Wind and light output scene generation method based on deep feature mining and adaptive clustering

Abstract

The application discloses a wind and light output scene generation method based on deep feature mining and adaptive clustering, which comprises the following steps: cleaning wind and light output data and performing normalization processing on the wind and light output data, so that the wind and light output data are mapped to a preset interval, so as to eliminate the dimension influence; a deep convolution feature extraction network is constructed to extract corresponding deep feature vectors from the normalized wind and light output data; a K-Means++ algorithm is used to initialize a clustering center, an improved ISODATA clustering algorithm is executed based on a density threshold dynamic splitting mechanism, and the optimization of clustering parameters is realized through Bayesian optimization to generate a typical scene. The application significantly improves the accuracy of the wind and light output scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of new energy power generation, in particular to a wind-solar power output scene generation method based on deep feature mining and adaptive clustering. BACKGROUND

[0002] Wind-solar energy has significant intermittency, randomness and volatility. Its output is affected by various factors such as weather conditions and geographical location, making it difficult to accurately predict. Accurate generation of wind-solar power output scenarios can help the power system better cope with the uncertainty of wind-solar power generation, improve the stability and reliability of the system, and is crucial for the planning, scheduling and operation of the power system.

[0003] Existing wind-solar power output scene generation methods have many shortcomings. On the one hand, many methods rely on complex meteorological data and physical models, which are computationally expensive and complex, making it difficult to adapt to the changing environment in practical applications. On the other hand, some data-driven methods have defects in feature extraction and scene clustering, such as insufficient deep feature mining of wind-solar power output data, using only simple statistical features (such as mean, standard deviation), failing to fully capture the complex patterns and potential laws in the data; the adaptability of the clustering algorithm to the data distribution is insufficient, and it is easily affected by the initial parameter settings, resulting in unstable clustering results and the generated typical scenes cannot accurately represent the actual wind-solar power output situation. These drawbacks seriously restrict the application and development of wind-solar power output scene generation methods in practical engineering, and there is an urgent need for more efficient, accurate and robust solutions. SUMMARY

[0004] In order to solve the problems of insufficient feature extraction, parameter dependence on experience and fixed number of scenes in traditional scene generation methods, the present application provides a wind-solar power output scene generation method based on deep feature mining and adaptive clustering.

[0005] The present application discloses a wind-solar power output scene generation method based on deep feature mining and adaptive clustering, which comprises:

[0006] Step 1: clean the wind-solar power output data and normalize it to map it to a pre-set interval to eliminate the effect of dimension;

[0007] Step 2: build a deep convolution feature extraction network to extract corresponding deep feature vectors from the normalized wind-solar power output data;

[0008] Step 3: use the K-Means++ algorithm to initialize the clustering center, execute the improved ISODATA clustering algorithm based on the density threshold dynamic splitting mechanism, optimize the clustering parameters through Bayesian optimization, and generate typical scenes.

[0009] Further, the step 2 comprises:

[0010] The constructed deep convolutional feature extraction network comprises an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a flattening layer, a full connection layer, a batch normalization layer, and a full connection layer connected in sequence;

[0011] The formula for extracting a deep feature vector from the normalized wind and light output data through the deep convolutional feature network is:

[0012]

[0013]

[0014] wherein, is the output of the first convolutional layer, is the normalized wind and light output, is the first layer convolutional layer, is the bias, is the deep feature vector of the first layer, is the convolutional kernel of the first layer, is the bias of the first layer, is the ReLU activation function, represents a convolution operation. Further, the input layer receives the normalized wind and light output time series , T is the time step, and 2 is the wind and light feature dimension; for the first convolutional layer, the number of convolutional kernels is set to 32, the size is 3, and the activation function is ReLU, which is used to maintain the time series dimension; for the first pooling layer, the pooling window size is 2, the step is 2, and the data dimension is reduced; for the second convolutional layer, the number of convolutional kernels is set to 64, the size is 3, and the activation function is ReLU, which is used to maintain the feature integrity; for the second pooling layer, the pooling window size is 2, and the step is 2; the full connection layer has a total of 2 layers, and the number of neurons is 128 and 64, respectively, and the output deep feature vector

[0015] is obtained.

[0016] Further, the step 3 comprises:

[0017] Step 31: setting clustering parameters; the clustering parameters include an expected number of cluster centers , a minimum number of samples in each cluster , a maximum distance standard deviation threshold in the cluster , a minimum distance threshold between clusters , and an iteration number ;

[0018] Step 32: initializing using the K-MEANS++ algorithm​​​ Cluster centers: Randomly select the first cluster center, calculate the distance of each sample to the nearest selected cluster center and the probability of selecting the next cluster center, repeat this process until a cluster center is selected. Cluster centers until;

[0019] Step 33: Cluster State Update: Calculate the distance between each sample and each cluster center using a distance formula. This formula dynamically weights the feature dimensions to accommodate the large fluctuations in wind power output and the intermittent output of photovoltaic power. It also introduces a local mean deviation product term to amplify the impact of samples deviating from the overall trend on the distance, preventing misclassification in extreme scenarios. The cluster to which a sample belongs is determined based on the distance, completing the cluster partitioning. Based on the samples contained in each partitioned cluster, the cluster center is recalculated to iteratively update the cluster center.

