New energy missing data completion method and system based on multi-model fusion

Abstract

The invention discloses a new energy missing data completion method and system based on multi-model fusion. The method comprises the following steps: performing linear and nonlinear correlation two-stage feature screening on a collected new energy data set; normalizing the data set, constructing a missing data mask according to a missing proportion parameter, and dividing the data set; the designed AFMFormer model is constructed and trained; a self-adaptive frequency domain feature extraction module constructed in the model utilizes a data-driven dominant frequency extraction and frequency spectrum noise suppression mechanism to enhance the modeling capability of the model for a long-sequence dominant mode structure; the multi-granularity time sequence coding module realizes semantic consistent feature fusion through short-term features and long-term trends of a multi-scale modeling time sequence and through a learnable weight and an adaptive alignment mechanism, and generalization ability training of the model under a complex time sequence structure is further improved. Compared with a traditional method, the method shows higher precision, robustness and generalization ability in a high-fluctuation and high-missing-rate data scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of new energy missing data completion, and particularly relates to a new energy missing data completion method and system based on multi-model fusion. BACKGROUND

[0002] In a new energy power generation system, data acquisition and monitoring is a core link for ensuring safe operation and energy efficiency optimization of the system. Key parameters such as bearing temperature of a wind turbine generator and component temperature of a photovoltaic array are closely related to power generation efficiency, equipment life and fault risk of new energy equipment. Due to the fact that new energy stations are mostly distributed in complex natural environments (such as sand erosion in a desert photovoltaic area and salt spray corrosion in a sea wind power station), problems such as sensor failure and communication interruption cause random missing and continuous missing of monitoring data to occur frequently. Completing new energy equipment operation data can provide an important reference for operation state assessment and fault warning.

[0003] New energy missing data completion is a time series interpolation task, which completes missing values of a new energy equipment at some time points or time periods by analyzing past new energy equipment operation data changing over time. In a traditional data completion method, a linear interpolation method is difficult to capture nonlinear characteristics in new energy data, a KNN interpolation method is prone to introduce pseudo fluctuation components when processing spatio-temporal correlation of wind power data, and a completion model based on LSTM can capture time sequence dependency but is insufficient in modeling cross-modal interaction of multi-source heterogeneous data of new energy.

[0004] The current technology faces three core challenges: first, new energy data has high complexity, and there can be complex mutual relationships between different sensors. A traditional Pearson correlation coefficient ignores a large number of nonlinear characteristics, resulting in failure of feature correlation modeling; second, new energy data is typical long time series data, and data at a current time often has a dependency relationship with data in a long period of time in the past. Traditional completion models are difficult to effectively capture this long sequence dependency relationship, resulting in inaccurate completion results; and finally, due to the instability of energy sources (such as solar and wind energy sources), new energy data has high volatility, which greatly interferes with model feature learning. How to consider nonlinear characteristics in the new energy data completion task, fully extract features of multi-source heterogeneous data such as time and multiple sensors of new energy units, reduce data volatility, and improve the accuracy and generalization ability of the model is a major challenge in current research. SUMMARY

[0005] Inventive purpose: In view of the problems of high data missing rate, complex time sequence dependence and insufficient accuracy of traditional completion method in new energy power generation scene, the purpose of the present application is to provide a new energy missing data completion method and system based on multi-model fusion, which learns the characteristics of new energy heterogeneous data through linear and nonlinear feature screening of input data, adaptive frequency domain feature extraction, multi-granularity time sequence modeling and other strategies, and improves the accuracy of missing data completion and system robustness of the model.

[0006] Technical scheme: In order to achieve the above inventive purpose, the present application adopts the following technical scheme:

[0007] In the first aspect, the present application provides a new energy missing data completion method based on multi-model fusion, comprising the following steps:

[0008] Linear and nonlinear correlation two-stage feature screening is performed on the collected new energy data set;

[0009] The data set is normalized, a missing data mask is constructed according to the missing proportion parameter, and the data set is divided;

[0010] A fusion model AFMFormer(Adaptive Frequency-aware Multi-scale Transformer) including an adaptive frequency domain feature extraction module, a multi-granularity time sequence encoding module and an adaptive fusion completion module is constructed; wherein the adaptive frequency domain feature extraction module is designed based on improved Fourier transform, has frequency selectivity and multi-resolution perception ability, realizes adaptive frequency spectrum decomposition, extracts the dominant frequency component in the sequence and suppresses high-frequency noise; the multi-granularity time sequence encoding module includes two parallel encoding paths: one is a Patch-based Transformer path, which divides the time sequence into several time windows and models local attention; the other is a Standard Transformer path, which is used to extract the trend information and overall dynamic change in the sequence; the adaptive fusion completion module is used to align the output features of different paths in the time dimension, and dynamically weight and fuse the output features of the two encoding paths;

[0011] The AFMFormer model is trained, and the trained model is used for missing data interpolation to output the missing data completion result.

