A wind speed prediction method and device based on a time sequence generative adversarial network and a medium

Abstract

The application discloses a wind speed prediction method and device based on a time sequence generative adversarial network and a medium, and relates to the technical field of wind speed prediction. The method comprises the following steps: collecting original wind speed time sequence data of a target wind field, and inputting the original wind speed time sequence data into a preset time sequence generative adversarial network to generate synthetic wind speed time sequence data; fusing the original wind speed time sequence data and the synthetic wind speed time sequence data according to a preset proportion and performing random shuffling to obtain an enhanced data set; training a wind speed prediction model by using the enhanced data set, and predicting to-be-predicted wind speed time sequence data based on the wind speed prediction model that has completed the training to obtain a wind speed prediction result. The application generates synthetic wind speed time sequence data that is the same as original data in distribution by using a time sequence generative adversarial network, makes up for small sample defects, and significantly improves the quality of the enhanced data set and the prediction ability of the model.

Description

Technical Field

[0001] This application relates to the field of wind speed prediction technology, and in particular to a wind speed prediction method, device and medium based on temporal generative adversarial networks. Background Technology

[0002] Wind speed forecasting is a crucial step in wind farm scheduling, turbine control, and grid power matching. Currently, wind speed forecasting mainly relies on historical observation data to train machine learning models. However, this approach has significant limitations in practical applications: Firstly, for newly built or remote wind farms, historical data is usually very limited, resulting in insufficient model training and poor generalization ability. Secondly, extreme wind conditions (such as typhoons and sudden gusts) occur infrequently and are scarce in training samples, making it difficult for models to effectively capture their nonlinear variation patterns, thus leading to low prediction accuracy in extreme scenarios. Furthermore, while traditional data augmentation methods (such as interpolation and resampling) can expand the sample size to some extent, they often fail to retain the dynamic characteristics of wind speed time series data, such as diurnal periodic fluctuations and seasonal trends. The generated data deviates from the distribution of real data, thus affecting the overall performance of the prediction model.

[0003] Therefore, how to utilize temporal generative adversarial networks for data augmentation and generate synthetic temporal data consistent with the distribution of real wind speed data to effectively expand the training samples, improve the sample coverage of extreme wind conditions, and thus enhance the generalization ability and accuracy of the prediction model has become an urgent problem to be solved. Summary of the Invention

[0004] This application provides a wind speed prediction method, device, and medium based on temporal generative adversarial networks (GANs) to solve the following technical problem: how to utilize temporal GANs for data augmentation and generate synthetic temporal data consistent with the distribution of real wind speed data to effectively expand training samples, improve sample coverage for extreme wind conditions, and thus enhance the generalization ability and accuracy of the prediction model.

[0005] In a first aspect, embodiments of this application provide a wind speed prediction method based on a time-series generative adversarial network. The method includes: collecting raw wind speed time-series data of a target wind field and inputting the raw wind speed time-series data into a preset time-series generative adversarial network to generate synthetic wind speed time-series data; fusing the raw wind speed time-series data and the synthetic wind speed time-series data according to a preset ratio and randomly shuffling them to obtain an augmented dataset; using the augmented dataset to train a wind speed prediction model, and based on the trained wind speed prediction model, predicting the wind speed time-series data to be predicted to obtain a wind speed prediction result.

[0006] In one implementation of this application, the method further includes: constructing a time-series generative adversarial network, specifically including: employing a time-adapted network structure and building an embedding network to map the original wind speed time-series data to a low-dimensional latent space; building a generator using a time-series grid structure matching the embedding network to receive random noise and learn the data distribution; building a discriminator to receive data and output probability results that distinguish between the original data and the synthetic data; building a recovery network using a fully connected layer structure to verify the reversibility of the synthetic features output by the generator; and associating and combining the embedding network, generator, discriminator, and recovery network to generate a time-series generative adversarial network.

[0007] In one implementation of this application, before inputting the original wind speed time series data into a preset time series generative adversarial network, the method further includes: identifying outliers in the original wind speed time series data to remove invalid and abruptly abnormal data; normalizing the removed original wind speed time series data; and arranging the normalized original wind speed time series data in timestamp order to generate standard wind speed time series data.

[0008] In one implementation of this application, the original wind speed time series data is input into a preset time series generative adversarial network (GAN) to generate synthetic wind speed time series data. Specifically, this includes: inputting standard wind speed time series data into an embedding network within the GAN to extract the temporal features of the standard wind speed time series data; inputting random noise into the generator and combining it with the temporal features to generate preliminary synthetic data; inputting the standard wind speed time series data and the preliminary synthetic data into a discriminator to output a probability result distinguishing the two types of data, and feeding the probability result back to the generator; mapping the preliminary synthetic data back to the original wind speed data space based on a recovery network to verify the reversibility of the temporal features; and adjusting the parameters of the embedding network, generator, discriminator, and recovery network to generate synthetic wind speed time series data with the same distribution as the standard wind speed time series data.

