Cross-domain adaptive wind power data analysis method, system and related device

Abstract

The invention discloses a cross-domain adaptive wind power data analysis method and system and a related device, which are used for more effectively processing complex dynamic change and scene diversity of wind power data and improving the accuracy and robustness of wind power prediction. The method comprises the steps that historical wind power data of a source domain and historical wind power data of a target domain are acquired respectively; dynamic features of historical wind power data are extracted by using bidirectional LSTM and an attention mechanism, and initial time sequence feature representations of a source domain and a target domain are generated; dynamically aligning the initial time sequence feature representation based on a sliding time window and a cross-domain double-discriminator adversarial training method to obtain a target time sequence feature representation; generating a multi-scale migration weight and a migration confidence coefficient according to the target time sequence feature representation; constructing a pre-training sub-model set according to the target time sequence feature representation, the multi-scale migration weight and the migration confidence; and calling the target sub-model according to the matching degree of the current time sequence feature and the sub-model in combination with the migration confidence, and generating a wind power prediction result of the target domain.

Description

Technical Field

[0001] This application relates to the field of meteorological data processing, and in particular to a method, system and related apparatus for cross-domain adaptive wind power data analysis. Background Technology

[0002] As a clean energy source, the accuracy of wind power forecasting is crucial for the stable operation and dispatch of the power system. One of the most critical issues in wind power forecasting is the dynamic change in wind conditions. Meteorological factors such as wind speed and direction are affected by seasonal variations, sudden weather changes, and other factors, causing the characteristic distribution of wind farms to change dynamically over time, thus impacting the forecast results.

[0003] In existing technologies, transfer learning or wind scene clustering methods are used to address the problem of dynamically changing wind conditions. However, traditional transfer learning methods typically assume that the feature distributions of the source and target domains have a static similarity. In reality, wind conditions are affected by seasons, sudden weather changes, and other factors, causing the feature distribution to change dynamically over time. This results in wind speed prediction errors fluctuating non-linearly with meteorological conditions, making it difficult for transfer learning methods to capture this dynamic correlation. Secondly, existing technologies often match the overall distribution at the domain level, ignoring the differences in the contribution of different time segments in the source domain to the target scene. This leads to insufficient dynamic adaptability of traditional transfer learning methods in wind power data analysis, ultimately affecting the accuracy of the prediction. Summary of the Invention

[0004] This application provides a method, system, and related apparatus for cross-domain adaptive wind power data analysis, which can more effectively handle the complex dynamic changes and scenario diversity of wind power data, and improve the accuracy and robustness of wind power prediction.

[0005] The first aspect of this application provides a method for cross-domain adaptive wind power data analysis, including:

[0006] Historical wind power data for the source and target domains are acquired separately. The historical wind power data includes long-term wind speed, wind direction, meteorological variables, and power output time series.

[0007] The dynamic features of the historical wind power data are extracted using a bidirectional LSTM and attention mechanism to generate initial temporal feature representations of the source domain and the target domain;

[0008] The initial temporal feature representation is dynamically aligned based on a sliding time window and a cross-domain dual discriminator adversarial training method to obtain a target temporal feature representation, which contains temporal feature representations of multiple time segments.

[0009] Based on the target time-series feature representation, the correlation between the source domain and the target domain in the whole time-series period and in each time segment is calculated through multi-level mutual information, and multi-scale transfer weights and transfer confidence are generated based on the calculation results.

[0010] A pre-trained sub-model set is constructed based on the target temporal feature representation, the multi-scale transfer weights, and the transfer confidence. The pre-trained sub-model set includes multiple sub-models optimized for different wind conditions. The transfer confidence is used to dynamically adjust the training priority of the sub-models.

[0011] The real-time wind power data of the target domain is obtained and the current time-series features are extracted using a sliding time window. Based on the matching degree between the current time-series features and the sub-model, combined with the migration confidence, the target sub-model is invoked to generate the wind power prediction result of the target domain.

[0012] Optionally, the step of extracting dynamic features from the historical wind power data using bidirectional LSTM and an attention mechanism to generate initial temporal feature representations of the source and target domains includes:

[0013] A bidirectional LSTM is used to perform time-series modeling on the historical wind power data to extract the long-term dependency features and short-term fluctuation features of the historical wind power data.

[0014] An attention mechanism is introduced to assign time step weights to the output of the bidirectional LSTM, and these time step weights are used to highlight the dynamic characteristics of sudden wind speed changes or meteorological events.

[0015] Based on the output of the bidirectional LSTM and the time step weights, initial temporal feature representations of the source domain and the target domain are generated by weighted combination.

[0016] Optionally, the method of dynamically aligning the initial temporal feature representation based on a sliding time window and cross-domain dual discriminator adversarial training to obtain the target temporal feature representation includes:

[0017] The initial temporal feature representations of the source domain and the target domain are segmented using a sliding time window to generate temporal feature representations of multiple time segments;

[0018] A cross-domain dual discriminator structure is constructed, which includes a global discriminator and a local discriminator. The global discriminator is used to optimize the consistency of the overall feature distribution between the source domain and the target domain, and the local discriminator is used to optimize the dynamic pattern differences within each time window.

[0019] By conducting adversarial training based on generative adversarial networks, the loss functions of the global discriminator and the local discriminator are jointly optimized, and the temporal feature representations of the source domain and the target domain are mapped to a shared feature space to generate the target temporal feature representation.

[0020] Optionally, the method further includes:

[0021] The length of the sliding time window is dynamically adjusted based on the wind speed fluctuation amplitude and meteorological event frequency in the historical wind power data.

[0022] Optionally, the step of calculating the correlation between the source domain and the target domain over the entire time series and in each time segment based on the target temporal feature representation using mutual information, and generating multi-scale transfer weights and transfer confidence based on the calculation results, includes:

[0023] Based on the target time-series feature representation, calculate the first mutual information value between the source domain and the target domain over the entire time-series period, and calculate the second mutual information value between the source domain and the target domain in each time segment;

[0024] The first mutual information value and the second mutual information value are fused to generate multi-scale transfer weights for each time segment of the source domain, and the transfer confidence is calculated based on the distribution characteristics of the second mutual information value. The multi-scale transfer weights are used to adjust the contribution of the time segment to the prediction task of the target domain.

