Short-term wind speed prediction method based on transfer learning and related equipment

Abstract

The invention belongs to the technical field of short-term wind speed prediction, and discloses a short-term wind speed prediction method based on transfer learning and related equipment. The short-term wind speed prediction method based on transfer learning comprises the following steps: acquiring an original wind speed sequence and related meteorological characteristic data, and inputting the original wind speed sequence and the related meteorological characteristic data into a constructed short-term wind speed prediction model based on transfer learning to obtain a short-term prediction wind speed sequence, the short-term wind speed prediction model based on transfer learning is constructed through CEEMDAN decomposition, a Bi-GRU-ED model and transfer learning; according to the method, noise interference in an original wind speed sequence can be self-adaptively and thoroughly stripped through CEEMDAN decomposition, an intrinsic mode component is accurately reconstructed, and noise residues and reconstruction distortion are remarkably reduced.

Description

Technical Field

[0001] This invention relates to the field of short-term wind speed forecasting technology, specifically to a short-term wind speed forecasting method and related equipment based on transfer learning. Background Technology

[0002] As the global energy system transitions towards sustainability, wind energy has become a core pillar of this transition. Wind energy is clean, abundant, and increasingly cost-effective; however, its efficient utilization heavily relies on accurate wind speed forecasting. Wind speed exhibits chaotic and non-stationary characteristics, and inaccurate forecasts can lead to energy imbalances, grid instability, and increased operating costs. Therefore, intelligent forecasting models capable of capturing complex wind speed dynamics under varying conditions are needed.

[0003] Currently, wind speed forecasting methods can be categorized into four types based on the forecast time domain: ultra-short-term (minutes to hours), short-term (hours to days), medium-term (days to weeks), and long-term (weeks to years); and broadly by methodology, they can be classified into physical methods, statistical methods, artificial intelligence methods, and hybrid methods.

[0004] Given the complexity and irregularity of wind speed signals, many studies use signal decomposition as a preprocessing step to reduce data complexity and improve model accuracy. Examples include using genetic algorithms (GA) to optimize discrete wavelet filters, combining adaptive noisy complete ensemble empirical mode decomposition (CEEMDAN) with a chimpanzee optimizer to tune extreme learning machines (ELM), combining ensemble empirical mode decomposition (EEMD) with seasonal autoregressive integral moving average (SARIMA) and long short-term memory (LSTM), and combining variational mode decomposition (VMD) with convolutional neural networks (CNN) and gated recurrent units (GRU).

[0005] However, the ensemble empirical mode decomposition (EEMD) algorithm cannot completely eliminate Gaussian white noise during signal reconstruction, and the variational mode decomposition (VMD) does not perfectly match the original dataset during reconstruction, usually resulting in reconstruction errors.

[0006] Therefore, there is an urgent need for a new technology that can solve the problem of not being able to completely eliminate Gaussian white noise and having reconstruction errors. Summary of the Invention

[0007] The purpose of this invention is to provide a short-term wind speed prediction method and related equipment based on transfer learning to overcome the problems existing in the prior art. This invention can adaptively and thoroughly remove noise interference from the original wind speed sequence through CEEMDAN decomposition and accurately reconstruct the intrinsic mode components, significantly reducing noise residue and reconstruction distortion.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a short-term wind speed prediction method based on transfer learning, comprising the following steps: Obtain the raw wind speed sequence and related meteorological characteristic data; The original wind speed sequence and related meteorological feature data are input into the constructed short-term wind speed prediction model based on transfer learning to obtain the short-term predicted wind speed sequence. The short-term wind speed prediction model based on transfer learning is constructed through CEEMDAN decomposition, Bi-GRU-ED model, and transfer learning.

