An artificial intelligence-based wind power processing prediction regulation method and system

Abstract

The application relates to the field of wind power grid connection regulation technology, and discloses a wind power processing prediction regulation method and system based on artificial intelligence; the method and system first acquire unit operation data, field area meteorological data, wide-area meteorological data, dispatching data and historical sample data, perform space-time alignment, field area correction and multi-scale feature layering on the above data, generate wind power probability prediction information covering multiple prediction scales, construct active regulation constraints according to quantiles, time-varying covariance and tail scenarios, determine backup capacity, and generate field-level active instructions and single-machine execution instructions; and the prediction model and regulation parameters are updated according to the actual output deviation backflow after regulation execution, so that the wind power prediction accuracy, backup configuration rationality and grid operation safety are improved.

Description

Technical Field

[0001] This invention relates to the field of wind power grid connection control technology, specifically to a wind power processing prediction and control method and system based on artificial intelligence. Background Technology

[0002] With the continuous increase in the proportion of large-scale wind power grid connection, wind power output is characterized by randomness, fluctuation, and intermittency, which can easily affect the power balance, reserve configuration, and frequency stability of the power system. In the prior art, the invention patent with authorization announcement number CN112990553B discloses a wind power ultra-short-term power prediction method using a self-attention mechanism and bilinear fusion. It extracts time-series features through numerical weather prediction data processing branch and historical power data processing branch, and outputs ultra-short-term wind power prediction results after information fusion. The invention patent with authorization announcement number CN102606395B discloses a wind farm active power optimization control method based on power prediction information. Based on the ultra-short-term wind power prediction results and wind farm operating constraints, it optimizes the allocation of active power output of each unit. While the aforementioned technologies can improve the accuracy of wind power forecasting or enhance the active power output allocation of wind farms, the forecasting and active power regulation processes still primarily operate sequentially, with forecasting preceding regulation. Forecast results are typically transmitted to the regulation stage as point values, trend values, or fixed confidence intervals. Uncertainties such as the probability distribution pattern, time-varying correlation, and tail risk of the forecasts are difficult to fully translate into regulatory constraints. This makes it challenging to quantify risk control based on forecast uncertainty, leading to a conservative allocation of spinning reserves. Consequently, reserve redundancy, increased operating costs, and higher wind curtailment rates result. Furthermore, the actual power output deviation after the execution of regulation commands lacks a clear direction for forecasting. The closed-loop feedback channel of the model backflow makes it difficult for the prediction model to perceive the impact of prediction errors on the control process in a timely manner. Under scenarios such as power output ramp-up, extreme weather, and out-of-training sample distribution, the prediction distribution is prone to cumulative deviation from the actual power output distribution, which further leads to an increase in the automatic generation control regulation mileage and an increased risk of system frequency and tie-line power exceeding limits. Therefore, there is an urgent need for a wind power prediction and control method that can quantitatively map prediction uncertainty information into control constraints and feed back actual power output deviations to the prediction model, so as to achieve a synergistic improvement in prediction accuracy and control efficiency, reduce spinning reserve occupancy rate and wind curtailment rate, and improve the safety and economy of wind power grid-connected operation. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a wind power processing prediction and control method and system based on artificial intelligence, which solves the technical problems mentioned in the background section.

[0004] To achieve the above objectives, the present invention provides the following technical solution: An artificial intelligence-based wind power forecasting and control method includes: S1. Acquire wind farm unit operation data, field meteorological observation data, wide-area meteorological data, superior dispatch data and historical sample data, perform time-series alignment of various types of data, and establish a unified field spatial coordinate mapping relationship; S2. Perform field-specific correction on wide-area meteorological data, and perform multi-time-scale feature layering processing on the corrected meteorological data and historical power data. S3. Based on the feature data after hierarchical processing, generate wind power probability prediction information covering day-ahead, intraday, ultra-short-term and real-time prediction scales; S4. Map the wind power probability prediction information to active power control constraints, determine the spinning reserve capacity based on the active power control constraints, and generate field-level active power commands and single-unit execution commands. S5. Collect the actual output deviation after the control is executed, feed the actual output deviation back to the corresponding prediction model according to the time scale corresponding to the deviation, and update the subsequent prediction and control process according to the deviation amplitude and deviation distribution.

[0005] Preferably, S1 includes: During the data acquisition process, rolling tasks are triggered according to the forecast scale, and the data is connected to the unit monitoring system, the wind measurement equipment in the field, the wide-area meteorological interface, the dispatch interface and the historical database. The data from different frequencies is timed, resampled, graded for data quality, and anomaly-filled. Based on the equipment space ledger, wide-area meteorological grid points are mapped to representative points in the field area. An aggregated index is established based on the power collection unit, wind direction sector, and unit health status to generate a unified spatiotemporal dataset with timestamps, spatial coordinates, equipment numbers, data quality identifiers, prediction scale identifiers, and data version identifiers.

[0006] Preferably, S2 includes: The availability of meteorological data, power data, and scheduling data is verified, and then the wide-area meteorological data is corrected based on topography, stability, and wind direction residuals to generate hub height wind speed sequence and hub height wind direction sequence. When performing feature layering, a basic feature matrix is ​​constructed using field-level power and corrected meteorological sequences. After modal decomposition, it is divided into trend layer, daily and semi-daily cycle layer, hourly fluctuation layer, short-term disturbance layer, abrupt change layer, and noise layer. The decomposition window and feature weights are adjusted according to the rapid power change state and the state of missing wide-area meteorological data.

[0007] Preferably, S3 includes: Input quality verification is performed on the multi-scale feature matrix; A multi-scale shared feature layer and a scale-segmented output layer are used to call the prediction model corresponding to the prediction scale. Quantile regression is used to generate quantile sequences and predict the mean. The time-varying covariance matrix and tail scene set are generated based on the predicted residuals; The generated results are subjected to cross-scale consistency verification, re-fusion, hill-climbing event identification, and low-confidence correction in the out-of-distribution state of the training samples.

[0008] Preferably, a multi-scale shared feature layer and a scale-specific output layer are used to call the prediction model corresponding to the prediction scale, including: The features of the trend layer, daily and semi-daily cycle layers, hourly fluctuation layer, short-term disturbance layer and abrupt change layer are encoded according to the time scale. Historical sample tags, unit availability status, power curtailment status, and field area spatial index are fused on the field-level time axis; Matching features are selected according to the prediction scale and sent to the day-ahead, intraday, ultra-short-term, and real-time prediction output channels.

[0009] Preferably, S4 includes: Opportunity constraints, tail risk boundaries, and correlation robustness constraints are constructed based on quantile sequences, tail scenario sets, and time-varying covariance matrices. Configure standby capacity according to response scale and the probability of ramp events; The field-level active power command is generated with the scheduling plan, tie line limits, reserve capacity, and achievable output as boundaries. Decompose the single-machine execution instructions according to the single-machine availability, health status, fatigue life and wake effect; The instruction correction branch is determined by combining the reachable upper limit check, further decomposition, and communication anomaly status.

[0010] Preferably, the single-machine execution instructions are decomposed according to the single-machine availability, health status, fatigue life, and wake effect, including: The usable capacity of a single unit is determined by the nameplate capacity, real-time wind speed, and power curve. Adjust the available capacity of a single unit based on the unit's alarm status, maintenance status, fault shutdown status, and grid connection status; The proportion of the instruction is limited based on the remaining fatigue life; the wake influence coefficient is calculated based on the field coordinates and the direction of the incoming wind; Normalized allocation weights are constructed using the available capacity of a single unit, the remaining fatigue life, and the wake influence coefficient, and field-level active power commands are allocated into active power commands, pitch coordination commands, and yaw coordination commands.

[0011] Preferably, S5 includes: The actual output deviation is divided into prediction deviation and command execution deviation; Deviation data is collected in seconds, minutes, and long time intervals; Incremental updates to partitions are triggered based on quantile out-of-bounds, normalized root mean square error, relative entropy, and bias moving average. The prediction bias is fed back to the prediction model parameters, and the command execution bias is fed back to the available capacity assessment, single-machine weight allocation and control constraint correction parameters. Sample labeling, training weight adjustment, and downgrade fusion processing are performed on retraining trigger states, extreme weather, out-of-distribution states, and data credibility states.

[0012] Preferably, the prediction bias is fed back to the prediction model parameters, and the command execution bias is fed back to the available capacity assessment, single-machine weight allocation, and control constraint correction parameters, including: Adjust the output bias parameters of the real-time prediction model, the quantile calibration parameters of the ultra-short-term prediction model, and the mid-frequency characteristic parameters of the intraday and day-ahead prediction models according to the deviation time scale. Determine the parameter update step size based on the deviation amplitude; Determine the quantile calibration direction based on the direction of quantile crossover; Based on the deviation of the single-unit instruction, the available capacity of the unit is updated and the weight of the single unit is allocated. The active power constraint parameters of the field level and the control parameters of the next cycle are also adjusted in conjunction.

[0013] On the other hand, the present invention provides a wind power processing prediction and control system based on artificial intelligence, comprising: Data aggregation and spatiotemporal mapping module: used to acquire wind farm unit operation data, field meteorological observation data, wide-area meteorological data, superior dispatch data and historical sample data, perform time-series alignment of various types of data, and establish a unified field spatial coordinate mapping relationship; Meteorological correction and feature layering module: It is used to perform field-based correction on wide-area meteorological data based on a unified field area spatial coordinate mapping relationship, and to perform multi-time-scale feature layering processing on the corrected meteorological data and historical power data. Probabilistic prediction generation module: used to generate wind power probabilistic prediction information covering day-ahead, intraday, ultra-short-term and real-time prediction scales based on the hierarchically processed feature data; Regulation constraint mapping module: used to map wind power probability prediction information into active power regulation constraints, and determine reserve capacity based on active power regulation constraints; Instruction generation and decomposition module: used to generate field-level active power instructions based on active power control constraints, reserve capacity, upper-level scheduling plan and unit operating status, and decompose field-level active power instructions into single-unit execution instructions; Deviation closed-loop feedback module: It is used to collect the actual output deviation after the control is executed, feed the actual output deviation back to the corresponding prediction model and control parameters according to the time scale corresponding to the actual output deviation, and update the subsequent prediction and control process according to the deviation amplitude and deviation distribution.

[0014] Compared with existing technologies, this invention provides a wind power processing prediction and control method and system based on artificial intelligence, which has the following beneficial effects: 1. This invention improves data consistency and feature adaptability across different prediction scales by uniformly processing wind farm unit operation data, field meteorological observation data, wide-area meteorological data, dispatch data, and historical sample data, and by performing field-specific correction and multi-time-scale feature layering on the wide-area meteorological data. It also generates quantile sequences, time-varying covariance matrices, and tail scenario sets, transforming prediction distribution patterns, time-varying correlations, and tail risks into opportunity constraints, tail risk boundaries, and correlation robustness constraints. Based on these constraints, it configures reserve capacity, generates field-level active power commands and individual unit execution commands, enabling the control process to perform constraint verification and command decomposition based on prediction uncertainty. By feeding back prediction deviations and command execution deviations after control execution to the prediction model and control parameters according to the time scale, it reduces the risks of prediction-control disconnect, reserve capacity redundancy, prediction distribution drift in extreme scenarios, and unreachability of individual unit commands. This achieves a synergistic improvement in prediction accuracy, reserve configuration rationality, field-level active power control stability, and wind power grid-connected operation safety.

