A method, device and system for dynamic spatiotemporal transfer learning of wind turbine operation
By using the method of centralized modeling of benchmark units and spatiotemporal migration of target units, combined with time-triggered and event-triggered mechanisms, the problem of high cost and low deployment efficiency of modeling wind turbines of the same type is solved, and the rapid construction and low-energy maintenance of multi-unit output models of wind farms are realized.
Patent Information
- Application Number
- CN202611253268.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-18
- Publication Date
- 2026-11-17
AI Technical Summary
In existing technologies, modeling wind turbines of the same type is costly, has low deployment efficiency, is difficult to maintain, and consumes a lot of energy for repeated full training. It also lacks dynamic spatiotemporal transfer learning schemes and cannot effectively utilize the spatial correlation and similarity of operating characteristics between units for model transfer.
A combined approach is adopted, which includes centralized modeling of benchmark units, spatiotemporal migration of target units, dual mechanisms of time-triggered and event-triggered mechanisms, sliding time window maintenance, and incremental learning. A basic model is established by using a small number of benchmark units and other similar units are updated in batches. By combining spatial correlation and similarity of operating characteristics, dynamic model migration and low-energy maintenance are achieved.
It significantly reduces the number of repetitive training sessions for the same type of wind turbine, improves the quality of model initialization, reduces computing energy consumption and communication overhead, meets the low-carbon operation requirements of wind farms, and enables rapid deployment and continuous optimization.
Smart Images

Figure CN123121834A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of new energy power generation prediction, wind farm group unit modeling and intelligent operation and maintenance technology, specifically to a dynamic spatiotemporal transfer learning method, device and system for the output of a group of wind turbine units of the same type. Background Technology
[0002] As the scale of new energy power generation continues to expand, the number of wind farm units continues to increase. Large wind farms typically contain dozens or even hundreds of wind turbine units. Although the units use the same or similar unit models, controller configurations, and rated capacities, their output characteristics often show significant differences due to differences in geographical location, prevailing wind direction, terrain undulations, wake obstruction, local air density variations, and equipment aging.
[0003] In existing technologies, it is typically necessary to build a separate power output prediction model for each wind turbine. While this turbine-by-turbine modeling approach can theoretically achieve precise fitting for a single turbine, it suffers from problems such as high modeling costs, low deployment efficiency, maintenance difficulties, and high computational energy consumption in engineering. For large-scale wind farms, if each turbine independently undergoes data cleaning, sample construction, model training, parameter optimization, and online validation, it will result in a large amount of repetitive work.
[0004] First, modeling on a per-machine basis is costly. Wind turbine operating data can be incomplete, abnormal, or experience shutdowns, power outages, and state transitions, requiring a lengthy process of cleaning and processing the raw SCADA data. Repeating this process for each turbine in a wind farm would significantly increase the data processing workload.
[0005] Secondly, the deployment efficiency of modeling on a per-unit basis is low. Training and deploying each unit's model independently is not conducive to establishing a unified model management mechanism at the wind farm level. When new units come online or existing units experience changes in operating conditions, maintenance personnel often need to reorganize the training process, making rapid batch deployment difficult.
[0006] Secondly, maintaining the model based on a per-machine approach is difficult. The operating environment of wind turbines changes with the seasons, control strategy adjustments, equipment maintenance, and sensor drift, leading to a decrease in the applicability of the original model. Existing solutions often require periodic retraining of the entire model, which increases the maintenance burden and makes it difficult to respond promptly to emergencies.
[0007] Finally, machine-by-machine modeling consumes a lot of computational energy. Repeatedly training a large number of models leads to significant server power consumption, edge computing load, and data communication overhead, which is detrimental to green and low-carbon operation. As wind farms place increasing demands on energy conservation and computing efficiency, reducing unnecessary repetitive training has become an urgent technical problem to be solved.
[0008] To overcome the shortcomings of unit-by-unit modeling, existing technologies have proposed a unified model scheme for the entire field, which involves mixing data from multiple units to construct a unified model. While this scheme can reduce the number of models to some extent, it often fails to take into account the individual adaptability of each unit because it ignores differences in unit location, wake effects, and local operating characteristics. When there is strong heterogeneity among units, the unified model is prone to averaging errors.
[0009] In addition, some transfer learning solutions attempt to directly copy or fine-tune models from existing units to new units, but most remain at the static one-time migration stage. Static one-time migration refers to completing the model migration at a certain point in time and then ceasing continuous updates based on subsequent operating conditions. This approach is effective during the initial commissioning phase of the unit, but as time goes on, when the target unit undergoes maintenance, changes in wind conditions, power curtailment switching, or sensor maintenance, the model will gradually deviate from actual operating patterns.
[0010] Most existing solutions lack clear time-triggered and event-triggered mechanisms. If the model is updated frequently at fixed short intervals, it will cause additional computing power and energy consumption; if event-triggered conditions are not established, the model may not be able to be updated in time when operational anomalies occur, affecting prediction accuracy and operation and maintenance decisions.
[0011] On the other hand, wind farm operational data exhibits a clear temporal evolution characteristic. Seasonal changes in wind resources, alterations in maintenance strategies, accumulated turbine wear, and revisions to control parameters can all lead to slow or abrupt changes in the dynamic relationship of turbine output. Therefore, model maintenance should not remain based on static historical data but should continuously reflect recent operational status through a sliding time window mechanism.
[0012] Meanwhile, during model maintenance, retraining the entire historical data for each update, while maintaining high theoretical accuracy, is prohibitively expensive in engineering. Especially when edge computing devices or local server resources in wind farms are limited, full retraining is not suitable as a routine update method. In contrast, incremental learning can use only newly added data to locally update the model, thus maintaining model performance at a lower computational cost.
[0013] Furthermore, wind turbines of the same type are not entirely independent. There are often spatial proximity relationships, upstream and downstream relationships with prevailing winds, and wake influence relationships between units. Historical operating curves of different units exhibit a certain degree of transferability under similar operating conditions. Therefore, if, after selecting a benchmark unit, the spatial correlation and similarity of operating characteristics between units can be fully utilized to batch migrate the output model of the benchmark unit to other units, and then incremental updates can be performed using a small number of local samples, it is expected to significantly reduce modeling costs and improve deployment efficiency.
[0014] In summary, current technologies lack a dynamic spatiotemporal transfer learning scheme for wind turbine clusters of the same type. Such a scheme should simultaneously possess the following capabilities: first, the ability to identify benchmark turbines and establish a shareable basic dynamic model; second, the ability to utilize the spatial relationships between turbines for model transfer; third, the ability to perform updates on demand based on time-triggered and event-triggered mechanisms; fourth, the ability to combine sliding time windows and incremental learning to achieve continuous optimization in later maintenance; and fifth, the ability to reduce computational energy consumption caused by repeated training, thereby meeting the requirements for low-carbon operation.
[0015] Therefore, there is an urgent need to propose a new technical approach for wind turbine operation modeling, so that after selecting one or a few benchmark turbines to establish a dynamic output model, it can be updated in batches and quickly deployed to other similar turbines, while maintaining dynamic update capability throughout the entire operation cycle. Summary of the Invention
[0016] To address the technical problems of high cost, low deployment efficiency, difficult maintenance, and high energy consumption of repeated full training for modeling identical wind turbines in existing technologies, this invention provides a method, device, and system for dynamic spatiotemporal transfer learning of wind turbine operation. Through a combination of "centralized modeling of benchmark units + spatiotemporal transfer of target units + dual mechanisms of time-triggered and event-triggered mechanisms + sliding time window maintenance + incremental learning optimization", it enables rapid construction, batch deployment, and low-energy maintenance of multi-unit output models in wind farms.
[0017] This invention provides a dynamic spatiotemporal migration learning system for wind turbine operation. The system includes: a data acquisition module, a benchmark turbine identification module, a benchmark model construction module, a spatiotemporal migration learning module, a trigger determination module, a sliding time window management module, an incremental learning module, a model deployment module, and a model maintenance management module.
[0018] Furthermore, the data acquisition module is used to collect operational data, spatial location data, and maintenance event data of the same type of wind turbine group within the wind farm. The operational data includes at least one or more of the following: wind speed, wind direction, turbine output power, rotational speed, pitch angle, status variables, alarm variables, temperature, vibration, and maintenance records. The spatial location data includes at least turbine coordinates, relative distance, upstream and downstream topology relationships, and wake influence relationships. The maintenance event data includes at least maintenance, shutdown, power curtailment, equipment replacement, fault recovery, and control strategy switching events.
[0019] Furthermore, the benchmark unit identification module is used to select at least one unit as a benchmark unit from multiple similar units in the wind farm based on the unit's operational stability, data integrity, operating condition coverage, failure rate, unit location representativeness, and average similarity with other units. The benchmark unit is used to undertake the full training task of the basic model.
