Offshore wind power intelligent prediction method and system fusing device operation and maintenance data

Abstract

The application provides a kind of offshore wind power intelligent prediction method and system of fusion equipment operation data, belong to wind power prediction technical field.The method obtains three kinds of multi-modal data of wind turbine operation time sequence, meteorological sea state and equipment operation text;The semantic analysis and structured extraction of operation text are carried out, the event time interval is aligned with each prediction window, and the time overlap ratio is calculated to generate the comprehensive event influence coefficient corresponding to each prediction step;Event-driven modulation and multi-modal feature fusion are carried out combined with time sequence modeling results, an event constraint loss function containing power prediction error, event time constraint and text semantic guidance constraint is used to train the model, and multi-step wind power prediction results are output.The application can effectively fuse unstructured operation information, improve the accuracy and interpretability of ultra-short-term power prediction in complex scenarios.

Description

[0001] This invention belongs to the field of wind power prediction technology, specifically relating to an intelligent prediction method and system for offshore wind power that integrates equipment operation and maintenance data. Background Technology

[0002] As the installed capacity of offshore wind power continues to expand, offshore wind power forecasting results are of great significance for power system dispatching, wind farm operation control, energy storage coordination, and power balance. Ultra-short-term wind power forecasting typically focuses on power change trends over the next few minutes to hours, requiring forecasting models to reflect the impact of wind turbine operating status and external environmental conditions on wind power in a timely manner.

[0003] Existing wind power prediction methods typically use historical wind turbine power, wind speed, wind direction, or numerical weather forecast data as primary inputs, employing statistical models, machine learning models, or deep learning models for prediction. While these methods can achieve certain predictive results under normal operating conditions, in actual offshore wind power operation, wind power is not only affected by meteorological and sea conditions but also significantly impacted by equipment maintenance events such as planned maintenance, temporary maintenance, fault alarms, outages, power limitations, site power curtailment, component replacement, and reset / recovery. These maintenance events are usually presented as unstructured text such as work orders, alarm records, maintenance records, outage records, power limitation records, inspection records, and operation logs. These texts contain information such as event type, event trigger time, event end time, event object, event level, and processing status. However, traditional power prediction models struggle to directly utilize this semantic information.

[0004] Specifically, existing methods have the following shortcomings: First, traditional time-series forecasting models are mainly inputted with numerical operational data and meteorological data, which cannot receive and process unstructured operation and maintenance text, resulting in the omission of event information such as maintenance, downtime, faults, and power limitation in the forecast. Second, even if some methods attempt to introduce operation and maintenance events as discrete markers into the model, they lack a mechanism to accurately align operation and maintenance events with each forecast time window, and cannot characterize the differentiated impact of the same operation and maintenance event on different forecast steps. Third, existing methods lack means to quantify the degree and direction of the impact of operation and maintenance events, the model prediction correction process lacks interpretability, and it is difficult to guarantee the reliability of predictions in complex operation and maintenance scenarios.

[0005] Therefore, there is an urgent need for an offshore wind power ultra-short-term power prediction scheme that can effectively integrate wind turbine operation sequence data, meteorological and sea condition data, and equipment operation and maintenance text information, so as to realize the effective utilization and accurate correction of operation and maintenance event information and improve the accuracy and interpretability of ultra-short-term power prediction in complex operation and maintenance scenarios. Summary of the Invention

[0006] To address the shortcomings and deficiencies of existing technologies, this invention provides an intelligent prediction method and system for offshore wind power by integrating equipment operation and maintenance data. By fusing three types of heterogeneous data—wind turbine operation sequence data, meteorological and sea condition data, and equipment operation and maintenance text information—into a unified prediction framework, the model can extract event semantic information from unstructured operation and maintenance text and transform it into feature representations that can participate in numerical calculations. Specifically, this invention performs semantic parsing and structured event extraction on the equipment operation and maintenance text information to obtain structured event descriptions and text semantic feature vectors containing fields such as event trigger time, event end time, event type, event level, event object, and processing status. The invention precisely aligns the event time interval of each operation and maintenance event with each prediction time window, calculates the time overlap ratio, and combines the matching relationship between event attributes and event objects and the wind turbines to be predicted to generate a comprehensive event impact coefficient corresponding to each prediction step. This coefficient automatically takes a value of zero when there is no time overlap between the operation and maintenance event and the prediction time window, or when there is no association between the operation and maintenance event and the prediction object. This invention uses wind turbine operating sequence data and meteorological and sea state data to perform time-series modeling to obtain an initial time-series prediction representation. It then uses a comprehensive event impact coefficient and textual semantic features to perform gating modulation and residual correction on the initial time-series prediction representation, resulting in a time-series enhanced feature that incorporates the impact of maintenance events. A two-stage gating fusion structure is used to hierarchically fuse multimodal coding features with the time-series enhanced feature, generating a base power prediction value, event gating weights, and residual corrections. Finally, the output power prediction value is adjusted by the comprehensive event impact coefficient, event gating weights, and residual corrections. When there are no maintenance events, the prediction value automatically degenerates into the base power prediction value. This invention also constructs an event-constrained loss function that includes a power prediction error loss term, an event time constraint term, and a textual semantic guidance constraint term. By referencing the event impact strength, the comprehensive event impact coefficient is constrained to maintain consistency with the time overlap relationship. Furthermore, by referencing the event gating strength and power impact direction, the event gating weights and residual corrections are constrained to maintain consistency with the maintenance textual semantics, thereby improving the interpretability and reliability of the prediction model in complex maintenance scenarios.

[0007] The specific technical solution adopted by this invention to solve its technical problem is as follows:

[0008] This invention first provides an intelligent prediction method for offshore wind power that integrates equipment operation and maintenance data. It aims to achieve effective utilization of unstructured equipment operation and maintenance text information, complete the accurate time sequence alignment and impact quantification of operation and maintenance events with different prediction step lengths, construct an event-driven prediction correction mechanism with interpretability, and improve the accuracy and reliability of ultra-short-term power prediction in complex operation and maintenance scenarios.

[0009] It acquires and preprocesses multimodal data of offshore wind turbines, including turbine operation sequence, meteorological and sea state information, and equipment maintenance text information; performs semantic parsing and structured extraction on the maintenance text to obtain text semantic features and structured event fields containing event start and end times, type, level, object, and processing status; aligns the event time intervals with each prediction window, calculates the time overlap ratio, and combines the structured event fields with the matching relationship between wind turbine objects to generate single event impact coefficients and aggregate them to obtain comprehensive event impact coefficients for each step; the coefficient is zero when there is no time overlap or no object association; and compiles the operation sequence and meteorological and sea state data... The system obtains an initial time-series prediction representation based on the coding features and modulates the time-series prediction using the comprehensive event influence coefficient and textual semantic features. It then fuses the multimodal coding features and the time-series enhancement features to generate a base power prediction value, event gating weights, and residual corrections. An event correction term is constructed based on the comprehensive event influence coefficient, event gating weights, and residual corrections, and this term is superimposed on the base power prediction value to obtain the final wind power prediction value. Finally, an event constraint loss function is constructed, including a power prediction error loss term, an event time constraint term, and a textual semantic guidance constraint term. The model is then trained, and multi-step prediction results are output.

[0010] In the above scheme, the structured parsing of unstructured operation and maintenance text can transform event information in work orders, logs, and alarm records into computable numerical features, making up for the lack of input modalities in traditional pure numerical prediction models. By accurately aligning the event time interval with the prediction window and calculating the overlap ratio, the scheme can characterize the difference in the impact intensity of the same operation and maintenance event under different prediction step sizes, avoiding the limitation that discrete flags cannot reflect temporal changes. The structure of superimposing the basic predicted value with the event correction amount can give the prediction correction process a clear physical meaning and reduce the black box nature of the model. The loss function with multiple constraints can optimize the prediction accuracy while guiding the model to learn the correction rules that conform to the operation and maintenance logic, improving the generalization performance in abnormal scenarios.

[0011] Further, the process of generating the single-event impact coefficient includes: concatenating the text semantic features, the time overlap ratio, the event type encoding, the event level normalized value, the event object matching flag, and the processing status encoding along the feature dimension to obtain an event impact input vector; inputting the event impact input vector into the event impact coefficient generation network to obtain candidate impact strengths; and multiplying the candidate impact strengths by the time overlap ratio and the event object matching flag to obtain the single-event impact coefficient. The event object matching flag is graded according to the degree of correlation between the event object and the wind turbine to be predicted: a value of 1 is taken when there is a complete match or the same wind turbine site; a preset weak matching coefficient is taken when there is a weak correlation in the same region, and the preset weak matching coefficient is greater than 0 and less than 1; and a value of 0 is taken when there is no correlation.

[0012] The multiplicative gating design allows the single-event impact coefficient to be subject to both time validity and object validity constraints. When the event does not overlap with the prediction window in time or is not related to the wind turbine to be predicted, the impact coefficient automatically returns to zero without the need for additional judgment logic. The hierarchical matching method can reflect the difference in the impact of events with different degrees of correlation, ensuring the correction strength of directly related events while also taking into account the weak impact of indirectly related events in the same area, thus adapting to the cluster operation characteristics of offshore wind farms.

[0013] Furthermore, when multiple operation and maintenance events exist within the same prediction time window, an attention input vector is constructed based on the event impact input vector of each operation and maintenance event and the single event impact coefficient. The attention weight of each operation and maintenance event is calculated through an attention network. The attention weight is used to weight and fuse multiple single event impact coefficients and then truncated to a preset interval to obtain the comprehensive event impact coefficient.

