Micro-grid typical scene generation method and system

Abstract

The invention discloses a micro-grid typical scene generation method and system, and the method comprises the steps: forming a plurality of load-photovoltaic output two-dimensional time sequences based on the historical load data and historical photovoltaic output data of a micro-grid, considering multiple types of objects, considering the complex mutual influence relation between photovoltaic output and load, and generating a typical scene of the micro-grid. And the diversity and accuracy of the generated scene are fully improved. According to the method, coding and dimensionality reduction are carried out on the load-photovoltaic two-dimensional output time sequence, two-dimensional deep-level features are extracted, and complex features of load-photovoltaic time sequence data and correlation between the two are fully considered, so that the generated scene is closer to the actual situation, and the accuracy and reliability of the scene are improved.

Description

Technical Field

[0001] The present invention belongs to the field of microgrid operation planning, and in particular relates to a method and system for generating typical scenarios of a microgrid. Background Art

[0002] A microgrid is an electric power system consisting of a variety of distributed energy resources (such as solar energy, wind energy, energy storage, etc.) and load equipment (such as buildings, factories, etc.) within a smaller geographical area. Microgrids are usually able to operate independently, and can be connected to the main grid or disconnected from the main grid, with a certain degree of autonomy and flexibility. The current large-scale access and consumption of distributed photovoltaics has caused a wide range of safety, stability and economic operation problems in microgrids, posing severe challenges to power quality and the reliability of grid operation. How to describe the typical operation scenarios of microgrids in different scenarios is of great significance for improving the operation indicators of distribution networks.

[0003] At present, scenario analysis is a common analysis method for describing the uncertainty of distributed energy. It mainly includes the study of scenario generation and simplification, which can be further divided into three steps: probability modeling, scenario sampling, and scenario simplification. Traditional scenario generation methods use statistical methods for modeling, such as Markov chain, Gaussian process, etc., which have simple structures and strong interpretability, but have weak mapping capabilities for time series data with multi-dimensional features. After completing probability modeling, typical scenarios should be obtained through sampling. Commonly used methods include Latin hypercube sampling, Monte Carlo simulation, scenario tree, and autoregressive moving average model. Among them, Latin hypercube sampling and Monte Carlo simulation are only for independent random variables and can only generate static scenarios. They have great limitations in processing wind and solar output data with time series characteristics. The number of scenarios generated by the scenario tree is exponentially related to the number of time periods, which is prone to dimensionality disaster. The prediction error of the autoregressive moving average model will gradually increase over time, affecting the results of scenario generation.

[0004] The traditional typical scenario generation method is generally applied to traditional distribution networks, so its construction object type is relatively single, such as only considering load data or a certain renewable energy output data. This single object type construction limits the diversity and accuracy of the scenario, because there is a complex mutual influence relationship between renewable energy output (such as photovoltaic output) and load in the actual system. For example, if only renewable energy output is considered without considering load data, the generated scenario may not accurately reflect the impact of actual load changes on photovoltaic power generation, thereby affecting the planning and operation of typical scenarios. Therefore, the traditional typical scenario generation method cannot meet the current microgrid requirements related to the construction object type.

[0005] At the same time, the traditional typical scene generation method lacks consideration of the deep-level characteristics of multi-dimensional time series data. The existing methods fail to fully consider the complex characteristics and interrelationships of multi-dimensional time series data in the scene generation process. Time series data may contain information in multiple dimensions, such as time, space, frequency, etc., and the existing methods often only consider the characteristics of some dimensions, while ignoring other important dimensional information. For example, in the photovoltaic power generation scenario, in addition to considering factors such as light intensity and power generation efficiency, multi-dimensional time series data characteristics such as weather changes and seasonal changes should also be considered to more accurately simulate the actual power generation situation. If there is a lack of consideration of the deep-level characteristics of multi-dimensional time series data, the generated scene may deviate greatly from the actual situation, thereby affecting the accuracy and reliability of the scene, and further affecting the effect of system planning and operation. The scenes obtained by the existing technology are not effective enough, and it is impossible to obtain better typical scenes for planning and operation. Summary of the invention

[0006] In order to solve the problems existing in the prior art, the present invention proposes a method and system for generating typical scenarios of a microgrid.

[0007] The technical solution of the present invention is as follows:

[0008] A method for generating a typical scenario of a microgrid, comprising:

[0009] Based on the historical load data and historical photovoltaic output data of the microgrid, several load-photovoltaic output two-dimensional time series are formed;

[0010] The load-photovoltaic two-dimensional output time series is encoded and dimensionally reduced by a deep neural network learning method based on Bayesian optimization to form a number of load-photovoltaic output one-dimensional time series;

[0011] The derivative dynamic time warping algorithm is used to measure the similarity between the load-PV output one-dimensional time series, and the derivative dynamic time warping distance between the load-PV output one-dimensional time series is obtained;

[0012] Based on the derivative dynamic time warping distance, the load-photovoltaic output one-dimensional time series is clustered by interpretable sequence tree to form an interpretable sequence clustering tree as the clustering result. The clustering result is decoded and dimensionally upgraded to form a typical scenario set of microgrid.

[0013] Furthermore, the specific steps of encoding and reducing the dimension of the load-photovoltaic two-dimensional output time series by the deep neural network learning method to form several load-photovoltaic output one-dimensional time series are as follows:

[0014] Constructing an autoencoder model, wherein the autoencoder model is composed of a CNN network module and a BiLSTM network module;

[0015] The load-photovoltaic two-dimensional output time series is input into the CNN network module, and the input data input into the CNN network module is recorded as a i First, the input load-photovoltaic two-dimensional output time series is reduced to one dimension; the reduced-dimensional data enters the activation layer after the convolution layer, pooling layer, and average pooling layer, and the activation layer uses a leaky corrected linear unit; then, the data activated by the activation layer is input into the BiLSTM network module to extract the front and back time series coupling features of the long time series as the encoder part of the autoencoder model; the output of the output node is the deep data feature h obtained by the encoder i ∈R 1×H , H represents the length of the deep feature; then the original data is reconstructed through a deconvolution layer as the decoder part, and the output data b of the decoder is obtained. i ;

[0016] The Bayesian optimization method is used to optimize the model parameters and model output of the autoencoder model;

[0017] The loss function of the autoencoder model is the mean square error between the reconstructed output data of the decoder and the original data of the input encoder, expressed as:

[0018]