[0020] Furthermore, the benchmark model construction module is used to construct a basic power output dynamic model based on the historical operating data of the benchmark unit. The basic power output dynamic model is used to characterize the dynamic mapping relationship between environmental variables, historical power output, operating status parameters and future power output, and preferably retains the transferable shared main structure and the adaptable local parameter structure.
[0021] Furthermore, the spatiotemporal transfer learning module is used to combine the spatial correlation between wind turbine units, the similarity of operating characteristics, and the most recent sliding time window data of the target unit to transfer the basic model of the benchmark unit to the target unit. In the scenario of multiple benchmark units, the spatiotemporal transfer learning module can also generate a multi-source fusion initial model of the target unit based on the comprehensive transfer weight between multiple benchmark units and the target unit.
[0022] Furthermore, the trigger determination module is used to determine whether to perform model updates based on a time-triggered mechanism or an event-triggered mechanism. The time-triggered mechanism is used to periodically refresh the model, while the event-triggered mechanism is used to quickly trigger updates after prediction errors exceed limits, maintenance events occur, wind resources change abruptly, power curtailment status switches, shutdown recovery, or sensor calibration.
[0023] Furthermore, the sliding time window management module is used to maintain the data window of the target unit for the most recent preset duration, and to assign weights to the samples in the window according to the data timeliness, operating condition similarity and sample quality; the incremental learning module is used to perform local parameter updates on the target unit model using only the newly added samples or a combination of newly added samples and replay samples after the update is triggered, so as to reduce update overhead and alleviate model forgetting.
[0024] Furthermore, the model deployment module is used to batch publish the updated model to the wind farm monitoring platform, edge computing node, industrial control computer or cloud-edge collaboration platform, and to uniformly record and manage the model version, update time, triggering reason, update benefits and rollback status.
[0025] This invention also provides a dynamic spatiotemporal transfer learning method for wind turbine operation, the method comprising the following steps:
[0026] The data acquisition steps involve collecting operational data, spatial location data, and maintenance event data from multiple identical wind turbine units in the wind farm, and performing time alignment and anomaly processing on the data. The operational data includes at least one or more of the following: wind speed, wind direction, turbine output power, speed, pitch angle, status variables, alarm variables, temperature, vibration, and maintenance records. The spatial location data includes at least turbine coordinates, relative distance, upstream and downstream topology relationships, and wake influence relationships. The maintenance event data includes at least maintenance, shutdown, power curtailment, equipment replacement, fault recovery, and control strategy switching events.
[0027] The benchmark unit identification process involves selecting at least one unit from multiple similar units within the wind farm as the benchmark unit, based on factors such as unit operational stability, data integrity, operating condition coverage, failure rate, unit location representativeness, and average similarity to other units.
[0028] The benchmark model construction steps involve building a basic dynamic output model based on the historical operating data of the benchmark unit. The basic dynamic output model is used to characterize the dynamic mapping relationship between environmental variables, historical output, operating status parameters and future output.
[0029] The spatiotemporal transfer learning step combines the spatial correlation between wind turbine units, the similarity of operating characteristics, and the most recent sliding time window data of the target unit to transfer the basic output dynamic model of the benchmark unit to the target unit.
[0030] The trigger determination step determines whether to perform a model update based on either a time-triggered mechanism or an event-triggered mechanism. The time-triggered mechanism is used to periodically refresh the model, while the event-triggered mechanism is used to trigger an update after the prediction error exceeds the limit, a maintenance event occurs, wind resources change abruptly, power curtailment status changes, shutdown recovery, or sensor calibration. The event-triggered mechanism is triggered by any of the following conditions: prediction error continuously exceeds the threshold, unit undergoes maintenance, blade cleaning or replacement, sensor calibration, control parameter adjustment, shutdown recovery, power curtailment status change, arrival or departure of extreme wind conditions, or a significant increase in data distribution drift.
[0031] The sliding time window management process involves maintaining a data window for the target unit with the most recent preset duration, and assigning weights to samples within the window based on data timeliness, operating condition similarity, and sample quality.
[0032] The incremental learning step involves using only newly added samples or a combination of newly added and replay samples to perform local parameter updates on the target unit model after an update is triggered, in order to reduce update overhead and alleviate model forgetting.
[0033] The model deployment steps involve batch publishing the updated model to the wind farm monitoring platform, edge computing nodes, industrial control computers, or cloud-edge collaboration platforms.
[0034] The present invention also provides a dynamic spatiotemporal migration learning device for wind turbine operation. The device includes a processor, a memory, and a communication interface for executing the above-mentioned method steps. When the processor executes the program stored in the memory, it can complete benchmark unit selection, migration weight calculation, update trigger determination, incremental learning, and model deployment.
[0035] The present invention also provides an electronic device, which includes one or more processors, a storage unit and a communication interface. When the processor executes the program in the storage unit, it implements the above-mentioned dynamic spatiotemporal migration learning method for wind turbine operation.
[0036] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the above-described dynamic spatiotemporal migration learning method for wind turbine operation.
[0037] (iv) Beneficial effects
[0038] Compared with the prior art, the present invention has at least the following beneficial effects:
[0039] Firstly, by centrally modeling a small number of benchmark units and migrating the remaining units as needed, the number of repetitive training sessions for the same type of units is significantly reduced.
[0040] Secondly, by jointly modeling spatial correlation and operational feature similarity, the initialization quality of the target unit model is improved, and the risk of negative migration is reduced.
[0041] Third, a balance between accuracy and computing power is achieved through a dual mechanism of time-triggered and event-triggered mechanisms;
[0042] Fourth, the model's long-term adaptability is maintained through sliding time windows and incremental learning;
[0043] Fifth, by deploying and updating in batches during low-load periods, the computational power consumption and communication overhead during model maintenance are reduced, thus meeting the low-carbon operation requirements of wind farms. Attached Figure Description
[0044] Figure 1 This is a schematic diagram of the structure of the wind turbine operation dynamic spatiotemporal transfer learning system provided in an embodiment of the present invention.
[0045] Figure 2 This is a schematic diagram of the dynamic spatiotemporal migration learning method for wind turbine operation provided in an embodiment of the present invention.
[0046] Figure 3 This is a schematic diagram of the structure of an electronic device or apparatus provided in an embodiment of the present invention.
[0047] Figure 4 This is a schematic diagram illustrating the time-triggered, event-triggered, sliding time window update, and incremental learning closed loop provided in the embodiments of the present invention.
[0048] Figure 5 This is a schematic diagram of benchmark unit identification and comprehensive scoring provided in an embodiment of the present invention.
[0049] Figure 6 This is a schematic diagram illustrating the calculation of the spatial correlation matrix and spatiotemporal migration weights of a wind turbine generator set, provided in an embodiment of the present invention. Detailed Implementation
[0050] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention; for those skilled in the art, equivalent substitutions, conventional adjustments or functional changes made without departing from the concept of the present invention should all fall within the scope of protection of the present invention.
[0051] (I) Overall System Structure
[0052] like Figure 1 The diagram shows the structure of a dynamic spatiotemporal migration learning system for wind turbine operation provided in an embodiment of the present invention. This system can be deployed on a wind farm master station server, a tower-based industrial control computer, an edge gateway, or a cloud-edge collaborative platform. Its functional modules include at least: a data acquisition module 110, a benchmark turbine identification module 120, a benchmark model construction module 130, a spatiotemporal migration learning module 140, a trigger determination module 150, a sliding time window management module 160, an incremental learning module 170, a model deployment module 180, and a model maintenance management module 190. The model maintenance management module 190 records model version, update logs, trigger reasons, rollback flags, and operational performance evaluation results to support full lifecycle maintenance of the model.
[0053] During system operation, the data acquisition module 110 first aggregates data from all turbines in the wind farm. Then, the benchmark turbine identification module 120 selects representative turbines from multiple candidate turbines. Following this, the benchmark model construction module 130 trains a basic dynamic power output model. After basic model training, the spatiotemporal transfer learning module 140 generates target turbine models by combining the spatial relationships and recent operational characteristics of the target turbines. During operation, the trigger determination module 150 continuously monitors update conditions. Once time-triggered or event-triggered conditions are met, the sliding time window management module 160 and the incremental learning module 170 perform continuous optimization. Finally, the model deployment module 180 completes batch replacement and release.
[0054] (II) Data Collection
[0055] In terms of data acquisition, the data acquisition module 110 is used to collect multi-source data from a group of wind turbines of the same type within the wind farm. The data sources can come from the turbine's main control system, the wind farm's monitoring platform, edge sensors, environmental monitoring equipment, and maintenance record systems. To ensure consistency in subsequent model training, the data acquisition module preferably uses a unified time base to synchronize all data.