[0012] Further, the feature screening includes: calculating the Pearson correlation coefficient and the maximum mutual information coefficient between the features in the collected new energy data set, if the correlation between any two features does not reach a preset threshold, it is considered that the data is redundant, and the features with small correlation with the target variable are ignored.

[0013] Further, the adaptive frequency domain feature extraction module comprises an adaptive frequency domain enhancement block and a cross-scale feature interaction block.

[0014] The adaptive frequency domain enhancement block introduces a learnable threshold parameter after performing fast Fourier transform on the input value, which is used to screen the high-energy region in the frequency domain; after screening the main frequency domain information, adaptive frequency spectrum enhancement is performed through local and global dual-channel learning filters.

[0015] The cross-scale feature interaction block inputs the data after adaptive frequency domain enhancement into two parallel convolution layers with different sizes of convolution kernels, and the outputs of the two convolution layers are interacted through element-wise multiplication to enhance the feature representation; the above two interacted features are fused and processed through a fusion convolution layer to obtain the final output feature.

[0016] Further, the calculation process of the adaptive frequency domain enhancement block is as follows:

[0017]

[0018] P = ‖F‖ 2

[0019] F filtered = F ⊙ (P > θ)

[0020] F enhanced = W G ⊙F + W L ⊙F filtered

[0021]

[0022] where χ t is the input observation value, denotes fast Fourier transform, F is the frequency domain representation, P is the power spectrum, θ is the trainable threshold, F filtered is the filtered frequency domain representation, W G is the global filter, W L is the local filter, χ t ′ is the time series representation obtained by inverse fast Fourier transform, N is the number of variables, T is the input sequence length, ‖·‖ 2 denotes the amplitude of each frequency component of the frequency domain representation F squared, and ⊙ denotes element-wise multiplication.

[0023] Further, in the multi-granularity time series encoding module, the Patch-based Transformer path first uses a Patch-based Embedding layer to divide the input time series data into a plurality of local blocks (patches) The original input is mapped to a high-dimensional space by a linear layer, where B is the batch size, T is the sequence length, N is the number of variables, and L is the length of each block; then a learnable position encoding is added to model the sequential relationship between the patches, obtaining the representation where d is the embedding dimension; then the high-dimensional space representation is encoded by a Patch-based Encoder; the Patch-based Encoder includes multiple stacked Patch-based Transformer encoder modules, each of which consists of a self-attention mechanism, a feedforward network, a residual connection, and a one-dimensional batch normalization, where the one-dimensional batch normalization is to transpose and expand the input in the time dimension, apply one-dimensional batch normalization to each feature channel, and realize standardization processing across time steps;

[0024] The Standard Transformer path first uses a Standard Embedding layer to map the time dimension of the original input to a high-dimensional space by a linear layer; then the high-dimensional space representation is encoded by a Standard Encoder; the Standard Encoder includes multiple stacked standard Transformer encoder modules, each of which consists of a self-attention mechanism, a feedforward network, a residual connection, and a standard layer normalization; the multi-granularity time sequence encoding module encodes the input in parallel through the above two paths.

[0025] Further, the adaptive fusion and completion module uses an adaptive average pooling operation to uniformly divide the sequence obtained by the Standard Transformer path and calculate the mean value on each sub-interval, and adjusts the output of the Standard Transformer path to the sequence length of the output of the Patch-based Transformer path; and on this basis, a learnable fusion weight parameter is introduced to adaptively adjust the contribution of the two aligned encoding paths to the final completion result, and then the parameter is fused to obtain the feature representation; finally, a linear mapping head is used to perform feature flattening and linear mapping on the fused features, and the high-dimensional features are restored to the predicted values in sequence form.

[0026] In a second aspect, the present application provides a new energy missing data completion system based on multi-model fusion, comprising: a feature screening module for performing two-stage linear and nonlinear correlation feature screening on the collected new energy data set; a data preprocessing module for normalizing the data set, constructing a missing data mask according to a missing proportion parameter, and dividing the data set; a model construction module for constructing a fusion model AFMFormer including an adaptive frequency domain feature extraction module, a multi-granularity time sequence encoding module, and an adaptive fusion completion module; wherein the adaptive frequency domain feature extraction module is designed based on an improved Fourier transform and has frequency selectivity and multi-resolution perception capability, realizes adaptive frequency spectrum decomposition, extracts dominant frequency components in the sequence and suppresses high-frequency noise; the multi-granularity time sequence encoding module includes two parallel encoding paths: one is a Patch-based Transformer path that models local attention by dividing the time sequence into a plurality of time windows; the other is a Standard Transformer path for extracting trend information and overall dynamic changes in the sequence; the adaptive fusion completion module is used to align the output features of different paths in the time dimension, and dynamically weight and fuse the output features of the two encoding paths; and a model training module for training the AFMFormer model, using the trained model to perform missing data imputation, and outputting the missing data completion result.

