Multi-scale contrast graph neural network time sequence interpolation method for wind power data

Abstract

The invention relates to a wind power data-oriented multi-scale contrast graph neural network time sequence interpolation method, which comprises the following steps of 1, data preprocessing and graph structure construction, 2, multi-scale contrast graph neural network model construction, namely constructing an end-to-end neural network model fusing multi-scale modeling, contrast learning and graph structure information, step 3, model training; step 4, missing data interpolation and output; according to the method, by constructing an end-to-end framework fusing multi-scale pyramid attention, season-trend contrast learning and a graph neural network, high-precision and high-efficiency interpolation of missing data of a wind power time sequence is realized, and the method is particularly suitable for long-sequence and multivariable wind power scenes.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and energy big data processing technology, specifically relating to a method for interpolating missing values ​​in time series of wind power data. Background Technology

[0002] With the advancement of "dual-carbon" goals, wind power is playing an increasingly important role in new power systems. However, because wind farms are typically deployed in harsh environments such as remote areas, high altitudes, or offshore, sensors are susceptible to corrosion, freezing, and communication interruptions. This results in significant random, piecemeal, or even systemic missing data in the collected key operating parameters, such as wind speed, power, wind direction, and temperature. These incomplete data severely limit the accuracy of advanced applications such as power prediction, fault diagnosis, and condition assessment.

[0003] Existing time series interpolation methods mainly include statistical methods (such as ARIMA and linear interpolation) and deep learning-based methods (such as RNNs and Transformers). Statistical methods assume strong stationarity of the data, making it difficult to capture the complex nonlinear dynamics and multi-scale periodicity (such as diurnal cycles and seasonal fluctuations) in wind power data. While RNN-based models (such as BRITS) can handle sequence dependencies, their ability to model long-range dependencies is limited, and the computational path length increases linearly with the sequence length, resulting in low efficiency. Transformer-based models, although capable of capturing long-range dependencies through self-attention mechanisms, have a computational complexity as high as O(L...). 2 When processing long-sequence wind power data, it faces huge computational and memory overhead.

[0004] In recent years, sparse attention mechanisms (such as Informer and LogTrans) and pyramid structures (such as Pyraformer) have been proposed to reduce computational complexity. Meanwhile, contrastive learning in self-supervised learning (such as CoST) constructs positive and negative samples through data augmentation, learning robust representations invariant to perturbations and improving the model's generalization ability. Furthermore, the spatial locations of multiple wind turbines in a wind farm are geographically correlated, and their operational data exhibit spatial dependence; however, most existing interpolation methods ignore this graph structure information.

[0005] Therefore, there is an urgent need for a wind power time series interpolation method that can effectively integrate multi-scale temporal features and spatial graph structure information, and has efficient long sequence modeling capabilities and robust representation learning. Summary of the Invention

[0006] Purpose of the Invention: The purpose of this invention is to overcome the shortcomings of existing technologies and provide a multi-scale comparative graph neural network time series interpolation method for wind power data. This method achieves high-precision and high-efficiency interpolation of missing wind power time series data by constructing an end-to-end framework that integrates multi-scale pyramid attention, seasonal-trend comparative learning, and graph neural networks. It is particularly suitable for long-sequence, multi-variable wind power scenarios.

[0007] Technical Solution: To achieve the above objectives, this invention provides a multi-scale contrastive graph neural network time series interpolation method for wind power data, comprising the following steps:

[0008] Step 1: Data Preprocessing and Graph Structure Construction

[0009] (1) Data Acquisition and Standardization: Time-series data of multiple wind turbines in the wind farm, including wind speed, power, and wind direction, are acquired. The sampling frequency is typically 10 or 30 minutes. The raw data is standardized using Z-score standardization to eliminate dimensional differences.

[0010]

[0011] in It is the standardized data value, x i,t,d μ represents the original measured value of the d-th variable at time point t for the i-th wind turbine. d σ d , where are the mean and standard deviation of the d-th dimension variable across all wind turbines and time, respectively, N is the number of wind turbines, D is the variable for wind turbines, and T is the time length.

[0012] (2) Missing Data Simulation and Partitioning: To train the model, missing data is introduced into the complete dataset in random blocks or points with a certain probability (2.5%), and a corresponding mask matrix M∈{0,1} is generated. N×T×D :

[0013]

[0014] Missing types include: punctate missing: with probability p p Randomly set m i,t,d =0; Block missing: with probability p b Randomly select the start time t0 and length l ~ Uniform(l) min ,l max Set l consecutive time points as missing, and divide the dataset into training, validation and test sets in a ratio of 70%:10%:20%.

