulation, full-cycle data, risk events, and actual revenue are stored in an experience base, driving the deep reinforcement learning agent to iterate offline. Through monthly lifespan verification and quarterly model update mechanisms, the aging model and revenue weights are continuously optimized, allowing the regulation strategy to dynamically evolve with equipment status and grid characteristics. Ultimately, this achieves long-term optimal operation of the substation cluster in terms of safety, lifespan, and efficiency, systematically solving the three core problems of low capacity utilization, uncontrolled lifespan loss, and slow risk response. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall system architecture of the present invention. Detailed Implementation

[0020] The present invention will be explained in detail through the following embodiments. The purpose of disclosing the present invention is to protect all technical improvements within the scope of the present invention.

[0021] The multi-transformer coordinated dynamic capacity adjustment multi-objective precise assessment and closed-loop control method according to this embodiment includes:

[0022] Step 1: Deploy edge sensing terminals on each main transformer in the substation cluster to collect real-time operating parameters of the main transformer, grid operating parameters, and environmental parameters. The operating parameters include load rate, winding hot spot temperature, top oil temperature, and furfural content in insulating oil. The grid operating parameters include node voltage, system frequency, and active / reactive power flow. The environmental parameters include ambient temperature, humidity, and wind speed.

[0023] The edge-sensing terminal employs a layered data processing and transmission mechanism:

[0024] Local storage layer: retains all fine-grained raw sampling data (sampling frequency ≥ 1 Hz) for high-precision insulation aging status assessment and lifetime loss calculation;

[0025] Abstract Upload Layer: Only preprocessed global status summary information is uploaded to the regional control master station at fixed intervals (≤5 s); the summary information includes: main transformer number, current load rate, hot spot temperature, insulation aging remaining rate, available dynamic margin, voltage influence coefficient and capacity adjustment feasibility indicator, but does not include fine-grained data such as original temperature curves, oil chromatography data, and voltage waveforms.

[0026] This hierarchical mechanism ensures global collaborative decision-making at the main station while significantly reducing communication bandwidth requirements. Actual measured communication traffic is more than 70% lower than the traditional full-upload method, meeting the engineering constraints of limited network resources at the substation edge.

[0027] Step 2: Based on the real-time power flow distribution and node voltage data collected from the substation cluster, construct a voltage fluctuation sensitivity matrix and a frequency deviation response dynamic model to predict the transient stability of candidate capacity adjustment schemes at the millisecond level.

[0028] The voltage fluctuation sensitivity matrix The value is obtained by inverting the system's Jacobian matrix, where n is the number of key monitoring nodes, m is the number of changes in the main transformer's capacity, and its elements are... Indicates the first The platform owner changes from having merit to contributing effort. For the Node voltage The influence coefficient;

[0029] The frequency deviation response model is established based on the system inertia H and damping coefficient D, and its dynamic equation is:

[0030]

[0031] in The power injection change caused by the capacity adjustment of the k-th main transformer. For load fluctuation compensation, H is in MW·s / MVA and D is in MW / Hz.

[0032] Based on the permissible voltage deviation limits (±5%) and permissible frequency deviation limits (±0.2 Hz) specified in the "Guidelines for the Safety and Stability of Power Systems" (DL / T 755-2001), real-time simulations were performed on all candidate capacity adjustment schemes.

[0033] If the predicted value of voltage deviation at any node or system frequency deviation prediction value If so, then the plan is excluded;

[0034] Only schemes where all node voltages and system frequencies meet the above limits are retained, forming the feasible scheme set F.

[0035] The prediction process is completed locally at the edge terminal or the main station, with a single round of evaluation taking ≤2.5 seconds, meeting the real-time control and response requirements of the power grid.

[0036] Step 3: Within the feasible solution set F generated in step (2), construct a multi-transformer global capacity margin collaborative model, and dynamically calculate the insulation aging residual rate of each transformer and the power flow transmission constraints of the power grid. Available dynamic margin Based on this, the total dynamic capacity margin of the region is calculated. .