[0027] In a third aspect, the present application provides a computer system comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the new energy missing data completion method based on multi-model fusion.

[0028] In a fourth aspect, the present application provides a computer readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the steps of the new energy missing data completion method based on multi-model fusion.

[0029] In a fifth aspect, the present application provides a computer program product comprising a computer program, wherein the computer program, when executed by a processor, implements the steps of the new energy missing data completion method based on multi-model fusion.

[0030] Beneficial effects: The new energy missing data completion method based on multi-model fusion provided by the application can effectively reduce the dimension of input data, reduce the complexity of input, and enhance the model's ability to capture the nonlinear relationship between features and the model's learning ability; the self-adaptive frequency domain feature extraction module designed by the application uses a data-driven main frequency extraction and spectrum noise suppression mechanism to enhance the model's ability to model long sequence main mode structure and improve the robustness to high volatility data; the multi-granularity time sequence coding module constructed by the application models the short-term features and long-term trends of the time sequence through multi-scale modeling, and realizes feature fusion with consistent semantics through a learnable weight and an adaptive alignment mechanism, further improving the model's generalization ability under complex time sequence structure, and effectively completing the completion of missing data between 8 hours in a single inference. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is a general flowchart of the embodiment of the application.

[0032] Figure 2 is a feature selection flowchart in the embodiment of the application.

[0033] Figure 3 is a Pearson correlation coefficient heat map of the data set in the embodiment of the application.

[0034] Figure 4 is a maximum mutual information coefficient heat map of the data set in the embodiment of the application.

[0035] Figure 5 is a schematic diagram of the overall structure of the model in the embodiment of the application.

[0036] Figure 6 is a structure diagram of the self-adaptive frequency domain enhancement block in the embodiment of the application.

[0037] Figure 7 is a structure diagram of the cross-scale feature interaction block in the embodiment of the application. DETAILED DESCRIPTION

[0038] The technical solutions of the application will be further described below in combination with the drawings and specific embodiments.

[0039] As shown in Figure 1 , the new energy missing data completion method based on multi-model fusion disclosed by the embodiment of the application includes the following steps:

[0040] (1) Linear and nonlinear two-stage feature selection is performed on the collected new energy data set (such as a wind turbine data set) to select feature data strongly correlated with the target variable (such as wind power).

[0041] Specifically, the embodiment adopts Pearson correlation coefficient to screen features with high linear correlation with the target variable, and adopts maximum mutual information coefficient to screen features with high nonlinear correlation with the target variable. Pearson correlation coefficients and maximum mutual information coefficients between all sensor features can be calculated, and if the correlation between any two features does not reach a preset threshold, it is considered that there is redundancy in the data, and the features with smaller correlation with the target variable are ignored.

[0042] The feature screening process of the embodiment will be described in detail below taking a public data set as an example. The data set corresponds to a Beberibe wind farm (UEBB) located in the coastal area of northeast Brazil, and the sampling interval of the SCADA system is 10 minutes; 32 Enercon E-48 wind turbines in the wind farm are all equipped with 100-meter metering masts in accordance with the IEC standard, and are equipped with five layers of first-class calibration cup anemometers and one layer (100 meters) of three-dimensional acoustic anemometers; the data set contains 3 kinds of time and location information, 16 kinds of wind speed and direction information, 6 kinds of meteorological condition information, 11 kinds of equipment operating state information, and 6 kinds of other auxiliary information. As shown in Figure 2 , the feature screening specifically includes:

[0043] (1.1) Linear correlation feature screening: calculate the Pearson correlation coefficient (PCC) between features, and set the threshold value to 0.8. If the linear correlation between any two features exceeds the threshold value, it is considered that there is redundancy in the data, and the features with smaller correlation with the target variable are ignored. Assuming that there are variables X and Y, the Pearson correlation coefficient of X and Y is calculated as follows:

[0044]

[0045] Where x i and y i are the values of the i th sample of X and Y, n is the number of samples, and are the sample means of X and Y.

[0046] The Pearson correlation coefficient heat map of the wind turbine data set is shown in Figure 3 . After linear feature screening, the features shown in Table 1 are obtained:

[0047] Table 1 Linear feature selection results

[0048]

[0049]

[0050] (1.2) Nonlinear correlation feature screening: For the features in Table 1 that have passed linear feature screening, the maximum mutual information coefficient between the features is calculated, and a threshold of 0.7 is set. If the nonlinear correlation between any two features exceeds the threshold, it is considered that there is redundancy in the data, and the features with smaller correlation with the target variable are ignored. Assuming that there are variables X and Y, the calculation method of the mutual information of X and Y is as follows:

[0051]

[0052] where, and are the value spaces of X and Y, P(x) and P(y) are the marginal probability distributions of X and Y, and P(x, y) is the joint probability distribution of X and Y.