[0020] Step 34: Cluster validity test: Check the number of samples within each cluster. If the number of samples in a cluster is less than... If the cluster is invalid, the cluster is cancelled and invalid small clusters are removed. The cluster density is calculated. If the cluster density is less than the preset density threshold, the "split operation" is performed, that is, step 35 is performed to adjust the cluster structure. Adjusting the cluster structure includes splitting the cluster and redistributing the samples.

[0021] Step 35: Splitting Process: For and Cluster center First, target its cluster. Extract the fluctuation rate and peak-to-valley difference of wind and solar power output, and dynamically adapt the split amplitude in combination with corresponding global indicators; then... To perform PCA dimensionality reduction, along the principal component direction, a decay term is introduced for the distance between the current cluster center and its nearest selected cluster center, as well as a decay term for the average distance between each cluster center and its nearest selected cluster center. This prevents the distance between the new center after splitting the current cluster center and the nearest selected cluster center from being less than a preset distance, and finally, the split is performed. It is the cluster density of the j-th cluster. It is the j-th cluster center;

[0022] Step 36: Merging Process: For For clusters that are closest to each other, merge that cluster into the cluster that is closest to it; for clusters that are far apart, merge that cluster into the cluster that is closest to it. Merge the cluster pairs with the highest similarity.

[0023] Step 37: In the clustering process, the average correlation between the sample and the cluster center is taken as the objective function when optimizing the clustering parameters by Bayesian optimization. In calculating the correlation, the volatility similarity between the sample and the center is introduced as a correction term on the basis of calculating the cosine similarity between the sample and the center, and an outlier penalty term is introduced to suppress the interference of extreme samples, so as to maximize the optimization objective and find the optimal density threshold in splitting;

[0024] Step 38: If the maximum number of iterations is reached , or the clustering result tends to be stable, stop iteration; otherwise, repeat steps 33 to 38 to continue iteration and optimization; the clustering result tends to be stable, including that the change value of the cluster center in continuous iteration is less than a first preset value, and the change value of the sample cluster attribution is less than a second preset value;

[0025] Step 39: Generating a typical scenario: after iteration termination, output the final clustering result to present the typical scenario of clustering analysis; the final clustering result includes the clustering center of each cluster, the sample distribution in the cluster and the number.

[0026] Further, in the step 32, the distance of each sample to the nearest selected cluster center is calculated by the following formula:

[0027]

[0028] Wherein, is the distance of the i-th sample to the nearest selected cluster center, is the i-th deep feature vector, is the i-th cluster center, is the set of selected cluster centers, is the Euclidean distance. The probability of selecting the next cluster center is calculated by the following formula:

[0029]

[0030]

[0031] Wherein, is the probability of selecting the next cluster center, is the number of samples.

[0032] In the step 33, the distance between each sample and each cluster center is calculated by the following formula:

[0033]

[0034] Wherein, is the distance between the i-th sample and the j-th cluster center, is the i-th sample, and​​​ a cluster center distance between, a first depth feature vector, a first weight of a sample, is the feature dimension number, a first feature dimension of a sample, a first feature dimension of a cluster center, a first feature dimension of a sample, a first local standard deviation of a feature of a sample, a first global standard deviation of a sample, a first global mean of a sample, a first , a first local mean of a feature of a sample, a first extreme scene adjustment coefficient.

[0035] Further, in the step 35, splitting along the principal component direction is realized by the following formula:

[0036]

[0037]

[0038]

[0039]

[0040]

[0041] wherein, is the cluster density, , is the PCA eigenvalue, is the intra-cluster standard deviation, is the minimum distance between clusters, is a first sample in the cluster where the cluster center is located, is a first sample in the cluster where the cluster center is located, is a first sample in the cluster where the cluster center is located, ​For PCA principal component direction, , For the first New centers after the split of a cluster center For the original undivided first Cluster centers, For scene adaptation coefficients, It is the j-th cluster Fluctuation rate of domestic wind and solar power output It is the time step. It is the first All samples within a cluster Average output of scenery at any given moment. It is the first Peak-valley difference within a cluster, It is the wind and light volatility of all samples globally. It is the peak-to-valley difference of all samples globally. It is the mean of the maximum output of all samples within the cluster. It is the mean of the minimum output of all samples within the cluster. It is a distance decay term to prevent the center from being too close to the neighboring center after splitting. The distance between cluster centers This represents the average distance between all cluster centers.

