A source-load dual attribute storage integrated energy system scenario generation method for power grid planning

Abstract

The application provides a source-load dual attribute storage integrated energy system scenario generation method for power grid planning, comprising: constructing a bidirectional storage integrated energy system model, preprocessing source-load historical data, modeling the charging and discharging behaviors of wind energy, solar energy, electric load, thermal load, cold load and energy storage based on a VAE-LSTM coupling model in combination with the integrated energy system model, and generating high-quality scenario data with time sequence dependence. The generated scenario data is evaluated and screened by using statistical indicators and visual indicators, and the screened data meeting the requirements is obtained. In combination with a real-time dispatching decision model, the real-time calculation of grid-connected power generation, power grid power purchase and energy storage charging and discharging is carried out according to the screened scenario data, and the optimal dispatching and management of the energy system are realized. The method not only improves the accuracy and efficiency of power grid planning and dispatching, but also effectively deals with the complex dynamic characteristics of the energy system.

Description

Technical Field

[0001] This invention belongs to the field of power system planning and operation technology, specifically relating to a method for generating integrated energy system scenarios with dual source and load attributes, including storage, for power grid planning. Background Technology

[0002] With the continued growth of global energy demand and increasingly stringent environmental requirements, the limitations of traditional energy systems are becoming increasingly apparent. To address issues such as low energy efficiency and environmental pollution, Integrated Energy Systems (CES), as a new, efficient, and flexible energy management model, have attracted widespread attention. By integrating multiple energy flows such as electricity, heat, and cooling, CES can improve energy efficiency while enhancing system reliability and flexibility. However, with the large-scale application and increasing proportion of renewable energy sources such as wind and solar power, the emergence of large-scale grid-connected integrated energy stations (i.e., electricity users with dual source and load attributes) on the user side has become a new challenge for distribution networks.

[0003] In such bidirectional integrated energy systems, both energy consumption (electricity load, heat load, cooling load, etc.) and energy production (wind power, photovoltaic power, etc.) are involved, resulting in highly uncertain and volatile input-output characteristics. This uncertainty presents significant challenges to the dispatching, operation, and planning of distribution networks. Especially when the power generation of fluctuating renewable energy sources such as wind and solar power is affected by various factors such as meteorological conditions and seasonal variations, effectively modeling these variable energy input-output characteristics becomes a major challenge in power grid planning and optimal dispatch. Against this backdrop, conducting research on scenario generation for bidirectional power users is doubly necessary: ​​firstly, effective scenario generation methods can effectively characterize their variable energy consumption patterns and power generation characteristics; secondly, the generated grid-connected power generation and grid-purchased power will provide important basis for the accurate planning and optimal dispatch of distribution networks. Therefore, in-depth research on the effectiveness of scenario generation methods has significant theoretical and practical value for improving the management and control capabilities of integrated energy stations under new power systems. Thus, scenario generation is necessary for such bidirectional integrated energy systems with energy storage.

[0004] In the prior art, Chinese patent CN119070283B discloses a grid energy storage method that considers the uncertainty and time-series characteristics of wind and solar power. This method includes reducing scenarios to typical seasonal wind and solar load scenarios using the SOM clustering method, establishing a multi-timescale energy storage model considering long-term energy storage, and constructing a grid-wind-solar-storage stochastic programming model that comprehensively considers the economic efficiency and low-carbon nature of grid investment. However, scenario generation methods for integrated energy systems with dual source and load attributes still face three major challenges:

[0005] 1. Applicability issues of probabilistic model assumptions: Many existing scenario generation methods rely on probabilistic models to describe system behavior. However, these assumptions often fail to accurately reflect the actual power output of renewable energy systems such as wind farms and photovoltaic farms, as well as the distribution characteristics of their prediction errors. In complex and ever-changing environmental conditions, constructing a model that is both universally applicable and accurately reflects the actual situation is extremely difficult and can easily lead to biases in scenario data.

[0006] 2. Difficulty in modeling spatiotemporal correlations: Most existing methods focus on modeling a single location or total output, making it difficult to fully consider the spatiotemporal correlations between renewable energy sources in multiple regions and from multiple sources. Because energy output is affected by various factors such as weather and seasons, the output of renewable energy in different regions exhibits significant spatiotemporal coupling characteristics, and existing scene generation methods have obvious limitations in modeling this characteristic.

[0007] 3. The balance between efficiency and accuracy: Large-scale sampling can ensure the accuracy of scene generation, but it is computationally expensive and inefficient. While scene reduction methods improve efficiency, they may compromise scene diversity to some extent, thus affecting the accuracy of the generated results. Therefore, how to improve computational efficiency while ensuring the accuracy of scene generation is an urgent problem to be solved.

[0008] Therefore, in order to effectively improve the planning, scheduling and management capabilities of the power distribution network, it is necessary to develop a more reliable, efficient, and accurate method for generating integrated energy system scenarios that can reflect spatiotemporal characteristics. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method for generating integrated energy system scenarios with dual source and load attributes, including storage, for power grid planning.

[0010] The objective of this invention can be achieved through the following technical solutions:

[0011] This invention provides a method for generating integrated energy system scenarios with dual source and load attributes, including storage, for power grid planning, comprising the following steps:

[0012] Construct a two-way integrated energy system model with energy storage, and determine the various parameters of the integrated energy system model based on the user's energy demand;

[0013] The historical source-load data is preprocessed, including power load, heat load, cooling load, photovoltaic power generation data, and wind power generation data.

[0014] The scene generation method based on the VAE-LSTM coupled model models the charging and discharging behavior of wind power, solar power, cooling / heating and power loads and energy storage based on preprocessed source-load historical data and integrated energy system model, generating high-quality scene data with time-dependent characteristics. The VAE-LSTM coupled model is a model based on variational autoencoder (VAE) and long short-term memory (LSTM) network.

[0015] The generated scene data is evaluated and filtered using statistical and visualization indicators to obtain the filtered scene data.