[0015] (3) Constructing a spatial adjacency matrix: using the geographical coordinates between wind turbines and the geographical coordinates of the wind turbines pi =(lat i ,lon i ), calculate the Euclidean distance d ij =||p i -p j ||, Construct a spatial adjacency matrix The threshold Gaussian kernel method is as follows:

[0016]

[0017] Where b is the bandwidth parameter and r0 is the connection radius. The Pearson correlation coefficient for each pair of wind turbines in the "weekly correlation" is calculated using the k-nearest neighbor correlation graph method. Construct an adjacency matrix using the top-k values:

[0018]

[0019] Final adjacency matrix normalization process: in Let A be the normalized adjacency matrix, which is an N×N matrix representing the connection relationships between N nodes in the graph, and S be the degree matrix to capture the spatial dependencies between wind turbines.

[0020] Step 2: Construction of a multi-scale contrastive neural network model

[0021] An end-to-end neural network model integrating multi-scale modeling, contrastive learning, and graph structure information is constructed. Its core architecture includes:

[0022] Step 2-1: Multi-scale Pyramid Attention Encoder (PAM Encoder)

[0023] Given an input sequence of length L, PAM divides it into N... s Each scale is measured by a downsampling factor s. k Gradually compress the sequence length. Let the sequence at the k-th scale be... in s is the step size factor (usually 2). Within each scale, a neighbor-tree attention mechanism is used to model local dependencies; between scales, feature aggregation is achieved through hierarchical skip connections to realize cross-scale information interaction. The pyramid attention output is defined as:

[0024] Z (k) =MSA(LN(W q Q (k) W k K (k) W v V (k) ))+DownSample(Z (k-1) (5)

[0025] Where: MSA is multi-head self-attention; LN is layer normalization; DownSample(·) is max pooling or convolution downsampling; skip connections implement an O(1) signal propagation path, and our total time complexity is O(L), which is significantly better than the Transformer's O(L) complexity. 2 ).

[0026] Step 2-2: Seasonal-Trend Comparison Learning Module

[0027] (1) Based on the intermediate representation of the encoder output, a trend feature decoupler (TFD) and a seasonal feature decoupler (SFD) are introduced;

[0028] (2) Trend Feature Decoupler (TFD): A hybrid autoregressive expert network is used to model trends.

[0029]

[0030] Where E represents the number of experts, w e (·) represents the gating weight. The e-th autoregressive expert is selected. Time-domain trend features are extracted using TFD, and time-domain contrastive loss (L...) is applied. time ) is optimized to ensure that the trend representation of the same time series is consistent after different enhancements;

[0031] (3) Seasonal Feature Decoupler (SFD): A learnable Fourier layer is used to extract seasonal features in the frequency domain.

[0032]

[0033] in For Fourier transform, and The amplitude and phase spectra are respectively. MLP is a multilayer perceptron, and frequency domain contrast loss (including amplitude loss L) is used. amp and phase loss L phase Optimize to ensure the robustness of the seasonal pattern in the frequency domain;

[0034] (4) By constraining the time correspondence of trend features in the time domain and constraining the amplitude and phase consistency of seasonal features in the frequency domain, the model can accurately align the trend and seasonal features of time series data and distinguish between matching and mismatched time series sample pairs.

[0035] Time-domain contrast loss, applied to trend representation:

[0036]

[0037] in The trend embedding vector for the first sample. Let be the trend embedding vector of the k-th sample, and τ be the temperature parameter. Frequency domain contrast loss is applied to the seasonal representation.

[0038]

[0039] Among them: amplitude loss: Phase loss: The total comparison loss is:

[0040]

[0041] Where γ is the equilibrium hyperparameter.

[0042] Steps 2-3: Graph Fusion Module

[0043] The hidden state of each wind turbine in the time dimension As a graph node feature:

[0044]

[0045] Spatial information aggregation is performed using a two-layer graph convolutional network (GCN):

[0046]

[0047] Among them W (0) W (1) The weight matrix is ​​a learnable weight matrix, and σ is the activation function (such as ReLU). This is the normalized adjacency matrix. Output the final spatiotemporal joint representation.

[0048] Step 3: Model Training

[0049] (1) Principal interpolation loss function:

[0050] Let the model output interpolation value be... The true value is X, and the loss is calculated only at the observation location:

[0051]

[0052] Where Ω={(i,t,d)∣mi ,t,d =1} is the observation set, and MAE is used to ensure robustness to outliers;

[0053] (2) Joint loss function

[0054] The total loss is the weighted sum of the reconstruction loss and the contrast loss:

[0055]

[0056] Where λ>0 is the weight coefficient, which controls the contribution of contrastive learning.

