Typhoon path-oriented offshore wind power prediction method and system

Abstract

The invention discloses a typhoon path-oriented offshore wind power prediction method and system, and the method comprises the steps: collecting typhoon path prediction data, offshore wind plant fan monitoring data and weather prediction data, forming a preprocessing data set, reducing the influence of communication jitter and missing measurement abnormity on input during a typhoon, and improving the prediction stability; typhoon path driving features and unit operation features are further extracted, a target feature set is obtained according to a preset fusion rule, external disturbances such as typhoon approaching and turning and unit operation constraints are subjected to correlation expression in the same feature space, and the sensitivity to non-stable working conditions is enhanced; inputting the target feature set into a control state transfer prediction model to output a control state probability sequence, describing limited power by using duration constraint, switching out state evolution of shutdown and climbing recovery, and reducing switching lag; and through shutdown threshold zero setting and climbing threshold amplitude limiting correction, power cutoff is suppressed and overshoot is recovered, so that an error peak value is reduced and scheduling availability is improved.

Description

Technical Field

[0001] This invention relates to the technical field of offshore wind power prediction in power system operation and new energy grid connection, and particularly to a method and system for offshore wind power prediction based on typhoon paths. Background Technology

[0002] With the continuous growth of offshore wind power installed capacity and the increasing grid connection rate, ultra-short-term and short-term power forecasting has gradually evolved from auxiliary references to a crucial foundation for grid dispatching, reserve configuration, offshore power collection, and safety verification of transmission channels. Existing offshore wind power forecasting technologies typically rely on physical mechanism prediction driven by numerical weather forecasting and data-driven prediction: the former depends on wind field forecasts, power curves, and turbine state constraints, and is interpretable; the latter uses multi-source monitoring and meteorological feature learning nonlinear mapping, and has strong fitting capabilities. However, in typhoon path impact scenarios, external disturbances such as wind speed, wind direction, gusts, and turbulence change rapidly with the relative position of the typhoon center, and offshore turbines exhibit a "power limiting—shutdown—recovery" pattern under protection strategies. The discrete control mode switching of "re-climbing" causes non-stationary characteristics in power response, such as abrupt changes, intermittent changes, and limited climbing. Traditional schemes that directly regress power based on wind speed or meteorological factors are prone to prediction lag and error peaks during the shutdown and recovery phases. At the same time, multi-source data are prone to communication jitter, missing data, and outliers during typhoons. Without a unified timestamp alignment, missing data completion, and outlier removal mechanism, inconsistent model inputs will further amplify the prediction instability under extreme weather conditions. Therefore, it is urgent to construct a unified input that integrates "path-driven characteristics and unit operation characteristics" for typhoon path scenarios, and to use the switching of unit control strategies as a key intermediate variable in the prediction link to form a power prediction method that is more sensitive to shutdown timing and recovery climbing constraints.

[0003] CN115360704A discloses a method for predicting the output of offshore wind power. This method focuses on the continuous mapping of "wind speed-output". It does not incorporate the unit protection control mode (such as shutdown and recovery ramp) into the prediction framework as a recursive state transition process, nor does it provide shutdown zeroing and ramp limit rules driven by control state probability. Therefore, when the approach of a typhoon causes frequent switching of the unit control mode, there may still be problems such as lag in identifying the shutdown time point, overestimation of the power ramp amplitude during the recovery phase, or aggravated fluctuations. It is difficult to directly suppress the error peak near the shutdown / recovery.

[0004] CN113962433A discloses a wind power prediction method and system that integrates causal convolution and separable temporal convolution. This method belongs to the data-driven power regression prediction paradigm. On the one hand, it does not construct a discrete control state set and its duration evolution mechanism for the sudden changes in external disturbances caused by typhoon paths and the switching of unit protection strategies. On the other hand, it does not provide a structured processing to introduce "control state probability sequence - threshold comparison correction" into the prediction link to form a pre-constraint on the shutdown zeroing and recovery ramping upper limit. As a result, in the typhoon scenario, the model may tend to extrapolate continuous power based on historical statistical relationships, making it difficult to make stable and interpretable prediction corrections for "power cliff caused by shutdown" and "upper limit of power rise slope caused by limited recovery ramping".

[0005] In summary, existing offshore wind power output / power prediction technologies for typhoon scenarios generally suffer from the following shortcomings: First, typhoon path information mostly remains at the level of wind speed or wind field prediction input, lacking the design of intermediate variables to introduce unit control mode switching into the prediction, resulting in prediction lag and error peaks that are difficult to suppress during the shutdown and recovery ramp-up phases; Second, during typhoons, the problems of missing, abnormal, and asynchronous multi-source monitoring and meteorological data are prominent. Without standardized preprocessing such as unified timestamp alignment, missing data completion, and anomaly removal, input drift and prediction instability are likely to occur.

[0006] To address the aforementioned issues, this invention proposes a method for predicting offshore wind power based on typhoon paths: A preprocessed dataset is formed by aligning typhoon path forecast data, offshore wind farm turbine monitoring data, and meteorological forecast data with unified timestamps, completing missing data, and removing anomalies. Based on this, typhoon path driving features and turbine operation features are extracted and normalized and fused according to preset fusion rules to obtain a target feature set. Furthermore, the target feature set is input into a control state transition prediction model constructed based on a semi-Markov control state transition algorithm, outputting a control state probability sequence within the prediction window. The initial value of the predicted power sequence is zeroed and the ramp-up increment is corrected according to a threshold comparison rule between the probability of shutdown and the probability of recovery ramp-up, thereby improving the prediction consistency of power constraints at shutdown time and during the recovery phase for typhoon path scenarios. Summary of the Invention

[0007] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0008] In view of the aforementioned existing problems, the present invention is proposed.

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for predicting offshore wind power based on typhoon paths, comprising: collecting typhoon path forecast data, offshore wind farm turbine monitoring data and meteorological forecast data, and aligning, filling in missing data and removing anomalies according to a unified timestamp to obtain a preprocessed dataset; Based on the preprocessed dataset, typhoon path-driving features and unit operation features are extracted, and the two types of features are normalized and spliced ​​together according to preset fusion rules to obtain the target feature set. The target feature set is input into the control state transition prediction model constructed based on the semi-Markov control state transition algorithm, and the control state probability sequence at each time point within the prediction window is output. Based on the control state probability sequence and the target feature set to be predicted, an initial value of the predicted power sequence is generated according to a preset power prediction mapping rule, and a threshold comparison correction is performed: When the probability of switching out of the shutdown state is greater than or equal to the preset shutdown threshold, the predicted power at the corresponding moment is set to zero. When the probability of resuming the climbing state is greater than or equal to the preset climbing threshold, the predicted power at the corresponding time is incrementally limited according to the preset climbing rate upper limit. Output the predicted power sequence after threshold comparison correction; The target feature set to be predicted is selected from the target feature set according to the prediction time window division rule.