[0009] In one implementation of this application, random noise is input into the generator and combined with time-series features to generate preliminary synthetic data. Specifically, this includes: adjusting the data distribution features of the random noise and performing dimensional matching between the data distribution features of the random noise and the time-series features to obtain the generator input data; and performing feature fusion and reconstruction on the generator input data based on a preset wind speed time-series data distribution pattern to generate preliminary synthetic data.

[0010] In one implementation of this application, the original wind speed time series data and the synthetic wind speed time series data are fused according to a preset ratio and then randomly shuffled to obtain an enhanced dataset. Specifically, this includes: determining the fusion ratio of the original wind speed time series data and the synthetic wind speed time series data based on the sample size of the original data, the data distribution characteristics, and the fidelity of the synthetic data; mixing the original wind speed time series data and the synthetic wind speed time series data based on the fusion ratio to generate an initial mixed dataset; and randomly shuffling the initial mixed dataset to obtain an enhanced dataset without order bias.

[0011] In one implementation of this application, the wind speed prediction model is trained using an augmented dataset. Specifically, this includes: dividing the augmented dataset into a training set and a validation set, and inputting historical time period wind speed data from the training set into the wind speed prediction model for training; and optimizing the training process of the wind speed prediction model using a prediction error function until the prediction error of the validation set converges, thereby obtaining a wind speed prediction model that has been trained.

[0012] In one implementation of this application, based on a trained wind speed prediction model, wind speed time series data to be predicted is predicted to obtain a wind speed prediction result. Specifically, this includes: inputting the wind speed time series data to be predicted into the trained wind speed prediction model to obtain the corresponding wind speed prediction output; performing inverse normalization processing on the wind speed prediction output to obtain the actual wind speed value; and when the actual wind speed value exceeds a preset threshold, performing smoothing correction on the actual wind speed value to generate the final wind speed prediction result.

[0013] Secondly, embodiments of this application also provide a wind speed prediction device based on a time-series generative adversarial network. The device includes: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor, which, when executed, enable the at least one processor to: collect raw wind speed time-series data of a target wind field and input the raw wind speed time-series data into a preset time-series generative adversarial network to generate synthetic wind speed time-series data; fuse the raw wind speed time-series data and the synthetic wind speed time-series data according to a preset ratio and randomly shuffle them to obtain an augmented dataset; train a wind speed prediction model using the augmented dataset, and based on the trained wind speed prediction model, predict the wind speed time-series data to be predicted to obtain a wind speed prediction result.

[0014] Thirdly, embodiments of this application also provide a non-volatile computer storage medium for wind speed prediction based on a time-series generative adversarial network, storing computer-executable instructions. The computer-executable instructions are configured to: collect raw wind speed time-series data of a target wind field, input the raw wind speed time-series data into a preset time-series generative adversarial network to generate synthetic wind speed time-series data; fuse the raw wind speed time-series data and the synthetic wind speed time-series data according to a preset ratio and randomly shuffle them to obtain an augmented dataset; use the augmented dataset to train a wind speed prediction model, and based on the trained wind speed prediction model, predict the wind speed time-series data to be predicted to obtain a wind speed prediction result.

[0015] The wind speed prediction method, device, and medium based on temporal generative adversarial networks provided in this application have the following beneficial effects: High-quality synthetic temporal data with the same distribution as the original wind speed data, generated by the temporal generative adversarial network, significantly increases the diversity and coverage of training samples, effectively mitigating overfitting caused by insufficient samples; the generated synthetic data highly overlaps with the original data in distribution and maintains consistency and authenticity in temporal features, thereby ensuring the information quality and training effectiveness of the augmented dataset and avoiding model performance degradation due to data distortion; by controlling noise input, synthetic sequences with extreme wind speed characteristics are generated in a targeted manner, effectively increasing the proportion of extreme samples in the augmented dataset, enabling the prediction model to better learn wind speed evolution patterns under extreme wind conditions. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart of a wind speed prediction method based on a temporal generative adversarial network provided for an embodiment of this application; Figure 2 This is a schematic diagram of the internal structure of a wind speed prediction device based on a temporal generative adversarial network, provided as an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] This application provides a wind speed prediction method, device, and medium based on temporal generative adversarial networks (GANs) to solve the following technical problem: how to utilize temporal GANs for data augmentation and generate synthetic temporal data consistent with the distribution of real wind speed data to effectively expand training samples, improve sample coverage for extreme wind conditions, and thus enhance the generalization ability and accuracy of the prediction model.