[0015] 2. This invention establishes a continuous and traceable data chain for wind farms by setting data quality identifiers, data version identifiers, abnormal status identifiers, and equipment space ledgers during the data access phase, and recording meteorological correction versions, feature-layered versions, prediction status identifiers, and command execution statuses during prediction, regulation, and deviation feedback processes. Furthermore, it records these data during prediction, regulation, and deviation feedback processes, enabling wind farms to form a continuous and traceable data chain from raw data acquisition, probability prediction generation, active power constraint conversion to single-unit command execution. By implementing degradation processing, sample labeling, weight adjustment, and local conservative control logic for time synchronization anomalies, wide-area meteorological gaps, sensor anomalies, communication anomalies, out-of-distribution training samples, and extreme weather conditions, the invention reduces the interference of abnormal samples on prediction models and regulation parameters, thereby improving the data traceability, accuracy of abnormal condition identification, reliability of model updates, and efficiency of anomaly handling in the wind power prediction and regulation process. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the wind power processing prediction and control method based on artificial intelligence according to the present invention. Figure 2 This is a schematic diagram of multi-source data access and unified spatiotemporal mapping according to the present invention; Figure 3 This is the wide-area meteorological field regionalization correction and multi-timescale feature layering map of the present invention; Figure 4 This is a structural diagram of the multi-scale probability prediction model of the present invention; Figure 5This is the mapping diagram from the probability prediction information to the active power regulation constraint of this invention; Figure 6 This is the field-level command decomposition and deviation closed-loop reflux diagram of the present invention; Figure 7 This is a module architecture diagram of a wind power processing prediction and control system based on artificial intelligence according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: Figure 1 - Figure 6 A wind power treatment prediction and control method based on artificial intelligence is presented, including: S1. Acquire wind farm unit operation data, field meteorological observation data, wide-area meteorological data, superior dispatch data and historical sample data, perform time-series alignment of various types of data, and establish a unified field spatial coordinate mapping relationship; S2. Perform field-specific correction on wide-area meteorological data, and perform multi-time-scale feature layering processing on the corrected meteorological data and historical power data. S3. Based on the feature data after hierarchical processing, generate wind power probability prediction information covering day-ahead, intraday, ultra-short-term and real-time prediction scales; S4. Map the wind power probability prediction information to active power control constraints, determine the spinning reserve capacity based on the active power control constraints, and generate field-level active power commands and single-unit execution commands. S5. Collect the actual output deviation after the control is executed, feed the actual output deviation back to the corresponding prediction model according to the time scale corresponding to the deviation, and update the subsequent prediction and control process according to the deviation amplitude and deviation distribution.

[0019] This method takes a centralized wind farm connected to a regional power grid as the application scenario. The centralized wind farm is equipped with a site-level data aggregation platform, a turbine monitoring system, a wind measurement tower, lidar, a numerical weather prediction interface, a regional dispatch data interface, and a historical sample database. These devices and interfaces work together to complete wind power prediction and active power regulation. The input data acquired by the site-level data aggregation platform includes at least the following fields: turbine operation data, site meteorological observation data, wide-area meteorological data, higher-level dispatch data, and historical sample data. These fields form a data foundation for prediction and regulation under a unified spatiotemporal reference. Based on the above data, the system generates wind power probability prediction information and includes the predicted distribution and tail. Uncertainty information such as risks, time-varying correlations, and ramp events is transformed into active power control constraints, reserve capacity configuration results, and single-unit execution basis. After processing, the system outputs field-level active power commands, single-unit execution commands, reserve capacity configuration results, prediction model update parameters, and degradation control commands under abnormal scenarios. The overall operation process sequentially completes data acquisition and spatiotemporal unification, wide-area meteorological field area correction and feature layering, wind power probability prediction, active power control constraint mapping and reserve configuration, field-level and single-unit command generation, and actual output deviation closed-loop feedback. The intermediate data generated in each link are transmitted step by step according to the time scale and the controlled object, so that the prediction process, control process, and execution feedback process form a closed-loop chain.

[0020] Specifically, such as Figure 2 As shown: When the station-level data aggregation platform receives the forecasting and control task start instruction issued by the superior dispatching system, or when the rolling trigger time of the corresponding forecast scale is reached, the data acquisition process is initiated. The day-ahead forecast is generated based on the day-ahead dispatching declaration cycle, and the meteorological input, forecast boundary, and reserve assessment parameters for the next 24 to 72 hours are corrected on an hourly rolling basis. This hourly rolling correction does not replace the already confirmed day-ahead power generation plan. The intraday forecast, ultra-short-term forecast, and real-time forecast are rolled on a 30-minute, 5-minute, and 30-second cycle, respectively, to adapt to the intraday plan verification, automatic power generation control connection, and station-level rapid active power correction requirements. The above cycle values ​​are based on the fact that the day-ahead forecast mainly serves the power generation plan and reserve assessment on a longer time scale, the intraday forecast needs to match the rolling plan verification, the ultra-short-term forecast needs to connect with minute-level power fluctuations, and the real-time forecast needs to meet the requirements of station-level rapid active power correction and second-level measurement update. The site-level data aggregation platform reads the unit's operating data from the monitoring and data acquisition devices of each wind turbine generator set. The acquisition cycle is 1 to 5 seconds, preferably 1 second. The unit's operating data includes at least the following fields: active power, reactive power, terminal voltage, terminal current, speed, pitch angle, yaw angle, generator winding temperature, converter temperature, main bearing temperature, gearbox oil temperature, tower vibration acceleration, fault word, and operating status word. The acquisition cycle is determined based on the fact that the active power regulation and automatic generation control of wind turbine generator sets have a second-level response requirement. A 1-second sampling can capture sudden changes in output and state switching, while a sampling within 5 seconds can meet the online update requirements and avoid excessive sampling causing continuous high load on the communication link. Meteorological observation data for the site is provided jointly by a wind measurement tower and a lidar system. The wind measurement tower collects wind speed, wind direction, temperature, humidity, air pressure, air density, and turbulence intensity at 10 meters, 30 meters, 50 meters, 70 meters, and at the hub height. The initial sampling period is 1 to 10 seconds, and 10-second statistical data is generated for prediction input. This data is consistent with the conventional sampling capability of the wind measurement tower for online monitoring and can retain short-term wind speed changes. The lidar performs sector scanning of the incoming wind speed and direction within a range of 300 to 1800 meters in front of the hub height, with a preferred scanning period of 5 seconds. The lower limit of 300 meters is used to avoid the near-wake of the turbine and... The impact of the near-range blind zone of lidar on wind identification is addressed by using an upper limit of 1800 meters to balance effective echo quality and short-term forecast lead. Based on common operating wind speeds of 5 to 12 meters per second at hub height, this forward-looking distance corresponds to approximately 25 seconds to 6 minutes of wind lead, primarily providing 1 to 4 minutes of effective wind forecast, serving as feedforward input for ultra-short-term and real-time forecasts. When dense fog, heavy precipitation, or dust storms cause a decrease in lidar echo quality, the data aggregation platform reduces the weight of that data channel based on the lidar quality indicator and supplements it with wind speed at the hub height of the wind measurement tower and wind trends from adjacent time periods. Wide-area meteorological data is provided by numerical weather prediction systems, satellite cloud imagery systems, weather radar systems, and short-term forecasting systems. The raw horizontal resolution of numerical weather prediction data is preferably 1 km to 3 km, which is then downscaled to form local meteorological grid data at 500 m to 1000 m, with a temporal resolution of 15 minutes to 1 hour. This resolution range can accommodate both regional weather process descriptions and corrections for topographic differences within the field area, and the 15-minute to 1-hour temporal resolution can meet the data requirements for intraday rolling forecasts and day-ahead trend forecasts. Wide-area meteorological data must include at least the following fields for the next 72 hours: wind speed, wind direction, temperature, air pressure, humidity, precipitation, boundary layer height, and atmospheric stability index. Upper-level dispatch data is provided by the energy management system and automatic power generation control system, and must include at least... Fields include tie-line power reference value, system frequency deviation, automatic generation control scaling commands, intraday rolling plan, day-ahead generation plan, reserve requirements, and power curtailment commands; historical sample data is retrieved from the site-level time-series database, with minute-level samples retained for more than 36 months to cover at least three full-year seasonal cycles; hour-level samples are retained for more than 60 months to cover low-frequency extreme weather events, annual wind resource fluctuations, and reference samples required for training distribution detection; when the sample accumulation of a newly built site or a site after a major overhaul is insufficient, the data aggregation platform calls historical samples of the same type, similar terrain, and similar climate zone as migration reference samples, and assigns data weights to migration reference samples that are lower than those of the site's actual measured samples. Once the accumulation of the site's actual measured samples meets the training requirements, the weights of migration reference samples are gradually reduced; After various data enter the site-level data aggregation platform, they undergo unified time reference processing. Within the site, GPS timing is the primary method, supplemented by Precision Time Protocol (PTP) network timing. Each data acquisition node adds a data acquisition timestamp, node number, and data quality identifier when uploading data. The time synchronization error between any two data acquisition nodes within the site is controlled within 10 milliseconds, preferably not exceeding 1 millisecond. The time synchronization error between the site and the regional dispatch center is controlled within 100 milliseconds, preferably not exceeding 5 milliseconds. This range is smaller than the unit sampling cycle and the automatic generation control command cycle, preventing data from different sources from being categorized into incorrect prediction windows. When the time synchronization error within the site exceeds 10 milliseconds but does not exceed 100 milliseconds, the data aggregation platform suspends cross-device second-level feature stitching, retaining only minute-level statistical features. When the time synchronization error within the site exceeds 100 milliseconds, the data in that time period is marked as a time anomaly sample and is not included in the real-time prediction model and model training pool until time synchronization is restored, at which point it re-participates in the subsequent prediction process. For data with different sampling frequencies, the data aggregation platform resamples the data based on the target time axis of the corresponding prediction task. For second-level data from the generator units, the average, maximum, minimum, standard deviation, and rate of change are extracted within the target time window. 10-second meteorological observation data is aligned to 5-minute or 15-minute time points using linear interpolation or time window statistics. 15-minute or hourly wide-area meteorological data is interpolated to form data sequences corresponding to the prediction scale. Data packets with missing timestamps or timestamp jumps exceeding one sampling period are first placed in an abnormal buffer queue. The abnormal buffer queue calculates candidate timestamps based on the data reception time, preceding and following valid timestamps, and the node's local clock. When the time interval deviation between the candidate timestamp and adjacent valid data is less than 20% of the sampling period, the time assignment of the data packet is restored. When the deviation reaches or exceeds 20% of the sampling period, the data packet is marked as timestamp-abnormal data and is not used as a valid sample for subsequent training and parameter updates. 20% is used as a tolerance ratio to allow for short-term jitter in the communication link while preventing misalignment of data across sampling windows. After completing the time alignment, a unified spatial coordinate mapping relationship for the wind farm area is established. Using the center point of the wind farm's booster station as the origin of the local rectangular coordinate system, a horizontal coordinate axis is established along the prevailing wind direction, a vertical coordinate axis is established perpendicular to the prevailing wind direction, and an altitude coordinate axis is established perpendicular to the ground. Each turbine, each meteorological tower, each lidar unit, and each collector unit is configured with a unique equipment number, local coordinates, hub height, altitude, associated power collection line, and availability status in the equipment spatial ledger. Wide-area meteorological grid points are mapped to representative points in the field area through spatial interpolation. These representative points are selected as the geometric center, boundary corner, and wind farm booster station location of each collector unit. Spatial interpolation uses bilinear interpolation or... Distance-inverse weighted interpolation is used; when using distance-inverse weighted interpolation, the weights are normalized according to the inverse of the square of the distance; meteorological grid points more than 5 kilometers away from the representative point of the field are not included in the interpolation; 5 kilometers is used as the screening distance because this distance is greater than the typical spatial scale of a single collector unit, which can cover the effective grid points near the field, while avoiding the introduction of irrelevant disturbances by meteorological grid points with large differences in distant topography and underlying surface; for wind farms with an east-west or north-south span of more than 10 kilometers, the grid screening distance can be adjusted to 0.5 to 1 times the maximum boundary length of the field, but the upper limit is no more than 10 kilometers, so as to balance the spatial coverage of large-scale stations and the exclusion of distant meteorological disturbances; For unit operation data, the data aggregation platform establishes an aggregation index based on power collection units, wind direction sectors, and unit health status; for site meteorological observation data, an observation index is established based on altitude layer, wind direction, and observation equipment type; data quality is graded according to the effective sampling ratio within the target time window: when the effective sampling ratio is not less than 95%, it is marked as normal data; when the effective sampling ratio is between 85% and 95%, it is marked as limited usable data, and is supplemented using adjacent time points, adjacent altitude layers, or adjacent measurement points of the same type; when the effective sampling ratio is less than 85%, it is marked as low-quality data, and this channel is not used as the core input of the prediction model within the current time window; 95% is used to ensure that the main input of the model has high integrity, and 85% is used as the minimum usable threshold to limit the proportion of missing data supplementation; when the effective sampling ratio is less than 85%, the proportion of missing or abnormal data exceeds 15%, and the proportion of supplemented data may change the statistical characteristics of this time window, so this channel is not used as the core input of the current time window; When a data acquisition node has no valid data for 30 consecutive seconds, the data aggregation platform calls nearby similar measurement points for weighted replacement. The replacement range is preferably 3 to 8 times the rotor diameter, or 3 to 5 closest units of the same model within the same collector unit for weighted supplementation. This range matches the common layout spacing and wake influence distance of wind turbine units, ensuring that the replacement measurement points have similar meteorological and operating conditions while avoiding data distortion caused by selecting units that are too far away. When performing weighted supplementation, the weight of candidate measurement points is determined by spatial distance, the angle of the incoming wind direction, and the consistency of equipment type. In one preferred method, spatial distance accounts for 60%, the similarity of the angle of the incoming wind direction accounts for 30%, and the consistency of equipment type accounts for 10%. The basis for this weight allocation is that spatial distance has the greatest impact on the similarity of wind speed and unit operating status in the same area, the incoming wind direction affects the consistency of wake and wind resources, and the consistency of equipment type is used to avoid inconsistencies in the dimensions and response characteristics of measurement points of different models. When there are fewer than 2 candidate measurement points, the supplemented value is only used for status display and alarm judgment and is not included in the core input of the prediction model. If the effective sampling rate of the same node is lower than 85% for three consecutive 15-minute windows, the data aggregation platform will switch the node's status to low-quality channel status, output an inspection task to the operation and maintenance system, and temporarily remove the corresponding channel of the node from the core input of the subsequent prediction model before the status is restored. Three consecutive 15-minute windows correspond to 45 minutes of continuous anomalies, which is used to distinguish between instantaneous communication packet loss and continuous measurement point anomalies, and avoid unnecessary device removal due to short-term network jitter. The 85% threshold is consistent with the minimum available threshold, which can ensure that the core channels entering the prediction model have stable data integrity. After the above processing, the site-level data aggregation platform generates a unified spatiotemporal dataset with timestamps, spatial coordinates, equipment numbers, data quality identifiers, prediction scale identifiers, and aggregation indexes, and sends it along with the original data version, the supplementary data version, and the anomaly identifier version to the subsequent site-specific correction process.

