Intelligent interpolation method and system for wind and light observation data of new energy station

Abstract

The application discloses a new energy station wind and light element observation data intelligent interpolation method and system, and belongs to the technical field of new energy observation data processing. The intelligent interpolation method comprises the following steps: acquiring a non-missing station data set; based on the non-missing station data set, constructing station missing simulation data by using an adaptive random mask mechanism; constructing a heterogeneous neural network model based on latent space feature alignment and training the constructed heterogeneous neural network model by using a two-stage training strategy; based on the trained heterogeneous neural network model, interpolating and correcting the missing observation data of the station, so as to interpolate the predicted value of a specified wind and light element at a specified wind and light element observation value missing new energy station site, and then outputting the non-missing station interpolation reconstruction data, and completing the interpolation processing of the missing observation data of the station. It can be seen that the application solves the problems that the prior art is difficult to process multiple intensity missing under irregular sampling and the spatial physical consistency of the interpolation result is poor.

Description

Technical Field

[0001] This invention relates to an intelligent interpolation method and system for wind and solar element observation data of new energy power stations, belonging to the technical field of new energy observation data processing. Background Technology

[0002] With the rapid development of the new energy industry, photovoltaic power plants, wind farms, and other new energy power stations are usually equipped with meteorological observation equipment to monitor key wind and solar elements such as irradiance and wind speed in real time. However, due to equipment failure, communication interruptions, extreme weather, and other reasons, observation data is often missing, which seriously affects the accuracy of new energy power prediction and the safety of power grid dispatch.

[0003] Existing data interpolation methods are mainly divided into two categories: traditional statistical interpolation and simple machine learning.

[0004] Traditional methods, such as linear interpolation and Kriging interpolation, only utilize observation data from a small number of new energy power stations around the missing sites (i.e., the sites where observation data is missing), making it difficult to handle scenarios with large-scale or continuous missing data.

[0005] In recent years, some studies have attempted to use deep learning methods for data completion. However, their training data often uses a fixed-ratio random mask, which cannot simulate the collective failure of regional observation stations caused by local extreme weather. The model architecture is mostly a single encoder-decoder structure, which does not decouple "global field distribution learning" from "incomplete feature mapping", resulting in insufficient physical consistency and local realism in the interpolation results.

[0006] Therefore, how to efficiently utilize the spatiotemporal correlation characteristics in historical observation data to achieve accurate and robust interpolation of observation data from new energy power plants has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent interpolation method and system for wind and solar element observation data of new energy power stations, so as to solve the problems of difficulty in handling large-scale missing data, continuous missing data, and poor physical consistency of interpolation results in the existing technology.

[0008] To achieve the above-mentioned technical objectives, the present invention will adopt the following technical solution:

[0009] A method for intelligent interpolation of wind and solar energy element observation data at new energy power stations includes the following steps:

[0010] To construct a dataset of stations with no missing measurements for a specified wind and solar element, the following steps are taken: First, obtain station data for all new energy power stations within a specified historical period within a specified area. Second, acquire station data for all new energy power stations within a specified area, including the station locations and observation values ​​for the specified wind and solar element. Third, define a time period with no missing measurements for a specified wind and solar element as follows: within a specified area, if all new energy power stations have complete and unmissing observation values ​​for the specified wind and solar element at observation time t, then observation time t can be considered a time period with no missing measurements for the specified wind and solar element.

[0011] Based on the observation values ​​of the specified landscape elements corresponding to each time point without missing measurements in the missing measurement data of the missing measurement site, the adaptive masking mechanism is used to perform several random masking processes to generate multi-sample missing measurement simulation data of the missing measurement site for the specified landscape elements.

[0012] A heterogeneous neural network model based on latent space feature alignment is constructed and trained. The heterogeneous neural network model includes a global feature extraction path, a missing feature mapping path, and a global field reconstruction path. During training, a two-stage training strategy is adopted.

[0013] The first stage of training: Self-supervised training of the global feature extraction path and the global field reconstruction path is carried out using samples from the missing test site dataset. The global feature extraction path is used to encode the samples of the missing test site dataset into latent variables, and the global field reconstruction path is used to reconstruct the latent variables into the missing test site reconstructed dataset. After training, the latent variables output by the global feature extraction path are recorded as the alignment target latent variables.

[0014] The second stage of training involves fixing the parameters of the global field reconstruction path, using the station's missing simulation data as input, and the alignment target latent variable as the supervision label to train the missing feature mapping path so that the predicted latent variable output by the missing feature mapping path is aligned with the supervision label in the latent space.

[0015] Based on a trained heterogeneous neural network model, the missing observation data of the station at any interpolation time is input. After the corresponding prediction latent variables are output through the missing feature mapping path, and then restored through the global field reconstruction path, the predicted values ​​of the specified wind and solar elements can be interpolated at the site location of the new energy station where the observation values ​​of the specified wind and solar elements are missing. Then, the interpolated reconstruction data of the station without missing observation data is output, thus completing the interpolation processing of the missing observation data of the station.

