Regional fan virtual inertia support scheduling method considering wind power plant group synergistic effect

Abstract

The invention discloses a regional fan virtual inertia support scheduling method considering a wind power plant group synergistic effect, and aims to solve the problems of non-uniform inertia distribution, response conflict and economy in wind power plant group scheduling. By introducing a hybrid multi-level cooperative control architecture, the inertia vacancy of the power grid is calculated in real time, the inertia demand of each wind power plant is allocated, and the scheduling strategy of the wind power plant group is optimized. And a time-varying inertia demand curve is generated by adopting a hybrid algorithm, the future inertia demand is accurately predicted, and the fan rotating speed recovery cost, the equipment loss and the wind curtailment amount are minimized in combination with a multi-target optimization model. A model predictive control framework and a Nash equilibrium theory are introduced, the game problem between wind power plants is solved, collaborative scheduling of a wind power plant group is optimized, and timeliness and stability of a scheduling instruction are ensured by embedding a time delay compensation algorithm. According to the method, the inertia supporting efficiency of the power grid can be remarkably improved, the equipment loss and the wind curtailment amount are reduced, the robustness of the system is improved, and the method has high application value and prospect.

Description

Technical Field

[0001] The present invention relates to the technical field of power system dispatching, and in particular to a regional fan virtual inertia support dispatching method considering the collaborative effect of a wind farm group. Background Technique

[0002] To address this challenge, the prior art has proposed a virtual inertia control technology that uses the converter control of wind turbines to simulate the inertia characteristics of synchronous generators. By introducing an analog inertia link into the control system of wind turbines, the wind turbines can release their rotor kinetic energy when the grid frequency is disturbed, providing a certain degree of frequency support. However, the existing wind power virtual inertia control technology mainly has the following problems: The inertia support capacity that a single wind turbine can provide is limited by its own rotor kinetic energy reserve and converter capacity. Excessive release of rotor kinetic energy may cause the rotational speed of the wind turbine itself to drop too fast, exceeding the safe operating range, and even triggering a stall protection action, which may instead have an adverse impact on grid stability. In addition, the response range and effect of a single unit are often limited for the overall inertia improvement of the regional power grid. When multiple wind turbines adopt independent virtual inertia control strategies, due to the differences in the frequency changes and their own operating states sensed by each unit, there may be "overload" or "under-compensation" phenomena in the inertia support behavior. For example, under the same frequency disturbance, some units may release excessive energy, resulting in too low a rotational speed of their own, while some other units may respond insufficiently and cannot effectively suppress the frequency deviation. More seriously, the lack of coordinated inertia injection among multiple wind turbines may be coupled with the inherent frequency of the power grid, inducing subsynchronous oscillation and posing a threat to the safe and stable operation of the power system. In an actual power grid, regional wind farm groups are usually distributed in different geographical locations and are affected by different wind resource conditions and power grid topologies. This results in significant spatio-temporal distribution differences in the available inertia potential (such as the rotor kinetic energy that can be released) of each wind farm and the demand for power grid frequency support. Existing centralized dispatching methods often have difficulty accurately evaluating and allocating the inertia support tasks of each wind farm, easily causing resource waste or insufficient support.

[0003] In view of the above analysis of the technical status quo, the key technical pain points of current wind power participating in power grid inertia support are mainly reflected in the following aspects:

[0004] Existing wind power virtual inertia control and scheduling methods mostly focus on optimizing individual turbines or wind farms, lacking effective consideration of the dynamic coupling characteristics and complementary inertia contributions between wind farms. Establishing a coordinated control mechanism across wind farms to fully tap their collective inertia support potential is a key issue that needs to be addressed. Existing extensive virtual inertia control strategies can cause some wind turbines to frequently participate in inertia response, accelerating wear and tear on mechanical components and electrical equipment, shortening equipment life, and potentially affecting wind turbine efficiency and reducing the overall power generation revenue of the wind farm. Providing sufficient inertia support while also ensuring the economic operation of wind turbines is a major challenge. Traditional centralized optimization and scheduling methods typically require real-time access to detailed information on all wind farms and the power grid across the entire region, placing extremely high demands on the reliability and real-time performance of the communication infrastructure. In scenarios with large-scale distributed wind power integration, high communication costs and potential communication failures limit the application of centralized scheduling. Designing a distributed collaborative scheduling method that can adapt to communication-constrained scenarios is of great practical significance.

[0005] After searching, the invention patent with patent publication number CN111105090B discloses a distributed energy system optimization scheduling method and device based on intelligent algorithm. It is aimed at distributed energy system scheduling, utilizes historical natural gas consumption, power generation and waste heat data, trains BP neural network model, establishes gas internal combustion unit replacement model and constructs profit objective function, and then realizes the optimization scheduling of system equipment operating parameters, focusing on the balance between energy distribution and overall economic benefits.

[0006] The technical comparison between the above-mentioned reference documents and this application is as follows:

[0007] This application addresses the uneven inertia distribution, response conflicts and economic issues in the dispatch of wind farm groups, and proposes a regional wind turbine virtual inertia support dispatching method that takes into account the synergistic effect of wind farm groups. A hybrid multi-level collaborative control architecture is used to calculate the grid inertia shortage in real time and reasonably allocate the inertia demand of each wind farm; a time-varying inertia demand curve is generated based on a hybrid algorithm, and a multi-objective optimization model is combined to minimize the wind turbine speed recovery cost, equipment loss and wind abandonment; at the same time, model predictive control and Nash equilibrium theory are introduced to solve the game problem between wind farms, ensuring the timeliness of dispatching instructions and system stability. While the above-mentioned comparative documents focus on energy distribution and economic dispatch, this application focuses more on solving the synergistic effect within the wind farm group and the real-time and refined control problems of virtual inertia support. Summary of the Invention

[0008] Objective of the Invention: The objective of the present invention is to provide a regional wind turbine virtual inertia support scheduling method considering the collaborative effect of a wind farm cluster. Through spatio-temporal collaborative optimization among wind farm clusters, dynamically evaluate and allocate virtual inertia support tasks, and make full use of the overall inertia potential of wind power resources within the region, so as to quickly and effectively provide sufficient inertia support when the power grid experiences frequency disturbances, thereby maximizing the equivalent inertia level of the regional power grid, significantly enhancing the frequency stability of the regional power grid; reducing the response frequency of the wind turbines and equipment wear; suppressing the risk of sub-synchronous oscillation caused by the concurrent response of multiple wind turbines; and enhancing the robustness of the system under limited communication conditions.

[0009] To achieve the above objective, the solution of the present invention includes the following steps:

[0010] S1. Dynamic inertia potential assessment of the wind farm cluster

[0011] Based on real-time wind speed, unit speed, and power grid frequency deviation, construct an inertia contribution capacity model at the wind farm level, quantify the maximum kinetic energy that each wind turbine can release and the rotational speed safety boundary, and evaluate the virtual inertia potential of each wind farm.

[0012] S2. Design of the regional collaborative scheduling architecture

[0013] Adopt a "centralized - decentralized" hybrid multi-level collaborative control architecture, which specifically includes:

[0014] Regional scheduling layer: Based on the data of the wide-area measurement system, calculate the power grid inertia deficit in real time, and calculate the inertia demand distribution coefficient of each wind farm according to the power grid frequency fluctuation situation.

[0015] Station coordination layer: Implement distributed decision-making within the wind farm, dynamically evaluate the adjustable inertia margin of each station, and optimize the action priority and output curve of the wind turbines in combination with local information.

[0016] Unit execution layer: According to the collaborative scheduling instructions of the wind farm cluster, adaptively adjust the virtual inertia control (VSG) parameters.

[0017] S3. Dynamic inertia demand prediction model

[0018] Adopt a hybrid algorithm of LSTM and ARIMA, fuse multi-source information such as wind speed prediction, load fluctuation, and unit status, generate a time-varying inertia demand curve, and predict the inertia demand in the future time period.

[0019] S4. Construction of a multi-objective optimization model

[0020] Through a multi-objective optimization method, considering constraint conditions such as power grid frequency deviation limit, unit speed safety threshold, and converter capacity limit, construct a comprehensive optimization model including objectives such as inertia support, wind turbine speed recovery cost, and equipment loss minimization.

[0021] S5, Dynamic Game and Rolling Correction

[0022] Introduce the model predictive control framework, update the wind farm cluster response instructions in real time, embed the communication delay compensation algorithm, and improve the timeliness and stability of instruction execution. At the same time, combine the Nash equilibrium theory to optimize the response trade-off and competitive game among all participants (wind farms) to achieve a collaborative equilibrium under steady state.

[0023] Furthermore, step S1 is specifically as follows:

[0024] Comprehensively utilize the real-time wind speed monitoring data, the real-time rotational speed information of the units, and the feedback of the grid frequency deviation to construct a set of accurate wind farm-level inertia contribution capacity models. This model can deeply analyze the maximum kinetic energy that each unit can release under the current operating state, and at the same time clarify its rotational speed safety boundary, providing a solid data basis for subsequent inertia allocation and scheduling decisions. Through the dynamic analysis of a large amount of real-time data, ensure the accuracy and timeliness of the inertia potential assessment, and accurately grasp the inertia supply capacity of the wind farm cluster.