[0053] Given the sample set of variables X and Y For any grid division G (dividing the x-y plane into x bin intervals in the x direction and y bin intervals in the y direction), the mutual information of the sample under this grid division is I(G). The calculation method of the maximum mutual information coefficient of the sample is as follows:

[0054]

[0055] where B is an upper limit related to the number of samples n, and B = n α (α∈(0.6,0.8)) is usually taken.

[0056] The maximum mutual information coefficient heat map of the wind turbine dataset is shown in Figure 4 Table 2 shows the features obtained after nonlinear feature screening:

[0057] Table 2 Nonlinear feature selection results

[0058] Variable name MIC Variable name MIC wind_speed_max 0.7492 rotor_rpm_max 0.7994 wind_speed_cube 1.0000 rotor_rpm_min 0.7119 sonic_wind_speed 0.9712 active_power_total 0.8595 wind_speed_nacelle 0.8418 active_power_total_max 0.8050 wind_speed_nacelle_max 0.8243 active_power_total_min 0.7116 rotor_rpm 0.8636 wind_speed 1.0000

[0059] (2) Normalize the dataset, construct a missing data mask according to the missing proportion parameter, and divide the dataset. In the dataset, the input is the value of multiple features containing missing data of the wind turbine within a period of time, and the expected output is the complete completion value of the missing data (such as wind speed) within the period of time.

[0060] In this embodiment, step (2) specifically includes:

[0061] (2.1) Data set normalization, using standard deviation normalization to process the data, converting the data into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating the dimensional influence of the data. The calculation formula of standard deviation normalization is as follows:

[0062]

[0063] where μ is the mean of the dataset and σ is the standard deviation of the dataset;

[0064] (2.2) Constructing the missing data mask according to the missing ratio parameter, creating an array of all 1s according to the shape of the data, representing that all data has no missing. Then, according to the missing ratio parameter, randomly selecting the corresponding proportion of element positions in the array, setting the values at these positions to 0 to represent data missing. Finally, the missing data mask is obtained. In this experiment, the missing ratio ε = 0.2 is taken.

[0065] (2.3) Dataset division, dividing the dataset and the corresponding mask into windows with a time step of 50. Among them, the input is the value of the wind turbine in 50 steps, including multiple features with missing data, and the supervision label is the complete completion value of the missing data in 50 steps. Among them, 80% of the data is divided into a training set, and 20% of the data is divided into a test set.

[0066] (3) Constructing a fusion model AFMFormer including an adaptive frequency domain feature extraction module, a multi-granularity time series encoding module, and an adaptive fusion completion module.

[0067] In this embodiment, the adaptive frequency domain feature extraction module performs frequency domain transformation on the input multi-dimensional time series data, realizes adaptive frequency spectrum decomposition combined with a data-driven dominant frequency selection mechanism, extracts the dominant frequency components in the sequence and suppresses high-frequency noise through a frequency spectrum attention mechanism, to improve the modeling ability of the main mode in long-period and strong volatility sequences. The module is designed based on an improved Fourier transform and has frequency selectivity and multi-resolution perception ability. The multi-granularity time series encoding module architecture includes two parallel encoding paths: one is a Patch-based Transformer path, which focuses on capturing short-term features and local structures by dividing the time series into several time windows and modeling local attention; the other is a Standard Transformer path, which is used for global modeling and learning of long-term dependencies to extract trend information and overall dynamic changes in the sequence. The adaptive fusion completion module aligns the output features of different paths in the time dimension through adaptive pooling, ensures the semantic consistency and time sequence synchronization of the fusion results, and then dynamically weights and fuses the output features of the two encoding paths. For example, Figures 5 to 7 , the construction process of each module in step (3) specifically includes:

[0068] (3.1) Constructing the adaptive frequency domain feature extraction module, which is divided into two steps:

[0069] First, construct an adaptive frequency domain enhancement block. For the input observation value χ tA Fast Fourier Transform (FFT) is performed to convert the time-domain representation χ t into a frequency-domain representation F. Further, a power spectrum P is computed, and the frequency-domain representation F is filtered by an adaptive mask to retain the frequency-domain components of the power spectrum P that are greater than a trainable threshold θ, resulting in a filtered frequency-domain representation F filtered . Then, two sets of learnable filters, a global filter W G and a local filter W L , are applied to the original frequency-domain representation F and the filtered frequency-domain representation F enhanced , respectively. The results of the two sets of filters are added to obtain a final frequency-domain feature F filtered . Finally, an Inverse Fast Fourier Transform (IFFT) is performed to convert the enhanced frequency-domain feature F enhanced back into a time-domain representation χ t '. The computation process of the adaptive frequency-domain enhancement block is shown in the following equations:

[0070]

[0071] P = || F ||2 2

[0072] F filtered = F ⊙ (P > θ)

[0073] F enhanced = W G ⊙ F + W L ⊙ F filtered

[0074]

[0075] where N is the number of variables, T is the length of the input sequence, ||·||2 2 denotes the squared magnitude of each frequency component of the frequency-domain representation F, and ⊙ denotes element-wise multiplication.