[0042] Further, in step 37, the formula for calculating the average correlation between a sample and its cluster center is:

[0043]

[0044]

[0045] in, The average correlation between the sample and the cluster center. The input data matrix has the following dimensions: , For the sample size, The characteristic number; It is a set of cluster centers. ; This is the current number of clusters; It is the mean cosine similarity between the sample and the cluster center; It is the first The sample set of each cluster For the sample The cluster number to which it belongs; It is the first The number of samples in each cluster; It is cosine similarity. It is a dynamic adjustment coefficient. is sample—centered volatility similarity, is time step, is global volatility benchmark, used for normalization , is outlier penalty term, suppresses extreme sample within cluster disturbance, is sample—centered maximum relative volatility within cluster, is average relative volatility of sample relative to center within cluster, is minimum, is cluster center wind and light output value at time t, is sample wind and light output value at time t.

[0046] Further, the step 3 further comprises:

[0047] Take the cluster center as the typical scene benchmark, and generate a quantifiable evaluation of the wind and light output scene set after reverse normalization.

[0048] Further, the step 3 further comprises:

[0049] Construct a scene generator, which is composed of an input module, a scene probability calculation module, a feature decoding module, and an output module; the input module includes a clustering model, a data preprocessor, and a feature extractor;

[0050] The data preprocessor is used for normalization and reverse normalization of wind and light output raw data to eliminate the influence of dimension and restore the data; the feature extractor is used for extracting deep feature vectors of wind and light output time series through a deep convolution feature extraction network CNN; the clustering model is used for executing an improved ISODATA clustering algorithm to output cluster centers and sample labels;

[0051] The scene probability calculation module calculates the number of samples within each cluster according to the cluster centers and sample labels and calculates the scene probability;

[0052] The clustering center output by the clustering model is input into a feature decoding module, and a three-dimensional time sequence feature sequence, i.e., the scene data after inverse normalization, is generated through layer-by-layer operation of the feature decoding module; a decoding path network structure symmetrical to the encoding path is extracted from the feature extractor, i.e., the feature decoding module, which includes an input layer, a first full connection layer, a second full connection layer, an output layer and a reshaping layer connected in turn; wherein the dimension of the input layer is consistent with the dimension of the deep feature vector output by the feature extractor, and is used to receive the clustering center output by the clustering model; the first full connection layer is symmetrical to the second-to-last full connection layer structure of the feature extractor; the second full connection layer is symmetrical to the last full connection layer structure of the feature extractor; the output layer is used to generate the intermediate representation of the time sequence feature; and the reshaping layer is symmetrical to the flattening layer structure of the feature extractor, and reshapes the one-dimensional feature vector of the output layer into a three-dimensional time sequence feature sequence with a dimension of sample number x time step x feature dimension.

[0053] The output module outputs a wind and light output scene set according to the scene probability output by the scene probability calculation module and the three-dimensional time sequence feature sequence output by the feature decoding module; the wind and light output scene set includes a wind power output scene set, a photovoltaic output scene set and a scene probability distribution.

[0054] Further, the scene probability is calculated by the following formula:

[0055]

[0056] wherein, is the number of samples of the i-th cluster, is the total number of clusters.

[0057] Due to the adoption of the above technical solutions, the present application has the following advantages:

[0058] 1. The present application uses a convolutional autoencoder to extract deep features of a wind and light output sequence for a period of time (e.g., 24 hours); an improved ISODATA algorithm is used for adaptive clustering, and the best clustering parameters are automatically determined through Bayesian optimization; finally, typical output scenes are generated and visualized. The present application solves the problems of insufficient feature extraction, parameter dependence on experience and fixed number of scenes in traditional scene generation methods, and improves the accuracy and representativeness of wind and light output scenes.

[0059] ​2、The method has three advantages: the CNN model is used for deep feature extraction, which can more comprehensively and accurately describe the wind and light output characteristics and improve the accuracy of scene generation; the improved ISODATA clustering algorithm improves the adaptability and stability of the clustering algorithm to different data distributions by optimizing the selection of initial center, introducing a density-based splitting strategy and self-adaptive parameter adjustment, and the generated typical scene can more accurately represent the actual wind and light output situation, and reduces the dependence on manual parameter setting; the generated wind and light output scene set not only contains typical output scenes, but also calculates the probability of each scene based on the clustering results, realizes the cooperative output of typical output scenes and probability information, and helps the power system to better cope with the uncertainty of wind and light power generation. BRIEF DESCRIPTION OF DRAWINGS

[0060] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments described in the embodiments of the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0061] Figure 1 It is a flowchart of a wind and light output scene generation method based on deep feature mining and adaptive clustering.

[0062] Figure 2 It is a schematic diagram of the constructed deep convolution feature extraction network (CNN).

[0063] Figure 3 It is a flowchart of the improved ISODATA clustering algorithm.

[0064] Figure 4 It is a flowchart of the scene generator.

[0065] Figures 5(a) to 5(f) All are wind power output curve diagrams of samples in clusters.

[0066] Figures 6(a) to 6(f) All are photovoltaic output curve diagrams of samples in clusters.

[0067] Figures 7(a) to 7(f) All are photovoltaic output clustering curve diagrams.