[0037] The available dynamic margin As defined in claim 2, it is calculated by the following formula:

[0038] For the first Current load capacity (MVA) of the main transformer. The maximum allowable overload capacity of the main transformer, provided it does not exceed the annual life loss threshold (0.01%), is dynamically determined by a thermal aging inverse model based on the current degree of polymerization (DP) of the insulation paper and the furfural content (ppm) in the oil. The expression is as follows:

[0039]

[0040] in For rated capacity, The current aggregation degree, For reference values ​​of new insulating paper, This is an empirical correction factor, set according to the State Grid's "Guidelines for Insulation Aging Assessment of Power Transformers" (Q / GDW 11745-2017).

[0041] The margin adjustable coefficient As defined in claim 3, by the main variable Sensitivity to critical node voltage With maximum power transmission capacity Joint weighted calculation:

[0042]

[0043] in The voltage regulation weight is set based on operational experience. This coefficient represents all main transformers within the cluster; it reflects their "coordinated capacity adjustment contribution weight" in the power grid topology.

[0044] Regional total dynamic capacity margin As defined in claim 3:

[0045]

[0046] Simultaneously, for each candidate solution, a thermal aging kinetic model is used to quantify the insulation life loss cost. According to claim 4, the winding hot spot temperature The relationship with the relative aging rate V is as follows:

[0047]

[0048] If the main transformer... The hotspot temperature remained above the threshold.

[0049] Then its cumulative aging amount for:

[0050]

[0051] in Let be the relative aging rate for the kth sampling period (Δt_k = 10s). ,

[0052] If the cumulative aging of any main transformer under this scheme (Assuming a certain percentage of its total design life), then this scheme is removed from set F, specifically for main transformers with severely aged insulation (i.e., current degree of aggregation). The system further introduces a safety redundancy factor. For the maximum permissible overload capacity Secondary revision:

[0053]

[0054] in, Dynamically set based on the remaining life assessment value of the main transformer:

[0055] like ,but (Under normal conditions, no restrictions); if ,but (Moderate old saying, moderately conservative); If ,but (Severe aging, forced conservative approach);

[0056] This mechanism originates from the mandatory recommendation in the State Grid's "Safety Regulations for Transformer Operation" (Q / GDW 11746-2017) that "high-risk equipment should have reduced operating margins," and is further derived from the analysis of 217 temperature rise mutation events caused by "compliant overload" in the historical experience database of this invention. Even if the conditions are met Its insulation material has entered the accelerated aging stage, and the actual life decay rate is more than 35% higher than the model prediction, forming the final "safety-economic dual-constraint feasible solution set" F∗.

[0057] Step 4: Model the candidate capacity adjustment actions in the "safety-economic dual-constraint feasible solution set" F∗ generated in step (3) as a Markov Decision Process (MDP), and use a deep reinforcement learning (DRL) agent to make long-term net benefit maximization decisions, outputting the target load rate, capacity adjustment start time and power flow allocation ratio of each main transformer.

[0058] The MDP quintuple is defined as follows:

[0059] state space : Contains the current state of n main transformers, each main transformer corresponding to 7-dimensional features:

[0060]

[0061] in:

[0062] : No. Current load rate of the main transformer (per unit value);

[0063] : Winding hot spot temperature;

[0064] Insulation aging residual rate ( (Total aging amount designed).

[0065] : Critical node voltage deviation (based on step 2) calculate);

[0066] System frequency deviation (Hz);

[0067] Yuan / kWh: Real-time electricity price (from electricity market data);

[0068] Safety redundancy coefficient (from step 3);

[0069] Action space For each main transformer Output two consecutive actions:

[0070]

[0071] Target load rate adjustment amount (e.g., +0.05 indicates an increase of 5%).

[0072] s: Capacity adjustment action start delay (unit: seconds), used to achieve timing coordination and avoid system impact caused by the simultaneous operation of multiple main transformers;

[0073] The reward function R is defined as follows according to claim 6:

[0074]

[0075] in: Total new transmittable power in the region (MW);

[0076] Real-time electricity price (RMB / kWh);

[0077] : Cost of insulation life loss calculated according to claim 4 (in yuan);

[0078] :

[0079] Voltage / frequency over-limit penalty item ( , (Unit: Yuan / per unit), exceeding ±3% or ±0.1Hz will trigger a penalty to strengthen the safety boundary; Yuan / time: Cost of a single capacity adjustment operation and communication (including relay action, fiber optic communication, and control command issuance);

[0080] Weighting coefficients: The method was determined through backtesting and optimization of historical data to ensure a balance between economic efficiency and security.