[0016] Preferably, the adaptive masking mechanism dynamically extracts the mask ratio from a preset mask ratio range for each random mask, which is 10% to 50%.

[0017] The random masking process specifically refers to: for the same time period without missing data, randomly selecting several new energy power stations according to the extracted masking ratio; replacing the observed values ​​of specified wind and solar elements of the selected new energy power stations at that time period with missing identifiers to form a missing data simulation sample for that time period; repeating the random masking process several times for the same time period without missing data to obtain multiple missing data simulation samples for that time period; and the missing data simulation samples from all time periods without missing data together constitute multi-sample power station missing data simulation data.

[0018] Preferably, in the heterogeneous neural network model, the global feature extraction path and the global field reconstruction path together constitute a variational autoencoder framework based on the Swin-Transformer; wherein:

[0019] The global feature extraction path includes a first mapping layer and a Swin-Transformer-based neural network encoder;

[0020] The global field reconstruction path includes a second mapping layer and a Swin-Transformer-based neural network decoder;

[0021] The missing feature mapping path includes a third mapping layer and a feature extraction network based on Swin-Transformer;

[0022] During training, the first mapping layer maps the site locations of each new energy power station and the observation values ​​of specified wind and solar elements to regular spatial grid points based on the received intact measurement station dataset, and outputs a regular gridded feature field.

[0023] The neural network encoder is used to perform hierarchical downsampling and shift window self-attention calculation on the regular gridded feature field, extract multi-scale spatial features, output the distribution parameters of latent variables, and use reparameterization techniques to sample latent variables that characterize the global collaborative features of all stations.

[0024] The neural network decoder is used to perform hierarchical upsampling and feature reconstruction on the input latent variables, and outputs a regular grid feature map corresponding to the geographical range of the new energy power station;

[0025] The second mapping layer is used to map the regular grid feature map back to the discretely distributed site locations of each new energy power station, to obtain a reconstructed dataset of the power station without missing data.

[0026] The third mapping layer is used to map the station missing data simulation data from discrete station locations to regular spatial grid points, generating a regular gridded feature map; wherein: for missing observations in the station missing data simulation data, a missing value is set at the corresponding grid point location when mapping to regular spatial grid points;

[0027] Feature extraction networks are used to encode features at multiple scales from regular gridded feature maps and output predicted latent variables.

[0028] Preferably, in the first stage of training, the training objective is to minimize the following loss function:

[0029] ;

[0030] in: For reconstruction error, This is the KL divergence term;

[0031] Reconstruction error Calculated using the following formula:

[0032] ;

[0033] In the above formula, This represents the reconstructed value of the dataset from the missing measurement stations. This represents the observations in the dataset of stations with no missing data. Indicates the sample sequence number index; This represents the total number of samples.

[0034] Preferably, in the second stage of training, the Euclidean distance between the predicted latent variable output by the missing feature mapping path and the supervision label is used as the alignment loss function.

[0035] Preferably, the alignment loss function Represented as:

[0036] ;

[0037] In the above formula: Indicates the latent variable being predicted; Indicates a supervisory label; sg( ) indicates stopping the gradient operation.

[0038] Preferably, the wind and light elements are specified as wind speed or irradiance.

[0039] Another technical objective of this invention is to provide an intelligent interpolation system for wind and solar energy element observation data at new energy power stations, comprising:

[0040] The data acquisition module is used to acquire the station data corresponding to each uninterrupted measurement time for the specified wind and solar elements of each new energy station in the specified area within the specified historical period, so as to construct the uninterrupted measurement station dataset.

[0041] The simulation data generation module is used to perform several random masking processes on the missing sub-station data in the missing measurement station dataset using an adaptive masking mechanism to generate multi-sample station missing measurement simulation data.

[0042] The model building and training module is used to build and train a heterogeneous neural network model based on latent space feature alignment. The heterogeneous neural network model includes a global feature extraction path, a missing test feature mapping path, and a global field reconstruction path. A two-stage training strategy is adopted during training: in the first stage of training, the global feature extraction path and the global field reconstruction path are self-supervised using a missing test site dataset; in the second stage of training, the parameters of the global field reconstruction path are fixed, the missing test site simulation data is used as input, and the alignment target latent variable output after the first stage of training is used as the supervision label to train the missing test feature mapping path so that the predicted latent variable output by the missing test feature mapping path is aligned with the supervision label in the latent space.

[0043] The interpolation module is used to take the missing observation data of the station at any interpolation time as input, based on the trained heterogeneous neural network model, and output the corresponding prediction latent variables through the missing feature mapping path. Then, it is restored through the global field reconstruction path to interpolate the corresponding predicted values ​​of the specified wind and solar elements at the site locations of new energy stations where the observation values ​​of the specified wind and solar elements are missing, and output the interpolated reconstruction data of the missing station.