[0010] Secondly, the present invention provides an offshore wind power prediction system for typhoon paths, comprising: a data acquisition module for acquiring typhoon path forecast data, offshore wind farm turbine monitoring data, and meteorological forecast data; The preprocessing module aligns the typhoon path forecast data, the offshore wind farm turbine monitoring data, and the weather forecast data according to a unified timestamp, completes missing data, and removes anomalies, and outputs a preprocessed dataset. The feature extraction module extracts typhoon path-driven features and unit operation features based on the preprocessed dataset, and normalizes and merges the two types of features according to preset fusion rules to output the target feature set. The control state prediction module inputs the target feature set into the control state transition prediction model constructed based on the semi-Markov control state transition algorithm, and outputs the control state probability sequence at each time point within the prediction window. The time window selection module selects a set of target features to be predicted from the target feature set that corresponds to the prediction time window according to the prediction time window division rules; The initial power value generation module generates an initial value for the predicted power sequence according to the control state probability sequence and the target feature set to be predicted, based on a preset power prediction mapping rule. The threshold correction module performs threshold comparison correction on the initial value of the predicted power sequence. When the probability of switching out of the shutdown state is greater than or equal to the preset shutdown threshold, the predicted power at the corresponding time is set to zero. When the probability of resuming the ramp state is greater than or equal to the preset ramp threshold, the predicted power at the corresponding time is incrementally limited according to the preset ramp rate upper limit, and the predicted power sequence after threshold comparison correction is output.

[0011] Preferably, the system further includes one or more processors; The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the offshore wind power forecasting method oriented towards typhoon paths as described above.

[0012] Thirdly, the present invention provides a computer-readable medium for storing software, characterized in that: the software includes instructions executable by one or more computers, the instructions causing the one or more computers to perform operations, the operations including the flow of the aforementioned method for predicting offshore wind power oriented towards typhoon paths.

[0013] The beneficial effects of this invention are as follows: This invention collects typhoon path forecast data, offshore wind farm turbine monitoring data, and meteorological forecast data, and forms a preprocessed dataset by aligning it with a unified timestamp, filling in missing data, and removing anomalies. This reduces the impact of communication jitter and missing data anomalies during typhoons on the input, improving prediction stability. Furthermore, it extracts typhoon path driving features and unit operation features from the preprocessed dataset, normalizes and fused them according to preset fusion rules to obtain a target feature set. This allows external disturbances such as typhoon approach and turning to be associated with unit operation constraints within the same feature space, enhancing sensitivity to non-stationary operating conditions. The target feature set is then input into a control state transition prediction model built based on a semi-Markov control state transition algorithm, outputting a control state probability sequence. Duration constraints are used to describe the state evolution of power limiting, shutdown, and recovery ramping, reducing switching lag. Finally, based on the control state probability sequence and the target feature set to be predicted, an initial predicted power value is generated. Power cliffs and recovery overshoots are suppressed by setting the shutdown threshold to zero and the ramping threshold to limit the amplitude, thereby reducing error peaks and improving scheduling availability. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1This is a flowchart illustrating the offshore wind power prediction method based on typhoon paths as described in this invention. Figure 2 This is a schematic diagram of the module structure distribution of the offshore wind power prediction system oriented towards typhoon paths as shown in this invention. Detailed Implementation

[0015] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0016] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0017] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0018] According to an embodiment of the present invention, in combination Figure 1 The flowchart shown illustrates a method for predicting offshore wind power based on typhoon paths, which specifically includes the following steps: S1. Collect typhoon path forecast data, offshore wind farm turbine monitoring data, and weather forecast data, and align, imput missing data, and remove anomalies according to a unified timestamp to obtain a preprocessed dataset. Note the following in this step: S1.1 Collect typhoon path forecast data, meteorological forecast data, and offshore wind farm turbine monitoring data, and write the collection timestamp and data source identifier for each typhoon path forecast data, meteorological forecast data, and offshore wind farm turbine monitoring data to obtain the original data set.

[0019] It should be noted that typhoon track forecast data is provided by the typhoon track operational release terminal, providing the typhoon center latitude and longitude, direction of movement, speed of movement, minimum central pressure, maximum wind speed near the center, radius of the 7-level wind circle, and radius of the 10-level wind circle for multiple valid moments in the future, according to the forecast release time; meteorological forecast data is provided by the marine grid forecast or wind farm numerical forecast terminal, and at least includes the wind speed forecast and wind direction forecast values ​​corresponding to the hub height of the wind farm, and may include the sea level pressure forecast value and temperature forecast value for power curve correction; offshore wind farm turbine monitoring data is provided by the wind farm SCADA system, and at least includes the turbine number, hub height wind speed, wind direction, pitch angle, yaw angle, speed, active power, nacelle vibration, and control mode code.

[0020] Preferably, in this embodiment, the collection timestamp for each piece of data is written in a uniform format, such as YYYY-MM-DD hh:mm:ss (UTC+8), and a data source identifier is written to represent the source channel and version; an example of the data source identifier is: typhoon path forecast data is written to TYP_TRACK_A, meteorological forecast data is written to MET_FCST_B, and wind turbine monitoring data is written to SCADA_FARM_01.

[0021] It should also be noted that this embodiment only selects typhoon path forecast data, meteorological forecast data, and offshore wind farm turbine monitoring data for processing. The reason is that power changes during a typhoon are jointly dominated by the evolution of the external wind field and the switching of the unit control mode. The evolution of the external wind field can be given by both typhoon path forecast and meteorological forecast, while the switching of the unit control mode and the unit response are given by SCADA monitoring data. Compared with the existing technology that directly regresses power using only meteorological forecasts, this embodiment incorporates the geometric relationship of the typhoon path and the persistence of the control state into the same link, so that the power limiting, shutdown, and ramp-up phases under the same wind speed conditions have distinguishable data basis.

[0022] S1.2. Generate a unified timestamp index based on the original data set, and map the typhoon path forecast data, meteorological forecast data and offshore wind farm turbine monitoring data to the unified timestamp index to obtain an aligned data set.

[0023] In a preferred embodiment, when generating a unified timestamp index based on the original data set, the alignment time range is first determined, with the starting point being the maximum value of the earliest available valid time of each data source and the ending point being the minimum value of the latest available valid time of each data source; then the time step is determined, which is consistent with the statistical period of the wind turbine monitoring data, such as 10 minutes, thereby obtaining a unified timestamp index sequence with one index point every 10 minutes.