[0068] Figures 8(a) to 8(f) All are wind power output clustering curve diagrams. DETAILED DESCRIPTION

[0069] The present application is further described in conjunction with the drawings and embodiments, and the described embodiments are only some of the embodiments of the present application, not all the embodiments. All other embodiments obtained by those skilled in the art shall belong to the scope of protection of the embodiments of the present application.

[0070] This application addresses the shortcomings of traditional clustering methods in capturing the temporal characteristics of wind and solar power output and their sensitivity to cluster numbers by integrating the deep feature extraction capabilities of one-dimensional convolutional autoencoders with an improved ISODATA dynamic clustering algorithm. The method is then applied to extract typical wind and solar power output scenarios in the Xiluodu area, constructing a database of typical wind and solar combined power output scenarios to improve the operational economy and renewable energy absorption capacity of multi-energy complementary systems under complex climatic conditions.

[0071] This application uses a typical landscape scene from a certain region as an analysis case. (See [link / reference]) Figure 1 This application provides an embodiment of a method for generating wind and solar power output scenes based on deep feature mining and adaptive clustering, which includes the following steps:

[0072] Step 1: Clean the wind and solar power output data and normalize it to map it to a preset range in order to eliminate the influence of dimensions.

[0073] Step 2: Construct a deep convolutional feature extraction network to extract the corresponding deep feature vectors from the normalized wind and solar power output data.

[0074] Step 2 includes:

[0075] See Figure 2 The constructed deep convolutional feature extraction network includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a flattening layer, a fully connected layer, a batch normalization layer, and a fully connected layer connected in sequence.

[0076] The formula for extracting deep feature vectors from normalized wind and solar power output data using a deep convolutional feature network is as follows:

[0077]

[0078]

[0079] in, It is the output of the first convolutional layer. It is the unified scenery that contributes to the effort. It is the first convolutional layer. It's a bias. For the first Layer depth feature vector, For the first Layer convolution kernel, For the first Layer bias, It is the ReLU activation function. This represents the convolution operation.

[0080] The input layer receives the normalized wind and solar power output time series. , T is time step, 2 is wind and light feature dimension; for the first convolution layer, the number of convolution kernels is set to 32, the size is 3, and the activation function is ReLU, which is used to maintain the time sequence dimension; for the first pooling layer, the pooling window size is 2, and the step is 2, which reduces the data dimension; for the second convolution layer, the number of convolution kernels is set to 64, the size is 3, and the activation function is ReLU, which is used to maintain the feature integrity; for the second pooling layer, the pooling window size is 2, and the step is 2; the fully connected layer has a total of 2 layers, and the number of neurons is 128 and 64 respectively, and the output depth feature vector .

[0081] Step 3: The K-Means++ algorithm is used to initialize the cluster center, the improved ISODATA clustering algorithm is executed based on the density threshold dynamic splitting mechanism, the clustering parameters are optimized through Bayesian optimization, and the typical scene is generated.

[0082] Referring to Figure 3 , the step 3 comprises:

[0083] Step 31: setting clustering parameters; the clustering parameters comprise expected number of cluster centers , minimum number of samples in each cluster , maximum distance standard deviation threshold in cluster class , minimum distance threshold between cluster classes , and iteration number ;

[0084] Step 32: using the K-MEANS++ algorithm to initialize cluster centers: randomly selecting the first cluster center, calculating the distance of each sample to the nearest selected cluster center and the probability of selecting the next cluster center, repeating the process until cluster centers are selected;

[0085] Step 33: cluster state updating: calculating the distance between each sample and each cluster center through the distance formula, the distance formula conforms to the characteristics of large wind power fluctuation and intermittent photovoltaic output by realizing dynamic weighting of feature dimension, and introducing a local mean deviation product term to amplify the influence of samples deviating from the overall trend on distance, avoiding extreme scenarios from being misclassified, determining the cluster to which the sample belongs according to the distance, and completing cluster division; based on the samples contained in each cluster after division, the center of the cluster is recalculated to realize iterative updating of the center of the cluster;

[0086] Step 34: cluster validity test: checking the number of samples in each cluster, if the number of cluster samples is less than If the cluster is invalid, the cluster is cancelled and invalid small clusters are removed. The cluster density is calculated. If the cluster density is less than the preset density threshold, the "split operation" is performed, that is, step 35 is performed to adjust the cluster structure. Adjusting the cluster structure includes splitting the cluster and redistributing the samples.

[0087] Step 35: Splitting Process: For and Cluster center First, target its cluster. Extract the fluctuation rate and peak-to-valley difference of wind and solar power output, and dynamically adapt the split amplitude in combination with corresponding global indicators; then... To perform PCA dimensionality reduction, along the principal component direction, a decay term is introduced for the distance between the current cluster center and its nearest selected cluster center, as well as a decay term for the average distance between each cluster center and its nearest selected cluster center. This prevents the distance between the new center after splitting the current cluster center and the nearest selected cluster center from being less than a preset distance, and finally, the split is performed. It is the cluster density of the j-th cluster. It is the j-th cluster center;

[0088] Step 36: Merging Process: For For clusters that are closest to each other, merge that cluster into the cluster that is closest to it; for clusters that are far apart, merge that cluster into the cluster that is closest to it. Merge the cluster pairs with the highest similarity.