[0014] The attention fusion mechanism can adaptively allocate aggregation weights based on the semantic features and impact intensity of events, so that events with more significant impact and higher relevance account for a higher proportion in the comprehensive coefficient. Compared with fixed weights or simple accumulation, it is more in line with the actual impact pattern in scenarios with multiple events superimposed. Truncation to a preset range can ensure that the numerical boundary of the comprehensive event impact coefficient is reasonable and avoid abnormal values ​​that lead to instability in the subsequent modulation and correction process.

[0015] Furthermore, the process of modulating the initial time-series prediction representation includes: generating an event modulation gating and an event residual representation by combining the comprehensive event influence coefficient and the text semantic features; weighting the event residual representation with the event modulation gating and the comprehensive event influence coefficient, and then superimposing it onto the initial time-series prediction representation to obtain the time-series enhancement feature. When the comprehensive event influence coefficient is zero, the time-series enhancement feature is consistent with the initial time-series prediction representation.

[0016] This modulation method controls the injection intensity of event information through a multiplicative switch, enabling adaptive activation and deactivation of the impact of maintenance events: when there are no effective maintenance events, the modulation effect automatically fails, and the model output remains consistent with the conventional time series prediction results; when maintenance events exist, the correction magnitude of the time series representation can be dynamically adjusted according to the intensity of the event's impact, so that the time series features can respond promptly to abnormal changes in the operating status.

[0017] Furthermore, the multimodal fusion adopts an event-condition-driven two-stage gating fusion structure, specifically including: a first stage, generating basic temporal environment fusion features based on runtime sequence coding features and meteorological and sea state coding features; a second stage, combining the basic temporal environment fusion features, the coding features corresponding to the text semantic features, the temporal enhancement features, and the comprehensive event influence coefficient to generate event correction features and event fusion gating weights, and then weighting the event correction features with the event fusion gating weights and superimposing them onto the basic temporal environment fusion features to obtain multimodal fusion features.

[0018] The dual-stage fusion structure can decouple the power variation pattern under normal operating conditions from the disturbance correction brought by operation and maintenance events at the feature level, enabling the model to learn the normal time-series environment mapping and the event correction mapping respectively, which is beneficial to improving the model convergence efficiency and performance stability under different operating conditions; the gated weighting superposition method can dynamically adjust the fusion ratio of correction features according to the event intensity, avoiding interference from redundant features in the absence of events.

[0019] Furthermore, the functions of each component of the event constraint loss function are as follows: the power prediction error loss term is used to constrain the deviation between the final wind power prediction value and the actual power value; the event time constraint term is used to constrain the comprehensive event influence coefficient to be consistent with the influence intensity of the reference event constructed based on the time overlap relationship; the text semantic guidance constraint term is used to constrain the event gating weight to be consistent with the gating intensity of the reference event generated based on the text semantic features, and to constrain the direction of the residual correction amount to be consistent with the power influence direction determined based on the event type.

[0020] The three types of constraints work synergistically: power prediction error loss ensures the basic accuracy of the prediction results and is the core objective of model optimization; event time constraints guide the event impact coefficients generated by the model to match the event's time coverage pattern, avoiding the problem of misaligned impact timing; textual semantic guidance constraints embed prior knowledge from operational texts into the training process, making the model's correction behavior consistent with the semantic connotation of the events, improving the interpretability of intermediate variables. The parameters of the event impact assessment module used to generate reference supervision signals are pre-trained and kept fixed, ensuring the stability of constraint anchor points during training, further improving the model's convergence reliability and constraint interpretability.

[0021] Furthermore, the text semantic guidance constraint includes a directional penalty term: when the power influence direction indicator power decreases and the residual correction amount is positive, or when the power influence direction indicator power recovers and the residual correction amount is negative, the directional penalty term generates a penalty value; when the power influence direction indicator has no significant influence, the directional penalty term does not take effect.

[0022] Directional penalty terms can constrain the rationality of the correction direction at the loss function level, avoiding the model from generating correction results that contradict the physical impact of the event. For example, for maintenance and downtime events, the model should not generate positive residual corrections for power decline trends. This constraint can reduce the probability of the model learning contradictory patterns, especially in small sample maintenance event scenarios, and can improve the logical reliability of the model output.

[0023] Furthermore, the reference event impact strength, reference event gating strength, and power impact direction are all uniformly generated by the pre-trained offline event impact assessment module. The event impact assessment module takes the textual semantic features of the operation and maintenance event, structured event fields, event object matching flags, and the time overlap ratio between the event and the prediction time window as input, and includes three output branches: event impact strength, reference event gating strength, and power impact direction. The module network parameters remain fixed throughout the training of the offshore wind power ultra-short-term intelligent prediction model and are not updated synchronously with the parameters of the main prediction model.

[0024] The event impact assessment module internally determines the event type weight and power impact direction based on the event type, determines the pre-trained fixed level weight parameters based on the event level, and determines the processing state weight based on the processing status. Combining the matching relationship between the event object and the wind turbine to be predicted and the time overlap ratio, the event type weight, event level weight, and processing state weight are multiplicatively combined to generate a single event reference impact intensity. In multi-event scenarios, the module internally pre-trained attention weights are aggregated to obtain a comprehensive reference value for the corresponding prediction step size.

[0025] This design is based on deterministic mapping rules for structured event fields. Combined with pre-trained and adapted weight parameters, it adapts to different operational scenarios, providing semantic anchors for model training that balance prior operational knowledge with scenario adaptability, guiding the model to converge in a direction consistent with actual operational patterns. The multiplicative combination generation method simultaneously reflects the combined impact of event severity, time validity, and object validity, maintaining consistency with the event impact coefficient generation logic. The reference event gating strength, generated through an internal mapping network combining structured event fields and textual semantic features, more precisely characterizes the guidance of event semantics on gating weights, ensuring the correction process aligns with the semantic connotations of operations and maintenance. The module parameter freezing design prevents constraint targets from drifting during training, strengthening the semantic anchoring effect of constraint terms.

[0026] Furthermore, the semantic parsing and structured extraction are performed using a large language model fine-tuned with low-rank adaptation based on samples from the equipment operation and maintenance domain. The backbone parameters of the large language model are frozen, and low-rank adaptation layers are inserted at the query matrix projection layer, key matrix projection layer, value matrix projection layer, and output projection layer of its attention layer. Only the parameters of the low-rank adaptation layers are trained. The structured event fields are output by the large language model constrained by operation and maintenance domain prompt word templates, and the text semantic features are obtained from the last hidden state of the large language model through average pooling, linear mapping, and layer normalization.

[0027] The low-rank adaptation fine-tuning method can complete the semantic adaptation in the field of wind power operation and maintenance with fewer parameters and data while retaining the general semantic understanding ability of large language models, thus reducing the computing power and data costs of engineering implementation; prompt word templates can standardize the output format of large language models and ensure the consistency of structured fields; the three-step extraction process of pooling, mapping and normalization can transform the variable-length text hidden state into a semantic feature vector with uniform dimension and stable distribution, which can adapt to the input requirements of subsequent networks.

[0028] This invention also provides an intelligent offshore wind power prediction system that integrates equipment operation and maintenance data, used to implement the above prediction method. The system includes:

[0029] The data preprocessing unit is used to acquire and preprocess multimodal data of offshore wind turbines, including wind turbine operation sequence, meteorological and sea conditions and equipment operation and maintenance text information.

[0030] The multimodal feature encoding unit is used to encode wind turbine operation sequence data, meteorological and sea condition data, and equipment operation and maintenance text information respectively; the encoding of equipment operation and maintenance text information includes semantic parsing and structured extraction to obtain text semantic features and structured event fields;

[0031] The operation and maintenance event text time sequence alignment and impact coefficient generation unit is used to align the event time interval with each prediction window, calculate the time overlap ratio, and generate the comprehensive event impact coefficient for each step based on the time overlap ratio, the matching relationship between the structured event fields and the event objects and the wind turbines to be predicted.

[0032] The temporal enhancement modeling unit is used to obtain an initial temporal prediction representation by temporal modeling based on the coding features of runtime sequence and meteorological and sea conditions, and to obtain temporal enhancement features by modulating the comprehensive event influence coefficient and text semantic features;

[0033] The feature fusion prediction unit is used to fuse multimodal coding features and temporal enhancement features to generate a base power prediction value, event gating weights and residual corrections. Based on the comprehensive event impact coefficient, event gating weights and residual corrections, an event correction term is constructed and superimposed on the base power prediction value to obtain the final wind power prediction value.

[0034] The model training unit is used to construct an event constraint loss function containing a power prediction error loss term, an event time constraint term, and a text semantic guidance constraint term, and to iteratively train the prediction model.

[0035] Each unit's function corresponds one-to-one with the aforementioned method steps, working together to complete the entire process from multimodal data input to event-aware power prediction, which can adapt to the actual operation and deployment needs of offshore wind farms.

[0036] Compared with the prior art, the present invention and its preferred embodiments have at least the following beneficial effects:

[0037] It expands the input mode boundary of power prediction, effectively mines event information in unstructured operation and maintenance texts, makes up for the information deficiency of traditional pure numerical prediction methods, and can comprehensively reflect the combined effect of operating status, meteorological and sea conditions and operation and maintenance events on power output, adapting to the complex and ever-changing operating scenarios of offshore wind power.

[0038] It achieves the quantification of the differentiated impact of operation and maintenance events on different prediction steps. Through precise time alignment and overlap ratio calculation, it can characterize the change in the degree of impact of the same event in different prediction periods. With the gating mechanism of object association, it automatically blocks the interference of irrelevant events and ensures the time and object validity of the event impact.

[0039] A hierarchical and decoupled interpretable prediction architecture was constructed, which decomposes the final prediction result into basic operating trend and event correction components. The correction process has a clear physical meaning, which makes it easier for operators to understand the impact of events on power and reduces the black box nature of model application.