[0024] Furthermore, when mapping typhoon track forecast data to a unified timestamp index, the valid time is used as the matching key: if the typhoon track forecast itself provides valid times in 1-hour increments, then the latitude, longitude, and intensity parameters of the typhoon center for the same hour are written to multiple unified timestamp index points falling within the same hour, and the valid time field from the typhoon track forecast is retained for backtracking; when mapping meteorological forecast data to a unified timestamp index, the valid times of the hub height wind speed forecast value and the wind direction forecast value are matched; when mapping offshore wind farm turbine monitoring data to a unified timestamp index, the SCADA recording time is merged into the nearest index point; for example, records with a difference of no more than 5 minutes from the index point are merged into that index point, records with a difference of more than 5 minutes are merged into the adjacent index point, and the original recording time field is retained.

[0025] For example, if the unified timestamp index contains 2025-08-18 02:00:00, 02:10:00, and 02:20:00, and the typhoon track forecast gives a valid value from 02:00 to 03:00 at 2025-08-18 02:00:00, then the above three index points are written with the same set of typhoon center latitude, longitude and intensity parameters, and their source valid time is written as 2025-08-18 02:00:00.

[0026] S1.3 Perform missing data completion processing on the aligned data set. Specifically, for data segments with consecutive missing lengths not exceeding a preset missing length threshold, interpolation is performed to complete the missing data segments based on adjacent valid sampling points. For data segments with consecutive missing lengths exceeding the preset missing length threshold, the missing data segments are marked as missing segments and the missing identifier is retained, thus obtaining the completed data set.

[0027] In a preferred embodiment, the missing data determination is based on a unified timestamp index: when a data field under a certain timestamp index is null, or the original record time corresponding to the field is missing and the value cannot be obtained according to the mapping rules, it is determined that the field is missing at that index point; the length of consecutive missing data is measured by the number of consecutive missing index points.

[0028] For example, in this embodiment, the preset missing length threshold is set to 3 index points, corresponding to 30 minutes (index step size of 10 minutes). The reason is that: the missing data of offshore wind farm SCADA and short-term forecasts under communication jitter and short-term maintenance often manifests as the missing of several adjacent index points. Such missing data can be filled by adjacent valid information without changing the trend. When the continuous missing data exceeds 30 minutes, the missing data is often related to communication interruption, equipment shutdown or forecast data gap. Direct interpolation is likely to introduce a false trend that is inconsistent with the typhoon process. Therefore, it is marked as a missing segment and the missing data is retained.

[0029] Preferably, the rule for determining adjacent valid sampling points in this embodiment is as follows: for a field to be filled, the nearest non-missing index point is retrieved forward from the left side of the missing segment as the left valid point, and the nearest non-missing index point is retrieved backward from the right side of the missing segment as the right valid point; if both left and right valid points exist and the missing segment is contained between them, they are considered to be adjacent valid sampling points; when the missing segment is located at the sequence boundary, resulting in only one valid point on one side, no interpolation is performed to fill in the missing points, and they are directly marked as missing segments.

[0030] Furthermore, in this embodiment, the interpolation completion method is differentiated according to field type: for continuous quantity fields (hub height wind speed, pitch angle, yaw angle, speed, active power, nacelle vibration, typhoon center latitude and longitude, minimum central pressure, and maximum near-center wind speed), linear transformation completion is performed on adjacent valid points; for angle fields, wind direction and typhoon movement direction are first transformed into a continuous angle sequence according to the angle circling rule before linear transformation completion is performed, and after completion, the angles are merged back into the range of 0 to 360 degrees; for discrete field control mode codes, no linear transformation completion is performed, but when the continuous missing length does not exceed the preset missing length threshold, the control mode code of the left valid point of the missing segment is taken and written to each index point of the missing segment, and the written record is marked as obtained by completion.

[0031] As an example, if the vibration of the nacelle of a wind turbine with the number WT_05 is null at the four index points of 02:00, 02:10, 02:20, and 02:30, then the consecutive missing length is 4, which exceeds the threshold of 3, and is determined to be a missing segment. The missing identifier corresponding to the vibration of the WT_05 nacelle at these four index points is written with a missing identifier of 1, and this field is kept as null in the preprocessed dataset. At the same time, the start and end indexes of the missing segment 02:00 to 02:30 are retained.

[0032] S1.4 Perform anomaly removal processing on the completed data set, wherein data points exceeding the preset physical range are removed according to preset physical constraint rules, and anomaly identifiers are written for the removed data points; the completed data set, missing identifiers and anomaly identifiers are summarized to generate a preprocessed dataset.

[0033] In a preferred embodiment, anomaly removal is performed based on preset physical constraint rules, which at least include value range constraints and jump constraints. Examples of value range constraints include: hub height wind speed 0 to 70 m / s; wind direction 0 to 360 degrees; propeller pitch angle -5 to 90 degrees; yaw angle -180 to 180 degrees; engine speed 0 to 20 rpm; active power 0 to rated power (e.g., rated power 10 MW); nacelle vibration 0 to 50 mm / s; and during a typhoon... The minimum central pressure is 850 to 1020 hPa; the maximum wind speed near the center is 0 to 80 m / s; the typhoon center's latitude and longitude are within the forecast sea area; examples of jump constraints are: the change in hub height wind speed between two adjacent unified timestamp index points does not exceed 20 m / s, the change in pitch angle does not exceed 30 degrees, the change in active power does not exceed 0.6 of the rated power, and the change in yaw angle does not exceed 60 degrees; if the above thresholds are exceeded and there is a lack of corresponding control mode code switching records, it is judged as an anomaly.

[0034] It should be noted that although there are rapid wind speed changes and control actions during a typhoon, the physical changes under the SCADA statistical period are still limited by the inertia of the unit and the rate of control action. Identifying large jumps that are obviously out of bounds or have no control basis as anomalies can reduce the interference of anomalies on subsequent feature weights and control state probability recursion. Compared with the existing technology that only performs statistical filtering without setting physical constraints, this embodiment introduces the feasible operating range of the unit and the rate of control action into the anomaly elimination rules, so that the anomaly identification has engineering interpretability and can directly participate in the weight reduction of S2.4.

[0035] In an optional implementation, when writing an anomaly identifier to the excluded data point, the anomaly identifier also adopts a feature component level identifier, whose index is determined together with the unified timestamp index and the fan number; for example, if the rotational speed of WT_03 is determined to be out of range at 02:10, then the anomaly identifier of WT_03 rotational speed@02:10 is written as 1, and the rotational speed value is set to null, while its original value is retained in the anomaly original value field for traceability.

[0036] Preferably, in this embodiment, the typhoon path field, weather forecast field and each wind turbine monitoring field under each unified timestamp index are combined into a single record, and two types of identifiers are added to the record: one is a missing identifier sequence corresponding to each field, and the other is an anomaly identifier sequence corresponding to each field, thereby forming a preprocessed dataset record structure of numerical field + missing identifier sequence + anomaly identifier sequence.

