Power grid dispatching adaptive cruise decision-making method based on large model

Abstract

The invention belongs to the technical field of intelligent power grid dispatching, and particularly relates to a power grid dispatching adaptive cruise decision-making method based on a large model. Performing power grid state prediction through an improved Transform large model; generating a self-adaptive scheduling instruction based on model predictive control MPC and a fuzzy self-adaptive PID algorithm; on-line iterative optimization of a large model is realized through a closed-loop feedback mechanism, and the state sensing and anti-disturbance capability of the power grid is enhanced through digital twin modeling and dynamic topology management; through laser radar three-dimensional modeling and graph database topological mapping, a power grid equipment geometric model, electrical parameters and a dynamic mirror image of a connection relation are constructed. The graph database updates the on-off state of the switch in real time, the power grid topology is automatically reconstructed, and prediction deviation caused by topology change of a traditional static model is avoided.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of smart grid dispatching, and particularly relates to a power grid dispatching adaptive cruise decision method based on a large model. BACKGROUND

[0002] With the promotion of the "double carbon" goal, the penetration rate of fluctuating power sources such as wind power and photovoltaic power continues to rise, and power grid dispatching faces challenges such as difficulty in processing multi-source heterogeneous data, large load and power prediction error, and real-time balance control lag. The traditional dispatching method based on rule engine or static model is difficult to adapt to the "second-level fluctuation" of new energy output and the "spatial and temporal coupling" characteristics of load, often leading to an increase in curtailment rate, an increase in equipment overload risk, or a decrease in power supply reliability.

[0003] The limitations of the prior art mainly include:

[0004] Insufficient data-driven capability: only relying on historical data statistical characteristics, lacking deep mining of the strong coupling relationship between weather, load and power;

[0005] Lack of model dynamics: fixed parameter prediction models cannot track real-time changes in power grid topology and equipment health status;

[0006] Weak decision robustness: open-loop control mode is difficult to cope with new energy power prediction error.

[0007] Therefore, the application provides a power grid dispatching adaptive cruise decision method based on a large model. SUMMARY

[0008] In order to make up for the deficiencies of the prior art and solve at least one technical problem proposed in the background art.

[0009] The technical scheme adopted by the application to solve its technical problems is: the power grid dispatching adaptive cruise decision method based on a large model, comprising the following steps:

[0010] Collecting multi-modal power grid data and constructing a digital twin database;

[0011] Performing power grid state prediction through an improved Transformer large model;

[0012] Generating adaptive dispatching instructions based on model predictive control (MPC) and fuzzy adaptive PID algorithm;

[0013] Realizing online iterative optimization of the large model through a closed-loop feedback mechanism.

[0014] Preferably, the multi-modal power grid data comprises:

[0015] Physical layer data: real-time measurement data of bus voltage, line flow and equipment temperature;

[0016] Environmental layer data: weather forecast, geological disaster warning data;

[0017] Behavior layer data: user load curve, electric vehicle charging cluster state, energy storage charging and discharging plan data;

[0018] The physical layer data, environmental layer data and behavior layer data are stored in a distributed time series database after being cleaned and normalized by an edge computing node.

[0019] Preferably, the improved Transformer large model is a TFT improved architecture, which realizes power grid state prediction in the following way:

[0020] The input is a multi-dimensional data sequence with a sliding time window length of 12-48 hours;

[0021] The attention mechanism is used to capture the spatio-temporal correlation features between data, and the output includes:

[0022] New energy output prediction value, node load prediction curve and key equipment health state warning in the next 15 minutes-1 hour.

[0023] Preferably, the improvement of the Transformer large model includes:

[0024] A multi-resolution time feature extraction module is added to the encoder layer to separate high-frequency fluctuations and low-frequency trends in the data;

[0025] An adversarial learning mechanism is introduced in the decoder layer to improve the prediction robustness under extreme weather conditions.

[0026] Preferably, the adaptive cruise decision mechanism includes:

[0027] Target setting module: dynamically generates multi-objective functions according to peak load, valley load and maintenance period data of power grid operation stage;

[0028] Dynamic adjustment module: based on the prediction results of the large model, a model predictive control (MPC) algorithm is used to generate a scheduling instruction sequence for the next 5-30 minutes, and the adjustment means includes generator power adjustment, energy storage charging and discharging optimization, and adjustable load scheduling;

[0029] Disturbance compensation module: real-time monitoring of frequency deviation Δf and tie-line power out-of-limit signal, triggering of fuzzy adaptive PID controller, adjustment parameters including proportional coefficient Kp, integral time Ti and derivative time Td, so that the frequency deviation is controlled within ±0.1 Hz.