[0040] By using a multi-dimensional event-constrained training mechanism to provide semantic anchors for the model, the model is guided to learn correction rules that conform to the physical logic of operation and maintenance. This effectively improves the model's generalization ability and prediction reliability in scenarios such as rare faults and maintenance, and reduces the risk of prediction bias in abnormal scenarios. Attached Figure Description

[0041] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0042] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention.

[0043] Figure 2 This is a schematic diagram of the structure of the ultra-short-term intelligent prediction model for offshore wind power provided in an embodiment of the present invention.

[0044] Figure 3 This is a graph showing the prediction results at the 15-minute level in an embodiment of the present invention.

[0045] Figure 4This is a graph showing the prediction results at the 30-minute level in an embodiment of the present invention. Detailed Implementation

[0046] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0047] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0048] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0049] This invention provides an intelligent prediction method and system for offshore wind power that integrates equipment operation and maintenance data, in order to solve the problems that existing offshore wind power prediction methods are unable to effectively utilize equipment operation and maintenance text information, are unable to characterize the temporal impact of equipment operation and maintenance events on different prediction step lengths, and are unable to dynamically correct events such as maintenance, shutdown, fault, power limitation and recovery.

[0050] To achieve the above objectives, this invention provides a method for ultra-short-term intelligent prediction of offshore wind power by integrating multi-modal data from equipment operation and maintenance, comprising the following steps:

[0051] S1. Acquire multimodal data on the operation and maintenance of offshore wind turbine equipment, and preprocess the data. The multimodal data includes wind turbine operation sequence data, meteorological and sea state data, and equipment operation and maintenance text information. Specifically, the wind turbine operation sequence data includes one or more of the following: wind power, wind speed, speed, pitch angle, yaw angle, generator temperature, operating status indicators, and fault status indicators; the meteorological and sea state data includes one or more of the following: sea wind speed, wind direction, temperature, air pressure, and sea state forecast data; and the equipment operation and maintenance text information includes one or more of the following: planned maintenance work orders, temporary maintenance records, fault alarm records, shutdown records, power limitation records, site power limitation records, equipment inspection records, component replacement records, and reset and recovery records.

[0052] S2. Construct training, validation, and test sets from the preprocessed equipment operation and maintenance multimodal data.

[0053] S3. Construct an ultra-short-term intelligent prediction model for offshore wind power that integrates multi-modal data from equipment operation and maintenance, and input the training set into the model for training. The specific implementation process of the ultra-short-term intelligent prediction model for offshore wind power is as follows:

[0054] S3.1 Construct a multimodal feature encoding layer to encode wind turbine operation sequence data, meteorological and sea state data, and equipment operation and maintenance text information respectively, obtaining corresponding multimodal encoded features. The multimodal feature encoding layer includes a wind turbine operation sequence data encoding sublayer, a meteorological and sea state data encoding sublayer, and an equipment operation and maintenance text data encoding sublayer. Specifically, the wind turbine operation sequence data encoding sublayer is used to encode wind turbine operation sequence data to extract the characteristics of wind turbine unit operating status changes; the meteorological and sea state data encoding sublayer is used to encode meteorological and sea state data to extract the characteristics of changes in the marine external environment; and the equipment operation and maintenance text data encoding sublayer is used to encode equipment operation and maintenance text information based on a large language model to extract the semantic features of the equipment operation and maintenance text.

[0055] As a preferred implementation, the wind turbine runtime sequence data encoding sublayer uses a pre-trained time series baseline model as its backbone network and employs an efficient parameter fine-tuning method for adaptation training. Specifically, the backbone parameters of the pre-trained time series baseline model are frozen, and a low-rank adaptation layer (LoRA) is inserted at the query matrix projection layer, key matrix projection layer, value matrix projection layer, and output projection layer in its encoding and / or decoding layers. Only the low-rank matrix parameters in the LoRA layer are trained to achieve model adaptation for offshore wind turbine runtime sequence data scenarios. The pre-trained time series baseline model is a Transformer-type model pre-trained on large-scale general time series data. Through LoRA adaptation, it acquires the ability to model wind turbine runtime sequence data while maintaining the general time series feature representation capabilities learned by the pre-trained model. The meteorological and sea state data encoding sublayer can adopt the same backbone network structure and adaptation method as the wind turbine runtime sequence data encoding sublayer, or use conventional encoding structures such as multilayer perceptrons and convolutional neural networks. The equipment operation and maintenance text data encoding sub-layer adopts a large language model fine-tuned by the LoRA low-rank adaptation layer of equipment operation and maintenance domain samples, and constrains the large language model with equipment operation and maintenance domain prompt word templates to output structured event fields in JSON format, while also outputting text semantic feature vectors.

[0056] S3.2 Construct a time-series alignment and impact coefficient generation layer for operation and maintenance event texts. Generate modal feature vectors for operation and maintenance event texts based on equipment operation and maintenance text information, and generate event impact coefficients corresponding to each prediction step size according to the time overlap relationship between operation and maintenance events and each prediction time window.

[0057] The O&M event text timing alignment and impact coefficient generation layer is used to perform semantic parsing and event structure extraction on equipment O&M text information, generating O&M event text modal feature vectors and O&M event prior information. The O&M event prior information includes one or more of the following: event type, event trigger time, event end time, planned end time, actual recovery time, event duration, wind turbine number, event object, affected components, processing status, event level, and power impact direction.

[0058] The O&M event text timing alignment and impact coefficient generation layer aligns the event trigger time and event end time in the prior information of the O&M event with the prediction time window corresponding to each prediction step size, and calculates the time overlap ratio of the O&M event within each prediction time window. The larger the time overlap ratio of the O&M event within the prediction time window, the larger the corresponding event impact coefficient; the smaller the time overlap ratio, the smaller the corresponding event impact coefficient; when there is no time overlap between the O&M event and a certain prediction time window, the corresponding event impact coefficient approaches zero.

[0059] Furthermore, for each operation and maintenance event, the text semantic feature vector obtained from the language large model encoding, the time overlap ratio, the event type encoding vector, the event level normalized value, the event object matching flag, and the processing status encoding vector are concatenated along the feature dimension to obtain the event impact input vector. This event impact input vector is then fed into the event impact coefficient generation network to generate candidate impact intensities. The candidate impact intensities are then multiplied by the time overlap ratio and the event object matching flag to obtain the single event impact coefficient. This multiplicative gating design ensures that the single event impact coefficient automatically resets to zero when the operation and maintenance event does not overlap with the prediction time window or is not associated with any object of the wind turbine currently being predicted.

[0060] When multiple operation and maintenance events exist within the same prediction time window, attention weights are calculated based on the event impact input vector and single event impact coefficient corresponding to each operation and maintenance event. Attention weighted fusion is then performed on multiple single event impact coefficients to obtain the comprehensive event impact coefficient corresponding to the prediction step size.

[0061] S3.3. Based on the large-scale time series model, time series coding features and meteorological and sea state coding features are used to perform time series modeling to obtain an initial time series prediction representation. Then, using the textual modal feature vector of maintenance events and the event influence coefficients corresponding to each prediction step size, the initial time series prediction representation is subjected to gating modulation and residual correction to obtain a time series enhancement feature that integrates the influence of equipment maintenance events. When the comprehensive event influence coefficient is zero, the time series enhancement feature degenerates into the initial time series prediction representation, and the modulation effect of maintenance events on the time series modeling results is automatically turned off; when the comprehensive event influence coefficient increases, the modulation effect of maintenance events on the time series modeling results is correspondingly enhanced.

[0062] S3.4 Construct a feature fusion layer to map multimodal coding features and temporal enhancement features to a unified feature space, generate multimodal fusion features corresponding to each prediction step, and obtain the final wind power prediction value corresponding to each prediction step based on the multimodal fusion features.

[0063] The feature fusion layer adopts an event-driven, two-stage gated fusion structure. The first stage generates basic time-series environmental fusion features based on runtime sequence coding features and meteorological and sea state coding features. The second stage generates event-corrected features and event fusion gating weights based on the basic time-series environmental fusion features, operation and maintenance text coding features, time-series enhancement features, and event impact coefficients corresponding to each prediction step. The event-corrected features are then weighted according to the event fusion gating weights and superimposed onto the basic time-series environmental fusion features to obtain the multimodal fusion features corresponding to each prediction step.

[0064] A base power prediction value is generated based on the fusion features of the basic time-series environment, and event gating weights and residual corrections are generated based on the fusion features of the multimodal environment. The final wind power prediction value corresponding to the h-th prediction step is obtained as follows: the final wind power prediction value equals the base power prediction value plus the product of the comprehensive event impact coefficient, the event gating weights, and the residual corrections. When the comprehensive event impact coefficient is zero, the final prediction value degenerates into the base power prediction value; when there are effective operation and maintenance events and their impact is strong, the event gating weights and residual corrections work together to perform event-driven corrections on the base power prediction value.

[0065] S3.5. Construct an event-constrained loss function associated with equipment operation and maintenance events, and train the ultra-short-term intelligent prediction model for offshore wind power using the training set. The event-constrained loss function includes a power prediction error loss term, an event time constraint term, and a text semantic guidance constraint term. The power prediction error loss term is used to constrain the deviation between the final wind power prediction value and the actual wind power value; the event time constraint term is used to constrain the comprehensive event impact coefficient generated by the model to be consistent with the reference event impact strength constructed from the event time overlap relationship; the text semantic guidance constraint term is used to constrain the event gating weights generated by the model to be consistent with the reference event gating strength generated by the equipment operation and maintenance text semantic features, and to constrain the direction of the residual correction amount to be consistent with the power impact direction determined by the event type.