[0037] S2. Extract typhoon path-driving features and unit operation features from the preprocessed dataset, and normalize and fuse the two types of features according to preset fusion rules to obtain the target feature set. Note that the following points should be noted in this step: S2.1 Read the preprocessed dataset and extract the typhoon center location, typhoon movement direction and typhoon intensity parameters at each time step according to the unified timestamp index. Calculate the typhoon's relative distance, relative azimuth and approach speed based on the wind farm's geographical coordinates to obtain the typhoon path-driven feature set.

[0038] In a preferred embodiment, when extracting the typhoon center location, typhoon movement direction and typhoon intensity parameters at each time step according to the unified timestamp index, the aligned data set obtained in S1.2 is used as input. Under each unified timestamp index, the typhoon center latitude, typhoon center longitude, typhoon movement direction angle, typhoon movement speed, minimum central pressure, maximum near-center wind speed, radius of the 7-level wind circle, and radius of the 10-level wind circle are read, and the wind circle coverage rate is further calculated.

[0039] For example, the wind circle coverage rate is calculated as follows: the radius of the influence range is set with the center point of the wind farm as the center (e.g., 150 kilometers), the area ratio of the seventh-level wind circle covering the diameter of the influence range is determined, and the ratio is written into the typhoon path driving feature set.

[0040] It should be noted that when calculating the relative distance, relative azimuth, and approach velocity of a typhoon based on the geographical coordinates of the wind farm, the great circle distance and azimuth are used for calculation, and the rate of change of distance is calculated at adjacent index points; the formula is as follows: in, R represents the relative distance of the typhoon at a unified timestamp index t; R is the Earth's radius. The latitude of the wind farm center; Longitude of the wind farm center; Latitude of the typhoon center; Longitude of the typhoon center; This refers to the relative azimuth of the typhoon. This is the two-parameter form of the arctangent function; To approximate velocity; The time interval between adjacent unified timestamp indices.

[0041] In this embodiment, the latitude and longitude of the wind farm center point are given by the wind farm mapping coordinates, such as 26.30°N, 119.90°E; the unified timestamp index step size is 10 minutes, then... Take 600 seconds; when A positive value indicates that the distance from the typhoon center is decreasing and the approaching trend is strengthening.

[0042] S2.2. Based on the preprocessed dataset, extract wind turbine hub height wind speed, wind direction, pitch angle, yaw angle, speed, active power, nacelle vibration and control mode code at each time step. Then, aggregate the wind turbine hub height wind speed, wind direction, pitch angle, yaw angle, speed, active power, nacelle vibration and control mode code by wind turbine number to obtain the unit operation feature set.

[0043] In a preferred embodiment, when extracting wind turbine hub height wind speed, wind direction, pitch angle, yaw angle, speed, active power, nacelle vibration and control mode code at each time step, the preprocessed dataset obtained in S1 is used as input, and the corresponding fields are read according to the wind turbine number under each unified timestamp index. When there are multiple original records for the same wind turbine at the same index point, the continuous quantity field takes the average value within the index point, the nacelle vibration can be written with both the average value and the maximum value at the same time, and the control mode code takes the mode code with the longest duration within the index point, and the duration percentage of the mode code is retained as an additional field.

[0044] Furthermore, when aggregating fields by turbine number, this embodiment arranges the turbine numbers in ascending order, such as "WT_01…WT_50", and sequentially writes the hub height, wind speed, wind direction, yaw angle, pitch angle, speed, nacelle vibration, control mode code, and active power corresponding to the previous unified timestamp index for each turbine under each unified timestamp index. The active power corresponding to the previous unified timestamp index is directly read from the historical index point of the same turbine. If the previous index point is missing, the component is written with a null value and a missing flag.

[0045] S2.3. Normalize the typhoon path driving feature set and the unit operation feature set respectively. In this process, the continuous quantities are scaled according to the preset dimension range, and the control mode code is discretely encoded according to the preset encoding rules to obtain the normalized typhoon path driving feature set and the normalized unit operation feature set.

[0046] In a preferred embodiment, when performing interval scaling on continuous quantities according to a preset dimensional range, a dimensional range is preset for each continuous quantity field. For example, the relative distance to a typhoon is 0 to 800 kilometers, the approach speed is -20 to +20 meters per second, the wind speed at hub height is 0 to 70 meters per second, the pitch angle is -5 to 90 degrees, the rotational speed is 0 to 20 revolutions per minute, the active power is 0 to the rated power, and the nacelle vibration is 0 to 50 millimeters per second. The interval scaling adopts the rule of mapping the values ​​to 0 to 1 according to the lower and upper limits of the field. When the value exceeds the range, it is truncated at the boundary, and its out-of-bounds record is retained in the anomaly identifier.

[0047] In a preferred embodiment, when discretely encoding the control mode code according to a preset encoding rule, a merge table of mode code and control state is first established. For example: when the control mode code belongs to the conventional grid-connected power generation category, the discrete code is written as 0; when the control mode code belongs to the limited power generation or derating category, the discrete code is written as 1; when the control mode code belongs to the yaw-limited or yaw-locked category, the discrete code is written as 2; when the control mode code belongs to the cut-out shutdown or overspeed protection shutdown category, the discrete code is written as 3; when the control mode code belongs to the grid-connected recovery and ramping category, the discrete code is written as 4. When multiple modes coexist at the same index point, the discrete code of the index point is determined according to the priority: "cut-out shutdown state is higher than recovery ramping state, higher than yaw-limited state, higher than power-limited state, higher than normal state".

[0048] S2.4. According to the preset fusion rules, the normalized typhoon path-driven feature set and the normalized unit operation feature set are concatenated at the same timestamp, and the missing and abnormal indicators in the preprocessed dataset are written into the concatenation result as input items for the fusion weight to obtain the target feature set.

[0049] Specifically, the preset fusion rules include: pairing the normalized typhoon path-driven feature set and the normalized unit operation feature set one-to-one according to the same unified timestamp index; under each unified timestamp index, concatenating the normalized typhoon path-driven feature set and the normalized unit operation feature set according to a preset feature arrangement order, and generating a feature validity identifier sequence corresponding to the concatenation result; generating a feature weight sequence based on missing and anomaly identifiers in the preprocessed dataset, wherein the feature components corresponding to missing identifiers are weighted by a first attenuation coefficient, and the feature components corresponding to anomaly identifiers are weighted by a second attenuation coefficient, and the second attenuation coefficient is less than the first attenuation coefficient; weighting the concatenation result according to the feature weight sequence, and writing the feature validity identifier sequence into the weighted result to obtain the target feature set.