[0030] The input of the fuzzy adaptive PID controller is preferably frequency deviation e and deviation change rate ec, and the proportional coefficient Kp, integral time Ti and differential time Td are dynamically adjusted through a fuzzy rule table, wherein the fuzzy rule table comprises 7*7=49 control rules, and the output is the active power adjustment amount of the frequency regulator.

[0031] Preferably, the closed-loop feedback mechanism comprises:

[0032] The execution layer collects actual operation data after the execution of the dispatching instruction through the SCADA system.

[0033] The evaluation module calculates the root mean square error RMSE of the predicted value and the actual value, and if the root mean square error RMSE of three consecutive periods is greater than 5%, the large model online incremental training is triggered, and the training data contains the latest extreme scenario sample.

[0034] Preferably, the generation of the dispatching instruction sequence needs to meet the power grid safety constraints, including node voltage constraints, line transmission power constraints and rotating reserve capacity constraints.

[0035] Preferably, it further comprises a transfer learning module, which realizes the rapid adaptation of power grid scenes of different voltage levels and different power supply structures through pre-trained model parameter transfer, and the model deployment period is less than or equal to 1 week.

[0036] Preferably, the power grid dispatching adaptive cruise decision system based on a large model comprises:

[0037] The multi-modal data acquisition terminal is used for accessing power grid real-time measurement, weather forecast and user behavior data.

[0038] The cloud large model server deploys an improved Transformer prediction model.

[0039] The adaptive cruise controller integrates an MPC algorithm and a fuzzy adaptive PID controller.

[0040] The closed-loop feedback execution unit comprises a SCADA system and a model online training module.

[0041] The beneficial effects of the present application are as follows:

[0042] 1. The power grid dispatching adaptive cruise decision method based on a large model disclosed in the present application, multi-modal data fusion and edge computing preprocessing, improve data quality and decision timeliness, through edge computing nodes, three types of data of physical layer, environment layer and behavior layer are cleaned, protocol conversion and feature extraction are carried out, and hybrid database architecture is used to realize efficient storage and query of data.

[0043] 2. The method for adaptive cruise decision of power grid dispatching based on large model, digital twin modeling and dynamic topology management, enhances the state perception and anti-disturbance ability of the power grid; through laser radar three-dimensional modeling and graph database topology mapping, a dynamic mirror of the geometric model, electrical parameters and connection relationship of the power grid equipment is constructed. The graph database updates the switch opening and closing state in real time, automatically reconfigures the power grid topology, and avoids the prediction deviation caused by the topology change of the traditional static model; the device health state label is bound with real-time data, the dispatching system can early warning of the overload risk of the main transformer, combined with the fuzzy adaptive PID controller, the frequency deviation is reduced, and the fault response time is shortened. BRIEF DESCRIPTION OF DRAWINGS

[0044] The application will be further described below with reference to the drawings.

[0045] Fig. 1 is a method flowchart of the method for adaptive cruise decision of power grid dispatching based on large model of the application;

[0046] Fig. 2 is a system block diagram of the adaptive cruise decision system of power grid dispatching of the application. DETAILED DESCRIPTION

[0047] In order to make the technical means, creative features, purposes and effects realized by the application easy to understand, the application will be further described below in combination with specific embodiments.

[0048] Embodiment one

[0049] As shown in Figs. 1-2 , the method for adaptive cruise decision of power grid dispatching based on large model, characterized in that it comprises the following steps:

[0050] Collecting multi-modal power grid data and constructing a digital twin database;

[0051] Making power grid state prediction through an improved Transformer large model;

[0052] Generating adaptive dispatching instructions based on model predictive control (MPC) and fuzzy adaptive PID algorithm;

[0053] Realizing online iterative optimization of the large model through a closed-loop feedback mechanism.

[0054] Specifically, the multi-modal power grid data comprises:

[0055] Physical layer data: real-time measurement data of bus voltage, line flow, and device temperature; data collection through intelligent electronic device (IED), synchronous phasor measurement unit (PMU), and temperature sensor;

[0056] Environmental layer data: weather forecast (wind speed, irradiance, temperature), geological disaster warning data; data collection through weather station API (such as China Weather Network), satellite remote sensing data, micro-meteorological sensors;

[0057] Behavioral layer data: user load curve, electric vehicle charging cluster state, energy storage charging and discharging plan data; data collection through smart meters (AMI), charging pile management system, energy storage BMS.

[0058] The physical layer data, environmental layer data and behavioral layer data are cleaned, normalized and stored in a distributed time series database through an edge computing node.