[0066] The reference values ​​in the event time constraint and text semantic guidance constraint are generated by a pre-trained event impact assessment module. Based on deterministic rules for structured event fields, these values ​​are calculated using pre-trained fixed level weight parameters and a semantic mapping network. Specifically, event type weights and power influence directions are determined based on event type, event level weights based on event level, processing state weights based on processing state, object validity based on event object matching flags, and time validity based on time overlap ratio. These factors are combined according to deterministic rules to generate reference event impact strength, reference event gating strength, and power influence direction, which serve as the supervision targets in the loss function.

[0067] S4. Use the trained offshore wind power ultra-short-term intelligent prediction model to perform multi-step predictions and obtain the wind power prediction values ​​corresponding to each prediction step.

[0068] This invention also provides an ultra-short-term intelligent prediction system for offshore wind power that integrates multi-modal data from equipment operation and maintenance, comprising:

[0069] The data preprocessing unit is used to acquire multimodal data on the operation and maintenance of offshore wind turbine equipment and to preprocess the multimodal data on equipment operation and maintenance.

[0070] The sample construction unit is used to divide the preprocessed equipment operation and maintenance multimodal data into training set, validation set and test set;

[0071] The multimodal feature encoding unit is used to encode wind turbine operation sequence data, meteorological and sea condition data and equipment operation and maintenance text information respectively to obtain the corresponding multimodal encoding features;

[0072] The operation and maintenance event text time sequence alignment and impact coefficient generation unit is used to generate operation and maintenance event text modal feature vectors based on equipment operation and maintenance text information, and generate event impact coefficients corresponding to each prediction step according to the time overlap relationship between the operation and maintenance event and each prediction time window.

[0073] The time-series enhancement modeling unit is used to perform time-series modeling on runtime coding features and meteorological and sea state coding features based on a large time-series model, and modulates the time-series modeling results using the text modal feature vector of operation and maintenance events and the event impact coefficient to obtain time-series enhanced features that fuse the impact of equipment operation and maintenance events;

[0074] The feature fusion prediction unit is used to map multimodal coding features and temporal enhancement features to a unified feature space, generate multimodal fusion features corresponding to each prediction step, and obtain the final wind power prediction value corresponding to each prediction step based on the multimodal fusion features.

[0075] The model training unit is used to construct the event constraint loss function associated with equipment operation and maintenance events, and to train the offshore wind power ultra-short-term intelligent prediction model using the training set to obtain the trained prediction model.

[0076] The model prediction unit is used to input the multimodal data of equipment operation and maintenance to be predicted into the trained offshore wind power ultra-short-term intelligent prediction model to obtain the wind power prediction value corresponding to each prediction step in the future.

[0077] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0078] like Figure 1 As shown in the figure, the offshore wind power ultra-short-term intelligent prediction method for fusion equipment operation and maintenance multimodal data provided in this embodiment of the invention includes the following steps: data acquisition and preprocessing, sample construction, model construction and training, and multi-step prediction result output. The model construction and training step is further divided into sub-steps such as multimodal feature encoding, operation and maintenance event time sequence alignment and influence coefficient generation, time sequence modeling and event influence modulation, two-stage gating fusion prediction, and event constraint loss function training, which constitute a complete end-to-end model training and prediction process.

[0079] like Figure 2 As shown in the figure, the ultra-short-term intelligent prediction model for offshore wind power provided in this embodiment of the invention adopts a layered modular architecture. From data input to result output, it includes a data acquisition layer, a multimodal feature encoding layer, a time-series alignment and impact coefficient generation layer for operation and maintenance event texts, a time-series modeling module, a feature fusion layer, and a prediction output layer. The data acquisition layer is used to access three types of heterogeneous source data: wind turbine operation sequence, meteorological and sea conditions, and equipment operation and maintenance texts. The multimodal feature encoding layer extracts and encodes features from each type of data to obtain modal features of a unified dimension. The time-series alignment and impact coefficient generation layer is responsible for extracting structured event information from unstructured texts and generating quantifiable event impact coefficients. The time-series modeling module performs deep modeling of time-series features and fuses event impacts to obtain enhanced time-series features. The feature fusion layer uses an event-condition-driven two-stage gating structure to achieve dynamic fusion of multimodal features. Finally, the prediction output layer generates wind power prediction results for each future step.

[0080] I. Structured Extraction Process of Equipment Operation and Maintenance Text

[0081] First, the equipment maintenance text information is input into a large language model constrained by prompt words in the equipment maintenance domain for semantic parsing and structured extraction. The equipment maintenance text information includes planned maintenance work orders, temporary maintenance records, fault alarm records, downtime records, power limit records, site power limit records, equipment inspection records, component replacement records, reset and recovery records, and operation logs.

[0082] The large language model employs a pre-trained large language model and undergoes LoRA fine-tuning based on samples from the equipment operation and maintenance domain. During LoRA fine-tuning, the core parameters of the large language model are frozen, and low-rank adaptation layers are inserted at the query matrix projection layer, key matrix projection layer, value matrix projection layer, and output projection layer in the attention layer. The rank, scaling factor, and random deactivation ratio of the low-rank adaptation layers are used as preset hyperparameters, which can be adjusted according to the training data scale and fitting effect. Domain adaptation of the large language model is achieved through LoRA fine-tuning to improve its ability to recognize the semantics of events such as faults, outages, power limitations, and reset / recovery in wind power operation and maintenance texts.

[0083] The prompt word template is used to constrain the large language model to output structured event information according to preset fields. Preset fields include event type, event trigger time, event end time, planned end time, actual recovery time, event duration, turbine number, event object, affected components, processing status, event level, and power impact direction. The structured results output by the large language model are in JSON format.

[0084] Event Type indicates whether the maintenance event belongs to planned maintenance, temporary maintenance, fault alarm, shutdown, power limitation, site power restriction, equipment inspection, component replacement, reset and recovery, reconnection to the grid, fault resolution, or no obvious abnormality. Event Trigger Time indicates the time when the maintenance event begins to affect the operation of the wind turbine or wind farm. Event End Time indicates the time when the maintenance event ends and affects the operation of the wind turbine or wind farm. Planned End Time indicates the estimated end time given in the work order or maintenance plan. Actual Recovery Time indicates the time when the equipment resumes operation, reconnects to the grid, or the fault is resolved. Event Duration indicates the length of time from triggering to ending the event. Turbine Number indicates the specific wind turbine number corresponding to the event. Event Object indicates whether the event affects a single turbine, multiple turbines, or the entire wind farm. Affected Components indicate the equipment components affected by the event. Processing Status indicates whether the event is in an unprocessed, processing, completed, recovered, closed, or confirmed state. Event Level indicates the severity of the event. The direction of power impact is used to indicate whether the event affects wind power in the direction of reduction, recovery, or no significant impact.

[0085] When the equipment maintenance documentation explicitly specifies the event trigger time and event end time, these extracted times are used directly. When the equipment maintenance documentation does not explicitly specify the event end time but provides a planned end time or actual recovery time, the actual recovery time is used first. If the actual recovery time is missing, the planned end time is used. If both the planned and actual recovery times are missing, the event end time is determined based on the event type and preset duration rules. For example, routine inspection events can be set to a shorter duration, planned maintenance events can be determined based on the work order's planned duration, and fault downtime events can continue until the fault is resolved or a reset recovery event occurs.

[0086] For wind turbine-level power curtailment, maintenance, or dispatch events, if the text does not specify a single wind turbine number, the event object will be set as a wind turbine-level object, and its scope will be set to the relevant wind turbine units within that wind turbine site. For single wind turbine failure, shutdown, or component replacement events, if the text includes a wind turbine number, the event object will be set as the corresponding wind turbine unit.

[0087] II. The Numericalization Process of Text Semantic Features and Structured Fields

[0088] For the m-th operation and maintenance event, the large language model outputs both the structured event field and a textual semantic feature vector representing the overall semantics of the event. Let the textual semantic feature vector of the m-th operation and maintenance event be denoted as... Text semantic feature vector It is obtained from the hidden state of the last layer of the large language model through average pooling, linear mapping and layer normalization.

[0089] First, average pooling is performed on the last hidden state of the large language model to obtain the original text semantic vector after pooling:

[0090]

[0091] in, This represents the hidden state of the large language model corresponding to the i-th text tag in the m-th operation and maintenance event text. This indicates the number of text tags corresponding to the operation and maintenance event text. This represents the original text semantic vector after pooling.

[0092] Subsequently, the original text semantic vectors are subjected to linear mapping and layer normalization:

[0093]

[0094] in, This represents the mapped text semantic feature vector. and These represent the trainable weight matrix and bias of the text semantic feature mapping layer, respectively. Presentation layer normalization operation.

[0095] Numerical processing is performed on the structured fields output by the large language model, specifically including the following steps.

[0096] Event type encoding. Event types are mapped to event type numbers, and the event type encoding vector is obtained through the event type embedding table.

[0097]

[0098] in, This represents the event type encoding vector for the m-th operation and maintenance event. The event type is embedded in the table. This represents the event type number of the m-th maintenance event. Event types are numbered as follows: planned maintenance, temporary maintenance, fault alarm, shutdown, power limitation, site power limitation, equipment inspection, component replacement, reset and recovery, reconnection to the grid, fault resolution, and no obvious abnormality correspond to different event type numbers. The above numbering is only used to convert discrete event types into a computable vector representation and does not limit the specific number of event types.