[0050] The preset feature arrangement order includes: first, writing the normalized typhoon path-driven feature set, which includes the typhoon relative distance, relative azimuth, approach speed, typhoon movement direction, typhoon intensity parameters, and wind circle coverage; then, writing the normalized unit operation feature set, and writing the hub height wind speed, wind direction, yaw angle, pitch angle, speed, nacelle vibration, control mode code, and active power corresponding to the previous unified timestamp index of each wind turbine in ascending order of turbine number.

[0051] In a preferred embodiment, when pairing the two types of normalized features one-to-one, a pairing record of "typhoon path-driven feature vector @ index point" and "unit operation feature vector @ index point" is generated using a unified timestamp index as the key. When splicing the features according to a preset feature arrangement order under each unified timestamp index, "typhoon relative distance, relative azimuth, approach speed, typhoon movement direction, typhoon intensity parameters, and wind circle coverage" are written first, followed by the feature components of each wind turbine arranged in ascending order of turbine number, forming a single spliced ​​feature vector.

[0052] When generating the feature validity identifier sequence, this embodiment maps the missing identifier obtained in S1.3 and the abnormal identifier obtained in S1.4 to the concatenated feature vector in a one-to-one correspondence order of feature components, and generates a validity identifier with the same index as each component.

[0053] For example, the validity identifier takes three values: when the component has neither a missing identifier nor an anomaly identifier, write "validity identifier is 2"; when the component has a missing identifier and no anomaly identifier, write "validity identifier is 1"; when the component has an anomaly identifier, write "validity identifier is 0"; when both a missing identifier and an anomaly identifier appear, write 0 according to the anomaly identifier.

[0054] Furthermore, when generating the feature weight sequence, this embodiment sets the basic weight of each component to 1; the weight of the component with a validity identifier of 1 is reduced by a first attenuation coefficient, such as 0.6; the weight of the component with a validity identifier of 0 is reduced by a second attenuation coefficient, such as 0.3.

[0055] For example, if WT_05 speed@02:10 is judged to be abnormal, its validity label is 0 and its weight is 0.3; if WT_12 engine room vibration@02:20 has a short gap and has been filled, its validity label is 1 and its weight is 0.6; the validity labels of other components without gaps or abnormalities are 2 and their weights are 1.

[0056] S3. Input the target feature set into the control state transition prediction model constructed based on the semi-Markov control state transition algorithm, and output the control state probability sequence at each time point within the prediction window. Note that the following should be noted in this step: S3.1 Read the target feature set and select the target feature set to be predicted corresponding to the prediction window according to the prediction window division rule to obtain the prediction input feature set.

[0057] It should be noted that the prediction window division rules in this embodiment are bound to the engineering update cycle: the unified timestamp index step size is consistent with the SCADA statistical cycle, such as 10 minutes; the typhoon path forecast and weather forecast are updated every 3 hours, so the most recent unified timestamp index corresponding to the release time of each forecast is used as the prediction starting point, and the prediction duration is set, for example, the next 24 hours, resulting in a prediction window containing 144 index points.

[0058] When selecting the target feature set to be predicted according to the prediction window division rule, the target feature set obtained in S2 is used as input, and the target feature records within the range from the prediction start point to the prediction end point are extracted to form the target feature set to be predicted; and the target feature set to be predicted is written into the input end of the control state transition prediction model as the prediction input feature set.

[0059] S3.2 In the control state transition prediction model, a set of control states is set. The set of control states includes at least the normal state, power-limited state, yaw-limited state, cut-out stop state, and recovery ramp state. The initial transition parameters and duration parameters of each control state are generated based on the changes in the control mode codes in the historical operation records.

[0060] In a preferred embodiment, the control state set is set to five categories: normal state, power-limited state, yaw-limited state, cut-out stop state, and recovery climb state; the control state and the control mode code are consistent with the discrete encoding merge table in S2.3, thereby ensuring that the control state sequence can be directly obtained from the historical control mode code sequence.

[0061] When generating initial transfer parameters, a historical typhoon process sample segment is selected. An example of the selection criteria for the sample segment is: the relative distance to the typhoon gradually decreases from 500 kilometers to a minimum and then increases again, and the wind farm experiences at least one power-limiting state or a cut-out shutdown state. For each sample segment, the discretely encoded control states are arranged into a state sequence in chronological order. The number of transitions between adjacent states is counted, and the initial transfer parameters are obtained by normalizing according to the initial state. To reflect the difference between the approach of the typhoon and the period before and after landfall, this embodiment further divides the sample segment into a far-distance segment, a near-distance segment, and a pre- and post-landfall segment according to the relative distance to the typhoon, and counts the transfer parameters for each segment. This results in the probability of the cut-out shutdown state entering the near-distance segment and the duration parameters being different from those in the far-distance segment.

[0062] When generating duration parameters, the number of continuous index points for each control state during the continuous holding period is counted for the above historical state sequence, and converted into duration samples according to a unified timestamp step size; then the duration samples are counted into discrete duration probability tables, corresponding to five types of control states, and stored in segments according to typhoon stages.

[0063] For example, before and after landfall, the duration of the shutdown state is mostly concentrated in the range of 6 to 10 hours, while the duration of the recovery and climbing state is mostly concentrated in the range of 1 to 3 hours. After this difference is written into the duration probability table, the semi-Markov recursion no longer approximates the state duration as a geometric distribution, thus better reflecting the continuous restriction and continuous shutdown phenomenon under typhoon process compared with the general Markov state transition.

[0064] S3.3 Input the predicted input feature set into the control state transition prediction model, calculate the control state observation probability at each time step according to the correspondence between the predicted input feature set and the control state set, and recursively update the model by combining the initial transition parameters and duration parameters to obtain the control state posterior probability at each time step.

[0065] In a preferred embodiment, when calculating the observation probability of the control state at each time step according to the correspondence between the predicted input feature set and the control state set, feature distribution parameters are first established for each control state. The parameters are derived from the target feature records of the historical sample segment. For example, a multidimensional normal distribution is set for each control state, and the mean vector and covariance matrix are estimated from the historical feature samples under that control state. Thus, given the target feature vector at a certain time step, the observation likelihood value under each control state at that time step can be calculated.

[0066] For example, let the unified timestamp index be t, the control state be j, and the observation vector be t. Then the observed likelihood is written as: In a semi-Markov recursion, a duration variable is introduced. The discrete duration probability table gives the duration of each control state during continuous operation. The probability at index point is then recursively updated as follows: And at every moment Normalizing the control state set, we obtain the posterior probability of the control state at each time step. The posterior probability is then written as: in, The input feature vector for prediction at time t with a unified timestamp index; This is the mean vector when the control state is j; The covariance matrix when the control state is j The observation likelihood at time t when the control state is j; For recursive variables; These are the transition parameters for transitioning from control state i to control state j; For control state j to continue The duration probability of each index point; The maximum number of index points for the duration; Let be the posterior probability of control state j at time t; m is the control state index variable.