[0016] Based on the selected scenario data and combined with the real-time grid dispatch decision model, the grid-connected power generation, grid-purchased electricity, and energy storage charging and discharging are calculated in real time to achieve optimized dispatch and management of the energy system.

[0017] Furthermore, the construction of a bidirectional integrated energy system model with energy storage, and the determination of various parameters of the integrated energy system model based on the user's energy demand, specifically includes: setting the capacity, efficiency, and operating mode of each device in the integrated energy system model according to the user's electricity, heat, and cooling demands; defining the power generation capacity, energy storage capacity, and charge / discharge efficiency of each device based on the characteristics of wind power, solar power, and energy storage devices; determining the interaction mode between the power grid and the integrated energy system based on the grid's dispatch requirements, including the grid's load response, power generation dispatch strategy, and the energy storage system's charge / discharge strategy; and forming a bidirectional integrated energy system model with energy storage by establishing the interrelationships and constraints between devices. The integrated energy system model includes energy supply devices, energy coupling devices, and energy storage devices.

[0018] Furthermore, the preprocessing of the source load historical data specifically includes:

[0019] Data cleaning is performed on the historical source-load data, including removing missing and outlier values; the cleaned historical source-load data is then standardized, including scaling the power load, heat load, cooling load, photovoltaic power generation data, and wind power generation data to a range with a mean of 0 and a variance of 1; thus obtaining the preprocessed historical source-load data.

[0020] Furthermore, the scenario generation method based on the VAE-LSTM coupled model models the charging and discharging behavior of wind power, solar power, cooling / heating and power loads, and energy storage based on preprocessed source-load historical data and a comprehensive energy system model, generating high-quality scenario data with time-dependent characteristics, specifically including:

[0021] A VAE-LSTM coupled model is constructed; the VAE-LSTM coupled model is trained based on the preprocessed source-load historical data to obtain the trained VAE-LSTM coupled model; based on the trained VAE-LSTM coupled model and combined with the integrated energy system model, the charging and discharging behavior of wind energy, solar energy, cooling / heating and power loads and energy storage is modeled to generate high-quality scene data with time-dependent characteristics.

[0022] Furthermore, the VAE-LSTM coupled model includes an encoder, a reparameterization module, and a decoder connected in sequence;

[0023] The encoder is used to map source payload data to the latent space, extract temporal features of input data, and output the mean and variance parameters of latent variables. The encoder includes two stacked LSTM layers and several fully connected layers, wherein the first LSTM layer is used to extract the original temporal features, the second LSTM layer is used to generate a high-dimensional representation of the latent space, and the fully connected layers are used to map the output to the mean and variance parameters of latent variables.

[0024] The reparameterization module is used to generate latent variables based on the mean and variance of the encoder output; the reparameterization module includes a random noise sampling unit and a latent variable calculation unit;

[0025] The decoder is used to map the latent variables generated by the reparameterization module back to the data space to reconstruct the time series data of wind energy, solar energy, cooling / heating and power loads and energy storage charging and discharging behavior. The decoder includes two stacked LSTM layers and several fully connected layers. The first LSTM layer is used to receive latent variables and recover time series features, the second LSTM layer is used to gradually generate complete time series output, and the fully connected layers are used to output reconstructed data with the same dimension as the original data.

[0026] Furthermore, the data processing procedure of the VAE-LSTM coupled model includes:

[0027] The source and load data are input into the encoder, including wind power generation, photovoltaic power generation, cooling load, heating load, and electrical load, forming an input sequence. ,in, For the first t Hourly source load feature vector;

[0028] The encoder processes the input sequence through two stacked LSTM layers. The first LSTM layer extracts the original temporal features, and the second LSTM layer generates a high-dimensional representation of the latent space. , represented as:

[0029]

[0030]

[0031]

[0032]

[0033]

[0034]

[0035]

[0036] in, This indicates the hidden state of the previous time step. Enter the current time step; , , These are the outputs of the forget gate, input gate, and output gate, respectively. Candidate memory states; This is the unit state; This indicates element-wise multiplication; These are the forget gate weight matrix and the bias vector, respectively; , These are the input gate weight matrix and the bias vector, respectively; , These are the output gate weight matrix and the bias vector, respectively; , These are the candidate memory unit weight matrix and bias vector, respectively; Output sequence for encoder; For activation functions;

[0037] High-dimensional representation of the encoder output The formula for mapping to the mean and log-variance of latent variables is:

[0038]

[0039]

[0040] in, These are the mean and log-variance mapped to the latent variables, respectively; , , Here are the weight matrix and bias vector of the fully connected layer;

[0041] The reparameterization module is based on the mean. Sum of logarithmic variance Generate latent variables The formula is:

[0042]

[0043]

[0044] in, Indicates variance; To obtain from the standard normal distribution Random noise in the mid-sample, It is the identity matrix;

[0045] latent variables The decoder is input to a given source, which consists of two stacked LSTM layers and several fully connected layers. The first LSTM layer receives latent variables and recovers temporal features, the second LSTM layer progressively generates a complete temporal output, and the fully connected layers output reconstructed data with the same dimension as the original input data. The formula is:

[0046]

[0047] in, These are the network parameters for the decoder; This represents the function mapping of the decoder.

[0048] Furthermore, the step of training the VAE-LSTM coupled model based on the preprocessed source-load historical data to obtain the trained VAE-LSTM coupled model specifically includes:

[0049] The preprocessed source load historical data is used to construct training samples through a time-series sliding window. The training samples are then input into the VAE-LSTM coupled model for processing to obtain the reconstructed data output by the model.

[0050] Based on the reconstructed output data and the actual input data, calculate the loss function, perform backpropagation using the Adam optimizer, and calculate and update the model parameters. , To minimize the loss function, the trained VAE-LSTM coupled model is obtained;

[0051] The loss function formula is as follows:

[0052]

[0053] in, The loss function; , These represent the network parameters of the encoder and decoder, respectively. This represents the total number of training samples; Indicates the first i The real input data for each training sample; Indicates the first iReconstructed data of training samples; The dimension of the latent space, i.e., the latent variables. The dimension; , The potential space is respectively j The mean and variance parameters of a dimension represent the statistical properties of each dimension in the latent space.