[0021] Specifically, such as Figure 3 As shown: Based on the unified spatiotemporal dataset formed previously, wide-area meteorological data, field meteorological observation data, historical power data, and superior dispatch data with timestamps, spatial coordinates, data quality identifiers, and aggregation indexes are read. After the single-channel data quality identifier is completed, the data set entering the field-specific correction process is checked for start-up conditions. When the effective sampling ratio of key meteorological channels within the target time window is not less than 85%, the effective sampling ratio of unit active power is not less than 90%, and the effective ratio of superior dispatch instruction timestamps is not less than 95%, field-specific correction and multi-timescale feature hierarchical processing are initiated. Among them, 85% is the lower limit for the initiation of key meteorological channels, and this value is based on wind speed, wind direction, temperature, and air pressure. Meteorological channels exhibit spatial and altitude-level correlations, which can be supplemented by adjacent altitude layers, adjacent measuring points, and adjacent grid points. 90% is the lower limit for the activation of active power channels. This value is based on the fact that historical power sequences directly participate in feature layering and prediction label construction; an excessively high missing proportion will introduce prediction label bias. 95% is the lower limit for the activation of timestamps of superior dispatch instructions. This value is based on the fact that dispatch instructions are used to determine power curtailment status, planning boundaries, and control constraints; timestamp bias will affect the generation of subsequent control constraints. If the above activation conditions are not met, the data aggregation platform retains valid data that has completed quality labeling, suspends model updates for the corresponding time window, and sends the abnormal status label to the subsequent probability prediction process. The field-specific correction of wide-area meteorological data is performed sequentially according to topographic correction, stability correction, and wind direction residual correction. Topographic correction utilizes the field digital elevation model, surface roughness data, local coordinates of the wind farm, and the mapping relationship between the field spatial coordinates to determine the roughness length corresponding to each surface type within the wind direction sector at each meteorological grid point. The roughness length is determined based on the surface type: 0.03 to 0.05 meters for bare land, 0.05 to 0.10 meters for low grassland, 0.10 to 0.20 meters for shrubland, and 0.30 to 0.05 meters for sparse woodland. The roughness length is 0.60 meters for woodlands or dense woodlands, and 0.80 to 1.20 meters for forests or dense woodlands. The above values ​​are based on the degree of obstruction of near-surface wind speed shear by different underlying surfaces. The higher and denser the surface obstacles, the greater the corresponding roughness length. The data aggregation platform weights the roughness length according to wind direction sectors to obtain the equivalent roughness length. The weight of the wind direction sector is determined by the area proportion of each surface type in the upwind sector, the distance from the representative point of the field area, and the angle of the incoming wind. The larger the area proportion of the surface type, the closer it is to the representative point of the field area, and the smaller the angle with the incoming wind direction, the higher the corresponding weight. The data aggregation platform corrects the reference height wind speed in wide-area meteorological data to the hub height of the field area based on the equivalent roughness length; when using power-law wind profiles, the hub height wind speed is determined according to the following formula:

[0022] Where Vh is the wind speed at hub height, Vr is the wind speed at reference height, Hh is the hub height, Hr is the reference height, and α is the wind profile index, the specific value of which is determined by stability correction; when using a logarithmic wind profile, the wind speed at hub height is calculated according to...

[0023] The following parameters are defined: z0 is the equivalent roughness length; power-law wind profiles are suitable for scenarios lacking complete underlying surface profile parameters but having stability determination results; logarithmic wind profiles are suitable for scenarios where the roughness length can be obtained stably from the surface type and digital elevation model; the hub height wind speed after terrain correction, the corresponding meteorological grid number, and the correction version number together form the terrain correction result and are sent to the stability correction process. Stability correction uses the near-surface temperature gradient, wind speed at 10 meters, and boundary layer height as inputs to perform state-specific corrections on the hub height wind speed after topographic correction. When the near-surface temperature gradient is greater than 0.01℃ / m and the wind speed at 10 meters is less than 4 m / s, it is considered a stable state, with a wind profile index ranging from 0.20 to 0.26. Specifically, 0.01℃ / m is used to identify near-surface inversion enhancement scenarios, and 4 m / s is used to exclude the disruption of stable stratification by strong mechanical turbulence. When the near-surface temperature gradient is less than -0.005℃ / m and the wind speed at 10 meters is greater than 6 m / s, it is considered an unstable state, with a wind profile index ranging from 0.10 to 0.12. Specifically, -0.005℃ / m is used to identify near-surface thermal convection enhancement scenarios, and 6 m / s is used to confirm an operating state with significant turbulent mixing. Other conditions are judged as... The system is defined as a neutral state, with a wind profiler index of 0.14. The boundary layer height is used to determine the specific value of the wind profiler index within the range corresponding to the stability state. When the boundary layer height is below 500 meters, the wind profiler index is taken closer to the upper limit within the range corresponding to the stability state. When the boundary layer height is above 1000 meters, the wind profiler index is taken closer to the lower limit within the range corresponding to the stability state. When the boundary layer height is between 500 and 1000 meters, it is determined by linear interpolation within the corresponding range. The range of 500 to 1000 meters is used as the boundary layer height segmentation range, matching the common diurnal boundary layer variation range of wind farms, to distinguish between the operating states of enhanced wind shear under shallow boundary layers and enhanced mixing under deep boundary layers. After the stability correction is completed, the data aggregation platform generates a hub height wind speed sequence with stability state identifiers. Wind direction residual correction uses the near-field wind direction from the lidar and the wind direction at the hub height of the wind measuring tower as references to calculate the residual between the wide-area meteorological wind direction and the measured wind direction in the field. The data aggregation platform statistically analyzes the median of the wind direction residuals over the past 30 days according to wind direction sectors, and uses this median as the wind direction correction for the next forecast period. The 30-day statistical window is used to balance the sample size and seasonal wind direction changes. A window that is too short is easily affected by local gusts, while a window that is too long may weaken the responsiveness to seasonal changes. The wind direction correction value is between -15 degrees and 15 degrees. When the median residual exceeds this range, it is truncated to 15 degrees. The truncation range of -15 degrees to 15 degrees is related to the wind farm's wind direction measurement error, yaw control tolerance, and local topographic disturbance. The range of motion is matched; when it exceeds the range, it is more likely to be a local gust, extreme weather, or observational anomaly, and should not be directly used as a long-term correction. When there are insufficient valid samples within 30 days in the same wind direction sector, the statistics are expanded to adjacent wind direction sectors. If adjacent wind direction sectors are still insufficient to form a valid statistic, the median of the wind direction residuals of historical samples in the same season is used as the initial correction. If there are still no valid samples, the wind direction correction is set to 0, and the wind direction residual correction status is marked as pending learning. After the wind direction residual correction is completed, the data aggregation platform forms field-specific corrected meteorological data, which includes at least the following fields: hub height wind speed, hub height wind direction, stability status, roughness correction parameters, wind direction correction, and data quality weight. After completing the field-specific correction, the data aggregation platform performs multi-time-scale feature layering processing on the corrected meteorological data and historical power data. The processing objects include at least the following fields: field-level active power, hub-height wind speed, hub-height wind direction, air density, turbulence intensity, temperature, humidity, air pressure, lidar wind forecast, power curtailment status, and dispatch plan boundary. Among these, field-level active power, hub-height wind speed, and hub-height wind direction are used as the main sequence, and the remaining fields are used as auxiliary sequences. The data aggregation platform aligns the main sequence and auxiliary sequences according to timestamps to construct a basic feature matrix. Each row in the basic feature matrix corresponds to a time point, and each column corresponds to a meteorological, power, or dispatch boundary feature. The basic feature matrix is ​​divided into four time scales: trend layer, daily and semi-daily cycle layer, hourly fluctuation layer, short-term disturbance layer, abrupt change layer, and noise layer. The trend layer corresponds to changes over 24 hours and is used to characterize weather-scale systems and long-term power output trends. The daily and semi-daily cycle layer corresponds to changes from 12 to 24 hours and is used to characterize day-night cycles and intraday periodic fluctuations. The hourly fluctuation layer corresponds to changes from 1 to 4 hours and is used to characterize mesoscale weather processes and intraday rolling forecast changes. The short-term disturbance layer corresponds to changes from 15 minutes to 1 hour and is used to characterize short-term wind speed fluctuations and gust disturbances. The abrupt change layer corresponds to changes from 1 minute to 15 minutes and is used to characterize rapid power increases, rapid power decreases, unit control responses, and short-term fault disturbances. The noise layer corresponds to measurement fluctuations of less than 1 minute and is used for data quality assessment, but is not used as the core input of the probabilistic prediction model. The above time scale division corresponds to the physical source of wind power fluctuations, enabling the subsequent probabilistic prediction process to call the matching feature layer according to the prediction scale. Layered processing can be implemented using variational mode decomposition, wavelet decomposition, or bandpass filters, with variational mode decomposition being preferred. When using variational mode decomposition, the data aggregation platform performs decomposition on the field-level active power sequence, hub height wind speed sequence, and hub height wind direction sequence respectively, and assigns the components of each sequence to the corresponding time scale layer according to the central period range. Auxiliary sequences are used to interpret and assess the quality of the layering results. The number of decomposition layers ranges from 5 to 8. Five layers are used to cover the main components such as trends, daily cycles, hourly fluctuations, short-term disturbances, and abrupt changes, while eight layers serve as an upper limit to avoid over-decomposition and the generation of false modes. When the residual energy accounts for less than 3% of the total energy and the difference in the center period between any two adjacent modal components is greater than 20%, the number of decomposition layers is stopped. The difference in center period is the ratio of the difference in center periods between two adjacent layers to the smaller center period. The 3% residual ratio is used to ensure that the main wind power fluctuation energy has been absorbed by the decomposition layers, and the 20% center period difference is used to confirm that adjacent modes have distinguishable time scales. If the above conditions are not met simultaneously even after the number of decomposition layers reaches 8, the decomposition results of 8 layers are retained, and the remaining residuals are assigned to the noise layer and marked as residual-limited state in the decomposition version number. During the tiered processing, the data aggregation platform continuously monitors the rapid changes in field-level active power. When the monotonically increasing or decreasing range of field-level active power exceeds 15% of the rated capacity within the last 15 minutes, it is determined that a rapid power increase or decrease is occurring, and the decomposition window is shortened from 36 hours to 18 hours. Changes exceeding 15% of the rated capacity within 15 minutes typically exceed the range of normal random fluctuations and are suitable as criteria for determining strong ramp-up or strong ramp-down. The 36-hour window can cover the entire day-night cycle and retain trend information, while the 18-hour window is used to reduce the smoothing effect of long windows on rapid changes in abrupt change scenarios. When the monotonically changing range of field-level active power is less than 5% of the rated capacity for 30 consecutive minutes, the 36-hour decomposition window is restored. The 30 consecutive minutes are used to confirm that the rapid change process has ended, and the 5% threshold is used to avoid frequent switching of the decomposition window caused by ordinary short-term fluctuations. During the field-based correction or feature stratification process, if wide-area meteorological data is missing, different processing methods are applied according to the duration of the missing data. When the missing length is less than 30 minutes, interpolation is used to fill in the missing data using adjacent valid meteorological grid points and adjacent valid time points. When the missing length reaches more than 30 minutes, since the missing time has reached one update cycle of the intraday rolling forecast, the data aggregation platform switches to the field observation data-dominated mode. The weight of wide-area meteorological features can be selected to be no higher than 0.3, and the weight of wind tower and lidar features can be selected to be no lower than 0.7. The weight values ​​of 0.3 and 0.7 are used to ensure that the field measured data has the main weight in short-term forecasts, while retaining a small amount of wide-area meteorological trend information. After the wide-area meteorological data is restored, the data aggregation platform restores the original weights using linear interpolation within two consecutive intraday rolling cycles. If the missing data occurs again during the restoration period, the field observation data-dominated mode is re-entered. After the above processing, the data aggregation platform generates a multi-scale feature matrix. The multi-scale feature matrix includes at least the following fields: timestamp, spatial index, prediction scale identifier, trend layer features, daily and semi-daily cycle layer features, hourly fluctuation layer features, short-term disturbance layer features, mutation layer features, noise layer quality assessment identifier, data quality weight, meteorological correction version number, feature layer version number, and abnormal state identifier. The multi-scale feature matrix is ​​then sent to the subsequent probability prediction process.