[0019] The technical solutions proposed in the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0020] Figure 1 A flowchart illustrating a wind speed prediction method based on a temporal generative adversarial network, provided as an embodiment of this application. Figure 1 As shown in the figure, the wind speed prediction method based on temporal generative adversarial networks provided in this application embodiment specifically includes the following steps: Step 10: Collect raw wind speed time series data of the target wind field and input the raw wind speed time series data into a preset time series generative adversarial network to generate synthetic wind speed time series data.

[0021] As an optional embodiment, the raw wind speed time series data of the target wind field is collected, and the raw wind speed time series data is input into a preset time series generative adversarial network to generate synthetic wind speed time series data. Specifically, it may include: Step 101: Constructing a time series generative adversarial network.

[0022] As an optional embodiment, constructing a time-series generative adversarial network may specifically include: Step 1011: adopting a time-adaptive network structure and building an embedded network to map the original wind speed time-series data to a low-dimensional latent space.

[0023] In this step, to accurately capture the inherent dynamic patterns of wind speed time-series data, a network structure highly adapted to the characteristics of time-series data is adopted. This embedded network is specifically designed for the temporal attributes of wind speed data, deeply matching the essential characteristics of wind speed changes over time, such as continuity and periodicity. This avoids feature loss due to mismatch between network structure and data type. The high-dimensional original wind speed time-series data is transformed into feature representations in a low-dimensional latent space through a specific mapping mechanism. During the dimensionality compression process, not only can redundant information and noise interference in the data be eliminated, but the core dynamic change patterns in the wind speed time series, such as key fluctuation patterns and trend features, can also be accurately preserved. This provides a high-fidelity feature foundation for the subsequent generation of synthetic data, ensuring that the generated synthetic data can maintain a high degree of consistency with the original data in terms of temporal characteristics, thus laying a core foundation for the effectiveness of the entire data augmentation chain.

[0024] Step 1012: Build a generator using a temporal grid structure that matches the embedded network to receive random noise and learn the data distribution.

[0025] In this step, to ensure the temporal consistency and distribution authenticity of the synthesized wind speed time series data, the generator adopts a temporal adaptation network structure that matches the embedded network. The two complement each other in network design logic, ensuring the smoothness of feature transfer and transformation. The generator takes random noise as input and, relying on the core architecture of the adapted time series data, can deeply mine the core features of the original wind speed data extracted by the embedded network and accurately learn the temporal distribution pattern of the original data. During the learning process, the generator can not only capture the dynamic change pattern of wind speed data, but also simulate its inherent correlation characteristics. Through reasonable transformation and feature mapping of random noise, it generates synthetic time series data that conforms to the distribution characteristics of the original data, providing high-quality sample support for subsequent data augmentation. At the same time, it complements the feature extraction capabilities of the embedded network, ensuring the coherence and reliability of the entire data generation process.

[0026] Step 1013: Build a discriminator to receive data and output the probability results that distinguish between the original data and the synthetic data.

[0027] In this step, to effectively distinguish between raw and synthetic wind speed time-series data, a discriminator adapted to the characteristics of the time-series data is built. Its network structure is collaboratively adapted with the embedded network and the generator to ensure accurate capture and judgment of time-series features. The discriminator analyzes the input data, which includes both the raw wind speed time-series feature data processed by the embedded network and the synthetic wind speed time-series data generated by the generator. By deeply mining the time-series distribution patterns and dynamic change characteristics of the input data, the discriminator performs discriminative analysis on the data source and ultimately outputs a result within a specific probability range. This result directly reflects the probability that the input data belongs to the raw or synthetic data. Through this discrimination process, the discriminator can supervise and guide the generator's training, prompting the generator to continuously optimize its generation strategy and improve the consistency of the distribution between synthetic and raw data. This provides an important guarantee for the generation of high-fidelity synthetic data and also provides the core discrimination basis for the entire adversarial training.

[0028] Step 1014: A recovery network is built using a fully connected layer structure to verify the reversibility of the synthesized features output by the generator.

[0029] In this step, to verify the correspondence and reversibility between the synthetic features output by the generator and the features of the original wind speed time series data, the recovery network is built with a fully connected layer structure. Its core function focuses on the reverse mapping and restoration verification of the feature space. This fully connected layer structure can accurately receive the synthetic features output by the generator. Through a specific mapping mechanism, the synthetic features in the low-dimensional latent space are reverse-transformed to the spatial dimension of the original wind speed data. In this restoration process, the recovery network not only completes the dimensional reconstruction of the features, but also verifies the effectiveness of the embedded network feature extraction and the generator feature generation by comparing the feature consistency between the restoration result and the original data. If the synthetic features have good reversibility, it indicates that the distribution pattern learned by the generator is highly consistent with the core features of the original data, and the generated synthetic features are real and reliable.

[0030] Step 1015: Combine the embedding network, generator, discriminator and recovery network to generate a time series generative adversarial network.