[0054] Furthermore, the trained VAE-LSTM coupled model, combined with the integrated energy system model, models the charging and discharging behavior of wind energy, solar energy, cooling / heating and power loads, and energy storage, generating high-quality scenario data with time-dependent characteristics, specifically including:

[0055] Based on the trained VAE-LSTM coupled model, 1000 sets of 64-dimensional latent variables are sampled from the latent space. z The latent variables are generated through a reparameterization module.

[0056] latent variables z The decoder of the VAE-LSTM coupled model, after input training, outputs time-series data of wind energy, solar energy, cooling / heating and power loads, and energy storage charging and discharging behavior. Based on the output time-series data, a standardized inverse processing is performed using the integrated energy system model. This includes: standardizing the time-series data generated by the VAE-LSTM decoder to restore the original data's dimensions and range; the standardized inverse processing includes performing an inverse transformation using the same standardization method as in the preprocessing stage, and adapting the inversely transformed time-series data according to the capabilities and energy demands of various devices in the integrated energy system model to ensure the data conforms to the characteristics and constraints of various devices in the system; the time-series data of wind energy, solar energy, cooling / heating and power loads, and energy storage charging and discharging behavior obtained after inverse processing are used as the generated scenario data; the scenario data includes wind energy data, solar energy data, electrical load data, heat load data, cooling load data, and energy storage charging and discharging data.

[0057] Furthermore, the process of evaluating and filtering the generated scene data using statistical and visualization indicators to obtain filtered scene data specifically includes:

[0058] Based on the generated scene data and historical source load data, the statistical and visualization indicators of the scene data are calculated, including autocorrelation coefficient, system peak-to-valley difference, Wasserstein distance (bulldozer distance) and JS divergence.

[0059] Based on calculated statistical and visualization metrics, the generated scene data is filtered to obtain filtered scene data. The filtering criteria include:

[0060]

[0061]

[0062]

[0063]

[0064] When all filtering criteria are met, retain the scene data. These are the autocorrelation coefficients of the generated scene data and the historical source load data, respectively. , The system peak-valley difference is represented by the generated scene data and the historical source load data, respectively. The Wasserstein distance represents the distribution of latent variables in the generated scene data and the distribution of latent variables in the historical source data. , The distribution of potential variables for the generated scene data and the historical source payload data, respectively; The JS divergence represents the distribution of latent variables in the generated scene data and the distribution of latent variables in the source payload historical data. , , , This is a preset threshold.

[0065] Furthermore, the autocorrelation coefficient is calculated using the following formula:

[0066]

[0067] in, The autocorrelation coefficient is... Indicates a point in time t Scene data or historical data, The mean of the data. The standard deviation of the data. For time intervals, This represents the expected value operation;

[0068] The peak-valley difference of the system is expressed by the formula:

[0069]

[0070] in, For the peak-valley difference of the system, This refers to the fluctuation range of load or power generation within the statistical period. The maximum load within the statistical period;

[0071] The Wasserstein distance is calculated using the following formula:

[0072]

[0073] in, This represents the joint distribution of the latent variable distributions of the generated scene data and the historical data. A set; This indicates taking the supremum of all joint distributions; Represents sample pairs in the distribution of latent variables; The distance order; For the dimension of the potential space;

[0074] The formula for the JS divergence is:

[0075]

[0076] in, The JS divergence represents the distribution of latent variables in the generated scene data and the distribution of latent variables in the source payload historical data. , These represent the probability distributions of the generated scene data and historical data, respectively.

[0077] Compared with the prior art, the present invention has the following advantages:

[0078] (1) In the prior art, scene generation methods generally rely on prior probability assumptions such as normal distribution to characterize the uncertainty of wind energy, solar energy and load. However, in actual operation, the output of renewable energy and the comprehensive energy load are affected by multiple factors such as meteorological conditions and user behavior. Their probability distribution often exhibits nonlinear, non-Gaussian and multi-peak characteristics, making it difficult for scenes generated based on traditional probability models to truly reflect the system's operating characteristics. This invention introduces a scene generation method based on a VAE-LSTM coupled model. It utilizes the adaptive learning capability of variational autoencoders for complex high-dimensional data distributions and combines the modeling capability of LSTM for time series dependencies. Without needing to preset a specific probability distribution form, it directly learns the inherent distribution characteristics from historical source-load data, thereby effectively solving the problem of mismatch in existing probability model assumptions and realizing an accurate characterization of the randomness and nonlinearity of wind energy, solar energy and multi-energy loads.

[0079] (2) Existing technologies mostly focus on scenario modeling of a single node or a single energy type, making it difficult to simultaneously characterize the coupling relationship and temporal correlation between electricity, heat, and cooling loads and multiple renewable energy sources. This results in the generated scenarios failing to reflect the dual attributes of source and load in an integrated energy system and the collaborative operation characteristics of multiple energy flows. This invention constructs a bidirectional integrated energy system model with energy storage and uses wind energy, solar energy, electricity load, heat load, cooling load, and energy storage charging and discharging behavior as unified multi-dimensional temporal inputs during scenario generation. Combined with the VAE-LSTM coupling model for joint modeling, it effectively characterizes the coupling relationship between multiple energy flows and the temporal correlation between source and load. This solves the technical problem that existing methods cannot reflect the overall operating characteristics of an integrated energy system and realizes the holistic and collaborative generation of integrated energy system scenarios with dual attributes of source and load.

[0080] (3) Existing scene generation methods often struggle to effectively characterize the temporal dependency between renewable energy output and load when considering data from multiple regions and time scales. The generated scenes are deficient in terms of temporal continuity and dynamic evolution characteristics, making it difficult to meet the requirements of power grid planning and dispatch for temporal consistency. This invention introduces a two-layer stacked LSTM structure in the encoder and decoder to extract and reconstruct temporal features from source-load data within typical time windows such as 24 hours. This ensures that the generated scenes maintain consistent correlation and evolutionary trends with historical data in the time dimension, effectively solving the problem of insufficient temporal modeling capabilities in existing technologies and achieving the generation of high-quality scene data with significant temporal dependencies.