[0025] Based on the information such as the real-time collected wind speed prediction data, the operating state of the wind turbines, and the grid frequency deviation, construct a wind turbine-level dynamic inertia capacity model to evaluate the kinetic energy that each wind turbine can release currently and its rotational speed safety boundary. The core goal of this process is to generate an "inertia resource map" to accurately allocate the virtual inertia support tasks of each wind turbine in the area.

[0026] The dynamic inertia capacity H of each wind turbine i (t) can be represented by its rotational speed and rotor kinetic energy. For wind turbines, the release of virtual inertia is closely related to the rotor kinetic energy E i (t). The calculation formula for the rotor kinetic energy is:

[0027]

[0028] Where:

[0029] I i [[ID=3l]]is the moment of inertia of the rotor of the wind turbine;

[0030] ω i (t) is the rotational speed of the i-th wind turbine.

[0031] According to the current rotational speed ω i (t), the range of kinetic energy that can be released can be calculated:

[0032]

[0033] Where

[0034] ω i,minis the minimum safe speed of the wind turbine. Speeds lower than this may cause the unit to become unstable or unable to operate properly;

[0035] ω i,max is the maximum safe speed of the wind turbine. Speeds exceeding this may cause damage to the unit or other failures.

[0036] Therefore, the range of available kinetic energy for the wind turbine is:

[0037] ΔE i (t) = E i,max (t) - E i,min (t).

[0038] The rotational speed and kinetic energy release capacity of the wind turbine are not only related to its operating state but also affected by external environmental factors, especially fluctuations in wind speed and grid frequency. Therefore, it is necessary to further correct the inertia assessment of the unit by combining wind speed prediction and grid frequency deviation.

[0039] Based on ARIMA, a wind speed prediction model is established to predict the wind speed v i (t) at several future moments, thereby predicting the output power and rotational speed changes of the future wind turbine.

[0040] The fluctuation of the grid frequency is closely related to the inertia support task. When the grid frequency deviates from the rated frequency, the wind turbine needs to release or absorb a certain amount of kinetic energy to maintain frequency stability. The grid frequency deviation is directly related to the kinetic energy release capacity of the wind turbine.

[0041] The virtual inertia response ability of the wind turbine under different conditions is:

[0042] H i (t) = ΔE i (t)·(1 + β1·Δf(t) + β2·v i (t))

[0043] where

[0044] β1 and β2 are weight coefficients that reflect the influence of frequency deviation and wind speed on the contribution of virtual inertia.

[0045] Define the total inertia contribution H total (t) of the wind farm group as the sum of the inertia contributions of all units:

[0046]

[0047] where

[0048] N is the total number of units in the wind farm group.

[0049] Through a dynamic inertia capacity model based on information such as wind speed prediction, generator set speed, and grid frequency deviation, a wind farm cluster can accurately evaluate the available kinetic energy and rotational speed safety margin of each wind turbine, generating an inertia "resource map". This process lays the foundation for subsequent coordinated dispatching of regional wind farm clusters and virtual inertia support task allocation, ensuring that each wind turbine can provide reasonable support according to its own inertia capacity, thereby enhancing the stability and efficiency of the overall system.

[0050] To quantify the inertia potential of a wind farm cluster in real time, it can be achieved through the following steps:

[0051] 1) Obtain real-time wind speed data: Acquire real-time wind speed through wind speed sensors or meteorological data, and calculate the output power and rotational speed of the generator set based on the relationship between wind speed and generator set speed.

[0052] 2) Calculate real-time rotational speed: Calculate the rotational speed of each generator set using the wind turbine power curve based on wind speed and power output.

[0053] 3) Estimate rotor kinetic energy: Calculate the rotor kinetic energy of each generator set based on rotational speed.

[0054] 4) Grid frequency deviation monitoring: Monitor the grid frequency deviation in real time, and adjust the virtual inertia support capacity of wind turbines according to the change of grid frequency.

[0055] 5) Evaluate inertia potential: Based on the above data, evaluate the inertia contribution capacity of the wind farm cluster in real time, determine the virtual inertia that each wind turbine can provide, and then calculate the total inertia contribution of the wind farm cluster.

[0056] 1] Further, step S2 innovatively adopts a "centralized - decentralized" hybrid multi - level collaborative control architecture mode.

[0057] Regional dispatching layer: Relying on the massive real - time data collected by the wide - area measurement system, quickly and accurately calculate the current inertia deficit of the power grid. Subsequently, based on a set of scientific and reasonable algorithms, calculate the inertia demand distribution coefficient of each wind farm. This coefficient fully considers various factors such as the geographical location, wind speed conditions, and generator set performance of each wind farm, ensuring that the inertia demand can be accurately allocated to each wind farm, achieving macroscopic overall planning at the regional level.

[0058] Station coordination layer: Inside the wind farm, implement a distributed decision - making mechanism. Based on the locally collected generator set operation information, dynamically evaluate the adjustable inertia margin of each station, that is, the remaining inertia supply capacity. At the same time, combined with local wind speed prediction, generator set status and other information, optimize the action priority sorting of the generator sets and the planning of the output power curve. Through local autonomous decision - making, achieve efficient utilization of internal resources of the wind farm, and improve the response speed and flexibility.

[0059] Unit execution layer: According to the instructions issued by the upper layer and the local real-time operating status, adaptively adjust the control parameters of the virtual synchronous generator. Ensure that the wind turbine can accurately and stably output the required virtual inertia to effectively support the grid frequency.

[0060] In the present invention, in order to effectively address the dynamic cooperative scheduling problem of wind farm clusters, a hierarchical control architecture is designed. This architecture consists of three layers: the regional scheduling layer, the station coordination layer, and the unit execution layer. Each layer plays a different role in the inertia distribution and scheduling process of the regional wind farm cluster and works together to ensure the accuracy and efficiency of virtual inertia support.

[0061] 1. Regional scheduling layer (regional layer)

[0062] The core task of the regional scheduling layer is to based on the data of the wide-area measurement system, evaluate the inertia gap of the regional power grid in real time, and by introducing factors such as wind speed distribution and geographical information, formulate reasonable inertia distribution weights for different wind farms, so as to achieve the optimal scheduling of inertia within the region. The steps are as follows

[0063] S211: According to the real-time grid frequency data and the grid status information provided by the WAMS, calculate the current inertia gap of the grid. The inertia gap of the grid is expressed as the difference between the required virtual inertia and the existing grid inertia, and the formula is as follows:

[0064] ΔH gap =H required -H available

[0065] Where:

[0066] H required is the total inertia currently required by the grid, usually determined according to the grid frequency deviation Δf and the time requirement of the frequency regulation response;

[0067] H available is the currently available inertia, including the inertia of traditional generating units and the virtual inertia provided by wind turbines.

[0068] S212: Combine the geographical information system (GIS) and real-time wind speed data to calculate the inertia support capacity of each wind farm, and dynamically allocate inertia tasks according to the inertia gap of the regional power grid. The inertia distribution weight of the wind farm can be expressed by the following formula:

[0069]

[0070] Where:

[0071] w i is the inertia distribution weight of the i-th wind farm; α iis a factor related to the distance from the i-th wind farm, and is inversely proportional to the distance from the wind farm to the center of the regional power grid; V i is the wind speed of the i-th wind farm; N is the total number of wind farms in the region.

[0072] Through the above allocation strategy, the regional dispatch layer can calculate the required virtual inertia output according to the inertia capacity of each wind farm, and instruct each wind farm to provide the virtual inertia support at that moment.

[0073] 2. Substation Coordination Layer (substation layer)

[0074] The substation coordination layer is responsible for calculating the inertia allocation priority of each unit inside the substation according to the inertia allocation weight provided by the regional dispatch layer, and combining local information such as the wind speed and unit health status inside the substation, and optimizing the virtual inertia response strategy of the unit. The goal of this layer is to ensure the coordinated response of each unit inside the substation, so as to optimally provide virtual inertia support while minimizing equipment losses.

[0075] S221. Based on the actual operating conditions of each wind farm and the inertia demand allocation of the regional dispatch layer, the substation coordination layer calculates the inertia allocation priority for each unit inside the substation. The specific priority scheduling formula is:

[0076]

[0077] Where:

[0078] P priority,i is the inertia allocation priority of the i-th fan; R health,i is the health coefficient of the i-th fan;

[0079] M is the number of units inside the substation.

[0080] S222. The health coefficient (R health,i ) of the fan is an index comprehensively evaluating the health status of the fan, considering multiple factors such as the operating state of the fan, failure frequency, maintenance history, wear condition of fan components, etc. The value range of the health coefficient R health,i is [0, 1], where 1 means the fan is completely healthy and 0 means the fan is completely ineffective. If the health coefficient is low (close to 0), it means the state of the fan is poor, which may affect the effect of virtual inertia support, and a lower inertia contribution weight should be given in the dispatch. A higher health coefficient (close to 1) means the fan is in good operating condition and can undertake a higher virtual inertia support task.

