An ai-based power and scheduling deviation early warning system for wind farm clusters

Abstract

The application relates to the technical field of new energy power dispatching, and discloses a wind farm group centralized power and dispatching deviation early warning system based on AI; the system comprises a multi-source heterogeneous data fusion unit, a physical information guided double-scale AI power prediction unit, a wind farm group dynamic adjustable capacity boundary evaluation unit, a dispatching deviation graded early warning unit, a closed-loop pre-control strategy generation and linkage execution unit and a model online self-iteration unit. The application embeds physical constraints into an AI power prediction model, improves prediction accuracy under extreme working conditions, initiates dynamic adjustable capacity boundary evaluation, distinguishes controllable and uncontrollable deviations, greatly reduces early warning false alarm rates, constructs a closed-loop early warning disposal system, realizes linkage execution of deviation early warning and pre-control strategies, effectively reduces wind farm group dispatching deviations, and supports safe and stable operation of a power grid and new energy consumption.

Description

Technical Field

[0001] This invention belongs to the field of new energy power dispatching technology, specifically an AI-based early warning system for centralized power and dispatching deviation in wind farm clusters. Background Technology

[0002] The power output of wind farm clusters is affected by meteorological factors such as wind speed, wind direction, and turbulence, exhibiting strong randomness, volatility, and intermittency. This can easily lead to deviations between actual power output and grid dispatch instructions, not only subjecting the grid to performance evaluations according to relevant regulations but also impacting the power balance and safe, stable operation of the regional power grid. Therefore, accurate power forecasting and dispatch deviation early warning for wind farm clusters have become core requirements in the field of new energy dispatch.

[0003] Existing technologies for wind farm power prediction and deviation early warning mainly suffer from the following core defects: 1. Poor generalization of power prediction models and sharp drop in accuracy under extreme conditions: Existing solutions mostly use pure data-driven neural network models, which are trained only on historical data and do not couple the physical laws such as wind turbine aerodynamic characteristics, wind farm wake effect, and power grid balance. Under extreme conditions such as typhoons, turbulence, and wind turbine icing, the prediction error increases significantly, which directly leads to high rates of missed warnings and false alarms, and fails to meet the accuracy requirements of power grid dispatch.

[0004] 2. The early warning logic has a fundamental flaw, resulting in an extremely high false alarm rate: Existing deviation early warning systems rely solely on the "static difference between predicted power and dispatch instructions," failing to consider the real-time adjustability of wind farm clusters and thus unable to distinguish between "uncontrollable deviations" and "controllable deviations." For example, if there is a deviation between the predicted output and the instruction, but this deviation can be eliminated by adjusting wind turbine parameters, the existing technology will still trigger an early warning, rendering the early warning information meaningless and preventing operators from identifying core risks.

[0005] 3. Open-loop early warning mode, no closed-loop handling capability: Existing systems can only output early warning signals, without the ability to generate and execute supporting pre-control strategies. After the early warning, they cannot provide effective handling solutions, nor can they directly link field equipment for execution. They cannot reduce dispatch deviations from the root and are difficult to adapt to the requirements of refined power grid dispatch.

[0006] 4. Single-site design cannot meet the centralized scheduling requirements of wind farm clusters: The existing solution does not consider the spatiotemporal coupling characteristics of wind farm clusters, including the wake effect between wind farms, power flow constraints at the grid connection section, and the collaborative adjustability of multiple sites. It cannot achieve centralized early warning and collaborative control of regional wind farm clusters, and it is difficult to support the scheduling and management of large-scale wind farm clusters by the power grid. Summary of the Invention

[0007] The purpose of this invention is to overcome the above-mentioned defects of the prior art and provide an AI-based early warning system for centralized power and scheduling deviations in wind farm clusters, which has high prediction accuracy, high early warning accuracy, closed-loop processing capability, and adaptability to the centralized scheduling needs of wind farm clusters. This system solves the core pain points of the prior art and supports the safe and stable operation of the power grid and the efficient consumption of new energy.

[0008] To achieve the above-mentioned objectives, the technical solution of the present invention is as follows: The core architecture of the system includes a multi-source heterogeneous data fusion unit, a physical information-guided dual-scale AI power prediction unit, a wind farm group dynamic adjustable capability boundary assessment unit, a scheduling deviation classification and early warning unit, a closed-loop pre-control strategy generation and linkage execution unit, a model online self-iteration and optimization unit, and a supporting human-machine interaction and scheduling communication unit.