[0024] Specifically, such as Figure 4As shown: After the multi-scale feature matrix is ​​generated, the data aggregation platform verifies the prediction input quality of each time-scale feature layer. When the effective feature ratios of the trend layer, daily and semi-daily cycle layers, and hourly fluctuation layer are all not less than 90%, and the effective feature ratios of the short-term disturbance layer and abrupt change layer are not less than 85%, wind power probability prediction is initiated. 90% is used as the lower limit of the effective ratio of the long-term feature layer. The basis for this value is that the trend layer, daily and semi-daily cycle layer, and hourly fluctuation layer are mainly used for day-ahead and intraday prediction. If the missing ratio is too high, it is easy to cause prediction bias over a long period of time. 85% is used as the lower limit of the effective ratio of the short-term feature layer. The basis for this value is that the short-term disturbance layer and abrupt change layer can be compensated by real-time measurement, persistence method, and features of adjacent time periods. When the above verification conditions are not met, the data aggregation platform marks the corresponding prediction scale as a feature-limited prediction state. When the long-term features are insufficient, the day-ahead prediction model parameter update is suspended, and the trend boundary that passed the quality verification most recently is used. When the short-term features are insufficient, the most recent effective power, real-time wind measurement data, and persistence method are used to generate temporary prediction inputs. The wind power probabilistic prediction model adopts a structure of multi-scale shared feature layer and scaled output layer. The shared feature layer receives features from the trend layer, daily and semi-daily cycle layers, hourly fluctuation layer, short-term disturbance layer, and abrupt change layer, and combines them with historical sample labels, unit availability status, dispatch curtailment status, and site spatial index for temporal encoding. The shared feature layer can be constructed using temporal convolutional networks, gated recurrent networks, long short-term memory networks, or attention networks to extract temporal and spatial correlations between features at different time scales. The scaled output layer includes at least a day-ahead prediction output layer, an intraday prediction output layer, an ultra-short-term prediction output layer, and a real-time prediction output layer. Each output layer calls the feature layer that matches the prediction scale to form the intermediate prediction representation at the corresponding scale. During model training, training labels are constructed from historical samples according to the prediction scale. The day-ahead prediction label is the field-level active power at the corresponding time point within the next 24 to 72 hours; the intraday prediction label is the field-level active power at the corresponding time point within the next 4 hours; the ultra-short-term prediction label is the field-level active power at the corresponding time point within the next 15 to 4 hours; and the real-time prediction label is the field-level active power at the corresponding time point within the next 1 to 15 minutes. The above label range is consistent with the output range of each prediction scale, so that the model training target corresponds to the online prediction target. Each output layer generates corresponding quantiles through quantile regression. During training, quantile loss is used to constrain the deviation between the predicted value and the true value. The quantile loss is assigned different weights according to whether the predicted value is higher or lower than the true value, so that the corresponding quantile output matches the target probability position in the historical error distribution. The current forecast output layer utilizes the trend layer, daily and semi-daily cycle layers, and wide-area meteorological correction features, with a forecast range of 24 to 72 hours and a time resolution of 15 minutes. The intraday forecast output layer utilizes the trend layer, hourly fluctuation layer, and latest dispatch plan features, with a forecast range of 4 hours, a time resolution of 15 minutes, and a rolling cycle of 30 minutes. The ultra-short-term forecast output layer utilizes the hourly fluctuation layer, short-term disturbance layer, lidar wind forecast, and unit status features, with a forecast range of 15 minutes to 4 hours, a time resolution of 5 minutes, and a rolling cycle of 5 minutes. Real-time forecast output... The system invokes short-term disturbance layer, abrupt change layer, measured power change rate in the most recent 15 minutes, and automatic generation control commands. The prediction range is from 1 minute to 15 minutes into the future, with a time resolution of 30 seconds and a rolling cycle of 30 seconds. These values ​​match the commonly used time scales for wind power prediction and dispatch control. The 15-minute resolution is suitable for day-ahead planning and intraday rolling verification, the 5-minute resolution is suitable for ultra-short-term power fluctuation tracking, and the 30-second resolution is suitable for rapid active power correction at the power plant level. When intraday prediction and ultra-short-term prediction overlap within the next 4 hours, intraday prediction provides the trend boundary, and ultra-short-term prediction provides near-end correction. Each prediction scale outputs a quantile sequence, predicted mean, time-varying covariance matrix, and tail scenario set. Day-ahead and intraday predictions output nine quantiles: 5%, 10%, 25%, 40%, 50%, 60%, 75%, 90%, and 95%. Of these, 5% to 95% are used to form a 90% confidence interval, 10% to 90% to form an 80% confidence interval, 25% to 75% to characterize the quartile range, 40% to 60% to characterize the distribution concentration near the median, and 50% serves as the median prediction benchmark. Ultra-short-term predictions add two tail quantiles, 2.5% and 97.5%, to the above nine quantiles to enhance the identification of extreme low-output and extreme high-output scenarios. Real-time predictions output point value predictions, a 10% to 90% confidence band, and instantaneous uncertainty. When a high quantile is lower than a low quantile at the same prediction time, the data aggregation platform performs monotonicity correction according to the quantile size from smallest to largest to ensure consistency in the output probability distribution. The time-varying covariance matrix is ​​generated based on the rolling statistics of the prediction residuals; the diagonal elements of the matrix are obtained by converting the historical residual variance or quantile width at the corresponding prediction time, and the off-diagonal elements are determined by the standard deviation and error correlation coefficient at the corresponding two prediction times; the error correlation coefficient is calculated according to the following formula:

[0025] Where Δt is the time interval between two prediction times, and τ is the decay time constant for the corresponding prediction scale; the decay time constant is 4 hours for day-ahead predictions, 1 hour for intraday predictions, 15 minutes for ultra-short-term predictions, and 5 minutes for real-time predictions. The above values ​​are based on the fact that the duration of prediction errors varies at different prediction scales. Day-ahead errors are mainly affected by the evolution of weather systems and have a longer duration; intraday errors are affected by rolling meteorological corrections and have a shorter correlation duration; ultra-short-term and real-time errors are mainly caused by gusts, turbulence, and control responses, and their correlation decays faster. When the generated time-varying covariance matrix does not meet the positive semi-definite condition, its eigenvalues ​​are truncated non-negatively, or a stabilizing term of no less than 1% of the historical residual variance is added to the diagonal of the matrix, and the off-diagonal elements in the form of the correlation coefficient are recalculated according to the corrected diagonal elements. This stabilizing term is used to improve the numerical stability of the matrix and does not significantly change the original covariance structure. The tail scene set is generated based on the predicted mean, quantile sequence, and time-varying covariance matrix. The data aggregation platform first generates candidate power trajectories based on the predicted mean and time-varying covariance matrix, and then truncates or shifts the candidate power trajectories according to the quantile sequence to ensure that the trajectory distribution at each prediction time is consistent with the corresponding quantile interval. Candidate power trajectories can be generated based on multivariate normal distribution sampling or empirical residual resampling. When the historical residuals are significantly skewed or have heavy tails, empirical residual resampling is preferred. The number of tail scene trajectories for day-ahead and intraday predictions is 500, and for ultra-short-term and real-time predictions... The number of predicted tail scenario trajectories is set to 200; 500 trajectories are used to improve the stability of tail statistics in day-ahead and intraday forecasts, suitable for offline or near real-time rolling calculations; 200 trajectories are used to meet the online calculation latency requirements of ultra-short-term and real-time forecasts, while maintaining the basic stability of tail probability estimation; the combined proportion of trajectories below the 5th percentile and above the 95th percentile in the tail scenario set is not less than 15%. This proportion is based on the theoretical 10% tail probability at both ends, increasing the tail sampling density to obtain more stable extreme scenario samples for subsequent conditional risk constraints and reserve capacity configuration. Before outputting probabilistic prediction information, the data aggregation platform performs cross-scale consistency verification. The verification targets are the mean trajectory, quantile trajectory, and time-varying covariance matrix of day-ahead, intraday, ultra-short-term, and real-time predictions within overlapping time periods. The mean deviation of adjacent prediction scales within the same time period does not exceed 5% of the rated capacity, the 90% confidence interval width ratio is between 0.75 and 1.30, and the positional deviation between the 5% quantile and the 95th quantile does not exceed 7% of the rated capacity. Among these, the 5% mean deviation threshold is used to identify cross-scale offsets near the sensitive area of ​​active power dispatching plan adjustments for wind farms; the confidence interval width ratio of 0.75 to 1.30 is used to allow reasonable uncertainty differences caused by different prediction durations, while also identifying abnormal states such as sudden contraction or expansion of confidence intervals; the 7% tail quantile deviation threshold is slightly higher than the mean deviation threshold, and its value is based on the fact that the tail quantile is more sensitive to sample perturbations. When cross-scale consistency verification fails, the data aggregation platform uses short-scale prediction to correct the near-term period and long-scale prediction to constrain the far-term trend for re-fusion. The re-fusion weight is between 0.30 and 0.70. The weight of short-scale prediction at the beginning of the overlapping time period is 0.70, and at the end of the time period it is 0.30. The intermediate time is determined by linear interpolation. The weight range of 0.30 to 0.70 is used to avoid a single-scale prediction completely dominating the fusion result, and to retain the combined effect of long-scale trend and short-scale correction. Re-fusion is performed in a maximum of 4 rounds. The upper limit of 4 rounds is used to avoid output delay caused by repeated iterations while meeting the online computing time limit. If the consistency verification condition is still not met after 4 rounds of re-fusion, the data aggregation platform selects the candidate prediction result with the smallest mean deviation from the adjacent prediction scale in the overlapping time period as the temporary output, and sets the prediction status to inconsistent and awaiting calibration. The probability prediction information synchronously generates ramp event identifiers; the data aggregation platform scans the tail scene set in three rolling windows of 15 minutes, 30 minutes, and 1 hour; when the power increase exceeds 10% of the rated capacity in the 15-minute window, 15% of the rated capacity in the 30-minute window, and 20% of the rated capacity in the 1-hour window, a ramp event is determined to exist in the corresponding window; when the power decrease reaches the same threshold, a ramp-down event is determined to exist in the corresponding window; the 15-minute, 30-minute, and 1-hour windows correspond to the response scales of real-time control, ultra-short-term regulation, and intraday rolling scheduling, respectively; the 10%, 15%, and 20% thresholds are used to identify power changes that may affect active power regulation in advance during the prediction stage, among which the 10% threshold in 15 minutes is lower than the strong rapid change threshold used in the feature layering decomposition window switching, and is used to capture potential ramp risks in advance; the probability of an event occurring is equal to the ratio of the number of trajectories that trigger the event to the total number of trajectories in the tail scene set, and the event output data includes at least the fields of occurrence time, duration, amplitude, direction, probability of occurrence, and associated prediction scale; When the current input feature is outside the training sample distribution, the data aggregation platform switches the prediction status to a low-confidence prediction status. The out-of-distribution status can be determined using Mahalanobis distance, kernel density estimation probability, or isolated forest anomaly score; Mahalanobis distance is preferred. When the Mahalanobis distance of the current input feature exceeds the 99th quantile of the historical training sample Mahalanobis distance distribution, it is determined to be outside the training sample distribution. The 99th quantile is used to control the misclassification rate of normal samples and to identify extreme or rare input states. When in a low-confidence prediction state, the quantile interval is expanded according to the error distribution of similar weather events over the past 7 days. The past 7 days reflect the error distribution under the current season and recent wind conditions. The expansion ratio ranges from 10% to 30%. When the out-of-distribution degree is less than 1.2 times the judgment threshold, the expansion ratio is 10%; when it is between 1.2 and 1.5 times, the expansion ratio increases from 10% to 30% through linear interpolation; and when it is greater than 1.5 times, the expansion ratio is 30%. After the above processing, the data aggregation platform generates wind power probability prediction information covering day-ahead, intraday, ultra-short-term, and real-time prediction scales. The wind power probability prediction information includes at least the quantile sequence, prediction mean, time-varying covariance matrix, tail scenario set, cross-scale consistency status, low confidence prediction status, and ramp-up event identifier for each prediction scale, and is sent to the subsequent active power regulation constraint generation process.