[0081] The following is the health coefficient calculation formula:

[0082] R health,i = w1·R operating,i + w2·R maintain,i+ w3·R wear,i + w4·R failure,i

[0083] Wherein:

[0084] R operating,i is the evaluation of the operating state of the wind turbine, usually represented by the ratio of the actual output power of the wind turbine to the expected output power, reflecting the current operating efficiency of the wind turbine.

[0085]

[0086] Wherein, P actual,i is the actual output power of the wind turbine, and P rated,i is the rated power of the wind turbine.

[0087] R maintain,i is the maintenance status of the wind turbine, indicating the maintenance history and frequency of the wind turbine. It can be evaluated according to the maintenance interval and maintenance quality of the wind turbine.

[0088] [[ID=�1]]

[0089] [[ID=3䲨]]Wherein, T last maintain,i is the duration from the last maintenance of the wind turbine to the current time, and T scheduled maintain,i is the scheduled maintenance period of the wind turbine.

[0090] R wear,i is the wear condition of the wind turbine components, usually evaluated through the component health monitoring data of the wind turbine (including at least the gearbox, blades, generator, etc.). According to the degree of component wear, the impact of this part on the overall health coefficient is calculated.

[0091] D

[0092] Wherein, W current,i is the current wear level of the wind turbine component, and W max,i is the maximum wear resistance of this component.

[0093] R failure,i is the failure rate of the wind turbine, indicating the number of failures of the wind turbine in the past period of time. Frequent failures will significantly reduce the health coefficient of the wind turbine.

[0094]

[0095] Wherein, F failures,i is the number of failures of the wind turbine, and F max,i is the maximum acceptable number of failures of the wind turbine.

[0096] w1, w2, w3, w4 are the weights of each index, which are adjusted according to the actual situation to ensure that the influence ratio of each factor on the health coefficient conforms to the actual application scenario.

[0097] Through this health coefficient, the station coordination layer can dynamically adjust the virtual inertia contribution of the units. Wind turbines with a lower health coefficient will be preferentially restricted from outputting virtual inertia in the dispatch to avoid overload caused by wind turbine failures or performance degradation, ensuring the overall dispatch efficiency of the wind farm group and the stability of the power grid.

[0098] Furthermore, in the collaborative dispatch of the wind farm group in step S3, accurate prediction of inertia demand is crucial, which can provide a reliable basis for subsequent dispatch decisions. To improve the prediction accuracy and reduce the response delay to future power grid frequency fluctuations, this paper adopts an improved hybrid LSTM-ARIMA model, which can combine multiple information sources and is used to dynamically predict the inertia demand of the power grid system in different time periods. This model can effectively integrate multi-source information such as wind speed prediction, load fluctuations, and the status of wind turbines, can effectively capture short-term and long-term trends in inertia demand changes, generate a time-varying inertia demand curve, and thus accurately predict the inertia demand of the power grid in the future time period, providing an advance dispatch deployment for the regional dispatch layer.

[0099] By introducing the LSTM-ARIMA hybrid algorithm, multi-source heterogeneous information such as wind speed prediction data, load fluctuation information, and the real-time and historical status of the units is deeply integrated. Through learning and mining of massive historical data and dynamic tracking of real-time data, a precise time-varying inertia demand curve is generated. This curve can prospectively reflect the trend of inertia demand changes in the power grid in the future for some time, providing a strong basis for formulating dispatch strategies in advance, and ensuring that the inertia support of the wind farm group can match the power grid demand in a timely and accurate manner.

[0100] The dynamic inertia demand changes with factors such as power grid load fluctuations, power grid frequency fluctuations, and the operating status of wind turbines. To accurately predict the inertia demand of the power grid, the present invention considers the following aspects: Power grid frequency deviation: The power grid frequency deviation is one of the main driving factors for inertia demand. Changes in the power grid load will cause fluctuations in frequency, and wind turbines need to adjust their output power according to the power grid frequency deviation to provide inertia support. Wind speed change: The wind speed directly affects the output power of the wind turbine, and thus affects the rotor kinetic energy of the wind turbine. The accuracy of wind speed prediction is crucial for predicting inertia demand. Power grid load fluctuation: Changes in the power grid load will cause fluctuations in the power grid frequency, which will affect the inertia support task of the wind turbine. Unit status: The operating status of the unit, such as start-stop status, fault status, etc., will also affect the virtual inertia it provides.

[0101] The steps of the model framework are as follows:

[0102] S31. Data preprocessing and feature extraction:

[0103] Collect real-time data such as wind speed, load, grid frequency, and unit status, and perform standardization processing to ensure data stability and uniformity. Extract relevant features, including wind speed change rate, load fluctuation amplitude, grid frequency deviation, etc.

[0104] S32. LSTM model building:

[0105] Use the LSTM neural network to process data on historical inertia demand and related factors. LSTM can effectively learn the impact of factors such as wind speed fluctuations and load changes on inertia demand. The input sequence is the time-series data of historical inertia demand and external factors such as wind speed, load, and frequency, and the output is the prediction of inertia demand for a future period.

[0106] S33. ARIMA model building:

[0107] Based on the known grid frequency fluctuation data, use the ARIMA model to model the grid frequency change. The ARIMA model is suitable for stationary time-series data and can capture the short-term change rules of frequency fluctuations. The input is the historical data of grid frequency, and the output is the predicted value of frequency at future time points.

[0108] S34. LSTM-ARIMA fusion model:

[0109] Fuse the prediction results of LSTM and ARIMA, and comprehensively consider multiple factors such as wind speed fluctuations, load changes, and grid frequency deviation. The fusion strategy usually adopts methods such as weighted average or recursive calculation. The finally output inertia demand curve can accurately predict the future inertia demand change and provide a reference for the advance deployment of the dispatching system.

[0110] The LSTM network can effectively capture the long-term and short-term dependencies in the time series. In the present invention, the LSTM model is mainly used to process information such as wind speed prediction data, short-term fluctuations of grid frequency, and unit status changes. The basic structure of LSTM includes an input gate, a forget gate, and an output gate, which can flexibly adjust the model's memory ability for historical information.

[0111] The basic formula of LSTM used in the present invention is as follows:

[0112] f t =σ(W f ·[h t-1 ,x t +b f )

[0113] i t =σ(W i ·[h t-1 ,x t +b i )

[0114] o t = σ(W o ·[h t-1 , x t + b o )

[0115] c t = f t ·c t-1 + i t ·tanh(W c ·[h t-1 , x t + b c )

[0116] h t = o t ·tanh(c t )

[0117] Among them, f t is the forget gate, i t is the input gate, o t is the output gate, c t is the cell state, h t is the output state, σ is the activation function, x t is the input data, W f , W, W o , W c are weight matrices, b f , b i , b o , b c are bias terms.

[0118] The LSTM model predicts the inertia demand by capturing long-term dependencies in the time series. However, under multi-level non-linear dynamic characteristics, the LSTM may still have certain limitations. Therefore, the improved LSTM model can incorporate more input features, such as the change trend of wind speed, the change pattern of load, etc., and locally process the features through the convolutional layer to improve the perception ability of short-term changes.

[0119] h t = tanh(W h ·[h t-1 , x t , CNN(x t )] + b h )

[0120] y t = W y ·h t + b y

[0121] Wherein:

[0122] The CNN(x t ) is the feature output after convolutional processing of the input data, further enhancing the response to short-term fluctuations;

[0123] y t is the output of the LSTM, i.e., the predicted value of the inertia demand.

[0124] For the ARIMA model, an optimization strategy for the differencing order is introduced, and seasonal factors are added to improve the model's ability to model periodic fluctuations. The enhanced ARIMA model formula is as follows:

[0125]

[0126] where: φ i is the autoregressive coefficient; θ j is the moving average coefficient; ε t is the error term; S t is the seasonal adjustment term, representing the impact of seasonal fluctuations on the frequency.

[0127] For the fused prediction results of the LSTM and ARIMA, an adaptive weighting mechanism is designed, which dynamically adjusts their weighting coefficients according to the real-time prediction errors of the two models. The specific weighting formula is as follows:

[0128]

[0129] where:

[0130] is the final predicted value of the inertia demand; y LSIM,t is the output of the LSTM model; y ARIMA,t is the output of the ARIMA model; α t is the adaptive weighting coefficient, dynamically adjusted according to the prediction error at each moment:

[0131]

[0132] where Err LSTM,t and Err ARIMA,t represent the prediction errors of the LSTM and ARIMA models respectively. Through this weighting coefficient, the fusion ratio can be dynamically adjusted according to the prediction ability of the model at a certain moment.

[0133] Furthermore, in the construction of the multi-objective optimization model in step S4, building an effective multi-objective optimization model is the key to ensuring the coordinated and efficient response of the wind turbine. This model needs to comprehensively consider the trade-offs between different objectives to achieve the best regional inertia support, while ensuring the stability of the grid frequency, reducing equipment wear, and improving system economy.

[0134] The core goal of the multi-objective optimization model is to achieve the following goals by rationally dispatching and controlling wind turbines:

[0135] Maximizing inertia response efficiency: This ensures that the wind farm's virtual inertia output can effectively compensate for grid frequency fluctuations, thereby improving the system's ability to respond to frequency changes. Achieving this goal contributes to stable grid operation, especially when large-scale wind farms are connected.