[0081] Transition probability The system is built offline using a power system flow simulation model (based on OpenDSS). It performs 10,000 Monte Carlo simulations on typical load-new energy scenarios (such as a sudden drop in photovoltaic output and a surge in charging load) to form a state transition statistics table for DRL training.

[0082] Discount factor: This encourages maximizing long-term benefits.

[0083] The DRL agent employs the Proximal Policy Optimization (PPO) algorithm, with a three-layer fully connected neural network structure (128→64→2n). The input is the state vector s, and the output is the action mean and variance (used for continuous action space sampling). The agent is pre-trained on an offline experience base (containing ≥1,500 historical control records) and the network parameters are updated after each control execution to achieve online policy adaptation.

[0084] The final output is:

[0085] Target load rate for each main transformer ;

[0086] Adjusting startup timing Execute in ascending order;

[0087] The power flow distribution ratio is automatically calculated by the system's power flow calculation module to ensure that the node voltage and line thermal limit are not exceeded.

[0088] The DRL agent output T1 is adjusted to 103% (Δρ=+0.03), with a 30-second delay before startup; T2 is reduced to 68% (Δρ=-0.02) after 60 seconds, while T3 remains unchanged. The power flow is recalculated using OpenDSS, reducing the voltage offset from +3.8% to +2.1%, resulting in a net benefit of NT$2,370, a 19% improvement over the traditional rule-based method.

[0089] Step 5: During the execution of the coordinated capacity adjustment command output in step (4), the edge sensing terminals deployed on each main transformer collect and calculate three dynamic risk indicators in real time at a period of 100 ms (Δt = 0.1 s):

[0090] (1) Winding hot spot temperature change rate:

[0091]

[0092] (2) Rate of change of voltage deviation at critical nodes:

[0093]

[0094] (3) System frequency deviation rate of change:

[0095]

[0096] The warning thresholds for each indicator are set according to the "Guidelines for the Safety and Stability of Power Systems" (DL / T 755-2001) and the historical risk event database of this invention, specifically as follows:

[0097] like This triggers a Level 1 warning for a sudden temperature rise.

[0098] like This triggers a voltage surge warning (Level 2).

[0099] like This triggers a warning for drastic frequency fluctuations (Level 3).

[0100] When any warning is triggered, the edge terminal immediately uploads an event identifier + timestamp + raw data fragment (sampling window: first 500ms) to the regional master station via the 5G private network. After receiving the warning signal, the master station completes the following operations within **≤200ms**:

[0101] ① Risk level assessment and response reconstruction:

[0102] If it is a Level 1 warning (sudden temperature rise): the main station calls the "emergency load reduction" strategy module of the DRL agent and outputs corrective actions. Forcefully reduce the capacity adjustment range of the main transformer as the hot spot temperature rises, and adjust the amplitude correction coefficient. That is, the new target load rate

[0103] ;

[0104] If it is a Level 2 or 3 warning (voltage / frequency sudden change): The main station initiates the "power flow redistribution" mode. Based on the current topology and impedance model, it redistributes the output ratio of each main transformer through fast power flow calculation (Newton-Raphson, iterations ≤ 3 times) to achieve this. and ;

[0105] ② Step-by-step retreat mechanism for extreme risks:

[0106] If any indicator continues to deteriorate within 1 second after the warning is triggered and meets any of the following conditions:

[0107] ,or ,or ,

[0108] If this is determined to be an extreme risk event, the main site will immediately initiate a three-tiered rollback process, executing in order of priority:

[0109] level Back action Execution time Target ① Adjust the load rate of the main transformer with the maximum temperature rise back to 90% of the value before capacity adjustment. 1.5 s Suppressing hotspot temperature rise ② Shut down the capacity adjustment actions of other main transformers in reverse order (in descending order of τi). 2.0 s Restoring tidal balance ③ If the voltage / frequency still exceeds the limits, activate the backup tie line for reactive power compensation. 3.0 s Stabilize system voltage

[0110] All rollback actions are controlled by the SCADA system. After the action is executed, the edge terminal provides feedback confirmation of the status, forming a closed loop of "command - execution - feedback" to ensure that the operation is traceable.