[0025] Optionally, constructing a pre-trained sub-model set based on the target temporal feature representation, the multi-scale transfer weights, and the transfer confidence includes:

[0026] Multiple sub-models are initialized based on the target temporal feature representation and the multi-scale transfer weights, and the parameters of the sub-models are optimized for different wind conditions.

[0027] Training resources are dynamically allocated based on the transfer confidence. A pre-trained sub-model set containing multiple sub-models is generated by training through a shared backbone network and scene-specific output layer structure. The pre-trained sub-model set is used to support prediction tasks for different wind conditions.

[0028] A second aspect of this application provides a system for cross-domain adaptive wind power data analysis, comprising:

[0029] The acquisition unit is used to acquire historical wind power data of the source domain and the target domain respectively. The historical wind power data includes long-term wind speed, wind direction, meteorological variables and power output time series.

[0030] The extraction unit is used to extract the dynamic features of the historical wind power data using bidirectional LSTM and attention mechanism, and generate the initial temporal feature representations of the source domain and the target domain.

[0031] An alignment unit is used to dynamically align the initial temporal feature representation based on a sliding time window and a cross-domain dual discriminator adversarial training method to obtain a target temporal feature representation, wherein the target temporal feature representation contains temporal feature representations of multiple time segments;

[0032] The calculation unit is used to calculate the correlation between the source domain and the target domain in the whole time series and in each time segment based on the target time series feature representation through multi-level mutual information, and generate multi-scale transfer weights and transfer confidence based on the calculation results.

[0033] The transfer unit is used to construct a pre-trained sub-model set based on the target temporal feature representation, the multi-scale transfer weights, and the transfer confidence. The pre-trained sub-model set includes multiple sub-models optimized for different wind conditions. The transfer confidence is used to dynamically adjust the training priority of the sub-models.

[0034] The prediction unit is used to acquire real-time wind power data of the target domain and extract the current time-series features using a sliding time window. Based on the matching degree between the current time-series features and the sub-model, combined with the migration confidence, the target sub-model is invoked to generate the wind power prediction result of the target domain.

[0035] A third aspect of this application provides an apparatus for cross-domain adaptive wind power data analysis, the apparatus comprising:

[0036] Processor, memory, input / output units, and bus;

[0037] The processor is connected to the memory, the input / output unit, and the bus;

[0038] The memory stores a program that the processor invokes to execute the first aspect and any optional method of cross-domain adaptive wind power data analysis in the first aspect.

[0039] The fourth aspect of this application provides a computer-readable storage medium storing a program that, when executed on a computer, performs the first aspect and any optional method of the first aspect for cross-domain adaptive wind power data analysis.

[0040] As can be seen from the above technical solutions, this application has the following advantages:

[0041] By extracting dynamic features from historical wind power data using a bidirectional LSTM combined with an attention mechanism, the system effectively captures the dependencies and nonlinear correlations of wind speed, wind direction, and meteorological variables over time, deeply extracting dynamic features. Subsequently, a sliding time window and cross-domain dual discriminator adversarial training method are employed to dynamically align the temporal feature representations of the source and target domains. Based on the aligned target temporal feature representation, the mutual information between the source and target domains throughout the entire time series and at each time segment is calculated to quantify the correlation strength between specific wind conditions in the target domain and different historical segments in the source domain, generating multi-scale transfer weights and transfer confidence scores. These parameters directly reflect the importance and reliability of source domain knowledge at different scales and scenarios. Finally, multiple sub-models optimized for different wind conditions are constructed based on this information, and the training of these sub-models is guided by multi-scale transfer weights to ensure that the most relevant knowledge from the source domain is transferred to the corresponding scenario model. Simultaneously, transfer confidence scores are used to dynamically adjust the training priority of sub-models and optimize resource allocation. Finally, during real-time prediction, the system dynamically selects the sub-model best suited for the current real-time wind condition scenario based on the matching degree between the current temporal features and the sub-model, combined with the transfer confidence scores.

[0042] This application overcomes the shortcomings of traditional transfer learning methods, such as static assumptions, neglect of local contributions, and insufficient dynamic adaptability. It can more effectively handle the complex dynamic changes and scene diversity of wind power data, thereby significantly improving the accuracy and robustness of prediction. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 A schematic flowchart of an embodiment of the cross-domain adaptive wind power data analysis method provided in this application;

[0045] Figure 2 A schematic flowchart of another embodiment of the method for cross-domain adaptive wind power data analysis provided in this application;

[0046] Figure 3 A schematic diagram of an embodiment of the system for cross-domain adaptive wind power data analysis provided in this application;

[0047] Figure 4 A schematic diagram of an embodiment of the apparatus for cross-domain adaptive wind power data analysis provided in this application. Detailed Implementation

[0048] This application provides a method, system, and related apparatus for cross-domain adaptive wind power data analysis, which can more effectively handle the complex dynamic changes and scenario diversity of wind power data, and improve the accuracy and robustness of wind power prediction.

[0049] Please see Figure 1 , Figure 1 An embodiment of the method for cross-domain adaptive wind power data analysis provided in this application includes:

[0050] 101. Obtain historical wind power data for the source and target domains respectively. Historical wind power data includes long-term wind speed, wind direction, meteorological variables, and power output time series.

[0051] In wind power data prediction, the source domain and the target domain represent different data sources or environments. The source domain typically refers to known wind farm data with abundant historical records and annotation information, while the target domain is the new wind farm to be predicted or an environment with less data. By acquiring historical wind power data from both, the data information from the source domain can be used to assist the prediction task in the target domain.

[0052] Specifically, the first step is to collect historical wind power data for both the source and target domains. This historical wind power data includes long-term time-series data for wind speed (m / s), wind direction (degrees), meteorological variables (such as temperature °C, humidity % and air pressure hPa), and power output (kW or MW). The time span should cover at least quarters to several years to reflect seasonality and long-term trends. The source domain data is extracted from existing wind farms or databases, covering long-term time series, while the target domain data is collected from the wind farm to be predicted, and should at least include representative periods (such as seasonal variations). The data types of both must be consistent.