[0009] Among them, the multi-source heterogeneous data fusion unit provides a standardized data foundation for the system; the dual-scale AI power prediction unit outputs accurate power prediction sequences; the dynamic adjustable capability boundary assessment unit provides core controllability judgment basis for early warning; the scheduling deviation classification and early warning unit realizes accurate deviation classification and classification early warning; the closed-loop pre-control strategy generation and linkage execution unit realizes the closed loop of early warning and disposal; and the online self-iteration model unit ensures the stability and accuracy of the system's long-term operation.

[0010] Compared with the prior art, the beneficial effects of the present invention are: This invention employs a physical information-guided dual-scale AI power prediction model, embedding physical laws such as wind turbine aerodynamic power equations, wake constraints, and grid power balance as regularization terms into a neural network. This addresses the shortcomings of purely data-driven models, such as poor generalization and sharp drops in accuracy under extreme conditions. Field tests have verified that the prediction accuracy under normal operating conditions can reach over 95%, and under extreme conditions, the prediction accuracy is improved by more than 30% compared to purely data-driven models, while the false alarm rate is reduced by more than 40%. Therefore, this invention significantly improves power prediction accuracy and generalization, fundamentally reducing the false alarm rate. This invention calculates the adjustable upper and lower limits of wind farm output under current operating conditions in real time, accurately distinguishing between "uncontrollable deviations" and "controllable deviations." A high-level warning is triggered only when the deviation exceeds the adjustable capacity range, solving the core defect of high false alarm rates in existing technologies. Field tests have verified that the false alarm rate of this invention is reduced by more than 60% compared to existing technologies, significantly improving the effectiveness and guidance of warning information and greatly reducing the workload of operators. Therefore, this invention pioneers a dynamic adjustable capacity boundary assessment mechanism, significantly reducing the false alarm rate. This invention not only achieves precise hierarchical early warning but also generates optimal pre-control strategies through deep reinforcement learning algorithms. These strategies can be directly linked to the AGC / AVC system for execution, while the execution results are fed back to the model for iterative optimization, forming a complete closed-loop management system. Field tests have verified that after adopting this system, the dispatch deviation rate of wind farm clusters has decreased by more than 50%, significantly reducing penalties under the "two detailed rules" for power grid assessment and improving the regional power grid's power balance capability. Therefore, this invention constructs a closed-loop early warning and response system, significantly reducing the dispatch deviation rate. This invention fully considers the spatiotemporal coupling characteristics of wind farm clusters, completes the pre-correction of wake effect during the data fusion stage, and incorporates grid section constraints and multi-site collaborative constraints in the adjustable boundary assessment. It can realize centralized early warning and collaborative control of more than 10 wind farms at the regional level, fully adapts to the development trend of centralized development of large-scale wind power clusters, and provides core technical support for the dispatch of new energy in large power grids. Therefore, this invention is adapted to the centralized dispatching needs of wind farm clusters and realizes collaborative management and control of multiple sites. The online self-iterative unit of the model in this invention can perform incremental training based on daily operating data and trigger emergency iteration under extreme conditions to ensure that the model continuously adapts to the aging of wind farm equipment, changes in meteorological conditions, etc., avoiding the accuracy decay of the model after long-term operation, and greatly improving the system's service life and long-term operational stability. Therefore, this invention has online self-iterative capability and high long-term operational stability. Attached Figure Description

[0011] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0012] In the attached diagram: Figure 1 This is a block diagram of the overall architecture of the AI-based wind farm cluster centralized power and scheduling deviation early warning system of the present invention. Figure 2 This is a core logic block diagram of the wind farm group power prediction and scheduling deviation graded early warning system of the present invention; Figure 3 This is a flowchart of the scheduling deviation closed-loop pre-control and online self-iteration of the model in this invention. Detailed Implementation

[0013] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0014] This embodiment is applied to a wind farm cluster under the jurisdiction of a regional power grid. The wind farm cluster includes 5 centrally connected wind farms with a total installed capacity of 1000MW, with each wind farm having an installed capacity of 200MW and equipped with 33 2MW doubly-fed wind turbine generators. All wind farms are connected to the same regional substation via 220kV lines. The power grid dispatching agency issues a daily dispatch plan for the next day at 96 points every day and issues rolling updated real-time dispatch instructions every 15 minutes. The system of this invention is used to realize centralized power prediction, dispatch deviation graded early warning and closed-loop pre-control and handling.