[0136] Minimizing speed recovery costs and equipment wear: When performing inertia support, wind turbines must release rotor kinetic energy to stabilize grid frequency. However, excessive kinetic energy release can cause turbine speeds to deviate from the optimal operating range, increasing wear and tear on the turbine. Therefore, optimizing inertia release methods is crucial to reduce losses during speed recovery and extend equipment life.

[0137] Minimizing frequency offsets caused by support fluctuations: Although virtual inertia can effectively suppress grid frequency fluctuations, the fluctuations of inertia injection may cause short-term overshoot or fluctuations in grid frequency. Therefore, the optimization algorithm needs to balance the smoothness of inertia output and avoid large frequency fluctuations.

[0138] Inertia response efficiency refers to the efficiency of releasing virtual inertia of a wind farm group during changes in grid frequency. In order to improve efficiency, it is necessary to ensure that each wind turbine responds as quickly as possible when frequency fluctuations occur, while avoiding excessive release of kinetic energy that leads to speed instability. The speed recovery cost of a wind turbine includes energy loss and mechanical wear caused by frequent inertia release and recovery. In order to minimize this goal, the optimization algorithm should try to avoid frequent inertia release or excessive kinetic energy release. There is a certain correlation between the speed recovery cost and the speed deviation of the wind turbine. Although the wind turbine supports the grid frequency through virtual inertia, it may cause short-term frequency deviations during the wind turbine control response process. In order to avoid unstable grid operation caused by frequency deviations, the optimization goal should minimize the amplitude of frequency fluctuations.

[0139] Furthermore, in step S5, because the coordinated scheduling of multiple wind farms requires consideration of their mutual influence, traditional centralized optimization methods are often unable to adapt to the complex dynamic interactions and time lags between multiple wind farms. To address these challenges, this paper proposes a dynamic scheduling strategy based on a model predictive control framework, while also incorporating a time lag compensation algorithm and Nash equilibrium theory to address multi-agent conflicts and scheduling lags between wind farms.

[0140] In each scheduling cycle, with the current state of the wind farm cluster as the initial condition, a rolling optimization problem is constructed to predict the inertia demand and system disturbances in multiple future time periods, and an optimal allocation strategy for the inertia response of wind turbines is obtained. Let the prediction time domain be N, and the control variable be the virtual inertia response power of wind farm i as P vi (k), then the optimization objective function can be expressed as:

[0141]

[0142] Where:

[0143] Δf(k) is the system frequency deviation at the k-th step;

[0144] Δω i (k) is the rotational speed deviation of wind turbine i at the k-th step;

[0145] C deg is the degradation cost function caused by the response of wind turbine equipment (such as bearing wear or power fluctuation penalty);

[0146] α1, α2, α3 are weight coefficients;

[0147] M is the number of wind farms.

[0148] The constraint conditions are as follows

[0149] Grid frequency change rate and deviation limit: The grid frequency change rate should be controlled within a certain range to avoid violent frequency fluctuations. The specific constraint is:

[0150] |Δf(t)| ≤ Δf max

[0151] Where, Δf max is the maximum allowable deviation of the grid frequency change.

[0152] Upper and lower limits of wind turbine rotational speed operation: The rotational speed of the wind turbine should not exceed its designed safety upper or lower limit to ensure that the wind turbine does not experience rotational speed instability due to excessive inertia release. The constraint condition is:

[0153] ω min ≤ ω i (t) ≤ ω max

[0154] Where, ω i (t) is the rotational speed of the i-th wind turbine, ω min 、ω max are the minimum and maximum safe rotational speeds of the wind turbine respectively.

[0155] Converter capacity constraint: The converter is an important connection component between the wind turbine and the power grid, and its capacity should meet the demand for the inertia release of the wind turbine. The constraint is

[0156] P VSG P(t) ≤ P converter,i

[0157] where P VSG P(t) is the power output of the virtual inertia controller at time t, and P converter,i is the maximum output power of the converter.

[0158] Matching of communication delay and scheduling period: In practical applications, communication delay and scheduling period affect the real-time performance of the system. The constraint is:

[0159] Delay i ≤ ΔT max

[0160] where Delay i is the communication delay of the i-th wind turbine, and ΔT max is the maximum allowable communication delay.

[0161] Limit of dead zone frequency modulation response mechanism: There is a certain dead zone range in the power grid, that is, when the frequency fluctuation is less than a certain critical value, the wind turbine does not inject or adjust inertia. The constraint is:

[0162] |Δf(t)| ≥ Δf dead

[0163] where Δf dead is the frequency modulation dead zone threshold to ensure that unnecessary inertia adjustment will not occur when the frequency change is small.

[0164] Constraint on active power injection ability:

[0165] 0 ≤ P vi P(k) ≤ P vi,max

[0166] Furthermore, the system frequency dynamic model is simplified to an equivalent frequency response model:

[0167]

[0168] Considering the influence of communication network bandwidth limitation and delay uncertainty on scheduling stability, a compensation model is introduced to pre-adjust the control variable P vi P(k). Let the communication delay be τ i , and its compensation mechanism can adopt a one-step prediction extension strategy

[0169]

[0170] where λ i is the dynamic adjustment coefficient, which is evaluated according to the link condition or historical fluctuation.

[0171] When the delay prediction cannot be accurately obtained, the gain attenuation method is used:

[0172]

[0173] In the coordinated dispatch of a wind farm cluster, there are competitive and cooperative relationships among wind farms. The goal of each wind farm is to help the power grid maintain frequency stability by providing virtual inertia while maximizing its own economic benefits. In this multi-agent environment, the decision-making of each wind farm not only depends on its own state but is also affected by the decisions of other wind farms. To solve the multi-agent interest conflicts in the wind farm cluster, the present invention introduces the Nash equilibrium theory. Nash equilibrium means that in a game, each participant knows the strategies of other participants, and no participant can obtain better results by unilaterally changing its own strategy. In the coordinated dispatch of a wind farm cluster, Nash equilibrium can be used to solve the optimal virtual inertia support strategy for each wind farm, so that each wind farm cannot obtain better benefits by changing its dispatch strategy after considering the dispatch behaviors of other wind farms.

[0174] To address the contradiction between the willingness of multiple wind farms in the region to respond autonomously and the system support requirements, a non-cooperative game model is introduced. Let the participants be N wind farms, and the profit function of wind farm i is defined as:

[0175] U i (P vi ) = γ1·S i (P vi ) - γ2·C i (P vi )

[0176] Where:

[0177] S i (P vi ) is the system compensation or frequency stability reward obtained for participating in the support;

[0178] C i (P vi ) is the power generation loss or equipment depreciation caused by the inertia response;

[0179] γ1, γ2 are weight factors, reflecting the negotiation mechanism between the power grid and the enterprise.

[0180] The game objective is to solve the Nash equilibrium point such that for all i ∈ {1, 2,..., N} it satisfies:

[0181]

[0182] Where: represents the equilibrium response of other wind farms except wind farm i.

[0183] This mechanism performs recalculation every T u seconds. During each optimization cycle, the output of the previous step is used as the current initial value, and at the same time, the fan response coefficient K v,i is adjusted to achieve closed-loop feedback correction of the state:

[0184] K v,i (k + 1) = K v,i (k) + η·(Δf(k) - Δf ref )

[0185] where: η: adaptive gain; Δf ref : set frequency deviation threshold.

[0186] Since the dispatching of a wind farm group depends on the real-time transmission and calculation of information such as wind speed, unit speed, and grid frequency, there is a certain time delay in the transmission of this information. Especially in a large-scale distributed wind farm group, the time delay may lead to the lag of dispatching instructions, thereby affecting the real-time performance of dispatching and the stability of the system.

[0187] To solve this problem, the present invention introduces a time-delay compensation algorithm. The specific method is to consider the influence of communication delay during the MPC optimization process and reduce the negative impact brought by the time delay through introducing prediction compensation. The dispatching instruction at the current moment is set as u(t), but due to the communication time delay, the actual wind farm state information received is x(t - Δt), where Δt is the communication time delay.

[0188] To compensate for this time delay, the following compensation model is used:

[0189] u(t) = u(t - Δt) + K·(x(t) - x(t - Δt))

[0190] where K is the compensation gain matrix, which is adjusted by learning the time-delay characteristics of the wind farm group. Through the time-delay compensation algorithm, the influence of communication delay on the dispatching result can be effectively reduced, and the stability and robustness of the system can be ensured.