[0031] In this step, the embedding network, as the core of feature extraction, first processes the original wind speed time-series data, mapping it to a low-dimensional latent space to provide an accurate temporal feature foundation for subsequent modules. The generator receives random noise, learns the distribution of the original data based on the temporal structure adapted to the embedding network, and generates synthetic features. Its output is simultaneously connected to the discriminator and the recovery network, forming a bidirectional verification link. The discriminator receives the original features output by the embedding network and the synthetic features output by the generator, and outputs probability results through discriminant analysis, providing supervision signals for the generator's optimization. The recovery network takes the synthetic features from the generator and reverses them back to the original data space through a fully connected layer structure, verifying the reversibility and authenticity of the features, and feeding the restoration results back to the model training process. The network modules support each other functionally and progress layer by layer in data flow. The feature output of the embedding network provides the learning basis for the generator, the supervision signal of the discriminator drives the generator to iteratively optimize, and the verification results of the recovery network ensure the reliability of feature extraction and generation. Through this interconnected combination, a time-series generative adversarial network capable of generating high-fidelity time-series data is finally formed, meeting the core requirement of wind speed data enhancement.

[0032] Step 102: Identify outliers in the original wind speed time series data to remove invalid and abruptly abnormal data.

[0033] In this step, outlier identification, taking into account the characteristics of wind speed data, will focus on the physical rationality and statistical regularity of the data: From a physical perspective, invalid data that does not have actual wind speed representation significance will be screened and removed to ensure that the data conforms to the natural attributes of wind speed; from the perspective of data change patterns, by judging whether the data deviates from the normal fluctuation range, abrupt abnormal data caused by factors such as sensor failure and extreme environmental interference will be identified and removed. Such data will disrupt the continuity and regularity of wind speed time series, and if retained, it will directly affect the subsequent feature extraction and model training effects; through a rigorous outlier identification process, the original dataset will be effectively purified, ensuring the accuracy and reliability of subsequent data processing, model training, and synthetic data generation, thus laying a solid data foundation for the accurate implementation of the entire wind speed prediction scheme.

[0034] Step 103: Normalize the original wind speed time series data after removing the data, and arrange the normalized original wind speed time series data in the order of timestamps to generate standard wind speed time series data.

[0035] In this step, after removing outliers from the original wind speed time series data, the data needs to be further normalized and ordered to generate standard wind speed time series data that meets the requirements of model training. The normalization transformation is carried out on the effective wind speed data after outlier removal. Through specific transformation rules, the data is mapped to a unified numerical range. The core purpose is to eliminate the interference caused by different data scales, so that the data has comparability and consistency, creating conditions for efficient training of the subsequent network model and avoiding model learning bias or slow convergence due to differences in data scale. After the normalization process is completed, the normalized wind speed data is ordered according to the natural order of time evolution based on the timestamp information of the data, ensuring the temporal continuity and logical integrity of the data, and finally forming standard wind speed time series data with a standardized structure, clear time sequence, and uniform scale.

[0036] Step 104: Input the standard wind speed time series data into the embedded network in the time series generative adversarial network to extract the time series features of the standard wind speed time series data.

[0037] In this step, after constructing the standard wind speed time series data, it is input into the embedding network in the time series generative adversarial network. The embedding network, relying on a dedicated structure adapted to the time series data, can deeply match the temporal evolution characteristics of the standard wind speed time series data and accurately capture the core dynamic patterns contained in the data. During the extraction process, the embedding network penetrates the surface numerical values ​​of the data, mines and extracts the key temporal features of wind speed changes over time, including the inherent periodic fluctuation patterns, long-term trends, and short-term dynamic changes. Through such feature extraction, the standard wind speed time series data is transformed into feature vectors that can accurately represent its essential attributes, providing solid feature support for the subsequent generator to learn the data distribution and generate high-fidelity synthetic data.

[0038] Step 105: Input random noise into the generator and combine it with temporal features to generate preliminary synthetic data.

[0039] As an optional embodiment, random noise is input into the generator and combined with temporal features to generate preliminary synthetic data. Specifically, this may include: Step 1051: Adjusting the data distribution features of the random noise and performing dimensional matching between the data distribution features of the random noise and the temporal features to obtain the generator input data.

[0040] In this step, combining the inherent characteristics of wind speed time series data with the learning requirements of the generator, the distribution pattern of random noise is reasonably adjusted to adapt to the subsequent fusion logic of time series features. On this basis, the dimensional matching processing of the adjusted random noise data distribution features and the time series features extracted by the embedded network is performed. By unifying the dimensional specifications of the two, the fusion obstacles caused by the difference in data dimensionality are eliminated, ensuring that the two types of features can be efficiently and collaboratively input into the generator. The combined data after distribution adjustment and dimensional matching is the input data adapted to the generator's training and synthetic data generation requirements, laying the foundation for the generator to accurately learn the data distribution and generate high-fidelity time series synthetic data.