[0019] The specific method of optimizing the model parameters and model output of the autoencoder model using the Bayesian optimization method includes:

[0020] Define the hyperparameters of the number of convolution kernels, window size, and number of LSTM layers and their search space; randomly select a set of hyperparameter configurations in the search space for evaluation to initialize the autoencoder model;

[0021] Select the proxy model;

[0022] Train the proxy model using the initial hyperparameter configuration;

[0023] Define the acquisition function to select the next set of hyperparameter configurations based on the current proxy model;

[0024] Optimize the acquisition function, optimize the acquisition function in the search space, and find the next set of hyperparameter configurations;

[0025] Evaluate the new sampling point, train the autoencoder model using the hyperparameter configuration found by the acquisition function, and record the average loss of the trained autoencoder model on the training set; use the hyperparameter configuration at this time as the sampling point, and add the sampling point and its corresponding objective function value to the dataset, and retrain the proxy model;

[0026] Repeat the optimization of the acquisition function and the evaluation of new sampling points until the preset number of iterations is reached; after the optimization is completed, the hyperparameter configuration with the smallest average loss is selected.

[0027] Furthermore, the specific steps of using the derivative dynamic time warping algorithm to measure the similarity between the load-photovoltaic output one-dimensional time series and obtaining the derivative dynamic time warping distance between the load-photovoltaic output one-dimensional time series include:

[0028] For any two load-PV output one-dimensional time series X = {x1, ..., x n} and Y = {y1, ..., y n}, the corresponding relationship sequence between sequence X and sequence Y is P = {p 1，1 ,…,p i，j ,…,p n，n};

[0029] Among them, x1 is the first element in the sequence X; x n is the nth element in sequence X; y1 is the first element in sequence Y; y n is the nth element in sequence Y; p 1，1 is the correspondence between the first element in sequence X and the first element in sequence Y; p i，j is the correspondence between the i-th element in sequence X and the j-th element in sequence Y; p n，n is the correspondence between the nth element in sequence X and the nth element in sequence Y;

[0030] Let the lengths of the one-dimensional time series X and Y be |X| and |Y| respectively, then the set of normalized paths between X and Y is S = s1, s2, ..., s L ,in

[0031] max(|X|,|Y|)≤L≤|X|+|Y|

[0032] For any s in the set of normalized paths S l , which is in the form of (i,j), then the set of normalized paths S starts from s1 = (1,1) to s L =(|X|,|Y|), then the element in the i-th row and j-th column of the normalized path distance matrix between X and Y is expressed as:

[0033] D(i,j)=Dist(i,j)+min[D(i-1,j)D(ij-1),D(i-1,j-1)]

[0034]

[0035] D x (0) = D x (1)

[0036] D x (m) = D x (m-1)

[0037] Dist(i,j)=(D x (i)-D y (j)) 2

[0038] Derivative Dynamic Time Warping via x i and j The square of the estimated derivative difference between them replaces the distance formula of the DTW algorithm. The calculation of the two endpoints of the time series is replaced by the second point and the second to last point respectively.

[0039] The obtained normalized path distance matrix D(|X|,|Y|) is used as the dynamic time warping distance of the derivatives of the one-dimensional time series X and Y.

[0040] Furthermore, based on the derivative dynamic time warping distance, interpretable sequence tree clustering is performed on the load-photovoltaic output one-dimensional time series to form an interpretable sequence clustering tree as a clustering result, and then the clustering result is decoded and dimensionally upgraded to form a typical microgrid scenario. The specific steps include:

[0041] Construct a decision tree with k leaf nodes and k-1 patterns. Each leaf node contains a one-dimensional time series of load-PV output belonging to the same clustering category. The interpretable sequence tree clustering algorithm includes the following two parts:

[0042] In the first part, initial clustering is performed based on a number of load-photovoltaic output one-dimensional time series and the dynamic time warping distance of the derivatives between the load-photovoltaic output one-dimensional time series to obtain an initial clustering result, wherein the initial clustering result includes k clustering centers;

[0043] In the second part, based on the initial clustering results, a decision tree is constructed according to the discriminative patterns in the load-PV output time series; the most discriminative patterns are identified from the k cluster centers as nodes of the decision tree, and the relative risk and internal similarity metrics are used to evaluate the discriminative ability of the patterns, and the optimal pattern is selected for data splitting; this process is recursively repeated until k leaf nodes representing different clusters are formed, and finally the construction of an interpretable sequence clustering tree is completed.

[0044] Furthermore, the specific steps of the first part include:

[0045] Let load-PV output one-dimensional time series S = {s1,s1,…,s n}The number of clusters is k, the number of patterns is numP, and the maximum random pattern length is maxS;

[0046] Step 1: Read data and read the sequence data set S into the set I;

[0047] Step 2: Generate a random pattern:

[0048] l = random(1,maxS)

[0049] Repeat numP times, each time randomly generating a pattern with a length between 1 and maxS, and randomly selecting a subsequence from set I

[0050] p = randomSubsequence(l,I)

[0051] Step 3: Add to the pattern set, add the generated random subsequence to the pattern set Pset;

[0052] Step 4: Longest common subsequence conversion, perform the longest common subsequence conversion on the database D and the pattern set Pset, where k is s i a dividing point of

[0053] lcs(s i ,s j )=max(lcs(s i [1:k],s j ),lcs(s i [k+1:],s j ))

[0054] Step 5: Principal component analysis: perform principal component analysis on the features after LCS conversion to reduce the dimension to k dimensions, which is achieved through the eigenvectors of the covariance matrix Σ of the feature matrix X:

[0055]

[0056] Where n is the number of samples and X is the matrix representation of the data points;

[0057] Step 6: K-medoids algorithm is applied to the obtained normalized path distance matrix D based on K-center clustering of derivative dynamic time warping distance; the K-medoids algorithm is provided with k clusters, thereby forming an index of k center points; the specific steps of step 6 are as follows:

[0058] 1) According to several load-PV output two-dimensional time series, a data set V is constructed, in which each sample corresponds to a load-PV output two-dimensional time series; k samples in the data set V are arbitrarily selected as the initial center points of the cluster;

[0059] 2) For each sample in the data set V, extract the distance from each sample to all center points in the normalized path distance matrix D;

[0060] 3) Each sample is assigned to the corresponding cluster according to its minimum distance to the center point, and an index array containing the cluster to which each sample belongs is obtained;

