Virtual power plant dynamic control method and system based on Times MLP

Abstract

The invention discloses a TimeMLP-based virtual power plant dynamic control method and system. The method comprises the steps of establishing a database of typical distributed flexible resources in a virtual power plant; performing classification adjustment characteristic analysis on the typical distributed flexible resources; inputting the data in the database into a time sequence analysis network Times MLP prediction model to obtain a positive power supply output prediction value; and based on the power market data and the positive power supply predicted value, calculating the comprehensive adjustment capability, analyzing the response effect of the virtual power plant, and dynamically adjusting the control strategy of the virtual power plant. According to the method, dynamic adjustment and intelligent scheduling can be realized in a complex power market environment, so that the operation efficiency and the response capability of the virtual power plant are optimized.

Description

Technical Field

[0001] This invention belongs to the field of smart grid technology and relates to a method and system for dynamic control of virtual power plants based on TimesMLP. Background Technology

[0002] Against the backdrop of building a new power system dominated by new energy sources, the drive for clean and efficient energy use is propelling distributed resources towards large-scale, highly aggregated, and deeply interactive development. Compared to microgrids and active distribution networks, virtual power plants (VPPs) can more effectively address the scheduling and control difficulties arising from differences in geographical location, operating characteristics, and response speed among various types of distributed resources, providing high-quality and flexible regulation resources for the power system. Several practical engineering cases of VPPs have been developed both domestically and internationally, primarily focusing on power market operation and peak-shaving ancillary services using VPPs.

[0003] Existing virtual power plant technologies have significant shortcomings in the following aspects: Insufficient dynamic response capability: Existing methods have weak real-time dynamic response capabilities for distributed resources and cannot efficiently cope with grid fluctuations. Lack of flexibility in regulating characteristic modeling: Most current research focuses only on a single resource type and fails to fully consider the heterogeneity of various distributed resources. Limited expressive power of time series analysis: Traditional methods are insufficient in capturing complex time series changes, especially performing poorly in periodic and nonlinear scenarios.

[0004] Current research has certain limitations in the application of virtual power plants (VPPs) in frequency regulation. The feasible region or response capability of VPP calculated in convergence theory studies only focuses on boundary values ​​under constraints, neglecting intermediate states and actual operating conditions, and cannot directly reflect the regulation methods of each resource. The control methods proposed in response strategy studies lack consideration of DER (distributed energy resources) types and cannot effectively monitor the changes in the operating state of the VPP during dynamic processes. DERs are heterogeneous, mainly because they can be any form of source, load, or storage, or a combination thereof. Due to the differences in the operating characteristics of each resource, the overall regulation characteristics of the VPP present a complex curve with multiple segments, and the control methods and response effect evaluation methods for different segments may vary.

[0005] Meanwhile, current research largely focuses on operating new energy sources like solar and wind power at their maximum power points, with little consideration given to utilizing energy storage and other virtual power plant equipment to achieve stable output in conjunction with solar and wind power. Furthermore, the prediction algorithms for solar and wind power output are not highly complex. Current mainstream prediction models for wind and solar power output mainly consist of LSTM models and combinations of various Transformer models and decomposition algorithms, lacking an integrated and efficient prediction algorithm. This results in low prediction accuracy, impacting the intraday participation of virtual power plants in peak-shaving ancillary services. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a dynamic control method for virtual power plants based on TimesMLP.

[0007] The present invention adopts the following technical solution.

[0008] The first aspect of this invention proposes a dynamic control method for a virtual power plant based on TimesMLP, comprising:

[0009] Step 1: Establish a database of typical distributed flexible resources within the virtual power plant, including historical output data, meteorological data, and electricity market data of typical distributed flexible resources;

[0010] Step 2: Analyze the regulation characteristics of typical distributed flexible resources in the virtual power plant and classify them into positive power sources, negative power sources, and energy storage power sources.

[0011] Step 3: Input the historical power output data and meteorological data of the positive source into the Time Series MLP prediction model to obtain the predicted power output value of the positive source;

[0012] Step 4: Calculate the comprehensive regulation capability of the positive power source, negative power source and energy storage power source in the virtual power plant based on the predicted output value of the positive power source and the power market data. Analyze the response effect of the virtual power plant based on the comprehensive regulation capability of the virtual power plant in order to dynamically adjust the virtual power plant control strategy.

[0013] Preferably, the typical distributed flexible resources include photovoltaic, wind power, energy storage, and load; the meteorological data includes wind direction, wind speed, light intensity, and temperature; and the electricity market data includes electricity price.