[0089] Step 37: During the clustering process, when optimizing the clustering parameters through Bayesian optimization, the average correlation between the sample and the cluster center is used as the objective function. When calculating this correlation, the volatility similarity between the sample and the center is incorporated as a correction term based on the cosine similarity between the sample and the center. At the same time, an outlier penalty term is introduced to suppress the interference of extreme samples, thereby maximizing the optimization objective and finding the optimal density threshold when splitting.

[0090] Step 38: If the maximum number of iterations is reached If the clustering results stabilize, the iteration stops; otherwise, steps 33 to 38 are repeated to continue iterative optimization. The clustering results stabilize when the change in cluster centers during consecutive iterations is less than the first preset value and the change in sample cluster affiliation is less than the second preset value.

[0091] Step 39: Generate typical scenarios: After the iteration terminates, output the final clustering results to present typical scenarios of cluster analysis; the final clustering results include the cluster centers of each cluster, the distribution and number of samples within the cluster.

[0092] Step 4: Using cluster centers as typical scenario benchmarks, generate a set of wind and solar power output scenarios that can be quantitatively evaluated after inverse normalization.

[0093] Optionally, in the step 32, the distance of each sample to the nearest selected cluster center is calculated by the following formula:

[0094]

[0095] wherein, is the distance of the i-th sample to the nearest selected cluster center, is the i-th depth feature vector, is the i-th cluster center, is the set of selected cluster centers, is the Euclidean distance. Optionally, in the step 33, the probability of selecting the next cluster center is calculated by the following formula:

[0096]

[0097]

[0098] wherein, is the probability of selecting the next cluster center, is the number of samples.

[0099] Optionally, in the step 33, the distance between each sample and each cluster center is calculated by the following formula:

[0100]

[0101] wherein, is the distance between the sample x and the i-th cluster center, is the i-th depth feature vector, is the weight of the i-th sample, is the number of feature dimensions, is the i-th feature dimension of the sample x, is the i-th feature dimension of the cluster center, is the local standard deviation of the i-th feature dimension of the sample x, is the global standard deviation of the sample, is the global mean of the sample, is the kurtosis of the sample x, , are the i-th and j-th feature dimensions of the sample x, respectively. ​​​​​​​​​​​​​​​Cluster Center In the Local mean of dimensional features This is an adjustment coefficient for extreme scenarios.

[0102] Optionally, in step 35, splitting along the principal component direction is achieved using the following formula:

[0103]

[0104]

[0105]

[0106]

[0107]

[0108]

[0109] in, It is cluster density. , For PCA eigenvalues, It is the standard deviation within the cluster. It is the minimum distance between clusters. Cluster center The first in the cluster One sample, Cluster center The first in the cluster One sample, For PCA principal component direction, , For the first New centers after the split of a cluster center For the original undivided first Cluster centers, For scene adaptation coefficients, It is the j-th cluster Fluctuation rate of wind and solar power output in the region It is the time step. It is the first All samples within a cluster Average output of scenery at any given moment. It is the first Peak-valley difference within a cluster, It is the wind and light volatility of all samples globally. It is the peak-to-valley difference of all samples globally. It is the mean of the maximum output of all samples within the cluster. is the average of the minimum power of all samples in the cluster, is the distance decay term, which avoids the post-split center being too close to the neighborhood center, is the distance between cluster centers, is the average distance of all cluster centers.

[0110] Optionally, in the step 37, the calculation formula of the average correlation between the sample and the cluster center is:

[0111]

[0112]

[0113] wherein, is the average correlation between the sample and the cluster center, is the input data matrix, the dimension of the input data matrix is , is the number of samples, is the number of features; is the set of cluster centers, ; is the current number of clusters; is the average cosine similarity between the sample and the cluster center; is the sample set of the th cluster, is the serial number of the cluster to which the sample belongs; is the number of samples in the th cluster; is the cosine similarity, is the dynamic adjustment coefficient, is the sample-center volatility similarity, is the time step, is the global volatility benchmark, used for normalizing , is the outlier penalty term, which suppresses the interference of extreme samples within the cluster, is the maximum relative volatility of the sample-center within the cluster, is the average relative volatility of the sample-center within the cluster, is the minimum value (for example, taking ), is the wind and light output value of the cluster center at t time, is the wind and light output value of the sample at t time.

[0114] After the step 3, further comprising:

[0115] With the cluster center as a typical scene reference, the reverse normalization generates a quantifiable evaluation of the wind and light output scene set.