[0053] 102. Utilize bidirectional LSTM and attention mechanism to extract dynamic features from historical wind power data, and generate initial temporal feature representations for the source and target domains;

[0054] Wind power forecasting is essentially a time-series forecasting problem, specifically involving the dynamic changes in meteorological variables such as wind speed and direction, as well as power output over time. However, wind power forecasting requires processing complex data with multiple variables, strong time-series dependencies, and the influence of sudden meteorological changes. Traditional LSTM can only capture positive time-series information, while in reality, wind power is affected by both historical and future meteorological conditions. For example, sudden changes in air pressure may indirectly cause power fluctuations the following day. Furthermore, the contributions of different variables to power output vary significantly at different times; for instance, wind speed is more important than temperature during periods of strong winds. Therefore, after obtaining historical wind power data, this embodiment chooses to combine bidirectional LSTM and an attention mechanism to extract complex time-series feature representations from wind power data.

[0055] Historical wind power data is organized into a fixed-length input sequence or sliding window sequence according to time series. The input sequence is first passed through one or a bidirectional LSTM layer, which is a variant of a recurrent neural network (RNN). Specifically, the bidirectional LSTM consists of two LSTM layers: a forward layer that processes information from the past to the future, and a backward layer that processes information from the future to the past. The backward layer can introduce the indirect influence of future meteorological conditions. For example, a sudden drop in air pressure at a certain moment may not immediately cause a change in power, but through the backward propagation of the bidirectional LSTM, it can be associated with a power decrease event the following day. For each time step in the historical wind power data, the forward LSTM generates a forward hidden state, and the backward LSTM generates a backward hidden state. Concatenating these two hidden states generates the hidden state representation of the current time step, which encodes the bidirectional contextual information of the current time point in the entire sequence.

[0056] Furthermore, considering that although bidirectional LSTM can capture temporal dependencies, it treats all time steps equally, failing to highlight critical moments more important for prediction. For example, data from sudden wind speed changes or meteorological events have a far greater impact on power prediction than data from ordinary times. Therefore, an attention mechanism is applied to the hidden state representation of the bidirectional LSTM output. This mechanism learns a weight distribution, calculates the importance of the hidden state at each time step to the final feature representation, and then uses these scores to weightedly sum the hidden states at all time steps, resulting in a fixed-dimensional vector. This vector is the initial temporal feature representation corresponding to the input sequence. This initial temporal feature representation is a compression and abstraction of the dynamic features of the entire input sequence. By performing the above process on each input historical time series window in both the source and target domains, corresponding initial temporal feature representation sets for the source and target domains are generated, respectively.

[0057] 103. Based on the sliding time window and cross-domain dual discriminator adversarial training method, the initial temporal feature representation is dynamically aligned to obtain the target temporal feature representation, which contains the temporal feature representations of multiple time segments;

[0058] Because the source and target domains may differ in geographical location, climate conditions, equipment status, etc., i.e., there is domain drift, the distribution of these initial feature representations in the feature space may still differ. Directly using features with such distribution differences for subsequent prediction model training and application will lead to a decrease in the generalization ability of a model that performs well in the source domain in the target domain. Therefore, in this embodiment, step 103 of domain adaptation is needed to reduce the distribution differences between the features of the source and target domains, so that they are aligned as much as possible in the shared feature space.

[0059] However, since wind power data is dynamically changing, domain differences may exhibit different patterns under different wind conditions or time periods. Simply aligning the static statistical characteristics of the entire long-term sequence may be insufficient. Therefore, this embodiment uses a sliding time window to segment the long sequence into multiple short feature sequence segments, enabling the alignment process to focus on and handle dynamic pattern differences in different local time segments, achieving dynamic alignment. The specific alignment relies on the idea of ​​Generative Adversarial Networks (GANs). A discriminator is introduced to distinguish the source domain of features (whether it's the source domain or the target domain) and is trained adversarially against a "generator" (a network that implements feature mapping). This forces the generator to learn a mapping, making the features of the source and target domains indistinguishable after transformation. Through alternating or joint optimization, the feature mapping network and the two discriminators engage in a game, ultimately driving the feature mapping network to map the features of the source and target domains to a shared feature space that the discriminators cannot distinguish. The initial temporal feature representation of the target domain is input into the trained feature mapping network, and the final output is the aligned target temporal feature representation located in the shared feature space. It should be noted that this target temporal feature representation specifically refers to the temporal feature representation of the target domain.

[0060] 104. Based on the target time series feature representation, calculate the correlation between the source domain and the target domain in the whole time series period and in each time segment through multi-level mutual information, and generate multi-scale transfer weights and transfer confidence based on the calculation results;

[0061] Although step 103 aligns the initial feature representations of the source and target domains to a shared feature space through cross-domain adversarial training, reducing the overall distributional differences, more targeted transfer learning requires quantifying the specific correlation between the source and target domains under different wind conditions or operational scenarios. Therefore, the core objective of step 104 is to quantify the correlation between the source and target domains at different granularities based on the aligned target temporal feature representation (and the corresponding source domain feature representation), and to generate key parameters guiding knowledge transfer, namely multi-scale transfer weights and transfer confidence. Specifically, step 104 employs multi-level mutual information (MI) as a correlation measurement tool because MI can capture complex nonlinear dependencies between variables and reflects the actual degree of information sharing better than linear correlation.

[0062] In practice, the mutual information between the source and target domains is first calculated over the entire time series to assess overall domain similarity and overall knowledge transferability. Then, for the temporal feature representations of the target domain generated through a sliding time window, the mutual information between these representations and relevant features of the source domain is calculated to identify the correlation strength between specific local scenes in the target domain and historical experience in the source domain. Based on these multi-level mutual information calculations, multi-scale transfer weights and transfer confidence are generated. Higher mutual information values ​​indicate a stronger correlation between the source and target domains (or their specific time segments), more information sharing, and correspondingly higher transfer weights and confidence. The multi-scale transfer weights are used in subsequent steps to determine the extent to which information from the source domain (e.g., sub-models trained under specific wind conditions, or source data related to a target segment) should be used for prediction in the target domain. The transfer confidence is used to assess the reliability of the entire transfer process or to dynamically adjust subsequent model selection and fusion strategies.

[0063] 105. Construct a pre-trained sub-model set based on the target temporal feature representation, multi-scale transfer weights, and transfer confidence. The pre-trained sub-model set includes multiple sub-models optimized for different wind conditions. The transfer confidence is used to dynamically adjust the training priority of the sub-models.