[0111] The mechanism was verified in Example 1: when the temperature change rate of T1 suddenly increased to 0.92°C / s, the master station initiated load reduction within 180 ms, reducing Δρ from +0.03 to +0.012. The temperature curve stabilized rapidly, avoiding triggering extreme rollback.

[0112] Compared to the traditional "threshold tripping" mode, this mechanism reduces the false tripping rate by 76%, shortens the extreme risk handling time to within 3 seconds (the traditional SCADA system takes an average of 8-15 seconds), and successfully avoids one risk of insulation damage caused by excessive temperature rise, ensuring the safe operation of the main transformer.

[0113] Step 6: After each capacity adjustment cycle, the system automatically stores the full cycle operation data, risk events and actual net income in the experience base E in a structured manner, as offline training samples for the deep reinforcement learning (DRL) agent, to achieve continuous self-optimization of the strategy; at the same time, the system performs insulation lifetime cumulative verification and aging model-income coefficient online fine-tuning at fixed intervals to ensure that the evaluation benchmark and the operating environment are dynamically synchronized.

[0114] (1) Experience base construction mechanism:

[0115] After each adjustment cycle k ends, the system extracts the following structured data and stores it in the experience base.

[0116] ,in:

[0117] The state at step t (same as claim 6, including load rate, hot spot temperature, remaining aging rate, voltage deviation, frequency, electricity price, and safety factor). );

[0118] : Corresponding DRL action (target load factor increment) Adjustment start-up delay );

[0119] Output of the real-time reward function as defined in claim 6;

[0120] The highest risk event level that occurred during this period (from step 5);

[0121] The final regional net profit (in yuan) realized in this period, according to the formula

[0122] calculate;

[0123] The experience base uses a sliding window + priority sampling structure, storing the most recent 2000 control records, and includes high-risk events (Level 3 / Fallback) and high-return samples (…). By assigning resampling weights, the learning efficiency of DRL in extreme scenarios can be improved.

[0124] (2) DRL agent offline self-updating mechanism:

[0125] From 2:00 to 4:00 daily, the system enters the offline training window and uses all samples in the experience base to fully retrain the DRL agent (PPO algorithm, neural network structure as in claim 6). The training consists of 10 rounds with a learning rate η = 0.0003 and a KL divergence threshold. To prevent strategy collapse.

[0126] The training objective function is:

[0127]

[0128] Where ϵ=0.2, and Aπ is the generalized advantage estimate (GAE), which ensures stable policy updates.

[0129] (3) Monthly lifespan accumulation verification mechanism:

[0130] On the 1st of each month, the system performs checks on each main transformer. Perform cumulative lifespan loss verification:

[0131]

[0132] in, The thermal aging rate (claim 4). The duration of the k-th capacity adjustment This represents the total number of times the main transformer participates in capacity adjustment.

[0133] Will Compared with model predictions Compare and calculate the deviation rate:

[0134]

[0135] like The system will then automatically initiate aging model parameter recalibration and update the reference temperature coefficient in the aging model. for:

[0136]

[0137] And update the remaining life parameters of the main transformer simultaneously. and safety redundancy coefficient .

[0138] 4) Quarterly return coefficients and model fine-tuning mechanism:

[0139] At the end of each quarter, the system uses transfer learning to fine-tune the following three sets of parameters online, based on all data from the past 90 days in the experience base:

[0140] Income weighting coefficient Based on historical net returns and risk event distribution, the objective function is optimized using gradient descent.

[0141]

[0142] Thermal aging model parameters The update is based on the monthly verification results and uses a weighted average (weight = number of monthly verification samples).

[0143] DRL Neural Network Structure: If the average reward increases by more than 15% for two consecutive quarters, the system will automatically try to expand the network layers (128→64→32→16 → 128→64→64→16) and make fine adjustments to improve the policy expression ability.

[0144] All updates are verified by the security verification module: the new strategy is run 1000 Monte Carlo tests in the simulation environment. If the average net profit decreases by more than 5% or the incidence of extreme events increases by more than 20%, the strategy will be rejected and the original strategy will be retained.

