SOM clustering algorithm-based new energy output prediction scene generation optimization method

Abstract

The invention relates to the related technical field of scene generation, in particular to a new energy output prediction scene generation optimization method based on an SOM clustering algorithm. The method comprises the following steps: acquiring new energy data; optimizing the new energy data through a sample optimization strategy based on Kalman gain; based on an SOM clustering algorithm, performing source load data matching on the new energy data, and clustering the source load data; generating a source load day sequence according to a Markov random process to obtain a source load data sequence of a future day; and through a GAN network, generating a new energy output scene based on the source load data sequence of the future day and the cluster of the source load data.

Description

Technical Field

[0001] The present application relates to the technical field related to scenario generation, and specifically to a new energy output prediction scenario generation optimization method based on a SOM clustering algorithm. Background Art

[0002] In the electricity market, the output characteristics of renewable energy make it the main uncertainty factor faced by market participants, which has become an important source of risk in operations under the current market rules. The problem of renewable energy consumption, that is, the waste of resources and inefficient system configuration caused by wind and solar power abandonment, has become one of the key issues restricting the development of renewable energy.

[0003] In order to meet these challenges, new energy output forecasting technology has emerged. Forecasting technology can help system dispatchers and market participants understand the output of new energy in advance, so as to make more reasonable dispatch plans and market decisions. New energy forecasting is an indispensable part of power system dispatching and power market, and scenario generation, as an important pre-process for new energy output forecasting, is becoming increasingly important. Summary of the invention

[0004] In view of this, the embodiments of the present application are directed to providing a new energy output prediction scenario generation optimization method based on the SOM clustering algorithm to perform scenario generation optimization.

[0005] The present application provides a method for optimizing the generation of new energy output prediction scenarios based on a SOM clustering algorithm, including:

[0006] Obtain new energy data;

[0007] Optimizing the new energy data by a sample optimization strategy based on Kalman gain;

[0008] Based on the SOM clustering algorithm, source-load data matching is performed on the new energy data, and the source-load data is clustered;

[0009] According to the Markov random process, the source-load daily sequence is generated to obtain the source-load data sequence of the future day;

[0010] Through the GAN network, based on the source-load data sequence of the future day and the clustering of the source-load data, the scenario of new energy output is generated.

[0011] In some embodiments, the energy data includes new energy output data and load data;

[0012] In some embodiments, the optimizing the new energy data by using a sample optimization strategy based on a Kalman gain includes:

[0013] Completing samples and expanding data of the new energy data based on time series supplementation and cross-section supplementation;

[0014] Based on the sample optimization strategy of Kalman gain, the parameters used in sample completion and data expansion are adjusted to make the completed and expanded data conform to statistical laws.

[0015] In some embodiments, the performing source-load data matching on the new energy data and clustering the source-load data based on the SOM clustering algorithm includes:

[0016] Construct dimensionality reduction indicators;

[0017] Performing dimension reduction on the new energy data based on the dimension reduction index to obtain dimension-reduced data;

[0018] Based on the SOM clustering algorithm, the data after dimensionality reduction is clustered to facilitate the division of scenes based on the clustering results.

[0019] In some embodiments, the dimension reduction indicators include: the total time period of a day, the load samples of the t time period, the average load rate of the day, the peak-to-valley difference ratio of the day, the maximum load of the day, and the average load of the day.

[0020] In some embodiments, the generating of the source-load day sequence according to the Markov random process to obtain the source-load data sequence of the future day includes:

[0021] Use the Markov process to describe the uncertainty of data corresponding to different scenarios and obtain the Markov state transfer matrix;

[0022] The source load data sequence for future days is obtained through Monte Carlo sampling.

[0023] In some embodiments, the generating of a scenario of new energy output based on the source-load data sequence of the future day and the clustering of the source-load data by the GAN network includes:

[0024] Input the source-load data sequence of the future day into the trained GAN network to generate a scenario of new energy output;

[0025] The GAN network is pre-trained and is used to automatically generate new energy output prediction scenarios under a given source-load data sequence for future days.

[0026] The present application also provides a new energy output prediction scenario generation optimization device based on the SOM clustering algorithm, comprising:

[0027] An acquisition module is used to acquire new energy data;

[0028] An optimization module, used for optimizing the new energy data by a sample optimization strategy based on Kalman gain;

[0029] A clustering module, used for matching the new energy data with source-load data and clustering the source-load data based on a SOM clustering algorithm;

[0030] The first generation module is used to generate a source-load daily sequence according to a Markov random process to obtain a source-load data sequence for future days;

[0031] The second generation module is used to generate a scenario of new energy output based on the source-load data sequence of the future day and the clustering of the source-load data through a GAN network.