[0057] (3) Training strategy: Use Adam or AdamW optimizer, set the initial learning rate to le-4, and adopt cosine annealing learning rate scheduler; set 200 as the number of training rounds, and adopt early stopping mechanism based on validation set performance; during training, additionally randomly mask 5% of the input data values ​​to enhance the robustness of the model to noise and missing values; for single-step prediction tasks, a pre-training strategy can be adopted: construct the network as an autoencoder in the initial rounds, and perform reconstruction and prediction at the same time to force the network to learn long-range dependencies.

[0058] Step 4: Imputing and Outputting Missing Data

[0059] The missing wind power sequence to be interpolated (χ) miss The pre-trained MSC-GNI model is input, and after encoding, contrastive learning, graph fusion, and decoding, a complete reconstructed sequence is output.

[0060] Decoders are typically simple fully connected layers:

[0061]

[0062] Where v i This represents the final characterization of the i-th wind turbine.

[0063] Compared with the prior art, the present invention has the following significant advantages:

[0064] (1) Efficient long sequence modeling: Using the PAM structure, O(L) time complexity and O(1) signal propagation path are achieved, which is significantly better than the O(L) time complexity of Transformer. 2 ),

[0065] (2) Robust feature learning: A seasonal-trend dual-channel contrast learning framework is proposed, which constructs contrast tasks in the time domain and frequency domain respectively, forcing the model to learn decoupled representations that are invariant to noise, scaling and jitter, thereby improving the interpolation stability.

[0066] (3) Integrating spatial information: The spatial topological relationship between wind turbines in the wind farm is explicitly modeled using graph neural networks, and spatial correlation is integrated into the interpolation process. This fully utilizes the collaborative information of multiple wind turbine data, and can still provide reliable estimates, especially when data in local areas is missing at the same time.

[0067] (4) End-to-end automation: The entire method is an end-to-end learning framework, which does not require complex feature engineering or parameter tuning, has a high degree of automation, and is easy to deploy and apply;

[0068] (5) Strong applicability: The method is universal and can be applied to wind farms of different sizes and geographical distributions. It has significant advantages in processing long-term, multi-variable, and high-missing-rate wind power data. Attached Figure Description

[0069] Figure 1 This is the overall flowchart of the present invention;

[0070] Figure 2 This is a schematic diagram of the Pyramid Attention Module (PAM) structure.

[0071] Figure 3 A visual representation of seasonal trends. Detailed Implementation

[0072] The following description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Those skilled in the art should understand that, without departing from the scope of the present invention, various modifications and improvements can be made when specific implementation conditions change, but all such modifications and improvements should fall within the scope of protection of the appended claims.

[0073] Example: This invention proposes a multi-scale contrastive graph neural network time series interpolation method for wind power data. Specifically, this invention includes the following steps:

[0074] Step 1: Data Preparation

[0075] Data was obtained from public datasets (such as GEFCom2014-W and Baidu KDD Cup 2022). After standardization, block missing (4-9 consecutive points in length) and random missing points were introduced with a probability of 2.5% to generate training, validation and test sets.

[0076] Step 2: Graph Construction

[0077] Based on the coordinates between wind power data, the Euclidean distance is calculated, and a 10×10 adjacency matrix A is constructed using a threshold Gaussian kernel function.

[0078] Step 3: Model Configuration

[0079] (1) Input sequence length H = 96 (corresponding to 48 hours);

[0080] (2) PAM module: Set 4 scales (N=4), each scale height S=4, number of heads H=6;

[0081] (3) Comparative learning: α = 1.0, λ = 0.5;

[0082] (4) Graph fusion: A 2-layer GCN is used with a hidden layer dimension of 128;

[0083] (5) Optimizer: AdamW, learning rate le-4, cosine annealing, batch size 32, maximum number of rounds 300, early stopping mechanism 40;

[0084] Step 4: Training and Evaluation

[0085] The model is trained on the training set, and the interpolation performance is evaluated using metrics such as MAE, RMSE, and MAPE on the test set.

[0086] Step 5: Output

[0087] Based on a well-trained state-of-the-art model, we can effectively process wind power time series data with missing values ​​and generate complete, high-quality prediction sequences, providing a solid foundation for subsequent wind farm analysis and decision-making. Experimental results show that our model achieves significant improvements in imputation performance on two important public datasets, GEFCom2014-W and Baidu KDD Cup 2022, as shown in Tables 1 and 2.