[0145] This mechanism was verified in the example: after T3 (DP=780) underwent 3 capacity adjustments, The result was 12.7% higher than the model prediction, and the system automatically updated α from 6.0 to 6.76. The value has been reduced from 0.85 to 0.7. In next month's capacity adjustment plan, T3 will be completely disabled to avoid premature retirement.

[0146] In actual testing, after 6 months of operation, the average reward of the DRL agent increased by 21%, the prediction lifetime error decreased from 18% to 5.3%, the annual main transformer retirement risk decreased by 43%, and the system became "smarter the more it is used", demonstrating significant long-term economic benefits and sustainability.

[0147] Example 1: As described above, the multi-transformer coordinated dynamic capacity adjustment multi-objective precise assessment and closed-loop control method can be used to calculate the dynamic margin using the following formula: in, The maximum allowable overload capacity of the i-th main transformer under the condition of not exceeding the preset life loss threshold is dynamically determined by back-calculating the thermal aging state from its current degree of polymerization of the insulation paper and furfural content.

[0148] The present invention further provides a dynamic calculation method for the available dynamic margin of the main transformer based on the insulation health status, which is used to break through the limitation of the "one-size-fits-all" rated capacity constraint in the traditional capacity adjustment technology, and realize personalized, state-driven accurate assessment of the capacity adjustment capability of each main transformer.

[0149] The available dynamic margin Defined as the maximum overload capacity increment that the i-th main transformer can safely bear without exceeding the preset insulation life loss threshold (e.g., 0.01% per hour), its calculation formula is:

[0150] in: The current actual load capacity (in MVA) of the i-th main transformer is collected in real time by the edge terminal; The maximum allowable overload capacity (unit: MVA) of the i-th main transformer under the current insulation health condition is not fixed, but dynamically calculated based on the real-time aging state of the main transformer insulation material. Specifically, it includes the following steps: (1) Collect the degree of polymerization (DP) of the main transformer insulation paper and the furfural content (FAL) in the insulating oil. These two are the core physical indicators reflecting the degree of thermal aging of paper insulation; (2) Based on the thermal aging life model recommended by the International Electrotechnical Commission standard IEC 60475 and IEEE C57.91, establish the exponential decay relationship between the degree of polymerization and aging time: in The initial degree of polymerization at the factory (typically 1200–1400), and k is the aging rate constant. The thermal aging rate is given by h⁻¹, and Th(τ) is the winding hot spot temperature; (3) Based on the current measured DPcurrent and FALcurrent, the current equivalent aging time of the main transformer is calculated in reverse. , and combined with the preset life end point (e.g., DPlim=250), calculate the remaining total thermal stress ΔAres; (4) Based on the nonlinear relationship between thermal stress and temperature rise, combined with the main transformer design thermal limit (e.g., the maximum allowable winding temperature 140℃), reverse-calculate the maximum continuous overload capacity that the main transformer can withstand under the condition that the life loss rate does not exceed δlim=0.01, Cmax−i, which is obtained by joint simulation of the main transformer temperature rise model and thermal time constant; (5) Finally, by Output the personalized available dynamic margin for each main transformer.

[0151] This method, for the first time, incorporates insulation aging status as a dynamic constraint into capacity margin calculation, realizing a new evaluation paradigm where "equipment health determines capacity authority." Compared to the traditional, crude method that uses a fixed overload factor of 110% or 120% of rated capacity, this method has the following significant technical advantages:

[0152] More precise assessment: The margin is no longer a "one-size-fits-all" approach, but varies independently according to the insulation health status of each main transformer. For example, a new main transformer (DP=1200) can be allowed +15% overload, while an aging main transformer (DP=750) can only be allowed +3%.

[0153] Safety Enhancement: By imposing hard constraints on lifespan loss thresholds, accelerated insulation degradation caused by frequent or excessive capacity adjustments is avoided, ensuring the safety of the main transformer throughout its entire lifespan.

[0154] Maximizing benefits: providing a quantifiable, comparable, and weighted input basis for the multi-objective collaborative optimization of steps (3)–(4), supporting the overall dynamic capacity margin of the region. The scientific construction;

[0155] Engineering compatibility: Compatible with existing transformer online monitoring systems (such as fiber optic temperature measurement and oil chromatography online monitoring), no new sensors are required, deployment only requires data interface integration.