[0014] Preferably, in step 2, the adjustment characteristics of typical distributed flexible resources are analyzed, and the resources are divided into three categories according to their response to frequency adjustment:

[0015] (1) Positive power supply: When the system frequency decreases, the positive power supply increases its output proportionally; when the system frequency increases, the positive power supply decreases its output proportionally.

[0016] (2) Negative power supply: For a continuously adjustable negative power supply, when the system frequency increases, the negative power supply increases power consumption proportionally, and when the system frequency decreases, the negative power supply decreases power consumption proportionally; For a stepped adjustable negative power supply, the segmented step function corresponding to the stepped adjustment is expressed as a continuous linear droop characteristic.

[0017] (3) Energy storage power supply: It does not operate within the frequency regulation dead zone, and stops frequency regulation when the SOC exceeds the limit.

[0018] Preferably, in step 3, the historical power output data and meteorological data of the positive power source are input into the TimesMLP prediction model to obtain the predicted power output value of the positive power source, including:

[0019] Step 3.1: Using the multidimensional time data composed of historical power output data and meteorological data of the positive power source as the original time series data, perform fast Fourier transform analysis on the original time series data to extract two-dimensional tensors;

[0020] Step 3.2: Based on the parameter-sharing Inception block, the TimesBlock module projects and reshapes the original time series data in two dimensions. Combined with the two-dimensional tensor from Step 3.1, multi-period two-dimensional convolutional features are obtained. The fused time series feature vector is obtained through amplitude-weighted feature fusion.

[0021] Step 3.3: The Multilayer Perceptron (MLP) performs output prediction on the fused temporal feature vector, and outputs the predicted output value after optimization by the loss function.

[0022] Preferably, step 3.1 specifically includes:

[0023] In the TimesMLP forecasting model, the historical power output data and meteorological data of the input positive power source constitute multidimensional time data as the raw time series data. Each dimension of the raw time series data X is then analyzed. lD Perform a fast Fourier transform according to equation (6) and filter the top k significant frequencies in the frequency domain, then calculate the corresponding period length for each significant frequency.

[0024] One-dimensional time data for each period length is reconstructed into a two-dimensional tensor, which includes variations during the period and variations within the period.

[0025]

[0026] Where FFT(·) and Amp(·) represent the calculation of FFT and amplitude values, respectively; Avg(·) is the averaging function. A Represents the calculated amplitude at each frequency; {f1,…,f k} represents the top k significant frequencies in the selected frequency domain, {p1,…,p k} represents the length of the k periods corresponding to the first k significant frequencies in the frequency domain.

[0027] Preferably, step 3.2 specifically includes:

[0028] 1) The TimesBlock module projects the original time series data into a d-dimensional space through an embedding layer to obtain d-dimensional embedding features;

[0029] 2) Reshape the d-dimensional embedding features into a two-dimensional tensor according to the period length obtained in step 3.1, apply the parameter-sharing Inception block to it, and align it with the two-dimensional tensor in step 3.1 period by period to obtain multi-period two-dimensional convolutional features;

[0030] 3) Based on the frequency domain amplitude weighted fusion of multi-period two-dimensional convolution features, output the d-dimensional fused temporal feature vector.

[0031] Preferably, in step 4, the formula for calculating the comprehensive regulation capability of the positive power source, negative power source, and energy storage power source within the virtual power plant is:

[0032] P VPP_adjustable (t)=P adjustable,gen (t)+P adjustable,dem (t)+P adjustable,storage (t)

[0033] P adjustable,gen (t)=α·P gen (t)

[0034] P adjustable,dem (t)=i∑β i ·△P price(t)

[0035] P adjustable,storage (t)=min(P discharge_max ,max(-P charge_max ,P storage(t) ))

[0036] Among them, P adjustable,gen (t) represents the regulation capability of the positive power supply, P adjustable,dem (t) represents the regulation capability of the negative power supply, P adjustable,storage (t) represents the regulation capability of the energy storage power source;

[0037] α is the adjustment coefficient, P gen (t) represents the predicted output of the positive power source at time t;

[0038] β i It is the demand response coefficient of user i, ΔP price(t) The demand adjustment caused by the price change at time t;

[0039] P charge_max and P discharge_max P represents the maximum power for charging and discharging the energy storage power supply, respectively. storage(t) Let t be the net power output of the stored energy.