[0009] In some embodiments, the steps for constructing the short-term wind speed prediction model based on transfer learning specifically include: Obtain historical raw wind speed sequences and historical related meteorological characteristic data; The historical raw wind speed sequence was decomposed into several intrinsic mode functions and residual components using a 6-level CEEMDAN decomposition, and the historical relevant meteorological characteristic data were normalized. A Bi-GRU-ED model is constructed by using several intrinsic mode functions, residual components, and normalized historical relevant meteorological feature data as the input dataset for training the Bi-GRU-ED model. The input dataset is structured using a time window to form an input matrix. The trained Bi-GRU-ED model is used as a pre-trained model. Based on transfer learning, the pre-trained model is transferred to the prediction of each intrinsic mode function to obtain the prediction model corresponding to each intrinsic mode function. New residual components are predicted using the trained Bi-GRU-ED model. New intrinsic mode functions are obtained through the prediction model corresponding to each intrinsic mode function. The new residual components and new intrinsic mode functions are aggregated to obtain the short-term predicted wind speed sequence, thus completing the construction of the short-term wind speed prediction model based on transfer learning.

[0010] In some embodiments, the relevant meteorological characteristic data includes temperature, air pressure, and wind speed; The historical meteorological data includes historical temperature, historical air pressure, and historical wind speed.

[0011] In some embodiments, the Bi-GRU-ED model includes: The bidirectional gated recurrent unit layer consists of two unidirectional gated recurrent units that process in the forward and reverse directions, and is used to simultaneously capture past and future contextual information of the input dataset; The encoder uses the bidirectional gated cyclic unit layer to encode the input dataset and generate a context vector with condensed time information. The decoder uses the bidirectional gated recurrent unit layer as input, taking the context vector and previous output as input, to progressively generate the prediction sequence.

[0012] In some embodiments, the transfer learning is homogeneous transfer learning, where the source task of homogeneous transfer learning is the prediction of residual components, and the target task of homogeneous transfer learning is the prediction of intrinsic mode functions.

[0013] In some embodiments, the feature space and label space of the source task and the target task are the same.

[0014] In some embodiments, the aggregation of the new residual components and each new intrinsic mode function to obtain the constructed short-term predicted wind speed sequence specifically includes the following formula: ; In the formula, This represents a short-term forecast wind speed sequence; This represents a predicted wind speed sequence; Indicates a time index; Indicates the new residual components; k This represents the current new intrinsic mode function; This represents the new intrinsic mode function.

[0015] Secondly, the present invention provides a short-term wind speed prediction system based on transfer learning, comprising: The data acquisition module is used to acquire raw wind speed sequences and related meteorological characteristic data. The short-term predicted wind speed sequence acquisition module is used to input the original wind speed sequence and related meteorological feature data into the constructed short-term wind speed prediction model based on transfer learning to obtain the short-term predicted wind speed sequence. The short-term wind speed prediction model based on transfer learning is constructed through CEEMDAN decomposition, Bi-GRU-ED model, and transfer learning.

[0016] Thirdly, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0017] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0018] The above technical solution has the following advantages or beneficial effects: Firstly, this invention provides a short-term wind speed prediction method based on transfer learning. By introducing the CEEMDAN decomposition method, it fundamentally solves the two major bottlenecks in traditional wind speed prediction: the difficulty in completely eliminating Gaussian white noise and the existence of errors in signal reconstruction. CEEMDAN decomposition can adaptively and thoroughly remove noise interference from the original wind speed sequence and accurately reconstruct the intrinsic mode components, significantly reducing noise residue and reconstruction distortion. On this basis, the Bi-GRU-ED model is combined to extract deep temporal features from the clean signal, and the transfer learning mechanism is used to ensure the strong generalization ability of the model under limited data conditions, thereby achieving high-precision and high-stability short-term wind speed prediction.

[0019] In some embodiments, a high-precision short-term wind speed prediction model is constructed by combining a 6-level CEEMDAN decomposition with a Bi-GRU-ED model and embedding a transfer learning mechanism. The core benefits are as follows: First, the 6-level CEEMDAN decomposition can more thoroughly and precisely separate Gaussian white noise from the effective signal in the wind speed sequence, fundamentally and significantly reducing noise residue and signal reconstruction error, and improving data quality. Second, by adopting a transfer learning strategy, the parameter knowledge of the fully trained Bi-GRU-ED model is transferred to the independent prediction tasks of each modal component, effectively solving the problem of insufficient fitting ability of a single model for complex multi-component sequences, enhancing the stability and accuracy of the overall prediction, and is especially suitable for complex wind field environments with scarce data.