[0116] Optionally, the wind and light output scene set generated by the reverse normalization with the cluster center as a typical scene reference includes:

[0117] Referring to Figure 4 , a scene generator is constructed, which consists of an input module, a scene probability calculation module, a feature decoding module, and an output module; the input module includes a clustering model, a data preprocessor, and a feature extractor;

[0118] The data preprocessor is used to normalize and reverse normalize the wind and light output raw data to eliminate the dimension effect and restore the data; the feature extractor is used to extract the deep feature vector of the wind and light output time series through a deep convolution feature extraction network CNN; the clustering model is used to execute an improved ISODATA clustering algorithm to output the cluster center and sample label;

[0119] The scene probability calculation module calculates the number of samples in each cluster and calculates the scene probability according to the cluster center and sample label;

[0120] The cluster center output by the clustering model is input into the feature decoding module, and a three-dimensional time series feature sequence, i.e., the scene data after reverse normalization, is generated through the layer-by-layer operation of the feature decoding module; a decoding path network structure symmetrical to the encoding path is extracted from the feature extractor, i.e., the feature decoding module, which includes an input layer, a first full connection layer, a second full connection layer, an output layer, and a reshaping layer connected in turn; wherein the dimension of the input layer is consistent with the dimension of the deep feature vector output by the feature extractor, and is used to receive the cluster center output by the clustering model; the first full connection layer is symmetrical to the second last full connection layer structure of the feature extractor; the second full connection layer is symmetrical to the last full connection layer structure of the feature extractor; the output layer is used to generate the intermediate representation of the time series feature; the reshaping layer is symmetrical to the flattening layer structure of the feature extractor, and reshapes the one-dimensional feature vector of the output layer into a three-dimensional time series feature sequence with a dimension of sample number x time step x feature dimension;

[0121] The output module outputs the wind and light output scene set according to the scene probability output by the scene probability calculation module and the three-dimensional time series feature sequence output by the feature decoding module; the wind and light output scene set includes a wind power output scene set, a photovoltaic output scene set, and a scene probability distribution.

[0122] Optionally, the scene probability is calculated by the following formula:

[0123]

[0124] wherein, is the Number of samples in a cluster, Total number of clusters.

[0125] The wind and light output sequence is adaptively clustered by the improved ISODATA algorithm, and finally 6 typical photovoltaic output scenarios are generated, and the probability distribution is shown in Table 1.

[0126] Table 1 Typical scenario probability distribution

[0127]

[0128] Figures 5(a) to 5(f) The wind power output curve of the cluster, Figures 6(a) to 6(f) The photovoltaic output curve of the cluster, Figures 7(a) to 7(f) The photovoltaic output clustering curve, Figures 8(a) to 8(f) The wind power output clustering curve. From the number of samples and probability distribution, scenario 6 (48.9%) and scenario 5 (29.7%) are dominant, with a total proportion of 78.6%, indicating that these two scenarios have the highest frequency in historical data and can represent the main mode of wind and light combined output; The number of samples of scenario 3 (0.6%) is significantly lower than that of other scenarios, corresponding to an extreme wind power output state. In the clustering process, the algorithm dynamically adjusts the density threshold (the optimal value is 0.5591) and the number of clusters (finally stabilized at 6), combined with weighted distance calculation and density splitting strategy, effectively identifies data clusters with different output characteristics. The final cluster sample distribution is [124, 124, 11, 125, 533, 879], covering the complete range from low output to high output.

[0129] From the clustering quality evaluation results, the silhouette coefficient is 0.3408, and the CH index is 1573. The clustering results perform well in cluster tightness and cluster separation, and can clearly distinguish different photovoltaic output modes. Compared with the results of the original ISODATA algorithm, the silhouette coefficient (SC) is improved from 0.3074 to 0.3408, with a relative improvement of +10.9%, indicating that the cluster tightness and cluster separation are significantly optimized; The Calinski-Harabasz index (CH) is improved from 1481.44 to 1573, with a relative improvement of +6.2%, reflecting that the ratio of cluster dispersion to cluster variance is better.

[0130] By analyzing the characteristics of typical scenarios, the photovoltaic output range of scenarios 1 and 2 is low (0.1-819.6 MW and 0.2-664.4 MW), corresponding to cloudy or insufficient light weather conditions; The photovoltaic output of scenario 3 is near the historical maximum (1.0-3519.3 MW), and the corresponding wind power output is also near the historical maximum, which is judged as an extreme weather event; Scenarios 5 and 6 respectively represent typical wind and light combined modes under medium and high output conditions.​​

[0131] In summary, the improved ISODATA algorithm effectively extracts typical scenarios of wind and solar power output by combining Bayesian optimization and dynamic clustering strategies. The generated scenarios not only cover the main distribution characteristics of historical data, but also reflect the frequency of different scenarios through probability weighting, which helps the power system better cope with the uncertainty of wind and solar power generation.

[0132] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application and not to limit them, although the present application has been described in detail with reference to the above examples, those skilled in the art should understand: the specific embodiments of the present application can still be modified or replaced by the equivalent, without departing from the spirit and scope of the present application. Any modification or equivalent replacement, which should be covered within the protection scope of the claims of the present application.