[0041] Step 1052: Based on the preset wind speed time series data distribution pattern, perform feature fusion and reconstruction on the generator input data to generate preliminary synthetic data.

[0042] In this step, after receiving the input data that has undergone distribution adjustment and dimensionality matching, the generator, leveraging its network structure adapted to time-series data, deeply interprets the distribution characteristics of random noise in the input data and the core time-series features extracted by the embedded network. Following the pre-defined inherent distribution logic of the wind speed data, it organically fuses these two types of features, preserving key information such as fluctuation patterns and trends in the time-series features while also imbuing the synthesized data with reasonable diversity through the characteristics of random noise. Based on feature fusion, the generator transforms the fused features into preliminary synthesized data with temporal continuity and logical consistency through a specific reconstruction mechanism. This process strictly follows the natural distribution patterns of wind speed data, ensuring that the preliminary synthesized data maintains a high degree of consistency with real wind speed data in terms of temporal structure and trends, providing high-quality basic samples for subsequent data optimization and the construction of augmented datasets.

[0043] Step 106: Input the standard wind speed time series data and the preliminary synthesized data into the discriminator to output the probability results that distinguish the two types of data, and feed the probability results back to the generator.

[0044] In this step, after generating the preliminary synthetic data, the standard wind speed time series data and the preliminary synthetic data are input into the discriminator. The discriminator will deeply mine the core information such as the time series distribution pattern and dynamic change characteristics of the input data. Based on these key criteria, it will judge the source of the data and finally output a probability result that can distinguish the categories to which the two types of data belong. This result intuitively reflects the similarity between the preliminary synthetic data and the standard wind speed time series data. Subsequently, this probability result is fed back to the generator in real time, providing the generator with a clear direction for optimization. If the probability result shows that there is a difference between the preliminary synthetic data and the standard data, the generator will adjust its feature fusion and reconstruction strategy according to the feedback, continuously optimize the generation logic of the synthetic data, and gradually improve the distribution consistency and time series authenticity of the synthetic data and the standard wind speed time series data.

[0045] Step 107: Based on the recovery network, the preliminary synthetic data is mapped back to the original wind speed data space to verify the reversibility of the time series features.

[0046] In this step, the recovery network reverse-transforms the synthetic features in the low-dimensional latent space to the spatial dimension of the original wind speed data according to the preset mapping rules. In this mapping process, the recovery network not only completes the reconstruction of the data dimension, but more importantly, it verifies whether the temporal features extracted by the embedded network and the synthetic features generated by the generator have good reversibility by comparing the feature fit between the mapping result and the original standard wind speed time series data. If the mapped result can accurately restore the core temporal pattern of the original wind speed data, it means that the generated preliminary synthetic data truly captures the essential features of the original data, and the reversibility is effectively verified. Otherwise, it provides feedback for further optimization of the model, ensuring that the subsequently generated synthetic data is highly consistent with the original data in terms of temporal characteristics.

[0047] Step 108: Adjust the parameters of the embedding network, generator, discriminator, and recovery network to generate synthetic wind speed time series data with the same distribution as the standard wind speed time series data.

[0048] In this step, after generating and verifying the reversibility of the initial synthetic data, the core parameters of the embedding network, generator, discriminator, and recovery network are coordinated and adjusted based on the probability feedback results output by the discriminator and the reversibility verification of the recovery network. For the embedding network, the accuracy of its feature extraction is optimized to ensure that it can capture the core dynamic patterns of the standard wind speed time series data more efficiently. For the generator, the weight logic of feature fusion and reconstruction is adjusted based on the feedback signal to improve the fit of the synthetic data to the distribution of the original data. For the discriminator, its sensitivity to the difference in features between the two types of data is optimized so that it can provide supervision signals more accurately. For the recovery network, its mapping mechanism is adjusted to enhance the reliability of the time series feature reversibility verification. Through dynamic adaptation and iterative optimization of the parameters of each network module, the entire time series generative adversarial network is driven to form a closed-loop optimization, ultimately generating synthetic wind speed time series data that is highly consistent with the standard wind speed time series data in terms of distribution patterns and time series characteristics.

[0049] Step 20: Merge the original wind speed time series data and the synthetic wind speed time series data according to a preset ratio and then randomly shuffle them to obtain an enhanced dataset.

[0050] As an optional embodiment, the original wind speed time series data and the synthetic wind speed time series data are fused according to a preset ratio and randomly shuffled to obtain an enhanced dataset. Specifically, it may include: Step 201: Based on the sample size of the original data, the data distribution characteristics and the fidelity of the synthetic data, determine the fusion ratio of the original wind speed time series data and the synthetic wind speed time series data.