[0032] The present application also provides an electronic device, comprising:

[0033] A processor, and a memory for storing a program executable by the processor;

[0034] The processor is used to implement the above-mentioned new energy output prediction scenario generation optimization method based on the SOM clustering algorithm by running the program in the memory.

[0035] The present application also provides a computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the processor executes the above-mentioned new energy output prediction scenario generation optimization method based on the SOM clustering algorithm.

[0036] The present application provides a method for optimizing the generation of new energy output prediction scenarios based on the SOM clustering algorithm, firstly obtaining new energy data; optimizing the new energy data through a sample optimization strategy based on the Kalman gain; matching the new energy data with the source-load data based on the SOM clustering algorithm, and clustering the source-load data; generating a source-load daily sequence according to the Markov random process to obtain a source-load data sequence for future days; and generating a new energy output scenario based on the source-load data sequence for future days and the clustering of the source-load data through the GAN network. In the scheme provided by the present application, the Kalman gain is an effective filtering technology that can reduce noise and uncertainty in the data and improve the accuracy and reliability of the data. The SOM clustering algorithm can effectively reduce the dimensionality of high-dimensional data, extract key features, and simplify the data structure. The clustering results can be used to predict the interaction between new energy output and load, and improve the accuracy and effectiveness of the prediction model. The Markov process can describe the uncertainty differences between different clusters and provide a probabilistic description for new energy output prediction. Through Monte Carlo sampling, source-load data sequences for future days can be generated based on the Markov state transition matrix to provide data support for long-term prediction. The GAN network can automatically generate new energy output prediction scenarios based on source-load data sequences for future days, improving the automation of the prediction process. The generated scenarios more accurately meet the principle of source-load interaction balance in power decision-making, improving the applicability and practicality of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0038] Figure 1 It is a flow chart of a new energy output prediction scenario generation optimization method based on a SOM clustering algorithm provided in an embodiment of the present application.

[0039] Figure 2 It is a structural schematic diagram of a new energy output prediction scenario generation optimization device based on a SOM clustering algorithm provided in an embodiment of the present application.

[0040] Figure 3 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0042] The main solution to the uncertainty of new energy is to predict its time-series output through various technical methods. New energy forecasting is an indispensable part of system scheduling and participation in the power market. As far as the output forecast of new energy is concerned, scenario generation is an important pre-process.

[0043] Compared with traditional statistical analysis methods, scenario generation and the use of scenario-corresponding probabilities to describe the time series output of renewable energy are more in line with the output characteristics of renewable energy and the actual needs of the business. On the one hand, from the perspective of the characteristics of renewable energy output, although the current research on the output distribution of photovoltaic and wind power is quite complete, its time series has problems such as instability and spatiotemporal transfer probability; on the other hand, from the perspective of business needs, we are more concerned about the characteristic information in renewable energy output that is more closely related to business goals, such as some output frequency bands associated with the dispatch plan, and some typical daily characteristics.

[0044] In general, the methods for generating scenarios of renewable energy output can be roughly divided into three categories: distribution model, clustering algorithm, and sample sampling method. Among them, the distribution model method requires us to accurately know the specific distribution of variables, and requires samples of rich types and sizes; clustering algorithms have the problem of slow convergence when processing large-scale samples when clustering accuracy is required; sample sampling methods include Monte Carlo, Latin hypercube and other sampling methods. Although these methods are easy to operate, the generated scenarios still contain a large number of scenarios that lack similarity and are difficult to use in actual analysis.

[0045] Therefore, the previous methods for generating output scenarios of new energy sources always have more or less defects, and it is difficult to directly apply them to the prediction of new energy output or even as a basis for subsequent decision-making. In fact, whether in power spot trading or system dispatching decisions, we cannot ignore the coupling between the source and the load. Specifically, in the prediction of new energy sources, we are more concerned about the prediction results of the output of new energy sources under different load conditions. In other words, the output scenario of new energy sources is closely related to the load conditions. We either need the conditional probability of the two, or use the load level as the boundary condition for the former prediction.

[0046] To sum up, the generation of scenarios for new energy output is a prerequisite for new energy prediction and the data basis for predictive analysis. However, the generation of new energy scenarios requires grasping the main characteristic information in the sample, eliminating scenarios that lack application value, and being consistent with business needs.

[0047] The present invention relates to a method for optimizing the generation of a new energy output prediction scenario based on a SOM clustering algorithm, which comprises the following steps: (1) sample optimization based on Kalman gain; (2) source-load data matching based on the SOM clustering algorithm; (3) source-load daily sequence generation based on a Markov random process; and (4) realizing the generation of new energy output scenarios through a GAN network.