[0081] (4) Existing technologies typically rely on extensive random sampling to improve accuracy during large-scale scene generation, resulting in high computational costs and low efficiency. Furthermore, subsequent scene reduction methods can easily introduce additional complexity and weaken scene diversity, making it difficult to achieve a balance between efficiency and accuracy. This invention generates scenes by sampling in the latent space using a VAE-LSTM model, enabling efficient generation of high-dimensional complex scenes with low-dimensional latent variables. Simultaneously, by combining statistical indicators such as autocorrelation coefficient, system peak-to-valley difference, Wasserstein distance, and JS divergence, the generated scenes are quantitatively evaluated and screened. This significantly reduces the number of redundant scenes while ensuring scene representativeness and diversity, thereby effectively solving the problems of low efficiency or reduction distortion in existing methods and achieving efficient and reliable scene generation and screening.

[0082] (5) The scenario data generated in existing studies often lacks effective connection with the physical constraints and equipment capabilities of actual integrated energy systems, resulting in a mismatch between the scenario and the actual system in terms of numerical range or operational feasibility, which limits its application value in power grid planning and dispatch. This invention introduces an integrated energy system model after scenario generation, performs standardized inverse processing and equipment capability constraint adaptation on the time series data output by the VAE-LSTM decoder, so that the generated wind energy, solar energy, cooling / heating and power load and energy storage charging and discharging data meet the system capacity, efficiency and operational constraint requirements, thereby solving the problem of existing scenarios being out of touch with actual systems and improving the engineering applicability and practical value of scenario data in distribution network planning and optimization dispatch. Attached Figure Description

[0083] Figure 1 This is a flowchart of the integrated energy system scenario generation method according to an embodiment of the present invention;

[0084] Figure 2 This is a logical framework diagram of the energy conversion relationship of a bidirectional integrated energy system with energy storage according to an embodiment of the present invention;

[0085] Figure 3 This is a historical data chart of cold, heat, electricity, photovoltaic, and wind turbines in an embodiment of the present invention;

[0086] Figure 4 This is a schematic diagram of the VAE-LSTM coupled model architecture according to an embodiment of the present invention;

[0087] Figure 5 This is a flowchart of the training process of the VAE-LSTM coupled model according to an embodiment of the present invention;

[0088] Figure 6 This is a schematic diagram comparing the fit of the VAE-LSTM scene before and after generation in an embodiment of the present invention;

[0089] Figure 7 This is a schematic diagram comparing the probability density of peak-valley differences in an embodiment of the present invention;

[0090] Figure 8 This is a schematic diagram of the Wasserstein distance according to an embodiment of the present invention;

[0091] Figure 9 This is a schematic diagram of JS divergence in an embodiment of the present invention;

[0092] Figure 10 This is a schematic diagram comparing peak-valley difference quantiles in an embodiment of the present invention;

[0093] Figure 11 This is a schematic diagram showing the seasonal comparison of time-of-use electricity prices according to an embodiment of the present invention;

[0094] Figure 12This is a schematic diagram of six types of scenario data after real-time scheduling in an embodiment of the present invention. Detailed Implementation

[0095] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0096] This embodiment specifically provides a method for generating integrated energy system scenarios with dual source and load attributes, including storage, for power grid planning. Figure 1 As shown, it includes the following steps:

[0097] Step S1: Construct a two-way integrated energy system model with energy storage, and determine the various parameters of the integrated energy system model according to the user's energy demand;

[0098] The established pure electric bidirectional integrated energy system with energy storage involves three energy forms: electricity, cooling, and heating. The bidirectional nature is reflected in the real-time acquisition of grid-connected power generation or grid-purchased electricity based on the power supply-demand difference. The bidirectional integrated energy system with energy storage includes heating equipment, cooling equipment, power generation equipment, and energy storage equipment. Specifically: Heating equipment consists of electric hot water boilers, which primarily use electrical energy converted into heat energy to directly heat the water inside the boiler; Cooling equipment consists of electric chiller units, which primarily use electrical energy to provide cooling; Power generation equipment consists of photovoltaic and wind turbines to supply the user's electricity load demand; Energy storage equipment consists of batteries, which play a role in energy storage and regulation within the system.

[0099] The electrothermal and electrocooling conversion coefficients used when converting user cooling and heating loads into virtual electrical loads are determined based on the different electrothermal and electrocooling equipment within the system. Therefore, this universal method is applicable to load modeling of various types of integrated energy stations on the market. In the pure electric grid-connected integrated energy system with energy storage constructed in this paper, the electrothermal conversion coefficient is determined based on the COP (Coefficient of Performance) of the electric hot water boiler's electrothermal conversion, and the electrocooling conversion coefficient is determined based on the COP of the electric chiller's electrocooling conversion. These two values ​​can be calculated from the basic parameters in the technical manual of actual integrated energy station equipment. The logical framework diagram of the energy conversion relationship of the bidirectional integrated energy system with energy storage is shown below. Figure 2 As shown.

[0100] Step S2: Preprocess the historical source-load data, which includes power load, heat load, cooling load, photovoltaic power generation data, and wind power generation data;

[0101] The historical source-load data is derived from the operational records of a real-world integrated regional energy system. This dataset includes data on wind power generation, photovoltaic power generation, electricity load, heat load, cooling load, and environmental characteristics. The historical dataset is as follows: Figure 3 As shown, the time span is one year, consisting of hourly data points, totaling 8760 time steps. To ensure the effectiveness and stability of model training, the raw data underwent preprocessing. First, data cleaning was performed to remove missing and outlier values, ensuring data integrity and accuracy. Second, a standardization method was used to scale the energy and load data to a range with a mean of 0 and a variance of 1, enabling the model to better learn the data distribution.