[0088] Table 1 compares the interpolation performance on the GEFCom2014-W dataset (H=96).

[0089]

[0090] Table 2: Comparison of interpolation performance on the Baidu KDD Cup 2022 wind collector (H=96)

[0091]

[0092] On the GEFCom2014-W dataset (H=96), the MSC-GNI(Ours) model achieved the best results in MAE, RMSE and MAPE metrics of 0.253, 0.362 and 11.4% respectively, demonstrating stronger imputation capabilities compared to other mainstream methods.

[0093] On the Baidu KDD Cup 2022 wind electronic collection (H=96), the MSC-GNI (Ours) model also performed excellently, ranking among the top in all performance indicators, with MAE of 0.271, RMSE of 0.391, and MAPE of only 13.2%.

[0094] These results fully demonstrate the effectiveness, robustness, and generalization ability of our model in handling the problem of missing wind power data, and show that it significantly outperforms traditional linear interpolation, ARIMA, and some advanced deep learning methods.

[0095] like Figure 1 The diagram shows the overall flowchart of the present invention, illustrating the entire process of data preprocessing, model architecture, and interpolation output.

[0096] like Figure 2 The diagram shows a schematic of the multi-scale pyramid attention module (PAM) structure of the present invention, which demonstrates intra-scale adjacency connections and inter-scale hierarchical aggregation.

[0097] like Figure 3 As shown, this invention illustrates a seasonal-trend visualization, where the module displays a visualization generated by selecting a single seasonality. Colors represent two different trends. Alternatively, a visualization can be generated by selecting a single trend. Colors represent three different seasonal patterns.

[0098] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention. Equivalent transformations or substitutions made based on the above technical solutions all fall within the scope of protection of the claims of the present invention.

Claims

1. A multi-scale contrastive graph neural network time series interpolation method for wind power data, characterized in that, Includes the following steps: Step 1: Data preprocessing and graph structure construction. Step 2: Construction of a multi-scale contrastive graph neural network model. This involves building an end-to-end neural network model that integrates multi-scale modeling, contrastive learning, and graph structure information. Step 3: Model training, Step 4: Imputing and outputting missing data.

2. The multi-scale contrastive graph neural network time series interpolation method for wind power data according to claim 1, characterized in that, Step 1: Data preprocessing and graph structure construction, as detailed below. 1-1. Data Acquisition and Standardization: Acquire multivariate time-series data on wind speed, power, and wind direction from multiple wind turbines in the wind farm. The sampling frequency is typically 10 or 30 minutes. Standardize the raw data using Z-score standardization to eliminate dimensional differences. in It is the standardized data value, x i,t,d μ represents the original measured value of the d-th variable at time point t for the i-th wind turbine. d σ d , where are the mean and standard deviation of the d-th dimension variable across all wind turbines and time, respectively, and N is the number of wind turbines. D represents the variable of the wind turbine, and T represents the time duration. 1-2. Missing Data Simulation and Partitioning: To train the model, missing data is introduced into the complete dataset in random blocks or points with a certain probability (2.5%), and a corresponding mask matrix M∈{0,1} is generated. N×T×D : Missing types include: punctate missing: with probability p p Randomly set m i,t,d =0; Block missing: with probability p b Randomly select the start time t0 and length l ~ Uniform(l) min ,l max Set l consecutive time points as missing, and divide the dataset into training, validation and test sets in a ratio of 70%:10%:20%. 1-3. Constructing a spatial adjacency matrix: Utilizing the geographical coordinates between wind turbines. i =(lat i ,lon i ), calculate the Euclidean distance d ij =||p i -p j ||, Construct a spatial adjacency matrix The threshold Gaussian kernel method is as follows: Where b is the bandwidth parameter and r0 is the connection radius, the Pearson correlation coefficient of each pair of wind turbines in the "weekly correlation" is calculated using the k-nearest neighbor correlation graph method. Construct an adjacency matrix using the top-k values: Final adjacency matrix normalization process: in Let A be the normalized adjacency matrix, which is an N×N matrix representing the connection relationships between N nodes in the graph, and S be the degree matrix to capture the spatial dependencies between wind turbines.