[0056] In one embodiment, the collected data includes, but is not limited to: wind speed, wind direction, active power output, reactive power output, generator speed, rotor speed, blade pitch angle, yaw angle, ambient temperature, estimated air density, unit operating status, fault alarm codes, downtime, power rationing indicators, maintenance records, and inspection logs. A portion of this data serves as the main input to the model, while another portion serves as auxiliary information for event triggering and maintenance determination.
[0057] The data acquisition module 110 can acquire data through the wind farm monitoring system interface or industrial communication protocols, such as OPC communication, Modbus communication, IEC 104 protocol, IEC 61400-25 protocol, and the internal API interface of the wind farm. Through a unified data interface, centralized acquisition and structured storage of unit data at the wind farm can be achieved.
[0058] In one embodiment, the data acquisition module 110 performs hierarchical acquisition of different types of data according to different acquisition frequencies: for core modeling input variables such as wind speed, wind direction, and active power output, high-frequency acquisition is used at 1 second or 100 milliseconds; for state monitoring variables such as temperature and vibration, lower-frequency acquisition is used at 1 minute or 10 minutes; for event-type data such as maintenance records and inspection logs, event-driven acquisition is used to acquire data in real time when it occurs. Data of different frequencies are aligned using a unified timestamp to form a multi-timescale modeling dataset.
[0059] In one embodiment, the data acquisition module 110 stores the acquired data in a structured manner according to the data model defined by the IEC 61400-25 standard, mapping the wind turbine operation data to logical nodes and data attributes in the standard information model, which facilitates subsequent cross-platform data exchange and model reuse.
[0060] In some implementations, the data acquisition module 110 further includes a data preprocessing unit for performing missing value imputation, outlier identification, state filtering, time alignment, and unit unification. Missing value imputation can employ interpolation, nearest neighbor imputation, or statistical substitution; outlier identification can combine physical threshold rules and statistical rules; state filtering can remove samples that are unsuitable for direct use in normal output modeling, such as those from shutdowns, maintenance, and grid switching.
[0061] In one implementation, the missing value imputation of the data preprocessing unit adopts the following hierarchical strategy: for data points with a continuous missing duration of less than 5 sampling periods, linear interpolation or cubic spline interpolation is used for imputation; for data segments with a continuous missing duration between 5 and 50 sampling periods, the nearest neighbor imputation method based on similar operating conditions is used, that is, the time period with the closest conditions such as wind speed, wind direction, and power output level is found from historical data for data imputation; for data segments with a continuous missing duration of more than 50 sampling periods, they are marked as unusable and removed without forced imputation.
[0062] In one implementation, outlier identification combines physical threshold rules and statistical rules for dual judgment: the physical threshold rule sets a reasonable range of values for each measuring point based on the design parameters of the wind turbine, and data exceeding the physical limits are directly marked as outliers; the statistical rule uses the 3σ criterion or interquartile range method based on a sliding window to mark data that exceeds the statistically normal range. Data points marked as outliers by both rules are removed; for data points marked as outliers only by the statistical rule but still within the physical threshold range, a secondary judgment is made based on the rate of change of data before and after the specified time.
[0063] In one embodiment, the status filtering unit classifies samples into multiple categories, such as normal modeling samples, shutdown samples, power outage samples, fault samples, and maintenance samples, based on the unit's operating status word and fault alarm code. Among them, normal modeling samples are used for model training and updating; shutdown samples and fault samples do not participate in normal output modeling, but are archived separately for training the fault diagnosis model; power outage samples can be selectively used after being labeled according to the power outage ratio.
[0064] (III) Benchmark Unit Identification
[0065] like Figure 5 The diagram illustrates the benchmark turbine identification and comprehensive scoring system provided in this embodiment of the invention. The benchmark turbine identification module 120 is used to select the most representative turbine from multiple turbines of the same type in a wind farm as the modeling benchmark. The benchmark turbine is not required to be absolutely optimal in all indicators, but rather it is required to have both good data quality and strong representativeness, so that its model can be easily transferred to other turbines.
[0066] In one embodiment, the benchmark unit identification module 120 performs a comprehensive evaluation based on the following indicators: operational stability, data integrity, operating condition coverage, failure rate, power fluctuation, unit representativeness, and average similarity with other units. Operational stability can be characterized by availability, unplanned downtime, and state switching frequency; data integrity can be characterized by missing data rate, anomaly rate, and the percentage of valid samples; operating condition coverage can be characterized by the coverage of operating condition samples across different wind speed ranges, wind direction ranges, and seasons.
[0067] Furthermore, the benchmark turbine identification module 120 can construct a comprehensive scoring function, weighting the above indicators according to preset weights to obtain a comprehensive score value for each candidate turbine. The turbine with the highest score can be selected as a single benchmark turbine; if the wind farm has complex terrain or large differences in turbine locations, multiple benchmark turbines can also be selected according to terrain zoning or wind direction zoning, serving different sub-regions respectively.
[0068] In one embodiment, the benchmark unit identification module 120 constructs the following comprehensive scoring function:
[0069] Suppose that there are N identical wind turbines in the wind farm. For the i-th turbine, the following scoring index is defined:
[0070] Operational stability index S_stab_i: Calculated based on unit availability η_avail_i, unplanned downtime T_unplanned_i, and state switching frequency F_switch_i: S_stab_i = w_avail × η_avail_i + w_stop × (1 - T_unplanned_i / T_total_i) + w_switch × (1 - F_switch_i / F_max) where w_avail, w_stop, and w_switch are the weights of each sub-index, T_total_i is the total duration of the statistical period, and F_max is the maximum state switching frequency among all units in the field.
[0071] The data completeness metric S_complete_i is calculated based on the missing test rate R_miss_i, the outlier rate R_abnormal_i, and the percentage of valid samples R_valid_i: S_complete_i = w_miss × (1 - R_miss_i) + w_abnormal × (1 - R_abnormal_i) + w_valid × R_valid_i
[0072] The operating condition coverage index S_coverage_i is calculated by dividing wind speed into m intervals (e.g., 0-3m / s, 3-6m / s, ..., 21-24m / s, >24m / s) and wind direction into n sectors (e.g., 8 45° sectors), and counting the number of valid samples for each unit under each wind speed interval and wind direction sector combination. The operating condition coverage score is defined as the proportion of the number of covered operating condition combinations to the total number of combinations, and a uniformity penalty is applied to the number of samples under each operating condition combination: S_coverage_i = (C_covered_i / C_total) × (1 - λ × D_KL(P_i || U)) where C_covered_i is the number of operating condition combinations covered by the i-th unit, C_total is the total number of operating condition combinations, P_i is the sample distribution of the unit under each operating condition combination, U is a uniform distribution, D_KL is the KL divergence, and λ is the uniformity penalty coefficient.
[0073] Fault rate metric S_fault_i: Calculated based on the number of faults and fault severity per unit time: S_fault_i = 1 - (∑_{k=1}^{K} severity_k × count_k) / (K × severity_max) where K is the total number of fault types, severity_k is the severity weight of the k-th type of fault, and count_k is the number of times this type of fault occurs within the statistical period.
[0074] The representative indicator of the aircraft position, S_position_i, is calculated based on the average distance between the aircraft unit and other aircraft positions in the field, the upstream and downstream relationship with the prevailing wind, and the terrain zoning: S_position_i = w_dist × (1 - d_i / d_max) + w_wind × connectivity_i + w_terrain × (1 - terrain_penalty_i) where d_i is the average distance from the i-th aircraft unit to other aircraft units, d_max is the maximum average distance in the field, connectivity_i is the connectivity of the aircraft unit in the upstream and downstream topology with the prevailing wind, and terrain_penalty_i is the terrain zoning edge penalty term.
[0075] Average similarity index S_similar_i: Calculates the average similarity of the operating characteristics of the i-th unit with all other units: S_similar_i = (1 / (N-1)) × ∑_{j≠i} Sim(F_i, F_j) where F_i and F_j are the operating characteristic vectors of the i-th and j-th units, respectively, and Sim is the similarity measurement function (such as cosine similarity or Pearson correlation coefficient).
[0076] Finally, the comprehensive score of the i-th unit is: Score_i = α_stab × S_stab_i + α_complete × S_complete_i + α_coverage × S_coverage_i + α_fault × S_fault_i + α_position × S_position_i + α_similar × S_similar_i, where each weight coefficient satisfies α_stab + α_complete + α_coverage + α_fault + α_position + α_similar = 1.
[0077] In one implementation, the aforementioned weighting coefficients are configured by the operation and maintenance personnel based on the actual operational needs of the wind farm. For newly built wind farms or wind farms with insufficient data accumulation, the weights of data integrity rate and operating condition coverage can be appropriately increased; for mature wind farms that have been operating for many years, the weights of operational stability and average similarity can be appropriately increased.
[0078] In one embodiment, the benchmark unit identification module 120 may also use the analytic hierarchy process (AHP) to determine the weight of each scoring indicator, and obtain a weight allocation scheme that conforms to expert experience by constructing a judgment matrix, calculating the maximum eigenvalue and performing a consistency check.