[0102] Step S3: A scene generation method based on the VAE-LSTM coupled model. Based on preprocessed source-load historical data and the integrated energy system model, the charging and discharging behavior of wind power, solar power, cooling / heating and power loads, and energy storage is modeled to generate high-quality scene data with time-dependent characteristics. Specifically, this includes:

[0103] Step S301: Construct the VAE-LSTM coupled model; the VAE-LSTM coupled model architecture is as follows: Figure 4 As shown, it includes an encoder, a reparameterization module, and a decoder connected in sequence;

[0104] The encoder is used to map source payload data to the latent space, extract temporal features of input data, and output the mean and variance parameters of latent variables. The encoder consists of two stacked LSTM layers and several fully connected layers. The first LSTM layer is used to extract the original temporal features, the second LSTM layer is used to generate a high-dimensional representation of the latent space, and the fully connected layers are used to map the output to the mean and variance parameters of latent variables.

[0105] The reparameterization module is used to generate latent variables based on the mean and variance of the encoder output; the reparameterization module includes a random noise sampling unit and a latent variable calculation unit;

[0106] The decoder is used to map the latent variables generated by the reparameterization module back to the data space to reconstruct the time series data of wind energy, solar energy, cooling / heating and power loads and energy storage charging and discharging behavior. The decoder includes two stacked LSTM layers and several fully connected layers. The first LSTM layer is used to receive latent variables and recover time series features, the second LSTM layer is used to gradually generate complete time series output, and the fully connected layers are used to output reconstructed data with the same dimension as the original data.

[0107] The data processing procedure for the VAE-LSTM coupled model includes:

[0108] The source and load data are input into the encoder. The input data includes wind power generation, photovoltaic power generation, cooling load, heating load, and electrical load, forming an input sequence. ,in, For the first t Hourly source load feature vector;

[0109] The encoder processes the input sequence through two stacked LSTM layers. The first LSTM layer extracts the original temporal features, and the second LSTM layer generates a high-dimensional representation of the latent space. ;

[0110] High-dimensional representation of the encoder output Mapped to the mean and log-variance of the latent variables;

[0111] The reparameterization module is based on the mean. Sum of logarithmic variance Generate latent variables ;

[0112] latent variables The input decoder consists of two stacked LSTM layers and several fully connected layers. The first LSTM layer receives latent variables and recovers temporal features, the second LSTM layer progressively generates the complete temporal output, and the fully connected layers output reconstructed data with the same dimension as the original input data. .

[0113] Step S302: Based on the preprocessed source-load historical data, train the VAE-LSTM coupled model to obtain the trained VAE-LSTM coupled model; the VAE-LSTM coupled model training process is as follows: Figure 5 As shown, it includes:

[0114] The preprocessed source load historical data is used to construct training samples through a time-series sliding window. The training samples are then input into the VAE-LSTM coupled model for processing to obtain the reconstructed data output by the model.

[0115] Based on the reconstructed output data and the actual input data, calculate the loss function, perform backpropagation using the Adam optimizer, and calculate and update the model parameters. , The trained VAE-LSTM coupled model is obtained by minimizing the loss function; the loss function formula is:

[0116]

[0117] in, The loss function; , These represent the network parameters of the encoder and decoder, respectively. This represents the total number of training samples; Indicates the first iThe real input data for each training sample; Indicates the first i Reconstructed data of training samples; The dimension of the latent space, i.e., the latent variables. The dimension; , The potential space is respectively j The mean and variance parameters of a dimension represent the statistical properties of each dimension in the latent space.

[0118] Step S303: Based on the trained VAE-LSTM coupled model and combined with the integrated energy system model, model the charging and discharging behavior of wind energy, solar energy, cooling / heating and power loads and energy storage to generate high-quality scenario data with time-dependent characteristics.

[0119] Step S4: Evaluate and filter the generated scene data using statistical and visualization indicators to obtain the filtered scene data;

[0120] First, it is necessary to verify the ability of the generated data to reproduce the distribution of historical data. The autocorrelation coefficient (ACF) is used to statistically validate the validity of the generated data. The formulas for calculating the autocorrelation coefficient of historical data and generated data are as follows:

[0121]

[0122] in, The autocorrelation coefficient is... Indicates a point in time t Scene data or historical data, The mean of the data. The standard deviation of the data. For time intervals, This represents the expected value operation;

[0123] The consistency of the VAE-LSTM coupled model before and after scene generation is as follows: Figure 6 As shown in Table 1-3 and Figure 6 When verifying the effectiveness of the generated scenarios, the data in the test set divided from historical data were real events, and the generator model of the VAE-LSTM coupled model was not trained using these data. However, some curves in the data generated by the generator were found to be highly similar to the curves in the test set. The autocorrelation coefficient was significantly improved compared to the VAE model, and the mean error was significantly reduced. This also proves that the data generated by the generator using the VAE-LSTM coupled model is very consistent with the real-world scenario and has a strong generalization ability.

[0124] Table 1 Comparison of Autocorrelation Coefficients of Cooling Load

[0125]

[0126] Table 2 Comparison of Autocorrelation Coefficients of Heat Load

[0127]

[0128] Table 3 Comparison of Autocorrelation Coefficients of Electrical Load

[0129]

[0130] In addition, it is also important to assess whether the scenarios before and after reduction are representative. Therefore, evaluation metrics such as system peak-to-valley difference and probabilistic distance (such as Wasserstein distance) can be used to assess the consistency of scenarios before and after reduction.