[0048] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be described in detail with reference to the accompanying drawings.

[0049] Exemplary Methods

[0050] Figure 1 FIG. 1 is a flow chart of a method for optimizing the generation of new energy output prediction scenarios based on a SOM clustering algorithm provided by an embodiment of the present application. Figure 1 As shown, the method includes the following contents.

[0051] Step S110: Acquire new energy data;

[0052] In this step, we first collect renewable energy power generation data, including but not limited to real-time output data of wind and solar energy. These data are the basis for subsequent analysis and prediction. Data collection can be obtained directly from the monitoring system of renewable energy power generation stations or from electricity market operators. The acquired data needs to include sufficient historical information for trend analysis and pattern recognition.

[0053] Step S120: optimizing the new energy data by a sample optimization strategy based on Kalman gain;

[0054] After acquiring the new energy data, we use the Kalman filter to optimize the data. The Kalman filter is a recursive algorithm that estimates the state of a dynamic system from a series of noisy measurements. In this step, we use the Kalman gain to adjust the data to reduce random noise and systematic errors, improving the accuracy and reliability of the data. In addition, for missing or incomplete data, the Kalman filter can predict and supplement these data points to ensure the integrity and continuity of the data set.

[0055] Step S130: Based on the SOM clustering algorithm, matching the new energy data with the source-load data is performed, and the source-load data is clustered;

[0056] After data optimization, we use the self-organizing map (SOM) clustering algorithm to match and cluster the new energy data and load data. SOM is an unsupervised neural network algorithm that can map high-dimensional data to low-dimensional space while maintaining the topological structure of the data. In this step, we reduce the dimensionality of the features of the new energy output data and load data, and classify the data into different categories based on these features, namely typical day and extreme day data. Such clustering helps us identify data sets with similar features, providing a more accurate data basis for subsequent predictions and analysis.

[0057] Step S140: Generate a source-load daily sequence according to the Markov random process to obtain a source-load data sequence for future days;

[0058] After completing the data clustering, we use the Markov random process to simulate and predict the source-load data sequence of future days. The Markov process is a random process whose future state depends only on the current state and has nothing to do with the previous historical state. In this step, we construct a state transition matrix to describe the transition probability between different cluster states. Through Monte Carlo simulation, we can generate a series of possible source-load data sequences for future days, which reflect the uncertainty and randomness of renewable energy output and load.

[0059] Step S150: Generate a scenario of new energy output based on the source-load data sequence of the future day and the clustering of the source-load data through the GAN network;

[0060] Finally, after obtaining the source load data sequence for future days, we use a generative adversarial network (GAN) to generate scenarios for renewable energy output. GAN consists of two parts: a generator and a discriminator. The generator is responsible for generating new data samples based on the input noise sequence, while the discriminator is responsible for distinguishing between generated data and real data. In this step, we use the source load sample data in each cluster to train the GAN model. When the model training is complete, we can input the source load data sequence for future days into the GAN model, and the generator will automatically generate renewable energy output prediction scenarios that match these data. These scenarios include not only output prediction values, but also corresponding prediction probabilities, providing support for the dispatch and decision-making of the power system.

[0061] Through these steps, we can generate more accurate and reliable new energy output forecast scenarios, providing a scientific basis for the stable operation of the power system and the efficient operation of the power market.

[0062] Specifically, the energy data includes new energy output data and load data; the energy data we collect includes two main parts:

[0063] New energy output data:

[0064] Wind energy output data: including real-time power output, wind speed, wind direction and other parameters of wind turbines.

[0065] Solar power output data: including real-time power output, sunshine intensity, temperature and other parameters of solar photovoltaic panels.

[0066] Historical output data: Output records of renewable energy sources over a period of time, used to analyze output patterns and trends.

[0067] Load data:

[0068] Electricity demand data: The electricity demand of the power system at different time periods, usually recorded at hourly or 15-minute intervals.

[0069] Historical load data: Electricity demand data at different time periods in history, used to analyze load patterns and predict future demand.

[0070] Weather and environmental data: Weather conditions that may affect the output of renewable energy, such as temperature, humidity, air pressure, etc.

[0071] In some embodiments, the new energy data is optimized through a sample optimization strategy based on Kalman gain, including: performing sample completion and data expansion on the new energy data based on time series supplementation and cross-section supplementation; based on the sample optimization strategy based on Kalman gain, adjusting the parameters used in sample completion and data expansion so that the completed and expanded data conform to statistical laws.