[0064] One of the challenges in wind power forecasting lies in the complexity and variability of wind conditions. The power output behavior of wind farms can exhibit drastically different nonlinear relationships under various scenarios, such as high wind speed, low wind speed, drastic wind speed changes, and wind direction variations. A single global model is unlikely to achieve optimal performance under all these scenarios simultaneously. Therefore, it is necessary to construct a series of sub-models optimized for different wind conditions. Different wind conditions can be addressed through cluster analysis of wind characteristics (such as K-means, DBSCAN, etc.), with each cluster representing a typical wind condition or operating state scenario. Alternatively, scenarios can be defined based on rules, such as thresholds for wind speed range, wind speed change rate, and wind direction stability. The defined scenarios should cover all possible situations that may occur in the target domain.

[0065] Specifically, based on the target time-series feature representation (and corresponding source domain aligned feature representation) obtained in step 103, various typical wind condition scenarios that the target wind farm may encounter are identified and defined. Subsequently, an independent prediction sub-model is constructed for each defined scenario. During the training of these sub-models, the multi-scale transfer weights calculated in step 104 guide the training process, enabling the sub-models to focus more on learning knowledge from historical wind power data or features in the source domain that are highly correlated with specific target scenarios. This effectively achieves targeted knowledge transfer and avoids the negative impact of irrelevant knowledge. Furthermore, during training, the training priority of the sub-models is dynamically adjusted based on the transfer confidence: for scenarios with higher transfer confidence, indicating that the knowledge provided by the source domain in that scenario is more reliable, the corresponding sub-model can be given higher training priority, allocated more computing resources, or adopt a more aggressive training strategy to ensure that the model can learn fully and converge quickly. Conversely, for scenarios with lower confidence, a more conservative training method or greater reliance on data from the target domain may be required for training.

[0066] 106. Obtain real-time wind power data of the target domain and extract the current time series features using a sliding time window. Based on the matching degree between the current time series features and the sub-model, combined with the transfer confidence, call the target sub-model to generate wind power prediction results for the target domain.

[0067] In the actual prediction phase, the system continuously acquires the latest real-time wind power data, including wind speed, wind direction, and meteorological variables, from the target wind farm's SCADA system or other real-time data sources. A sliding time window is applied to the acquired real-time data; the size and step size of this window should be consistent with those used during offline training (especially steps 102 and 103). Using the data within the window as input, the system extracts the current initial temporal feature representation through the same bidirectional LSTM and attention mechanism model as in step 102 (whose parameters have been learned or fine-tuned during the offline phase). This initial feature representation is then input into the feature mapping network trained in step 103, transforming it into the same shared feature space as during sub-model training to obtain the current temporal features. In the shared feature space, the matching degree between the current temporal features and the scenarios represented by each sub-model in the pre-trained sub-model set is calculated to determine the target sub-model for prediction. The target sub-model is used to perform the prediction task, outputting the wind power prediction result for a future point in time or time period.

[0068] This method can accurately identify the scene based on the current wind conditions and dynamically select and call the most suitable professional sub-model for prediction based on the transfer confidence score. This effectively addresses the variability of wind conditions and outputs more accurate and robust wind power prediction results for the target domain. For example, if the match with scene A is high and the confidence score is also high, model A is directly called; if the match is high but the confidence score is low, a more robust backup model can be selected, or the prediction results of model A can be post-processed, thus avoiding over-reliance on source domain knowledge in scenarios where transfer is unreliable.

[0069] Taking the prediction of summer gusts in the target domain using source domain data dominated by winter data as an example: Existing technologies may coarsely align all winter stable wind data in the source domain with the overall data of the target domain, and use these low-correlation winter data with equal weight or without distinction when predicting summer gusts. This results in the model failing to capture the unique dynamic patterns of summer gusts, and the prediction error increases due to the nonlinear fluctuations of meteorological conditions. This embodiment extracts the dynamic features of the data through bidirectional LSTM and attention mechanism; it uses a cross-domain dual discriminator adversarial training method including a local discriminator to align features within the sliding time window, achieving dynamic alignment at the time segment level; and then uses multi-level mutual information to accurately calculate the correlation between the current "summer gust" scene segment in the target domain and each historical segment in the source domain, identifying the phenomenon that winter stable wind data contributes little while a few summer gust data contribute highly, and generating multi-scale transfer weights and transfer confidence based on this. When constructing the pre-trained sub-model set, these weights are used to weight the training data, ensuring that the sub-model specifically trained for the "summer gust" scenario learns primarily from the most relevant summer data in the source domain, effectively avoiding negative transfer caused by a large amount of irrelevant winter data; at the same time, the training priority is dynamically adjusted using confidence. Finally, in the real-time prediction stage, this specially optimized "summer gust" sub-model is accurately matched and invoked based on the current real-time features. This mechanism enables this application to dynamically adapt to different wind conditions, capture the nonlinear fluctuations in wind speed prediction error with meteorological conditions, and fully utilize the most relevant historical experience in the source domain, overcoming the shortcomings of traditional methods and significantly improving the accuracy and robustness of wind power prediction.

[0070] In this embodiment, a bidirectional LSTM combined with an attention mechanism is used to extract dynamic features from historical wind power data. This effectively captures the dependencies and nonlinear correlations of wind speed, wind direction, and meteorological variables over time, deeply extracting dynamic features. Subsequently, a sliding time window and cross-domain dual discriminator adversarial training method are used to dynamically align the temporal feature representations of the source and target domains. Based on the aligned target temporal feature representation, the mutual information between the source and target domains throughout the entire time series and at each time segment is calculated to quantify the correlation strength between specific wind conditions in the target domain and different historical segments in the source domain. Multi-scale transfer weights and transfer confidence scores are generated; these parameters directly reflect the importance and reliability of source domain knowledge at different scales and scenarios. Finally, multiple sub-models optimized for different wind conditions are constructed based on this information, and the training of these sub-models is guided by multi-scale transfer weights to ensure that the most relevant knowledge from the source domain is transferred to the corresponding scenario model. Simultaneously, the transfer confidence score is used to dynamically adjust the training priority of the sub-models and optimize resource allocation. Finally, during real-time prediction, the most suitable sub-model for the current real-time wind condition scenario is dynamically selected based on the matching degree between the current temporal features and the sub-model, combined with the transfer confidence score.

[0071] This application overcomes the shortcomings of traditional transfer learning methods, such as static assumptions, neglect of local contributions, and insufficient dynamic adaptability. It can more effectively handle the complex dynamic changes and scene diversity of wind power data, thereby significantly improving the accuracy and robustness of prediction.