[0131] The peak-to-valley difference is a crucial indicator in power system operation, measuring the fluctuation range of load or generation power within a day. It reflects the difference between peak and valley demand for electricity and directly impacts the grid's peak-shaving capacity, equipment utilization, and operational economy. The formula is:

[0132]

[0133] in, For the peak-valley difference of the system, This refers to the fluctuation range of load or power generation within the statistical period. The maximum load within the statistical period;

[0134] The Wasserstein distance is a metric that measures the difference between two probability distributions, and it provides a more reasonable measure of the geometric difference between the distributions. The formula is:

[0135]

[0136] in, This represents the joint distribution of the latent variable distributions of the generated scene data and the historical data. A set; This indicates taking the supremum of all joint distributions; Represents sample pairs in the distribution of latent variables; The distance order; For the dimension of the potential space;

[0137] The JS divergence is a metric for measuring the similarity between two probability distributions. It is an improvement upon the KL divergence, and compared to the KL divergence, the JS divergence provides a more stable measure of distribution differences. The formula is:

[0138]

[0139] in, The JS divergence represents the distribution of latent variables in the generated scene data and the distribution of latent variables in the source payload historical data. , These represent the probability distributions of the generated scene data and historical data, respectively.

[0140] from Figure 7 The results show that the mean of the original data (103kW) differs from the mean of the reduced data (101kW) by only 2kW, a difference rate of approximately 1.94%. This indicates that the reduction scenario based on the VAE-LSTM coupled model preserves the overall level of the original peak-valley difference more completely than the pure VAE model. There is no distortion in the peak-valley difference characteristics between the reduced scenario and the original scenario. Figure 8 The Wasserstein distance graph shows that all indicators, such as user electricity load, wind turbines, and photovoltaic power generation, are less than 0.2, which is within an acceptable range. Furthermore, two-thirds of the Wasserstein indicators are better than those obtained using the VAE model alone. Figure 9 The overall low JS divergence value indicates that the reduced scene based on the VAE-LSTM coupled model has a high statistical similarity to the original scene and is representative. Furthermore, all three indicators listed in the JS divergence obtained based on the VAE-LSTM coupled model are lower than the JS divergence value of the pure VAE model. Figure 10 The comparison of peak-valley quantiles shows that the median value of the reduced scenario (94 kW) is 4 kW higher than that of the original scenario (98 kW), with a difference rate of approximately 4.08%, which is extremely small. This indicates that the typical scenario obtained based on the VAE-LSTM coupled model retains the median trend of the original data. The comparison of sparse distribution locations shows that the generated scenario obtained by the VAE-LSTM coupled model is more similar to the original data in terms of coverage of sparse regions compared to the pure VAE model. The fundamental reason for the good clustering effect is that the generated data itself has low noise and high structure, which indirectly verifies the effectiveness of the VAE-LSTM coupled model in scene generation.

[0141] Table 4 Comparison of Comprehensive Evaluation Indicators for Scene Generation

[0142]

[0143] comprehensive Figures 7-10As shown in Table 4, the statistical indicators (Wasserstein distance, JS divergence) and visual comparisons (peak distribution, number of locations) used to evaluate the scene generation fit indicate that the scenes reduced based on the VAE-LSTM coupling model have a high degree of consistency with the original data in the main feature dimensions. Therefore, through the above dual experimental verification, it can be concluded that compared with the VAE model, the VAE-LSTM coupling model generates scene sequences that are closer to the time series patterns of historical data in the generation of cold, heat, electricity, load data and renewable energy data of photovoltaics and wind turbines. The fit between historical data and generated data is higher, the generated data has a stronger ability to restore the distribution of historical data, and is more representative.

[0144] Step S5: Based on the selected scenario data and combined with the real-time grid dispatch decision model, perform real-time calculations of grid-connected power generation, grid-purchased electricity, and energy storage charging and discharging to achieve optimized dispatch and management of the energy system. Specifically, this includes:

[0145] In the real-time dispatch decision-making framework of a two-way integrated energy system with energy storage, dispatch decisions are based on multi-objective optimization and a dynamic electricity price response mechanism. The time-of-use pricing mechanism is as follows: Figure 11As shown. The decision-making process first generates typical probabilistic scenarios of renewable energy output and load demand using a VAE-LSTM coupled model. Then, a hierarchical decision-making logic is adopted: on the time scale, the system dynamically assesses the net power balance (total renewable energy generation minus electricity / heat / cooling load demand) at one-hour intervals, combined with the real-time state of charge (SOC) constraint of the battery storage system; on the spatial scale, a differentiated time-of-use pricing mechanism is first established based on seasonal divisions (summer, winter, autumn, and other seasons), dividing the entire day into four price ranges: peak, mid-peak, normal, and off-peak, and formulating energy storage charging and discharging strategies for different ranges. For example, during off-peak periods, when net energy > 0, renewable energy is used to charge the remaining energy storage; any shortfall is supplemented by purchasing electricity from the grid at a low price, and the remaining energy is sold; when net energy < 0, additional electricity is purchased from the grid to cover the current demand gap, forcibly supplementing the current hourly energy storage charging power, and no renewable energy is available for sale. During peak electricity price periods, when net energy > 0, all renewable energy is sold, and additional energy storage power is released and sold at a high price. When net energy < 0, energy storage is prioritized to fill the gap; electricity is purchased only when the gap exceeds the storage capacity. During normal electricity price periods, when net energy > 0, energy storage is prioritized for charging, and the remaining electricity is sold to the grid. When net energy < 0, energy storage is prioritized for discharging to fill the gap; electricity is purchased only when the gap exceeds the storage capacity. During peak electricity price periods, when net energy > 0, all renewable energy is sold; when net energy < 0, energy storage is released to fill the gap; electricity is purchased only when the gap exceeds the storage capacity. Ultimately, the optimal power allocation under multiple constraints is achieved through the `_energy_allocation` method, whose output includes grid interaction power and energy storage charging / discharging plans. Finally, the energy allocation logic is invoked in real-time, utilizing an improved decision model to calculate information such as grid-connected generation, grid-purchased electricity, and energy storage charging / discharging amounts. The scenario data generation categories are shown in Table 5. Figure 12 This represents six typical scenario data obtained from the generation and reduction of bidirectional integrated energy system scenarios based on the VAE-LSTM coupled model, followed by real-time scheduling decisions. It includes the grid-connected power generation and grid-purchased electricity under each scenario after real-time scheduling decisions. The formula is:

[0146]

[0147]

[0148] In the formula: jia represents the four types of electricity prices: peak, valley, flat, and peak; P 净 P represents the size of renewable energy generation and load. storage P represents energy storage capacity. ele P represents the power generation from photovoltaic and wind turbines within the integrated energy station. load P represents user load. charge P represents battery charging. dischargeP represents battery discharge. buy P represents the amount of electricity purchased from the power grid. sell Represents grid-connected power generation;

[0149] Table 5. Scene Data Generation Categories

[0150]

[0151] Integrated energy systems have attracted widespread attention as an efficient and flexible energy management model. By integrating various forms of energy such as electricity, heat, and cooling, they achieve optimized energy allocation and efficient utilization, significantly improving the overall efficiency and reliability of the energy system. However, with the increasing proportion of renewable energy sources such as wind and solar power in bidirectional integrated energy systems, the uncertainty of input and output at a certain node of the distribution network poses new challenges to grid stability and economic dispatch. The flexibility and variability of bidirectional energy stations make them an uncertain factor in the distribution network, having a greater impact on grid operation and load distribution than other relatively stable parts. Therefore, more effective scenario generation is needed for such bidirectional integrated energy systems. To this end, the VAE-LSTM coupled model proposed in this invention improves the robustness and adaptability of the model while generating high-quality and diverse scenarios for integrated energy stations. Experimental results show that using the VAE-LSTM coupled model for scenario generation of bidirectional integrated energy systems is more practically significant.

[0152] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0153] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for generating scenarios of integrated energy systems with dual source and load attributes, including energy storage, oriented towards power grid planning, characterized in that, Includes the following steps: Construct a two-way integrated energy system model with energy storage, and determine the various parameters of the integrated energy system model based on the user's energy demand; The historical source-load data is preprocessed, including power load, heat load, cooling load, photovoltaic power generation data, and wind power generation data. The scene generation method based on the VAE-LSTM coupled model models the charging and discharging behavior of wind power, solar power, cooling / heating and power loads and energy storage based on preprocessed source-load historical data and integrated energy system model, generating high-quality scene data with time-dependent characteristics. The generated scene data is evaluated and filtered using statistical and visualization indicators to obtain the filtered scene data. Based on the selected scenario data and combined with the real-time grid dispatch decision model, the grid-connected power generation, grid power purchase, and energy storage charging and discharging are calculated in real time to achieve optimized dispatch and management of the energy system. The scenario generation method based on the VAE-LSTM coupled model models the charging and discharging behavior of wind power, solar power, cooling / heating and power loads, and energy storage based on preprocessed source-load historical data and a comprehensive energy system model, generating high-quality scenario data with time-dependent characteristics, specifically including: Construct a VAE-LSTM coupled model; The VAE-LSTM coupled model is trained based on the preprocessed source load historical data to obtain the trained VAE-LSTM coupled model. Based on the trained VAE-LSTM coupled model and combined with the integrated energy system model, the charging and discharging behavior of wind energy, solar energy, cooling / heating and power loads and energy storage is modeled to generate high-quality scenario data with time-dependent characteristics. The VAE-LSTM coupled model includes an encoder, a reparameterization module, and a decoder connected in sequence. The encoder is used to map source payload data to the latent space, extract temporal features of input data, and output the mean and variance parameters of latent variables. The encoder includes two stacked LSTM layers and several fully connected layers, wherein the first LSTM layer is used to extract the original temporal features, the second LSTM layer is used to generate a high-dimensional representation of the latent space, and the fully connected layers are used to map the output to the mean and variance parameters of latent variables. The reparameterization module is used to generate latent variables based on the mean and variance of the encoder output; the reparameterization module includes a random noise sampling unit and a latent variable calculation unit; The decoder is used to map the latent variables generated by the reparameterization module back to the data space to reconstruct the time series data of wind energy, solar energy, cooling / heating and power loads and energy storage charging and discharging behavior. The decoder includes two stacked LSTM layers and several fully connected layers. The first LSTM layer is used to receive latent variables and recover time series features, the second LSTM layer is used to gradually generate complete time series output, and the fully connected layers are used to output reconstructed data with the same dimension as the original data.

2. The method for generating a scenario of a power grid planning-oriented integrated energy system with dual source and load attributes, including energy storage, as described in claim 1, is characterized in that... The construction of a two-way integrated energy system model with energy storage, and the determination of various parameters of the integrated energy system model based on the user's energy demand, specifically includes: Based on the user's electricity, heat, and cooling needs, the capacity, efficiency, and operating mode of each device in the integrated energy system model are set; based on the characteristics of wind power, solar power, and energy storage devices, the power generation capacity, energy storage capacity, and charge / discharge efficiency of each device are defined; based on the grid's dispatch requirements, the interaction mode between the grid and the integrated energy system is determined, including the grid's load response, power generation dispatch strategy, and the energy storage system's charge / discharge strategy; by establishing the interrelationships and constraints between devices, a two-way integrated energy system model with energy storage is formed, which includes energy supply devices, energy coupling devices, and energy storage devices.

3. The method for generating a scenario of a power grid planning-oriented integrated energy system with dual source and load attributes, including energy storage, as described in claim 1, is characterized in that... The preprocessing of historical source load data specifically includes: Perform data cleaning on the source load historical data, including removing missing values ​​and outliers; The cleaned source-load historical data is standardized, including scaling the power load, heat load, cold load, photovoltaic power generation data, and wind power generation data to a range with a mean of 0 and a variance of 1; thus obtaining the pre-processed source-load historical data.