[0026] Specifically, such as Figure 5 As shown: After the cross-scale consistency verification of wind power probability prediction information is completed, or after being re-fused to form temporary usable probability prediction information, the data aggregation platform sends the quantile sequence, prediction mean, time-varying covariance matrix, tail scenario set, ramp event identifier, unit available capacity, upper-level scheduling plan, tie line power limit, reserve requirements, and site spatial coordinates to the site controller; the site controller maps the prediction uncertainty into executable active power control constraints, and generates site-level active power instructions and single-unit execution instructions after the constraints are verified. Wind power probability prediction information is mapped to quantile opportunity constraints, tail risk constraints, and correlation robustness constraints. The quantile opportunity constraints are set with feasibility confidence levels according to the prediction scale: 0.90 for day-ahead dispatch, 0.92 for intraday dispatch, 0.95 for ultra-short-term dispatch, and 0.98 for real-time control. These values ​​are based on the following: day-ahead dispatch has a longer correction time, and 0.90 meets the needs of planning and reserve assessment; intraday dispatch has a shorter correction time, and 0.92 is used to improve the reliability of rolling plans; ultra-short-term dispatch directly affects minute-level reserve and automatic generation control tracking, and 0.95 is used to improve the feasibility of near-end regulation; real-time control has the shortest error correction window, and 0.98 is used to improve the security of immediate control. Full margin; the target value of the field-level active power command for each scheduling period in the future is checked according to the quantile safety range of the corresponding forecast scale. The 10% to 90% quantile range is used for the day-ahead and intraday periods, and the 2.5% to 97.5% quantile range is used for ultra-short-term and real-time control. When the target value deviates from the quantile safety range, the station controller generates the reserve call amount, the limited power generation correction amount, or the reachable output correction amount. The above quantile range is based on the fact that the day-ahead and intraday stages have a rolling correction margin, and the use of the 10% to 90% quantile range can take into account both plan tracking and adjustment economy. The error correction time is short in the ultra-short-term and real-time stages, and the use of the 2.5% to 97.5% tail range can improve the coverage of the short-term control boundary. Tail risk constraints are determined based on the tail scenario set; the station controller extracts extreme low-output trajectories below the 5th percentile and uses the mean of such trajectories as the low-output representative boundary; it also extracts extreme high-output trajectories above the 95th percentile and uses the mean of such trajectories as the high-output representative boundary; the low-output representative boundary is used to characterize the plan tracking capability under extreme low-output conditions, and the high-output representative boundary is used to characterize the delivery safety boundary under extreme high-output conditions. The correlation robust constraint is determined based on the time-varying covariance matrix; the station controller uses the predicted mean trajectory as the center of the uncertainty set, and the time-varying covariance matrix characterizes the correlation between prediction errors in different time periods, constructing an ellipsoidal power deviation set; the ellipsoidal power deviation set satisfies the following relationship:

[0027] Where P is the candidate power trajectory, μ is the predicted mean trajectory, Σ is the time-varying covariance matrix, and r is the robust radius; the robust radius for day-ahead scheduling is 1 standard deviation, for intraday scheduling it is 1.5 standard deviation, and for ultra-short-term and real-time control it is 2 standard deviation; under extreme weather or low-confidence forecast conditions, the robust radius is expanded to 3 standard deviations; the basis for this value is that the closer to real-time control, the more direct the impact of forecast error on control actions, requiring a wider uncertainty coverage range; under extreme weather or low-confidence forecast conditions, the fluctuation of forecast distribution is aggravated, and using 3 standard deviations can improve the coverage of abnormal scenarios; when Σ is irreversible or the condition number exceeds the numerical stability upper limit, the covariance matrix after adding the stability term is used to participate in the robust constraint verification, or a pseudo-inverse matrix is ​​used for calculation, to avoid numerical instability in the robust constraint verification process; Based on the aforementioned active power regulation constraints, the power station controller determines the spinning reserve capacity or active power reserve capacity in layers. Reserve capacity is calculated separately for second-level response reserve, minute-level rolling reserve, and hourly planned reserve. Second-level response reserve is oriented towards primary frequency regulation and automatic generation control, using the weighted sum of the square roots of the diagonal elements of the covariance matrix within the first 15 minutes of the ultra-short-term forecast as a measure of uncertainty, and taking the larger of this sum and the lower limit verification value corresponding to 5% of the wind farm's real-time output. During normal operation, 8% of the wind farm's real-time output is used as the upper limit for reasonableness verification. In extreme weather, low-confidence forecast conditions, or when a ramp-up event is triggered, exceeding the 8% upper limit is allowed, and reserve increment correction is initiated. For the n forecast times within the first 15 minutes of the ultra-short-term forecast, the weight of the i-th forecast time can be determined according to...

[0028] It is determined that the summation range is j = 1 to n, where, Here, represents the weight corresponding to the i-th prediction time, n is the total number of prediction times divided according to the prediction time resolution within the first 15 minutes of the ultra-short-term prediction, i is the current prediction time index participating in the weighting, j is the summation index, i=1 represents the prediction time closest to the current time, and i=n represents the prediction time farthest from the current time. In this formula... This indicates that the summation is performed on all terms from j to n. Since the prediction error closer to the current time has a more direct impact on the second-level response reserve, the above weighting method is used to give higher weight to the near-end prediction variance. 5% to 8% is used to cover instantaneous prediction errors and automatic generation control tracking deviations. If the proportion is too low, it will be difficult to cover rapid fluctuations. If the proportion is too high, it will cause unnecessary occupation of the output margin. Minute-level rolling reserve is used for intraday scheduling, with half the maximum width of the 90% confidence interval for the next 4 hours as the capacity basis, and the larger of the lower limit check value corresponding to the 7% forecast average is taken. During normal operation, the upper limit of the reasonableness check value is 12% of the forecast average. 7% to 12% is used to cover the cumulative forecast error and short-term weather changes within the intraday rolling cycle. Hourly-level planned reserve is used for day-ahead planning, with half the maximum width of the 95% confidence interval for the next 24 hours as the capacity basis, and the larger of the lower limit check value corresponding to the 10% forecast average is taken. During normal operation, the upper limit of the reasonableness check value is 18% of the forecast average. 10% to 18% is used to cover the longer timescale deviation between the day-ahead plan and actual operation. Thus, the reserve capacity is determined by the probability forecast uncertainty as the base value, and engineering verification is carried out using a proportional range to avoid fixing the reserve capacity to a conservative proportion in the long term. When a ramp event indicator shows that a ramp-up or ramp-down event will occur in the future scheduling period, the reserve capacity at the corresponding level is incrementally adjusted according to the probability of the event. When the probability of a ramp event is greater than 0.60, it indicates that most tail scenarios have experienced rapid changes in power in the same direction, and the reserve capacity is increased by 20%. When the probability of a ramp event is greater than 0.90, it indicates that the event has a high confidence of occurrence, and the reserve capacity is increased by 40%. When the probability of the event is between 0.60 and 0.90, the reserve increment is determined by linear interpolation between 20% and 40%. This rule enables the reserve capacity to be dynamically adjusted with the probability of ramp risk, avoiding long-term reservation based on a fixed worst-case boundary. Field-level active power instructions are generated based on the scheduling plan, using quantile opportunity constraints, tail risk constraints, correlation robustness constraints, and reserve capacity as boundary conditions. Candidate field-level active power instructions can be generated in steps of 0.5% to 1% of rated capacity near the scheduling plan, or a candidate set can be formed by the upper and lower boundaries of the quantile safety interval, the predicted mean, and the scheduling plan. The 0.5% to 1% step size matches the field-level active power adjustment accuracy, reducing online computation while ensuring search accuracy. When the predicted mean, quantile safety interval, tail representative boundary, and tie-line constraints all satisfy the scheduling plan, the field-level active power instructions are executed according to the scheduling plan. When the planned output exceeds the low quantile safety boundary and the available output under the low output representative boundary is insufficient to meet the planned tracking margin, the station controller generates an achievable output correction and sends the reserve requirement upward to the regional dispatch center; when in-situ adaptive correction is required, the station-level active power command is lowered; when the high output representative boundary exceeds the tie-line power limit or the station's outgoing capacity, the station controller generates a power curtailment correction and a wind curtailment risk indicator, and lowers the station-level active power command accordingly; the adjustment range of the station-level active power command is based on the principle of minimizing deviation from the dispatch plan, and among multiple candidate commands that meet the control constraints, the candidate command with the smallest deviation from the dispatch plan and the lowest reserve capacity occupancy is selected; After the power generation command at the plant level is generated, the plant controller executes the decomposition of individual unit commands. The inputs for individual unit command decomposition include at least the available capacity of each unit, real-time wind speed, unit power curve, health status, wake influence coefficient, fatigue life remaining, and grid connection status. The available capacity of a single unit is capped at the nameplate capacity, based on the power curve value corresponding to the real-time wind speed, and derating is performed according to pitch angle deviation, yaw error, main bearing temperature, tower vibration, and converter temperature. A yellow alarm reduces the available capacity by 5% to 10%, an orange alarm reduces the available capacity by 15% to 20%, and a red alarm reduces the available capacity by 25% to 30%. A yellow alarm indicates a minor abnormality and is mainly used for preventative derating. An orange alarm indicates that component temperature rise, vibration, or yaw error has significantly affected the available power. A red alarm indicates the presence of protection actions or shutdown risk, and therefore the reduction is the largest. When the unit is under maintenance, out of service due to a fault, or in an abnormal grid connection state, the available capacity is set to 0. After the available capacity of a single unit is determined, the station controller further limits the command share of each unit based on the fatigue life remaining value. The fatigue life remaining value is a cumulative damage index formed based on blade load, tower vibration, and number of start-ups and shutdowns, and is normalized to 0% to 100%. Units with a fatigue life remaining value of less than 20% are close to the high fatigue operating zone, and their command share shall not exceed 60% of the available capacity. Units with a fatigue life remaining value of more than 80% have a higher load margin, and their command share can be increased to 95% of the available capacity. When the fatigue life remaining value is between 20% and 80%, the command share is determined by linear interpolation. The wake effect is calculated based on the spatial coordinates of the wind farm area and the wind direction after site-specific correction. The station controller identifies upwind and downwind turbines. Significant wake coupling is determined when the angle between the wind direction of the downwind and upwind turbines is less than 15 degrees and the distance between the turbines is less than 10 times the rotor diameter. The 15-degree angle is used to determine whether two turbines are in the same wind-blown area, and the 10-times rotor diameter is used to cover the common wake effect distances in wind farms. The wake effect coefficient is calculated based on the distance between the upwind and downwind turbines, the angle between the wind direction, and the wind direction. The hub height difference and turbulence intensity are determined, with values ​​ranging from 0 to 1. The smaller the unit spacing and the smaller the angle of the incoming wind direction, the greater the wake influence coefficient. The higher the turbulence intensity, the faster the wake spreads, and the lower the wake influence coefficient accordingly. When the wake influence coefficient is greater than 0.12, the downwind unit's command share is reduced by 5% to 10%. 0.12 is used to identify wake disturbances that are sufficient to affect the achievable power of downwind units. When the wake influence coefficient is not greater than 0.12, the wake influence is included in the allocation weight calculation as a general spatial difference. The initial allocation weights for a single unit are determined jointly by the unit's available capacity, fatigue life remainder, and wake influence coefficient; for example, the initial allocation weights for a single unit are determined according to...