[0051] In this step, when determining the fusion ratio of the original wind speed time-series data and the synthetic wind speed time-series data, three core factors will be comprehensively considered: the sample size of the original data, the data distribution characteristics, and the fidelity of the synthetic data. The sample size of the original data determines the support strength of the basic data. If the number of original samples is limited, the fusion ratio needs to be appropriately tilted towards the synthetic data to fully leverage the supplementary role of data augmentation. The data distribution characteristics focus on the completeness of the temporal patterns of the original data. If there are distribution gaps in certain temporal dimensions of the original data, the fusion ratio will be adjusted to allow the synthetic data to fill these gaps, making the overall dataset more balanced. The fidelity of the synthetic data is a key constraint. Only when the temporal characteristics and distribution patterns of the synthetic data are highly consistent with those of the original data will its fusion ratio be increased accordingly to avoid the reliability of the overall dataset being affected by low-quality synthetic data. Through comprehensive evaluation and dynamic balancing of these three factors, a scientifically reasonable fusion ratio is finally determined, providing an augmented dataset with both scale and quality for subsequent prediction model training.

[0052] Step 202: Based on the fusion ratio, the original wind speed time series data and the synthetic wind speed time series data are mixed to generate an initial mixed dataset.

[0053] In this step, after determining the fusion ratio of the original wind speed time-series data and the synthetic wind speed time-series data, the two types of data are mixed according to the preset ratio to generate an initial mixed dataset. During the mixing process, the determined ratio rules are strictly followed to ensure that the proportion of the original data and the synthetic data in the dataset meets the expectations. This fully utilizes the supplementary role of the synthetic data to the original data, making up for possible insufficient sample size or uneven distribution, while preserving the core value and true temporal characteristics of the original data. At the same time, attention is paid to maintaining the temporal integrity of each type of data during the mixing operation to avoid the disruption of the temporal logic of the data due to mixing. This ensures that the initial mixed dataset can balance sample size and data quality, providing a well-structured and balanced basic data support for subsequent random shuffling and prediction model training.

[0054] Step 203: Randomize the initial mixed dataset to obtain an augmented dataset without order bias.

[0055] In this step, after generating the initial mixed dataset, it needs to be randomly shuffled. The core purpose is to eliminate any inherent order imprints in the data, preventing the subsequent prediction model from mistakenly learning order patterns unrelated to the data's inherent features during training, thus causing order bias. The shuffling process breaks the original arrangement logic of the data in the initial mixed dataset, redistributing the position of each data point according to random principles. This ensures that the original and synthetic wind speed time series data are evenly distributed in the new dataset without a fixed arrangement pattern. After random shuffling, the final augmented dataset no longer carries any order-related interference information. The distribution of the data is determined solely by its own time series characteristics and distribution patterns, providing an unbiased, high-quality data foundation for fair training of the subsequent prediction model. This ensures that the model can purely learn the core patterns of the wind speed data rather than irrelevant order features.

[0056] Step 30: Use the augmented dataset to train the wind speed prediction model, and based on the trained wind speed prediction model, make predictions on the time series data of the wind speed to be predicted to obtain the wind speed prediction results.

[0057] As an optional embodiment, the wind speed prediction model is trained using an augmented dataset, and the wind speed prediction model is trained to predict the time series data of the wind speed to be predicted in order to obtain the wind speed prediction result. Specifically, it may include: Step 301: Divide the augmented dataset into a training set and a validation set, and input the historical time period wind speed data in the training set into the wind speed prediction model for training.

[0058] In this step, after obtaining the unbiased augmented dataset, it is split into a training set and a validation set according to a reasonable partitioning rule. The training set serves as the core data foundation for model learning, containing rich and comprehensive wind speed time-series information. During training, wind speed data corresponding to historical time periods in the training set are selected as input. This data covers the characteristics and temporal patterns of wind speed changes under different scenarios. This type of historical wind speed data is input into a preset wind speed prediction model. The model will deeply learn the inherent logic, fluctuation patterns, and correlation characteristics of wind speed evolution over time based on the input data, gradually optimizing its own parameters to adapt to the data characteristics, and continuously improving its ability to capture and fit the temporal patterns of wind speed, laying a solid model foundation for subsequent accurate wind speed prediction.

[0059] Step 302: Optimize the training process of the wind speed prediction model using the prediction error function until the prediction error of the validation set converges, so as to obtain the wind speed prediction model that has been trained.

[0060] In this step, a prediction error function is introduced as the core of optimization during the training of the wind speed prediction model. This function accurately quantifies the deviation between the model's prediction results and the actual data in the validation set, providing a clear basis for adjusting model parameters. During training, the model continuously receives learning signals from the training set and monitors the prediction performance of the validation set in real time using the prediction error function. Based on the error feedback, it dynamically adjusts its internal parameters. The optimization process always revolves around reducing prediction errors and improving the model's generalization ability. Through continuous iterative adjustments, the model's fitting effect on the wind speed time series patterns is gradually optimized. When the prediction error of the validation set tends to stabilize and reaches a convergence state, it indicates that the model has fully learned the core features and changing patterns of the wind speed data. At this point, the training process is terminated, and a stable and reliable wind speed prediction model is finally obtained.