[0099] Event level coding. Converting event levels to numerical levels and then normalizing them:

[0100]

[0101] in, This represents the normalized event level of the m-th operation and maintenance event. This represents the original level of the m-th operational event. This indicates the total number of event levels. If the event level is given in text form such as severe, important, general, or warning, it is first mapped to a numerical level according to a preset level table, and then normalized.

[0102] Event object matching flag. The event object matching flag indicates whether the m-th maintenance event applies to the currently predicted wind turbine or its associated site. Its value is determined as follows: when the maintenance event object matches the current predicted wind turbine number, it is set to [value missing]. When the operation and maintenance event is a site-level power curtailment, site-level maintenance, site-level dispatch, or site-level fault event, and the wind turbine to be predicted belongs to that site, take... When the maintenance event object and the wind turbine to be predicted belong to the same collection line, the same transformer substation, the same subsystem, or the same area, take... ,in For preset weak matching coefficients and When the operation and maintenance event object is not related to the wind turbine unit to be predicted, take... .

[0103] As a preferred implementation, a weak matching coefficient is preset. The value can be set according to the topological coupling degree or wake effect intensity of the wind farm's collector lines. In this embodiment, it is preferably set to a value between 0.3 and 0.7, for example, 0.5.

[0104] Processing state encoding. Processing states are mapped to processing state numbers, and the processing state encoding vector is obtained through a processing state embedding table.

[0105]

[0106] in, This represents the processing status encoding vector for the m-th operation and maintenance event. The table represents the embedded processing status. This represents the processing status number of the m-th operation and maintenance event. The processing status can include unprocessed, processing, processed, recovered, and closed.

[0107] Time field standardization. The event trigger time, event end time, prediction time window start time, and prediction time window end time are all uniformly converted to timestamps under the same time base. The time unit can be uniformly set to seconds, minutes, or hours. The event trigger time of the m-th maintenance event is denoted as... The end time of the event is recorded as The start time of the h-th prediction time window is denoted as . The end time is recorded as .

[0108] III. Calculation Process of Time Overlap Ratio

[0109] For the m-th maintenance event and the h-th prediction time window, first calculate the overlap start time and overlap end time of the two:

[0110]

[0111] in, This indicates the start time of the overlap between the m-th maintenance event and the h-th prediction time window. This represents the event trigger time of the m-th operation and maintenance event. This indicates the start time of the h-th prediction time window.

[0112]

[0113] in, This represents the end time of the overlap between the m-th maintenance event and the h-th prediction time window. This represents the end time of the m-th maintenance event. This represents the end time of the h-th prediction time window.

[0114] The overlap between maintenance events and the predicted time window is:

[0115]

[0116] in, This represents the overlap duration between the m-th maintenance event and the h-th prediction time window.

[0117] The window length of the h-th prediction time window is:

[0118]

[0119] in, This represents the window length of the h-th prediction time window. As a preferred implementation, when the window length... When it is zero, let the time overlap ratio be zero. This is to avoid division by zero exceptions.

[0120] Therefore, the time overlap ratio between the m-th maintenance event and the h-th prediction time window is:

[0121]

[0122] in, This represents the time overlap ratio between the m-th maintenance event and the h-th prediction time window, with a value ranging from 0 to 1. and These represent the event trigger time and event end time of the m-th operation and maintenance event, respectively. and These represent the start and end times of the h-th prediction time window, respectively. When the operational events completely cover the h-th prediction time window, When the maintenance event only covers a portion of the h-th prediction time window, When the operation and maintenance event does not overlap with the h-th prediction time window, .

[0123] IV. The impact of events on the input vector concatenation process

[0124] After extracting semantic features from the text and quantifying the structured fields, the features of the m-th operational event at the h-th prediction step are concatenated along the feature dimension to obtain the event impact input vector. This concatenation operation does not concatenate the original text string, but rather performs a feature-dimensional connection on the already quantified or vectorized features:

[0125]

[0126] in, This represents the impact of the m-th operational event on the h-th prediction step in the input vector. This indicates a vector concatenation operation along the feature dimension. The textual semantic feature vector representing the m-th operation and maintenance event. Indicates the proportion of time overlap. Represents an event type encoding vector. This represents the normalized event level. Indicates the event object matching flag, This represents the processing state encoding vector. If the text semantic feature vector... The dimension is Event type encoding vector The dimension is Processing state encoding vectors The dimension is The event then affects the input vector. The dimension is The additional three scalars correspond to the time overlap ratios, respectively. Event Level Matching flags for event objects Before being input into the neural network, the aforementioned scalars are expanded into one-dimensional features and concatenated with other vector features in the same floating-point format.

[0127] V. The process of generating the single-event influence coefficient

[0128] The effect of events on the input vector Input event impact coefficient generation network. The event impact coefficient generation network uses a two-layer fully connected network to generate candidate impact strengths:

[0129]

[0130] in, This represents the hidden layer features of the event impact coefficient generation network. and This represents the trainable parameters of the first layer of the event impact coefficient generation network. This represents the linear rectification activation function.

[0131]

[0132] in, This represents the candidate impact strength generated by the event's semantic features, event type, event level, event object, and processing status, with a value ranging from 0 to 1. and This represents the trainable parameters of the second layer of the event impact coefficient generation network. This represents the Sigmoid function.

[0133] To ensure that the event impact coefficient satisfies both time validity and object validity, the candidate impact strength is multiplied by the time overlap ratio and the event object matching flag to obtain the single-event impact coefficient of the m-th maintenance event at the h-th prediction step:

[0134]

[0135] in, This represents the single-event impact coefficient. Because... explicit inclusion and When the m-th maintenance event and the h-th prediction time window do not overlap in time, ,thereby When the m-th maintenance event is not related to the current wind turbine unit to be predicted, ,thereby When the operational event and the prediction time window have a high degree of temporal overlap and are highly correlated with the current prediction object, the single-event impact coefficient increases with the increase of the candidate impact strength. Therefore, the single-event impact coefficient... It also reflects the textual semantic strength, time scope, event type, event level, event object association, and processing status of the operation and maintenance event.

[0136] VI. Multi-Event Attention Weighted Fusion Process

[0137] When multiple operation and maintenance events exist within the h-th prediction time window, attention-weighted fusion of the impact coefficients of multiple single events is performed to obtain the comprehensive event impact coefficient corresponding to the prediction step size.

[0138] First, construct the attention input vector. For the m-th operational event, the event's influence input vector is... Compared with the single event impact coefficient Concatenate them again to obtain the attention input vector:

[0139]

[0140] in, This represents the attention input vector for the m-th maintenance event at the h-th prediction step.

[0141] Then, calculate the attention score for the m-th operational event at the h-th prediction step:

[0142]

[0143] in, Indicates attention score, , and Represents the trainable parameters of the attention network. This represents the hyperbolic tangent function.

[0144] Normalize multiple operational events within the same prediction time window to obtain attention weights:

[0145]

[0146] in, This represents the attention weight of the m-th operational event at the h-th prediction step. This represents the number of valid operational events included in the calculation within the h-th prediction time window. Valid operational events include those that overlap with the h-th prediction time window and are directly or indirectly related to the current prediction object.

[0147] The comprehensive event impact coefficient corresponding to the h-th prediction step is calculated as follows:

[0148]

[0149] in, This represents the comprehensive event impact coefficient corresponding to the h-th prediction step. This represents the truncation function, used to limit the comprehensive event impact coefficient to between 0 and 1. When there are no valid maintenance events within the h-th prediction time window, let... The value is 0. `clip` is a numeric truncation function used to restrict input values ​​to a specified range, taking the corresponding boundary value when the value exceeds the range boundary.

[0150] VII. The process for determining the intensity of the impact of a reference event

[0151] Reference event impact strength in event time constraints The event impact assessment module, which is pre-trained, generates data based on the time overlap ratio of operation and maintenance events, the importance of event types, event levels, event object matching relationships, and processing status.

[0152] For the m-th maintenance event, first determine the event type weight based on the event type. The type weights are used to characterize the basic impact of different types of operation and maintenance events on wind power output. For example, events that directly affect power output, such as faults, outages, power curtailment, and power curtailment at the wind farm, correspond to higher event type weights; events such as planned maintenance, component replacement, and temporary maintenance correspond to medium to high event type weights; and events such as routine inspections and status confirmations correspond to lower event type weights.

[0153] Determine the event level weight based on the event level:

[0154]

[0155] in, This represents the event level weight of the m-th operation and maintenance event. This represents the event level weight coefficient, a pre-trained fixed parameter of the event impact assessment module. It is obtained through offline training and fitting based on historical operation and maintenance data and power data, and does not participate in parameter updates during the training phase of the main prediction model. This represents the normalized event level of the m-th operation and maintenance event.

[0156] As a preferred option, the event type weight and the processing status weight are based on the prior rules of the operation and maintenance field, or they can be obtained by offline training and fitting in combination with historical power data. Both are pre-trained fixed parameters of the event impact assessment module.

[0157] Determine the processing status weight based on the processing status. For example, unprocessed or processing states correspond to higher weights, while recovered or closed states correspond to lower weights.

[0158] The importance weight of the m-th maintenance event can be represented as:

[0159]

[0160] in, This represents the importance weight of the m-th maintenance event. Indicates the event type weight. Indicates the event level weight. Indicates the processing state weights. Indicates the event object matching flag.

[0161] The influence strength of the reference event corresponding to the h-th prediction step is calculated as follows:

[0162]

[0163] in, This represents the influence intensity of the reference event, constructed from the proportion of time overlap between the equipment maintenance event and the h-th prediction time window. This represents the attention weight of the m-th operational event within the h-th prediction time window. This represents the importance weight of the m-th maintenance event. This represents the time overlap ratio between the m-th maintenance event and the h-th prediction time window. When there are no valid maintenance events within the h-th prediction time window, let... .