Claims

1. A method for generating wind and solar power output scenes based on deep feature mining and adaptive clustering, characterized in that, include: Step 1: Clean the wind and solar power output data and normalize it to map it to a preset range in order to eliminate the influence of dimensions; Step 2: Construct a deep convolutional feature extraction network to extract the corresponding deep feature vectors from the normalized wind and solar power output data; Step 3: Initialize cluster centers using the K-Means++ algorithm, execute the improved ISODATA clustering algorithm based on the density threshold dynamic splitting mechanism, optimize clustering parameters through Bayesian optimization, and generate typical scenarios; Step 3 includes: Step 31: Set clustering parameters; clustering parameters include the expected number of cluster centers. Minimum number of samples in each cluster Maximum distance standard deviation threshold within a cluster Minimum distance threshold between clusters and the number of iterations ; Step 32: Initialize using the K-MEANS++ algorithm Cluster centers: Randomly select the first cluster center, calculate the distance of each sample to the nearest selected cluster center and the probability of selecting the next cluster center, repeat this process until a cluster center is selected. Cluster centers until; Step 33: Cluster State Update: Calculate the distance between each sample and each cluster center using the distance formula. The distance formula dynamically weights the feature dimensions to suit the characteristics of large fluctuations in wind power output and intermittent photovoltaic output. It also introduces a local mean deviation product term to amplify the impact of samples that deviate from the overall trend on the distance. Determine the cluster to which the sample belongs based on the distance, thus completing the cluster division. Based on the samples contained in each cluster after division, recalculate the cluster center to achieve iterative update of the cluster center. Step 34: Cluster validity test: Check the number of samples within each cluster. If the number of samples in a cluster is less than... If the cluster is invalid, the cluster is cancelled and invalid small clusters are removed. The cluster density is calculated. If the cluster density is less than the preset density threshold, the "split operation" is performed, that is, step 35 is performed to adjust the cluster structure. Adjusting the cluster structure includes splitting the cluster and redistributing the samples. Step 35: Splitting Process: For and Cluster center First, target its cluster. Extract the fluctuation rate and peak-to-valley difference of wind and solar power output, and dynamically adapt the split amplitude in combination with corresponding global indicators; then... To perform PCA dimensionality reduction, along the principal component direction, a decay term is introduced for the distance between the current cluster center and its nearest selected cluster center, as well as a decay term for the average distance between each cluster center and its nearest selected cluster center. This prevents the distance between the new center after splitting the current cluster center and the nearest selected cluster center from being less than a preset distance, and finally, the split is performed. It is the cluster density of the j-th cluster. It is the j-th cluster center; Step 36: Merging Process: For For clusters that are closest to each other, merge that cluster into the cluster that is closest to it; for clusters that are far apart, merge that cluster into the cluster that is closest to it. Merge the cluster pairs with the highest similarity. Step 37: During the clustering process, when optimizing the clustering parameters through Bayesian optimization, the average correlation between the sample and the cluster center is used as the objective function. When calculating this correlation, the volatility similarity between the sample and the center is incorporated as a correction term based on the cosine similarity between the sample and the center. At the same time, an outlier penalty term is introduced to suppress the interference of extreme samples, thereby maximizing the optimization objective and finding the optimal density threshold when splitting. Step 38: If the maximum number of iterations is reached If the clustering results stabilize, the iteration stops; otherwise, steps 33 to 38 are repeated to continue iterative optimization. The clustering results stabilize when the change in cluster centers during consecutive iterations is less than the first preset value and the change in sample cluster affiliation is less than the second preset value. Step 39: Generate typical scenarios: After the iteration terminates, output the final clustering results to present typical scenarios of cluster analysis; the final clustering results include the cluster centers of each cluster, the distribution and number of samples within the cluster.

2. The method according to claim 1, characterized in that, Step 2 includes: The constructed deep convolutional feature extraction network includes an input layer, a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a flattening layer, a fully connected layer, a batch normalization layer, and a fully connected layer connected in sequence. The formula for extracting deep feature vectors from normalized wind and solar power output data using a deep convolutional feature network is as follows: in, It is the output of the first convolutional layer. It is the unified scenery that contributes to the effort. It is the first convolutional layer. It's a bias. For the first Layer depth feature vector, For the first Layer convolution kernel, For the first Layer bias, It is the ReLU activation function. This represents the convolution operation.

3. The method according to claim 2, characterized in that, The input layer receives the normalized wind and solar power output time series. T represents the time step, and 2 represents the landscape feature dimension; For the first convolutional layer, the number of kernels is set to 32, the kernel size is 3, and the activation function is ReLU to preserve temporal dimension. For the first pooling layer, the pooling window size is 2, and the stride is 2 to reduce data dimensionality. For the second convolutional layer, the number of kernels is set to 64, the kernel size is 3, and the activation function is ReLU to preserve feature integrity. For the second pooling layer, the pooling window size is 2, and the stride is 2. There are two fully connected layers with 128 and 64 neurons respectively, outputting a deep feature vector. .