[0061] Step 303: Input the time series data of the wind speed to be predicted into the wind speed prediction model that has been trained to obtain the corresponding wind speed prediction output.

[0062] In this step, after obtaining a well-trained and stable wind speed prediction model, the time series data of the wind speed to be predicted is input into the trained wind speed prediction model. The model will rely on the wind speed time series variation patterns, distribution characteristics, and correlation logic learned during the training process to perform in-depth analysis and feature calculation on the input data to be predicted. Through the capture of time series features, pattern matching, and result inference within the model, the final output is the wind speed prediction result corresponding to the time period to be predicted. This result can accurately reflect the changing trend and possible values ​​of wind speed during the time period to be predicted, providing reliable prediction support for wind farm scheduling, safe operation, and other related decisions.

[0063] Step 304: Perform inverse normalization on the wind speed prediction output to obtain the actual wind speed value.

[0064] In this step, after obtaining the prediction results output by the wind speed prediction model, inverse normalization processing needs to be carried out. Through the inverse mapping mechanism, the prediction output that is in a unified numerical range is restored to the inherent magnitude range of the original wind speed data. After inverse normalization processing, the originally standardized prediction data is transformed into actual wind speed values ​​that can directly represent the wind speed conditions of the real wind field, so that the prediction results have clear physical meaning and provide intuitive and usable data support for subsequent wind speed threshold judgment, result correction and actual wind field operation decisions.

[0065] Step 305: When the actual wind speed value exceeds the preset threshold, the actual wind speed value is smoothed and corrected to generate the final wind speed prediction result.

[0066] In this step, after obtaining the actual wind speed value through inverse normalization, it needs to be compared with a preset wind speed threshold. If the actual wind speed value exceeds the preset threshold, it indicates that there may be abrupt fluctuations related to extreme wind conditions in the prediction result. If such fluctuations are directly used as the output result, they may affect the rationality of subsequent decisions such as wind farm scheduling and safe operation. Therefore, it is necessary to perform a smoothing correction on such actual wind speed values ​​that exceed the threshold. The correction process will follow the natural variation law of extreme wind conditions, weaken the abrupt characteristics of the predicted value, and make the corrected result fit the continuous evolution logic of extreme wind speed in the actual wind farm. Through this smoothing correction operation, a final wind speed prediction result that is accurate, stable, and meets the needs of actual applications is finally generated.

[0067] The above are embodiments of the method proposed in this application. Based on the same inventive concept, embodiments of this application also provide a wind speed prediction device based on a temporal generative adversarial network, the structure of which is as follows: Figure 2 As shown.

[0068] Figure 2 This is a schematic diagram of the internal structure of a wind speed prediction device based on a temporal generative adversarial network, provided as an embodiment of this application. Figure 2 As shown, the device includes: At least one processor 201; And a memory 202 that is communicatively connected to at least one processor; The memory 202 stores instructions executable by at least one processor. These instructions are executed by at least one processor 201 to enable the at least one processor 201 to: collect raw wind speed time-series data of the target wind field and input the raw wind speed time-series data into a preset time-series generative adversarial network to generate synthetic wind speed time-series data; fuse the raw wind speed time-series data and the synthetic wind speed time-series data according to a preset ratio and randomly shuffle them to obtain an augmented dataset; use the augmented dataset to train a wind speed prediction model, and based on the trained wind speed prediction model, predict the wind speed time-series data to be predicted to obtain a wind speed prediction result.

[0069] Some embodiments of this application provide corresponding to Figure 1 A non-volatile computer storage medium for wind speed prediction based on a time-series generative adversarial network is disclosed. The computer-executable instructions are configured to: collect raw wind speed time-series data of a target wind field and input the raw wind speed time-series data into a preset time-series generative adversarial network to generate synthetic wind speed time-series data; fuse the raw wind speed time-series data and the synthetic wind speed time-series data according to a preset ratio and randomly shuffle them to obtain an augmented dataset; use the augmented dataset to train a wind speed prediction model, and based on the trained wind speed prediction model, predict the wind speed time-series data to be predicted to obtain the wind speed prediction result.

[0070] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for IoT devices and media are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0071] The systems, media, and methods provided in this application are one-to-one correspondences. Therefore, the systems and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the systems and media will not be repeated here.