[0164] It should be noted that the event type weights, event level calibration coefficients, and processing status weights can be predetermined based on historical samples before training the main prediction model. During the training phase of the offshore wind power ultra-short-term intelligent prediction model, these weights remain unchanged as internal parameters of the event impact assessment process and are not updated synchronously with the parameters of the time-series coding network, feature fusion network, event gating network, and residual correction network in the main prediction model. Therefore, the reference target in the event time constraint term remains stable during training, thereby improving the computability and training stability of the event constraint function.

[0165] VIII. The process of generating reference event gating strength

[0166] The reference event gating strength in the text semantic guidance constraint is used to constrain the event gating weights generated by the model, ensuring consistency with the event impact expressed in the equipment operation and maintenance text. It is generated by the gating strength output branch of the event impact assessment module.

[0167] For the m-th maintenance event, the text semantic feature vector, event type encoding vector, event level, event object matching flag, and processing status encoding vector are concatenated along the feature dimension to obtain the reference gating strength input vector, and candidate reference gating strengths are generated through a two-layer fully connected network:

[0168]

[0169]

[0170] in, This represents the hidden layer features of the reference-gated strength generation network. and This represents the parameters of the first layer of the reference gating strength generation network. These are pre-trained fixed parameters of the event impact assessment module and are not updated during the training phase of the main prediction model. This indicates a vector concatenation operation.

[0171]

[0172] in, Indicates the candidate reference gating strength, and This indicates that the pre-trained fixed parameters of the event impact assessment module in the second layer of the reference gating strength generation network are not updated during the training phase of the main prediction model.

[0173] To ensure that the reference gating strength simultaneously reflects the validity of the event object, the validity of the time, and the severity of the event, the candidate reference gating strength is multiplied by the event object matching flag, the time overlap ratio, and the event level to obtain the single-event reference gating strength:

[0174]

[0175] in, This represents the single-event reference gating strength corresponding to the m-th maintenance event. Indicates the event object matching flag, Indicates the proportion of time overlap. This represents the normalized event level. When multiple valid operational events exist within the h-th prediction time window, they are weighted and aggregated according to the pre-trained attention weights within the event impact assessment module to obtain the reference event gating strength corresponding to the h-th prediction step size.

[0176]

[0177] in, This represents the gating strength of the reference event corresponding to the h-th prediction step. Let represent the attention weight of the m-th operational event at the h-th prediction step. When there are no valid operational events within the h-th prediction time window, let .

[0178] The fully connected mapping network described in this section is an internal network of the event impact assessment module, and its parameters are frozen after the module as a whole completes pre-training; the attention weights of multi-event aggregation are also pre-trained parameters within the module and are not updated with the main model.

[0179] IX. The process of determining the direction of power influence

[0180] The power impact direction is used to represent the trend of the impact of maintenance events on wind power. For the m-th maintenance event, the power impact direction... Determined based on event type and processing status. Direction of power impact. It is generated by the power impact direction output branch of the event impact assessment module, and the judgment rule is based on the deterministic mapping between event type and processing status.

[0181] When the event type is maintenance, shutdown, fault, power limitation, load reduction, or station power restriction, set... This indicates that the maintenance event has a reducing effect on wind power output. When the event type is reset / recovery, reconnection to grid, fault clearing, or power limitation removal, the following settings are configured: This indicates that the maintenance event has a restorative effect on wind power output. When the event type is routine inspection, status confirmation, no obvious abnormalities, or no direct impact on power output, the following settings are configured: This indicates that the maintenance event had no significant direct impact on wind power output.

[0182] When multiple operational events exist within the same prediction time window, the power impact direction corresponding to the h-th prediction step is obtained by weighting the events according to the pre-trained attention weights within the event impact assessment module and then taking the sign of the weights.

[0183]

[0184] in, This indicates the direction of power influence corresponding to the h-th prediction step, with a value of [value missing]. , Represents a symbolic function. Indicates attention weights, This indicates the direction of the power impact of the m-th maintenance event. When the weighted result is greater than zero, When the weighted result is less than zero, When the weighted result is zero or there are no valid maintenance events within the current prediction time window, In this way, the direction of power impact is determined by the event type, processing status, and power impact direction in the equipment operation and maintenance text, which can convert the power reduction, power recovery, or no significant impact in the text semantics into numerical signals that can be used for loss function calculation.

[0185] 10. The Application of Event Influence Coefficient in Temporal Enhancement Features

[0186] After obtaining the comprehensive event impact coefficient corresponding to the h-th prediction step size Then, it is used to modulate the initial time series prediction representation of the output of the large-scale time series model. Let the initial time series prediction representation obtained by the large-scale time series model based on runtime time series coding features and meteorological / sea state coding features be: The modal feature vector of the operation and maintenance event text is .in, The results are obtained by weighting the textual semantic features of each operation and maintenance event within the h-th prediction time window according to attention weights:

[0187]

[0188] in, This represents the text modal feature vector of the operation and maintenance events within the h-th prediction time window. Indicates attention weights. Let represent the textual semantic feature vector of the m-th maintenance event. When there are no valid maintenance events within the h-th prediction time window... Take the zero vector.

[0189] Based on the comprehensive event impact coefficient And operation and maintenance event text modal feature vector generation event modulation gating:

[0190]

[0191] in, Indicates event modulation gating, and Indicates trainable parameters, This indicates a vector concatenation operation.

[0192] Generate event residual representations based on the feature vectors of operation and maintenance event texts:

[0193]

[0194] in, Represents the residual characterization of the event. This represents the mapping function of a multilayer perceptron, which is also known as a nonlinear activation function. and This represents the trainable parameters.

[0195] The event residual representation is superimposed onto the initial time series prediction representation according to the comprehensive event impact coefficient and event modulation gating to obtain the time series enhancement features:

[0196]

[0197] in, This indicates the time-series enhancement characteristics of the impact of maintenance events on converged equipment. This represents element-wise multiplication. Therefore, when... At that time, the modulating effect of operational events on the time series modeling results is turned off. equal ;when As the size increases, the modulating effect of operation and maintenance events on the time series modeling results is correspondingly enhanced.

[0198] XI. Event-Driven Two-Stage Gated Fusion Process

[0199] In the feature fusion stage, runtime sequence coding features, meteorological and sea state coding features, operation and maintenance text coding features, and time-series enhancement features are first mapped to a unified feature space. For the h-th prediction step, the unified feature space mapping process is as follows:

[0200]

[0201]

[0202]

[0203]

[0204] in, , , , These represent the runtime sequence coding input, meteorological and sea state coding input, operation and maintenance text coding input, and time series enhancement input corresponding to the h-th prediction step size, respectively. , , , These represent runtime sequence features, meteorological and sea condition features, operation and maintenance text features, and time-series enhanced features, respectively, mapped to a unified feature space. , , , and , , , These represent the trainable weight matrix and bias of each modality mapping layer, respectively.

[0205] In the first stage, basic time-series environmental fusion features are generated based on runtime sequence features and meteorological and sea state features. The runtime sequence features and meteorological and sea state features are then concatenated and input into the basic fusion gating network to obtain the fusion weights.

[0206]

[0207] in, This represents the fusion weight of runtime sequence features at the h-th prediction step. The fusion weights of meteorological and sea state characteristics at the h-th prediction step are represented by the following: , and This represents the trainable parameters of the basic fusion gating network. This represents the Softmax normalization function.

[0208] Based on the fusion weights, the runtime sequence features and meteorological and sea state features are weighted and fused to obtain the basic time-series environmental fusion features:

[0209]

[0210] in, This represents the basic temporal environment fusion feature corresponding to the h-th prediction step.

[0211] The second stage involves generating event correction features based on basic time-series environment fusion features, operation and maintenance text features, time-series enhancement features, and comprehensive event impact coefficients.

[0212]

[0213] in, This represents the event correction feature corresponding to the h-th prediction step. Represents a non-linear activation function. and This represents the trainable parameters of the event-corrected feature network. This indicates a vector concatenation operation.

[0214] Event fusion gating weights are generated based on basic time-series environment fusion features, operation and maintenance text features, time-series enhancement features, and comprehensive event impact coefficients.

[0215]

[0216] in, This represents the event fusion gating weight corresponding to the h-th prediction step, with a value ranging from 0 to 1. and This represents the trainable parameters of the event fusion gating network.

[0217] The event-modified features are weighted using event fusion gating weights and then superimposed onto the basic temporal environment fusion features to obtain the multimodal fusion features corresponding to the h-th prediction step size:

[0218]

[0219] in, This represents the multimodal fusion feature corresponding to the h-th prediction step. This indicates element-wise multiplication.

[0220] 12. The process of generating the final power prediction value

[0221] Generate basic power prediction values ​​based on the fusion features of the underlying temporal environment:

[0222]

[0223] in, This represents the base power prediction value corresponding to the h-th prediction step. and This represents the trainable parameters of the base power prediction layer.

[0224] Event gating weights are generated based on multimodal fusion features:

[0225]

[0226] in, This represents the event gating weight corresponding to the h-th prediction step generated by the model, with a value ranging from 0 to 1. and This represents the trainable parameters of the event gating network. This represents the event fusion gating weights in the feature fusion stage, used to control the injection strength of event-corrected features in multimodal feature fusion. This represents the event gating weight in the prediction output stage, used to control the correction strength of the residual correction to the final power prediction value. The two have different levels of function to avoid confusion.

[0227] Residual corrections are generated based on multimodal fusion features:

[0228]

[0229] in, This represents the residual correction amount corresponding to the h-th prediction step. and This represents the trainable parameters of the residual correction layer.