[0191] Beneficial effects: The present invention has the following advantages:

[0192] By constructing a dynamic inertia collaborative allocation mechanism for a wind farm cluster, the overall inertia potential of dispersed wind power resources in the region is fully exploited and integrated. Compared with the traditional single-machine control or disorderly response mode, the present invention can dynamically and reasonably allocate inertia support tasks according to the real-time inertia demand of the power grid, significantly improve the equivalent inertia level of the power grid, thereby more effectively suppressing the amplitude and change rate of frequency deviation and enhancing the power grid's ability to resist frequency disturbances. The present invention adopts a hierarchical distributed optimization architecture and a multi-objective optimization model, taking into account the operating economy of the wind turbines themselves while meeting the inertia demand of the power grid. Through refined inertia allocation and control strategies, excessive actions of individual or a few units are avoided, the fluctuation amplitude of the unit speed and the frequent adjustment of the active power are reduced, the stress on the mechanical components and electrical equipment of the unit is effectively reduced, the service life of the equipment is extended, and the maintenance cost is reduced. The distributed control architecture reduces the dependence on the central communication system. Even in the case of failures in some communication links, each wind farm can still operate autonomously based on local information, providing a certain amount of inertia support, ensuring the stability and reliability of the system in the face of emergencies. At the same time, the embedded time-delay compensation algorithm further improves the system's adaptability to communication delays. Through the excavation and collaborative optimization of the spatio-temporal complementary characteristics between the stations, the present invention can more flexibly utilize the wind power resources in different regions to participate in power grid frequency support, avoiding the situation where other wind farms are forced to curtail power due to the limited inertia potential of some wind farms, thereby improving the overall utilization rate of wind power resources to a certain extent and reducing the phenomenon of wind curtailment. The scheduling method proposed in the present invention is designed to consider compatibility with the existing wind farm SCADA system and can be deployed by software upgrade without large-scale hardware transformation, reducing the implementation cost and technical risks, and having good engineering application prospects. Through the collaborative control strategy, the disorderly inertia injection of multiple wind turbines is avoided, and the risk of coupling with the inherent frequency of the power grid or the wind farm itself is reduced, which helps to ensure the safe and stable operation of the power system. The regional wind turbine virtual inertia support scheduling method considering the collaborative effect of the wind farm cluster proposed in the present invention can achieve significant beneficial effects in aspects such as improving power grid frequency stability, optimizing the operation of wind turbines, enhancing system robustness, increasing wind power utilization rate, and reducing implementation costs, providing effective technical support for the safe and stable operation of a high-proportion new energy power system. Description of the Drawings

[0193] Figure 1 is the framework diagram of the collaborative virtual inertia support scheduling system for the wind farm cluster;

[0194] Figure 2 is the schematic diagram of the dynamic inertia potential evaluation model for the wind farm cluster;

[0195] Figure 3 is the schematic diagram of dynamic game and rolling correction based on the MPC framework;

[0196] Figure 4 is the flow chart of the multi-LSTM-ARIMA hybrid algorithm;

[0197] Figure 5 is the flow step diagram of the present invention. Specific embodiments

[0198] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0199] Embodiment

[0200] This embodiment provides a regional fan virtual inertia support scheduling method considering the collaborative effect of a wind farm group, including the following steps:

[0201] S1. Evaluation of the dynamic inertia potential of the wind farm group

[0202] Based on the real-time wind speed, unit speed, and grid frequency deviation, a wind farm-level inertia contribution capacity model is constructed to quantify the maximum kinetic energy that each unit can release, the rotational speed safety boundary, and the inertia potential;

[0203] The maximum kinetic energy that each unit can release is calculated as follows:

[0204] The calculation of the rotor kinetic energy of each unit is as follows:

[0205]

[0206] Where:

[0207] I i is the moment of inertia of the wind turbine rotor; ω i (t) is the rotational speed of the i-th fan;

[0208] According to the current rotational speed ω i (t), the range of kinetic energy that can be released is calculated:

[0209]

[0210] Where:

[0211] ω i,min is the minimum safe rotational speed of the wind turbine; ω i,max is the maximum safe rotational speed of the wind turbine;

[0212] The kinetic energy release interval of each unit is:

[0213] ΔE i (t) = E i,max (t) - E i,min (t);

[0214] The quantization steps of the inertia potential are as follows:

[0215] 1) Obtain real-time wind speed data

[0216] Obtain real-time wind speed through a wind speed sensor or meteorological data, and calculate the output power and speed of the unit according to the relationship between the wind speed and the unit speed;

[0217] 2) Calculate real-time speed

[0218] Calculate the speed of each unit using the wind turbine power curve according to the wind speed and power output;

[0219] 3) Estimate rotor kinetic energy

[0220] Calculate the rotor kinetic energy of each unit according to the speed;

[0221] 4) Grid frequency deviation monitoring

[0222] Monitor the grid frequency deviation in real time, and adjust the virtual inertia support ability of the wind turbine according to the change of the grid frequency;

[0223] 5) Evaluate inertia potential

[0224] According to the above data, evaluate the inertia contribution ability of the wind farm group in real time, determine the virtual inertia that each unit can provide, and then calculate the total inertia contribution of the wind farm group.

[0225] It also includes the correction of unit inertia evaluation, specifically as follows:

[0226] Based on ARIMA, establish a wind speed prediction model to predict the wind speed v i (t) in the next few moments, so as to predict the output power and speed change of the future wind turbine;

[0227] Calculate the virtual inertia response ability of each unit under different conditions as:

[0228] H i (t) = ΔE i (t)·(1 + β1·Δf(t) + β2·v i (t))

[0229] Where:

[0230] β1 and β2 are weight coefficients;

[0231] Define the total inertia contribution H total (t) of the wind farm group as the sum of the inertia contributions of each unit:

[0232]

[0233] Where:

[0234] N is the total number of units in the wind farm group.

[0235] S2. Regional Collaborative Scheduling Architecture Design

[0236] Adopt a "centralized - decentralized" hybrid multi - level collaborative control architecture mode:

[0237] Regional Scheduling Layer: Based on the Wide - Area Measurement System (WAMS) data, calculate the grid inertia deficit in real - time and calculate the inertia demand distribution coefficient of each wind farm;

[0238] S211: According to the real - time grid frequency data and the grid state information provided by WAMS, calculate the current inertia gap of the grid as follows:

[0239] ΔH gap =H required -H available

[0240] Where:

[0241] H required is the total inertia required by the grid currently; H available is the available inertia currently;

[0242] S212: Combine the Geographic Information System (GIS) and real - time wind speed data to calculate the inertia support capacity of each wind farm, and dynamically allocate inertia tasks according to the inertia gap of the regional grid;

[0243] The inertia allocation weight of the wind farm is calculated as follows:

[0244]

[0245] Where:

[0246] w i is the inertia allocation weight of the i - th wind farm; α i is the factor related to the distance from the i - th wind farm;

[0247] V i is the wind speed of the i - th wind farm;

[0248] Station Coordination Layer: Distributed decision - making within the wind farm, dynamically evaluate the adjustable inertia margin of each station, and optimize the unit action priority and output curve in combination with local information;

[0249] S221: Based on the actual operating conditions of each wind farm and the inertia demand distribution of the regional scheduling layer, calculate the inertia allocation priority for each unit in the station. The priority scheduling formula is:

[0250]

[0251] Where:

[0252] P priority,iis the inertia allocation priority of the i-th wind turbine;

[0253] R health,i is the health coefficient of the i-th wind turbine;

[0254] S222. Dynamically adjust the virtual inertia contribution of the unit through the health coefficient as follows:

[0255] R health,i = w1·R operating,i + w2·R maintain,i + w3·R wear,i + w4·R failure,i

[0256] Where:

[0257] R health,i is the health coefficient of the wind turbine;

[0258] R operating,i is the operation status evaluation of the wind turbine;

[0259]

[0260] Where, P actual,i is the actual output power of the wind turbine, P rated,i is the rated power of the wind turbine;

[0261] R maintain,i is the maintenance status of the wind turbine;

[0262]

[0263] Where, T last maintain,i is the duration from the last maintenance of the wind turbine to the current time, T scheduled maintain,i is the scheduled maintenance period of the wind turbine;

[0264] R wear,i is the wear condition of the wind turbine components;

[0265]

[0266] Where, W current,i is the current wear level of the wind turbine component, W max,i is the maximum wear resistance of the component;

[0267] R failure,i is the failure rate of the wind turbine;

[0268]

[0269] Where, F failures,i is the number of failures occurred by the wind turbine, F max,i is the maximum acceptable number of failures of the wind turbine;

[0270] w1, w2, w3, and w4 are the weights of each item index.

[0271] Unit execution layer: adaptively adjust the VSG control parameters;

[0272] S3. Dynamic inertia demand prediction model

[0273] Adopt the LSTM-ARIMA hybrid algorithm, fuse wind speed prediction, load fluctuation, and unit status to generate a time-varying inertia demand curve;

[0274] S31. Data preprocessing and feature extraction

[0275] Collect real-time data of wind speed, load, grid frequency, and unit status, and perform standardization processing; extract relevant features, including wind speed change rate, load fluctuation amplitude, and grid frequency deviation;

[0276] S32. LSTM model building

[0277] Use the LSTM neural network to process data of historical inertia demand and related factors. The input sequence is the historical inertia demand and the time series data of external factors including wind speed, load, and frequency, and the output is the inertia demand prediction for a future period;

[0278] The LSTM model includes an input gate, a forget gate, and an output gate. The model formula is as follows:

[0279] f t = σ(W f · [h t-1 , x t + b f )

[0280] i t = σ(W i · [h t-1 , x t + b i )

[0281] o t = σ(W o · [h t-1 , x t + b o )

[0282] c t = f t · c t-1 + i t · tanh(W c · [h t-1 , x t + b c )

[0283] h t = o t ·tanh(c t )

[0284] Among them, f t is the forget gate, i t is the input gate, o t is the output gate, c t is the cell state, h t is the output state, σ is the activation function, x t is the input data, W f , W, W o , W c are the weight matrices, b f , b i , b o , b c are the bias terms;

[0285] h t = tanh(W h ·[h t-1 , x t , CNN(x t )] + b h )

[0286] y t = W y ·h t + b y

[0287] Among them:

[0288] CNN(x t ) is the feature output after convolving the input data;

[0289] y t is the output of the LSTM;

[0290] The described ARIMA model introduces an optimization strategy for the differencing order. The ARIMA model formula is as follows:

[0291]

[0292] Among them: φ i is the autoregressive coefficient, θ j is the moving average coefficient, ε t is the error term, S t is the seasonal adjustment term.