[0015] 1. Multi-source heterogeneous data fusion unit This unit is the data foundation of the system, responsible for collecting multi-source data from all dimensions of the wind farm cluster, completing preprocessing, spatiotemporal alignment and wake effect pre-correction, and outputting a standardized fusion dataset.

[0016] (1) Data collection The collected multi-source data includes: ① Meteorological data: NWP data from the European Centre for Medium-Range Weather Forecasts (ECMWF) (temporal resolution 1h, spatial resolution 1km, including wind speed, wind direction, temperature, air pressure, humidity, and turbulence intensity); each wind farm is equipped with 2 wind measurement towers to collect real-time wind speed, wind direction, and temperature data at heights of 10m-140m, with a sampling frequency of 1Hz; ② Wind turbine operating data: SCADA data for each wind turbine, including impeller speed, pitch angle, generator speed, active / reactive power, gearbox temperature, generator temperature, pitch / yaw system status, and fault alarm information, with a sampling frequency of 1Hz; ③ Equipment health status data: CMS status monitoring data of wind turbines, including gearbox vibration, generator vibration, bearing temperature, etc., are used to assess the health status and maximum allowable output of a single wind turbine; ④ Power Grid and Dispatch Data: Daily dispatch plans issued by the power grid, 15-minute rolling dispatch instructions, maximum power flow limits of 220kV grid-connected sections, voltage upper and lower limits, and operating status of reactive power compensation devices; ⑤ Site control data: Real-time setpoints, execution status, and upper and lower limit constraints of the AGC / AVC system for each wind farm.

[0017] (2) Data preprocessing For multi-source heterogeneous data, perform the following preprocessing: ①Outlier removal: The 3σ criterion and box plot method are combined to remove abnormal data such as wind speed exceeding the fan cut-in / cut-out range (3m / s-25m / s), negative power, and sensor fault jumps; ② Missing value completion: Short-term missing data (≤5 minutes) are completed using linear interpolation; long-term missing data (>5 minutes) are completed using the Random Forest (RF) algorithm combined with data from adjacent measurement points. ③ Data normalization: The min-max normalization method is used to normalize all input data to the [0,1] interval, eliminating the influence of different units on the model. The normalization formula is: =(X- ) / ( - ) Where x is the original data, , These are the historical minimum and maximum values ​​of this feature.

[0018] (3) Spatiotemporal alignment and wake effect pre-correction ① Spatiotemporal alignment: All data are unified to Beijing time, with a time resolution of 5 minutes and a spatial reference of WGS84 coordinate system, to achieve accurate spatiotemporal alignment of multi-source data; ② Wake effect pre-correction: Based on the coordinates, hub height, and impeller diameter of all fans, combined with the incoming wind speed and direction of NWP, the Jensen wake model is used to calculate the wake speed attenuation of a single fan. The superimposed wake of multiple fans is corrected by the LES correction coefficient of large eddy simulation to obtain the actual inflow wind speed of each fan.

[0019] The Jensen wake model formula for calculating wind speed attenuation is: = ·(1-(1- )·(D / (D+2kx)) 2 ) in, The wind speed is the result of the wake effect. For the incoming wind speed, denoted as the wind turbine thrust coefficient, D as the impeller diameter, k as the wake attenuation coefficient, and x as the axial distance between the upstream and downstream wind turbines.

[0020] The wake of multiple superimposed fans is calculated using the energy conservation method. In this embodiment, the LES correction coefficient is set to 0.85-1.15 (dynamically adjusted according to the turbulence intensity). The corrected inflow velocity is input to the subsequent power prediction and adjustable boundary assessment module to eliminate the impact of wake effect on accuracy.

[0021] 2. Physical Information Guided Dual-Scale AI Power Prediction Unit This unit communicates with the multi-source heterogeneous data fusion unit, constructs a dual-scale power prediction model with embedded physical constraints based on the fused dataset, and outputs the predicted power output sequence of wind farm clusters, providing an accurate basis for deviation early warning.