[0059] In one embodiment, the edge computing preprocessing is:

[0060] In a substation, a new energy station and the like, an edge computing node (such as Atlas 500) is deployed on site to achieve:

[0061] Data cleaning: eliminate jump values (such as abnormal points where the voltage exceeds the rated value ± 30%), and fill in missing values (based on historical data interpolation method);

[0062] Protocol conversion: unify different manufacturers' private protocols (such as converting ModbusRTU to MQTT);

[0063] Feature extraction: calculate active / reactive power fluctuation variance, load peak-valley difference and other derived features to reduce cloud computing pressure;

[0064] Local cache: temporarily store data with low real-time requirements (such as daily load curve) in local SSD and periodically upload in batches.

[0065] Specifically, the improved Transformer large model is a TFT improved architecture, which realizes power grid state prediction in the following ways:

[0066] The input is a multi-dimensional data sequence with a sliding time window length of 12-48 hours;

[0067] The attention mechanism is used to capture the spatio-temporal correlation features between data, and the output includes:

[0068] New energy output prediction value for the next 15 minutes to 1 hour (error rate ≤7%);

[0069] Node load prediction curve (considering temperature-price-holiday influence factors);

[0070] Key equipment health state warning (such as main transformer overload probability, line icing risk).

[0071] Specifically, the improvement of the Transformer large model includes:

[0072] A multi-resolution temporal feature extraction module is added at the encoder layer to separate high-frequency fluctuations (e.g., second-level new energy output changes) from low-frequency trends (e.g., daily load cycles) in the data.

[0073] An adversarial learning mechanism is introduced at the decoder layer to improve the prediction robustness under extreme weather scenarios.

[0074] Specifically, the adaptive cruise decision mechanism includes:

[0075] The target setting module dynamically generates a multi-objective function based on the peak load, valley load, and maintenance period data of the power grid operation stage, such as the "maximum renewable energy consumption" as the target during the high new energy penetration period, allowing a voltage deviation of ±5%;

[0076] The dynamic adjustment module generates a scheduling instruction sequence for the next 5-30 minutes based on the large model prediction results using the model predictive control (MPC) algorithm, and the adjustment methods include generator power adjustment, energy storage charging and discharging optimization, and adjustable load scheduling.

[0077] The disturbance compensation module monitors the frequency deviation Δf and the tie-line power limit signal in real time, triggers the fuzzy adaptive PID controller, and adjusts the parameters including the proportional coefficient Kp, the integral time Ti, and the derivative time Td to control the frequency deviation within ±0.1 Hz.

[0078] Specifically, the input of the fuzzy adaptive PID controller is the frequency deviation e and the deviation change rate ec, and the proportional coefficient Kp, the integral time Ti, and the derivative time Td are dynamically adjusted through the fuzzy rule table, which contains 7x7=49 control rules, and the output is the active power adjustment of the frequency regulation unit.

[0079] Specifically, the closed-loop feedback mechanism includes:

[0080] The execution layer collects the actual operation data (such as voltage and curtailment rate) after the execution of the scheduling instruction through the SCADA system.

[0081] The evaluation module calculates the root mean square error RMSE between the predicted value and the actual value, and if the root mean square error RMSE of the last 3 cycles is greater than 5%, the large model online incremental training is triggered, and the training data includes the latest extreme scenario samples (such as lightning-induced photovoltaic sudden drop).

[0082] Specifically, the generation of the scheduling instruction sequence needs to meet the power grid safety constraints, including:

[0083] Node voltage constraint (rated voltage ±5%);

[0084] Line transmission power constraint (not exceeding 90% of the thermal stability limit);

[0085] Spinning reserve capacity constraint (≥15% of maximum load during peak power supply periods).

[0086] The invention also includes a transfer learning module, which enables rapid adaptation to power grid scenarios with different voltage levels (10kV-500kV) and different power supply structures (wind and solar power accounting for 10%-60%) through the transfer of pre-trained model parameters, with a model deployment cycle of ≤1 week.

[0087] Example 2

[0088] like Fig. 2 As shown in Example 1, another embodiment of the present invention is as follows:

[0089] An adaptive cruise decision-making system for power grid dispatching, which implements a large-model-based adaptive cruise decision-making method, includes:

[0090] Multimodal data acquisition terminal, used to access real-time power grid measurements, weather forecasts, and user behavior data;

[0091] Deploy improved Transformer prediction models on cloud-based large model servers;

[0092] Adaptive cruise controller, integrating MPC algorithm and fuzzy adaptive PID controller;

[0093] The closed-loop feedback execution unit includes the SCADA system and the online model training module.