[0072] The method for cross-domain adaptive wind power data analysis provided in this application is described in detail below. Please refer to [link / reference]. Figure 2 , Figure 2 Another embodiment of the method for cross-domain adaptive wind power data analysis provided in this application includes:

[0073] 201. Obtain historical wind power data for the source and target domains respectively. Historical wind power data includes long-term wind speed, wind direction, meteorological variables, and power output time series.

[0074] In this embodiment, step 201 is similar to step 101 in the previous embodiment, and will not be described again here.

[0075] 202. Use bidirectional LSTM to perform time series modeling on historical wind power data to extract long-term dependency features and short-term fluctuation features of historical wind power data;

[0076] Historical wind power data is organized chronologically into a fixed-length input sequence or multiple sequence samples are constructed using a sliding time window. The input at each time step is a vector containing wind speed, wind direction, meteorological variables, and any other relevant features at that moment. One or more bidirectional LSTM layers are constructed, each consisting of a forward LSTM and a backward LSTM. The forward LSTM processes the input sequentially from the start to the end of the sequence, its hidden state capturing contextual information from before each time step (including the current step). The backward LSTM processes the input sequentially from the end to the start of the sequence, its hidden state capturing contextual information from after each time step (including the current step). At each time step, the hidden states of the forward and backward LSTMs are concatenated to form the output vector for that time step. This output vector incorporates the bidirectional contextual information from the entire input sequence at the current time point. The output of the bidirectional LSTM layer is an abstract feature extracted from the original data, containing both long-term dependency features and short-term fluctuation features.

[0077] The long-term dependency characteristic specifically refers to the pattern reflected by the dependency relationship between data points that are far apart in a time series, such as trends, seasonality, or long-term climate effects. The short-term fluctuation characteristic refers to the pattern reflected by rapid, localized changes between adjacent or nearby data points in a time series, such as turbulence, gusts, and instantaneous power changes.

[0078] 203. An attention mechanism is introduced to assign time step weights to the output of the bidirectional LSTM. The time step weights are used to highlight the dynamic characteristics of sudden wind speed changes or meteorological events.

[0079] By performing time-series modeling on historical wind power data, a sequence of hidden state outputs can be generated. Each hidden state vector encodes the corresponding time step and its bidirectional contextual information within the sequence. However, subsequent prediction tasks typically require a fixed-dimensional vector to represent the features of the entire time window or sequence, rather than a variable-length sequence output. Simply averaging, summing, or using only the hidden state of the last time step as the representation of the entire sequence may lose important information contained in key time points (such as sudden changes in wind speed, sudden rises / falls in temperature, and severe wind direction fluctuations). These critical wind speed abrupt changes or meteorological events are often the direct cause of large fluctuations in wind power, and their corresponding dynamic features are crucial for accurate prediction. Based on this, this embodiment employs an attention mechanism to automatically learn the importance of each time step and assign it corresponding time step weights, thereby enhancing the impact of key dynamic information such as sudden wind speed changes and important meteorological events.

[0080] 204. Based on the output of the bidirectional LSTM and the time step weights, generate the initial temporal feature representations of the source and target domains through weighted combination;

[0081] The output of the bidirectional LSTM and the importance weights calculated based on the attention mechanism are effectively aggregated into a fixed-dimensional vector that can represent the core dynamic features of the entire sequence through weighted combination. This allows the model to pay more attention to time steps that are assigned higher weights and contain key dynamic information (such as the moment of sudden wind speed change), thereby generating an initial temporal feature representation that both compresses information and highlights importance.

[0082] 205. A sliding time window is used to segment the initial temporal feature representations of the source and target domains to generate temporal feature representations of multiple time segments;

[0083] To achieve finer-grained cross-domain alignment capable of capturing local dynamic patterns, these continuous feature representation sequences need to be further decomposed. In this embodiment, a sliding time window mechanism is used to segment the initial temporal feature representation sequences of the source and target domains. Specifically, by setting a fixed window size and sliding step, a time window is moved across the feature sequence, and each time a continuous initial feature representation vector within the window is extracted to form an independent temporal feature representation of a time segment. This process is applied to the entire initial feature representation sequence of the source and target domains respectively, thereby generating a set containing a large number of source domain time segments and target domain time segments.

[0084] Considering that wind speed and meteorological conditions in wind power data do not change at a constant rate, wind speed may remain within a range for a long time under stable weather conditions with small fluctuations, while wind speed may change drastically in a short period of time before a storm, during periods of frequent gusts, or when a weather front passes, accompanied by other meteorological events (such as significant changes in temperature and air pressure). A fixed-size sliding time window is insufficient to effectively capture the temporal characteristics under both extreme conditions simultaneously. Therefore, in some specific embodiments, the length of the sliding time window can be dynamically adjusted based on the amplitude of wind speed fluctuations and the frequency of meteorological events in historical wind power data. By adjusting the window length, the extracted initial temporal feature representation can better reflect the actual changes in the current local time period: a longer window is used for stable periods (low fluctuations / low event frequency) to obtain a more global feature representation; a shorter window is used for periods of drastic change (high fluctuations / high event frequency), thereby focusing on rapid dynamic changes in a short period of time, accurately capturing the morphology and instantaneous impact of events, and avoiding the smoothing out of these key and transient features in a long window.

[0085] 206. Construct a cross-domain dual discriminator structure that includes a global discriminator and a local discriminator. The global discriminator is used to optimize the consistency of the overall feature distribution between the source domain and the target domain, and the local discriminator is used to optimize the dynamic pattern differences within each time window.

[0086] To achieve more comprehensive and refined domain alignment, this embodiment constructs a cross-domain dual discriminator structure, which specifically includes a global discriminator and a local discriminator, which are described below:

[0087] Global Discriminator: The input to the global discriminator is a large batch or an entire sequence of feature representations from the source and target domains. The core task of the global discriminator is to learn to distinguish the differences in the overall distribution of these features. By training the global discriminator, the feature mapping network is driven to adjust the features so that the macroscopic statistical properties (such as mean, variance, and higher-order moments) of the features in the source and target domains are as consistent as possible in the shared feature space. It can be understood that the global discriminator focuses on "whether, on average, the features of the two domains look the same."