[0029] It is confirmed that, among them, This represents the normalized single-machine usable capacity. This represents the normalized fatigue life remaining. This refers to the wake effect coefficient; the station controller applies this parameter to all participating units. Normalization is performed, and field-level active power commands are allocated according to the normalization weights; The single-unit execution command includes at least the fields of active power command, pitch coordination command, and yaw coordination command. After receiving the single-unit execution command, the unit sends back the current achievable upper limit. When the current achievable upper limit is lower than 95% of the issued active power command, the station controller triggers a station-level re-decomposition, allocating the difference to units with higher available capacity, better health status, and less wake impact. The 95% threshold allows for small errors during unit control execution; a value below this threshold indicates a significant mismatch between the command and the unit's actual achievable capacity. The re-decomposition is performed in a maximum of two rounds. This two-round limit is designed to meet the real-time requirements of station control while avoiding command delays caused by repeated decomposition. If the station-level command still cannot be met after two rounds of re-decomposition, the station controller sends the station's achievable output upper limit and the difference in capacity to the regional dispatch center. When communication with the upper-level dispatcher is interrupted for more than 5 minutes, the station controller enters the dispatch communication anomaly branch. The station-level active power commands are temporarily executed according to the last valid dispatch plan and the upper limit of 70% of the current available capacity. The 5-minute interval is used to distinguish between short-term communication jitter and continuous dispatch link interruption. The 70% upper limit is used to reserve sufficient active power regulation margin when dispatch communication is abnormal. After communication is restored, normal control is restored when the command handshake is effective for three consecutive 4-second cycles. The 4-second cycle matches the commonly used command cycle of automatic generation control. The three consecutive cycles are used to confirm the stability of communication restoration and avoid immediately switching back to normal control after a single successful handshake. After the above processing, the station controller generates the station-level active power command, single-unit execution command, standby capacity configuration result and command execution status, and sends them to the subsequent actual output deviation analysis process.

[0030] Specifically, such as Figure 6 As shown: After the field-level active power command and single-unit execution command are issued, and the field-level data aggregation platform receives the actual active power output measurement value after the control execution, deviation acquisition and feedback processing are initiated. Input data includes at least the following fields: field-level active power measurement value, single-unit active power measurement value, preceding probability prediction information, issued field-level active power commands and single-unit execution commands, standby call records, unit status change records, power curtailment status, data quality identifier, and abnormal branch status identifier. Actual output deviation includes prediction deviation and command execution deviation. Prediction deviation characterizes the deviation between actual active power output and the corresponding predicted value, while command execution deviation characterizes the deviation between actual active power output and the issued active power command. It can be represented as:

[0031] in, The actual active power output at time t. For the predicted mean, point value prediction, or quantile boundary at the corresponding prediction scale; instruction execution deviation. It can be represented as

[0032] in, The field-level active power command or single-unit execution command at time t; by distinguishing between prediction deviation and command execution deviation, it is possible to avoid mistakenly treating power curtailment, standby dispatch, changes in automatic generation control commands, or insufficient unit reachability as simple prediction errors in model updates; Actual output deviation is collected in parallel at the second, minute, and long-term time intervals. The second-level deviation is collected in 30-second intervals, corresponding to the real-time prediction cycle and the station-level rapid active power correction cycle. The second-level prediction deviation is the difference between the actual active power output at the station level and the real-time prediction point value. The second-level command execution deviation is the difference between the actual active power output at the station level and the active power command at the station level during the same period. At the same time, the changes in the active power command at the station level, the changes in the automatic generation control command, and the standby call status are recorded within the 30-second cycle. The second-level deviation is processed by a first-order low-pass filter with a filter time constant of 10 seconds. The 10-second time constant is less than the 30-second real-time prediction cycle, which can filter out communication jitter, sampling spikes, and short-term measurement glitches, thereby preserving the true power fluctuation trend. Minute-level deviations are collected on a 5-minute cycle, corresponding to the ultra-short-term forecast rolling cycle. Minute-level forecast deviations are the difference between the 5-minute field-level active power average and the ultra-short-term forecast quantile trajectory. The data aggregation platform determines whether this average falls within the 10% to 90% quantile range and records the direction of the out-of-bounds movement, the out-of-bounds amplitude, and the number of consecutive out-of-bounds movements. The 10% to 90% quantile range corresponds to the 80% confidence interval, used to identify whether the short-term forecast deviation has exceeded the normal fluctuation range. Minute-level command execution deviations are the difference between the 5-minute field-level active power average and the average field-level active power command within this time window, used to determine whether the active power commands issued by the station controller can be stably executed by the generator group. Long-term forecast bias is assessed using a 30-minute rolling period to cover multiple ultra-short-term forecast periods and suppress single-event anomaly triggers. Long-term forecast bias is characterized by the normalized root mean square error (RMSE) between the actual power output sequence and the intraday forecast mean sequence over the past 6 hours, with the wind farm's rated capacity as the normalization benchmark. The 6-hour window covers multiple intraday rolling forecast periods, suitable for distinguishing between short-term disturbances and persistent distribution drift. The data aggregation platform also calculates the distribution divergence index between the actual and predicted power output distributions, determined using relative entropy, expressed as follows:

[0033] in, This represents the actual power output probability density. To predict the power output probability density, and Obtained through kernel density estimation; to avoid computational anomalies caused by a probability density of 0, a value less than 0 is added to the probability density estimate. The smoothing term; relative entropy is used to determine whether the predicted distribution has experienced systematic drift, and not just to determine the magnitude of the error at a single time point; Deviation feedback is routed to corresponding parameter ranges according to time scale and deviation type; second-level prediction deviations are fed back to the output bias parameters of the real-time prediction model, minute-level prediction deviations are fed back to the quantile calibration parameters of the ultra-short-term prediction model, and long-term prediction deviations are fed back to the mid-frequency characteristic parameters of the intraday and day-ahead prediction models; command execution deviations are fed back to the parameters for unit available capacity assessment, single-unit weight allocation, and field-level active power constraint correction. The reason for using zonal feedback is that short-term deviations may come from automatic generation control actions, turbulence disturbances, or measurement noise, and should not be directly changed to the long-term trend parameters of the day-ahead model; long-term deviations better reflect the systematic differences between weather patterns, forecast distribution, and actual output distribution, and are suitable for intraday and day-ahead forecast corrections. Model parameter updates employ an incremental approach, keeping the core model parameters frozen and adjusting only output bias parameters, quantile calibration parameters, or highly sensitive adjustable parameters relevant to the corresponding time scale. Highly sensitive adjustable parameters are those contributing the top 10% to the prediction error in the most recent validation samples, or those ranking in the top 10% for absolute gradient value. The update direction is determined by the gradient direction of the corresponding parameter with respect to the prediction loss function. When updating quantile calibration parameters, the update direction is determined by the out-of-bounds direction where the actual output falls above or below the target quantile interval. The update step size is linked to the deviation magnitude: when the deviation is less than 3% of the rated capacity, the step size is 0.1% of the current parameter's absolute value; when the deviation... When the deviation is between 3% and 6% of the rated capacity, the step size is 0.5% of the absolute value of the current parameter; when the deviation is greater than 6% of the rated capacity, the step size is 1% to 2% of the absolute value of the current parameter, and short-term anomaly records are generated simultaneously; within 3% is usually considered to be a normal prediction error or control tracking error, and a small step size can avoid parameter oscillation; 3% to 6% indicates that the deviation has entered a significant correction range, and a medium step size is used; when it exceeds 6%, there may be sudden weather changes, power rationing changes, or model drift, and a larger step size is used to improve the response speed; when the absolute value of the current parameter is lower than the lower limit of the normalized parameter scale, the normalized scale unit is used as the basis for step size calculation to avoid the parameter approaching 0 and failing to update; When the single-machine instruction execution deviation of a certain unit exceeds 5% of its issued instructions for two consecutive 5-minute windows, the data aggregation platform reduces the allocation weight of the unit in the next cycle and increases the priority of the unit's available capacity review; when the single-machine instruction execution deviation of a certain unit is less than 2% of its issued instructions for three consecutive 5-minute windows, the allocation weight of the unit is gradually restored; two consecutive 5-minute windows are used to avoid misjudgment of a single disturbance, 5% is used to identify a significant deviation between the unit's achievable capacity and the issued instructions; three consecutive 5-minute windows below 2% is used to confirm that the unit's execution capacity has returned to stability; Incremental updates employ a dual-threshold trigger. The first threshold is the moving average of the deviation. When the minute-level moving average of the deviation exceeds 5% of the rated capacity over the past 30 minutes, updates to the real-time and ultra-short-term prediction models are triggered. The 30-minute window is used to avoid updates triggered by a single 5-minute anomaly, and the 5% threshold is used to identify persistent deviations that have regulatory significance. The second threshold is the distribution divergence. When the relative entropy index exceeds 0.15 over the past 6 hours, updates to the intraday and day-ahead prediction models are triggered. The 0.15 threshold is determined based on the average and fluctuation range of relative entropy in historical stable operating samples. During stable operation, the relative entropy is usually below 0.10, and exceeding 0.15 indicates a persistent deviation between the predicted and actual distributions. A 2-hour cooldown period is set after each incremental update of any model level. During the cooldown period, only the deviation is recorded and the deviation statistics are updated; updates to parameters at the same level are not repeated. The 2-hour cooldown period can cover multiple ultra-short-term and intraday rolling update cycles, avoiding frequent fluctuations in model parameters. If the same prediction model triggers three consecutive incremental updates within 24 hours and the deviation still does not return to the corresponding threshold, the data aggregation platform determines that the incremental update is insufficient to adapt to the current data distribution and initiates the retraining process. Three consecutive triggers within 24 hours are used to identify persistent mismatches on the same day, avoiding direct entry into retraining due to occasional anomalies. The retraining sample is taken from all valid samples in the past 90 days, and the validation sample is taken from valid samples in the past 7 days. The 90-day sample is used to cover recent seasonal changes and major weather patterns, and the 7-day sample is used to reflect recent wind conditions and operating status. After retraining, if the normalized root mean square error of the validation set decreases by no less than 5% compared to the original model, and the normalized deviation of the actual output of the field level relative to the rated capacity improves by no less than 0.5 percentage points compared to the original model within 1 hour of shadow operation, the new model is allowed to take over online operation; otherwise, it is rolled back to the original model, and the retraining results are retained as offline candidate models. When extreme weather conditions or conditions outside the training sample distribution are identified, the deviation feedback process does not directly update the model parameters. Instead, it marks the corresponding deviation sample as an abnormal operating condition sample and sends extreme mode identifiers, low confidence prediction identifiers, and backup correction identifiers to subsequent data processing, probability prediction, and active power control constraint generation processes. Extreme weather conditions are identified by meteorological triggering conditions. These conditions include: when a typhoon path enters within 150 kilometers of the field area; when a strong convective cell is expected to pass through the field area within 6 hours; when icing conditions reach a temperature below -10℃ and relative humidity above 90%; or when sandstorms or dense fog reduce visibility to below 500 meters. Extreme mode: 150 km is used for identifying the outer impact of typhoons on coastal or near-shore wind farms; 6 hours is used for short-term severe convection warnings; -10℃ and relative humidity greater than 90% are used to identify high-risk icing conditions; 500 meters of visibility is used to identify low-visibility scenarios that affect lidar, video monitoring, and operational safety. The above thresholds can be adjusted according to the climate zone of the site and operating procedures. In extreme mode, the reserve increment can be checked against the baseline by an upper limit of 50% to 100%. This increment serves as the upper limit for reserve capacity check in extreme mode, and the actual increment is still determined based on forecast uncertainty, ramp-up probability, and available capacity. When the Mahalanobis distance between the current input feature and the mean of the training samples over the past 90 days exceeds 1.5 times the 99th quantile of the historical Mahalanobis distance distribution, the data aggregation platform enters out-of-distribution mode. The 99th quantile is used to control the misclassification rate of normal samples, and 1.5 times is used to further identify extreme or rare input states. In this mode, the current abnormal samples do not participate in online parameter updates, and the prediction output switches to a weighted fusion of the continuous method and model prediction. The fusion output can be expressed as...