4. The method according to claim 1, characterized in that, In step 32, the distance from each sample to the nearest selected cluster center is calculated using the following formula: in, For the first The distance from each sample to the nearest selected cluster center For the first A deep feature vector, For the first Cluster centers, For the selected set of cluster centers, The distance is Euclidean. The probability of selecting the next cluster center is calculated using the following formula: in, The probability of selecting the next cluster center, The number of samples; In step 33, the distance between each sample and each cluster center is calculated using the following formula: in, It is a sample To the Cluster centers The distance between them For the first A deep feature vector, For the first The weights of each sample, It is the number of feature dimensions. For the sample The Feature dimension Cluster center The Feature dimension For the sample The Local standard deviation of dimensional features The global standard deviation of the sample. The global mean of the sample is given. For the sample peak, , Samples Cluster Center In the Local mean of dimensional features This is an adjustment coefficient for extreme scenarios.

5. The method according to claim 1, characterized in that, In step 35, splitting along the principal component direction is achieved using the following formula: in, It is cluster density. , For PCA eigenvalues, It is the standard deviation within the cluster. It is the minimum distance between clusters. Cluster center The first in the cluster One sample, Cluster center The first in the cluster One sample, For PCA principal component direction, , For the first New centers after the split of a cluster center For the original undivided first Cluster centers, For scene adaptation coefficients, It is the j-th cluster Fluctuation rate of domestic wind and solar power output It is the time step. It is the first All samples within a cluster Average output of scenery at any given moment. It is the first Peak-valley difference within a cluster, It is the wind and light volatility of all samples globally. It is the peak-to-valley difference of all samples globally. It is the mean of the maximum output of all samples within the cluster. It is the mean of the minimum output of all samples within the cluster. It is a distance attenuation term. The distance between cluster centers This represents the average distance between all cluster centers.

6. The method according to claim 1, characterized in that, In step 37, the formula for calculating the average correlation between a sample and its cluster center is: in, The average correlation between the sample and the cluster center. The input data matrix has the following dimensions: , For the sample size, The characteristic number; It is a set of cluster centers. ; This is the current number of clusters; It is the mean cosine similarity between the sample and the cluster center; It is the first The sample set of each cluster For the sample The cluster number to which it belongs; It is the first The number of samples in each cluster; It is cosine similarity. It is a dynamic adjustment coefficient. It is the sample-center volatility similarity. It is the time step. It is a global volatility benchmark used for normalization. , It is an outlier penalty term that suppresses interference from extreme samples within the cluster. It is the maximum relative fluctuation of the sample-center within the cluster. It is the average relative fluctuation of samples within a cluster relative to the center. It is the minimum value. Cluster center The wind power output value at time t. For the sample in The output value of the scenery at any moment.

7. The method according to claim 1, characterized in that, After step 3, the following is also included: Using cluster centers as typical scenario benchmarks, a set of wind and solar power output scenarios that can be quantitatively evaluated is generated after inverse normalization.

8. The method according to claim 7, characterized in that, The set of wind and solar power output scenarios that can be quantitatively evaluated after inverse normalization, using cluster centers as typical scenario benchmarks, includes: A scene generator is constructed, which consists of an input module, a scene probability calculation module, a feature decoding module, and an output module; the input module includes a clustering model, a data preprocessor, and a feature extractor. The data preprocessor is used to normalize and denormalize the raw wind and solar power output data to eliminate the influence of dimensions and restore the data; the feature extractor is used to extract the deep feature vector of the wind and solar power output time series through the deep convolutional feature extraction network (CNN); the clustering model is used to execute the improved ISODATA clustering algorithm and output the cluster centers and sample labels. The scene probability calculation module counts the number of samples in each cluster and calculates the scene probability based on the cluster center and sample label; The cluster centers output by the clustering model are input into the feature decoding module. Through layer-by-layer operations of the feature decoding module, a three-dimensional temporal feature sequence is generated, which is the scene data after inverse normalization. A decoding path network structure symmetrical to the encoding path is extracted from the feature extractor, namely the feature decoding module. The feature decoding module includes an input layer, a first fully connected layer, a second fully connected layer, an output layer, and a reshaping layer connected in sequence. The dimension of the input layer is consistent with the dimension of the depth feature vector output by the feature extractor, and it is used to receive the cluster centers output by the clustering model. The first fully connected layer is symmetrical to the penultimate fully connected layer of the feature extractor. The second fully connected layer is symmetrical to the last fully connected layer of the feature extractor. The output layer is used to generate an intermediate representation of the temporal features. The reshaping layer is symmetrical to the flattening layer structure of the feature extractor and reshapes the one-dimensional feature vector of the output layer into a three-dimensional temporal feature sequence with the dimension of sample number × time step × feature dimension. The output module outputs a set of wind and solar power output scenarios based on the scenario probability output by the scenario probability calculation module and the three-dimensional time-series feature sequence output by the feature decoding module. The set of wind and solar power output scenarios includes a set of wind power output scenarios, a set of photovoltaic power output scenarios, and a scenario probability distribution.

9. The method according to claim 8, characterized in that, The scenario probability is calculated using the following formula: in, For the first Number of samples in each cluster The total number of clusters.

Images