[0079] In another embodiment, the benchmark unit identification module 120 can also employ cluster analysis, first grouping the units according to their location and operational characteristics, and then selecting representative units from each group as local benchmarks. This method is suitable for mountainous wind farms or wind farm scenarios where wake effects are significant.
[0080] (iv) Benchmark Model Construction
[0081] The benchmark model building module 130 is used to establish a dynamic power output model of the wind turbine based on historical operating data of the benchmark unit. The goal of the dynamic power output model is to depict the power output variation of the wind turbine over a future period under given environmental and operating conditions. Unlike static power curve modeling, the dynamic power output model emphasizes historical time series information and state evolution processes.
[0082] In one embodiment, the model input variables include: current and historical wind speed, wind direction, historical power output, rotational speed, pitch angle, ambient temperature, and unit status variables; the model output variable is the predicted power output value within a preset future time period. Depending on the needs, prediction results for multiple time steps can also be output simultaneously to meet the requirements of short-term prediction and rolling updates.
[0083] The benchmark model construction module 130 can organize training samples using a sliding time window approach. For example, it can use the most recent sampling times as the input window and the output times of the next few sampling times as the output window. This structure is suitable for both deep time series models and traditional sequence regression models.
[0084] In one embodiment, the benchmark model construction module 130 may employ a neural network, an ensemble learning model, or a hybrid model. If a neural network is used, a long short-term memory network, a gated recurrent unit network, a temporal convolutional network, a Transformer-like temporal model, or a multi-scale hybrid model may be used; if an ensemble learning model is used, gradient boosting trees may be used; if a hybrid model is used, it may combine mechanistic constraints and data-driven structures.
[0085] To reduce the difficulty of subsequent migration, the benchmark model construction module 130 preferably retains a transferable general main structure and an adaptable local parameter structure. The general main structure is used to characterize the shared output dynamics of similar units and can be widely reused during migration; the local parameter structure is used to reflect the individual differences of units and can be updated on the target unit with a small number of samples.
[0086] (v) Spatiotemporal transfer learning
[0087] The spatiotemporal transfer learning module 140 is one of the core modules of this invention, used to quickly transfer the benchmark unit model to other units. Unlike traditional transfer methods that rely solely on parameter replication, this invention emphasizes considering both spatial relationships and temporal evolution characteristics, hence the name dynamic spatiotemporal transfer learning.
[0088] Spatial correlation analysis is primarily used to characterize the physical location and incoming flow relationships between wind turbine units. In one embodiment, the spatiotemporal transfer learning module 140 constructs a spatial correlation matrix based on the wind turbine unit location coordinates, relative distance, prevailing wind direction, terrain elevation difference, and wake propagation path. This spatial correlation matrix can reflect the degree to which the target unit and the benchmark unit share wind resource conditions.
[0089] In one embodiment, the spatial correlation matrix C_space constructed by the spatiotemporal transfer learning module 140 is an N×N matrix (N is the total number of units), where the element c_{ij} represents the degree of spatial correlation between the i-th unit and the j-th unit, and the value range is [0,1].
[0090] The calculation of c_{ij} takes into account the following factors:
[0091] (1) Distance factor d_factor_{ij}: Calculated based on the Euclidean distance between units: d_factor_{ij} = exp(-α_d × L_{ij} / L_ref) where L_{ij} is the spatial distance between the i-th and j-th units, L_ref is the reference distance (such as the average distance between all units in the field), and α_d is the distance attenuation coefficient.
[0092] (2) Wind direction consistency factor w_factor_{ij}: calculated based on the angle between the main wind direction and the line connecting the units: w_factor_{ij} = cos²(θ_{ij} - θ_main) where θ_{ij} is the direction angle from the i-th unit to the j-th unit, and θ_main is the direction angle of the main wind direction of the wind farm.
[0093] (3) Wake factor: Based on the upstream and downstream relationship between units and the wake model, wake_factor_{ij} = 1 - wake_loss_{j←i} where wake_loss_{j←i} represents the estimated wake loss of the i-th unit to the j-th unit, which can be calculated using the Jensen wake model or a more complex wake model.
[0094] (4) Terrain zone factor t_factor_{ij}: If the two units are located in the same terrain zone (such as the same ridge or the same slope), then t_factor_{ij}=1; otherwise, t_factor_{ij}=t_decay(0 <t_decay<1)。
[0095] Comprehensive spatial correlation: c_{ij} = w_d × d_factor_{ij} + w_w × w_factor_{ij} + w_wake × wake_factor_{ij} + w_t × t_factor_{ij} where w_d, w_w, w_wake, and w_t are the weights of each factor, and their summation is 1.
[0096] Operational feature similarity analysis is mainly used to characterize the dynamic similarity of units within a recent operating window. In one embodiment, the spatiotemporal transfer learning module 140 extracts statistical features, fluctuation features, state distribution features, and error features from the most recent sliding time windows of the target unit and the benchmark unit, and calculates operational similarity based on these features.
[0097] In one embodiment, the spatiotemporal transfer learning module 140 extracts the following statistical features from the most recent sliding time windows of the target unit and the benchmark unit:
[0098] Time-series statistical characteristics include: mean wind speed, standard deviation of wind speed, mean wind speed change rate, mean power output, standard deviation of power output, mean power output change rate, mean speed, and mean pitch angle.
[0099] Fluctuation characteristics include peak-to-peak power output, fluctuation intensity (cumulative absolute value of power output change per unit time), proportion of high-frequency component energy, and proportion of low-frequency component energy.
[0100] State distribution characteristics: including the time percentage of the unit in each operating state (start-up, grid connection, rated, power restriction, shutdown, etc.) and the frequency of state transitions.
[0101] Operating condition distribution characteristics: including the percentage of operating time of the unit in each wind speed range and power range, operating condition transition matrix, etc.
[0102] Let the feature vector of the benchmark unit be F_source and the feature vector of the target unit be F_target. Then the operational feature similarity Sim_run is: Sim_run = ρ(F_source, F_target), where ρ is the Pearson correlation coefficient or cosine similarity.
[0103] In one embodiment, to further improve the accuracy of similarity calculation, the above features can be reduced in dimensionality by principal component analysis (PCA) before similarity calculation, so as to remove redundant information and noise interference.
[0104] Furthermore, the spatiotemporal transfer learning module 140 can integrate spatial correlation and operational feature similarity into a comprehensive transfer weight. For target units with high comprehensive transfer weights, the system can directly inherit the model parameters of multiple benchmark units and adjust only the local parameters; for target units with low comprehensive transfer weights, the learning weight of local samples is increased to avoid inappropriate strong transfer causing negative transfer.
[0105] In one implementation, the spatiotemporal transfer learning module 140 fuses the spatial correlation matrix C_space and the runtime similarity Sim_run into a comprehensive transfer weight W_transfer:
[0106] W_transfer(i, j) = β_space × c_{ij} + β_run × Sim_run(i, j) where β_space and β_run are fusion weights, satisfying β_space + β_run = 1.
[0107] The values of β_space and β_run can be dynamically adjusted according to the data availability of the target unit: when the number of effective samples of the target unit is sufficient, β_run (operational similarity weight) can be appropriately increased to make the migration more dependent on recent actual operation characteristics; when the number of effective samples of the target unit is insufficient, β_space (spatial correlation weight) can be appropriately increased to make the migration more dependent on physical spatial relationships.
[0108] In one implementation, the integrated migration weight matrix W_transfer is normalized so that for each target unit j, the sum of its migration weights with those of all benchmark units is 1.
[0109] In scenarios involving multiple benchmark units, the spatiotemporal transfer learning module 140 can also perform multi-source transfer based on the comprehensive transfer weights between the target unit and multiple benchmark units. Specifically, the parameters or intermediate representations of multiple benchmark models can be weighted and fused to generate the initial model of the target unit.
[0110] In one implementation, in a multi-benchmark scenario (with M benchmark units), the initial model parameters θ_j of the target unit j are obtained in the following way:
[0111] θ_j = ∑_{k=1}^{M} w_transfer(k, j) × θ_k^{adapt} where θ_k^{adapt} is the adaptation layer parameter of the k-th benchmark unit model (the main layer parameter is directly reused), and w_transfer(k, j) is the comprehensive migration weight from the k-th benchmark unit to the target unit j.
[0112] For benchmark-target unit pairs with a comprehensive transfer weight higher than the first threshold (e.g., 0.7), the system directly reuses the main layer parameters of the benchmark unit and only makes minor adjustments to the adaptation layer. For benchmark-target unit pairs with a comprehensive transfer weight between the first and second thresholds (e.g., 0.3), the system reuses the main layer parameters and increases the training rounds of the adaptation layer. For benchmark-target unit pairs with a comprehensive transfer weight lower than the second threshold, the system adopts a weak transfer strategy—using only the benchmark model as an initialization reference and mainly relying on the local data of the target unit for training.