[0044] Preferably, the adaptive masking mechanism dynamically extracts the masking ratio for each random mask within the range of 10% to 50%.

[0045] Another technical objective of this invention is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned intelligent interpolation method for wind and solar element observation data at new energy power stations.

[0046] Based on the above-mentioned technical objectives, the present invention has the following advantages compared with the prior art:

[0047] 1. The intelligent interpolation method for wind and solar element observation data of new energy power stations described in this invention adopts an adaptive random masking mechanism (the masking ratio can be dynamically extracted from 10% to 50%) to perform multiple masking processes on the power station data corresponding to each time without missing measurements in the dataset of power stations without missing measurements. This can effectively simulate various intensities of missing measurements, from occasional failure of single-station sensors to large-area regional communication link failures, and significantly improve the robustness of the model in real complex missing measurement scenarios.

[0048] 2. The intelligent interpolation method for wind and solar element observation data of new energy power stations described in this invention uses a heterogeneous neural network model based on latent space feature alignment to decouple "global distribution learning" (global feature extraction path) from "incomplete feature mapping" (missing measurement feature mapping path). Through latent space alignment, the interpolation results are highly consistent with the real meteorological field in terms of physical magnitude and spatial manifold.

[0049] 3. The intelligent interpolation method for wind and solar element observation data of new energy power stations described in this invention adopts a two-stage training strategy when training heterogeneous neural network models. In the first stage of training, a stable latent space is constructed. In the second stage, the reconstruction path parameters are fixed and only the missing feature mapping path is trained. This ensures the stability of the latent space, reduces the difficulty of mapping learning, and improves the model convergence efficiency.

[0050] 4. The intelligent interpolation method for wind and solar element observation data of new energy power stations described in this invention follows the principle of "observation retention and missing measurement backfilling" during interpolation inference. It only interpolates for missing measurement locations (i.e., the site locations of new energy power stations where the observation values ​​of specified wind and solar elements are missing), retaining the original valid observation data, taking into account the authenticity and integrity of the data, and avoiding the contamination of the original valid observations. Attached Figure Description

[0051] Figure 1 This is a flowchart of the intelligent interpolation method for wind and solar element observation data of new energy power stations as described in this invention;

[0052] Figure 2 For the present invention Figure 1 A schematic diagram of the architecture of a heterogeneous neural network model based on latent space feature alignment;

[0053] Figure 3 This is a schematic diagram illustrating the architecture for interpolating missing observation data from a field station using the heterogeneous neural network model based on latent space feature alignment described in this invention, and outputting reconstructed data from the field station without missing data.

[0054] Figure 4 This is a comparison chart of wind speed interpolation effects at a certain station in an embodiment of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Unless otherwise specifically stated, the relative arrangement, expressions, and values ​​of components and steps set forth in these embodiments do not limit the scope of the present invention. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0056] Example 1

[0057] like Figure 1 As shown, the intelligent interpolation method for wind and solar element observation data of new energy power stations according to the present invention includes the following steps:

[0058] Step 1: Obtain the dataset of flawless test stations:

[0059] To construct a dataset of stations with no missing measurements for a specified wind and solar element within a specified historical period, the following steps are involved:

[0060] Step 1.1: Obtain station data for each renewable energy power station within a specified historical period within the designated area. The station data includes the station location (latitude and longitude coordinates) and renewable energy observation data (i.e., the observed values ​​of wind and solar elements) of each renewable energy power station at any observation time t, as well as the timestamp (i.e., the corresponding observation time), to construct a historical observation dataset for the power stations. .

[0061] New energy observation data should include at least wind speed and irradiance. In this invention, however, the new energy observation data only includes these two wind and solar elements. Therefore, the historical observation dataset from the site... In the formula, The coordinates of the new energy power station are represented by latitude and longitude (i.e., the location of the station), t represents the observation time, and v represents the wind speed; Indicates irradiance.

[0062] Step 1.2: Based on the historical observation dataset of the station Each scenic element k was measured without defects for a specified number of times. Filtering is performed to construct a complete dataset of survey stations for the corresponding landscape element k. When k is 1, it represents wind speed; when k is 2, it represents irradiance.

[0063] For the time without missing measurements of specified scenic elements (the specified scenic elements can be wind speed or irradiance), Defined as follows: If all new energy power stations within a specified area have complete and unmissing observations of the specified wind and solar elements at observation time t, then observation time t can be considered as the unmissing observation time for the specified wind and solar elements. .

[0064] By implementing steps 1.1 and 1.2 above, a complete dataset of survey stations for the specified landscape elements can be constructed.