[0088] Local Discriminator: This local discriminator focuses on fine-grained dynamic patterns. The input to the local discriminator is a segment of feature representations corresponding to a specific sliding time window. The local discriminator learns to distinguish whether there are differences in the dynamic behavior patterns exhibited by the source and target domain features within these specific time segments. For example, the local discriminator might attempt to distinguish subtle differences in the morphology, rate of change, and other aspects of the feature sequences from the source and target domains during similar wind speed increases. By optimizing the local discriminator, the inconsistencies in the micro-dynamics of the two domain features within a specific wind condition or time period are eliminated as much as possible.

[0089] With this dual discriminator structure, this embodiment can simultaneously measure and constrain the feature differences between the source and target domains from both global statistical and local dynamic perspectives.

[0090] 207. By adversarial training based on generative adversarial networks, the loss functions of the global discriminator and the local discriminator are jointly optimized to map the temporal feature representations of the source domain and the target domain to a shared feature space to generate the target temporal feature representation;

[0091] After constructing the dual discriminator structure, the domain adaptation process can be actually performed through the adversarial training mechanism. That is, a feature mapping network (which can be regarded as the generator in the generative adversarial network) is trained to transform the initial temporal feature representations of the source domain and the target domain into a shared feature space that the discriminator cannot distinguish.

[0092] Specifically, the feature mapping network receives initial temporal feature representations from the source or target domain, transforms them using internally learned parameters, and outputs the mapped target temporal feature representation. The training objective of this feature mapping network is to enable its output features to successfully "deceive" the global and local discriminators, preventing them from accurately determining the original domain (source or target domain) of the input features. The dual discriminators receive the target temporal feature representations output by the feature mapping network and attempt to distinguish whether these features originate from the source or target domain. The global discriminator focuses on determining the source of the overall distribution, while the local discriminator focuses on determining the source of dynamic patterns within each time window. The adversarial training process involves jointly optimizing a combined objective function that includes the feature mapping network loss function and the loss functions of the two discriminators to achieve adversarial training. Training can be performed alternately: first, fix the feature mapping network and update the discriminators to improve discrimination ability; then, fix the discriminators and update the feature mapping network to generate features that are more difficult to distinguish. Ideally, training will eventually reach an equilibrium point where the feature mapping network can map the initial feature representations of the source and target domains to a shared feature space. In this space, the global discriminator cannot effectively distinguish the overall distribution, and the local discriminator cannot effectively distinguish the dynamic patterns within the window, thus obtaining an aligned target temporal feature representation.

[0093] 208. Based on the target time series feature representation, calculate the first mutual information value between the source domain and the target domain in the entire time series period, and calculate the second mutual information value between the source domain and the target domain in each time segment;

[0094] After obtaining the aligned target temporal feature representation, to achieve more accurate and evidence-based transfer learning, it is necessary to quantify the correlation strength between the source and target domains at different time scales. Mutual information, as a powerful information-theoretic metric, can capture complex nonlinear dependencies between variables and reflects the true level of association better than simple linear correlation. This embodiment specifically requires two levels of mutual information calculation:

[0095] First, consider the entire aligned temporal feature sequence of the source and target domains. A first mutual information value is obtained by calculating the mutual information between these two complete sequences. This first mutual information value assesses the overall similarity or information overlap between the source and target domains over the entire temporal period, reflecting the fundamental fit for knowledge transfer from the source domain. A high first mutual information value typically indicates a strong macroscopic correlation between the two domains. Second, to capture the correlation between dynamic changes and local scenarios, for each time segment of the target domain, the mutual information between the aligned feature representation of that segment and its corresponding or most relevant aligned feature representation in the source domain is calculated, generating a second mutual information value for each target domain time segment. The second mutual information value represents the correlation strength between the target domain and historical data of the source domain under different local time periods or specific weather conditions. A high second mutual information value for some time segments indicates a strong correlation between the target domain and historical data of the source domain, while a low value for others indicates a weak correlation or pattern mismatch.

[0096] 209. The first mutual information value and the second mutual information value are fused to generate the multi-scale transfer weights for each time segment of the source domain, and the transfer confidence is calculated based on the distribution characteristics of the second mutual information value. The multi-scale transfer weights are used to adjust the contribution of the time segment to the prediction task of the target domain.

[0097] For each time segment in the target domain, the first and second mutual information values ​​need to be fused to determine a corresponding multi-scale transfer weight. The core purpose of this multi-scale transfer weight is to adjust the contribution of source domain knowledge related to that time segment to the target domain prediction task in subsequent steps. The higher the transfer weight, the more relevant and important the corresponding knowledge in the source domain should be. It should be noted that there can be various fusion strategies, such as weighted average, product, etc., which are not limited here, but the core idea of ​​fusion is to consider both the overall domain similarity and the local scene matching degree. For example, even if the global similarity is average and the first mutual information value is not high, if a target segment is highly related to a certain part of the source domain, that is, the second mutual information value is high, then the transfer weight corresponding to that segment should also be relatively high. Conversely, even if the global similarity is high, if a specific segment is not strongly related to the source domain, its transfer weight should be lowered.

[0098] In addition to calculating transfer weights, it is also necessary to evaluate the reliability of the transfer process or the transfer in specific scenarios. The calculation of transfer confidence is based on the distribution characteristics of the second mutual information values. For example, statistical indicators such as the mean, variance, kurtosis, or the proportion of high values ​​among all second mutual information values ​​can be analyzed. If the second mutual information values ​​are relatively high and concentrated in most time segments, it indicates that there is a stable and strong correlation between the source and target domains in various local scenarios, and the transfer confidence is high. Conversely, if the second mutual information values ​​are generally low or very scattered, it means that the applicability of source domain knowledge to the target domain is unstable or inapplicable in many scenarios, and the transfer confidence is low. Transfer confidence can be used to dynamically adjust training strategies (such as training priority and learning rate) or model selection fusion strategies in subsequent steps. For scenarios with high confidence, source domain knowledge can be utilized more actively, while for scenarios with low confidence, more reliance on target domain data or a more conservative strategy can be adopted.

[0099] 210. Initialize multiple sub-models based on the target temporal feature representation and multi-scale transfer weights, and optimize the parameters of the sub-models for different wind conditions.