[0022] (1) Dual-scale prediction architecture ① Short-term power prediction subunit: Outputs a power output prediction sequence for the next 24-72 hours with a step size of 15 minutes, matching the time scale of the daily power grid dispatch plan, and supporting short-term deviation early warning; ② Ultra-short-term power prediction sub-unit: Outputs a power output prediction sequence for the next 15 minutes to 4 hours with a step size of 5 minutes, matching the real-time rolling dispatch time scale of the power grid, and supporting real-time early warning and pre-control.

[0023] (2) Transformer-PINN model architecture Both short-term and ultra-short-term prediction models adopt the Transformer-PINN architecture guided by physical information. The basic network uses a Transformer encoder to capture the long-term time-series dependency features of wind power sequences, and embeds physical laws as regularization terms into the loss function through physical constraint embedding sub-units.

[0024] The model input features include 28 dimensions such as corrected inflow wind speed, wind direction, air temperature, turbulence intensity, pitch angle, impeller speed, historical active power, dispatch instructions, and cross-sectional constraints. The model output is the predicted active power of a single wind turbine at the corresponding time step, which is then superimposed to obtain the predicted sequence of total output of the wind farm group.

[0025] (3) Embedding of physical constraints and design of loss function The total loss function is: L = λ1· +λ2· In this embodiment, λ1 is 0.7 and λ2 is 0.3; For data loss, This is a physical loss.

[0026] ① Data loss The mean square error between predicted power and actual power is calculated using the following formula: =(1 / N)·Σ( ,i- ,i) 2 Where N is the sample size. , where i is the predicted power of the i-th sample. , i represents the actual power.

[0027] ②Physical loss Losses due to aerodynamic power constraints of the fan Wake constraint loss Power balance constraint losses The composition, and the formula are: =ω1· +ω2· +ω3· In this embodiment, ω1=0.5, ω2=0.3, and ω3=0.2.

[0028] The constrained predicted power conforms to the wind turbine aerodynamic power equation. =0.5·ρ·S·Cp(λ,β)·v 3 To avoid predictive results that do not conform to physical logic; : Constrain the predicted output of downstream wind turbines to conform to the wake influence of upstream wind turbines, and avoid contradictions in output prediction; The total predicted output is constrained to conform to the power flow constraints of the grid section, thus avoiding invalid predictions that exceed the grid constraints.

[0029] (4) Model training and rolling updates The model adopts a pre-training + incremental training mode: the pre-training uses the historical data of wind farm clusters for the past 3 years, batch_size=64, learning rate 1e-4, Adam optimizer, epoch=200, and early stopping mechanism to prevent overfitting. Rolling updates after deployment: The ultra-short-term model optimizes parameters every 15 minutes based on the latest 1-hour data; the short-term model is incrementally trained every 24 hours based on the complete data from the previous day, ensuring that the prediction accuracy remains stable.

[0030] 3. Assessment Unit for the Dynamic Adjustable Capacity Boundary of Wind Farm Clusters This unit communicates with the multi-source heterogeneous data fusion unit and calculates the adjustable upper and lower limits of the wind farm group's output under the current operating conditions in real time based on the fused dataset. It outputs the dynamic adjustable capability boundary and is the core module for distinguishing between controllable and uncontrollable deviations.

[0031] (1) Optimization objective ① Adjustable upper limit optimization target: max =Σ( ), i=1,2,...,n (n=165 wind turbines) ② Adjustable lower limit optimization objective: min =Σ( ), i=1,2,...,n in, Contribute to the overall power of the wind farm cluster. The active power output of the i-th wind turbine.

[0032] (2) Constraints ① Health constraints for wind turbine equipment: ≤ , Based on the evaluation of wind turbine CMS data and fault information, the rated power is 2MW when there is no fault, and the rated power is reduced according to the level when there is a defect (such as reducing the rated power to 60% when the gearbox overheats). ② Wind resource output constraints: ≤ , The maximum aerodynamic output at the current inflow wind speed is given by the formula: =0.5·ρ·S· · In this embodiment Take 0.48; ③ Power flow constraints at power grid sections: ≤ ≤ , This represents the maximum allowable input power of the cross-section. This represents the minimum power absorption capacity of the power grid. ④ Wind farm cluster wake coupling constraint: The output coupling relationship of upstream and downstream wind turbines is incorporated into the wake correction model to ensure that the solution results conform to physical laws; ⑤ Constraints for stable operation of the wind turbine: ≥ (In this embodiment, it is 0.2MW, 10% of the rated power), while the output change rate is ≤20% of the rated power per minute to avoid equipment damage and grid fluctuations.