[0065] For ease of explanation, wind speed will be used as an example below.

[0066] During the screening process, the first step is to determine the number of missing measurements, thereby compiling the historical observation dataset of the stations. The included data is divided into two categories according to the observation time: observations with no missing wind speed data and other observations. Observations with no missing wind speed data refer to the observation times when the wind speed data observed by all new energy power stations in the specified area are complete and without any missing data. Therefore, the data from each power station corresponding to the observations with no missing wind speed data can be used to construct a wind speed missing data station dataset. Similarly, a dataset of irradiance-free measurement stations can also be constructed. .

[0067] Step 2: Constructing simulation data for missing measurements at the site:

[0068] For missing measurement station datasets Missing station data at each missing time t An adaptive random masking mechanism is used for masking, mapping the data to the corresponding station-missing simulation data. ( (This indicates the number of renewable energy power stations retained for observation at that time), and uses data from power stations without missing data. For full space field labels and station missing measurement simulation data Training sample pairs are constructed for locally incomplete inputs. Thus, this step, by actively constructing a closed-loop sample library of "observation-masking-reconstruction," enables the model to simulate and learn the complex distribution patterns of missing data in real new energy observation data.

[0069] In this step, an adaptive random masking mechanism is used to process the data from the unmissing stations. Masking is performed to map the data to the corresponding station's missing measurement simulation data. For each time t without missing data, the mask ratio is dynamically extracted from the preset mask ratio range of 10%-50%; the upper limit of the mask ratio range is determined based on the maximum missing data rate of the station data without missing data, and the lower limit is set to 10%.

[0070] In this step, the random masking process specifically refers to: for the same time period t without missing data, randomly selecting several renewable energy power stations according to the extracted mask ratio; replacing the observation values ​​of specified wind and solar elements of the selected renewable energy power stations with missing data indicators at that time period, while leaving the observation values ​​of unselected renewable energy power stations unchanged. This processed data at that time period becomes a simulated missing data sample. To increase the diversity of training samples, for the same time period without missing data t, the above random masking process is repeated several times (the specific number can be set according to actual needs, for example, 5-20 times), each time independently re-extracting the mask ratio and selecting different combinations of renewable energy power stations, thereby obtaining multiple simulated missing data samples corresponding to that time period. Traversing all time periods without missing data, the multiple simulated missing data samples generated at each time period are aggregated together to constitute the multi-sample power station simulated missing data. This simulated data is used for supervised learning of the missing feature mapping path in the subsequent second-stage training.

[0071] Step 3: Construct a heterogeneous neural network model based on latent space feature alignment:

[0072] like Figure 2 As shown, this step constructs a heterogeneous neural network model based on latent space feature alignment, aiming to solve the problem of precise interpolation of new energy observation data under irregular sampling by decoupling spatial feature extraction and global field reconstruction. This heterogeneous neural network model includes a global feature extraction path, a global field reconstruction path, and a missing feature mapping path. The global feature extraction path and the global field reconstruction path constitute a variational autoencoder framework based on Swin-Transformer. Wherein:

[0073] A global feature extraction path is used to extract features from a dataset of undamaged test sites. Extract latent variables characterizing the global collaborative features of the stations. It includes a first mapping layer (Set ConvLayer1) and a Swin-Transformer-based neural network encoder. Specifically, spatial feature encoding is first performed through the first mapping layer (Set ConvLayer1), which receives the latitude and longitude coordinates of N new energy power stations as input. and its corresponding new energy observation data Kernel density estimation is used to map discrete new energy observation data on latitude and longitude coordinates to regular spatial grid points. A regular grid covering the geographical area of ​​new energy power stations is defined, with a grid resolution of H×W (H=64, W=64 in this example). The first mapping layer uses kernel density estimation to map discrete new energy observation data to regular spatial grid points: For any grid point (lon,lat) in the regular grid, its mapped eigenvalue F(lon,lat) is:

[0074] ;

[0075] Where K( ) represents the Gaussian kernel function, and b is the adaptive bandwidth (determined by the average distance between all renewable energy power stations). The output is the size. The regular gridded feature field, where H and W represent the height and width of the regular gridded feature field, respectively. This represents the number of feature channels at each grid point, in this example... =1. Then, this regular gridded feature field is input into a Swin-Transformer-based neural network encoder, which performs downsampling and multi-scale feature extraction through a hierarchical self-attention mechanism. Its specific components are as follows:

[0076] Stage 1: Divide the input size H×W×1 into 4×4 non-overlapping image blocks. Each image block is flattened and linearly mapped to a feature vector. The output size is... × × ,in =96.