[0020] In some embodiments, by incorporating key meteorological features such as temperature, air pressure, and historical wind speed into the prediction system, this method significantly enhances the physical completeness of the model and the dimensionality of the input information. These features form a strong coupling relationship with wind speed, which helps the Bi-GRU-ED model to more accurately capture the dynamic evolution of atmospheric motion, thereby effectively compensating for the inadequacy of single wind speed sequence information and further improving the accuracy and reliability of the prediction results.

[0021] In some embodiments, the Bi-GRU-ED model structure employs a bidirectional gated recurrent unit (Bi-GRU) layer to simultaneously capture the forward and backward temporal dependencies of the input data, achieving deep fusion and comprehensive perception of past and future contextual information of wind speed and meteorological feature sequences. The encoder-decoder (ED) architecture uses this highly condensed context vector as the state basis, and the decoder gradually and autoregressively generates high-precision prediction sequences. The above settings enable the model to not only effectively learn the complex long-term and short-term patterns of wind speed evolution, but also accurately model its dynamic correlation with multi-dimensional features such as temperature and air pressure, thereby significantly improving the temporal coherence and physical accuracy of the prediction, and is particularly adept at handling wind speed sequences with strong non-stationarity and randomness.

[0022] In some embodiments, by employing a homogeneous transfer learning strategy, knowledge from a mature model (source task) that predicts residual components is creatively transferred to the prediction of each intrinsic mode function (target task). This fully utilizes the inherent similarity between residual components and each mode component in terms of mathematical properties and evolutionary laws, achieving effective reuse of the model's core prediction capabilities. This not only significantly improves the efficiency and accuracy of independently predicting multiple complex components, but also greatly alleviates the model training difficulties caused by insufficient data or distribution differences, enhancing the robustness and generalization ability of the overall prediction framework.

[0023] In some embodiments, this feature clarifies the consistency between the source task and the target task in terms of data structure and learning objectives. Through highly homogeneous settings, it ensures that the deep feature representations and sequence evolution rules learned from the residual prediction model can be smoothly and directly transferred to the prediction of each modality component, minimizing knowledge loss or negative transfer during the transfer process, thereby efficiently and stably improving the training starting point and final performance of each component prediction model.

[0024] In some embodiments, the aggregation formula is the final integration step in the prediction process. Its core advantage lies in achieving lossless reconstruction of the signal. By directly linearly superimposing the independent new intrinsic mode functions and new residual components obtained after CEEMDAN decomposition, it ensures the complete preservation of all effective prediction information. This process completely reverses the previous decomposition process, thereby completely avoiding the reconstruction accumulation error that may be introduced by intermediate calculations at the final output level, and directly generating a high-fidelity short-term predicted wind speed sequence.

[0025] Secondly, this invention provides a short-term wind speed forecasting system based on transfer learning. First, the system effectively removes and suppresses Gaussian white noise in the original wind speed sequence from the source using CEEMDAN decomposition, significantly reducing the prediction bias caused by noise interference. Second, it fully explores the long-term and short-term spatiotemporal dependence characteristics of meteorological data through the Bi-GRU-ED model, and combines transfer learning to achieve rapid adaptation and robust prediction of the model under different data distributions. This system not only improves the prediction accuracy, but also has good engineering applicability and scalability, providing reliable technical support for wind power forecasting and grid dispatching.

[0026] Thirdly, the present invention provides a computer device that, through a processor executing a specific computer program, can efficiently implement the steps of the method of the present invention. When performing data processing tasks, the computer device can accurately perform numerical calculations and logical judgments, avoiding errors caused by human factors. At the same time, since the computer program has high stability and reliability, it can ensure the accuracy and consistency of the data processing results.