[0034] in, To fuse the predicted output, For continuous method prediction output, The model predicts the output, and λ represents the persistence weight. The persistence weight λ ranges from 0.60 to 0.80, with a higher λ value for a larger Mahalanobis distance. The range of 0.60 to 0.80 is used to ensure that the recent measured force is dominated by the out-of-distribution state while retaining the trend information of the model prediction. During this period, outlier samples are retained but their training weights are reduced, with the training weights of outlier samples reduced to 0.10 to 0.30 of the weights of normal samples. This range is used to retain information from extreme samples while avoiding outlier samples from dominating the normal model parameters. After the Mahalanobis distance returns to below the historical experience threshold and remains below it for 2 hours, the normal bias reflux is restored. When sensor, field communication, or time synchronization status reduces the reliability of deviation samples, the deviation feedback process does not immediately update the model. Sensor anomalies are defined as key measurement points remaining invalid or maintaining constant values ​​for 30 consecutive seconds. This 30-second period matches the real-time prediction cycle and can be used to identify persistent sampling anomalies. Field communication anomalies are defined as communication interruptions in the power collection unit or field acquisition link exceeding 2 seconds. When communication interruptions exceed 2 seconds, the weight of deviation samples is reduced; when communication interruptions exceed 30 seconds, the relevant power collection unit enters local conservative control. The 2-second interval is used to identify a decrease in the reliability of data from the rapid field acquisition link, and the 30-second interval is used to trigger control degradation. Unlike communication interruptions at the higher-level dispatch center, time synchronization anomalies are defined as a synchronization error of more than 10 milliseconds at the core acquisition node, or a time synchronization error between the field station and the dispatch center exceeding 100 milliseconds. When the synchronization error exceeds the preferred synchronization requirement but does not exceed the above-mentioned anomaly threshold, the weight of the deviation sample is reduced, and the sample is not immediately removed. During sensor anomalies, the weighted substitute value of the neighboring unit of the same model is used for status display and control reference, but it is not used as a model update label. During time synchronization anomalies, only local prediction and local control are performed. Local prediction uses the effective measurement of this node and the boundary of the most recent effective field-level prediction, and cross-node sample stitching is not performed. After the above processing, the data aggregation platform generates updated prediction model parameters, quantile calibration parameters, deviation status identifiers, abnormal operating condition sample identifiers, command execution deviation identifiers, and next cycle control correction identifiers, and sends them to the subsequent data processing, probability prediction, and active power control constraint generation processes. Through the above deviation feedback mechanism, the actual output deviation can enter the corresponding update path according to the time scale and deviation type, so that the prediction model correction, reserve capacity configuration, and single-machine command decomposition form a closed-loop update.

[0035] Example 2: Figure 7 The present invention discloses a wind power processing prediction and control system based on artificial intelligence, comprising: Data aggregation and spatiotemporal mapping module: used to acquire wind farm unit operation data, field meteorological observation data, wide-area meteorological data, superior dispatch data and historical sample data, perform time-series alignment of various types of data, and establish a unified field spatial coordinate mapping relationship; Meteorological correction and feature layering module: It is used to perform field-based correction on wide-area meteorological data based on a unified field area spatial coordinate mapping relationship, and to perform multi-time-scale feature layering processing on the corrected meteorological data and historical power data. Probability prediction generation module: used to generate wind power probability prediction information covering day-ahead, intraday, ultra-short-term and real-time prediction scales based on the hierarchically processed feature data; Regulation constraint mapping module: used to map wind power probability prediction information into active power regulation constraints, and determine reserve capacity based on active power regulation constraints; Instruction generation and decomposition module: used to generate field-level active power instructions based on active power control constraints, reserve capacity, upper-level scheduling plan and unit operating status, and decompose field-level active power instructions into single-unit execution instructions; Deviation closed-loop feedback module: It is used to collect the actual output deviation after the control is executed, feed the actual output deviation back to the corresponding prediction model and control parameters according to the time scale corresponding to the actual output deviation, and update the subsequent prediction and control process according to the deviation amplitude and deviation distribution.

[0036] Example 3: Based on Examples 1 and 2, the specific application process of an artificial intelligence-based wind power forecasting and control method and system is further explained: In a specific application scenario, a centralized wind farm is connected to the regional power grid. The farm is equipped with multiple wind turbine generators, wind measurement towers, lidar, a farm-level data aggregation platform, a farm controller, a prediction model server, a historical sample database, and a dispatch communication interface connected to the regional dispatch center. The data aggregation and spatiotemporal mapping module, the meteorological correction and feature layering module, the probability prediction generation module, the regulation constraint mapping module, the instruction generation and decomposition module, and the deviation closed-loop feedback module are deployed on the farm-level data aggregation platform and the farm controller. The modules interact with each other through a unified timestamp, equipment number, data quality identifier, and prediction scale identifier. After the regional dispatch center issues the power generation plan, reserve requirements, and tie-line power constraints for the future dispatch period, the data aggregation and spatiotemporal mapping module initiates data access processing. It reads the operating data of each wind turbine from the unit monitoring system. This operating data includes at least the fields of active power, reactive power, terminal voltage, terminal current, speed, pitch angle, yaw angle, temperature, vibration, and operating status. Simultaneously, it reads meteorological observation data and wide-area meteorological data from the wind farm, lidar, numerical weather prediction interface, and short-term meteorological interface. The data aggregation and spatiotemporal mapping module performs timestamp verification and resampling on data from different sampling periods, merging the unit's second-level data, meteorological observation data, wide-area meteorological grid data, and dispatch plan data into the target time axis of the corresponding prediction scale. It also establishes a spatial coordinate mapping relationship for the wind farm based on the wind farm's booster station center point, unit coordinates, wind measurement equipment coordinates, and collector line topology. After the above processing, all types of raw data are converted into a unified spatiotemporal dataset and sent to the meteorological correction and feature layering module. After receiving the unified spatiotemporal dataset, the meteorological correction and feature layering module performs field-specific corrections on the wide-area meteorological data according to the site topography, underlying surface roughness, anemometer tower height data, and lidar wind direction. For example, when the wide-area weather forecast indicates a significant increase in wind speed upwind in the site over the next 4 hours, the module combines the actual wind measurement results to correct the hub height wind speed and hub height wind direction, and then layers the corrected meteorological sequence with the historical power sequence across multiple time scales. For slowly changing weather trends, the corresponding features are assigned to the trend layer or the daily and semi-daily cycle layers; for wind speed fluctuations within 1 to 4 hours, the corresponding features are assigned to the hourly fluctuation layer; for gust disturbances within 15 minutes to 1 hour, the corresponding features are assigned to the short-term disturbance layer; and for rapid power changes within 1 to 15 minutes, the corresponding features are assigned to the abrupt change layer. The multi-scale feature matrix after layering is then sent to the probabilistic prediction generation module as input to models at different prediction scales. The probability prediction generation module calls the day-ahead prediction model, intraday prediction model, ultra-short-term prediction model, and real-time prediction model respectively based on the multi-scale feature matrix. The day-ahead prediction model outputs wind power probability prediction information for the next 24 to 72 hours, the intraday prediction model outputs rolling prediction information for the next 4 hours, the ultra-short-term prediction model outputs short-term prediction information for the next 15 to 4 hours, and the real-time prediction model outputs fast-correction prediction information for the next 1 to 15 minutes. The probability prediction generation module not only outputs point value prediction results, but also quantile sequences, prediction means, time-varying covariance matrices, and tail scenario sets. For example, when the ultra-short-term prediction model judges that there is a high probability of power ramp-up within the next 30 minutes, the probability prediction generation module generates a corresponding ramp-up event identifier and gives the probability of occurrence of the ramp-up event, the expected start and end times, and the power change amplitude. If there is a significant deviation between the day-ahead, intraday, and ultra-short-term predictions within the overlapping time period, the probability prediction generation module performs cross-scale consistency verification and forms probability prediction information that can be used for regulation by correcting the near-term period through short-scale prediction and constraining the far-term trend through long-scale prediction. After receiving the probabilistic prediction information, the control constraint mapping module transforms the prediction uncertainty into active power control constraints that the station controller can directly use. During scheduling periods under normal weather conditions, the control constraint mapping module determines the safe range of the station-level active power command based on the quantile sequence. In low-output and high-output scenarios at the tail end, it determines the representative boundaries of low output and high output based on the tail scenario set. When there is a correlation between prediction errors at different times, a correlation robust constraint is constructed using a time-varying covariance matrix. The station controller determines whether the upper-level scheduling plan is within the reachable range based on the above constraints, and determines the second-level response reserve, minute-level rolling reserve, and hour-level planned reserve accordingly. When the prediction results show that there is a high probability of a ramp-up or ramp-down event in the short term, the control constraint mapping module incrementally corrects the reserve capacity based on the probability of the ramp-up event to reduce the risk of reserve redundancy caused by a fixed proportion of long-term conservative reserve. After completing constraint verification, the instruction generation and decomposition module generates field-level active power instructions based on the scheduling plan. When the forecast mean, quantile safety interval, tail representative boundary, and tie-line power limit all meet the scheduling plan, the field-level active power instructions are executed according to the scheduling plan. When the low quantile of the forecast indicates insufficient short-term achievable output of the power station, the instruction generation and decomposition module generates an achievable output correction and sends the reserve requirement to the regional dispatch center. When the high quantile of the forecast indicates a possible exceedance of tie-line power limit or power station output capacity, the instruction generation and decomposition module generates a power curtailment correction and a wind curtailment risk indicator. After the field-level active power instructions are determined, the instruction generation and decomposition module... The system reads the available capacity, real-time wind speed, power curve, health status, wake influence coefficient, and fatigue life remainder of each generating unit, and decomposes the field-level active power command into individual unit execution commands according to the normalized allocation weight. For units with poor health status, significant wake influence, or low fatigue life remainder, their command share is reduced; for units with high available capacity, normal health status, and minimal wake influence, their share is increased. After the individual unit execution command is issued, each unit reports its current achievable upper limit. When there is a situation where the current achievable upper limit of a unit is lower than the issued command, the site controller triggers further decomposition, allocating the difference to other units with higher available capacity. After the control command is executed, the deviation closed-loop feedback module collects the actual active power output at the field level and the actual active power output of a single unit, and divides the actual output deviation into prediction deviation and command execution deviation. The prediction deviation is used to determine whether there is a system offset between the output of the probabilistic prediction model and the actual output, and the command execution deviation is used to determine whether the field-level active power command and the single unit execution command are being executed stably. For real-time deviations on a 30-second scale, the deviation closed-loop feedback module corrects the output bias of the real-time prediction model. For short-term deviations on a 5-minute scale, it corrects the quantile calibration parameters of the ultra-short-term prediction model. For distribution offsets that occur continuously within a 6-hour window, it determines whether the intraday prediction model or the day-ahead prediction model needs to be incrementally updated based on the normalized root mean square error and relative entropy. If a unit cannot track the single unit execution command for multiple consecutive time windows, the deviation closed-loop feedback module reduces the allocation weight of the unit in the next cycle and increases the priority of the unit's available capacity review, thereby feeding back the execution-side deviation to the subsequent single unit decomposition process. When extreme weather, out-of-distribution of training samples, sensor anomalies, on-site communication anomalies, or time synchronization anomalies occur during operation, the deviation closed-loop feedback module does not directly use abnormal samples for online model updates. Instead, it marks the corresponding samples as abnormal operating condition samples and sends low-confidence prediction labels, extreme mode labels, or backup correction labels to the probability prediction generation module and the control constraint mapping module. For example, when the echo quality of the lidar deteriorates due to dense fog, the data aggregation and spatiotemporal mapping module reduces the weight of that data channel in short-term prediction and supplements it with wind tower and adjacent time period trends. When the input features differ too much from the historical training sample distribution, the probability prediction generation module uses a weighted fusion of the continuous method and model prediction to generate temporary prediction output, while the deviation closed-loop feedback module reduces the training weight of abnormal samples. When a continuous communication anomaly occurs in the on-site acquisition link, the site controller only uses local valid measurements for conservative control, and the data aggregation and spatiotemporal mapping module does not use the data of that time period as a model update label. Through the aforementioned collaborative operation process, the data aggregation and spatiotemporal mapping module provides a unified spatiotemporal benchmark for subsequent processing, the meteorological correction and feature layering module provides scaled features for probabilistic prediction, the probabilistic prediction generation module provides prediction results containing distribution information, the regulation constraint mapping module transforms prediction uncertainty into executable regulation boundaries, the instruction generation and decomposition module implements field-level control objectives at the individual unit execution level, and the deviation closed-loop feedback module feeds back the actual execution results to the prediction model and regulation parameters. Thus, wind power prediction, reserve configuration, field-level active power control, and individual unit execution control form a closed-loop prediction and regulation process based on continuous correction of actual output deviation.