[0077] Stage 2 consists of two Swin-Transformer blocks (STBs) and one downsampling layer. Each STB executes the following steps sequentially: Layer Normalization (LN) → Window Multi-Head Self-Attention (W-MSA) → Residual Connection → LN → Multilayer Perceptron (MLP, GELU Activation) → Residual Connection; the second STB uses Shift-Window Multi-Head Self-Attention (SW-MSA) to achieve cross-window information interaction. The downsampling layer concatenates and reduces the dimensionality of features within a 2×2 neighborhood, outputting a size... × × ,in =192.

[0078] Stage 3: Contains 6 STBs (structure as above) and one downsampling layer, output size × × ,in =384.

[0079] Stage 4: Contains 2 STBs (structure as above), no downsampling is performed, output size is preserved. × × .

[0080] The feature map output by the neural network encoder is then processed by global average pooling (GAP) and fed into two parallel fully connected layers, which output the mean of the latent variable distribution. and variance Subsequently, latent variables characterizing the global collaborative features of the stations were obtained through sampling. :

[0081] ;

[0082] in, This involves element-wise multiplication. ϵ is a distribution derived from the standard normal distribution. (mean is 0, covariance is the identity matrix) The noise vector randomly sampled from the middle.

[0083] The global field reconstruction path includes a Swin-Transformer-based neural network decoder and a second mapping layer (SetConvLayer2) used to reconstruct the latent variables. Reconstruct the dataset by restoring it to a complete set of measurement stations (i.e., the attached dataset) Figure 2 (Reconstructed dataset of missing monitoring stations in the data).

[0084] Specifically, in the global field reconstruction path, latent variables will be... The input is a Swin-Transformer-based neural network decoder, which employs a hierarchical upsampling structure symmetrical to the neural network encoder. After upsampling and feature reconstruction, the output is a regular grid feature map corresponding to the geographical range of the new energy power stations. The second mapping layer maps the regular grid feature map to the latitude and longitude coordinates of each new energy power station, thus obtaining the reconstructed dataset of the power stations without missing data. For the i-th new energy power station, its reconstructed value is... The pixel values ​​at the corresponding locations of the stations were extracted from the regular grid feature map using bilinear interpolation. The reconstructed values ​​of all new energy power stations constituted the incomplete test station reconstruction dataset. .

[0085] For any new energy power station i, let its latitude and longitude coordinates be... The longitude range covered by the regular grid is The latitude range is The grid resolution is H×W. First, calculate the normalized coordinates of the site in the regular grid:

[0086] ; ;

[0087] Its reconstructed value It can be represented as:

[0088] ;

[0089] ; ; , , , Let be the feature values ​​of the four grid points surrounding the corresponding position of the i-th new energy power station in the regular grid feature map.

[0090] The missing data feature mapping path, including the third mapping layer (Set ConvLayer3) and a Swin-Transformer-based feature extraction network, is used to extract predictive latent variables from the station's missing data simulation data. .

[0091] Specifically, the structure of the third mapping layer is exactly the same as that of the first mapping layer, which incorporates the missing simulation data. The feature map is mapped to a regular grid of size H×W×1. During kernel density estimation, observations corresponding to missing sites are not included in the calculation (assigned zero weights), thus naturally forming low-confidence regions in the grid features. The Swin-Transformer feature extraction network uses the exact same hierarchical structure as the neural network encoder in the global feature extraction path (i.e., Stage 1 to Stage 4, containing the same number of STBs, channels, and downsampling operations), but the output features are not parameterized. Specifically, the final output size of the feature extraction network is... × × The feature maps, after global average pooling, are fed into a fully connected layer to directly output the predicted latent variables. .

[0092] Step Four: Two-Phase Training

[0093] Phase 1: Self-supervised training of the global feature extraction path and global field reconstruction path is performed using samples from the incomplete test site dataset. The training objective is to minimize the following loss function:

[0094] ;

[0095] in: For reconstruction error, This is the KL divergence term.

[0096] Reconstruction error Dataset reconstruction using missing measurement stations Compared with the original intact survey station dataset The mean square error (MSE) between them is:

[0097] ;

[0098] KL divergence term Used to constrain the latent variable distribution q(c|X) to the standard normal distribution. The difference between them is specifically expressed using the standard KL divergence calculation formula for variational autoencoders. Where: the latent variable distribution q(c|X) represents the distribution of latent variables given input X. c The posterior distribution. middle, Represents the mean vector. This represents the unit covariance matrix.

[0099] By minimizing the loss function Force latent variables It can accurately reconstruct complete station data (due to reconstruction error) This ensures that a complete reconstructed dataset of the measurement stations is obtained, while also maintaining a continuous and regular probability structure (derived from the KL divergence term). (to ensure), thereby building a high-quality potential space.

[0100] The training parameters were set to 200 iterations, a batch size of 32, and an initial learning rate of 1e-3, enabling the model to accurately represent the spatial topological evolution of wind speed and irradiance. At this stage, the model grasped the "background characteristics" of the full field distribution of new energy power stations. The reconstruction error was calculated using the mean squared error between the reconstructed dataset and the dataset containing all missing measurements.