[0076] In the second step, a cross-scale feature interaction block is constructed. This block is composed of two parallel convolutional layers that use different kernel sizes. The convolutional layer Conv small uses a small kernel size (e.g., kernel size = 1) and focuses on extracting fine-grained local features in the time series. The convolutional layer Conv big uses a large kernel size (e.g., kernel size = 3) and focuses on capturing long-term dependencies in the sequence. The input data χ tThe features are fed into parallel convolutional layers, and the outputs of the two convolutional layers interact through element-wise multiplication (Hadamard product) to enhance the feature representation. The two interacting features are then fused and passed through a Convolutional layer. fusion The process is then refined to obtain the final output features. The calculation process for cross-scale feature interaction blocks is shown in the formula:

[0077] A ˋ1 =GELU(Conv small (χ t ′))⊙Conv big (χ t ′)

[0078] A2 = GELU(Conv) big (χ t ′))⊙Conv small (χ t ′)

[0079] χ t "=Conv fusion (A ˋ1 +A ˋ2 )

[0080] GELU stands for Gaussian Error Linear Unit.

[0081] (3.2) Constructing a multi-granularity timing coding module involves two steps:

[0082] The first step is to construct a short-term temporal embedding code. First, short-term temporal embedding is performed using a patch-based embedding layer. Specifically, the original input with a batch size of B, a length of T, and N variables is... The time step is divided into multiple local blocks (Patches), each of length L, and there are a total of Each block, after being divided Then, a linear layer is used to map it to a high-dimensional space H, and finally, learnable positional encoding is added to model the order relationship between patches, resulting in a representation. where d is the embedding dimension. Second, we use a Patch-based Encoder to encode the short-term time series. Specifically, the Patch-based Encoder contains multiple stacked Transformer encoder modules. Each encoder module is composed of four parts: self-attention mechanism (Self-Attention), feed-forward network (FFN) residual connection, and one-dimensional batch normalization (BatchNorm1d). Among them, the cross-time step normalization normalizes each feature channel in the time dimension. For the input tensor First, we transpose it to make the value of each channel unfold in the time dimension. Then, in the unit of feature channel, we use one-dimensional batch normalization to calculate its mean and variance on all samples and time steps for normalization, and finally transpose it back to the original shape. The cross-time step normalization is calculated by the following formula:

[0083] BN T (x)=Transpose -1 °BatchNorm1d°Transpose(x)

[0084] where Transpose -1 is the inverse transpose transformation, BatchNorm1d is one-dimensional batch normalization, Transpose is the transpose transformation, and ° is the function composition operator.

[0085] The short-term time series embedding encoding formula is as follows:

[0086]

[0087] FFN(H)=ReLU(Linear1(H))·Linear2(H)

[0088] H′=BN T (H+Self-Attention(H))

[0089] H short =BN T (H′+FFN(H′))

[0090] where is the high-dimensional space embedding feature, Q, K, and V are the query (Query), key (Key), and value (Value) vectors, respectively, d k is the dimension of the key vector, Linear1 and Linear2 represent linear layers, and ReLU represents the rectified linear unit (ReLU).

[0091] The second step is to construct long-term temporal embedding encoding. First, long-term temporal embedding is performed using a Standard Embedding layer. Specifically, each time series is independently embedded as a variable label. Specifically, the original input X is mapped from the time dimension to a high-dimensional space H through a linear layer. Second, long-term temporal encoding is performed using a Standard Encoder. Specifically, the Standard Encoder contains multiple stacked Transformer encoder modules. Each encoder module consists of four parts: a self-attention mechanism, feedforward network residual connections, and layer normalization. Layer normalization calculates the mean μ of the feature vector at each time step for each sample along dimension d. b,t and variance They are then used to normalize the vectors, thereby helping the model training to be more stable. The calculation formula is as follows:

[0092]

[0093] Where, x b,t,i Let μ be the embedding feature of the b-th sample in the batch at the t-th time step in the i-th dimension. b,t and respectively Let be the feature mean and variance of the b-th sample at time step t, d be the embedding dimension, and ∈ be a stabilizing term to prevent division by zero.

[0094]

[0095] FFN(H)=ReLU(Linear1(H))·Linear2(H)

[0096] H′=LayerNorm(H+Attention(H))

[0097] H long =LayerNorm(H′+FFN(H′))

[0098] in, For high-dimensional space embedding features, Q, K, and V are the query, key, and value vectors, respectively, and d k is the dimension of the key vector.