[0072] Specifically, when we conduct short-term electricity trading and unit scheduling, the forecast of renewable energy output and the load level in the same period are not separate processes, but coupled. Therefore, the relevant decisions are part of the interactive balance between system sources and loads.

[0073] Based on this consideration, in the process of generating new energy output forecast scenarios, we will coordinate the load time series and new energy samples of the same time scale and size.

[0074] Taking into account that many small and medium-sized new energy stations may have problems such as missing samples and insufficient samples, sample completion and data expansion must be carried out during data processing. When expanding the data, it is necessary to ensure data accuracy and avoid problems such as pseudo-randomness of samples due to improper avoidance methods.

[0075] To this end, we use the idea of ​​Kalman gain to optimize sample data. During the data optimization process, we do not distinguish between new energy output data and load data, but process the source and load data samples in a unified manner.

[0076] First, we supplement the samples in time series and cross-section. Cross-section supplementation refers to the sample data that we add or supplement using the data of two adjacent periods on the same day, and time series supplementation refers to the sample data that we add or supplement using the samples of the same period on two consecutive days. The details are as follows

[0077]

[0078] In the above formula, x i,t represents the source load sample data of the tth period of the i-th day, i∈(1,2,…365), t∈(1,2,…96), the sample value range is 365 days in a year and 96 periods in a day (sampling is performed every 15 minutes); x c,t represents cross-sectional supplementary data expanded by data from two adjacent periods on the same day; x s,t It indicates the time series supplementary data expanded by two adjacent samples of the same period; γ c represents a real number randomly drawn in the interval [0,1] when the cross section is supplemented; γ s Represents a real number randomly drawn in the interval [0,1] when supplementing the time series.

[0079] In order to improve the accuracy of the augmented data, we merge the two types of augmented data given in the above formula as follows

[0080] x f,t =x c,t +β K (x s,t -x c,t ) (2)

[0081] Among them, x f,t represents the supplementary sample data after Kalman gain fusion; β K represents the Kalman gain coefficient, which can be given by the following formula

[0082]

[0083] Among them, σ 2 (x c,t ) and σ 2 (x s,t ) represent the variance of the two types of supplementary samples, which can be obtained by the following formula

[0084]

[0085] The above formula is the variance formula of the expanded sample, where and Respectively represent the means of the two expanded samples; and Represent the total number of two types of expanded samples respectively.

[0086] The supplementary data after Kalman gain fusion can reflect the horizontal trend of samples in each time period within the day and the vertical trend of samples in the same time period on different dates, which is more in line with statistical laws and improves the accuracy and effectiveness of subsequent algorithms.

[0087] In some embodiments, the source-load data matching and source-load data clustering of the new energy data based on the SOM clustering algorithm include: constructing a dimensionality reduction index; reducing the dimension of the new energy data based on the dimensionality reduction index to obtain the reduced-dimensional data; clustering the reduced-dimensional data based on the SOM clustering algorithm to facilitate the division of scenes based on the clustering results. The dimensionality reduction index includes: the total time period of a day, the load sample of the t time period, the average load rate of the day, the peak-to-valley difference ratio of the day, the maximum load of the day, and the average load of the day.

[0088] Specifically, as mentioned above, in our short-term decision-making of the power system, the output of new energy and the load changes in the same sequence are not isolated, but the source-load balance must be considered. At the same time, due to the SCED constraint principle of the entire grid, we pay more attention to those time period data with typical scenario analysis significance, such as the source-load situation during peak and valley periods.

[0089] Therefore, we need to use the load scenario as the conditional distribution or boundary for generating new energy output scenarios. This requires us to cluster the new energy output samples and load samples of the same scale and time series after data optimization in the previous part.

[0090] The goal of clustering is to classify the optimized load samples into typical day data and extreme day data according to the characteristic information, and then merge the corresponding new energy output samples into clusters with the same characteristics. Since the samples are sampled every 15 minutes, there are 96 samples per day. If we cluster directly by Euclidean distance or Kantorovich distance, the correlation between sample points will be weakened. To solve this problem, we construct the following indicators to reduce the dimension of the sample features.

[0091]

[0092] In the above formula, T represents the total time period of a day; l(t) represents the load sample in time period t; ζ represents the average load rate of the day; η represents the peak-to-valley difference ratio of the day; θ represents the maximum load of the day; Indicates the average load for the day.