[0293] S33, ARIMA model modeling

[0294] Based on the known power grid frequency fluctuation data, use the ARIMA model to model the power grid frequency change. The input is the historical data of the power grid frequency, and the output is the frequency prediction value at future time points;

[0295] S34. LSTM-ARIMA Fusion Model

[0296] Fuse the prediction results of LSTM and ARIMA and output the inertia demand curve.

[0297] After fusing the prediction results of LSTM and ARIMA in the above steps, an adaptive weighting mechanism is carried out. The weighting formula is as follows:

[0298]

[0299] Where:

[0300] is the final inertia demand prediction value; y LSIM,t is the output of the LSTM model; y ARIMA,t is the output of the ARIMA model.

[0301] α t is the adaptive weighting coefficient;

[0302]

[0303] Where: Err LSTM,t and Err ARIMA,t are the prediction errors of the LSTM and ARIMA models respectively.

[0304] S4. Multi-objective Optimization Model Construction

[0305] Construct a multi-objective optimization model. Its objective function comprehensively considers maximizing the equivalent inertia level of the regional power grid, minimizing the fan speed recovery cost and equipment loss, and suppressing the secondary oscillation risk caused by the operation of multiple units. The constraint conditions include the power grid frequency deviation limit, the unit speed safety threshold, the converter capacity limit, the communication delay tolerance, and the frequency modulation dead zone threshold;

[0306] S5. Dynamic Game and Rolling Correction

[0307] Based on the model predictive control framework, rollingly update the wind farm cluster scheduling instructions, embed the time-delay compensation algorithm to improve the robustness, and combine the Nash equilibrium theory to solve the multi-agent interest conflict.

[0308] The specific method is as follows:

[0309] S51. Dynamic scheduling strategy based on the model predictive control framework, introducing the time-delay compensation algorithm and the Nash equilibrium theory at the same time to solve the multi-agent conflict and scheduling time-delay problems among wind farm clusters;

[0310] S52. In each scheduling period, taking the current state of the wind farm cluster as the initial condition, construct a rolling optimization problem, predict the inertia demand and system disturbances in multiple future time periods, and obtain the optimal allocation strategy of the wind turbine inertia response;

[0311] Let the prediction horizon be N, and the control variable be the virtual inertia response power of wind farm i as P vi (k). Then the optimization objective function is:

[0312]

[0313] Where:

[0314] Δf(k) is the system frequency deviation at the k-th step; Δω i (k) is the rotational speed deviation of wind turbine i at the k-th step; C deg is the degradation cost function caused by the response of the wind turbine equipment; α1, α2, α3 are weight coefficients; M is the number of wind farms;

[0315] The constraint conditions are as follows:

[0316] Grid frequency change rate and deviation limit:

[0317] |Δf(t)| ≤ Δf max

[0318] where, Δf max is the maximum allowable deviation of the grid frequency change;

[0319] Lower and upper limits of the wind turbine rotational speed operation:

[0320] ω min ≤ ω i (t) ≤ ω max

[0321] where, ω i (t) is the rotational speed of the i-th wind turbine, ω min , ω max are the minimum and maximum safe rotational speeds of the wind turbine respectively;

[0322] Converter capacity constraint:

[0323] P VSG (t) ≤ P converter,i

[0324] where, P VSG (t) is the power output of the virtual inertia controller at time t, P converter,i is the maximum output power of the converter;

[0325] Communication delay and scheduling period matching:

[0326] Delayi ≤ΔT max

[0327] where Delay i is the communication delay of the i-th fan, and ΔT max is the maximum allowable communication time delay;

[0328] Dead zone frequency modulation response mechanism limitation:

[0329] |Δf(t)|≥Δf dead

[0330] where Δf dead is the frequency modulation dead zone threshold;

[0331] Active power injection capacity constraint:

[0332] 0≤P vi (k)≤P vi,max < / /

[0333] S53. The system frequency dynamic model is simplified to an equivalent frequency response model:

[0334]

[0335] Assume the communication time delay is τ i , and adopt a one-step prediction extension strategy as the compensation mechanism:

[0336]

[0337] where: λ i is the dynamic adjustment coefficient;

[0338] When the delay prediction cannot be accurately obtained, use the gain attenuation method:

[0339]

[0340] S??. To address the contradiction between the autonomous response willingness of multiple wind farms within the region and the system support requirements, introduce a non-cooperative game model;

[0341] Define the profit function of wind farm i as: ?

[0342] U i (P vi )=γ1·S i (P vi )-γ2·C i (P vi )

[0343] where:

[0344] S i (P viSystem compensation or frequency stability rewards obtained for participating in support;

[0345] C i (P vi ) Generation losses or equipment depreciation due to inertia response;

[0346] γ1 and γ2 are weighting factors;

[0347] S55. The game objective is to solve the Nash equilibrium point such that for all i ∈ {1, 2,..., N} it satisfies:

[0348]

[0349] where: is the equilibrium response of other wind farms except wind farm i;

[0350] S56. Recalculation is performed every T u seconds. In each optimization cycle, the previous output is used as the current initial value, and at the same time, the fan response coefficient K v,i is adjusted to achieve state closed-loop feedback correction:

[0351] K v,i (k + 1) = K v,i (k) + η·(Δf(k) - Δf ref )

[0352] where:

[0353] η is the adaptive gain; Δf ref is the set frequency deviation threshold.

[0354] As Figure 1 shown is the framework diagram of the collaborative virtual inertia support scheduling system for a wind farm cluster, demonstrating how this embodiment realizes the virtual inertia support of the wind farm cluster through a multi-level collaborative control architecture. The core objective of this system is to achieve dynamic collaborative scheduling among different wind farms, thereby optimizing the inertia support capacity of the wind farm cluster, ensuring the stability of the power grid frequency, and reducing equipment losses.

[0355] The overall framework of this system is divided into three main levels: the regional dispatching level, the station coordination level, and the unit execution level. The regional dispatching level is responsible for collecting power grid data in real time through the wide-area measurement system WAMS, including information such as power grid frequency and load fluctuations. By analyzing the real-time state of the power grid, the regional dispatching level calculates the inertia demand of the wind farm cluster and allocates this demand to different wind farms.

[0356] In the substation coordination layer, there is a distributed decision-making system in each wind farm. This layer evaluates the inertia contribution ability of each wind farm and the priority of the virtual inertia support task based on local wind speed, unit speed and other information. The substation coordination layer is responsible for dynamically adjusting the virtual inertia control (VSG) parameters of each unit and feeding back the adjustment instructions to each wind farm. Through this mechanism, each unit in the wind farm can dynamically adjust its speed and output power according to the current demand, so as to provide the best inertia support without affecting the grid frequency.

[0357] The unit execution layer is the most basic execution unit. Each wind turbine responds to the instructions from the substation coordination layer by adjusting the speed, power output and virtual inertia control parameters in real time. The unit execution layer mainly executes the virtual inertia control strategy to simulate the inertia characteristics of the synchronous generator to help the power grid recover stability when frequency fluctuations occur.

[0358] Through spatio-temporal collaborative optimization, the system maximizes the equivalent inertia of the regional power grid, reduces the number of frequent starts and stops of equipment, reduces equipment wear, and thus improves the economy of the wind farm group. In addition, the system can also operate autonomously in the case of partial communication link interruption, enhancing the robustness of the system.

[0359] Through the design of the hierarchical architecture, each layer of the wind farm group can flexibly adjust the scheduling strategy with the support of real-time data to cope with the changing grid state and ensure the stability of the grid frequency and the efficient operation of the wind farm group.

[0360] In this flowchart, the grid data is transmitted to the regional dispatch layer in real time through the wide area measurement system WAMS, which calculates the grid inertia gap and allocates the inertia demand of the wind farm group accordingly. The substation coordination layer optimizes the unit scheduling strategy based on local information and transmits the optimized instructions to the unit execution layer. The unit execution layer adjusts the virtual inertia control according to the instructions and finally responds to the inertia demand of the grid. The system continuously optimizes the scheduling strategy through a rolling correction mechanism to ensure the efficient operation of the wind farm group in a dynamic environment and the stability of the grid frequency.

[0361] As Figure 2 shown is a schematic diagram of the dynamic inertia potential evaluation model of the wind farm group in this embodiment, which shows how to evaluate the inertia potential of the wind farm group based on real-time wind speed, unit speed and grid frequency deviation data. The core goal of this model is to quantify the inertia support ability of each wind turbine in different working states in real time, and through this evaluation result, help the regional dispatch layer reasonably allocate the inertia support tasks.