[0099] (3.3) Constructing an adaptive fusion completion module: Since encoders for different paths use different modeling granularities, the encoded features differ in the time dimension. Adaptive average pooling is used to uniformly adjust the output of longer paths to the target time length. Specifically, for long sequences Adaptive average pooling divides it evenly into a sub-interval, and calculate the mean value on each sub-interval, and finally output the result A learnable fusion weight parameter a e (0, 1) is introduced to adaptively adjust the contribution of the two paths to the final completion result, and the fused result is H fused = aH short +(1-a)H' long Finally, the high-dimensional features are restored to the predicted values in sequence form using the linear mapping head (FlattenHead). This module includes two steps of feature flattening (Flatten) and linear projection (Linear Projection):

[0100]

[0101] (4) Use the model to train and infer to obtain the new energy missing data completion value, and test the model result.

[0102] In this embodiment, the model performance quantification evaluation is selected on R 2 , MSE, RMSE, MAE and MAPE five indexes, wherein: R 2 (determination coefficient) is an index used to measure the fitting degree of the data prediction model, measures the fitting degree of the regression model to the data, and explains the proportion of the variation of the target variable. R 2 closer to 1, the better the model fitting effect. MSE (Mean Square Error) is an index used to measure the fitting degree of the data prediction model, and the calculation method is to square the difference between the predicted value and the actual value and then take the average. The smaller the MSE, the smaller the difference between the predicted value and the actual value, and the better the fitting degree of the model. RMSE (Root Mean Square Error) is the square root of MSE, which is sensitive to outliers and can reflect the distribution of prediction errors. MAE (Mean Absolute Error) is an index used to measure the accuracy of a prediction model or estimation method in statistics, which reflects the average level of absolute error between predicted values and true values, and the smaller the MAE, the better. MAPE (Mean Absolute Percentage Error) is an index for measuring prediction accuracy, which is obtained by calculating the absolute value of the percentage error between the predicted value and the actual value.

[0103] To evaluate the prediction effect of the model (AFMFormer) of the present embodiment, the following six time series completion prediction models are selected: 1) Transformer: a neural network that relies on self-attention mechanism to capture long-range dependencies in sequences; 2) iTransformer: a Transformer structure that adopts variable separation modeling, effectively enhancing the structural perception ability between variables and improving the modeling effect of multivariate time series; 3) PatchTST: a Transformer structure based on sliding window to divide the time series into multiple local patches, using the self-attention mechanism within the patch to model short-term information, and maintaining local temporal consistency between patches, suitable for modeling short-term local fluctuation characteristics; 4) Informer: a Transformer variant optimized for long sequence time series, using ProbSparse attention mechanism to reduce the complexity of traditional self-attention mechanism; 5) FreTS: a neural network that converts time series from time domain to frequency domain, using frequency domain multilayer perceptron to capture global dependencies; 6) DLiner: a minimalist linear model, the core is a single linear layer, directly regressing the history sequence, avoiding the information loss and computational overhead brought by the self-attention mechanism of Transformer, and having obvious advantages in long sequence tasks with significant error accumulation.

[0104] Comparative experiment: using the UEBB dataset to test the AFMFormer and the other six mainstream time series completion models, the five evaluation indicators are shown in Table 3. AFMFormer achieves the best in the five indicators, because the model realizes dynamic filtering of dominant frequency components and deep interaction of cross-scale features through the adaptive frequency domain feature extraction module (AFEB+CSFIB), combined with the Patch-based and Standard Transformer dual-path structure, realizing the collaborative modeling of short-term local changes and long-term trends, significantly improving the structural decoupling ability and expression accuracy of complex time series structure.

[0105] Table 3 Completion accuracy under different missing rates

[0106]

[0107] Ablation experiment: AFMFormer is composed of an adaptive frequency domain feature extraction module and a multi-granularity time series completion backbone network. To verify the effectiveness of each part, the following models are designed: 1) Base; 2) Patch-only Base; 3) Standard-only Base; 4) AFMFormer. Among them, the Base model is a time series completion backbone network (without the adaptive frequency domain feature extraction module). Patch-only Base is a Base model without long-term time series embedding coding. Standard-only Base is a Base model without short-term time series embedding coding.

[0108] Table 4 Ablation experiment

[0109]

[0110] The ablation experiment results show that after introducing the adaptive frequency domain feature extraction module (AFDFEM) on the basis model (Base), AFMFormer shows significant improvement in each index. This result shows that the AFDFEM module significantly enhances the model's ability to model multi-scale time series structures in new energy data by performing data-driven frequency domain decomposition and dominant frequency extraction on the input sequence, especially in the decoupling of long-term trends and short-term fluctuations. Removing any coding path will cause a significant decline in performance. The Patch-only Base model has some ability to model short-term fluctuations, but compared with the complete fusion model, its R 2 dropped by 5.07%, and RMSE and MAE increased to 0.7063 and 0.4869, respectively, with a significant performance degradation; similarly, the Standard-only Base model performs poorly in modeling long-term dependencies. This result shows that the Patch-based path and the Standard path have obvious complementarity in modeling ability, and any single path cannot independently perform the task of modeling complex multi-scale time series features.