[0061] 4) For each sample, the probability density function of the Cauchy distribution is used to calculate its probability density relative to each cluster center and normalized to obtain a probability soft assignment matrix Q. The element Q in the i-th row and j-th column of the probability soft assignment matrix Q is ij Represents characteristic data h i Belongs to cluster center c j The probability of j∈[1,k], element Q ij The calculation formula is

[0062]

[0063] In the formula, D(h i ,c j ) is the distance from the i-th sample to the j-th center point, γ j is the cluster center c j The scale parameter, Z i is the normalization factor;

[0064] 5) Normalizing the square of the probability soft assignment matrix Q to obtain the target distribution matrix T, where the element T in the i-th row and j-th column of the target distribution matrix T is ij The calculation formula is

[0065]

[0066] 6) During the model training process, the calculated soft assignment matrix Q and target distribution matrix T are used to calculate the KL divergence as the clustering loss, and the reconstruction loss is the mean square error MSE, which is used to measure the difference between the reconstructed time series and the original time series; the clustering loss is used to measure the difference between the probability soft assignment matrix Q and the target distribution matrix T;

[0067] The reconstruction loss is: i is the input data, b i Output data for the decoder;

[0068]

[0069] The clustering loss is:

[0070]

[0071] 7) In each training cycle, forward propagation is performed to obtain model output and intermediate output; the derivative dynamic time warping distance matrix, soft assignment matrix Q and target distribution matrix T are calculated; the total loss is calculated, and back propagation and optimization of model parameters are performed;

[0072] 8) After each training cycle, the adjusted Rand index is used to evaluate the consistency between the current clustering result and the previous clustering result. If the ARI is greater than 0.98 for 7 consecutive times, it is considered that the clustering result does not change and the model converges. The cluster center C = {c1, c1, …, c k};

[0073] Furthermore, the specific steps of the second part include:

[0074] Step 1: Initialize clustering. Given the initial cluster C = {c1, c1, …, c k}, each cluster center corresponds to a leaf node;

[0075] Step 2: Pattern recognition, find representative patterns for each k-1 cluster, which will be used to further build a decision tree;

[0076] Step 3: Split the data set S into two groups S according to whether it contains the sequence pattern p p and

[0077] Step 4: Pattern selection, using a relative risk-based splitting metric to select a discriminative pattern p for tree splitting;

[0078] Step 5: Relative risk assessment and internal similarity measurement. Calculate the relative risk RR of pattern p, which is the ratio of its support in the positive class S+ to its support in the negative class S-:

[0079]

[0080] To evaluate the quality of pattern p, we use the internal similarity measure SIM, which takes into account the pattern length and the number and length of sequences in the positive class;

[0081]

[0082] where |p| is the length of pattern p, |S + | is a positive class S + The number of sequences in |s i |For S + medium sequence s i Length;

[0083] Step 6: Multi-class expansion, for output k categories, expand the concept of relative risk and define the positive class S +To support the cluster with the highest pattern p, and use the one-vs-rest strategy to obtain the negative class S - ;

[0084]

[0085] Step 7: Mode selection: if multiple modes have the same relative risk, the mode with the largest SIM value is selected as the basis for dividing the data set;

[0086] Step 8: Tree construction, using the top-1 distinguishing sequence pattern as the split point for building the tree, and assigning the sequences containing this pattern to the same cluster;

[0087] According to the k cluster centers obtained in the first part, the root node N is initialized and the pattern tree is constructed. The tree construction follows the top-down binary split method;

[0088] Determine whether the required number of leaf nodes has been reached, and decide whether to continue splitting based on the minimum number of sequences contained. The judgment criteria are:

[0089] k<2

[0090] |S|≤min(k,minS)

[0091] Where minS is the minimum sequence number;

[0092] Apply the sequential pattern mining algorithm to each cluster center c i Extract sequence patterns, assign sequences containing pattern p to the right node, and those not containing pattern p to the left node, until the stopping condition is met, and the construction of an interpretable sequence tree is completed.

[0093] Furthermore, a plurality of sample points are stored in the load-photovoltaic output two-dimensional time series, and each sample point records load data and photovoltaic output data of the same time series.

[0094] A microgrid typical scenario generation system, comprising: a sequence generation module, an encoding and dimensionality reduction module, a derivative dynamic time warping distance acquisition module, a clustering module and a decoding and dimensionality increase module;

[0095] The sequence generation module is used to form a plurality of load-photovoltaic output two-dimensional time series based on the microgrid historical load data and the historical photovoltaic output data;

[0096] The encoding and dimensionality reduction module is used to encode and reduce the dimensionality of the load-photovoltaic two-dimensional output time series through a deep neural network learning method to form a plurality of load-photovoltaic output one-dimensional time series;

[0097] The derivative dynamic time warping distance acquisition module is used to measure the similarity between the load-photovoltaic output one-dimensional time series using the derivative dynamic time warping algorithm to obtain the derivative dynamic time warping distance between the load-photovoltaic output one-dimensional time series;

[0098] The clustering module is used to perform interpretable sequence tree clustering on the load-photovoltaic output one-dimensional time series based on the derivative dynamic time warping distance, form an interpretable sequence clustering tree as a clustering result, decode and upgrade the clustering result to form a set of typical scenarios of a microgrid; an electronic device, the electronic device includes a memory and a processor, the memory stores a computer program, and the processor is used to call and run the computer program stored in the memory to execute any of the methods described above.

[0099] A computer-readable storage medium stores a computer program, wherein the computer program implements the steps of any of the above methods when executed by a processor.

[0100] Compared with the prior art, the present invention has the following beneficial effects:

[0101] The present invention provides a method and system for generating typical scenarios of a microgrid. The method forms a plurality of load-photovoltaic output two-dimensional time series based on the historical load data and historical photovoltaic output data of the microgrid, takes into account multiple types of objects, considers the complex mutual influence relationship between photovoltaic output and load, and fully improves the diversity and accuracy of the generated scenarios.

[0102] The method of the present invention encodes and reduces the dimension of the load-photovoltaic two-dimensional output time series, extracts two-dimensional deep features, and fully considers the complex characteristics of the load-photovoltaic time series data and the mutual correlation between the two, so that the generated scene is closer to the actual situation, thereby improving the accuracy and reliability of the scene.