[0033] (3) Solution algorithm and rolling update The interior point method is used to solve the above nonlinear programming problem, and real-time adjustable upper and lower limits are obtained. The adjustable boundary is updated every 5 minutes, and all constraint parameters are updated synchronously to ensure the accuracy of the boundary.

[0034] 4. Dispatch Deviation Graded Early Warning Unit This unit communicates with the power prediction unit and the adjustable boundary assessment unit to obtain the scheduling instruction sequence, calculates the deviation by combining the predicted output and the adjustable boundary, determines the type and level of the deviation, and outputs a graded early warning signal.

[0035] (1) Calculation of scheduling deviation Calculate the absolute and relative deviations step by step: Absolute deviation ΔP= - Relative deviation δ=ΔP / 100% in, To predict total output, This is the value of the scheduling instruction.

[0036] (2) Determination of Deviation Type Based on adjustable boundaries, they are divided into two categories: ① Uncontrollable deviation: The predicted output exceeds the adjustable boundary and cannot be eliminated through on-site adjustments. This is a core risk that must be reported to the dispatch center. ② Controllable deviation: If the predicted output is within the adjustable boundary, it can be eliminated through pre-control strategies and is a risk that can be handled on-site.

[0037] (3) Tiered early warning triggering logic Based on the type, magnitude, and duration of the deviation, it is divided into four levels, and a rolling assessment is performed every 5 minutes: ① Level 1 Red Emergency Warning: The predicted output exceeds the adjustable boundary and the relative deviation exceeds ±20% for more than 30 minutes. This is the highest level warning and requires immediate reporting to the dispatch center and activation of emergency control measures. ② Level 2 Orange Important Warning: If the predicted output is within the adjustable boundary and the relative deviation exceeds ±10% for more than 15 minutes, the pre-control strategy must be activated immediately to prevent the deviation from escalating. ③ Level 3 Yellow General Warning: The relative deviation of the predicted output exceeds ±5% and lasts for more than 10 minutes. Real-time tracking and pre-control preparations are required. ④ Level 4 Blue Normal Status: Relative deviation within ±5%, operating normally, continuous monitoring.

[0038] 5. Closed-loop pre-control strategy generation and linkage execution unit This unit communicates with the hierarchical early warning unit. Based on the early warning signal, with the core objective of minimizing scheduling deviation, it generates the optimal pre-control strategy by combining adjustable boundaries, and links the AGC / AVC system to execute it, thus realizing a closed loop of early warning and response.

[0039] This unit uses the Deep Reinforcement Learning (DQN) algorithm to construct a policy generation model, as detailed below: (1) Design of DQN model ①State space S: 24-dimensional features, including predicted output, scheduling instructions, adjustable upper and lower limits, wind turbine operating status, inflow wind speed, cross-sectional constraints, etc., comprehensively characterizing the system operating status; ②Action Space A: 32 discrete actions, including AGC setting adjustment for each wind farm, pitch angle / speed adjustment for a single wind turbine, start-up and shutdown of standby units, switching on and off of reactive power compensation devices, etc., covering all power output adjustment methods; ③ Reward function R: Multi-objective reward function, the formula is: R = -α1·|ΔP|-α2· +α3·ΔE In this embodiment, α1=0.6, α2=0.2, α3=0.2; ΔP is the absolute value of the scheduling deviation. Let ΔE represent the change in wind turbine fatigue load, and ΔE represent the change in power generation. The reward logic is: the smaller the deviation, the smaller the load, and the higher the power generation, the higher the reward, guiding the model to generate the optimal strategy.

[0040] (2) Model training and policy generation The DQN model is trained using an experience replay and target network mechanism. Pre-training uses historical running data from the past two years and typical deviation scenario data, with 1 million iterations, a γ discount factor of 0.95, and a learning rate of 1e-4. Upon receiving an early warning signal, the model outputs the optimal combination of actions based on the current state: For Level 2 and Level 3 early warnings: Automatically generate AGC setpoint optimization schemes, adjust the wind turbine pitch angle and speed, coordinate the power distribution of the station, and adjust the total output to near the dispatch command within the adjustable boundary to eliminate deviations; For Level 1 warnings: generate emergency pre-control strategies, including starting and stopping standby units and adjusting the voltage of reactive power units, and generate suggestions for revising dispatch plans to assist operators in communicating with dispatching agencies.