[0113] (vi) Trigger determination
[0114] The trigger determination module 150 is used to determine whether an update should be performed on the target unit model. This invention does not require fixed high-frequency updates, but rather employs an on-demand update strategy that combines time-triggered and event-triggered updates. This strategy achieves a balance between accuracy and energy consumption.
[0115] Time-triggered mechanisms are used for periodic updates. For example, updates can be performed daily, weekly, monthly, quarterly, or by rolling boundaries within a sliding time window. Time-triggered mechanisms are suitable for handling slowly changing seasonal wind resource variations and long-term wear effects.
[0116] Event triggering mechanisms are used to respond to sudden or significant changes in operating conditions. Event triggers can be caused by one or more of the following conditions: prediction errors continuously exceeding a threshold, unit maintenance, blade cleaning or replacement, sensor calibration, control parameter adjustment, shutdown recovery, power outage switching, arrival or departure of extreme wind conditions, and a significant increase in data distribution drift.
[0117] In one embodiment, the trigger determination module 150 continuously monitors the error statistics sequence and event logs of each target unit. When the error of a target unit exceeds a threshold for multiple consecutive time periods, or when a critical maintenance event appears in the event log, an update instruction is immediately generated.
[0118] In one embodiment, the trigger determination module 150 uses a sliding window error statistics method to determine whether the prediction error continuously exceeds the threshold: the error monitoring window length is set to W_e (e.g., prediction samples from the most recent 24 hours), and the absolute percentage error (APE) or root mean square error (RMSE) of each prediction point within the window is statistically analyzed. When the APE of N_e consecutive (e.g., 6 consecutive) prediction points within the window exceeds a preset threshold T_e (e.g., 15%), or the average RMSE within the window exceeds a preset threshold T_rmse, it is determined that the prediction error continuously exceeds the threshold, triggering a model update.
[0119] In one implementation, the error thresholds T_e and T_rmse are dynamically set based on the baseline error of the benchmark turbine on the validation set: T_e = μ_base + k × σ_base, where μ_base is the mean of the error in the benchmark turbine validation set, σ_base is the standard deviation, and k is an adjustable coefficient (default value is 2~3). This dynamic threshold setting method can adapt to the differences in error levels in different wind farms and different seasons.
[0120] In one embodiment, the trigger determination module 150 quantifies the determination of wind resource mutations by: setting the wind resource monitoring window length to W_wind (e.g., the last 6 hours), and extracting the average wind speed V_avg, wind speed standard deviation V_std, prevailing wind direction θ_main, and turbulence intensity TI within the window. A wind resource mutation is determined when any of the following conditions are met:
[0121] (1) The deviation between the current wind speed and the average wind speed of the window exceeds a preset percentage (e.g., 30%); (2) The angle between the current wind direction and the main wind direction of the window exceeds a preset angle (e.g., 60°); (3) The deviation between the current turbulence intensity and the average turbulence intensity of the window exceeds a preset percentage (e.g., 50%); (4) The rate of change of wind speed exceeds a preset threshold (e.g., 5 m / s / 10 min).
[0122] In one embodiment, the trigger determination module 150 detects data distribution drift using the following method: It maintains the historical data distribution of the target unit for the most recent N_day days (e.g., 30 days) as the reference distribution P_ref, and uses the data distribution for the most recent N_hour days (e.g., 7 days) as the current distribution P_cur. The distance between P_ref and P_cur is calculated using KL divergence or maximum mean difference (MMD). When this distance exceeds a preset threshold, it is determined that a significant data distribution drift has occurred, triggering a model update.
[0123] In one implementation, data distribution drift detection is performed on the following key variables: wind speed distribution, power distribution, combined wind speed and power distribution, pitch angle distribution, and rotational speed distribution. An update is triggered when the distribution drift of any variable exceeds a threshold.
[0124] (vii) Sliding Time Window Management
[0125] The sliding time window management module 160 is used to maintain the timeliness of data used in the modeling and maintenance of the target wind turbine. Because wind turbine operating conditions exhibit significant temporal evolution, consistently using excessively long historical data for training can easily lead to the model over-reliance on old operating conditions and neglect of recent changes. Therefore, a sliding time window mechanism is needed to ensure the time freshness of the samples.
[0126] In one embodiment, the sliding time window management module 160 can set the most recent 7 days, 15 days, 30 days, 90 days, or 180 days as candidate window lengths, and automatically select the window length currently used for updating based on the unit's operational stability and the rate of seasonal change. For scenarios with significant seasonal changes, a longer window can be used; for scenarios with frequent maintenance or sudden changes in operating conditions, a shorter window can be used.
[0127] The sliding time window management module 160 adds the latest sample to the window when new data arrives, removes the oldest sample from the window, and updates the sample weights within the window. To enhance responsiveness to recent operating conditions, samples within the window can be weighted using time decay, with samples closer to the current time having higher weights.
[0128] In one embodiment, the sliding time window management module 160 adopts a three-layer window structure:
[0129] Short window: The length is the most recent 3 to 7 days, with the highest weight (e.g., 0.6), used to capture the rapidly changing operating characteristics of the unit in the near future;
[0130] Medium window: The length is the most recent 15 to 30 days, with a medium weight (e.g., 0.3), used to capture the medium-term operating trend of the unit;
[0131] Long window: The length is the most recent 60 to 90 days, with a low weight (e.g., 0.1), used to preserve the long-term historical patterns and seasonal characteristics of the unit.
[0132] The samples from the three windows together constitute the training dataset, but the sampling weights of the samples in each window are allocated according to the above proportions.
[0133] In one embodiment, the specific value of the window length is automatically adjusted according to the operating characteristics of the wind turbine: for turbines that operate smoothly and have slow changes in operating conditions, a longer window can be used (short window 7 days, medium window 30 days, long window 90 days); for turbines that have large fluctuations in operation and frequent changes in operating conditions, a shorter window can be used (short window 3 days, medium window 15 days, long window 60 days).
[0134] In one embodiment, the sliding time window management module 160 assigns a comprehensive weight w_sample to the samples within the window, which is composed of the product of the following three parts:
[0135] The time decay weight w_time is calculated as: w_time(t) = exp(-γ × (t_cur - t) / T_window), where t_cur is the current time, t is the sample time, T_window is the total window length, and γ is the time decay coefficient (default value is 1~2). Samples closer to the current time have a higher time decay weight.
[0136] The similarity weight of the working conditions, w_cond, is: w_cond(t) = exp(-δ × ||F(t) - F_cur|| 2 ), where F(t) is the operating condition feature vector of sample at time t (including wind speed, wind direction, power output level, etc.), F_cur is the current operating condition feature vector, and δ is the operating condition similarity decay coefficient. The more similar the sample is to the current operating condition, the higher the operating condition similarity weight.
[0137] Data quality weight w_quality: w_quality(t) = 1 - R_miss(t) - R_abnormal(t), where R_miss(t) is the missing data rate of sample t in the time period, and R_abnormal(t) is the anomalous rate. Samples with higher data quality have a higher data quality weight.
[0138] Final sample weighting: w_sample(t) = w_time(t) × w_cond(t) × w_quality(t).
[0139] In one embodiment, the sliding time window management module 160 performs window scrolling updates when any of the following conditions are met: (1) the amount of new data accumulated reaches a preset threshold (e.g., 1,000 new valid samples are added); (2) the time since the last window update reaches a preset period (e.g., 24 hours); (3) the window update is performed synchronously when an event triggers the update.
[0140] (viii) Incremental learning
[0141] The incremental learning module 170 is used to perform lightweight maintenance on the target unit model based on the new data after the update conditions are met. Compared with retraining the complete model, incremental learning only updates some parameters or lightweight adaptation layers, thus reducing training time and edge device resource consumption.
[0142] In one implementation, the incremental learning module 170 updates only the prediction head parameters and adaptation layer parameters of the target unit model, while freezing the main feature extraction layer. This preserves the shared temporal knowledge of the benchmark model while incorporating individual variations of the target unit.
[0143] In one embodiment, the output dynamic model constructed by the benchmark model construction module 130 adopts the following hierarchical structure:
[0144] Main feature extraction layer: Composed of multiple temporal feature extraction modules (such as LSTM layers, GRU layers, or Transformer encoding layers), this layer is responsible for extracting general temporal features from the input time series data. The parameters of this layer are completely frozen during transfer learning and do not participate in incremental updates.
[0145] Adaptation Layer: Composed of 1-3 fully connected network layers, this layer is responsible for adapting the general features extracted by the main layer to the personalized feature space of a specific unit. The parameters of this layer are initialized during the transfer process based on the comprehensive transfer weights and updated during incremental updates.
[0146] Prediction layer: This is the output layer, responsible for mapping the output of the adaptation layer to the final predicted output value. The parameters of this layer are updated during incremental updates.