[0101] The second stage involves fixing all parameters of the Swin-Transformer-based neural network decoder and the second mapping layer in the global field reconstruction path. Since this path has already learned the precise mapping from latent variables to complete observations, fixing these parameters ensures the latent space geometry remains unchanged, allowing the missing-measure feature mapping path to focus on learning latent variable alignment. Specifically, during the joint training of the model, the latent variables output by the global feature extraction path are... For supervised labeling, the training objective is to predict the latent variables output by the missing feature mapping path. Align the supervision label in the latent space The alignment loss function used is based on the prediction of latent variables. Constructed using the Euclidean distance between the supervisory label and the supervisory label, it can be represented as:

[0102] ;

[0103] Among them: sg( () indicates stopping the gradient operation, i.e., supervision label. The gradient from the alignment loss is not accepted. Through this alignment term, the missing feature mapping path can predict latent variables consistent with complete observations from incomplete observations, thereby establishing a closed loop of mapping from local incompleteness to global field features in the latent space.

[0104] In the second training phase, the latent variable c output from the first phase is used as the supervision label to train the missing feature mapping path, minimizing its output predicted latent variable. The Euclidean distance between the model and the supervised labels is used for 100 iterations, with a batch size of 32 and an initial learning rate of 5e-4, to align the incomplete features with the global features in the latent space. The core significance of this process lies in establishing an "information compensation" logic, enabling the model to learn to accurately retrieve and activate the corresponding global feature parameters in the latent space using a Swin-Transformer-based feature extraction network, even with only partial field observation data.

[0105] Step 5: Interpolate and correct missing observation data from the field stations:

[0106] Based on a pre-trained heterogeneous neural network model, inputting the station's missing observation data (i.e., reference data) at any interpolation time. Figure 3 The missing observation data from the stations is used to output the corresponding predicted latent variables through the missing feature mapping path. After being restored through the global field reconstruction path, the predicted values ​​of the specified wind and solar elements can be interpolated at the site locations of the new energy power stations where the observed values ​​of the specified wind and solar elements are missing. This results in the output of the missing station interpolated and reconstructed data (i.e., the missing station data). Figure 3 (Reconstructing data from missing stations), completing the interpolation of missing station observation data.

[0107] Specifically, the missing observation data of the power stations at the interpolation time (i.e., in the new energy observation data obtained at the interpolation time, there are some new energy power stations with missing observation data of specified wind and solar elements) is input into the pre-trained missing feature mapping path to output the corresponding predicted latent variables. The predictor latent variable The system directly inputs the global field reconstruction path, driving the neural network decoder within it to generate predicted values ​​for specified wind and solar elements. These predicted values ​​are then spatially mapped through a second mapping layer, transforming them into predicted values ​​perfectly aligned with the latitude and longitude coordinates of all new energy power stations, thus obtaining the corresponding missing-measurement-free reconstructed dataset. In the final interpolation and correction stage, the system adopts the principle of "observation retention and missing measurement backfilling": for station data present in the original missing-measurement observation data, the original observation values ​​are retained to maintain local authenticity; for new energy power stations marked as missing, the predicted values ​​of the specified wind and solar elements at the corresponding positions in the aforementioned missing-measurement-free reconstructed dataset are extracted for precise filling, outputting missing-measurement-free interpolated reconstructed data, completing the interpolation processing of the missing-measurement observation data. Through this logic of "feature inference - global reconstruction - local correction," this invention achieves intelligent completion of new energy power station sequences. The output missing-measurement-free interpolated reconstructed data not only fills the spatiotemporal gaps but also maintains a high degree of consistency with the real meteorological field in terms of physical magnitude and spatial manifold.

[0108] Application examples

[0109] Using 15-minute anemometer wind speed observation data from 163 renewable energy power plants in a certain region in 2024 as the subject, the method described in this invention was used for data interpolation. Taking one of the renewable energy power plants as an example, the original missing data rate of this station was 38.46%, which seriously affected the integrity of the wind speed sequence and the accuracy of subsequent power prediction. After applying the method of this invention, the interpolation integrity rate reached 100%. Figure 4 The figure shows a comparison of the interpolation results for this station from March 1st to March 7th, 2024. The black dots represent the original observations, the green solid line represents the complete sequence after interpolation, and the gray shaded area marks the original missing periods. It can be seen that for sporadic missing data (such as 11:45 on March 1st), the interpolated values ​​smoothly connect the preceding and following observations without significant abrupt changes; for large periods of missing data lasting several hours (such as 22:15 on March 3rd to 08:30 on March 4th), the interpolation method described in this invention can still output a complete curve conforming to physical laws based on historical meteorological patterns and spatial correlation characteristics, verifying the robustness of the interpolation method described in this invention under extreme missing data scenarios.