[0362] The model first collects real-time data from each unit in the wind farm group, including wind speed, unit speed, and grid frequency deviation. This data is collected through a real-time monitoring system (such as a SCADA system) and transmitted to the evaluation model. Based on this information, the model constructs an inertia contribution capacity model for wind turbines to quantify the maximum kinetic energy that each turbine can release and the rotational speed safety boundary under the current operating conditions.

[0363] To accurately evaluate the inertia potential of each unit, the model also considers multiple factors, such as the impact of wind speed fluctuations on the unit speed, the mechanical stability of the unit, and the response ability of the unit to grid frequency fluctuations. The model combines these factors and dynamically calculates the amount of inertia support that each unit can provide, thereby obtaining the overall inertia potential of the wind farm group. Finally, the evaluation results will provide input to the regional dispatching layer to help it make more accurate dispatching decisions.

[0364] This model can evaluate the inertia potential of the wind farm group in real time under different wind speeds and grid conditions, and can dynamically adjust the evaluation strategy to adapt to the changes in the operating conditions of the wind farm group. This dynamic evaluation can improve the accuracy of the system's inertia support ability for the wind farm group, thereby optimizing the overall dispatching effect, ensuring grid frequency stability, and maximizing the utilization rate of the wind farm.

[0365] In this flowchart, the real-time data of wind turbines (including wind speed, speed, and grid frequency deviation) is first collected and transmitted to the dynamic inertia potential evaluation model. The model analyzes this data, evaluates the impact of wind speed on the unit speed, and calculates the inertia contribution capacity of each unit. By calculating the maximum kinetic energy that each unit can release, the inertia potential of the entire wind farm group is finally obtained. These evaluation results will be transmitted to the regional dispatching layer to guide the inertia task allocation of the wind farm group and ensure the accuracy and real-time nature of the evaluation through dynamic adjustment strategies.

[0366] As Figure 3 shown is a schematic diagram of dynamic game and rolling correction based on the MPC framework, demonstrating how to achieve effective coordination and strategy optimization among wind farm groups through the combination of dynamic game theory and model predictive control in a multi-agent environment. The design goal of this flowchart is to ensure that the wind farm group can dynamically adjust the response strategy based on its own inertia potential and operating status while meeting the grid frequency stability, so as to achieve the overall optimum of the system.

[0367] First, the system performs rolling optimization through the model predictive control framework. This process includes predicting key factors such as grid frequency demand and wind speed changes in the short term in the future. In each rolling control cycle, the system dynamically updates the dispatching strategy according to the new prediction information, thereby adjusting the virtual inertia support behavior of the wind farm group.

[0368] The game control part plays a key role in this process. By introducing the Nash equilibrium theory, the system optimizes the responses among multiple wind farms to ensure that each wind farm selects the optimal response strategy according to its own local interests and the overall goal. In each control cycle, each wind farm, as an independent entity, decides its response behavior based on the current grid state and its own inertia contribution. After these response behaviors go through the game process, a stable equilibrium point is reached, enabling each wind farm to meet the needs of the grid frequency while maximizing its own benefits when providing virtual inertia support.

[0369] At the end of each rolling control cycle, the system continues to perform prediction and optimization adjustment for the next time step through the updated game solution and MPC adjustment strategy. Through this rolling correction mechanism, the system can effectively cope with real-time disturbances and environmental changes, improving the collaborative efficiency of the wind farm group and the system robustness.

[0370] As Figure 4 shown is the flowchart of the LSTM-ARIMA hybrid algorithm in this embodiment. It shows the whole process from the start of multi-source data collection, through a series of refined processing steps, and finally generating an accurate time-varying inertia demand curve. At the beginning of the flowchart, the multi-source data collection link presents a scene of extensive and complex information aggregation. On the one hand, high-precision wind speed prediction data is obtained from professional meteorological departments. Relying on advanced meteorological prediction models and a dense network of meteorological monitoring stations, these data can give a relatively accurate estimate of the wind speed change trend of wind farms in different regions in the future period. On the other hand, a close data interaction channel is established with the power grid dispatching center to introduce load fluctuation information in real time. The dynamic change of the load reflects the real-time demand of the power grid's power consumption side and has a key impact on the determination of inertia demand. At the same time, a large number of sensors deployed by the wind farm group continuously collect unit state data, covering all aspects such as the real-time speed, power output, and equipment health status of the units, comprehensively reflecting the operation status of the wind farm group.

[0371] The collected raw data then enters the preprocessing stage. During this process, data cleaning techniques are used to remove noise interference, outliers, and obvious incorrect data points in the data to ensure the purity of the data. Subsequently, the data normalization operation unifies data with different magnitudes and physical meanings into the same scale range, facilitating the processing and analysis of subsequent algorithms. The preprocessed multi-source data is diverted to different paths according to its characteristics. The wind speed prediction data and load fluctuation data mainly flow to the ARIMA model training module. The ARIMA model, with its strong modeling and prediction capabilities for time series data, deeply mines and learns the time series information with obvious periodicity and trend such as wind speed and load, and constructs a preliminary prediction model framework.

[0372] The unit status data, on the other hand, enters the LSTM model training module more. Due to its unique long short-term memory cell structure, the LSTM is good at processing unit status sequence data with long-term dependence relationships and obvious non-linear characteristics, and can accurately capture the subtle changes and internal correlations of the unit operation status over time. During the training process of each model, repeated learning and optimization are carried out based on a large amount of historical data, continuously adjusting the model parameters to improve the prediction accuracy of the model. When the two models are initially trained, they enter the fusion prediction link. At this stage, the prediction results of the ARIMA model for wind speed and load and the prediction results of the LSTM model for unit status are organically combined through a specific fusion algorithm, giving full play to the advantages of the two models, complementing and verifying each other.

[0373] Finally, the fused comprehensive data is input into the time-varying inertia demand curve generation module. Based on a set of pre-set rules and algorithms, this module converts the multi-source fusion information into a visual and time-dynamically changing inertia demand curve, presenting the demand changes of the power grid for the inertia support of the wind farm group in a future period of time, providing a basis for the subsequent dispatching decision-making of the wind farm group, and ensuring the real-time matching of the inertia support of the wind farm group and the power grid demand.

[0374] The above has shown and described the basic principles, main features and advantages of the present invention. The present invention will also have various changes and improvements, and these changes and improvements fall within the scope of the present invention claimed.

Claims

1. A regional fan virtual inertia support scheduling method considering the collaborative effect of a wind farm group, characterized in that, It includes the following steps: S1. Dynamic inertia potential assessment of wind farm group Based on real-time wind speed, unit speed and grid frequency deviation, a wind farm-level inertia contribution capacity model is constructed to quantify the maximum kinetic energy that can be released by each unit, the rotational speed safety boundary and the inertia potential; S2. Design of regional collaborative scheduling architecture Adopt a "centralized-decentralized" hybrid multi-level collaborative control architecture mode: Regional scheduling layer: Based on wide-area measurement WAMS data, calculate the grid inertia deficit in real time and calculate the inertia demand distribution coefficient of each wind farm; Substation coordination layer: Distributed decision-making within the wind farm, dynamically evaluate the adjustable inertia margin of each substation, and optimize the unit action priority and output curve in combination with local information; Unit execution layer: Adaptive adjustment of VSG control parameters; S3. Dynamic inertia demand prediction model Adopt the LSTM-ARIMA hybrid algorithm, fuse wind speed prediction, load fluctuation and unit status to generate a time-varying inertia demand curve; S4. Construction of multi-objective optimization model Construct a multi-objective optimization model, the objective function of which comprehensively considers maximizing the equivalent inertia level of the regional power grid, minimizing the fan speed recovery cost and equipment loss, and suppressing the secondary oscillation risk caused by multi-unit actions. The constraint conditions include grid frequency deviation limit, unit speed safety threshold, converter capacity limit, communication delay tolerance, and frequency modulation dead zone threshold; S5. Dynamic game and rolling correction Based on the model predictive control framework, roll-update the wind farm group scheduling instruction, embed the time-delay compensation algorithm to improve the robustness, and solve the multi-agent interest conflict in combination with the Nash equilibrium theory.

2. The regional fan virtual inertia support scheduling method considering the collaborative effect of wind farm clusters according to claim 1, characterized in that In step S1, the maximum kinetic energy that can be released by each unit is calculated as follows: The rotor kinetic energy of each unit is calculated as follows: Where: I i is the moment of inertia of the rotor of the wind turbine; ω i (t) is the rotational speed of the i-th fan; Based on the current rotational speed ω i (t), calculate the range of kinetic energy that can be released: Where: ω i,min is the minimum safe rotational speed of the wind turbine; ω i,max is the highest safe rotational speed of the wind turbine; The range of kinetic energy that can be released by each unit is: ΔE i (t) = E i,max (t) - E i,min (t); The quantization steps of inertia potential are as follows: 1) Obtain real-time wind speed data Obtain real-time wind speed through a wind speed sensor or meteorological data, and calculate the output power and speed of the unit according to the relationship between the wind speed and the unit speed; 2) Calculate real-time speed According to the wind speed and power output, calculate the speed of each unit using the fan power curve; 3) Estimate rotor kinetic energy Calculate the rotor kinetic energy of each unit according to the speed; 4) Grid frequency deviation monitoring Real-time monitor the grid frequency deviation, and adjust the virtual inertia support capacity of the wind turbine according to the change of the grid frequency; 5) Evaluate inertia potential According to the above data, evaluate the inertia contribution capacity of the wind farm group in real time, determine the virtual inertia that each unit can provide, and then calculate the total inertia contribution of the wind farm group.