4. The method for generating a scenario of a power grid planning-oriented integrated energy system with dual source and load attributes, including energy storage, as described in claim 1, is characterized in that... The data processing procedure of the VAE-LSTM coupled model includes: The source and load data are input into the encoder, including wind power generation, photovoltaic power generation, cooling load, heating load, and electrical load, forming an input sequence. ,in, For the first t Hourly source load feature vector; The encoder processes the input sequence through two stacked LSTM layers. The first LSTM layer extracts the original temporal features, and the second LSTM layer generates a high-dimensional representation of the latent space. , represented as: in, This indicates the hidden state of the previous time step. Enter the current time step; , , These are the outputs of the forget gate, input gate, and output gate, respectively. Candidate memory states; This is the unit state; This indicates element-wise multiplication; These are the forget gate weight matrix and the bias vector, respectively; , These are the input gate weight matrix and the bias vector, respectively; , These are the output gate weight matrix and the bias vector, respectively; , These are the candidate memory unit weight matrix and bias vector, respectively; Output sequence for encoder; For activation functions; High-dimensional representation of the encoder output The formula for mapping to the mean and log-variance of latent variables is: in, These are the mean and log-variance mapped to the latent variables, respectively; , , Here are the weight matrix and bias vector of the fully connected layer; The reparameterization module is based on the mean. Sum of logarithmic variance Generate latent variables The formula is: in, Indicates variance; To obtain from the standard normal distribution Random noise in the mid-sample, It is the identity matrix; latent variables The decoder is input to a given source, which consists of two stacked LSTM layers and several fully connected layers. The first LSTM layer receives latent variables and recovers temporal features, the second LSTM layer progressively generates a complete temporal output, and the fully connected layers output reconstructed data with the same dimension as the original input data. The formula is: in, These are the network parameters for the decoder; This represents the function mapping of the decoder.

5. The method for generating a scenario of a power grid planning-oriented integrated energy system with dual source and load attributes, including energy storage, as described in claim 1, is characterized in that... The step of training the VAE-LSTM coupled model based on the preprocessed source-load historical data to obtain the trained VAE-LSTM coupled model specifically includes: The preprocessed source load historical data is used to construct training samples through a time-series sliding window. The training samples are then input into the VAE-LSTM coupled model for processing to obtain the reconstructed data output by the model. Based on the reconstructed output data and the actual input data, calculate the loss function, perform backpropagation using the Adam optimizer, and calculate and update the model parameters. , To minimize the loss function, the trained VAE-LSTM coupled model is obtained; The loss function formula is as follows: in, The loss function; , These represent the network parameters of the encoder and decoder, respectively. This represents the total number of training samples; Indicates the first i The real input data for each training sample; Indicates the first i Reconstructed data of training samples; The dimension of the latent space, i.e., the latent variables. The dimension; , The potential space is respectively j The mean and variance parameters of a dimension represent the statistical properties of each dimension in the latent space.

6. The method for generating a scenario of a power grid planning-oriented integrated energy system with dual source and load attributes, including energy storage, as described in claim 1, is characterized in that... The trained VAE-LSTM coupled model, combined with the integrated energy system model, models the charging and discharging behavior of wind power, solar power, cooling / heating and power loads, and energy storage, generating high-quality, time-dependent scenario data, specifically including: Based on the trained VAE-LSTM coupled model, 1000 sets of 64-dimensional latent variables are sampled from the latent space. z The latent variables are generated through a reparameterization module. latent variables z The decoder of the trained VAE-LSTM coupled model outputs time-series data on wind energy, solar energy, cooling / heating and power loads, and energy storage charging and discharging behavior. Based on the output time-series data, the standardized inverse processing through the integrated energy system model includes: standardizing the time-series data generated by the VAE-LSTM decoder to restore the original data's dimensions and range; the standardized inverse processing includes performing an inverse transformation using the same standardization method as in the preprocessing stage, and adapting the inversely transformed time-series data according to the capabilities and energy requirements of various devices in the integrated energy system model to ensure that the data conforms to the characteristics and constraints of various devices in the system; After inverse processing, the time-series data of wind energy, solar energy, cooling / heating and power loads, and energy storage charging and discharging behavior are obtained and used as the generated scene data; the scene data includes wind energy data, solar energy data, power load data, heat load data, cooling load data, and energy storage charging and discharging data.

7. The method for generating a scenario of a power grid planning-oriented integrated energy system with dual source and load attributes and storage, as described in claim 1, is characterized in that... The process of evaluating and filtering the generated scene data using statistical and visualization indicators to obtain filtered scene data specifically includes: Based on the generated scene data and historical source load data, calculate the statistical and visualization indicators of the scene data, including autocorrelation coefficient, system peak-to-valley difference, Wasserstein distance and JS divergence; Based on calculated statistical and visualization metrics, the generated scene data is filtered to obtain filtered scene data. The filtering criteria include: When all filtering criteria are met, retain the scene data. These are the autocorrelation coefficients of the generated scene data and the historical source load data, respectively. , The system peak-valley difference is represented by the generated scene data and the historical source load data, respectively. The Wasserstein distance represents the distribution of latent variables in the generated scene data and the distribution of latent variables in the historical source data. , The distribution of potential variables for the generated scene data and the historical source payload data, respectively; The JS divergence represents the distribution of latent variables in the generated scene data and the distribution of latent variables in the source payload historical data. , , , This is a preset threshold.

8. The method for generating a scenario of a power grid planning-oriented integrated energy system with dual source and load attributes, including energy storage, as described in claim 7, is characterized in that... The autocorrelation coefficient is formulated as follows: in, The autocorrelation coefficient is... Indicates a point in time t Scene data or historical data, The mean of the data. The standard deviation of the data. For time intervals, This represents the expected value operation; The peak-valley difference of the system is expressed by the formula: in, For the peak-valley difference of the system, This refers to the fluctuation range of load or power generation within the statistical period. The maximum load within the statistical period; The Wasserstein distance is calculated using the following formula: in, This represents the joint distribution of the latent variable distributions of the generated scene data and the historical data. A set; This indicates taking the supremum of all joint distributions; Represents sample pairs in the distribution of latent variables; The distance order; For the dimension of the potential space; The formula for the JS divergence is: in, The JS divergence represents the distribution of latent variables in the generated scene data and the distribution of latent variables in the source payload historical data. , These represent the probability distributions of the generated scene data and historical data, respectively.

Images