[0067] It should be noted that in this embodiment... The example uses 72, corresponding to 12 hours (index step size 10 minutes), to cover the long-term restrictions and shutdown periods that are common during typhoon processes.

[0068] S3.4 Output the posterior probabilities of the control states at each time step as a control state probability sequence according to the order of the control state set.

[0069] For example, the order is fixed as "normal state, power-limited state, yaw-limited state, cut-out shutdown state, and recovery ramp state"; for each unified timestamp index within the prediction window, a set of five-element probability values ​​is output and written together with the timestamp index into the control state probability sequence table.

[0070] S4. Based on the control state probability sequence and the target feature set to be predicted, generate initial values ​​for the predicted power sequence according to the preset power prediction mapping rules, and perform threshold comparison correction. Note that the following should be noted in this step: S4.1 Read the control state probability sequence and the target feature set to be predicted, and match them one-to-one according to the unified timestamp index to obtain the power prediction input set.

[0071] In a preferred embodiment, when reading the control state probability sequence and the target feature set to be predicted, the two are matched one-to-one using a unified timestamp index as the key: if a certain index point has a corresponding record in the control state probability sequence and a corresponding record in the target feature set to be predicted, a power prediction input record is generated; if there is a missing record, the index point is not included in the power prediction input set, and the input missing flag is written as 1 so as to calculate the prediction coverage.

[0072] For example, if the prediction window starts at 00:00:00 on 2025-08-18, and both the control state probability sequence and the target feature record exist at 00:10:00, then a power prediction input record is written; if the target feature record is marked as a long missing segment and has no valid value at 00:20:00, then this index point is not included in the power prediction input set.

[0073] S4.2. Based on the power prediction input set, candidate power values ​​corresponding to each control state are generated according to the preset power prediction mapping rules. The preset power prediction mapping rules select mapping parameters related to wind speed, pitch angle, rotational speed and control mode code under different control states to obtain the candidate power set.

[0074] In a preferred embodiment, the preset power prediction mapping rules set the sources and constraints of mapping parameters according to control state classification: When the control state is normal, the candidate power value is taken from the output of the normal power curve corresponding to the wind speed at the hub height at the index point, and can be written into the candidate power value after correcting the air density according to the temperature and air pressure in the weather forecast. When the control state is the power-limited state, the candidate power value is obtained by multiplying the candidate power value of the normal state by the power-limiting coefficient. The power-limiting coefficient is determined by the discrete encoding of the control mode code and the pitch angle level. For example, when the discrete encoding is 1 and the pitch angle is greater than 20 degrees, the power-limiting coefficient is 0.6; when the discrete encoding is 1 and the pitch angle is not greater than 20 degrees, the power-limiting coefficient is 0.8. When the control state is the yaw-limited state, the candidate power value is obtained by multiplying the normal state candidate power value by the yaw loss coefficient. The yaw loss coefficient is determined by the range of the difference between the wind direction and the yaw angle; for example, 0.95 is taken when the difference is within 15 degrees, 0.85 is taken when the difference is between 15 and 30 degrees, and 0.7 is taken when the difference is greater than 30 degrees. When the control state is the cut-out shutdown state, the candidate power value is written to 0, while the shutdown candidate flag is retained as 1; When the control state is in the recovery ramp state, the candidate power value is written based on the active power corresponding to the previous unified timestamp index, and its increment is limited according to the ramp rate limit in the subsequent S4.4. This generates five candidate power values ​​at each index point, forming a candidate power set, which, together with the control state probability sequence at the same index point, is written into the power prediction input set.

[0075] S4.3. Based on the control state probability sequence, perform probability weighted fusion on the candidate power set to obtain the candidate initial value of the predicted power sequence.

[0076] In a preferred embodiment, when performing probability-weighted fusion of the candidate power set based on the control state probability sequence, the initial candidate value is calculated for each unified timestamp index point in the following manner: in, The initial candidate power value is a unified timestamp index at time t; J represents the number of control states. Let be the posterior probability of the control state j output by S3 at time t; Let be the candidate power value generated at time t in control state j.

[0077] S4.4 When the difference between the probability of the highest probability control state and the probability of the second highest probability control state in the control state probability sequence is less than the preset difference threshold, the weight of the highest probability control state in the probability weighted fusion is reduced by a preset smoothing coefficient, and the reduced weight difference is proportionally allocated to the weight of the second highest probability control state, and the initial value of the updated prediction power sequence is output. When the probability of switching out of the shutdown state is greater than or equal to the preset shutdown threshold, the predicted power at the corresponding moment is set to zero. When the probability of resuming the climbing state is greater than or equal to the preset climbing threshold, the predicted power at the corresponding time is incrementally limited according to the preset climbing rate upper limit. Output the predicted power sequence after threshold comparison correction; The target feature set to be predicted is selected by the target feature set according to the prediction time window division rule.

[0078] In a preferred embodiment, when the difference between the probabilities of the most probable control state and the second most probable control state in the control state probability sequence is less than a preset difference threshold, the weights of the probability weighted fusion are smoothed; let the index of the most probable control state be... The second most likely control state index is The preset difference threshold is The preset smoothing coefficient is Then when When updating weights, an example is written as: Then, keeping the remaining control state weights unchanged, the updated candidate initial values ​​are calculated using the updated weights: in, These are the smoothed control state weights; This is the initial value of the smoothed predicted power; The preset difference threshold; The preset smoothing coefficient; This is the index for the maximum probability control state; This is the index for the second most probable control state.

[0079] As an example, the preset difference threshold is 0.1. Specifically, during the typhoon boundary impact phase, there are often short-term alternations between the power-limited state and the yaw-limited state. If only the control state with the highest probability is taken, the predicted power will exhibit unnecessary sawtooth fluctuations. Introducing weighted smoothing for moments with a probability difference of less than 0.1 can reduce the impact of control state boundary jitter on the power sequence.

[0080] As an example, the preset smoothing coefficient is 0.3, which means that 30% of the weights are transferred from the highest probability control state to the second highest probability control state, making the initial power value more continuous between the two high probability control states.

[0081] For example, if the power-limited state probability is 0.46 and the yaw-limited state probability is 0.4 at a certain time, and the difference between the two is 0.06, which is less than 0.1, then according to... After the update, the power-limited state weight is 0.322, the yaw-limited state weight is 0.538, and the remaining weights remain unchanged. Based on this, the initial predicted power value at that moment is recalculated.

[0082] Furthermore, the threshold comparison correction includes: When the posterior probability of switching out of the shutdown state is greater than or equal to the preset shutdown threshold, the predicted power at the corresponding time is written to 0. The preset shutdown threshold is 0.70. Specifically, in the sample segment before and after the typhoon landfall, when the probability of switching out of the shutdown state reaches 0.70 or higher, the control mode code and power monitoring often show continuous shutdown intervals. Writing 0 according to this threshold can reduce the situation where the shutdown interval is diluted to non-zero power.