[0110] Example 2

[0111] This embodiment provides an intelligent interpolation system for wind and solar element observation data of new energy power stations, used to implement the above-mentioned intelligent interpolation method for wind and solar element observation data of new energy power stations, including:

[0112] The data acquisition module is used to acquire the data of each new energy power station in a specified area for each time period without missing test within a specified historical period, so as to construct a dataset of power stations without missing test.

[0113] The simulation data generation module is used to perform multiple random masking processes on the station data in the missing test station dataset using an adaptive masking mechanism to generate multi-sample station missing test simulation data.

[0114] The model building and training module is used to build and train a heterogeneous neural network model based on latent space feature alignment. The heterogeneous neural network model includes a global feature extraction path, a missing test feature mapping path, and a global field reconstruction path. A two-stage training strategy is adopted during training: the first stage uses a dataset of missing test sites to perform self-supervised training on the global feature extraction path and the global field reconstruction path; the second stage fixes the parameters of the global field reconstruction path, uses the missing test simulation data of the test sites as input, and uses the latent variables output by the first stage as labels to train the missing test feature mapping path, so that the predicted latent variables output by the missing test feature mapping path are aligned with the latent variables output by the global feature extraction path.

[0115] The interpolation module is used to take the missing observation data of the station at any interpolation time as input, based on the trained heterogeneous neural network model, and output the corresponding prediction latent variables through the missing feature mapping path. Then, it is restored through the global field reconstruction path to interpolate the corresponding predicted values ​​of the specified wind and solar elements at the site locations of new energy stations where the observation values ​​of the specified wind and solar elements are missing, and output the interpolated reconstruction data of the missing station.

[0116] Example 3

[0117] The present invention also provides a storage medium, wherein the computer program stored in the storage medium executes the above-mentioned intelligent interpolation method for wind and solar element observation data of new energy power stations when it is running.

[0118] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the above-described intelligent interpolation method for wind and solar element observation data of new energy power stations through the computer program.

[0119] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0120] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0121] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0122] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0123] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0124] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0125] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent interpolation of wind and solar element observation data at new energy power stations, characterized in that, Includes the following steps: To construct a dataset of stations with no missing measurements for a specified wind and solar element, the following steps are taken: First, obtain station data for all new energy power stations within a specified historical period within a specified area. Second, acquire station data for all new energy power stations within a specified area, including the station locations and observation values ​​for the specified wind and solar element. Third, define a time period with no missing measurements for a specified wind and solar element as follows: within a specified area, if all new energy power stations have complete and unmissing observation values ​​for the specified wind and solar element at observation time t, then observation time t can be considered a time period with no missing measurements for the specified wind and solar element. Based on the observation values ​​of the specified landscape elements corresponding to each time point without missing measurements in the missing measurement data of the missing measurement site, the adaptive masking mechanism is used to perform several random masking processes to generate multi-sample missing measurement simulation data of the missing measurement site for the specified landscape elements. A heterogeneous neural network model based on latent space feature alignment is constructed and trained. The heterogeneous neural network model includes a global feature extraction path, a missing feature mapping path, and a global field reconstruction path. During training, a two-stage training strategy is adopted. The first stage of training: Self-supervised training of the global feature extraction path and the global field reconstruction path is carried out using samples from the missing test site dataset. The global feature extraction path is used to encode the samples of the missing test site dataset into latent variables, and the global field reconstruction path is used to reconstruct the latent variables into the missing test site reconstructed dataset. After training, the latent variables output by the global feature extraction path are recorded as the alignment target latent variables. The second stage of training involves fixing the parameters of the global field reconstruction path, using the station's missing simulation data as input, and the alignment target latent variable as the supervision label to train the missing feature mapping path so that the predicted latent variable output by the missing feature mapping path is aligned with the supervision label in the latent space. Based on a trained heterogeneous neural network model, the missing observation data of the station at any interpolation time is input. After the corresponding prediction latent variables are output through the missing feature mapping path, and then restored through the global field reconstruction path, the predicted values ​​of the specified wind and solar elements can be interpolated at the site location of the new energy station where the observation values ​​of the specified wind and solar elements are missing. Then, the interpolated reconstruction data of the station without missing observation data is output, thus completing the interpolation processing of the missing observation data of the station.

2. The intelligent interpolation method for wind and solar element observation data of new energy power stations according to claim 1, characterized in that, The adaptive masking mechanism dynamically extracts the mask ratio from a preset mask ratio range for each random mask, which is 10% to 50%. The random masking process specifically refers to: for the same time period without missing data, randomly selecting several new energy power stations according to the extracted masking ratio; replacing the observed values ​​of specified wind and solar elements of the selected new energy power stations at that time period with missing identifiers to form a missing data simulation sample for that time period; repeating the random masking process several times for the same time period without missing data to obtain multiple missing data simulation samples for that time period; and the missing data simulation samples from all time periods without missing data together constitute multi-sample power station missing data simulation data.