3. The multi-scale contrastive graph neural network time series interpolation method for wind power data according to claim 2, characterized in that, Step 2: Construction of a multi-scale contrastive graph neural network model, as detailed below. Step 2-1: Multi-scale Pyramid Attention Encoder (PAM Encoder) Given an input sequence of length L, PAM divides it into N... s Each scale is measured by a downsampling factor s. k Gradually compress the sequence length, let the sequence at the k-th level be... in s is the step size factor. Within each scale, a neighbor-tree attention mechanism is used to model local dependencies; between scales, feature aggregation is performed through hierarchical skip connections to achieve cross-scale information interaction. The pyramid attention output is defined as: Z (k) =MSA(LN(W q Q (k) ,W k K (k) ,W v V (k) ))+DownSample(Z (k-1) ) (5) Where: MSA is multi-head self-attention; LN is layer normalization; DownSample(·) is max pooling or convolution downsampling; skip connections implement an O(1) signal propagation path, with a total time complexity of O(L), which is significantly better than the Transformer's O(L) time complexity. 2 Step 2-2: Seasonal-Trend Comparison Learning Module (1) Based on the intermediate representation of the encoder output, a trend feature decoupler (TFD) and a seasonal feature decoupler (SFD) are introduced; (2) Trend Feature Decoupler (TFD): A hybrid autoregressive expert network is used to model trends. Where E represents the number of experts, w e (·) represents the gating weight. For the e-th autoregressive expert, time-domain trend features are extracted using TFD, and time-domain contrastive loss (L) is applied. time ) is optimized to ensure that the trend representation of the same time series is consistent after different enhancements; (3) Seasonal Feature Decoupler (SFD): A learnable Fourier layer is used to extract seasonal features in the frequency domain. in For Fourier transform, and The amplitude and phase spectra are respectively. MLP is a multilayer perceptron, and frequency domain contrast loss (including amplitude loss L) is used. amp and phase loss L phase Optimize to ensure the robustness of the seasonal pattern in the frequency domain; (4) By constraining the time correspondence of trend features in the time domain and constraining the amplitude and phase consistency of seasonal features in the frequency domain, the model can accurately align the trend and seasonal features of time series data and distinguish between matching and mismatched time series sample pairs. Time-domain contrast loss, applied to trend representation: in The trend embedding vector for the first sample. Let τ be the trend embedding vector of the k-th sample, τ be the temperature parameter, and frequency domain contrast loss be applied to seasonal representation. Among them: amplitude loss: Phase loss: The total comparison loss is: Where γ is the equilibrium hyperparameter. Steps 2-3: Graph Fusion Module The hidden state of each wind turbine in the time dimension As a graph node feature: Spatial information aggregation is performed using a two-layer graph convolutional network (GCN): Among them W (0) W (1) Let σ be the learnable weight matrix, and σ be the activation function. To normalize the adjacency matrix, output the final spatiotemporal joint representation.

4. The multi-scale contrastive graph neural network time series interpolation method for wind power data according to claim 2, characterized in that, Step 3: Model training, as detailed below. (1) Principal interpolation loss function: Let the model output interpolation value be... The true value is X, and the loss is calculated only at the observation location: Where Ω={(i,t,d)∣m i,t,d =1} is the observation set, and MAE is used to ensure robustness to outliers; (2) Joint loss function The total loss is the weighted sum of the reconstruction loss and the contrast loss: Where λ>0 represents the weighting coefficient, which controls the contribution of contrastive learning. (3) Training strategy: Use Adam or AdamW optimizer, set the initial learning rate to le-4, and use cosine annealing learning rate scheduler; Set the number of training rounds to 200 and adopt an early stopping mechanism based on the performance on the validation set. During training, additionally randomly mask 5% of the input data values ​​to enhance the model's robustness to noise and missing values. For single-step prediction tasks, a pre-training strategy can be adopted: construct the network as an autoencoder in the initial rounds and perform reconstruction and prediction simultaneously to force the network to learn long-range dependencies.

5. The multi-scale contrastive graph neural network time series interpolation method for wind power data according to claim 2, characterized in that, Step 4: Imputing and outputting missing data, as detailed below. The missing wind power sequence to be interpolated (χ) miss The pre-trained MSC-GNI model is input, and after encoding, contrastive learning, graph fusion, and decoding, a complete reconstructed sequence is output. The decoder is a simple fully connected layer: Where v i This represents the final characterization of the i-th wind turbine.

6. An electronic 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 program, it implements the multi-scale contrastive graph neural network time series interpolation method for wind power data as described in any one of claims 1 to 5.

7. A computer-readable storage medium storing computer instructions thereon, characterized in that, When the computer instructions are executed by the processor, they implement the multi-scale contrastive graph neural network time series interpolation method for wind power data as described in any one of claims 1-5.

Images