[0083] When the posterior probability of recovering the ramp state is greater than or equal to the preset ramp threshold, an upper limit is imposed on the increment of the predicted power at the corresponding time. The preset ramp threshold is 0.60 in example, and the upper limit of the ramp rate is "not exceeding 0.08 of the rated power every 10 minutes" in example. The increment limit is based on "the predicted power corresponding to the previous unified timestamp index". When the increment of the smoothed initial value of the predicted power relative to the previous predicted power exceeds the upper limit, the increment is truncated to the upper limit value.

[0084] It should also be noted in this embodiment that the posterior probability of the cut-out shutdown state and the posterior probability of the recovery ramp state are both derived from the control state probability sequence output by S3.4. After pairing in S4.1, they are directly read by the timestamp index without further calculation, thus ensuring that the threshold comparison is consistent with the semi-Markov recursion. Compared with the scheme of directly shutting down or ramping based on the wind speed threshold, this embodiment uses the control state probability as the decision quantity, which can explicitly write the uncertainty of the control logic under the typhoon process into the prediction method, reducing misjudgments when the wind speed is close to the cut-out point but the unit is still under power limitation or yaw limitation.

[0085] Preferably, after outputting the predicted power sequence after threshold comparison correction, this embodiment writes the predicted power sequence into the power prediction result table according to a unified timestamp index, and writes it into the result storage area that can be read by the scheduling side according to the existing scheduling data interface of the wind farm. At the same time, it retains the control state probability sequence and validity identification sequence corresponding to each prediction point, so as to perform post-event verification and parameter backtracking update for prediction deviation.

[0086] In applying the above embodiments, other aspects of the present invention also propose a typhoon-path-oriented offshore wind power prediction system, including: The data acquisition module collects typhoon path forecast data, offshore wind farm turbine monitoring data, and weather forecast data; The preprocessing module aligns typhoon path forecast data, offshore wind farm turbine monitoring data, and meteorological forecast data according to a unified timestamp, completes missing data, and removes anomalies, outputting a preprocessed dataset. The feature extraction module extracts typhoon path-driven features and unit operation features based on the preprocessed dataset, and normalizes and fused the two types of features according to preset fusion rules to output the target feature set. The control state prediction module inputs the target feature set into the control state transition prediction model built based on the semi-Markov control state transition algorithm, and outputs the control state probability sequence at each time point within the prediction window. The time window selection module selects the target feature set to be predicted from the target feature set according to the prediction time window division rules; The initial power value generation module generates the initial value of the predicted power sequence according to the control state probability sequence and the target feature set to be predicted, and according to the preset power prediction mapping rule. The threshold correction module performs threshold comparison correction on the initial value of the predicted power sequence. Specifically, when the probability of switching out of the shutdown state is greater than or equal to the preset shutdown threshold, the predicted power at the corresponding time is set to zero. When the probability of resuming the ramp state is greater than or equal to the preset ramp threshold, the predicted power at the corresponding time is incrementally limited according to the preset ramp rate upper limit, and the predicted power sequence after threshold comparison correction is output.

[0087] The system also includes one or more processors and memory.

[0088] The memory is used to store operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the offshore wind power forecasting method for typhoon paths described in the foregoing embodiments, particularly... Figure 1 The flowchart of the method is shown.

[0089] Other aspects disclosed in the embodiments of the present invention also propose a computer-readable medium for storing software including instructions executable by one or more computers, which, upon execution, cause the one or more computers to perform operations including the flow of the offshore wind power forecasting method for typhoon paths described in the foregoing embodiments, particularly... Figure 1 The flowchart of the method is shown.

[0090] It should be recognized that embodiments of the present invention may be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium.

[0091] The method can be implemented using standard programming techniques, including a non-transitory computer-readable storage medium configured with a computer program in the computer program, wherein the storage medium is configured such that the computer operates in a specific and predefined manner.

[0092] Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system; however, if required, the program can be implemented in assembly or machine language.

[0093] In any case, the language can be either compiled or interpreted.

[0094] Furthermore, for this purpose, the program can run on programmed application-specific integrated circuits.

[0095] The processes described herein (or variations and / or combinations thereof) can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program includes a plurality of instructions executable by one or more processors.

[0096] Furthermore, the method can be implemented in any suitable computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices.

[0097] Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether portable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein.

[0098] Furthermore, machine-readable code, or parts thereof, can be transmitted via wired or wireless networks.

[0099] When such media includes instructions or programs that combine with a microprocessor or other data processor to implement the steps described above, the invention described herein includes these and other different types of non-transitory computer-readable storage media.

[0100] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting offshore wind power based on typhoon paths, characterized in that, include: Typhoon path forecast data, offshore wind farm turbine monitoring data, and meteorological forecast data were collected and aligned according to a unified timestamp, missing data was filled in, and anomalies were removed to obtain a preprocessed dataset. Based on the preprocessed dataset, typhoon path-driving features and unit operation features are extracted, and the two types of features are normalized and spliced ​​together according to preset fusion rules to obtain the target feature set. The target feature set is input into the control state transition prediction model constructed based on the semi-Markov control state transition algorithm, and the control state probability sequence at each time point within the prediction window is output. Based on the control state probability sequence and the target feature set to be predicted, an initial value of the predicted power sequence is generated according to a preset power prediction mapping rule, and a threshold comparison correction is performed: When the probability of switching out of the shutdown state is greater than or equal to the preset shutdown threshold, the predicted power at the corresponding moment is set to zero. When the probability of resuming the climbing state is greater than or equal to the preset climbing threshold, the predicted power at the corresponding time is incrementally limited according to the preset climbing rate upper limit. Output the predicted power sequence after threshold comparison correction; The target feature set to be predicted is selected from the target feature set according to the prediction time window division rule.

2. The offshore wind power prediction method based on typhoon paths according to claim 1, characterized in that, The preprocessed dataset is obtained by: Collect typhoon path forecast data, meteorological forecast data, and offshore wind farm turbine monitoring data, and write a collection timestamp and data source identifier for each of the typhoon path forecast data, meteorological forecast data, and offshore wind farm turbine monitoring data to obtain the original data set; A unified timestamp index is generated based on the original data set, and the typhoon path forecast data, the weather forecast data, and the offshore wind farm turbine monitoring data are mapped to the unified timestamp index to obtain an aligned data set. The aligned data set is subjected to missing data completion processing, wherein data segments with consecutive missing lengths not exceeding a preset missing length threshold are interpolated and completed by adjacent valid sampling points, and data segments with consecutive missing lengths exceeding the preset missing length threshold are marked as missing segments and the missing identifier is retained, thus obtaining the completed data set. Anomaly removal processing is performed on the completed data set, wherein data points exceeding a preset physical range are removed according to preset physical constraint rules, and anomaly identifiers are written for the removed data points; the completed data set, the missing identifiers, and the anomaly identifiers are aggregated to generate the preprocessed dataset.