[0041] (3) Joint execution and effect feedback The pre-control strategy is automatically executed by sending it to the monitoring system and AGC / AVC system of each wind farm through a dedicated power communication channel. At the same time, the actual output data after execution is collected to track the effect of deviation elimination. If the expected results are not achieved, the strategy is continuously optimized. The execution effect data is fed back to the model self-iteration unit for model optimization.

[0042] 6. Online self-iteration and optimization unit of the model This unit communicates with the power prediction unit and the pre-control strategy unit respectively. Based on actual operating data and strategy execution effects, it performs incremental training and online optimization of the model to ensure long-term operating accuracy.

[0043] (1) Iterative subunit of the prediction model Daily iteration: Every 24 hours, based on the complete running data of the previous day, incrementally train the short-term and ultra-short-term prediction models, update parameters, and adapt to situations such as equipment aging and changes in meteorological patterns; Emergency Iteration: When extreme conditions are detected (sudden wind speed changes exceeding 5 m / s, typhoons, icing, etc.), or when the relative deviation of the prediction exceeds 30% for more than 15 minutes, emergency incremental training is immediately triggered to quickly improve the prediction accuracy of extreme conditions.

[0044] (2) Iterative subunit of strategy model Based on the performance of the pre-control strategy, the reward function and network parameters of the DQN model are optimized. If the strategy performs well, the reward weight of the corresponding action is increased, and vice versa. Full training is performed every quarter based on all historical execution data to improve the adaptability to complex scenarios.

[0045] 7. Human-Computer Interaction and Scheduling Communication Unit This unit communicates with the hierarchical early warning unit and the pre-control strategy unit to achieve two-way communication for operation monitoring and scheduling: ① Centralized monitoring interface: Real-time display of core data such as total output of wind farm cluster, output of each station, forecast sequence, dispatch instructions, adjustable boundaries, and early warning status, and supports status query of individual wind turbines; ② Early Warning Management Interface: Displays early warning level, trigger time, deviation value, cause analysis, and handling suggestions, and supports early warning confirmation, handling, and closed-loop management; ③ Strategy Management Interface: Displays the generated pre-control strategies, supports automatic / manual execution modes, and operators can modify and confirm them before issuing them; ④ Dispatch and communication function: Through bidirectional communication with the power grid EMS system via the power dispatch data network, it automatically reports early warning information, pre-control schemes, and operating data, and receives rolling updated dispatch instructions and cross-section constraints.

[0046] 8. System Hardware Architecture This system adopts a cloud-edge collaborative hardware architecture and complies with the security protection specifications for power secondary systems. ① Edge acquisition layer: Deployed at various wind farm sites, including wind turbine PLC, SCADA system, meteorological tower acquisition device, AGC / AVC system, responsible for on-site data acquisition and pre-control strategy execution; ② Communication layer: Power-dedicated optical fiber + 5G redundant communication is adopted to realize bidirectional data transmission between the edge layer and the platform layer; ③ Platform layer: Deployed in the regional new energy control center, including data server, AI computing server, scheduling and communication server, and redundant backup server, responsible for core functions such as data storage, model calculation, early warning analysis, and strategy generation; ④ Application layer: Deployed on the human-machine interface terminal of the central control center, running the monitoring interface for operation and management by operators.

[0047] Based on field measurements, this embodiment achieved a power prediction accuracy of 96.2% under normal operating conditions and a 32.5% improvement in prediction accuracy under extreme operating conditions compared to traditional models. The false alarm rate was reduced by 64.8%, and the missed alarm rate was reduced by 41.2%. The wind farm cluster dispatch deviation rate was reduced by 53.7%, significantly reducing grid assessment penalties and significantly improving the regional grid's renewable energy absorption capacity and safe and stable operation level.