[0147] During incremental updates, only the parameters of the adaptation and prediction layers participate in gradient updates, while the parameters of the main feature extraction layer remain frozen. This strategy ensures that the number of trainable parameters for each incremental update is only 5% to 20% of the total number of model parameters, significantly reducing the computational cost of updates.
[0148] In one implementation, the historical playback unit of the incremental learning module 170 maintains a fixed-size playback sample pool (e.g., 2000-5000 samples), which employs the following update strategy:
[0149] Sample selection strategy: From the samples within the sliding time window, samples are selected to be added to the replay pool based on the principles of maximizing operational coverage and representativeness. Specifically, the k-centroid algorithm or core set selection algorithm is used to select the subset of samples that best represents the overall data distribution from the samples in the current window.
[0150] Sample replacement strategy: When a new sample is added, if the replay pool is full, the sample is replaced according to its "contribution"—the sample with the lowest contribution is eliminated first. The contribution can be calculated based on the rarity of the sample's operating conditions, its contribution to model performance, and the sample's timeliness.
[0151] Replay Participation Strategy: During each incremental update, a number of replay samples, roughly equal to the number of new samples, are randomly selected from the replay pool and combined with the new samples to form a training batch. The loss function weights of the replay samples can be set lower than those of the new samples (e.g., a weight of 1.0 for new samples and a weight of 0.5 for replay samples) to appropriately emphasize recent changes while preserving historical knowledge.
[0152] In one embodiment, the trigger suppression mechanism of the incremental learning module 170 is used to avoid wasting computing power due to repeated triggering within a short period of time. Specifically:
[0153] Set a minimum update interval T_min (e.g., 1 hour), with at least T_min of time remaining between two updates;
[0154] Set a window T_merge for merging similar events (e.g., 10 minutes). Similar triggering events occurring within the T_merge window will be merged into a single update.
[0155] Establish an update priority queue, sorted by event urgency: maintenance event recovery > prediction error exceeding limits > sudden wind resource changes > time-triggered events.
[0156] In one embodiment, the version rollback mechanism of the incremental learning module 170 is executed according to the following process: (1) Before each incremental update, a complete copy of the parameters of the current model is saved as the "previous version"; (2) After the incremental update is performed, the performance of the updated model is evaluated on the validation set; (3) If the key performance indicators (such as RMSE, MAE) of the updated model deteriorate by more than a preset threshold (such as 5%) compared with the previous version, the rollback is automatically triggered and restored to the previous version; (4) If the performance of the updated model is improved, the previous version is retained as the "historical version" for subsequent rollback as needed; (5) The system retains the most recent K (such as 10) historical versions, and when there are more than K, the oldest version is eliminated.
[0157] In one implementation, the rollback operation also records the rollback reason, triggering conditions, and timestamp, forming a rollback log for operation and maintenance personnel to analyze.
[0158] In one embodiment, after the trigger determination module 150 determines that a model update is needed and before the incremental learning module 170 performs the parameter update, the system performs the following data quality verification and operational condition validity screening steps:
[0159] Valid grid connection status verification: Check the operating status word of the target unit during the period to be updated. Only when the unit is in "grid-connected generation" status (not out of service, not under maintenance, not due to fault, not subject to power curtailment) will the corresponding sample be marked as valid. For samples in power curtailment status, if the curtailment ratio is known and stable, they can be selectively used after the curtailment ratio is marked.
[0160] Key measurement point integrity verification: Check whether the data of each key measurement point (at least including wind speed, wind direction, active power output, and rotational speed) in the sample to be updated are complete. If the missing rate of any key measurement point exceeds a preset threshold (e.g., 5%) within the sample period, the sample is marked as invalid.
[0161] Timestamp continuity check: Checks whether the timestamps of the samples to be updated are continuous. If there are timestamp jumps within a sample period (intervals exceeding twice the preset sampling period), the samples in that period are marked as suspicious and require further investigation to determine if there is data loss or communication interruption.
[0162] Physical quantity range rationality verification: Check whether the values of each measuring point are within the physically reasonable range. For example, wind speed should be within the range of 0~40m / s, active power output should be within the range of 0~rated power, rotational speed should be within the range of 0~rated rotational speed, and propeller pitch angle should be within the range of 0~90°. Data points that exceed the physical range are marked as abnormal and removed.
[0163] Operating condition validity screening: Only samples with wind speeds between the cut-in and cut-out wind speeds and where the unit is in normal power generation conditions are retained. Samples with wind speeds lower than the cut-in wind speed or higher than the cut-out wind speed, as well as samples where the unit is in abnormal power generation conditions such as startup, shutdown, or fault, are isolated or archived separately and are not included in the incremental update process.
[0164] In another embodiment, the incremental learning module 170 also includes a history playback unit. This unit stores a small number of representative old samples and participates in training together with new samples during each incremental update, thereby mitigating the model's forgetting of old conditions.
[0165] Furthermore, the incremental learning module 170 can incorporate a version management mechanism to record the model version number, update time, triggering reason, usage window length, and key performance indicators after each update. When the update effect is not as expected, it can be rolled back to the previous stable version.
[0166] (ix) Model Deployment
[0167] The model deployment module 180 is used to quickly distribute updated models to wind farm monitoring systems, edge nodes, tower-based industrial control computers, or cloud-edge collaborative platforms. The model deployment module 180 supports batch model generation, automatic model replacement, unified model management, and status feedback.
[0168] In one implementation, when performing batch deployment, the model deployment module 180 selects an appropriate deployment period based on the wind farm's computing and communication loads. For example, model updates can be performed centrally during low-load periods at night, low-computing-power periods, or preset low-carbon scheduling periods to reduce additional energy consumption during deployment.
[0169] The model deployment module 180 can also synchronously upload the update time, deployment status, rollback status and operation indicators of the target unit model to the operation and maintenance platform, which facilitates the unified management of the entire field model by operation and maintenance personnel.
[0170] (x) Methods and Procedures
[0171] like Figure 2 As shown, the method of the present invention can be specifically executed according to the following process:
[0172] S101 performs data acquisition; the system collects multi-source data from a group of wind turbines of the same type within the wind farm, including operation data, spatial location data, and operation and maintenance event data, and performs time alignment and anomaly processing on the data;
[0173] S102 selects benchmark units; the system identifies benchmark units from similar units using a comprehensive scoring method;
[0174] S103 performs benchmark unit modeling; the system trains a basic output dynamic model based on the historical operating data of the benchmark unit.
[0175] S104 performs spatiotemporal transfer learning; the system generates transfer weights based on the spatial correlation matrix, operational similarity matrix and the most recent sliding time window data of the target unit, and transfers the basic output dynamic model of the benchmark unit to each target unit.
[0176] S105 enables multi-unit model generation; the system generates initial models for each target unit in batches, serving as the basis for online deployment.
[0177] S106 performs time-triggered updates; the system decides whether to initiate model updates based on the time-triggered conditions.
[0178] S107 performs an event-triggered update; the system decides whether to initiate a model update based on the event triggering conditions.
[0179] S108 adopts sliding time window management; it maintains the data window of the target unit for the most recent preset duration, and assigns weights to the samples within the window based on data timeliness, operating condition similarity, and sample quality;
[0180] S109 performs incremental learning maintenance; the system performs incremental learning updates on the target unit model.
[0181] The S110 completes model deployment; the system publishes the new model to deployment platforms such as the power plant monitoring platform, edge computing nodes, industrial control computers, or cloud-edge collaboration platforms, and records the version and running results.
[0182] (xi) Dynamic updates and model maintenance
[0183] In a preferred embodiment, the dynamic update closed loop includes data quality verification and operating condition validity screening steps before performing model updates. The system preferably only performs parameter updates on samples that meet the requirements of valid grid connection, complete key measurement points, continuous timestamps, and reasonable physical quantity ranges. Samples that are experiencing prolonged shutdowns, severe faults, communication interruptions, or sensor anomalies can be isolated, marked, or archived separately to prevent invalid samples from entering the update process and contaminating the target unit model.
[0184] Furthermore, the model maintenance management module can establish an independent maintenance file for each target unit. This file records at least the trigger type, trigger time, usage window length, number of incremental samples, key error indicators before and after the update, rollback flag, and corresponding operation and maintenance event number. Using this maintenance file, operation and maintenance personnel can trace the basis and effects of any round of model updates, providing support for subsequent parameter tuning, benchmark replacement, and operation and maintenance responsibility analysis.
[0185] In one specific embodiment, the dynamic update mechanism of the present invention is not a simple fixed-period retraining, but a closed-loop maintenance system composed of a time-triggered mechanism, an event-triggered mechanism, a sliding time window management mechanism, and an incremental learning mechanism. For example... Figure 4 As shown, the time-triggered mechanism is used to perform periodic refreshes during the stable phase of model operation, while the event-triggered mechanism is used to quickly initiate updates when prediction errors suddenly increase, maintenance events occur, power rationing status changes, shutdowns resume, control parameters are adjusted, or wind resources change significantly, thereby achieving a balance between response speed and computational cost.