[0027] Fourthly, the present invention provides a computer-readable storage medium. By programming the steps of the method of the present invention into a computer program and storing it on the computer-readable storage medium, users can easily load these programs onto any compatible computer device and execute them without rewriting or converting the code, which greatly improves the convenience and flexibility of program execution. Attached Figure Description

[0028] Figure 1 This is a schematic flowchart illustrating a short-term wind speed prediction method based on transfer learning, according to some embodiments of this specification. Figure 2 This is a schematic diagram of the structure of a computer device according to some embodiments of this specification. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to specific embodiments. These descriptions are for explanation purposes only and are not intended to limit the scope of the invention. To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention. It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] Currently, wind speed forecasting methods can be categorized into four types based on the forecast time domain: ultra-short-term (minutes to hours), short-term (hours to days), medium-term (days to weeks), and long-term (weeks to years); and broadly by methodology, they can be classified into physical methods, statistical methods, artificial intelligence methods, and hybrid methods.

[0031] Physical methods, based on numerical weather prediction (NWP), use topographic datasets, temperature, humidity, and air pressure to make weather forecasts through mathematical formulas. Although advanced NWP models can make relatively accurate predictions of weather at specific locations, the dynamic nature of wind speed requires large datasets and lengthy computation times. While suitable for long-term wind speed forecasting, their dependence on large-scale datasets, high processing requirements, and complex models limit their applicability to small-scale forecasts.

[0032] Statistical methods include models such as Autoregressive Moving Average (ARMA) and Autoregressive Integrated Moving Average (ARIMA), which make predictions by analyzing the temporal correlations in historical data. For example, some studies have explored the use of ARMA and generalized autoregressive conditional heteroscedasticity models for predicting the mean and volatility of wind speed, showing some effectiveness in structured time series analysis. However, traditional statistical models often have shortcomings in handling typical nonlinear and nonstationary patterns of wind speed. Although improved versions such as Nonlinear Autoregressive Exogenous Models (NARX) and Nonlinear Autoregressive Moving Average Models (NARMA) have been proposed to capture complex patterns, they are sensitive to data noise and outliers.

[0033] Artificial intelligence methods have demonstrated strong capabilities in modeling the complex nonlinear characteristics of wind speed time series. Artificial neural networks (ANNs) approximate nonlinear functions using historical meteorological data, but are susceptible to local minima and overfitting. Support vector machines (SVMs) exhibit good generalization ability and show potential in short-term wind speed forecasting with limited and noisy datasets. Extreme learning machine models (ELMs) combined with error correction mechanisms improve the training speed and accuracy of short-term wind power forecasting. Stacked long short-term memory models (S-LSTMs) and integrated gated recurrent units (GRUs) based on interval prediction frameworks, among other deep learning architectures, have also demonstrated effectiveness in time-related forecasting. Multidimensional convolutional neural networks (CNNs) emphasize the role of spatial correlation in modeling.

[0034] To overcome the limitations of single models, hybrid methods have emerged, typically combining artificial intelligence models with preprocessing techniques such as decomposition. For example, hybrid convolutional neural networks (CNNs) and long short-term memory (LSTMs) are used for wind speed prediction of autonomous sailboats; CNN encoder-decoders are combined with LSTMs to improve sequence learning; gradient boosting trees (GBOOST), LSTMs, and Transformer-based bias correction are combined with static covariates to enhance the geographic adaptability of wind speed predictions. However, while these hybrid models can improve performance, they often involve significant computational costs.

[0035] Given the complexity and irregularity of wind speed signals, many studies use signal decomposition as a preprocessing step to reduce data complexity and improve model accuracy. Examples include using genetic algorithms (GA) to optimize discrete wavelet filters, combining adaptive noisy complete ensemble empirical mode decomposition (CEEMDAN) with a chimpanzee optimizer to tune extreme learning machines (ELM), combining ensemble empirical mode decomposition (EEMD) with seasonal autoregressive integral moving average (SARIMA) and long short-term memory (LSTM), and combining variational mode decomposition (VMD) with convolutional neural networks (CNN) and gated recurrent units (GRU).

[0036] However, the ensemble empirical mode decomposition (EEMD) algorithm cannot completely eliminate Gaussian white noise during signal reconstruction, and the variational mode decomposition (VMD) does not perfectly match the original dataset during reconstruction, usually resulting in reconstruction errors. In addition, it also suffers from insufficient modeling accuracy and weak generalization.