Claims

1. An AI-based early warning system for centralized power and dispatch deviation in wind farm clusters, characterized in that, include: Multi-source heterogeneous data fusion unit, physical information-guided dual-scale AI power prediction unit, wind farm group dynamic adjustable capability boundary assessment unit, scheduling deviation classification early warning unit, closed-loop pre-control strategy generation and linkage execution unit, and model online self-iteration and optimization unit. The multi-source heterogeneous data fusion unit is used to collect multi-source operation data of wind farm clusters, complete data preprocessing, spatiotemporal alignment and pre-correction of wind farm cluster wake effect, and output a standardized fusion dataset. The physical information-guided dual-scale AI power prediction unit is communicatively connected to the multi-source heterogeneous data fusion unit. It is used to construct a short-term and ultra-short-term dual-scale wind farm power prediction model with embedded physical constraints based on the fused dataset, and output the predicted power output sequence of the wind farm group. The wind farm cluster dynamic adjustable capability boundary assessment unit is communicatively connected to the multi-source heterogeneous data fusion unit. It is used to calculate the upper and lower adjustable limits of the wind farm cluster under the current operating conditions in real time based on the fused dataset, and output the dynamic adjustable capability boundary. The scheduling deviation classification and early warning unit is communicatively connected to the physical information-guided dual-scale AI power prediction unit and the wind farm group dynamic adjustable capacity boundary assessment unit, respectively. It is used to obtain the power grid scheduling instruction sequence, combine the predicted power output sequence and the dynamic adjustable capacity boundary, calculate the scheduling deviation value, determine the deviation type and level, and output the corresponding classification and early warning signal. The closed-loop pre-control strategy generation and linkage execution unit is communicatively connected to the scheduling deviation hierarchical early warning unit. It is used to generate the optimal pre-control strategy based on the hierarchical early warning signal, with the core objective of minimizing scheduling deviation, combined with the dynamic adjustable capability boundary, and to link the AGC / AVC system of the wind farm group for execution. The model online self-iteration and optimization unit is communicatively connected to the physical information-guided dual-scale AI power prediction unit and the closed-loop pre-control strategy generation and linkage execution unit, respectively, and is used to perform incremental training and online optimization of the power prediction model and the pre-control strategy model based on actual operating data and the execution effect of the pre-control strategy.

2. The AI-based early warning system for centralized power and dispatch deviation in wind farm clusters according to claim 1, characterized in that: The multi-source heterogeneous data fusion unit collects multi-source operational data including numerical weather prediction (NWP) data, real-time meteorological data from wind farm group anemometers, SCADA operational data for each wind turbine, equipment health status data, power grid dispatch command data, power grid section constraint data, and AGC / AVC operational data for the wind farm. The data preprocessing includes outlier removal, missing value completion, and data normalization. The aforementioned spatiotemporal alignment and wake effect pre-correction are used to unify multi-source data to the same spatiotemporal reference, and calculate the actual inflow wind speed of each wind turbine in the wind farm group based on the Jensen wake model combined with large eddy simulation correction, thus completing the pre-correction of the wake effect.

3. The AI-based early warning system for centralized power and dispatch deviation in wind farm clusters according to claim 1, characterized in that: The physical information-guided dual-scale AI power prediction unit includes a short-term power prediction subunit, an ultra-short-term power prediction subunit, and a physical constraint embedding subunit. The short-term power prediction subunit outputs a wind farm cluster power output prediction sequence for the next 24-72 hours with a step size of 15 minutes. The ultra-short-term power prediction subunit outputs a wind farm cluster power output prediction sequence for the next 15 minutes to 4 hours with a step size of 5 minutes. The physical constraint embedding subunit uses the wind turbine aerodynamic power equation, wind farm cluster wake constraints, and grid power balance constraints as regularization terms, embedding them into the loss function of the neural network to construct a Physical Information Neural Network (PINN) model. The loss function includes data loss and physical loss, with the total loss function being L = λ1· +λ2· Where λ1 and λ2 are weighting coefficients, To determine the mean square error between predicted power and actual power, This refers to the deviation loss due to physical constraints.

4. The AI-based early warning system for centralized power and dispatch deviation in wind farm clusters according to claim 1, characterized in that: The wind farm cluster dynamic adjustable capability boundary assessment unit is used to obtain the real-time adjustable upper and lower limits of the wind farm cluster based on the fused dataset, with the optimization objective being the maximum / minimum total adjustable output of the wind farm cluster, and with constraints such as wind turbine health constraints, wind resource output constraints, power grid cross-section flow constraints, wind farm cluster wake coupling constraints, and wind turbine stable operation constraints. The adjustable boundary is updated every 5 minutes.