[0072] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0073] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0074] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0075] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0076] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0077] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0078] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0079] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0080] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A wind speed prediction method based on temporal generative adversarial networks, characterized in that, The method includes: Collect raw wind speed time-series data of the target wind field and input the raw wind speed time-series data into a preset time series generative adversarial network to generate synthetic wind speed time-series data; The original wind speed time series data and the synthesized wind speed time series data are fused according to a preset ratio and then randomly shuffled to obtain an enhanced dataset; The wind speed prediction model is trained using the augmented dataset, and based on the trained wind speed prediction model, the wind speed time series data to be predicted is predicted to obtain the wind speed prediction result.

2. The wind speed prediction method based on temporal generative adversarial networks according to claim 1, characterized in that, The method further includes: Constructing the time-series generative adversarial network specifically includes: A time-adaptive network structure is adopted, and an embedded network is built to map the original wind speed time-series data to a low-dimensional potential space. A generator is built using a temporal grid structure that matches the embedded network to receive random noise and learn the data distribution; Build a discriminator to receive data and output the probability results that distinguish the original data from the synthetic data; A fully connected layer structure was used to build a recovery network to verify the reversibility of the synthesized features output by the generator; The embedded network, the generator, the discriminator, and the recovery network are combined and linked to generate the time series generative adversarial network.

3. The wind speed prediction method based on a temporal generative adversarial network according to claim 2, characterized in that, Before inputting the raw wind speed time-series data into a preset time-series generative adversarial network, the method further includes: Outlier identification is performed on the original wind speed time series data to remove invalid and abruptly abnormal data. The original wind speed time series data after removal is normalized and then arranged in timestamp order to generate standard wind speed time series data.

4. The wind speed prediction method based on a temporal generative adversarial network according to claim 3, characterized in that, The raw wind speed time-series data is input into a preset time-series generative adversarial network to generate synthetic wind speed time-series data, specifically including: The standard wind speed time series data is input into the embedded network in the time series generative adversarial network to extract the time series features of the standard wind speed time series data; Random noise is input into the generator and combined with the time-series features to generate preliminary synthetic data; The standard wind speed time series data and the preliminary synthesized data are input into the discriminator to output a probability result that distinguishes the two types of data, and the probability result is fed back to the generator; Based on the recovery network, the preliminary synthesized data is mapped back to the original wind speed data space to verify the reversibility of the temporal features; The parameters of the embedded network, the generator, the discriminator, and the recovery network are adjusted to generate the synthetic wind speed time series data that is distributed in the same way as the standard wind speed time series data.

5. The wind speed prediction method based on a temporal generative adversarial network according to claim 4, characterized in that, Random noise is input into the generator and combined with the time-series features to generate preliminary synthetic data, specifically including: Adjust the data distribution characteristics of the random noise, and perform dimensional matching between the data distribution characteristics of the random noise and the temporal characteristics to obtain the generator input data; Based on the preset wind speed time series data distribution pattern, the input data of the generator is subjected to feature fusion and reconstruction to generate preliminary synthetic data.

6. The wind speed prediction method based on temporal generative adversarial networks according to claim 1, characterized in that, The original wind speed time-series data and the synthesized wind speed time-series data are fused according to a preset ratio and then randomly shuffled to obtain an enhanced dataset, specifically including: Based on the sample size of the original data, the data distribution characteristics, and the fidelity of the synthetic data, the fusion ratio of the original wind speed time series data and the synthetic wind speed time series data is determined. Based on the fusion ratio, the original wind speed time series data and the synthetic wind speed time series data are mixed to generate an initial mixed dataset; The initial mixed dataset is randomly shuffled to obtain an enhanced dataset without order bias.

7. The wind speed prediction method based on temporal generative adversarial networks according to claim 1, characterized in that, The wind speed prediction model is trained using the augmented dataset, specifically including: The augmented dataset is divided into a training set and a validation set, and the historical wind speed data for different time periods in the training set is input into the wind speed prediction model for training. The training process of the wind speed prediction model is optimized using a prediction error function until the prediction error of the validation set converges, thereby obtaining a wind speed prediction model that has been successfully trained.

8. The wind speed prediction method based on a temporal generative adversarial network according to claim 1, characterized in that, Based on the trained wind speed prediction model, the wind speed time series data to be predicted is used to predict the wind speed to obtain the wind speed prediction result, specifically including: The wind speed time series data to be predicted is input into the wind speed prediction model that has been trained to obtain the corresponding wind speed prediction output. The wind speed prediction output is inversely normalized to obtain the actual wind speed value; When the actual wind speed value exceeds a preset threshold, the actual wind speed value is smoothed and corrected to generate the final wind speed prediction result.

9. A wind speed prediction device based on a temporal generative adversarial network, characterized in that, The device includes: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform a method as described in any one of claims 1-8.

10. A non-volatile computer storage medium for wind speed prediction based on a temporal generative adversarial network, storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, they implement a method as described in any one of claims 1-8.

Images