[0037] Transfer learning, as a developing trend in deep learning, plays a crucial role in addressing data scarcity and model generalization issues across various domains. For example, a transfer learning model based on convolutional neural networks and gated recurrent units (CNN-GRU-TL) is used for short-term wind speed prediction in canyon areas, where a bidirectional long short-term memory (Bi-LSTM) model pre-trained at four wind farm locations is transferred to a central control center. However, the success of transfer learning heavily relies on the domain similarity between the source and target domains. When the source and target domains have different distributions, there is a risk of negative transfer. While retraining some layers can reduce computation, it may limit adaptability to the target task. Retraining all layers can avoid negative transfer and achieve better performance, but it increases computational cost and training complexity. Furthermore, in existing hybrid models, each decomposed component is often trained independently, leading to computational redundancy and suboptimal convergence, and a lack of joint learning strategies.

[0038] This invention constructs a short-term wind speed prediction model based on transfer learning (CEEMDAN-Bi-GRU-ED-TL). This model innovatively integrates adaptive noise complete set empirical mode decomposition (CEEMDAN), bidirectional gated cyclic unit encoder-decoder (Bi-GRU-ED), and transfer learning to form a unified prediction framework. It effectively solves the problems existing in the prior art, such as the inability to completely eliminate Gaussian white noise, the existence of reconstruction errors, large computational load, insufficient prediction accuracy, and limited generalization ability, and achieves high-precision and efficient short-term wind speed prediction.

[0039] Example: This embodiment provides a short-term wind speed prediction method based on transfer learning. See [link to relevant documentation]. Figure 1 This includes the following steps: Step 1: Obtain the original wind speed sequence and related meteorological characteristic data.

[0040] In some embodiments, the relevant meteorological characteristic data includes temperature, air pressure, and wind speed.

[0041] Step 2: Input the original wind speed sequence and related meteorological feature data into the constructed short-term wind speed prediction model (CEEMDAN-Bi-GRU-ED-TL) based on transfer learning to obtain the short-term predicted wind speed sequence; wherein, the short-term wind speed prediction model based on transfer learning is constructed through CEEMDAN decomposition, Bi-GRU-ED model and transfer learning.

[0042] In some embodiments, the steps for constructing the short-term wind speed prediction model based on transfer learning specifically include: Step 2.1: Obtain historical raw wind speed sequences and historical related meteorological characteristic data.

[0043] In some embodiments, the historical relevant meteorological data includes historical temperature T, historical air pressure P, and historical wind speed W at 10 m. 10 .

[0044] Step 2.2: The historical original wind speed sequence is decomposed into 6 intrinsic mode functions (IMF1-IMF6) and residual components (Res) using a 6-level CEEMDAN decomposition. The residual components are mainly used to capture low-frequency trends. The historical relevant meteorological feature data are normalized to the [0,1] interval using a formula to avoid model bias towards specific numerical ranges and ensure that the influence of each feature on model performance is balanced.

[0045] In some embodiments, the formula specifically includes: ; In the formula, This represents the value after normalization. This represents the original value before normalization. This represents the minimum value of the relevant historical meteorological characteristic data; This represents the maximum value of the relevant historical meteorological characteristic data.

[0046] In some embodiments, the CEEMDAN decomposes into an adaptive noise-complete ensemble empirical mode decomposition. As an improved empirical mode decomposition (EMD) method, CEEMDAN effectively mitigates mode aliasing by adding adaptive noise and performing ensemble averaging, generating more stable and reliable intrinsic mode functions and residual components. The specific decomposition steps are as follows: Step 2.2.1, in the i-th iteration, input the original time series dataset Add normal Gaussian white noise ,get: ; In the formula, Represents a time series signal; t Indicates time; Indicates the noise weighting coefficient; N This indicates the number of times noise was added.