[0093] The optimized daily load samples are processed according to formula (5). Assuming that the sample sampling still uses one year of samples, the 96×365 samples of the intraday load are replaced by a daily load index sequence with a length of 365, which we record as

[0094]

[0095] Where z1, z2, z3, z4 represent the corresponding indicator variables ζ, η, θ, They are time series of equal length.

[0096] Next, we use the SOMA (Self-Organizing Map) clustering algorithm to process the indicators after dimensionality reduction.

[0097] SOMA clustering algorithm is a competitive learning algorithm based on neural network. We use Kohonen network structure, which includes two layers, input layer and calculation layer. The calculation layer is a square matrix composed of neurons, each neuron is connected to all source nodes of the input layer.

[0098] First, initialize the weights of each neuron. For the i-th input node, we input a four-dimensional vector, then the connection weight between input node i and neuron j can be expressed as

[0099]

[0100] Where N represents the total number of neurons in the computational layer. The Euclidean distance between input node i and neuron j is

[0101]

[0102] The Euclidean distance in the above formula is used to determine the winning node. That is, in the same iteration step, we input samples at different positions in the time series to the input layer nodes at the same time. For the neuron j in the same connection layer, the winning node is the node with the smallest Euclidean distance. Then the input sample of the winning node will be mapped to the discrete space of the neuron.

[0103] In order to describe the discrete space of neurons, we need to construct its topological structure. As mentioned earlier, the computational layer neurons of the Kohonen network are a square matrix of rows and columns. We regard it as a mesh graph with a geometric structure, then the topological domain of neurons i and j can be expressed as

[0104]

[0105] In the above formula, represents the topological domain kernel function between neurons i and j, where the kernel function adopts the form of normal distribution; I(z) represents the input index of the winning input node after Euclidean distance judgment according to formula (8); represents the square of the horizontal distance between neuron j and the winning node in the mesh graph; σ 2 represents the variance parameter of the kernel function, which decays with the iteration time, specifically:

[0106]

[0107] Among them, σ0 represents the initial value of σ; t represents the iteration step; τ σ is the decay rate parameter of σ, which is a built-in parameter of the algorithm.

[0108] Based on the topological domain of neurons and nodes, in order to achieve the mapping of feature information from input to output, the weights of the winning nodes and connected neurons will be updated with the iteration process. The specific process is as follows:

[0109]

[0110] in, is the learning rate, which is a parameter that depends on the iteration time. The specific expression is

[0111]

[0112] in, Represents the initial learning rate; Represents the decay rate parameter of the learning rate.

[0113] After the above process, the neurons in the computing layer divide the disordered input features into an ordered topological network through an adaptive learning process, thereby achieving the purpose of clustering discrimination.

[0114] When the algorithm reaches convergence in the set maximum iteration step, we get the best clustering result for the daily load index. For each cluster, we merge the new energy output samples in the same time period into one cluster, so that we get the clustering of new energy output samples based on different load index clusters.

[0115] In some embodiments, the source load daily sequence is generated according to the Markov random process to obtain the source load data sequence for future days, including: using the Markov process to describe the uncertainty of data corresponding to different scenarios to obtain the Markov state transfer matrix; and obtaining the source load data sequence for future days through Monte Carlo sampling.

[0116] In the previous part, we completed the clustering of source-load data. In fact, we hope to obtain a set of renewable energy under different daily load characteristics, and then divide the scene. However, there is no division of uncertainty or probability between clusters, so we also need to characterize the uncertainty differences between different clusters, and then generate a new daily source-load sequence as the scene generation material for renewable energy output prediction.

[0117] We do not want to examine the specific distribution of source load data, so we can use a Markov process to describe the uncertainty between them.

[0118] Define the Markov state transfer matrix of the source load-day sequence type, specifically:

[0119]

[0120] The elements in the matrix represent the transition probability between two states, specifically:

[0121] ψ kl =p(λ s+1 = l|λ s =k) ​​(14)

[0122] Among them, ψ kl represents the conditional probability of the source load variable transferring from state k to state l; p(·) represents the probability function; λ s Indicates the state value corresponding to the sampling step s. Here, the state can be understood as the specific clustering of the source load data obtained by the SOM algorithm in the previous part, that is, the daily classification type of the sample.

[0123] The state transition probability can be obtained through sampling fitting, and the specific expression is

[0124]

[0125] Among them, n kl It indicates the cumulative frequency of the variable transferred from k to l at the end of sampling; It represents the sum of all frequencies of the variable transferring from other states to k.