5. The AI-based early warning system for centralized power and dispatch deviation in wind farm clusters according to claim 1, characterized in that: The dispatch deviation classification and early warning unit is used to calculate the deviation between the predicted output sequence and the power grid dispatch command sequence step by step. Combined with the dynamic adjustable capacity boundary, the deviation is divided into uncontrollable deviation and controllable deviation. Among them, the predicted output exceeding the adjustable boundary is uncontrollable deviation, and the predicted output within the adjustable boundary is controllable deviation. Based on the deviation amplitude and duration, the early warning is divided into four levels: Level 1 Red Emergency Warning, Level 2 Orange Important Warning, Level 3 Yellow General Warning, and Level 4 Blue Normal State.

6. The AI-based early warning system for centralized power and dispatch deviation in wind farm clusters according to claim 5, characterized in that: The triggering conditions for the Level 1 Red Emergency Warning are: the predicted output exceeds the adjustable boundary and the relative deviation from the dispatch command exceeds ±20%, lasting for more than 30 minutes; the triggering conditions for the Level 2 Orange Important Warning are: the predicted output is within the adjustable boundary and the relative deviation from the dispatch command exceeds ±10%, lasting for more than 15 minutes; the triggering conditions for the Level 3 Yellow General Warning are: the relative deviation between the predicted output and the dispatch command exceeds ±5%, lasting for more than 10 minutes; the Level 4 Blue Normal State is defined as a deviation within ±5%.

7. The AI-based early warning system for centralized power and dispatch deviation in wind farm clusters according to claim 1, characterized in that: The closed-loop pre-control strategy generation and linkage execution unit uses the deep reinforcement learning (DQN) algorithm to construct a pre-control strategy generation model. It takes minimizing scheduling deviation, minimizing wind turbine fatigue load, and maximizing the power generation of the wind farm cluster as multiple optimization objectives, and uses the dynamic adjustable capability boundary as a constraint condition to output the optimal pre-control strategy. The pre-control strategy includes adjusting the AGC setpoints of each wind farm, optimizing the wind turbine pitch angle and speed, starting and stopping standby units, and coordinating the control of reactive power compensation devices. The pre-control strategy is then distributed to the monitoring systems of each wind farm for execution via a dedicated power communication channel, while simultaneously collecting actual operational data after execution.

8. The AI-based early warning system for centralized power and dispatch deviation in wind farm clusters according to claim 1, characterized in that: The online self-iteration and optimization unit of the model includes a prediction model iteration subunit and a strategy model iteration subunit; the prediction model iteration subunit is used to incrementally train the dual-scale power prediction model every 24 hours based on the actual operating data of the previous day. The policy model iteration subunit is used to optimize and update the reward function of the deep reinforcement learning model based on the execution effect of the pre-control policy; at the same time, when extreme working conditions or prediction deviations exceed the threshold are detected, emergency incremental training is triggered to update the model parameters in real time.

9. The AI-based early warning system for centralized power and dispatch deviation in wind farm clusters according to claim 1, characterized in that, It also includes a human-machine interaction and scheduling communication unit, which is connected to the scheduling deviation classification and early warning unit and the closed-loop pre-control strategy generation and linkage execution unit, respectively. It is used to display the operating status of the wind farm group, early warning information, and pre-control strategies in real time, and at the same time realize two-way communication with the power grid dispatching agency, report early warning information and pre-control schemes, and receive rolling updated dispatch instructions.

10. The AI-based early warning system for centralized power and dispatch deviation in wind farm clusters according to claim 2, characterized in that: The wake effect pre-correction is specifically as follows: based on the coordinate position, hub height, and rotor diameter of each wind turbine in the wind farm group, combined with the incoming wind speed and direction of the NWP, the Jensen wake model is used to calculate the wake wind speed attenuation of a single wind turbine, and then the wake effect of multiple wind turbines is corrected by the correction coefficient of the Large Eddy Simulation (LES) to obtain the actual inflow wind speed of each wind turbine in the wind farm group, which is then input into the power prediction model and the adjustable capacity boundary assessment unit.

Images