[0230] The final wind power forecast is obtained as follows:

[0231]

[0232] in, This represents the final predicted wind power value corresponding to the h-th prediction step. This represents the base power prediction value. This represents the overall impact coefficient of the event. Indicates the event gating weight, This represents the residual correction amount. When there are no valid maintenance events within the h-th prediction time window... The final predicted value degenerates into the basic power predicted value. When valid operational events exist and their time overlap ratio, event level, event object matching relationship, and text semantics have a strong impact, Increase event gating weight and residual correction Together, they perform event-driven corrections to the baseline power forecast.

[0233] 13. Calculation process of event-constrained loss function

[0234] During the model training phase, an event-constrained loss function is constructed:

[0235]

[0236] in, This represents the total loss function of the model. This represents the power prediction error loss term. This indicates the event timing constraint. This indicates textual semantic guidance constraints. and These represent the weight coefficients of the event timing constraint and the text semantic guidance constraint, respectively.

[0237] The power prediction error loss term is used to constrain the deviation between the final predicted wind power value and the actual wind power value:

[0238]

[0239] in, Indicates the total number of prediction steps. This represents the final predicted wind power value corresponding to the h-th prediction step. This represents the actual wind power value corresponding to the h-th prediction step.

[0240] The event time constraint term is used to constrain the comprehensive event impact coefficient generated by the model. Influence intensity of the reference event constructed from the event time overlap relationship Maintain consistency:

[0241]

[0242] in, This represents the comprehensive event impact coefficient corresponding to the h-th prediction step generated by the model. This represents the intensity of the influence of a reference event, constructed based on the time overlap ratio and the importance of the events.

[0243] Textual semantic guidance constraints are used to constrain the event gating weights generated by the model. Reference event gating strength Maintain consistency and constrain residual correction amounts. Direction and power influence direction To maintain consistency, the specific approach is to first normalize the residual correction amount, and then apply a squared penalty.

[0244]

[0245]

[0246] in, This represents the event gating weight corresponding to the h-th prediction step generated by the model. This represents the gating strength of the reference event corresponding to the h-th prediction step. Indicates the intensity of the impact of the reference event. This indicates the direction of power influence corresponding to the h-th prediction step. Indicates rated power. This represents the residual correction amount corresponding to the h-th prediction step. If the power value and the residual correction amount have been normalized, then... Can be directly taken as The first term in this formula constrains the consistency between the event gating weight and the reference gating strength, while the second term is a directional constraint term. A penalty is applied only when the direction of the residual correction is inconsistent with the direction of the power effect. Specifically, when When this occurs, it indicates that maintenance events within the predicted time window generally have a reducing effect on power consumption. If at this time... A positive value results in a penalty, prompting the model to generate a negative residual correction to reflect the power reduction trend; when... When this occurs, it indicates that the overall operation and maintenance event has a restorative effect on power; if at this time... A negative value incurs a penalty, prompting the model to generate a positive residual correction to reflect the power recovery trend; when At this time, the direction constraint term has no effect. The weighting coefficients used as directional constraints impose stricter directional constraints on prediction step sizes for events with stronger impact, while automatically disabling directional constraints when there are no events. When equipment maintenance text represents routine inspections or no obvious abnormal events, the reference event gating strength is used. A smaller value is used to constrain the model to generate weaker event gating weights and residual corrections.

[0247] The impact intensity of the above reference events Reference event gating strength Direction of power influence All outputs are uniformly generated by the event impact assessment module, which has been pre-trained offline. During the iterative training of the main prediction model, all network parameters of the event impact assessment module remain fixed and are not updated synchronously with the parameters of the main model. This ensures the stability of the supervision anchor point of the loss function and avoids the drift of the constraint target.

[0248] XIV. Training and Reasoning Process

[0249] During the training process, training samples are first constructed based on wind turbine operation sequence data, meteorological and sea condition data, and equipment operation and maintenance text information. Then, a large language model is used to perform structured extraction and semantic feature encoding of the equipment operation and maintenance text. Next, the time overlap ratio between each operation and maintenance event and each prediction time window is calculated. And generate the event-affected input vector. Then, the comprehensive event impact coefficient is obtained by using an event impact coefficient generation network and multi-event attention weighted fusion. Simultaneously generate reference event gating strength Direction of power influence and the intensity of the influence of the reference event The reference event impact strength, reference event gating strength, and power impact direction are all generated by a pre-trained offline event impact assessment module. The parameters of this module are frozen throughout the main model training and do not participate in gradient updates. Subsequently, runtime sequence coding features, meteorological and sea state coding features, operation and maintenance text coding features, and time-series enhancement features are input into the feature fusion prediction layer to obtain the final wind power prediction value. Finally, the model parameters are optimized using the event-constrained loss function. Trainable parameters include the LoRA low-rank matrix parameters in the wind turbine runtime sequence data encoding sublayer, the meteorological and sea state data encoding sublayer parameters, the LoRA parameters in the equipment operation and maintenance text data encoding sublayer, the event impact coefficient generation network parameters, the attention network parameters, the event modulation gating and event residual representation network parameters, the two-stage gating fusion network parameters, the base power prediction layer parameters, the event gating network parameters, the residual correction layer parameters, and the text semantic feature mapping layer parameters. The above trainable parameters do not include the network parameters of the event impact assessment module, whose parameters were pre-trained offline and fixed before the main model training started.

[0250] During inference, the input includes wind turbine operating sequence data, meteorological and sea condition data, and equipment operation and maintenance text information prior to the time period to be predicted. Following the same method as the training process, structured event fields are extracted, time overlap ratios are calculated, comprehensive event impact coefficients are generated, multimodal fusion features are obtained, and the predicted wind power values ​​for each future prediction step are output. The inference process does not require calculating the influence intensity of reference events. Reference event gating strength and the direction of power influence The above reference values ​​are only used for loss function constraints during the training phase.

[0251] Through the above specific calculation process, a complete calculation link can be formed, from unstructured equipment operation and maintenance text to structured event fields, from structured event fields to event impact coefficients, from event impact coefficients to time-series enhancement features and final prediction corrections, and then to event constraint loss functions. This makes the key technical aspects of the present invention have clear engineering implementation methods and reproducibility.

[0252] To verify the effectiveness of the ultra-short-term wind power forecasting method provided above, historical SCADA operation data of wind turbines from a wind farm were selected for testing. The test data included power data from the wind turbines from January 1, 2025 to February 28, 2025. Specifically, the January 2025 data and the February data obtained before each forecast time were used to construct the historical input sequence of the model, while the data from February 1, 2025 to February 28, 2025 were used for rolling forecast testing. This embodiment employs a Chronos-2 time series forecasting large-scale model with fine-tuned parameters.

[0253] The prediction process is executed every 15 minutes. For missing data in the historical input sequence, time interpolation, forward imputation, and backward imputation are used sequentially to complete the data. During the entire test, a total of 2,687 valid prediction start samples were generated, each sample containing 16 prediction steps, resulting in a total of 42,992 predicted values.

[0254] The tests were conducted on a computer equipped with an Intel Core-i5 processor, an NVIDIA GeForce RTX 4080 SUPER graphics processor with 16GB of video memory and 32GB of RAM, running Windows 11 and Python version 3.11. During the tests, the historical input length was fixed at 96, the prediction length at 16, and the batch sample size at 128. The final output was the predicted value.

[0255] Depend on Figure 3 and Figure 4 As can be seen from the prediction curves, the power change trend predicted by the model implemented using the present invention is basically consistent with the actual power change trend, and it can track the rise, fall, and low-power operation of wind turbine power. As the prediction time domain increases, the prediction error increases somewhat, but the model still exhibits high trend-following ability and small system bias in short-term predictions. This indicates that the method of the present invention can utilize historical SCADA power sequences and maintenance log text to extract the time-related features of wind power, providing technical support for ultra-short-term power prediction of wind turbines.

[0256] In summary, the advantages of the present invention include:

[0257] First, this invention integrates wind turbine operation sequence data, meteorological and sea state data, and equipment operation and maintenance text information, enabling it to simultaneously characterize the impact of operating status, external environment, and equipment operation and maintenance events on offshore wind power output, thus improving the comprehensiveness of ultra-short-term forecasts in complex operating scenarios. Traditional methods only utilize numerical operating data and meteorological data, neglecting the significant impact of operation and maintenance events such as maintenance, shutdowns, faults, and power limitations on power output. This invention compensates for this information gap by introducing operation and maintenance text modalities.

[0258] Secondly, this invention performs semantic parsing and event structure extraction on equipment operation and maintenance text information based on a large language model fine-tuned by LoRA in the field of equipment operation and maintenance. It can extract information such as event type, event time, event object, event level and power impact direction from unstructured operation and maintenance text, and output structured fields in JSON format by constraining the model through prompt word templates. This realizes the automatic transformation from unstructured text to computable features and improves the model's ability to understand operation and maintenance events.

[0259] Third, this invention aligns equipment maintenance events with the prediction time windows corresponding to each prediction step through a text-based time alignment and impact coefficient generation layer. It then generates an event impact coefficient based on the time overlap ratio. Through interval intersection calculation, the same maintenance event achieves different time overlap ratios at different prediction steps, enabling the model to distinguish the varying degrees of impact of the same event at different prediction steps. Furthermore, the single-event impact coefficient employs a multiplicative gating design of candidate impact strength, time overlap ratio, and event object matching flag. When there is no time overlap between the event and the prediction window, or when the event is unrelated to the prediction object, the impact coefficient automatically resets to zero, ensuring the temporal and object-specific validity of the event's impact.