3. The regional wind turbine virtual inertia support scheduling method considering the collaborative effect of wind farm clusters according to claim 2, characterized in that, Step S1 also includes the correction of unit inertia evaluation, which is specifically as follows: Establish a wind speed prediction model based on ARIMA to predict the wind speed v i (t) at several future moments, so as to predict the output power and rotational speed changes of future wind turbines; The virtual inertia response ability of each unit under different conditions is calculated as: H i H(t) = ΔE i H(t)·(1 + β1·Δf(t) + β2·v i (t)) Where: β1 and β2 are weight coefficients; Define the total inertia contribution \(H\) of the wind farm group total (t) is the sum of the inertia contributions of each unit: Where: N is the total number of units in the wind farm group.

4. The regional fan virtual inertia support scheduling method considering the collaborative effect of wind farm clusters according to claim 1, characterized in that The specific scheduling method of the regional scheduling layer in step S2 is as follows: S211: According to the real-time grid frequency data and the grid state information provided by WAMS, calculate the current inertia gap of the grid as follows: ΔH gap = H required - H available Where: H required is the total inertia required by the power grid currently; H available is the currently available inertia; S212. Combine the geographic information system and real-time wind speed data, calculate the inertia support capacity of each wind farm, and dynamically allocate inertia tasks according to the inertia gap of the regional power grid; The inertia allocation weight of the wind farm is calculated as follows: Where: w i is the inertia distribution weight for the i-th wind farm; α i is a factor related to the distance from the i-th wind farm; V i is the wind speed of the i-th wind farm; The specific scheduling method of the station coordination layer is as follows: S221. Calculate the inertia allocation priority for each unit in the station based on the actual operating conditions of each wind farm and the inertia demand allocation of the regional scheduling layer. The priority scheduling formula is: Where: P priority,i is the inertia allocation priority of the i-th wind turbine; R health,i is the health coefficient of the i-th wind turbine; S222. Dynamically adjust the virtual inertia contribution of the unit through the health factor, as follows: R health,i = w1·R operating,i + w2·R maintain,i + w3·R wear,i + w4·R failure,i Where: R health,i is the health factor of the fan; R operating,i For the operation status evaluation of the fan Among them, P actual,i is the actual output power of the fan, and P rated,i is the rated power of the fan; R maintain,i is the maintenance status of the fan; Among them, T lastmaintain,i is the duration from the last maintenance of the fan to the current time, and T scheduledmaintain,i is the predetermined maintenance cycle of the fan; R wear,i is the wear condition of the fan components; where W current,i is the current wear level of the fan component, and W max,i is the maximum wear resistance of the component; R failure,i is the failure rate of the fan; Among them, F failures,i is the number of faults that occur in the fan, and F max,i is the maximum number of faults that the fan can accept; w1, w2, w3, and w4 are the weights of each index.

5. The regional fan virtual inertia support scheduling method considering the collaborative effect of wind farm clusters according to claim 1, characterized in that The specific method of step S3 is as follows: S31. Data preprocessing and feature extraction Collect real-time data of wind speed, load, grid frequency, and unit status, and perform standardization processing; extract relevant features, including wind speed change rate, load fluctuation amplitude, and grid frequency deviation; S32. LSTM model building Use the LSTM neural network to process the data of historical inertia demand and related factors. The input sequence is the historical inertia demand and the time series data of external factors including wind speed, load, and frequency, and the output is the predicted inertia demand for a future period of time; S33. ARIMA model building Based on the known grid frequency fluctuation data, use the ARIMA model to model the grid frequency change. The input is the historical data of the grid frequency, and the output is the predicted frequency value at a future time point; S34. LSTM-ARIMA fusion model Fuse the prediction results of LSTM and ARIMA to output the inertia demand curve.

6. The regional fan virtual inertia support scheduling method considering the collaborative effect of wind farm groups according to claim 5, characterized in that The LSTM model includes an input gate, a forget gate, and an output gate. The model formula is as follows: f t = σ(W f · [h t-1 , x t + b f ) i t = σ(W i · [h t-1 , x t + b i ) o t = σ(W o · [h t-1 , x t + b o ) c t = f t · c t-1 + i t · tanh(W c · [h t-1 , x t + b c ) h t = o t ·tanh(c t ) Among them, f t is the forget gate, i t is the input gate, o t is the output gate, c t is the cell state, h t is the output state, σ is the activation function, x t is the input data, W f , W, W o , W c are the weight matrices, b f , b i , b o , b c are the bias terms; h t = tanh(W h ·[h t-1 , x t , CNN(x t )] + b h ) y t = W y ·h t + b y Where: CNN(x t ) is the feature output after convolutional processing of the input data; y t is the output of the LSTM; The ARIMA model introduces an optimization strategy for the difference order. The ARIMA model formula is as follows: Where: φ i is the autoregressive coefficient, θ j is the moving average coefficient, ε t is the error term, S t is the seasonal adjustment term.

7. The regional fan virtual inertia support scheduling method considering the collaborative effect of wind farm clusters according to claim 6, characterized in that In step S34, after fusing the prediction results of LSTM and ARIMA, an adaptive weighting mechanism is performed. The weighting formula is as follows: Where: is the final predicted value of the inertia requirement; y LSIM,t is the output of the LSTM model; y ARIMA,t Is the output of the ARIMA model. α t is the adaptive weighting coefficient; Where: Err LSTM,t and Err ARIMA,t are the prediction errors of the LSTM and ARIMA models respectively.

8. The regional fan virtual inertia support scheduling method considering the collaborative effect of wind farm clusters according to claim 1, characterized in that The specific method of step S5 is as follows: S51. Based on the dynamic scheduling strategy of the model predictive control framework, introduce the time-delay compensation algorithm and the Nash equilibrium theory to solve the multi-agent conflict and scheduling time-delay problems between wind farm groups; S52. In each scheduling period, with the current state of the wind farm group as the initial condition, construct a rolling optimization problem, predict the inertia demand and system disturbances in multiple future time periods, and obtain the optimal fan inertia response allocation strategy; Let the prediction horizon be \(N\), and the control variable be the virtual inertia response power \(P(k)\) of wind farm \(i\). Then the optimization objective function is as follows: vi (k), then the optimization objective function is: Where: Δf(k) is the system frequency deviation at the kth step; Δω i (k) is the rotational speed deviation of the i-th fan at the k-th step; C deg The degradation cost function caused by the response of the fan equipment; α1, α2, and α3 are weight coefficients; M is the number of wind farms; The constraint conditions are as follows: Grid frequency change rate and deviation limit: |Δf(t)|≤Δf max where Δf max is the maximum allowable deviation of the power grid frequency change; Upper and lower limits of fan speed operation: ω min ≤ ω i (t) ≤ ω max where ω i (t) is the rotational speed of the i-th fan, and ω min , ω max are the minimum and maximum safe rotational speeds of the fan, respectively; Converter capacity constraint: P VSG (t) ≤ P converter,i where, P VSG (t) is the power output of the virtual inertia controller at time t, P converter,i is the maximum output power of the converter; Matching of communication delay and scheduling period: Delay i ≤ΔT max Among them, Delay i is the communication delay of the i-th fan, and ΔT max is the maximum allowable communication time delay; Dead zone frequency modulation response mechanism limit: |Δf(t)|≥Δf dead where, Δf dead is the frequency modulation dead zone threshold; Active power injection capacity constraint: 0 ≤ P vi (k) ≤ P vi,max S53. Simplify the system frequency dynamic model into an equivalent frequency response model: Let the communication delay be τ i , and adopt a one-step prediction extension strategy as a compensation mechanism: where: λ i is the dynamic adjustment coefficient; When the delay prediction cannot be accurately obtained, use the gain attenuation method: S54. To address the contradiction between the autonomous response willingness of multiple wind farms in the region and the system support requirements, introduce a non-cooperative game model; Define the revenue function of wind farm i as: U i (P vi ) = γ1·S i (P vi ) - γ2·C i (P vi ) Where: S i (P vi ) is the system compensation or frequency stability reward obtained for participating in support C i (P vi ) is the power generation loss or equipment depreciation due to inertia response; γ1 and γ2 are weight factors; S55. The game objective is to solve for the Nash equilibrium point such that for all i ∈ {1, 2,..., N}, it holds that: Wherein: is the balanced response of other wind farms except wind farm i; S56, every T u seconds for recalculation, using the previous step output as the current initial value in each optimization cycle, and at the same time adjusting the fan response coefficient K v,i to achieve closed-loop feedback correction of the state: K v,i (k + 1) = K v,i (k) + η·(Δf(k) - Δf ref ) Where: η is the adaptive gain; Δf ref is the set frequency deviation threshold value.

Images