[0047] Step 2.2.2, for the time series signal of N iterations EMD decomposition is performed, and the first intrinsic mode function is obtained by calculating the average value of the decomposition results: ; In the formula, Denotes the first eigenmode function; E represents the EMD decomposition operator, and is obtained through... Obtain residual components ; Step 2.2.3, to the residual components Add an adaptive white noise sequence to form a new time series signal: ; In the formula, This represents the first noise weighting coefficient; This represents the first IMF obtained by EMD; The signal was decomposed using EMD and the average value was calculated to obtain the second intrinsic mode function. and residual components : ; ; Step 2.2.4, for k =2,..., k Repeat the above process using the formula: ; ; In the formula, Indicates the first k +1 eigenmode function; Indicates the first k One residual component; Indicates the first k +1 residual component; Indicates the first k Noise weighting coefficient; Get the first k +1 intrinsic mode function and residual component; Step 2.2.5, repeat step 2.2.4 until the residual components cannot be further decomposed (i.e., the residuals contain at least two extreme values). At this point, the original time series is: .

[0048] Step 2.3: Construct a Bi-GRU-ED (Bidirectional Gated Cyclic Unit-Encoder-Decoder) model. Use several intrinsic mode functions, residual components, and normalized historical relevant meteorological feature data as the input dataset for the Bi-GRU-ED model to train the Bi-GRU-ED model. The input dataset is structured using a time window to form an input matrix.

[0049] In some embodiments, the Bi-GRU-ED model includes: The bidirectional gated recurrent unit layer consists of two unidirectional gated recurrent units that process in the forward and reverse directions, and is used to simultaneously capture past and future contextual information of the input dataset; The encoder uses the bidirectional gated cyclic unit layer to encode the input dataset and generate a context vector with condensed time information. The decoder uses the bidirectional gated recurrent unit layer as input, taking the context vector and previous output as input, to progressively generate the prediction sequence.

[0050] In some embodiments, Bi-GRU-ED model training (residual component prediction) involves constructing a Bi-GRU-ED model (Bi-GRU-EDRes) to predict the residual components Res. The inputs to this model include the intrinsic mode functions obtained from the decomposition, the residual components, and the normalized meteorological features (T). n P n W 10 The model is structured using a 10-time-step time window to form a 10×10 input matrix. After training, the model's weights and biases are saved as a pre-trained model.

[0051] In some embodiments, the Bi-GRU-ED model includes: Gated Recurrent Unit (GRU): As an improved version of the traditional Recurrent Neural Network (RNN), the GRU updates the gates... and reset door To address the vanishing gradient problem in long-term dependencies, the update gate determines the extent to which historical information is passed to subsequent time steps, while the reset gate determines the extent to which information from the previous time step is discarded. The mathematical formulas are as follows: Reset Door: ; Update Gate: ; Candidate hidden state: ; Hidden state: ; In the formula, Represents the sigmoid function; This indicates the weight parameters for resetting the door; This indicates the previously hidden state; Indicates the input to the gate; This indicates the offset item corresponding to the reset door; This indicates that the weight parameters of the gate are being updated; This indicates that the bias term corresponding to the updated gate is being updated; Indicates the candidate hidden state; h Indicates that the content is hidden; Weight parameters of candidate hidden states; It represents the Hadamardi (or Hadama) stack; Bias terms representing candidate hidden states; Indicates a hidden state; Bidirectional GRU (Bi-GRU): Consists of two unidirectional GRU layers, one of which processes the input sequence in the forward direction. A reverse-processing input sequence The final hidden state is obtained by concatenating the outputs of the two layers. It can capture contextual information from both the past and the future simultaneously; in the formula, This indicates that the input sequence is being processed in the forward direction. f Indicates a positive direction; Indicates a forward-gated loop unit; This indicates the positive hidden state of the previous moment; This indicates that the input sequence is processed in reverse. Indicates a reverse-gated loop unit; This indicates the reverse hidden state at the previous moment; Indicates splicing; Encoder-decoder architecture: The encoder uses a bidirectional GRU to perform bidirectional analysis on the input sequence and generate a context vector that condenses the temporal information of the input dataset; the decoder also uses a bidirectional GRU, taking the context vector and previously generated values ​​as input, and gradually generating the output sequence, using the encoded information of past and future context to construct the output.

[0052] Step 2.4: Use the trained Bi-GRU-ED model as a pre-trained model, and transfer learning to transfer the pre-trained model to the prediction of each intrinsic mode function, so as to obtain the prediction model corresponding to each intrinsic mode function.