[0100] Based on the target time-series feature representation obtained in step 207, clustering analysis (such as K-means) or preset rules are used to divide the historical wind power data of the target domain into different wind condition scenarios. An independent sub-model structure is created for each identified typical scenario, specifically a sub-model sharing a common underlying structure. Before formal training, these sub-models are initialized using the multi-scale transfer weights generated in step 209. Specific methods include, but are not limited to, using weights to guide the selection or weighting of pre-trained parameters from the source domain; adjusting the distribution of initial parameters of the sub-model using weights to make it more likely to reflect the characteristics of highly correlated scenarios in the source domain; and during the initial training phase, using weights to weight training samples from the source or target domain, so that samples highly correlated with the current sub-model scenario receive greater attention.

[0101] 211. Training resources are dynamically allocated based on transfer confidence. A pre-trained sub-model set containing multiple sub-models is generated by training through a shared backbone network and scene-specific output layer structure. The pre-trained sub-model set is used to support prediction tasks for different wind conditions.

[0102] In this embodiment, the prediction model can specifically adopt a structure of a shared backbone network and scene-specific output layers. The shared backbone network is responsible for processing the input target temporal feature representation and extracting deep temporal features and patterns common to all scenes. The scene-specific output layers are connected after the shared backbone network, and an independent output layer is set up for each wind condition scene. Each output layer is specifically responsible for making the final prediction of wind power for that scene based on the shared features.

[0103] During training, the transfer confidence calculated in step 209 can assess the overall reliability of source domain knowledge transfer across different scenarios. Therefore, it can guide resource allocation during training to achieve more efficient and robust learning. For scenarios with high transfer confidence, since source domain knowledge is more reliable in that scenario, more training epochs can be allocated, a higher learning rate can be used, or its weight in the loss function can be increased. For scenarios with low transfer confidence, a more conservative training strategy can be adopted, such as reducing the learning rate, relying more on target domain data, or introducing regularization to prevent unreliable source domain knowledge from negatively impacting the model.

[0104] The pre-trained sub-model set exists as a single model, but logically includes the capabilities of multiple sub-models optimized for different wind conditions. The pre-trained sub-model set not only effectively supports accurate prediction tasks for various wind conditions in the target domain, but its training process also considers the reliability of knowledge transfer, thus improving the robustness and efficiency of the overall prediction system.

[0105] 212. Obtain real-time wind power data of the target domain and extract the current time series features using a sliding time window. Based on the matching degree between the current time series features and the sub-model, combined with the transfer confidence, call the target sub-model to generate wind power prediction results for the target domain.

[0106] In this embodiment, step 212 is similar to step 106 in the previous embodiment, and will not be described again here.

[0107] In this embodiment, a bidirectional LSTM and attention mechanism are combined in the feature extraction stage. This not only effectively captures the long-term dependencies and short-term fluctuations in wind power data, but also highlights the impact of sudden wind speed changes or key meteorological events through attention weights, generating an information-rich and focused initial time-series feature representation. In the domain adaptation stage, a dual-discriminator structure with global and local components is used for adversarial training, overcoming the limitations of a single discriminator. This dual-discriminator structure can simultaneously align the overall distribution of features in the source and target domains with the dynamic pattern differences within local time windows, achieving a more thorough domain adaptation that better reflects the dynamic characteristics of wind power data and effectively mitigating the domain drift problem. In the knowledge transfer stage, multi-level mutual information is calculated, and multi-scale transfer weights and transfer confidence scores are generated based on this, providing clear guidance and reliability assessment for subsequent knowledge transfer. The training process uses multi-scale transfer weights for initialization and dynamically allocates training resources based on transfer confidence scores, making the model training process more efficient and robust. It intelligently focuses on knowledge with high transfer reliability, avoiding negative transfer risks, thereby significantly improving the accuracy, robustness, and scene adaptability of predictions.

[0108] The following provides a detailed description of the cross-domain adaptive wind power data analysis system provided in this application. Please refer to [link / reference]. Figure 3 , Figure 3 Another embodiment of the system for cross-domain adaptive wind power data analysis provided in this application, the system includes:

[0109] The acquisition unit 301 is used to acquire historical wind power data of the source domain and the target domain respectively. The historical wind power data includes long-term wind speed, wind direction, meteorological variables and power output time series.

[0110] Extraction unit 302 is used to extract dynamic features of historical wind power data using bidirectional LSTM and attention mechanism, and generate initial temporal feature representations of source and target domains;

[0111] Alignment unit 303 is used to dynamically align the initial temporal feature representation based on the sliding time window and cross-domain dual discriminator adversarial training method to obtain the target temporal feature representation, which contains the temporal feature representation of multiple time segments;

[0112] The calculation unit 304 is used to calculate the correlation between the source domain and the target domain in the whole time series and in each time segment based on the target time series feature representation and through multi-level mutual information, and to generate multi-scale transfer weights and transfer confidence based on the calculation results.

[0113] The transfer unit 305 is used to construct a pre-trained sub-model set based on the target temporal feature representation, multi-scale transfer weights and transfer confidence. The pre-trained sub-model set includes multiple sub-models optimized for different wind conditions. The transfer confidence is used to dynamically adjust the training priority of the sub-models.

[0114] Prediction unit 306 is used to acquire real-time wind power data of the target domain and extract the current time series features using a sliding time window. Based on the matching degree between the current time series features and the sub-model, combined with the transfer confidence, the target sub-model is called to generate wind power prediction results for the target domain.

[0115] In this embodiment, the functions of each unit are the same as described above. Figure 1 or Figure 2 The steps in the method embodiments shown correspond to those in the examples, and will not be repeated here.

[0116] This application also provides an apparatus for cross-domain adaptive wind power data analysis; please refer to [link to relevant documentation]. Figure 4 , Figure 4 One embodiment of the apparatus for cross-domain adaptive wind power data analysis provided in this application includes:

[0117] Processor 401, memory 402, input / output unit 403, bus 404;

[0118] The processor 401 is connected to the memory 402, the input / output unit 403, and the bus 404;

[0119] The memory 402 stores a program, and the processor 401 calls the program to execute any of the methods described above for cross-domain adaptive wind power data analysis.

[0120] This application also relates to a computer-readable storage medium on which a program is stored, which, when run on a computer, causes the computer to perform any of the methods described above for cross-domain adaptive wind power data analysis.