[0156] verify:

[0157] T1 (New)

[0158] T3 (Old)

[0159] The significant differences in capacity allowance lead to fundamentally different capacity adjustment schemes, avoiding the risk of "over-squeezing" aging equipment in traditional methods.

[0160] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0161] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for multi-transformer collaborative dynamic capacity adjustment with multi-objective precise evaluation and closed-loop control, characterized in that, Includes the following steps: (1) Deploy edge sensing terminals on each main transformer in the substation cluster to collect main transformer operating parameters, power grid operating parameters and environmental parameters in real time. Adopt a hierarchical transmission mechanism: the edge terminal stores fine-grained data locally for lifetime loss calculation and only uploads the global status summary to the regional control master station. (2) Based on the real-time power flow and voltage distribution of the power grid, a voltage fluctuation sensitivity model and a frequency deviation response model are constructed to make a rapid prediction of the transient stability of the candidate capacity adjustment schemes and to select a set of feasible schemes that meet the constraints of the "Guidelines for the Safety and Stability of Power Systems". (3) Within the set of feasible solutions, a global capacity margin coordination model for multiple main transformers is constructed. The insulation aging status of each main transformer and the power flow constraints of the power grid are combined to calculate the available dynamic margin of each main transformer. The total dynamic capacity margin of the region is obtained by weighted summation through the adjustable margin coefficient. At the same time, the insulation life loss cost of each main transformer during the capacity adjustment process is quantified based on the thermal aging dynamics model, and the schemes with life loss exceeding the preset threshold are eliminated. (4) The remaining schemes are modeled as Markov decision processes. Deep reinforcement learning agents are used to maximize the net regional revenue. The target load rate, capacity adjustment timing and power flow allocation ratio of each main transformer are output. The net revenue is the capacity adjustment gain minus the lifetime loss conversion cost and operating cost. (5) During the capacity adjustment process, the edge terminal monitors the temperature change rate, voltage deviation rate and frequency deviation rate every 100ms. If any indicator exceeds the limit, an early warning is triggered and the main station reconstructs the collaborative strategy in real time to suppress the risk. If an extreme risk occurs, a step-by-step rollback mechanism will be triggered to restore the original state; (6) After each regulation ends, the full cycle operation data, risk events and actual benefits are stored in the experience library for offline self-updating of the deep reinforcement learning agent. The cumulative life loss of the main transformer is checked monthly, and the aging model and benefit coefficient are updated quarterly to achieve continuous optimization of the regulation strategy. For the main transformer with severe insulation aging, its maximum allowable overload capacity is dynamically reduced through the safety redundancy coefficient to form a hierarchical safety constraint mechanism based on the aging stage.

2. The method for multi-transformer coordinated dynamic capacity adjustment and multi-objective precise evaluation and closed-loop control as described in claim 1, characterized in that, The formula for calculating the available dynamic margin is as follows: in, For the first The maximum allowable operating capacity of the main transformer is dynamically determined based on its current insulation aging state, provided that it does not exceed the preset life loss threshold. For the first The current actual load capacity of the main transformer.

3. The method for multi-transformer coordinated dynamic capacity adjustment and multi-objective precise evaluation and closed-loop control as described in claim 1 or 2, characterized in that, The formula for calculating the total dynamic capacity margin of the region is as follows: （2） in, For the first The margin adjustment coefficient of the main transformer is defined as in claim 2; For the first The margin adjustment coefficient of the main transformer is determined by its comprehensive contribution weight to the voltage sensitivity and power flow transmission capacity of key nodes in the power grid, reflecting the cooperative capacity adjustment potential of the main transformer in the regional power grid topology.

4. The method for multi-transformer coordinated dynamic capacity adjustment and multi-objective precise evaluation and closed-loop control as described in claim 1, characterized in that, The cost of insulation life loss The calculations are based on a thermal aging kinetic model: in, Temperature of the main transformer winding hot spot (°C). The relative aging rate is used to calculate the cumulative aging amount, which is calculated in conjunction with the duration of capacity adjustment, T (in hours). And it is calculated as an equivalent economic loss, denoted as If the cumulative life loss in a single cycle exceeds 0.01% of the total life of the main transformer, the capacity adjustment scheme will be eliminated.