3. The intelligent interpolation method for wind and solar element observation data of new energy power stations according to claim 1, characterized in that, In the heterogeneous neural network model, the global feature extraction path and the global field reconstruction path together constitute a variational autoencoder framework based on the Swin-Transformer; wherein: The global feature extraction path includes a first mapping layer and a Swin-Transformer-based neural network encoder; The global field reconstruction path includes a second mapping layer and a Swin-Transformer-based neural network decoder; The missing feature mapping path includes a third mapping layer and a feature extraction network based on Swin-Transformer; During training, the first mapping layer maps the site locations of each new energy power station and the observation values ​​of specified wind and solar elements to regular spatial grid points based on the received intact measurement station dataset, and outputs a regular gridded feature field. The neural network encoder is used to perform hierarchical downsampling and shift window self-attention calculation on the regular gridded feature field, extract multi-scale spatial features, output the distribution parameters of latent variables, and use reparameterization techniques to sample latent variables that characterize the global collaborative features of all stations. The neural network decoder is used to perform hierarchical upsampling and feature reconstruction on the input latent variables, and outputs a regular grid feature map corresponding to the geographical range of the new energy power station; The second mapping layer is used to map the regular grid feature map back to the discretely distributed site locations of each new energy power station, to obtain a reconstructed dataset of the power station without missing data. The third mapping layer is used to map the station missing data simulation data from discrete station locations to regular spatial grid points, generating a regular gridded feature map; wherein: for missing observations in the station missing data simulation data, a missing value is set at the corresponding grid point location when mapping to regular spatial grid points; Feature extraction networks are used to encode features at multiple scales from regular gridded feature maps and output predicted latent variables.

4. The intelligent interpolation method for wind and solar element observation data of new energy power stations according to claim 3, characterized in that, In the first phase of training, the training objective is to minimize the following loss function: ； in: For reconstruction error, This is the KL divergence term; Reconstruction error Calculated using the following formula: ； In the above formula, This represents the reconstructed value of the dataset from the missing measurement stations. This represents the observations in the dataset of stations with no missing data. Indicates the sample sequence number index; This represents the total number of samples.

5. The intelligent interpolation method for wind and solar element observation data of new energy power stations according to claim 3, characterized in that, In the second stage of training, the Euclidean distance between the predicted latent variables output by the missing feature mapping path and the supervision label is used as the alignment loss function.

6. The intelligent interpolation method for wind and solar element observation data of new energy power stations according to claim 5, characterized in that, Alignment loss function Represented as: ； In the above formula: Indicates the latent variable being predicted; Indicates a supervisory label; sg( ) indicates stopping the gradient operation.

7. The intelligent interpolation method for wind and solar element observation data of new energy power stations according to claim 1, characterized in that, The specified wind and light elements are wind speed or irradiance.

8. An intelligent interpolation system for wind and solar energy element observation data at new energy power stations, characterized in that, include: The data acquisition module is used to acquire the station data corresponding to each uninterrupted measurement time for the specified wind and solar elements of each new energy station in the specified area within the specified historical period, so as to construct the uninterrupted measurement station dataset. The simulation data generation module is used to perform several random masking processes on the missing sub-station data in the missing measurement station dataset using an adaptive masking mechanism to generate multi-sample station missing measurement simulation data. The model building and training module is used to build and train a heterogeneous neural network model based on latent space feature alignment. The heterogeneous neural network model includes a global feature extraction path, a missing test feature mapping path, and a global field reconstruction path. A two-stage training strategy is adopted during training: in the first stage of training, the global feature extraction path and the global field reconstruction path are self-supervised using a missing test site dataset; in the second stage of training, the parameters of the global field reconstruction path are fixed, the missing test site simulation data is used as input, and the alignment target latent variable output after the first stage of training is used as the supervision label to train the missing test feature mapping path so that the predicted latent variable output by the missing test feature mapping path is aligned with the supervision label in the latent space. The interpolation module is used to take the missing observation data of the station at any interpolation time as input, based on the trained heterogeneous neural network model, and output the corresponding prediction latent variables through the missing feature mapping path. Then, it is restored through the global field reconstruction path to interpolate the corresponding predicted values ​​of the specified wind and solar elements at the site locations of new energy stations where the observation values ​​of the specified wind and solar elements are missing, and output the interpolated reconstruction data of the missing station.

9. The intelligent interpolation system for wind and solar element observation data of new energy power stations according to claim 8, characterized in that, The adaptive masking mechanism dynamically extracts the masking ratio from the range of 10% to 50% for each random mask.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the intelligent interpolation method for wind and solar element observation data of new energy power stations as described in any one of claims 1 to 7.

Images