[0111] Based on the same inventive concept, the embodiment of the present application discloses a new energy missing data completion system based on multi-model fusion, comprising: a feature screening module for linear and nonlinear correlation two-stage feature screening of the collected new energy data set; a data preprocessing module for normalizing the data set, constructing a missing data mask according to a missing proportion parameter, and dividing the data set; a model construction module for constructing a fusion model AFMFormer including an adaptive frequency domain feature extraction module, a multi-granularity time sequence encoding module and an adaptive fusion completion module; wherein the adaptive frequency domain feature extraction module is designed based on an improved Fourier transform and has frequency selectivity and multi-resolution perception ability, realizes adaptive frequency spectrum decomposition, extracts the dominant frequency component in the sequence and suppresses high-frequency noise; the multi-granularity time sequence encoding module includes two parallel encoding paths: one is a Patch-based Transformer path, which divides the time sequence into a plurality of time windows and models local attention; the other is a Standard Transformer path, which is used to extract trend information and overall dynamic changes in the sequence; the adaptive fusion completion module is used to align the output features of different paths in the time dimension, and dynamically weight and fuse the output features of the two encoding paths; a model training module is used to train the AFMFormer model, and the trained model is used for missing data imputation to output the missing data completion result.

[0112] The specific working processes of the above-described modules can refer to the corresponding processes in the foregoing method embodiments, and will not be described here. The division of the modules is only a logical functional division, and another division mode can be used in actual implementation, for example, a plurality of modules can be combined or integrated into another system.

[0113] The embodiment of the present application also discloses a computer system, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the computer program is executed by the processor to realize the steps of the new energy missing data completion method based on multi-model fusion.

[0114] The embodiment of the present application also discloses a computer readable storage medium, which stores a computer program, wherein the computer program is executed by a processor to realize the steps of the new energy missing data completion method based on multi-model fusion.

[0115] The embodiment of the present application also discloses a computer program product, comprising a computer program, wherein the computer program is executed by a processor to realize the steps of the new energy missing data completion method based on multi-model fusion.

[0116] Program code for carrying out the methods of the present application can be written in any combination of one or more programming languages. The program code can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the program code, when executed by the processor or controller, causes the steps of the methods of the present application to be implemented. The program code can execute entirely on a machine, partly on the machine, as a stand-alone software package, partly on the machine and partly on a remote machine or entirely on the remote machine or server. The present application is not limited by the

[0117] While the present application has been illustrated and described in detail in the preferred embodiments, the skilled in the art will appreciate that the application can be modified and varied without departing from the scope of the application as defined by the appended claims.

Claims

1. A new energy missing data completion method based on multi-model fusion, characterized in that, The method comprises the following steps: Two-stage feature screening of linear and nonlinear correlation is performed on the collected new energy data set; The data set is normalized, a missing data mask is constructed according to a missing proportion parameter, and the data set is divided; A fusion model AFMFormer including an adaptive frequency domain feature extraction module, a multi-granularity time sequence encoding module and an adaptive fusion completion module is constructed; The adaptive frequency domain feature extraction module is designed based on an improved Fourier transform, has frequency selectivity and multi-resolution perception capability, realizes adaptive frequency spectrum decomposition, extracts dominant frequency components in the sequence and suppresses high-frequency noise; The multi-granularity time sequence encoding module includes two parallel encoding paths: one is a Patch-based Transformer path, which divides the time sequence into a plurality of time windows and models local attention; the other is a Standard Transformer path, which is used to extract trend information and overall dynamic changes in the sequence; The adaptive fusion completion module is used to align the output features of different paths in the time dimension, and dynamically weight and fuse the output features of the two encoding paths; The AFMFormer model is trained, and the trained model is used for missing data interpolation to output the missing data completion result. The feature screening includes: calculating the Pearson correlation coefficient and the maximum mutual information coefficient between the features in the collected new energy data set; if the correlation between any two features does not reach a preset threshold, it is considered that the data is redundant, and the features with small correlation with the target variable are ignored.

2. The new energy missing data completion method based on multi-model fusion according to claim 1, characterized in that, The adaptive frequency domain feature extraction module includes an adaptive frequency domain enhancement block and a cross-scale feature interaction block.