[0103] The method of the present invention uses a derivative dynamic time warping algorithm to measure the similarity between load-photovoltaic output one-dimensional time series, obtains the derivative dynamic time warping distance between load-photovoltaic output one-dimensional time series, and clusters the load-photovoltaic output one-dimensional time series, avoiding the "singularity" problem generated by DTW when the y-axis is distorted to align the x-axis, that is, a point in one time series will be mapped to a part of another time series, which will lead to the wrong calculation of the warping distance. In this way, the morphological distance between the load time series and the photovoltaic output time series can be better measured, thereby obtaining a more accurate typical scenario of the microgrid.

[0104] The method of the present invention constructs an autoencoder model (AE model) composed of a CNN network module (convolutional neural network model) and a BiLSTM network module (bidirectional long short-term memory network module), thereby better extracting the deep-level features of the load-photovoltaic output two-dimensional time series variables, making the feature extraction more accurate and comprehensive.

[0105] The method of the present invention uses Bayesian optimization to optimize and adjust the parameters of the CNN-BiLSTM model, which can more effectively search for the optimal parameter combination and improve the performance and generalization ability of the model.

[0106] The method of the present invention uses the Interpretable Sequence Tree Clustering (ISCT) algorithm to further refine and optimize the generated typical microgrid scenarios. The ISCT algorithm groups similar load-photovoltaic output time series together by constructing a tree structure, thereby achieving a more detailed classification and identification of typical microgrid scenarios. This method not only improves the diversity of typical microgrid scenarios, but also enhances the interpretability of typical microgrid scenarios, allowing users to more intuitively understand the differences and connections between different typical microgrid scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0107] Figure 1 A schematic diagram of a process for generating a typical microgrid scenario in an embodiment;

[0108] Figure 2 It is a time series curve diagram in the embodiment;

[0109] Figure 3 A schematic diagram for regular understanding of distance in the embodiment;

[0110] Figure 4 It is the structure diagram of CNN-BiLSTM in the embodiment.

[0111] Figure 5 Flow chart of ISCT algorithm in the embodiment. DETAILED DESCRIPTION

[0112] Embodiment 1:

[0113] A typical scenario generation method of a microgrid of the present invention is provided, wherein a photovoltaic power plant is provided in the microgrid, such as Figure 1 As shown, including:

[0114] S1. Based on the historical load data and historical photovoltaic output data of the microgrid, several load-photovoltaic output two-dimensional time series are formed;

[0115] S2. Encode and reduce the dimension of the load-photovoltaic two-dimensional output time series through a deep neural network learning method to form several load-photovoltaic output one-dimensional time series;

[0116] S3, using a dynamic time warping algorithm to measure the similarity between the load-PV output one-dimensional time series, and obtaining the derivative dynamic time warping distance between the load-PV output one-dimensional time series;

[0117] S4. Based on the derivative dynamic time warping distance, interpretable sequence tree clustering (ISCT) is performed on the load-PV output one-dimensional time series to form a concise and interpretable tree structure as the clustering result. The clustering result is then decoded and dimensionally upgraded to form a typical microgrid scenario.

[0118] Embodiment 2:

[0119] This example is further designed on the basis of the first embodiment in that the specific steps of encoding and reducing the dimensionality of the load-photovoltaic output two-dimensional time series by a deep neural network learning method to form several load-photovoltaic output one-dimensional time series are as follows:

[0120] Construct an autoencoder model (AE model). The autoencoder model is composed of a CNN network module (convolutional neural network model) and a BiLSTM network module (bidirectional long short-term memory network module). The structure is as follows: Figure 4 shown.

[0121] The load-PV two-dimensional output time series is input into the CNN network module, and the input data into the CNN network module is recorded as a i In order to fully extract the deep features in the multi-dimensional time series variables, the input load-photovoltaic two-dimensional output time series is first reduced to one dimension; in order to retain the data information in the time window as much as possible, the reduced-dimensional data enters the activation layer after the convolution layer, pooling layer, and average pooling layer, and the activation layer uses the leaky-rectified linear unit (Leaky-ReLU); the convolutional neural network module reduces the dimension of the multi-dimensional time series data and extracts preliminary features. Then, the data activated by the activation layer is input into the BiLSTM network module to extract the front and back time series coupling features of the long time series as the encoder part of the autoencoder model; compared with LSTM, BiLSTM can mine the bidirectional time features of the time series. The output of the output node is the deep data feature h obtained by the encoder i ∈R 1×H , H represents the length of the deep feature; then the original data is reconstructed through a deconvolution layer as the decoder part, and the output data b of the decoder is obtained. i .

[0122] Embodiment three:

[0123] This example is further designed on the basis of the second embodiment in that the loss function of the automatic encoder model (AE model) in this example is the mean square error (MSE) between the reconstructed output data of the decoder and the original data of the input encoder, and the expression is:

[0124]

[0125] Embodiment 4:

[0126] This example is further designed on the basis of the second embodiment in that this example also includes: using the Bayesian optimization method to optimize the model parameters and model output of the autoencoder model, and the specific method includes:

[0127] Use the Optuna library based on Bayesian optimization to perform hyperparameter optimization;

[0128] Define the hyperparameters of the number of convolution kernels, window size, and number of LSTM layers and their search space; randomly select a set of hyperparameter configurations (such as convolution kernel data, window size, or number of LSTM layers) in the search space for evaluation to initialize the autoencoder model; this can be achieved by randomly selecting a set of hyperparameter configurations and training;

[0129] Select a surrogate model. Bayesian optimization approximates the true objective function through the surrogate model.

[0130] Train the proxy model, using the initial hyperparameter configuration to train the proxy model so that it can predict the performance of the target function in the search space; define the acquisition function to select the next set of hyperparameter configurations based on the current proxy model;

[0131] Optimize the acquisition function, optimize the acquisition function in the search space, and find the next set of hyperparameter configurations; the acquisition function strikes a balance between the mean and uncertainty estimates of the surrogate model, considering both exploring new areas and utilizing the current best area;

[0132] Evaluate the new sampling point, use the hyperparameter configuration found by the acquisition function to train the autoencoder model and evaluate its performance, record the average loss of the trained autoencoder model on the training set; use the hyperparameter configuration at this time as the sampling point, and add the sampling point and its corresponding objective function value to the dataset, and retrain the proxy model;

[0133] Repeat the optimization of the acquisition function and the evaluation of new sampling points (training steps) until the preset number of iterations is reached; after the optimization is completed, the hyperparameter configuration with the smallest average loss during the evaluation process is selected.