[0094] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A power grid dispatch adaptive cruise decision-making method based on a large model, characterized by: Includes the following steps: Collect multimodal power grid data and construct a digital twin database; Power grid state prediction using an improved Transformer large model; Generate adaptive scheduling instructions based on Model Predictive Control (MPC) and Fuzzy Adaptive PID algorithm; A closed-loop feedback mechanism enables online iterative optimization of large models.

2. The adaptive cruise decision-making method for power grid dispatch based on a large model according to claim 1, characterized in that, The multimodal power grid data includes: Physical layer data: Real-time measurement data of bus voltage, line power flow, and equipment temperature; Environmental layer data: weather forecasts and geological disaster early warning data; Behavioral layer data: user load curves, electric vehicle charging cluster status, and energy storage charging and discharging plan data; Physical layer data, environmental layer data, and behavioral layer data are cleaned and normalized through edge computing nodes before being stored in a distributed time-series database.

3. The adaptive cruise decision-making method for power grid dispatch based on a large model according to claim 2, characterized in that, The improved Transformer large model is a TFT improved architecture, which achieves power grid state prediction in the following ways: The input is a multi-dimensional data sequence with a sliding time window, the window length of which is 12-48 hours. By using an attention mechanism to capture spatiotemporal correlation features between data, the output includes: Forecast values ​​of new energy output, node load forecast curves, and early warning of the health status of key equipment for the next 15 minutes to 1 hour.

4. The adaptive cruise decision-making method for power grid dispatch based on a large model according to claim 3, characterized in that, The improvements to the Transformer large model include: A multi-resolution temporal feature extraction module is added to the encoder layer to separate high-frequency fluctuations from low-frequency trends in the data; An adversarial learning mechanism is introduced at the decoder layer to improve the prediction robustness in extreme weather scenarios.

5. The adaptive cruise decision-making method for power grid dispatch based on a large model according to claim 1, characterized in that, The adaptive cruise decision-making mechanism includes: Target setting module: Dynamically generates multi-objective functions based on peak load, valley load, and maintenance period data during power grid operation. Dynamic adjustment module: Based on the prediction results of the large model, the model predictive control MPC algorithm generates a sequence of scheduling instructions for the next 5-30 minutes. The adjustment methods include generator power adjustment, energy storage charging and discharging optimization, and adjustable load scheduling. Disturbance compensation module: Real-time monitoring of frequency deviation Δf and tie line power over-limit signal, triggering fuzzy adaptive PID controller, adjusting parameters including proportional coefficient Kp, integral time Ti, and derivative time Td, to keep the frequency deviation within ±0.1Hz.

6. The adaptive cruise decision-making method for power grid dispatch based on a large model according to claim 5, characterized in that, The inputs of the fuzzy adaptive PID controller are frequency deviation e and deviation change rate ec. The proportional coefficient Kp, integral time Ti, and derivative time Td are dynamically adjusted through a fuzzy rule table, which contains 7×7=49 control rules. The output is the active power regulation of the frequency modulation unit.

7. The adaptive cruise decision-making method for power grid dispatch based on a large model according to claim 1, characterized in that, The closed-loop feedback mechanism includes: The execution layer collects actual operational data after the execution of scheduling instructions through the SCADA system; The evaluation module calculates the root mean square error (RMSE) between the predicted and actual values. If the RMSE exceeds 5% for three consecutive periods, online incremental training of the large model is triggered, and the training data includes the latest extreme scenario samples.

8. The adaptive cruise decision-making method for power grid dispatch based on a large model according to claim 1, characterized in that, The generation of the scheduling instruction sequence must meet power grid security constraints, including: node voltage constraints, line transmission power constraints, and spinning reserve capacity constraints.

9. The adaptive cruise decision-making method for power grid dispatch based on a large model according to claim 1, characterized in that, It also includes a transfer learning module, which enables rapid adaptation to power grid scenarios with different voltage levels and power supply structures through the transfer of pre-trained model parameters, with a model deployment cycle of ≤1 week.

10. A large-model-based adaptive cruise decision-making system for power grid dispatch, used to implement the large-model-based adaptive cruise decision-making method for power grid dispatch as described in any one of claims 1-9, characterized in that, include: Multimodal data acquisition terminal, used to access real-time power grid measurements, weather forecasts, and user behavior data; Deploy improved Transformer prediction models on cloud-based large model servers; Adaptive cruise controller, integrating MPC algorithm and fuzzy adaptive PID controller; The closed-loop feedback execution unit includes the SCADA system and the online model training module.

Images