3. The new energy missing data completion method based on multi-model fusion according to claim 1, characterized in that, The adaptive frequency domain enhancement block introduces a learnable threshold parameter after performing fast Fourier transform on the input value, which is used to filter the high-energy region in the frequency domain; after filtering the main frequency domain information, adaptive frequency spectrum enhancement is performed through local and global dual-channel learning filters; The cross-scale feature interaction block inputs the data after adaptive frequency domain enhancement into two parallel convolution layers with different size convolution kernels, the outputs of the two convolution layers are interacted by element-wise multiplication to enhance the feature representation; the above two interacted features are fused and processed through a fusion convolution layer to obtain the final output feature. The calculation process of the adaptive frequency domain enhancement block is as follows:

4. The new energy missing data completion method based on multi-model fusion according to claim 3, characterized in that, The Standard Transformer path first uses a Standard Embedding layer to map the time dimension of the original input to a high-dimensional space through a linear layer; then the high-dimensional space representation is encoded by a Standard Encoder; the Standard Encoder includes a plurality of stacked standard Transformer encoder modules, each encoder module is composed of four parts: self-attention mechanism, feedforward network, residual connection and standard layer normalization; the multi-granularity time sequence encoding module encodes the input in parallel through the above two paths. P = ||F|| 2 F filtered = F O (P > θ) F enhanced = W G ⊙ F + W L ⊙ F filtered where χ t is the input observation, denotes the fast Fourier transform, F is the frequency domain representation, P is the power spectrum, θ is a trainable threshold, F filtered is the filtered frequency domain representation, W G is the global filter, W L is the local filter, χ t ' is the time series representation obtained by the inverse fast Fourier transform, N is the number of variables, T is the input sequence length, ‖·‖ 2 denotes the amplitude squared of each frequency component of the frequency domain representation F, and denotes the element-wise multiplication.

5. The new energy missing data completion method based on multi-model fusion according to claim 1, characterized in that, In the multi-granularity time series encoding module, the Patch-based Transformer path first uses a Patch-based Embedding layer to convert the input time series data into a high-dimensional space is divided into a plurality of local blocks (Patch) is mapped to a high-dimensional space through a linear layer, where B is the batch size, T is the sequence length, N is the number of variables, and L is the length of each block; then a learnable position encoding is added to model the sequential relationship between the patches, obtaining the representation where d is the embedding dimension; the high-dimensional space representation is then encoded by a Patch-based Encoder; the Patch-based Encoder includes a plurality of stacked Patch-based Transformer encoder modules, each of which is composed of a self-attention mechanism, a feedforward network, a residual connection, and a one-dimensional batch normalization; the one-dimensional batch normalization is to transpose and unfold the input in the time dimension, apply one-dimensional batch normalization to each feature channel, and realize standardization processing across time steps; ​ 6. The new energy missing data completion method based on multi-model fusion according to claim 1, characterized in that, The adaptive fusion completion module uses an adaptive average pooling operation to uniformly divide the sequence obtained by the Standard Transformer path, calculates the mean on each sub-interval, and uniformly adjusts the output of the Standard Transformer path to the sequence length of the output of the Patch-based Transformer path; and on this basis, a learnable fusion weight parameter is introduced to adaptively adjust the contribution of the two aligned encoding paths to the final completion result, and then the parameter is used to fuse the feature representation; finally, a linear mapping head is used to perform feature flattening and linear mapping on the fused features, and the high-dimensional features are restored to the predicted values in sequence form.

7. A new energy missing data completion system based on multi-model fusion, characterized in that, Comprise: The feature screening module is used for two-stage feature screening of linear and nonlinear correlation of the collected new energy data set; The data preprocessing module is used for normalizing the data set, constructing a missing data mask according to a missing proportion parameter, and dividing the data set; The model construction module is used for constructing a fusion model AFMFormer including an adaptive frequency domain feature extraction module, a multi-granularity time series encoding module and an adaptive fusion completion module; Wherein The adaptive frequency domain feature extraction module is designed based on an improved Fourier transform, has frequency selectivity and multi-resolution perception ability, realizes adaptive frequency spectrum decomposition, extracts dominant frequency components in the sequence and suppresses high-frequency noise; The multi-granularity time series encoding module includes two parallel encoding paths: one is a Patch-based Transformer path, which divides the time sequence into several time windows and models local attention; the other is a Standard Transformer path, which is used to extract trend information and overall dynamic changes in the sequence; The adaptive fusion completion module is used for aligning the output features of different paths in the time dimension, and dynamically weighting and fusing the output features of the two encoding paths; The model training module is used for training the AFMFormer model, and the trained model is used for missing data imputation to output the missing data completion result.

8. A computer system comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The computer program is executed by the processor to realize the steps of the new energy missing data completion method based on multi-model fusion according to any one of claims 1-6.

9. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 8. The computer program is executed by the processor to realize the steps of the new energy missing data completion method based on multi-model fusion according to any one of claims 1-6.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the new energy missing data completion method based on multi-model fusion according to any one of claims 1-6.

Images