[0134] Embodiment five:

[0135] This example is further designed on the basis of Example 1 in that this example uses the derivative dynamic time warping algorithm to measure the similarity between the load-photovoltaic output one-dimensional time series, and obtains the derivative dynamic time warping distance between the load-photovoltaic output one-dimensional time series. The derivative dynamic time warping algorithm is the DDTW algorithm, and the DDTW algorithm has high robustness and accuracy when processing time series data sequences. Unlike the Euclidean distance, the DDTW algorithm can match the data points of the time series by bending the time points of the time series, and can measure signals with equal or unequal data lengths, and has good adaptability to the missing or dynamics of the sequence data. The core of the DDTW algorithm is to construct a corresponding relationship between the shapes of the two time series (for example, the corresponding relationship between the peak and the valley, respectively), and then use the square of the estimated derivative difference between the corresponding points of the two time series on the basis of the established corresponding relationship instead of the calculation of traditional similarity measures such as the Euclidean distance. The specific steps include:

[0136] For any two load-PV output one-dimensional time series X = {x1, ..., x n} and Y = {y1, ..., y n}, the corresponding relationship sequence between sequence X and sequence Y is P = {p 1，1 ,…,p i，j ,…,p n，n};

[0137] Among them, x1 is the first element in the sequence X; x n is the nth element in sequence X; y1 is the first element in sequence Y; y n is the nth element in sequence Y; p 1，1 is the correspondence between the first element in sequence X and the first element in sequence Y; p i，j is the correspondence between the i-th element in sequence X and the j-th element in sequence Y; p n，n is the correspondence between the nth element in sequence X and the nth element in sequence Y;

[0138] In the DDTW algorithm, it is stipulated that the effective path P needs to satisfy two constraints, namely, boundary, continuity and monotonicity;

[0139] In time series, the lengths of time series in the same time period may not be equal. The displacements on different time axes lead to differences between time series. If two time series want to reach the same state, the displacements must be restored. In order to better obtain the similarity between two time series, DDTW expands and shortens the time series. Figure 2As shown, the two time series are represented by the upper and lower solid lines, and the dotted line represents the similarity between the two time series. DDTW measures the similarity between two time series by replacing the distance formula of the DTW algorithm with the square of the estimated derivative difference between the corresponding points of the two time series, which is called the normalized path distance;

[0140] DDTW calculation method: Let the lengths of the one-dimensional time series X and Y be |X| and |Y| respectively, then the set of normalized paths between X and Y is S = s1, s2, ..., s L ,in

[0141] max(|X|,|Y|)≤L≤|X|+|Y|

[0142] For any s in the set of normalized paths S l , which is in the form of (i,j), then the set of normalized paths S starts from s1 = (1,1) to s L =(|X|,|Y|), such as Figure 3 The curve shown represents the normalized distance between time series X and Y. The element in the i-th row and j-th column of the normalized path distance matrix between X and Y is expressed as:

[0143] D(i,j)=Dist(i,j)+min[D(i-1,j)D(ij-1),D(i-1,j-1)]

[0144]

[0145] D x (0) = D x (1)

[0146] D x (m) = D x (m-1)

[0147] Dist(i,j)=(D x (i)-D y (j)) 2

[0148] Derivative Dynamic Time Warping via x i and j The square of the estimated derivative difference between them replaces the distance formula of the DTW algorithm. The second point and the second to last point are used to calculate the two endpoints of the time series respectively.

[0149] The obtained normalized path distance matrix D(|X|,|Y|) is used as the dynamic time warping distance of the derivatives of the one-dimensional time series X and Y.

[0150] Embodiment six:

[0151] This example is further designed on the basis of the first embodiment in that, Figure 5 As shown in the figure, this example uses the interpretable sequence tree clustering (ISCT) algorithm on the load-PV output one-dimensional time series based on the derivative dynamic time warping distance to form a concise and interpretable tree structure as the clustering result. The clustering result is then decoded and dimensionally upgraded to form a typical microgrid scenario. The specific steps include:

[0152] The specific algorithm steps are described as follows:

[0153] Construct a decision tree with k leaf nodes and k-1 patterns. Each leaf node contains a one-dimensional time series of load-PV output belonging to the same clustering category. The interpretable sequence tree clustering algorithm includes the following two parts:

[0154] In the first part, initial clustering is performed based on a number of load-photovoltaic output one-dimensional time series and the dynamic time warping distance of the derivatives between the load-photovoltaic output one-dimensional time series to obtain an initial clustering result, wherein the initial clustering result includes k clustering centers;

[0155] In the second part, based on the initial clustering results, a decision tree is constructed according to the discriminative patterns in the load-PV output time series; the most discriminative patterns are identified from the k cluster centers as nodes of the decision tree, and the relative risk and internal similarity metrics are used to evaluate the discriminative ability of the patterns, and the optimal pattern is selected for data splitting; this process is recursively repeated until k leaf nodes representing different clusters are formed, and finally the construction of an interpretable sequence clustering tree is completed.

[0156] Furthermore, the specific steps of the first part include:

[0157] Let load-PV output one-dimensional time series S = {s1,s1,…,s n}The number of clusters is k, the number of patterns is numP, and the maximum random pattern length is maxS;

[0158] Step 1: Read data and read the sequence data set S into the set I;

[0159] Step 2: Generate a random pattern:

[0160] l = random(1,maxS)

[0161] Repeat numP times, each time randomly generating a pattern with a length between 1 and maxS, and randomly selecting a subsequence from set I

[0162] p = randomSubsequence(l,I)

[0163] Step 3: Add to the pattern set, add the generated random subsequence to the pattern set Pset;

[0164] Step 4: Longest common subsequence conversion, perform the longest common subsequence conversion on the database D and the pattern set Pset, where k is s i a dividing point of

[0165] lcs(s i ,s j )=max(lcs(s i [1:k],s j ),lcs(s i [k+1:],s j ))

[0166] Step 5: Principal component analysis: perform principal component analysis on the features after LCS conversion to reduce the dimension to k dimensions, which is achieved through the eigenvectors of the covariance matrix Σ of the feature matrix X:

[0167]

[0168] Where n is the number of samples and X is the matrix representation of the data points;

[0169] Step 6: K-medoids algorithm is applied to the obtained normalized path distance matrix D based on K-center clustering of derivative dynamic time warping distance; the K-medoids algorithm is provided with k clusters, thereby forming an index of k center points; the specific steps of step 6 are as follows:

[0170] 1) According to several load-PV output two-dimensional time series, a data set V is constructed, in which each sample corresponds to a load-PV output two-dimensional time series; k samples in the data set V are arbitrarily selected as the initial center points of the cluster;

[0171] 2) For each sample in the data set V, extract the distance from each sample to all center points in the normalized path distance matrix D;

[0172] 3) Each sample is assigned to the corresponding cluster according to its minimum distance to the center point, and an index array containing the cluster to which each sample belongs is obtained;

[0173] 4) For each sample, the probability density function of the Cauchy distribution is used to calculate its probability density relative to each cluster center and normalized to obtain a probability soft assignment matrix Q. The element Q in the i-th row and j-th column of the probability soft assignment matrix Q is ij Represents characteristic data h i Belongs to cluster center c j The probability of j∈[1,k], element Q ij The calculation formula is

[0174]

[0175] In the formula, D(h i ,c j ) is the distance from the i-th sample to the j-th center point, γ j is the cluster center c j The scale parameter, Z i is the normalization factor;

[0176] 5) Normalizing the square of the probability soft assignment matrix Q to obtain the target distribution matrix T, where the element T in the i-th row and j-th column of the target distribution matrix T is ij The calculation formula is

[0177]

[0178] 6) During the model training process, the calculated soft assignment matrix Q and target distribution matrix T are used to calculate the KL divergence as the clustering loss, and the reconstruction loss is the mean square error MSE, which is used to measure the difference between the reconstructed time series and the original time series; the clustering loss is used to measure the difference between the probability soft assignment matrix Q and the target distribution matrix T;

[0179] The reconstruction loss is: i is the input data, b i Output data for the decoder;

[0180]

[0181] The clustering loss is:

[0182]

[0183] 7) In each training cycle, forward propagation is performed to obtain model output and intermediate output; the derivative dynamic time warping distance matrix, soft assignment matrix Q and target distribution matrix T are calculated; the total loss is calculated, and back propagation and optimization of model parameters are performed;

[0184] 8) After each training cycle, the adjusted Rand index is used to evaluate the consistency between the current clustering result and the previous clustering result. If the ARI is greater than 0.98 for 7 consecutive times, it is considered that the clustering result does not change and the model converges. The cluster center C = {c1, c1, …, c k};

[0185] Furthermore, the specific steps of the second part include:

[0186] Step 1: Initialize clustering. Given the initial cluster C = {c1, c1, …, c k}, each cluster center corresponds to a leaf node;

[0187] Step 2: Pattern recognition, find representative patterns for each k-1 cluster, which will be used to further build a decision tree;

[0188] Step 3: Split the data set S into two groups S according to whether it contains the sequence pattern p p and

[0189] Step 4: Pattern selection, using a relative risk-based splitting metric to select a discriminative pattern p for tree splitting;

[0190] Step 5: Relative risk assessment and internal similarity measurement. Calculate the relative risk RR of pattern p, which is the ratio of its support in the positive class S+ to its support in the negative class S-:

[0191]

[0192] To evaluate the quality of pattern p, we use the internal similarity measure SIM, which takes into account the pattern length and the number and length of sequences in the positive class;

[0193]

[0194] where |p| is the length of pattern p, |S + | is a positive class S + The number of sequences in |s i |For S + medium sequence s i Length;

[0195] Step 6: Multi-class expansion, for output k categories, expand the concept of relative risk and define the positive class S + To support the cluster with the highest pattern p, and use the one-vs-rest strategy to obtain the negative class S-;

[0196]

[0197] Step 7: Mode selection: if multiple modes have the same relative risk, the mode with the largest SIM value is selected as the basis for dividing the data set;

[0198] Step 8: Tree construction, using the top-1 distinguishing sequence pattern as the split point for building the tree, and assigning the sequences containing this pattern to the same cluster;

[0199] According to the k cluster centers obtained in the first part, the root node N is initialized and the pattern tree is constructed. The tree construction follows the top-down binary split method;

[0200] Determine whether the required number of leaf nodes has been reached, and decide whether to continue splitting based on the minimum number of sequences contained. The judgment criteria are:

[0201] k<2

[0202] |S|≤min(k,minS)

[0203] Where minS is the minimum sequence number;

[0204] Apply the sequential pattern mining algorithm to each cluster center c i Extract sequence patterns, assign sequences containing pattern p to the right node, and those not containing pattern p to the left node, until the stopping condition is met, and the construction of an interpretable sequence tree is completed.

[0205] Embodiment seven:

[0206] This example is further designed on the basis of the first embodiment in that a number of sample points are stored in the load-photovoltaic output two-dimensional time series in this example, and each sample point records the load data and photovoltaic output data of the same time series.

[0207] Embodiment eight:

[0208] The present invention discloses a typical scene generation system for a microgrid, in which a photovoltaic power plant is provided, comprising: a sequence generation module, an encoding and dimensionality reduction module, a derivative dynamic time warping distance acquisition module, a clustering module and a decoding and dimensionality increase module.

[0209] A sequence generation module is used to form several load-photovoltaic output two-dimensional time series based on the microgrid historical load data and historical photovoltaic output data;

[0210] The encoding and dimensionality reduction module is used to encode and reduce the dimensionality of the load-photovoltaic two-dimensional output time series through a deep neural network learning method to form several load-photovoltaic output one-dimensional time series;

[0211] The derivative dynamic time warping distance acquisition module is used to measure the similarity between the load-photovoltaic output one-dimensional time series using the dynamic time warping algorithm to obtain the derivative dynamic time warping distance between the load-photovoltaic output one-dimensional time series;

[0212] Clustering module, used to cluster the load-PV output one-dimensional time series based on the derivative dynamic time warping distance;

[0213] The decoding and dimension upgrading module is used to decode and upgrade the clustering results to form a typical scenario of the microgrid.

[0214] Embodiment nine:

[0215] An electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor is used to call and run the computer program stored in the memory to execute any of the methods described above.

[0216] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of any of the above methods are implemented.

Claims

1. A method for generating a typical microgrid scenario, characterized in that: include: Based on the historical load data and historical photovoltaic output data of the microgrid, several load-photovoltaic output two-dimensional time series are formed; The load-photovoltaic two-dimensional output time series is encoded and dimensionally reduced by a deep neural network learning method based on Bayesian optimization to form a number of load-photovoltaic output one-dimensional time series; The derivative dynamic time warping algorithm is used to measure the similarity between the load-PV output one-dimensional time series, and the derivative dynamic time warping distance between the load-PV output one-dimensional time series is obtained; Based on the derivative dynamic time warping distance, the load-photovoltaic output one-dimensional time series is clustered by interpretable sequence tree to form an interpretable sequence clustering tree as the clustering result. The clustering result is decoded and dimensionally upgraded to form a typical scenario set of microgrid.