[0037] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0038] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination. When implemented in software, the above embodiments can be implemented in whole or in part by a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions of the embodiments of this application are implemented in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted wirelessly or wiredly from one website, computer, server, or data center to another website, computer, server, or data center. Wired methods include optical fiber, twisted pair, coaxial cable, etc. Wireless methods include infrared, microwave, etc. Available media include any available media that can be accessed by a computer or data storage devices such as servers and data centers that contain one or more sets of available media. Available media can be magnetic media (floppy disks, hard disks, magnetic tapes), optical media (DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0039] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0040] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A wind power processing prediction and control method based on artificial intelligence, characterized in that, include: S1. Acquire wind farm unit operation data, field meteorological observation data, wide-area meteorological data, superior dispatch data and historical sample data, perform time-series alignment of various types of data, and establish a unified field spatial coordinate mapping relationship; S2. Perform field-specific correction on wide-area meteorological data, and perform multi-time-scale feature layering processing on the corrected meteorological data and historical power data. S3. Based on the feature data after hierarchical processing, generate wind power probability prediction information covering day-ahead, intraday, ultra-short-term and real-time prediction scales; S4. Map the wind power probability prediction information to active power control constraints, determine the spinning reserve capacity based on the active power control constraints, and generate field-level active power commands and single-unit execution commands. S5. Collect the actual output deviation after the control is executed, feed the actual output deviation back to the corresponding prediction model according to the time scale corresponding to the deviation, and update the subsequent prediction and control process according to the deviation amplitude and deviation distribution.

2. The wind power processing prediction and control method based on artificial intelligence according to claim 1, characterized in that, S1 includes: During the data acquisition process, rolling tasks are triggered according to the forecast scale, and the data is connected to the unit monitoring system, the wind measurement equipment in the field, the wide-area meteorological interface, the dispatch interface and the historical database. The data from different frequencies is timed, resampled, graded for data quality, and anomaly-filled. Based on the equipment space ledger, wide-area meteorological grid points are mapped to representative points in the field area. An aggregated index is established based on the power collection unit, wind direction sector, and unit health status to generate a unified spatiotemporal dataset with timestamps, spatial coordinates, equipment numbers, data quality identifiers, prediction scale identifiers, and data version identifiers.

3. The wind power processing prediction and control method based on artificial intelligence according to claim 1, characterized in that, S2 includes: The availability of meteorological data, power data, and scheduling data is verified, and then the wide-area meteorological data is corrected based on topography, stability, and wind direction residuals to generate hub height wind speed sequence and hub height wind direction sequence. When performing feature layering, a basic feature matrix is ​​constructed using field-level power and corrected meteorological sequences. After modal decomposition, it is divided into trend layer, daily and semi-daily cycle layer, hourly fluctuation layer, short-term disturbance layer, abrupt change layer, and noise layer. The decomposition window and feature weights are adjusted according to the rapid power change state and the state of missing wide-area meteorological data.

4. The wind power processing prediction and control method based on artificial intelligence according to claim 1, characterized in that, S3 includes: Input quality verification is performed on the multi-scale feature matrix; A multi-scale shared feature layer and a scale-segmented output layer are used to call the prediction model corresponding to the prediction scale. Quantile regression is used to generate quantile sequences and predict the mean. The time-varying covariance matrix and tail scene set are generated based on the predicted residuals; The generated results are subjected to cross-scale consistency verification, re-fusion, hill-climbing event identification, and low-confidence correction in the out-of-distribution state of the training samples.

5. The wind power processing prediction and control method based on artificial intelligence according to claim 4, characterized in that, The method employs a multi-scale shared feature layer and a scale-based output layer to call the prediction model corresponding to the prediction scale, including: The features of the trend layer, daily and semi-daily cycle layers, hourly fluctuation layer, short-term disturbance layer and abrupt change layer are encoded according to the time scale. Historical sample tags, unit availability status, power curtailment status, and field area spatial index are fused on the field-level time axis; Matching features are selected according to the prediction scale and sent to the day-ahead, intraday, ultra-short-term, and real-time prediction output channels.

6. The wind power processing prediction and control method based on artificial intelligence according to claim 1, characterized in that, S4 includes: Opportunity constraints, tail risk boundaries, and correlation robustness constraints are constructed based on quantile sequences, tail scenario sets, and time-varying covariance matrices. Configure standby capacity according to response scale and the probability of ramp events; The field-level active power command is generated with the scheduling plan, tie line limits, reserve capacity, and achievable output as boundaries. Decompose the single-machine execution instructions according to the single-machine availability, health status, fatigue life and wake effect; The instruction correction branch is determined by combining the reachable upper limit check, further decomposition, and communication anomaly status.

7. The wind power processing prediction and control method based on artificial intelligence according to claim 6, characterized in that, Decompose the single-machine execution instructions according to the single-machine availability, health status, fatigue life, and wake effect, including: The usable capacity of a single unit is determined by the nameplate capacity, real-time wind speed, and power curve. Adjust the available capacity of a single unit based on the unit's alarm status, maintenance status, fault shutdown status, and grid connection status; The proportion of the instruction is limited based on the remaining fatigue life; the wake influence coefficient is calculated based on the field coordinates and the direction of the incoming wind; Normalized allocation weights are constructed using the available capacity of a single unit, the remaining fatigue life, and the wake influence coefficient, and field-level active power commands are allocated into active power commands, pitch coordination commands, and yaw coordination commands.

8. The wind power processing prediction and control method based on artificial intelligence according to claim 1, characterized in that, S5 includes: The actual output deviation is divided into prediction deviation and command execution deviation; Deviation data is collected in seconds, minutes, and long time intervals; Incremental updates to partitions are triggered based on quantile out-of-bounds, normalized root mean square error, relative entropy, and bias moving average. The prediction bias is fed back to the prediction model parameters, and the command execution bias is fed back to the available capacity assessment, single-machine weight allocation and control constraint correction parameters. Sample labeling, training weight adjustment, and downgrade fusion processing are performed on retraining trigger states, extreme weather, out-of-distribution states, and data credibility states.

9. The wind power processing prediction and control method based on artificial intelligence according to claim 8, characterized in that, The prediction bias is fed back to the prediction model parameters, and the command execution bias is fed back to the available capacity assessment, individual machine weight allocation, and regulation constraint correction parameters, including: Adjust the output bias parameters of the real-time prediction model, the quantile calibration parameters of the ultra-short-term prediction model, and the mid-frequency characteristic parameters of the intraday and day-ahead prediction models according to the deviation time scale. Determine the parameter update step size based on the deviation amplitude; Determine the quantile calibration direction based on the direction of quantile crossover; Based on the deviation of the single-unit instruction, the available capacity of the unit is updated and the weight of the single unit is allocated. The active power constraint parameters of the field level and the control parameters of the next cycle are also adjusted in conjunction.

10. An artificial intelligence-based wind power treatment prediction and control system, used to implement the artificial intelligence-based wind power treatment prediction and control method according to any one of claims 1-9, characterized in that, include: Data aggregation and spatiotemporal mapping module: used to acquire wind farm unit operation data, field meteorological observation data, wide-area meteorological data, superior dispatch data and historical sample data, perform time-series alignment of various types of data, and establish a unified field spatial coordinate mapping relationship; Meteorological correction and feature layering module: It is used to perform field-based correction on wide-area meteorological data based on a unified field area spatial coordinate mapping relationship, and to perform multi-time-scale feature layering processing on the corrected meteorological data and historical power data. Probabilistic prediction generation module: used to generate wind power probabilistic prediction information covering day-ahead, intraday, ultra-short-term and real-time prediction scales based on the hierarchically processed feature data; Regulation constraint mapping module: used to map wind power probability prediction information into active power regulation constraints, and determine reserve capacity based on active power regulation constraints; Instruction generation and decomposition module: used to generate field-level active power instructions based on active power control constraints, reserve capacity, upper-level scheduling plan and unit operating status, and decompose field-level active power instructions into single-unit execution instructions; Deviation closed-loop feedback module: It is used to collect the actual output deviation after the control is executed, feed the actual output deviation back to the corresponding prediction model and control parameters according to the time scale corresponding to the actual output deviation, and update the subsequent prediction and control process according to the deviation amplitude and deviation distribution.

Images