[0126] Based on the Markov state transfer matrix, the specific process of obtaining the source-load data daily sequence of a given interval in the future through Monte Carlo sampling involves the following steps:

[0127] Constructing the Markov state transition matrix

[0128] First, it is necessary to determine the states of the Markov model, which represent different categories or clustering results of the source-load data. Then, based on historical data, the transition probabilities between these states are calculated to construct a state transition matrix. Each element P_ij in the state transition matrix represents the probability of transitioning from state i to state j.

[0129] 1. Determine the initial state

[0130] Determine the initial state at the start of the simulation. This can be the most likely state based on historical data, or a specific state preset based on specific business needs.

[0131] 2. Monte Carlo Sampling Process

[0132] Monte Carlo sampling is a statistical method that computes possible outcomes in a system by repeatedly taking random samples. In the context of Markov chains, this process involves:

[0133] 3. Initialization: Set the total number of simulation steps (i.e. the number of future days) and the initial state.

[0134] Iterative sampling:

[0135] For each step, the next state is randomly chosen based on the current state and the state transition matrix. This can be done by generating a random number and looking up the cumulative probability distribution to determine the next state.

[0136] Updates the current state to the newly selected state.

[0137] Repeat: Repeat the above sampling steps until all simulation days are completed.

[0138] 4. Generate day sequence

[0139] In each iteration, one day's source load data is generated based on the source load data characteristics corresponding to the current state. This can be achieved in the following ways:

[0140] Mapping of state to data: Each state corresponds to a specific set of source-load data features. Based on these features, specific source-load data can be generated.

[0141] Random fluctuations: To simulate real fluctuations, randomness can be introduced into the generated data. For example, random noise can be added based on the state features.

[0142] 5. Repeat the simulation

[0143] In order to improve the reliability of the results and reduce the impact of randomness, the above process needs to be repeated many times. Each simulation starts from the initial state and independently generates a future day sequence.

[0144] 6. Summarize the results

[0145] Finally, the results of all simulations are summarized to obtain the possible distribution of source load data within a given interval in the future. This can be used for further analysis, such as calculating statistical indicators (mean, variance, etc.) or conducting risk assessment.

[0146] In this way, Monte Carlo sampling combined with Markov chains can effectively simulate the uncertainty and randomness of future source and load data, providing valuable insights for the planning and operation of power systems.

[0147] In some embodiments, the generating of a scenario of new energy output based on the source-load data sequence of the future day and the clustering of the source-load data by the GAN network includes:

[0148] The source-load data sequence of the future day is input into the trained GAN network to obtain the scenario generation of new energy output; wherein the GAN network is pre-trained and is used to automatically generate the new energy output prediction scenario under the given source-load data sequence of the future day.

[0149] After completing the above two parts, we obtained the clustering of source-load samples and generated the source-load data sequence for future days through the Markov chain.

[0150] Our goal is to train the neural network to automatically generate renewable energy output forecast scenarios under given load constraints. To this end, we can use the source-load sample data in each cluster to train a GAN model. When we use the method in the third part to obtain the source-load data sequence for future days, we input it into the trained GAN model, and we can automatically output the corresponding scenarios and corresponding forecast probabilities that match the future day data.

[0151] For the GAN to be trained, its business structure consists of two parts, namely the discriminator and the generator. The noise sequence is input into the generator, and the generator generates the output result based on the noise sequence; the result generated by the generator is input into the discriminator together with the real sample sequence, and then the discriminator is asked to make a judgment. The error between the generator result and the real result is the loss function. The implementation logic of the training is to continuously iterate and optimize the prediction level of the generator so that it is close to the real sample, so that the real sample plays the role of the training sample.

[0152] We use the source-load sample data in any cluster as training samples to illustrate the training process. The source-load data in the cluster is denoted as

[0153]

[0154] Where m represents the total number of source-load samples in this cluster. The distribution function of

[0155] The noise sequence in the input generator is denoted as ε, and its distribution function is denoted as f(ε); the output of the generator is denoted as G(ε), and the probability distribution of the generated data is denoted as The real sample data Input discriminator, its output data is recorded as Represents the judgement of the is the probability of a true sample. Then, the loss functions of the generator and the discriminator are

[0156]

[0157] Among them, L G and LD They represent the loss functions of the generator and discriminator respectively; E(·) represents the expectation; D(G(ε)) represents the probability of judging the generator data as real data after inputting it into the discriminator.

[0158] The game training model of the generator and discriminator can be expressed as

[0159]

[0160] In order to improve the prediction accuracy and convergence effect, we improve the original GAN ​​and introduce the mean square error into the loss function of the model.