2. The method for generating a typical microgrid scenario according to claim 1, characterized in that: The specific steps of encoding and reducing the dimension of the load-photovoltaic two-dimensional output time series by the deep neural network learning method to form several load-photovoltaic output one-dimensional time series are as follows: Constructing an autoencoder model, wherein the autoencoder model is composed of a CNN network module and a BiLSTM network module; The load-photovoltaic two-dimensional output time series is input into the CNN network module, and the input data input into the CNN network module is recorded as a i First, the input load-photovoltaic two-dimensional output time series is reduced to one dimension; the reduced-dimensional data is sequentially passed through the convolution layer, the pooling layer, and the average pooling layer before entering the activation layer, and the activation layer uses a leaky corrected linear unit; Then, the data activated by the activation layer is input into the BiLSTM network module to extract the front and back temporal coupling features of the long time series as the encoder part of the autoencoder model; the output of the output node is the deep data feature h obtained by the encoder i ∈R 1×H , H represents the length of the deep-level feature; Reconstruct the original data through a deconvolution layer as the decoder part, and get the output data b of the decoder i。 3. The method for generating a typical microgrid scenario according to claim 2, characterized in that: The loss function of the autoencoder model is the mean square error between the reconstructed output data of the decoder and the original data of the input encoder, expressed as:

4. The method for generating a typical microgrid scenario according to claim 2, characterized in that: Also includes: The Bayesian optimization method is used to optimize the model parameters and model output of the autoencoder model. The specific methods include: Define the hyperparameters of the number of convolution kernels, window size, and number of LSTM layers and their search space; randomly select a set of hyperparameter configurations in the search space for evaluation to initialize the autoencoder model; Select the proxy model; Train the proxy model using the initial hyperparameter configuration; Define the acquisition function to select the next set of hyperparameter configurations based on the current proxy model; Optimize the acquisition function, optimize the acquisition function in the search space, and find the next set of hyperparameter configurations; Evaluate the new sampling point, train the autoencoder model using the hyperparameter configuration found by the acquisition function, and record the average loss of the trained autoencoder model on the training set; use the hyperparameter configuration at this time as the sampling point, and add the sampling point and its corresponding objective function value to the dataset, and retrain the proxy model; Repeat the optimization of the acquisition function and the evaluation of new sampling points until the preset number of iterations is reached; after the optimization is completed, the hyperparameter configuration with the smallest average loss is selected.

5. The method for generating a typical microgrid scenario according to claim 1, characterized in that: The specific steps of using the derivative dynamic time warping algorithm to measure the similarity between the load-photovoltaic output one-dimensional time series and obtaining the derivative dynamic time warping distance between the load-photovoltaic output one-dimensional time series include: For any two load-PV output one-dimensional time series X = {x1, ..., x n } and Y = {y1, ..., y n }, the corresponding relationship sequence between sequence X and sequence Y is P = {p 1，1 ,…,p i，j ,…,p n，n }; Among them, x1 is the first element in the sequence X; x n is the nth element in sequence X; y1 is the first element in sequence Y; y n is the nth element in sequence Y; p 1，1 is the correspondence between the first element in sequence X and the first element in sequence Y; p i，j is the correspondence between the i-th element in sequence X and the j-th element in sequence Y; p n，n is the correspondence between the nth element in sequence X and the nth element in sequence Y; Let the lengths of the one-dimensional time series X and Y be |X| and |Y| respectively, then the set of normalized paths between X and Y is S = s1, s2, ..., s L ,in max(|X|,|Y|)≤L≤|X|+|Y| For any s in the set of normalized paths S l , which is in the form of (i,j), then the set of normalized paths S starts from s1 = (1,1) to s L =(|X|,|Y|), then the element in the i-th row and j-th column of the normalized path distance matrix between X and Y is expressed as: D(i,j)=Dist(i,j)+min[D(i-1,j)D(ij-1),D(i-1,j-1)] D x (0)=D x (1) D x (m)=D x (m-1) Dist(i,j)=(D x (i)-D y (j)) 2 Derivative Dynamic Time Warping via x i and j The square of the estimated derivative difference between them replaces the distance formula of the DTW algorithm. The calculation of the two endpoints of the time series is replaced by the second point and the second to last point respectively. The obtained normalized path distance matrix D(|X|,|Y|) is used as the dynamic time warping distance of the derivatives of the one-dimensional time series X and Y.

6. The method for generating a typical microgrid scenario according to claim 1, characterized in that: The specific steps of performing interpretable sequence tree clustering on the load-photovoltaic output one-dimensional time series based on the derivative dynamic time warping distance to form an interpretable sequence clustering tree as a clustering result, and then decoding and upgrading the clustering result to form a typical microgrid scenario include: Construct a decision tree with k leaf nodes and k-1 patterns. Each leaf node contains a one-dimensional time series of load-PV output belonging to the same clustering category. The interpretable sequence tree clustering algorithm includes the following two parts: In the first part, initial clustering is performed based on a number of load-photovoltaic output one-dimensional time series and the dynamic time warping distance of the derivatives between the load-photovoltaic output one-dimensional time series to obtain an initial clustering result, wherein the initial clustering result includes k clustering centers; In the second part, based on the initial clustering results, a decision tree is constructed according to the discriminative patterns in the load-PV output time series; the most discriminative patterns are identified from the k cluster centers as nodes of the decision tree, and the relative risk and internal similarity metrics are used to evaluate the discriminative ability of the patterns, and the optimal pattern is selected for data splitting; this process is recursively repeated until k leaf nodes representing different clusters are formed, and finally the construction of an interpretable sequence clustering tree is completed.