[0053] In some embodiments, the transfer learning is homogeneous transfer learning, where the source task of homogeneous transfer learning is the prediction of residual components, the target task of homogeneous transfer learning is the prediction of intrinsic mode functions, and the feature space and label space of the source task and the target task are the same.

[0054] In some embodiments, transfer learning is applied (intrinsic mode function prediction): the trained Bi-GRU-EDRes model is used as a pre-trained model, and by fine-tuning all its layers, corresponding prediction models (Bi-GRU-ED1 to Bi-GRU-ED6) are constructed for each intrinsic mode function (IMF1-IMF6); each model uses the same type of input data as the Bi-GRU-EDRes model, and utilizes the knowledge learned during the residual prediction process to efficiently predict the corresponding intrinsic mode function; Transfer learning framework: Homogeneous transfer learning is adopted because both the source task (residual component prediction) and the target task (intrinsic mode function prediction) are sub-components of short-term wind speed prediction, and their feature spaces and label spaces are the same. In the formula, Represent the feature space; Indicates the source task; Indicate the target task; Represents the tag space; Represents the feature space of the source task; The feature space representing the target task; The label space representing the source task; The label space represents the target task; by reusing the Bi-GRU-ED model obtained from predicting the residual components, we avoid developing and training a deep network separately for each intrinsic mode function, reducing the computational burden, and achieving stable and effective initialization of model parameters based on previous learning.

[0055] Step 2.5: Predict new residual components using the trained Bi-GRU-ED model, obtain each new intrinsic mode function (IMF) using the prediction model corresponding to each IMF, aggregate the new residual components and each new IMF to obtain the short-term predicted wind speed sequence, and complete the construction of the short-term wind speed prediction model based on transfer learning.

[0056] In some embodiments, the results are aggregated and the wind speed is reconstructed as follows: Based on the additive reconstruction principle of CEEMDAN, the predicted residual components and each intrinsic mode function are aggregated to reconstruct the final short-term predicted wind speed sequence, as shown in the following formula: ; In the formula, This represents a short-term forecast wind speed sequence; This represents a predicted wind speed sequence; Indicates a time index; Indicates the new residual components; k This represents the current new intrinsic mode function; This represents the new intrinsic mode function.

[0057] This step requires no additional complex calculations, and its core guarantee is the consistency of "decomposition-reconstruction": because there is no information loss during CEEMDAN decomposition (original signal = ... Therefore, the reconstructed predicted value can completely restore all frequency components of wind speed (from high-frequency fluctuations to low-frequency trends), avoiding the accumulation of errors caused by improper reconstruction methods.

[0058] The core innovation of the short-term wind speed prediction model based on transfer learning constructed in this embodiment lies in the deep integration of CEEMDAN decomposition, Bi-GRU-ED time series modeling, and homogeneous transfer learning through a collaborative design of "decomposition-modeling-transfer-aggregation". First, CEEMDAN is used to decompose the complex nonlinear and non-stationary wind speed signal into eigenmode functions and residual components with well-defined characteristics, effectively reducing the modeling difficulty and noise interference. Then, the residual components are accurately modeled using the Bi-GRU-ED model and used as a pre-trained model. Transfer learning is used to reuse its knowledge to perform lightweight fine-tuning of each eigenmode function, avoiding the computational redundancy caused by independent training. Finally, the prediction results are aggregated by addition reconstruction. This not only leverages the ability of bidirectional time series modeling to capture complex dependencies, but also improves the model's generalization ability and training efficiency through transfer learning. Overall, it solves the problems of insufficient modeling accuracy, high computational cost, and weak generalization of traditional methods, and has both technical innovation and engineering practicality.

[0059] In one embodiment of the present invention, a short-term wind speed prediction system based on transfer learning is provided, comprising: The data acquisition module is used to acquire raw wind speed sequences and related meteorological characteristic data. The short-term predicted wind speed sequence acquisition module is used to input the original wind speed sequence and related meteorological feature data into the constructed short-term wind speed prediction model based on transfer learning to obtain the short-term predicted wind speed sequence. The short-term wind speed prediction model based on transfer learning is constructed through CEEMDAN decomposition, Bi-GRU-ED model, and transfer learning.