[0161]

[0162] Where R represents the total number of samples. In addition, since the original GAN ​​uses Jensen–Shannon (JS) as the loss function in the discriminator D, it is prone to training instability. Therefore, the Wasserstein distance is introduced to measure the distance between the real data and the generated data distribution, which is defined as follows:

[0163]

[0164] In the above formula, sup(·) represents the upper bound of the function value; K is the Lipschitz constant; ‖f‖L≤K means that the component functions of the discriminator D must satisfy K-Lipschitz continuity.

[0165] Introducing a penalty term in formula (20), the training model can be reformulated as

[0166]

[0167] In the above formula, τ represents the penalty coefficient; is the gradient operator; Represents the linear processing term of the real sample, and the expression is

[0168]

[0169] Among them, ξ represents the assignment coefficient, which ranges from 0 to 1.

[0170] In the early stages of training, there is a large difference between the data generated by the generator and the real data, and the discriminator can easily distinguish them. Then, through continuous optimization and training, the generation and discrimination capabilities of the generator and discriminator continue to improve until the discriminator has difficulty distinguishing between the generated data and the real data, and the model training is completed.

[0171] After completing the model training, when we input the source load data sequences of future days into these trained GAN models, they can automatically generate corresponding scenarios and give scenario probabilities.

[0172] In summary, the present invention proposes a method for optimizing the generation of new energy output prediction scenarios based on the SOM clustering algorithm. Starting from the coupling of new energy output prediction scenarios and load data in time series and the business needs for typical scenarios and extreme scenarios, the SOM clustering algorithm is used to realize the integrated clustering of source-load data; then the Markov chain and Monte Carlo sampling are used to describe the transfer probability of the clustering results, so as to obtain the future daily source-load data for a given period; finally, the results of the previous clustering work are used to train the GAN model. After the training is completed, the GAN can automatically generate corresponding scenarios based on the future daily source-load sequence. The scenarios generated by this method avoid the defects of low scenario applicability, little feature information, and unclear clustering standards in a single clustering algorithm, and accurately meet the principle that power decision-making is the interactive balance of source and load. .

[0173] Exemplary Devices

[0174] The device embodiments of the present application can be used to execute the method embodiments of the present application. For details not disclosed in the device embodiments of the present application, please refer to the method embodiments of the present application.

[0175] Figure 2 FIG. 1 is a block diagram of a new energy output prediction scenario generation optimization device based on a SOM clustering algorithm provided by an embodiment of the present application. Figure 2 As shown, the device comprises:

[0176] An acquisition module 21 is used to acquire new energy data;

[0177] An optimization module 22, configured to optimize the new energy data by using a sample optimization strategy based on Kalman gain;

[0178] A clustering module 23 is used to match the new energy data with source-load data and cluster the source-load data based on a SOM clustering algorithm;

[0179] The first generating module 24 is used to generate a source-load daily sequence according to a Markov random process to obtain a source-load data sequence for future days;

[0180] The second generation module 25 is used to generate a scenario of new energy output based on the source-load data sequence of the future day and the clustering of the source-load data through a GAN network.

[0181] Exemplary Electronic Devices

[0182] Below, reference Figure 3To describe an electronic device according to an embodiment of the present application. Figure 3 A block diagram of an electronic device according to an embodiment of the present application is illustrated.

[0183] like Figure 3 As shown, electronic device 300 includes one or more processors 310 and memory 320 .

[0184] The processor 310 may be a central processing unit (CPU) or other forms of processing units having data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 300 to perform desired functions.

[0185] The memory 320 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 310 may run the program instructions to implement the above-described various embodiments of the present application, the new energy output prediction scenario generation optimization method based on the SOM clustering algorithm and / or other desired functions. Various contents such as category correspondences may also be stored in the computer-readable storage medium.

[0186] In one example, the electronic device 300 may further include: an input device 330 and an output device 340 , and these components are interconnected via a bus system and / or other forms of connection mechanisms (not shown).

[0187] In addition, the input device 330 may also include, for example, a keyboard, a mouse, an interface, etc. The output device 340 may output various information to the outside, including analysis results, etc. The output device 340 may include, for example, a display, a speaker, a printer, a communication network and a remote output device connected thereto, etc.

[0188] Of course, to simplify, Figure 3 Only some of the components in the electronic device related to the present application are shown, and components such as a bus, an input / output interface, etc. are omitted. In addition, the electronic device may further include any other appropriate components according to specific application conditions.

[0189] Exemplary computer program products and computer-readable storage media

[0190] In addition to the above-mentioned methods and devices, an embodiment of the present application may also be a computer program product, which includes computer program instructions, which, when executed by a processor, enable the processor to execute the steps of the new energy output prediction scenario generation optimization method based on the SOM clustering algorithm according to various embodiments of the present application described in the above "Exemplary Method" section of this specification.