7. The method for generating a typical microgrid scenario according to claim 6, characterized in that: The specific steps of the first part include: Let load-PV output one-dimensional time series S = {s1,s1,…,s n }The number of clusters is k, the number of patterns is numP, and the maximum random pattern length is maxS; Step 1: Read data and read the sequence data set S into the set I; Step 2: Generate a random pattern: l = random(1,maxS) Repeat numP times, each time randomly generating a pattern with a length between 1 and maxS, and randomly selecting a subsequence from set I p = randomSubsequence(l,I) Step 3: Add to the pattern set, add the generated random subsequence to the pattern set Pset; Step 4: Longest common subsequence conversion, perform the longest common subsequence conversion on the database D and the pattern set Pset, where k is s i a dividing point of lcs(s i ,s j )=max(lcs(s i [1:k],s j ),lcs(s i [k+1:],s j )) Step 5: Principal component analysis: perform principal component analysis on the features after LCS conversion to reduce the dimension to k dimensions, which is achieved through the eigenvectors of the covariance matrix Σ of the feature matrix X: Where n is the number of samples and X is the matrix representation of the data points; Step 6: K-medoids algorithm is applied to the obtained normalized path distance matrix D based on K-center clustering of derivative dynamic time warping distance; the K-medoids algorithm is provided with k clusters, thereby forming an index of k center points; the specific steps of step 6 are as follows: 1) According to several load-PV output two-dimensional time series, a data set V is constructed, in which each sample corresponds to a load-PV output two-dimensional time series; k samples in the data set V are arbitrarily selected as the initial center points of the cluster; 2) For each sample in the data set V, extract the distance from each sample to all center points in the normalized path distance matrix D; 3) Each sample is assigned to the corresponding cluster according to its minimum distance to the center point, and an index array containing the cluster to which each sample belongs is obtained; 4) For each sample, the probability density function of the Cauchy distribution is used to calculate its probability density relative to each cluster center and normalized to obtain a probability soft assignment matrix Q. The element Q in the i-th row and j-th column of the probability soft assignment matrix Q is ij Represents characteristic data h i Belongs to cluster center c j The probability of j∈[1,k], element Q ij The calculation formula is In the formula, D(h i ,c j ) is the distance from the i-th sample to the j-th center point, γ j is the cluster center c j The scale parameter, Z i is the normalization factor; 5) Normalizing the square of the probability soft assignment matrix Q to obtain the target distribution matrix T, where the element T in the i-th row and j-th column of the target distribution matrix T is ij The calculation formula is 6) During the model training process, the calculated soft assignment matrix Q and target distribution matrix T are used to calculate the KL divergence as the clustering loss, and the reconstruction loss is the mean square error MSE, which is used to measure the difference between the reconstructed time series and the original time series; the clustering loss is used to measure the difference between the probability soft assignment matrix Q and the target distribution matrix T; The reconstruction loss is: i is the input data, b i Output data for the decoder; The clustering loss is: 7) In each training cycle, forward propagation is performed to obtain model output and intermediate output; the derivative dynamic time warping distance matrix, soft assignment matrix Q and target distribution matrix T are calculated; the total loss is calculated, and back propagation and optimization of model parameters are performed; 8) After each training cycle, the adjusted Rand index is used to evaluate the consistency between the current clustering result and the previous clustering result. If the ARI is greater than 0.98 for 7 consecutive times, it is considered that the clustering result does not change and the model converges. The cluster center C = {c1, c1, …, c k }.

8. The method for generating a typical microgrid scenario according to claim 6, characterized in that: The specific steps of the second part include: Step 1: Initialize clustering. Given the initial cluster C = {c1, c1, …, c k }, each cluster center corresponds to a leaf node; Step 2: Pattern recognition, find representative patterns for each k-1 cluster, which will be used to further build a decision tree; Step 3: Split the data set S into two groups S according to whether it contains the sequence pattern p p and Step 4: Pattern selection, using a relative risk-based splitting metric to select a discriminative pattern p for tree splitting; Step 5: Relative risk assessment and internal similarity measurement. Calculate the relative risk RR of pattern p, which is the ratio of its support in the positive class S+ to its support in the negative class S-: To evaluate the quality of pattern p, we use the internal similarity measure SIM, which takes into account the pattern length and the number and length of sequences in the positive class; where |p| is the length of pattern p, |S + | is a positive class S + The number of sequences in |s i |For S + medium sequence s i Length; Step 6: Multi-class expansion, for output k categories, expand the concept of relative risk and define the positive class S + To support the cluster with the highest pattern p, and use the one-vs-rest strategy to obtain the negative class S - ; Step 7: Mode selection: if multiple modes have the same relative risk, the mode with the largest SIM value is selected as the basis for dividing the data set; Step 8: Tree construction, using the top-1 distinguishing sequence pattern as the split point for building the tree, and assigning the sequences containing this pattern to the same cluster; According to the k cluster centers obtained in the first part, the root node N is initialized and the pattern tree is constructed. The tree construction follows the top-down binary split method; Determine whether the required number of leaf nodes has been reached, and decide whether to continue splitting based on the minimum number of sequences contained. The judgment standard is k<2 |S|≤min(k,minS) Where minS is the minimum sequence number; Apply the sequential pattern mining algorithm to each cluster center c i Extract sequence patterns, assign sequences containing pattern p to the right node, and those not containing pattern p to the left node, until the stopping condition is met, and the construction of an interpretable sequence tree is completed.

9. The method for generating a typical microgrid scenario according to claim 1, characterized in that: A plurality of sample points are stored in the load-photovoltaic output two-dimensional time series, and each sample point records load data and photovoltaic output data of the same time series.

10. A microgrid typical scenario generation system, characterized in that: include: Sequence generation module, encoding and dimension reduction module, derivative dynamic time warping distance acquisition module, clustering module and decoding and dimension enhancement module; The sequence generation module is used to form a plurality of load-photovoltaic output two-dimensional time series based on the microgrid historical load data and the historical photovoltaic output data; The encoding and dimensionality reduction module is used to encode and reduce the dimensionality of the load-photovoltaic two-dimensional output time series through a deep neural network learning method to form a plurality of load-photovoltaic output one-dimensional time series; The derivative dynamic time warping distance acquisition module is used to measure the similarity between the load-photovoltaic output one-dimensional time series using the derivative dynamic time warping algorithm to obtain the derivative dynamic time warping distance between the load-photovoltaic output one-dimensional time series; The clustering module is used to perform interpretable sequence tree clustering on the load-photovoltaic output one-dimensional time series based on the derivative dynamic time warping distance, form an interpretable sequence clustering tree as a clustering result, decode and upgrade the clustering result, and form a microgrid typical scenario set; The decoding and dimensionality-increasing module is used to decode and upgrade the clustering results to form a typical scenario of a microgrid.

Images