[0060] See Figure 2In one embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, and is suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to realize a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used for the operation of a short-term wind speed prediction method based on transfer learning.

[0061] In one embodiment of the present invention, a computer-readable storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space containing the terminal's operating system; and the storage space also contains one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the short-term wind speed prediction method based on transfer learning in the embodiment.

[0062] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented 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.

[0063] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should 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 illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0064] 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.

[0065] 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.

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A short-term wind speed prediction method based on transfer learning, characterized in that, Includes the following steps: Obtain the raw wind speed sequence and related meteorological characteristic data; The original wind speed sequence and related meteorological feature data are input into the constructed short-term wind speed prediction model based on transfer learning to obtain the short-term predicted wind speed sequence. The short-term wind speed prediction model based on transfer learning is constructed through CEEMDAN decomposition, Bi-GRU-ED model, and transfer learning.

2. The short-term wind speed prediction method based on transfer learning according to claim 1, characterized in that, The construction steps of the short-term wind speed prediction model based on transfer learning specifically include: Obtain historical raw wind speed sequences and historical related meteorological characteristic data; The historical raw wind speed sequence was decomposed into several intrinsic mode functions and residual components using a 6-level CEEMDAN decomposition, and the historical relevant meteorological characteristic data were normalized. A Bi-GRU-ED model is constructed by using several intrinsic mode functions, residual components, and normalized historical relevant meteorological feature data as the input dataset for training the Bi-GRU-ED model. The input dataset is structured using a time window to form an input matrix. The trained Bi-GRU-ED model is used as a pre-trained model. Based on transfer learning, the pre-trained model is transferred to the prediction of each intrinsic mode function to obtain the prediction model corresponding to each intrinsic mode function. New residual components are predicted using the trained Bi-GRU-ED model. New intrinsic mode functions are obtained through the prediction model corresponding to each intrinsic mode function. The new residual components and new intrinsic mode functions are aggregated to obtain the short-term predicted wind speed sequence, thus completing the construction of the short-term wind speed prediction model based on transfer learning.

3. The short-term wind speed prediction method based on transfer learning according to claim 2, characterized in that, The relevant meteorological data include temperature, air pressure, and wind speed; The historical meteorological data includes historical temperature, historical air pressure, and historical wind speed.

4. The short-term wind speed prediction method based on transfer learning according to claim 2, characterized in that, The Bi-GRU-ED model includes: The bidirectional gated recurrent unit layer consists of two unidirectional gated recurrent units that process in the forward and reverse directions, and is used to simultaneously capture past and future contextual information of the input dataset; The encoder uses the bidirectional gated cyclic unit layer to encode the input dataset and generate a context vector with condensed time information. The decoder uses the bidirectional gated recurrent unit layer as input, taking the context vector and previous output as input, to progressively generate the prediction sequence.

5. The short-term wind speed prediction method based on transfer learning according to claim 2, characterized in that, The transfer learning is homogeneous transfer learning, where the source task is the prediction of residual components and the target task is the prediction of intrinsic mode functions.

6. The short-term wind speed prediction method based on transfer learning according to claim 5, characterized in that, The feature space and label space of the source task and the target task are the same.

7. The short-term wind speed prediction method based on transfer learning according to claim 2, characterized in that, The aggregation of the new residual components and each new intrinsic mode function yields the constructed short-term predicted wind speed sequence. The specific formula includes: ； In the formula, This represents a short-term forecast wind speed sequence; This represents a predicted wind speed sequence; Indicates a time index; Indicates the new residual components; k This represents the current new intrinsic mode function; This represents the new intrinsic mode function.

8. A short-term wind speed prediction system based on transfer learning, characterized in that, include: The data acquisition module is used to acquire raw wind speed sequences and related meteorological characteristic data. The short-term predicted wind speed sequence acquisition module is used to input the original wind speed sequence and related meteorological feature data into the constructed short-term wind speed prediction model based on transfer learning to obtain the short-term predicted wind speed sequence. The short-term wind speed prediction model based on transfer learning is constructed through CEEMDAN decomposition, Bi-GRU-ED model, and transfer learning.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images