[0121] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0122] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0123] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0124] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0125] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for cross-domain adaptive wind power data analysis, characterized in that, The method includes: Historical wind power data for the source and target domains are acquired separately. The historical wind power data includes long-term wind speed, wind direction, meteorological variables, and power output time series. The dynamic features of the historical wind power data are extracted using a bidirectional LSTM and attention mechanism to generate initial temporal feature representations of the source domain and the target domain; The initial temporal feature representation is dynamically aligned based on a sliding time window and a cross-domain dual discriminator adversarial training method to obtain a target temporal feature representation, which contains temporal feature representations of multiple time segments. Based on the target time-series feature representation, the correlation between the source domain and the target domain in the whole time-series period and in each time segment is calculated through multi-level mutual information, and multi-scale transfer weights and transfer confidence are generated based on the calculation results. A pre-trained sub-model set is constructed based on the target temporal feature representation, the multi-scale transfer weights, and the transfer confidence. The pre-trained sub-model set includes multiple sub-models optimized for different wind conditions. The transfer confidence is used to dynamically adjust the training priority of the sub-models. The real-time wind power data of the target domain is obtained and the current time-series features are extracted using a sliding time window. Based on the matching degree between the current time-series features and the sub-model, combined with the migration confidence, the target sub-model is invoked to generate the wind power prediction result of the target domain.

2. The method according to claim 1, characterized in that, The step of extracting dynamic features from the historical wind power data using a bidirectional LSTM and attention mechanism to generate initial temporal feature representations for the source and target domains includes: A bidirectional LSTM is used to perform time-series modeling on the historical wind power data to extract the long-term dependency features and short-term fluctuation features of the historical wind power data. An attention mechanism is introduced to assign time step weights to the output of the bidirectional LSTM, and these time step weights are used to highlight the dynamic characteristics of sudden wind speed changes or meteorological events. Based on the output of the bidirectional LSTM and the time step weights, initial temporal feature representations of the source domain and the target domain are generated by weighted combination.

3. The method according to claim 1, characterized in that, The method based on sliding time windows and cross-domain dual discriminator adversarial training dynamically aligns the initial temporal feature representation to obtain the target temporal feature representation, including: The initial temporal feature representations of the source domain and the target domain are segmented using a sliding time window to generate temporal feature representations of multiple time segments; A cross-domain dual discriminator structure is constructed, which includes a global discriminator and a local discriminator. The global discriminator is used to optimize the consistency of the overall feature distribution between the source domain and the target domain, and the local discriminator is used to optimize the dynamic pattern differences within each time window. By conducting adversarial training based on generative adversarial networks, the loss functions of the global discriminator and the local discriminator are jointly optimized, and the temporal feature representations of the source domain and the target domain are mapped to a shared feature space to generate the target temporal feature representation.

4. The method according to claim 3, characterized in that, The method further includes: The length of the sliding time window is dynamically adjusted based on the wind speed fluctuation amplitude and meteorological event frequency in the historical wind power data.

5. The method according to claim 1, characterized in that, The step of calculating the correlation between the source domain and the target domain across the entire time series and at various time segments using mutual information based on the target temporal feature representation, and generating multi-scale transfer weights and transfer confidence based on the calculation results, includes: Based on the target time-series feature representation, calculate the first mutual information value between the source domain and the target domain over the entire time-series period, and calculate the second mutual information value between the source domain and the target domain in each time segment; The first mutual information value and the second mutual information value are fused to generate multi-scale transfer weights for each time segment of the source domain, and the transfer confidence is calculated based on the distribution characteristics of the second mutual information value. The multi-scale transfer weights are used to adjust the contribution of the time segment to the prediction task of the target domain.

6. The method according to any one of claims 1 to 5, characterized in that, The step of constructing a pre-trained sub-model set based on the target temporal feature representation, the multi-scale transfer weights, and the transfer confidence includes: Multiple sub-models are initialized based on the target temporal feature representation and the multi-scale transfer weights, and the parameters of the sub-models are optimized for different wind conditions. Training resources are dynamically allocated based on the transfer confidence. A pre-trained sub-model set containing multiple sub-models is generated by training through a shared backbone network and scene-specific output layer structure. The pre-trained sub-model set is used to support prediction tasks for different wind conditions.

7. A system for cross-domain adaptive wind power data analysis, characterized in that, The system includes: The acquisition unit is used to acquire historical wind power data of the source domain and the target domain respectively. The historical wind power data includes long-term wind speed, wind direction, meteorological variables and power output time series. The extraction unit is used to extract the dynamic features of the historical wind power data using bidirectional LSTM and attention mechanism, and generate the initial temporal feature representations of the source domain and the target domain. An alignment unit is used to dynamically align the initial temporal feature representation based on a sliding time window and a cross-domain dual discriminator adversarial training method to obtain a target temporal feature representation, wherein the target temporal feature representation contains temporal feature representations of multiple time segments; The calculation unit is used to calculate the correlation between the source domain and the target domain in the whole time series and in each time segment based on the target time series feature representation through multi-level mutual information, and generate multi-scale transfer weights and transfer confidence based on the calculation results. The transfer unit is used to construct a pre-trained sub-model set based on the target temporal feature representation, the multi-scale transfer weights, and the transfer confidence. The pre-trained sub-model set includes multiple sub-models optimized for different wind conditions. The transfer confidence is used to dynamically adjust the training priority of the sub-models. The prediction unit is used to acquire real-time wind power data of the target domain and extract the current time-series features using a sliding time window. Based on the matching degree between the current time-series features and the sub-model, combined with the migration confidence, the target sub-model is invoked to generate the wind power prediction result of the target domain.

8. The system according to claim 7, characterized in that, The extraction unit is specifically used for: A bidirectional LSTM is used to perform time-series modeling on the historical wind power data to extract the long-term dependency features and short-term fluctuation features of the historical wind power data. An attention mechanism is introduced to assign time step weights to the output of the bidirectional LSTM, and these time step weights are used to highlight the dynamic characteristics of sudden wind speed changes or meteorological events. Based on the output of the bidirectional LSTM and the time step weights, initial temporal feature representations of the source domain and the target domain are generated by weighted combination.

9. A device for cross-domain adaptive wind power data analysis, characterized in that, The device includes: Processor, memory, input / output units, and bus; The processor is connected to the memory, the input / output unit, and the bus; The memory stores a program, which the processor invokes to perform the method as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains a program that, when executed on a computer, performs the method as described in any one of claims 1 to 6.

Images