[0191] The computer program product may be written in any combination of one or more programming languages ​​to write program codes for performing the operations of the embodiments of the present application, including object-oriented programming languages, such as Java, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as an independent software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0192] In addition, an embodiment of the present application may also be a computer-readable storage medium having computer program instructions stored thereon, which, when executed by a processor, enables the processor to execute the steps of the new energy output prediction scenario generation optimization method based on the SOM clustering algorithm according to various embodiments of the present application described in the above "Exemplary Method" section of this specification.

[0193] The computer readable storage medium can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can include, for example, but is not limited to, a system, device or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0194] The above description has been given for the purpose of illustration and description. In addition, this description is not intended to limit the embodiments of the present application to the forms disclosed herein. Although multiple example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.

Claims

1. A method for optimizing the generation of new energy output prediction scenarios based on a SOM clustering algorithm, characterized in that: include: Obtain new energy data; Optimizing the new energy data by a sample optimization strategy based on Kalman gain; Based on the SOM clustering algorithm, matching the new energy data with the source-load data is performed, and the source-load data is clustered; According to the Markov random process, the source-load daily sequence is generated to obtain the source-load data sequence of the future day; Through the GAN network, based on the source-load data sequence of the future day and the clustering of the source-load data, the scenario of new energy output is generated.

2. The method for generating and optimizing new energy output prediction scenarios based on the SOM clustering algorithm according to claim 1 is characterized in that: The energy data includes new energy output data and load data.

3. The method for generating and optimizing new energy output prediction scenarios based on the SOM clustering algorithm according to claim 1 is characterized in that: The optimizing the new energy data by using a sample optimization strategy based on Kalman gain includes: Completing samples and expanding data of the new energy data based on time series supplementation and cross-section supplementation; Based on the sample optimization strategy of Kalman gain, the parameters used in sample completion and data expansion are adjusted to make the completed and expanded data conform to statistical laws.

4. The method for generating and optimizing new energy output prediction scenarios based on the SOM clustering algorithm according to claim 1 is characterized in that: The method of matching the new energy data with the source-load data and clustering the source-load data based on the SOM clustering algorithm includes: Construct dimensionality reduction indicators; Performing dimension reduction on the new energy data based on the dimension reduction index to obtain dimension-reduced data; Based on the SOM clustering algorithm, the data after dimensionality reduction is clustered to facilitate the division of scenes based on the clustering results.

5. The method for generating and optimizing new energy output prediction scenarios based on the SOM clustering algorithm according to claim 4 is characterized in that: The dimension reduction indicators include: the total time period of a day, the load sample of the t period, the average load rate of the day, the peak-to-valley difference ratio of the day, the maximum load of the day, and the average load of the day.

6. The method for generating and optimizing new energy output prediction scenarios based on the SOM clustering algorithm according to claim 4 is characterized in that: The source-load daily sequence is generated according to the Markov random process to obtain the source-load data sequence for the future day, including: Use the Markov process to describe the uncertainty of data corresponding to different scenarios and obtain the Markov state transfer matrix; The source load data sequence for future days is obtained through Monte Carlo sampling.

7. The method for generating and optimizing new energy output prediction scenarios based on the SOM clustering algorithm according to claim 6 is characterized in that: The generating of a scenario of new energy output based on the source-load data sequence of the future day and the clustering of the source-load data by using the GAN network includes: Input the source-load data sequence of the future day into the trained GAN network to generate a scenario of new energy output; The GAN network is pre-trained and is used to automatically generate new energy output prediction scenarios under a given source-load data sequence for future days.

8. A new energy output prediction scenario generation optimization device based on SOM clustering algorithm, characterized in that: include: An acquisition module is used to acquire new energy data; An optimization module, used for optimizing the new energy data by a sample optimization strategy based on Kalman gain; A clustering module, used for matching the new energy data with source-load data and clustering the source-load data based on a SOM clustering algorithm; The first generation module is used to generate a source-load daily sequence according to a Markov random process to obtain a source-load data sequence for future days; The second generation module is used to generate a scenario of new energy output based on the source-load data sequence of the future day and the clustering of the source-load data through a GAN network.

9. An electronic device, characterized in that: include: A processor, and a memory for storing a program executable by the processor; The processor is used to implement the new energy output prediction scenario generation optimization method based on the SOM clustering algorithm as described in any one of claims 1 to 7 by running the program in the memory.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, enables the processor to execute the new energy output prediction scenario generation optimization method based on the SOM clustering algorithm as described in any one of claims 1 to 7.

Images