[0260] Fourth, this invention utilizes the textual modal feature vectors of maintenance events and the comprehensive event impact coefficient to perform gating modulation and residual correction on the initial time series prediction representation output by the large time series model, obtaining time series enhancement features that fuse the impact of equipment maintenance events. The comprehensive event impact coefficient acts as a multiplicative switch to control the on and off of event modulation. When there is no event impact, the time series enhancement features automatically degenerate into the initial time series representation; when the event impact is enhanced, the modulation effect is correspondingly enhanced, thereby improving the model's dynamic perception capability for events such as maintenance, downtime, power limiting, faults, and recovery.

[0261] Fifth, this invention employs an event-driven, two-stage gated fusion structure. The first stage generates basic temporal environment fusion features based on runtime sequence features and meteorological / sea state features, reflecting the foundation for power prediction under normal operating conditions. The second stage generates event correction features based on maintenance text features, time-series enhancement features, and event impact coefficients, and then overlays these features onto the basic temporal environment fusion features through gating, thereby achieving hierarchical dynamic fusion of normal operating trends and equipment maintenance event correction information. This structure decouples basic prediction and event correction at the feature level, facilitating the model's separate learning of temporal patterns under normal operating conditions and correction strategies under the influence of maintenance events.

[0262] Sixth, this invention decomposes the final wind power forecast into a combination of the base power forecast, the comprehensive event impact coefficient, the event gating weight, and the residual correction amount. The comprehensive event impact coefficient reflects the degree of event influence, the event gating weight reflects the correction strength learned by the model, and the residual correction amount reflects the correction direction and magnitude. This decomposition makes the forecast correction process clearly interpretable, facilitating operation and maintenance personnel's understanding of the composition of the forecast results and the contribution of event impacts.

[0263] Seventh, this invention constructs an event-constrained loss function, introducing event time constraints and textual semantic guidance constraints on top of power prediction error constraints. The event time constraint maintains consistency between the comprehensive event impact coefficient generated by the model and the event time overlap relationship by referencing the event impact strength constraint. The textual semantic guidance constraint maintains consistency between the event gating weights and the operational text semantics by referencing the event gating strength constraint, and maintains consistency between the direction of the residual correction amount and the event type semantics by constraining the direction of the power impact direction constraint. These reference quantities are based on operational prior rules, combined with pre-trained and adapted weight parameters and a semantic mapping network adapted to different operational scenarios. Through module pre-training and parameter freezing, the constraint supervision anchor point is kept fixed throughout training, effectively avoiding constraint target drift and further improving the model training convergence stability and the interpretability of event constraints.

[0264] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent prediction of offshore wind power output by integrating equipment operation and maintenance data, characterized in that, Includes the following steps: Acquire and preprocess multimodal data of offshore wind turbines, including turbine operation sequence, meteorological and sea conditions, and equipment operation and maintenance text information; Semantic parsing and structured extraction are performed on the operation and maintenance text to obtain text semantic features and structured event fields containing event start and end time, type, level, object, and processing status; Align the event time interval with each prediction window, calculate the time overlap ratio, combine the structured event fields with the wind turbine object matching relationship, generate the single event impact coefficient, and aggregate to obtain the comprehensive event impact coefficient for each step; the value is zero when there is no time overlap or no object association. The runtime sequence and meteorological and sea state data are encoded, and an initial time series prediction representation is obtained based on the encoding feature time series modeling. The time series enhancement feature is obtained by modulating the comprehensive event impact coefficient and text semantic features. By integrating multimodal coding features and temporal enhancement features, a base power prediction value, event gating weights, and residual corrections are generated. An event correction term is constructed based on the comprehensive event impact coefficient, event gating weights, and residual corrections. The event correction term is then superimposed on the base power prediction value to obtain the final wind power prediction value. Construct an event-constrained loss function that includes a power prediction error loss term, an event time constraint term, and a text semantic guidance constraint term. Train the model and output multi-step prediction results.

2. The intelligent prediction method for offshore wind power based on integrated equipment operation and maintenance data according to claim 1, characterized in that: The process of generating the single-event influence coefficient includes: The text semantic features, the time overlap ratio, the event type encoding, the event level normalization value, the event object matching flag, and the processing status encoding are concatenated along the feature dimension to obtain the event impact input vector; Input the event impact vector into the event impact coefficient generation network to obtain the candidate impact strength; The single-event influence coefficient is obtained by multiplying the candidate influence intensity by the time overlap ratio and the event object matching flag; The event object matching flag is graded according to the degree of correlation between the event object and the wind turbine to be predicted: it is 1 when there is a complete match or the same site; it is a preset weak matching coefficient when there is a weak correlation in the same region, and the preset weak matching coefficient is greater than 0 and less than 1; it is 0 when there is no correlation.

3. The intelligent prediction method for offshore wind power based on integrated equipment operation and maintenance data according to claim 2, characterized in that: When multiple operation and maintenance events exist within the same prediction time window, an attention input vector is constructed based on the event impact input vector and the single event impact coefficient of each operation and maintenance event. The attention weight of each operation and maintenance event is calculated through an attention network. The attention weight is used to weight and fuse multiple single event impact coefficients and then truncated to a preset interval to obtain the comprehensive event impact coefficient.

4. The intelligent prediction method for offshore wind power based on integrated equipment operation and maintenance data according to claim 1, characterized in that: The process of modulating the initial time-series prediction representation includes: The event modulation gating and event residual representation are generated by combining the comprehensive event impact coefficient with the text semantic features; The event residual representation is weighted by event modulation gating and comprehensive event influence coefficient, and then superimposed onto the initial time series prediction representation to obtain the time series enhancement feature; When the overall event impact coefficient is zero, the time series enhancement features are consistent with the initial time series prediction characterization.

5. The intelligent prediction method for offshore wind power based on integrated equipment operation and maintenance data according to claim 1, characterized in that: The fusion adopts an event-condition-driven, two-stage gating fusion structure, specifically including: In the first stage, basic temporal environment fusion features are generated based on runtime sequence coding features and meteorological and sea state coding features; In the second stage, by combining the basic temporal environment fusion features, the encoding features corresponding to the text semantic features, the temporal enhancement features, and the comprehensive event influence coefficient, event correction features and event fusion gating weights are generated. The event correction features are then weighted by the event fusion gating weights and superimposed on the basic temporal environment fusion features to obtain multimodal fusion features.

6. The intelligent prediction method for offshore wind power based on integrated equipment operation and maintenance data according to claim 1, characterized in that: The power prediction error loss term is used to constrain the deviation between the final wind power prediction value and the actual power value. The event time constraint term is used to ensure that the comprehensive event impact coefficient is consistent with the impact intensity of the reference event constructed based on the time overlap relationship; The text semantic guidance constraint is used to ensure that the event gating weight is consistent with the reference event gating strength generated based on the text semantic features, and to ensure that the direction of the residual correction is consistent with the power influence direction determined based on the event type.

7. The intelligent prediction method for offshore wind power based on integrated equipment operation and maintenance data according to claim 6, characterized in that: The text semantic guidance constraint includes a directional penalty term: When the power influence direction indicator power decreases and the residual correction is positive, or when the power influence direction indicator power recovers and the residual correction is negative, the directional penalty term generates a penalty value. When the power influence on the direction indication has no significant effect, the directional penalty term is not effective.

8. The intelligent prediction method for offshore wind power based on integrated equipment operation and maintenance data according to claim 6, characterized in that: The reference event impact strength, reference event gating strength, and power impact direction are all generated by the pre-trained event impact assessment module. The event impact assessment module is pre-trained offline based on historical equipment operation and maintenance data and historical wind power data. During the training phase of the offshore wind power ultra-short-term intelligent prediction model, its network parameters remain fixed and are only used to generate reference supervision signals in the event constraint loss function. The event impact assessment module takes the textual semantic features of the operation and maintenance event, structured event fields, event object matching flags, and the time overlap ratio between the event and the predicted time window as inputs. It includes three output branches: event impact intensity, reference event gating intensity, and power impact direction. Internally, the module determines the event type weight and power impact direction based on the event type, the event level weight based on the event level, and the processing status weight based on the processing status. Combining the event object matching relationship and the time overlap ratio, the various weights are multiplicatively combined to generate the reference event impact intensity. In multi-event scenarios, the comprehensive reference value for the corresponding step size is obtained by aggregating the attention weights pre-trained within the module.

9. The intelligent prediction method for offshore wind power based on integrated equipment operation and maintenance data according to claim 1, characterized in that: The semantic parsing and structured extraction are performed using a large language model fine-tuned with low-rank adaptation based on samples from the equipment operation and maintenance domain. The backbone parameters of the large language model are frozen, and low-rank adaptation layers are inserted at the query matrix projection layer, key matrix projection layer, value matrix projection layer, and output projection layer of its attention layer. Only the parameters of the low-rank adaptation layers are trained. The structured event fields are output by the large language model constrained by operation and maintenance domain prompt word templates. The text semantic features are obtained from the last hidden state of the large language model through average pooling, linear mapping, and layer normalization.

10. A smart offshore wind power prediction system integrating equipment operation and maintenance data, used to implement the method as described in claim 1, characterized in that, include: The data preprocessing unit is used to acquire and preprocess multimodal data of offshore wind turbines; The multimodal feature encoding unit is used to perform semantic parsing and structured extraction on operation and maintenance text to obtain text semantic features and structured event fields; The operation and maintenance event text time sequence alignment and impact coefficient generation unit is used to align the event time interval with the prediction window and generate the comprehensive event impact coefficient for each step. The temporal enhancement modeling unit is used for temporal modeling to obtain an initial temporal prediction representation and modulating it to obtain temporal enhancement features; The feature fusion prediction unit is used to fuse multimodal features and generate the final wind power prediction value; The model training unit is used to construct the event constraint loss function and train the model.

Images