[0186] Specifically, the time-triggered mechanism can be triggered on a daily, weekly, monthly, or preset window rolling cycle to address seasonal wind resource changes, long-term wear accumulation, and slow control strategy drift. The event-triggered mechanism can determine the trigger based on error statistics and operation and maintenance logs. When the error of the target unit exceeds the threshold in multiple consecutive windows, or when critical events such as blade maintenance, sensor calibration, component replacement, or grid connection restoration occur, the system immediately initiates a partial update for that unit. To avoid wasting computing power due to repeated triggering in a short period, trigger suppression intervals and priority queues can be set to perform merged judgments on similar events.
[0187] To ensure the model consistently reflects the recent operating status of the target unit, the sliding time window management module continuously maintains a data window for the target unit with the most recent preset duration. This data window can be a single window or a multi-scale window structure combining short, medium, and long windows. Samples closer to the current time, more similar to the current operating conditions, and of higher quality can be assigned higher weights; samples with older histories, abnormal conditions, or weak representativeness can have their weights reduced or be gradually eliminated. This design enhances the model's sensitivity to recent changes in operating conditions while retaining necessary historical information.
[0188] After the update conditions are met, the incremental learning module performs local parameter updates on the target unit model using only newly added samples or a combination of newly added and replayed samples. Preferably, the main layer of the base model can be frozen, and only the adaptation layer, prediction layer, or a small number of efficient fine-tuning parameters are updated. Combined with historical typical sample replay, constraint regularization terms, knowledge distillation, and version rollback mechanisms, the risk of model forgetting is reduced, and the stability of the update process is ensured. Thus, this invention can maintain the long-term applicability of the model without full retraining, making it suitable for deployment on field-side equipment with limited computing power and power consumption.
[0189] During the later maintenance phase, operation and maintenance personnel can configure the time trigger cycle, event trigger threshold, sliding time window length, sample weight rules, replay sample pool size, and incremental update parameters based on the seasonal characteristics, terrain characteristics, maintenance strategies, and operating load of the wind farm. For newly launched units, a small number of initial samples can be used to quickly complete migration initialization; for units that have undergone major overhauls, newly added samples after maintenance can be prioritized for rapid re-adaptation after event triggering; for benchmark units whose representativeness has declined after long-term operation, the benchmark unit selection process can be re-executed to complete benchmark replacement and model system reconstruction.
[0190] (xii) Spatiotemporal migration and low-carbon deployment
[0191] In a further embodiment, before performing the target unit migration initialization, the system can also perform a migration feasibility pre-verification on the candidate benchmark units and the target unit. The pre-verification can be based on indicators such as the consistency of unit rated parameters, the consistency of prevailing wind direction, the proximity of unit locations, the proportion of recent valid samples, and the estimation of historical error lower bounds. When the pre-verification result is lower than a preset threshold, the system can automatically reduce the migration intensity, switch to a multiple benchmark unit fusion mode, or directly prompt maintenance personnel to adopt a more conservative local update strategy.
[0192] In terms of low-carbon deployment, the model deployment module can not only perform batch replacements during low-load periods, but also dynamically adjust the concurrency and priority of update tasks based on edge device computing power reserves, CPU utilization, memory usage, network congestion levels, and temperature rise status. For units with poor communication conditions or high hardware loads, the system can first distribute a lightweight version of the model or only distribute the adaptation layer parameter package, and then supplement the complete update after the equipment load recovers, so as to ensure the stability of the deployment process.
[0193] Furthermore, in multi-wind farm collaborative management scenarios, this invention can also complete benchmark selection and target unit migration at the farm level, and complete energy efficiency statistics, policy distribution, and version archiving at the regional center or cloud, thereby constructing a layered deployment model with rapid response at the farm level and unified management at the regional level. This model retains the advantages of real-time updates at the edge while facilitating unified governance of the model lifecycle for group-based wind power assets.
[0194] In one specific implementation, the spatiotemporal transfer learning module jointly models spatial correlation and operational feature similarity. For example... Figure 6 As shown, spatial correlation can be determined by the distance between generator sites, the prevailing wind angle, the wake link, the same-area topographic attributes, and the upstream and downstream topological relationships; operational characteristic similarity can be determined by the slope of the power curve, fluctuation intensity, state distribution, error sequence similarity, and operating condition coverage within the most recent sliding time window. The system can normalize and merge these two factors into a comprehensive migration weight, thereby controlling the migration ratio of the benchmark model parameters to the target generator model.
[0195] When the combined migration weight of the target turbine and the benchmark turbine is high, the system can reuse a high proportion of shared main parameters and only adjust the local adaptation layer. When the combined migration weight is in the middle range, the number of local sample training rounds can be increased and the migration intensity can be appropriately reduced. When the combined migration weight is low, multiple benchmark turbine weighted initialization or a weak migration strategy can be adopted to reduce the risk of negative migration. For wind farms with complex terrain or significant wake effects, the system can also first establish multiple benchmark turbine sets by region or wind direction, and then generate the initial model of the target turbine based on multi-source weight fusion.
[0196] This invention not only focuses on prediction accuracy and migration efficiency, but also on energy consumption during model operation and maintenance. The model deployment module can be linked with the site energy consumption management system, computing resource monitoring module, or edge device operation monitoring module to perform batch model updates and replacements when the site load is low, communication bandwidth usage is low, edge device temperature rise is low, or during preset low-carbon time slots. In scenarios triggered by daytime emergencies, single-machine lightweight updates are prioritized to avoid additional energy consumption and business disruptions caused by large-scale concurrent updates.
[0197] Furthermore, the system can statistically analyze the training time, computing power consumption, update package size, amount of data transmitted back, and accuracy improvement benefits corresponding to each round of model maintenance, and record these as low-carbon operation indicators. Using these indicators, wind farm maintenance personnel can compare the energy efficiency of different update strategies, further reducing unnecessary repetitive training and data transmission overhead while ensuring predictive capabilities. Therefore, this invention is suitable for deployment in large-scale wind farms, centralized wind power bases, and multi-site collaborative management scenarios that emphasize low-carbon operation and operational computing power management.
[0198] In typical application scenarios, this invention is applicable to cold start modeling of new units in wind farm construction or expansion, as well as to scenarios involving seasonal changes, frequent power curtailment, concentrated maintenance events, control strategy adjustments, and benchmark reselection after multiple years of operation. Through centralized modeling of benchmark units, on-demand migration of target units, and dynamic maintenance, the model deployment cycle can be significantly shortened, and the consistency and maintainability of the group unit models can be improved.
[0199] (xiii) Deployment of electronic equipment and software
[0200] In one embodiment, the electronic device may further include a log management unit, a task scheduling unit, and an anomaly recovery unit. The log management unit is used to record model update logs, operational alarms, and communication status; the task scheduling unit is used to arrange model training, parameter distribution, and result feedback according to preset priorities; the anomaly recovery unit is used to automatically trigger rollback or switch to the previous stable version when update fails, interface times out, model verification fails, or errors abnormally increase after deployment, thereby ensuring the continuous availability of the production system.
[0201] At the software engineering implementation level, this invention preferably adopts a modular, containerized, or service-oriented encapsulation approach. Different services can collaborate through message queues, remote call interfaces, or shared time-series databases, making it suitable for both local deployment at a single airport site and collaborative cloud-edge deployment across multiple sites. By decoupling transfer learning, trigger determination, sample management, incremental updates, and deployment control, the system's subsequent scalability, testing convenience, and operational maintainability can be improved.
[0202] like Figure 3The diagram shown is a schematic representation of the structure of an electronic device or apparatus provided in an embodiment of the present invention. The electronic device may include one or more processors 301, a storage unit 302, a communication interface 303, and a bus structure 304. The processor is used to execute programs related to model migration, update trigger determination, sliding time window maintenance, incremental learning, and deployment control. The storage unit is used to store model parameters, program code, operation and maintenance logs, trigger records, and historical samples. The communication interface is used to establish data connections with a data acquisition module, a status monitoring module, an operation and maintenance platform, and edge devices.
[0203] The electronic equipment can be deployed on the wind farm master station server, the industrial control computer at the base of the tower, the edge gateway, or the cloud-edge collaborative platform. The processor can be a general-purpose CPU, DSP, FPGA, edge AI acceleration chip, or other programmable processor; the storage unit can be memory, solid-state drive, or other computer-readable storage media; the communication interface can support Ethernet, serial port, OPC, Modbus, IEC 104, IEC 61400-25, or the internal API interface of the wind farm. When the processor executes the program in the storage unit, the aforementioned dynamic spatiotemporal migration learning method for wind turbine operation can be implemented.