5. The method for multi-transformer coordinated dynamic capacity adjustment and multi-objective precise evaluation and closed-loop control as described in claim 1, characterized in that, The region's net income The calculation formula is: in, The new transmittable power is represented by T, where T is the duration of capacity adjustment and P is the real-time electricity price. Costs related to the operation and communication of the capacity adjustment switch; The The value is the cost of insulation life loss calculated according to claim 4. ,in The unit aging cost coefficient is determined based on the purchase cost, expected lifespan, and residual value of the main transformer.

6. The method for multi-transformer coordinated dynamic capacity adjustment and multi-objective precise evaluation and closed-loop control as described in claim 1, characterized in that, The state space of the deep reinforcement learning agent includes: the current load rate of each main transformer, hotspot temperature, remaining insulation aging rate, grid node voltage deviation, system frequency, and electricity price signal; the action space consists of the target capacity adjustment amplitude and adjustment timing for each main transformer; the reward function is defined as: in, The newly added transmittable power of the system after capacity adjustment, in MW; p represents the real-time electricity price, expressed in yuan / (kW·h); The cost is calculated based on the insulation life loss as defined in claim 4, in yuan. The system stability penalty term is defined as the weighted sum of squares of the voltage deviation rate and the frequency deviation rate, i.e.: in The percentage represents the voltage deviation at critical nodes. Rated voltage, The system frequency deviation is expressed in Hz. The rated frequency is 50 Hz. , These are preset weighting coefficients, with typical values ​​of 0.7 and 0.3 respectively; A positive reward weighting coefficient is used to balance the relative importance of economic gains, lifetime losses, and system stability. Its value is dynamically configured by the power grid operation strategy and is adjusted online based on historical returns and operational risks during the model self-update phase.

7. The method for multi-transformer coordinated dynamic capacity adjustment and multi-objective precise evaluation and closed-loop control as described in claim 1, characterized in that, In the edge-master closed-loop control mechanism, the edge terminal has local risk identification and millisecond-level early warning capabilities, while the master station has strategy reconstruction and step-by-step rollback control capabilities. When the winding hot spot temperature exceeds 140℃ or the voltage deviation exceeds ±5%, the master station initiates a three-level response: ① adjust the power flow ratio; ② shorten the capacity adjustment time; ③ gradually restore the original operating state step by step.

8. The method for multi-transformer coordinated dynamic capacity adjustment and multi-objective precise evaluation and closed-loop control as described in claim 1, characterized in that, The model self-updating mechanism includes: monthly full-cycle lifetime cumulative verification of all main transformers participating in capacity adjustment, and updating their remaining lifetime parameters; quarterly online fine-tuning of thermal aging model parameters, electricity price revenue coefficient and deep reinforcement learning neural network structure based on historical experience database, to improve the strategy generalization ability.

9. The method for multi-transformer coordinated dynamic capacity adjustment and multi-objective precise evaluation and closed-loop control as described in claim 1, characterized in that, In the aforementioned layered transmission mechanism, the edge terminal only uploads the following summary data to the main station: main transformer number, current load rate, hot spot temperature, insulation aging remaining rate, adjustable margin, voltage influence coefficient, and capacity adjustment feasibility indicator. The remaining fine-grained data is retained locally for high-precision life assessment, reducing communication bandwidth by more than 70%.

10. The method for multi-transformer coordinated dynamic capacity adjustment and multi-objective precise evaluation and closed-loop control as described in any one of claims 1 to 9, characterized in that, This method is applied to substation clusters containing three or more main transformers, and is suitable for distribution network scenarios with high penetration of new energy sources and drastic load fluctuations. Compared with traditional rated capacity constraint methods, it improves regional capacity utilization by 15% to 25%, reduces the average annual lifespan loss of main transformers by more than 40%, and reduces the probability of voltage / frequency exceeding limits by more than 80%. The control method is applicable to main transformer clusters with different insulation aging states, and the system utilizes a safety redundancy coefficient. Capacity suppression is implemented for severely aged main transformers to ensure that while improving capacity utilization, the risk of equipment overload due to model bias or extreme operating conditions is avoided.

Images