3. The method for predicting offshore wind power based on typhoon paths according to claim 1 or 2, characterized in that, Obtaining the target feature set includes: Read the preprocessed dataset and extract the typhoon center location, typhoon movement direction and typhoon intensity parameters time by time according to the unified timestamp index. Calculate the typhoon relative distance, relative azimuth and approach speed based on the geographical coordinates of the wind farm to obtain the typhoon path driving feature set. Based on the preprocessed dataset, wind turbine hub height wind speed, wind direction, pitch angle, yaw angle, speed, active power, nacelle vibration and control mode code are extracted time-by-time. The wind turbine hub height wind speed, wind direction, pitch angle, yaw angle, speed, active power, nacelle vibration and control mode code are aggregated according to the wind turbine number to obtain the unit operation feature set. Normalization processing is performed on the typhoon path driving feature set and the unit operation feature set respectively. In this process, continuous quantities are scaled according to a preset dimension range, and the control mode code is discretely encoded according to a preset encoding rule to obtain the normalized typhoon path driving feature set and the normalized unit operation feature set. According to the preset fusion rules, the normalized typhoon path-driven feature set and the normalized unit operation feature set are concatenated at the same timestamp, and the missing and abnormal identifiers in the preprocessed dataset are written into the concatenation result as input items for the fusion weights to obtain the target feature set.

4. The offshore wind power prediction method based on typhoon paths according to claim 3, characterized in that, The preset fusion rules include: The normalized typhoon path driving feature set and the normalized unit operation feature set are paired one-to-one using the same unified timestamp index; Under each unified timestamp index, the normalized typhoon path driving feature set and the normalized unit operation feature set are spliced ​​together according to a preset feature arrangement order, and a feature validity identifier sequence corresponding to the splicing result is generated. A feature weight sequence is generated based on the missing and anomaly identifiers in the preprocessed dataset, wherein the feature components corresponding to the missing identifiers are weighted by a first attenuation coefficient, and the feature components corresponding to the anomaly identifiers are weighted by a second attenuation coefficient, and the second attenuation coefficient is less than the first attenuation coefficient. The splicing result is weighted according to the feature weight sequence, and the feature validity identifier sequence is written into the weighted result to obtain the target feature set.

5. The offshore wind power prediction method based on typhoon paths according to claim 4, characterized in that, The preset feature arrangement order includes: First, write the normalized typhoon path-driven feature set, which includes the typhoon relative distance, relative azimuth, approach speed, typhoon movement direction, typhoon intensity parameters, and wind circle coverage. Then, the normalized unit operating characteristic set is written, and the hub height, wind speed, wind direction, yaw angle, pitch angle, speed, nacelle vibration, control mode code, and active power corresponding to the previous unified timestamp index of each wind turbine are written in ascending order according to the wind turbine number.

6. The offshore wind power prediction method based on typhoon paths according to claim 4, characterized in that, The method for generating the control state probability sequence includes: Read the target feature set, and select the target feature set to be predicted corresponding to the prediction window according to the prediction window division rule to obtain the prediction input feature set; In the control state transition prediction model, a set of control states is set, which includes at least the normal state, power-limited state, yaw-limited state, cut-out stop state, and recovery climb state. The initial transition parameters and duration parameters of each control state are generated based on the changes in the control mode codes in the historical operation records. The predicted input feature set is input into the control state transition prediction model. The control state observation probability at each time step is calculated according to the correspondence between the predicted input feature set and the control state set. The model is then recursively updated by combining the initial transition parameter and the duration parameter to obtain the control state posterior probability at each time step. The posterior probabilities of the control states at each time point are output as a control state probability sequence according to the order of the control state set.

7. The offshore wind power prediction method based on typhoon paths according to claim 6, characterized in that, The initial values ​​of the predicted power sequence are generated according to the preset power prediction mapping rules, including: Read the control state probability sequence and the target feature set to be predicted, and match them one-to-one according to a unified timestamp index to obtain the power prediction input set; Based on the power prediction input set, candidate power values ​​corresponding to each control state are generated according to the preset power prediction mapping rules. The preset power prediction mapping rules select mapping parameters related to wind speed, pitch angle, rotational speed and control mode code under different control states to obtain the candidate power set. Based on the control state probability sequence, the candidate power set is subjected to probability weighted fusion to obtain candidate initial values ​​of the predicted power sequence; When the difference between the probability of the highest probability control state and the probability of the second highest probability control state in the control state probability sequence is less than a preset difference threshold, the weight of the highest probability control state in the probability weighted fusion is reduced by a preset smoothing coefficient, and the reduced weight difference is proportionally allocated to the weight of the second highest probability control state, and the initial value of the updated prediction power sequence is output.

8. A typhoon-path-oriented offshore wind power prediction system, characterized in that, include: The data acquisition module collects typhoon path forecast data, offshore wind farm turbine monitoring data, and weather forecast data; The preprocessing module aligns the typhoon path forecast data, the offshore wind farm turbine monitoring data, and the weather forecast data according to a unified timestamp, completes missing data, and removes anomalies, and outputs a preprocessed dataset. The feature extraction module extracts typhoon path-driven features and unit operation features based on the preprocessed dataset, and normalizes and merges the two types of features according to preset fusion rules to output the target feature set. The control state prediction module inputs the target feature set into the control state transition prediction model constructed based on the semi-Markov control state transition algorithm, and outputs the control state probability sequence at each time point within the prediction window. The time window selection module selects a set of target features to be predicted from the target feature set that corresponds to the prediction time window according to the prediction time window division rules; The initial power value generation module generates an initial value for the predicted power sequence according to the control state probability sequence and the target feature set to be predicted, based on a preset power prediction mapping rule. The threshold correction module performs threshold comparison correction on the initial value of the predicted power sequence. When the probability of switching out of the shutdown state is greater than or equal to the preset shutdown threshold, the predicted power at the corresponding time is set to zero. When the probability of recovering to the climbing state is greater than or equal to the preset climbing threshold, the predicted power at the corresponding time is incrementally limited according to the preset upper limit of the climbing rate, and the predicted power sequence after threshold comparison correction is output.

9. The offshore wind power prediction system oriented towards typhoon paths according to claim 8, characterized in that, The system also includes one or more processors; The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the offshore wind power prediction method for typhoon paths as described in any one of claims 1 to 7.

10. A computer-readable medium for storing software, characterized in that: The software includes instructions executable by one or more computers, which, upon execution, cause the one or more computers to perform operations including the flow of the offshore wind power prediction method for typhoon paths as described in any one of claims 1 to 7.

Images