[0204] At the software implementation level, this invention can be encapsulated into multiple service units, including model service, event triggering service, window management service, incremental update service, deployment service, and version management service. These service units can be centrally deployed on the same device or distributed across wind farm master stations, edge nodes, and cloud platforms to adapt to the varying real-time, stability, and computing power requirements of different wind farms. In practical applications, maintenance personnel only need to configure the benchmark turbine selection strategy, time triggering cycle, event triggering threshold, sliding time window length, and incremental learning parameters; the system can then automatically complete subsequent batch migration, model updates, and deployment management.
[0205] It should be noted that this invention does not limit the specific use of a particular time-series prediction model, a particular migration weight calculation formula, or a particular trigger threshold setting method. Those skilled in the art can adapt and adjust the benchmark selection rules, migration strategies, sliding time window scales, incremental learning methods, and deployment architecture according to the different wind farm models, terrain complexity, control strategies, and operation and maintenance resource conditions, without affecting the core technical concept of this invention regarding "group migration of similar units + dynamic maintenance + low-carbon deployment".
[0206] In summary, this invention addresses the batch modeling and long-term maintenance needs of similar wind turbine clusters by constructing an integrated technical solution comprising benchmark turbine identification, basic model construction, spatiotemporal transfer learning, dual-trigger dynamic updates, sliding time window management, incremental learning maintenance, and low-carbon deployment control. This solution reduces redundant training, shortens deployment cycles, improves the long-term adaptability of the model, and balances the computational constraints at the wind farm site with the requirements of low-carbon operation, providing a stable and scalable implementation path for wind farm cluster turbine output prediction and intelligent operation and maintenance.
[0207] Those skilled in the art will understand that the module division in the above embodiments is only for illustrating the system's functional structure. In actual implementation, several modules can be merged, split, or deployed in parallel according to hardware and software conditions. All equivalent substitutions, improvements, and modifications made within the spirit and principles of this invention should be included within the scope of protection of this invention.
Claims
1. A dynamic spatiotemporal transfer learning method for wind turbine operation, characterized in that, Includes the following steps: The data acquisition steps involve collecting operational data, spatial location data, and maintenance event data from a group of similar wind turbine units within the wind farm, and performing time alignment and anomaly processing on the data. The operational data includes at least one or more of the following: wind speed, wind direction, turbine output power, speed, pitch angle, status variables, alarm variables, temperature, vibration, and maintenance records. The spatial location data includes at least turbine coordinates, relative distance, upstream and downstream topology relationships, and wake influence relationships. The maintenance event data includes at least maintenance, shutdown, power curtailment, equipment replacement, fault recovery, and control strategy switching events. The benchmark unit identification process involves selecting at least one unit from multiple similar units within the wind farm as the benchmark unit, based on factors such as unit operational stability, data integrity, operating condition coverage, failure rate, unit location representativeness, and average similarity to other units. The benchmark model construction steps involve building a basic dynamic output model based on the historical operating data of the benchmark unit. The basic dynamic output model is used to characterize the dynamic mapping relationship between environmental variables, historical output, operating status parameters and future output. The spatiotemporal transfer learning step combines the spatial correlation between wind turbine units, the similarity of operating characteristics, and the most recent sliding time window data of the target unit to transfer the basic output dynamic model of the benchmark unit to the target unit. The trigger determination step determines whether to perform a model update based on either a time-triggered mechanism or an event-triggered mechanism. The time-triggered mechanism is used to periodically refresh the model, while the event-triggered mechanism is used to trigger an update after the prediction error exceeds the limit, a maintenance event occurs, wind resources change abruptly, power curtailment status changes, shutdown recovery, or sensor calibration. The event-triggered mechanism is triggered by any of the following conditions: prediction error continuously exceeds the threshold, unit undergoes maintenance, blade cleaning or replacement, sensor calibration, control parameter adjustment, shutdown recovery, power curtailment status change, arrival or departure of extreme wind conditions, or a significant increase in data distribution drift. The sliding time window management process involves maintaining a data window for the target unit with the most recent preset duration, and assigning weights to samples within the window based on data timeliness, operating condition similarity, and sample quality. The incremental learning step involves using only newly added samples or a combination of newly added and replay samples to perform local parameter updates on the target unit model after an update is triggered, in order to reduce update overhead and alleviate model forgetting. The model deployment steps involve batch publishing the updated model to the wind farm monitoring platform, edge computing nodes, industrial control computers, or cloud-edge collaboration platforms.
2. The method according to claim 1, characterized in that, The benchmark unit identification step adopts a weighted scoring strategy, which uses operational stability, data integrity rate, failure rate, operating condition coverage, unit representativeness and average similarity as scoring indicators to rank candidate units and select one or more units with the highest scores as benchmark units.
3. The method according to claim 1, characterized in that, The spatiotemporal transfer learning steps include: A spatial correlation matrix is constructed based on the wind turbine location coordinates, relative distance, prevailing wind direction, terrain elevation difference, and wake propagation path; Statistical features, fluctuation features, state distribution features, and error features are extracted from the most recent sliding time windows of the target unit and the benchmark unit, and the similarity of operating features is calculated. The spatial correlation and the operational feature similarity are integrated into a comprehensive migration weight, and the migration ratio of the benchmark model parameters to the target unit model is controlled according to the comprehensive migration weight.
4. The method according to claim 3, characterized in that, In the scenario of multiple benchmark units, the spatiotemporal transfer learning step further includes: weighted fusion of the parameters or intermediate representations of multiple benchmark models based on the comprehensive transfer weight between the target unit and multiple benchmark units to generate a multi-source fusion initial model of the target unit.
5. The method according to any one of claims 1-4, characterized in that, In the sliding time window management step, the data window adopts a multi-scale window structure that combines short windows, medium windows and long windows. Samples that are closer to the current time, more similar to the current working conditions and have higher data quality are given higher weights.
6. The method according to claim 5, characterized in that, The incremental learning step employs an efficient fine-tuning strategy of freezing the main layer and updating only the parameters of the adaptation and prediction layers, and introduces a historical sample replay mechanism, a trigger suppression mechanism, and a version rollback mechanism to improve the stability of model updates.
7. The method according to claim 6, characterized in that, Before the triggering judgment step, there are also data quality verification and operating condition validity screening steps. Parameter updates are only performed on samples that meet the requirements of valid grid connection status, complete key measurement points, continuous timestamps, and reasonable range of physical quantities.
8. The method according to claim 6, characterized in that, In the model deployment step, the concurrency and priority of the update tasks are dynamically adjusted according to the computing load of the site, communication load, computing power margin of edge devices, CPU utilization, memory usage, network congestion level and temperature rise status, and batch model updates are executed in a concentrated manner during low load periods.
9. A dynamic spatiotemporal transfer learning system for wind turbine operation, characterized in that, include: The data acquisition module is used to collect operational data, spatial location data, and maintenance event data of the same type of wind turbine group within the wind farm; The benchmark unit identification module is used to select at least one unit as a benchmark unit from multiple similar units in the wind farm based on the unit's operational stability, data integrity, operating condition coverage, failure rate, unit location representativeness, and average similarity with other units. The benchmark model construction module is used to construct a basic dynamic output model based on the historical operating data of the benchmark unit. The basic dynamic output model is used to characterize the dynamic mapping relationship between environmental variables, historical output, operating status parameters and future output. The spatiotemporal transfer learning module is used to transfer the basic model of the benchmark unit to the target unit by combining the spatial correlation between wind turbine units, the similarity of operating characteristics, and the most recent sliding time window data of the target unit. The trigger determination module is used to determine whether to perform a model update based on a time-triggered mechanism or an event-triggered mechanism. The sliding time window management module is used to maintain the data window of the target unit for the most recent preset duration, and to assign weights to the samples within the window based on data timeliness, operating condition similarity and sample quality. The incremental learning module is used to perform local parameter updates on the target unit model using only the newly added samples or a combination of newly added and replay samples after an update is triggered. The model deployment module is used to batch publish the updated model to the wind farm monitoring platform, edge computing nodes, industrial control computers, or cloud-edge collaboration platform. and / or The spatiotemporal transfer learning module constructs a wind turbine spatial correlation matrix and a turbine feature similarity matrix, and merges the two into a comprehensive transfer weight matrix to control the transfer ratio of benchmark model parameters to the target turbine model.
10. The system according to claim 9, characterized in that, It also includes a model maintenance and management module, which records model version, update log, trigger reasons, rollback flags, and performance evaluation results to support the maintenance of the model throughout its entire lifecycle.
11. A dynamic spatiotemporal transfer learning device for wind turbine operation, characterized in that, It includes a processor, a memory, and a communication interface. When the processor executes a program stored in the memory, it implements the method as described in any one of claims 1 to 8.
12. An electronic device, characterized in that, It includes one or more processors, a storage unit, and a communication interface, wherein when the processor executes a program in the storage unit, it implements the method as described in any one of claims 1 to 8.
13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.