[0040] Preferably, in step 4, the response effect of the virtual power plant is analyzed based on its comprehensive regulation capability to dynamically adjust the virtual power plant control strategy, including:

[0041] If the overall regulation capability of the virtual power plant is lower than the set value, the response effect of the virtual power plant is considered to be below expectations, and the virtual power plant control strategy will be dynamically adjusted in the following way:

[0042] Optimize the synergy between photovoltaics, wind power, energy storage, and load through a multi-level dynamic interaction model;

[0043] Resources and information are shared through a cross-virtual power plant collaboration mechanism;

[0044] Deep reinforcement learning is used to dynamically adjust energy storage and renewable energy output in order to optimize market response to dynamic changes and emergencies in the electricity market.

[0045] A second aspect of this invention proposes a dynamic control system for a virtual power plant based on TimesMLP, comprising:

[0046] The database construction module is used to build a database of typical distributed flexible resources within the virtual power plant, including historical output data, meteorological data, and electricity market data of typical distributed flexible resources.

[0047] The resource classification and analysis module is used to analyze the regulation characteristics of typical distributed flexible resources in the virtual power plant and classify them into positive power sources, negative power sources and energy storage power sources.

[0048] The power output prediction module is used to input the historical power output data and meteorological data of the positive power source into the Time Series MLP prediction model to obtain the predicted power output value of the positive power source.

[0049] The dynamic control module is used to calculate the comprehensive regulation capability of the positive power source, negative power source and energy storage power source in the virtual power plant based on the predicted output value of the positive power source and the power market data. Based on the comprehensive regulation capability of the virtual power plant, the response effect of the virtual power plant is analyzed so as to dynamically adjust the control strategy of the virtual power plant.

[0050] A third aspect of the present invention provides a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to perform operations according to the instructions to execute the steps of the method.

[0051] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method.

[0052] The beneficial effects of this invention are compared with those of the prior art:

[0053] This invention, for the first time, proposes a joint design of TimesBlock and Multilayer Perceptron (MLP) to construct a Time Series Analysis Network (TimeMLP) prediction model. Utilizing the TimesNet architecture, time series data is transformed from one-dimensional to two-dimensional, effectively improving the ability to capture periodic and cross-period changes. By combining Fast Fourier Transform (FFT), significant periodic features are extracted and efficiently modeled in two-dimensional space, thereby enhancing the analysis and prediction capabilities of complex time series patterns. This improves prediction accuracy while reducing model size for real-time prediction. It addresses the problem that existing virtual power plant regulation feasible regions or response capabilities only focus on boundary values ​​under constraints, neglecting intermediate states and actual operating conditions. Furthermore, existing response strategy research lacks consideration of distributed energy resources (DER) types in control methods, failing to effectively monitor and predict changes in the operating state of virtual power plants during dynamic processes. Current mainstream output prediction models lack both prediction accuracy and speed. The TimesMLP prediction model of this invention, through joint training of TimesNet and MLP, can accurately predict multi-dimensional feature data, effectively improving the model's adaptability to complex input data and thus enhancing prediction accuracy.

[0054] This invention unifies the modeling of the heterogeneous characteristics of various distributed resources and constructs a static integrated regulation characteristic model to calculate the integrated regulation capability of a virtual power plant, overcoming the limitations of existing methods in adapting to the regulation of heterogeneous resources. Based on this model, this invention calculates the integrated regulation capability of positive power sources, negative power sources, and energy storage power sources within the virtual power plant. It analyzes the response effect of the virtual power plant based on its integrated regulation capability, dynamically adjusting the virtual power plant control strategy to achieve dynamic adjustment of the virtual power plant's operation strategy, significantly improving the virtual power plant's rapid response capability to grid demand. This invention effectively solves the problems of numerous resource types, large quantities, and diverse regulation characteristics encountered in the dynamic assessment and prediction of operating power of virtual power plants with massive distributed flexible resources. By introducing deep reinforcement learning, more efficient resource coordination and optimized scheduling can be achieved in real-time dynamic regulation, thereby enhancing the dynamic response capability and market adaptability of the virtual power plant. This addresses the limitations of existing technologies where the dynamic response capability of virtual power plants is restricted by traditional static models, failing to fully capture complex time series characteristics and lacking real-time collaborative optimization mechanisms. Furthermore, existing methods cannot effectively balance economy and stability when facing scenarios such as power market fluctuations and sudden changes in the state of distributed resources.

[0055] This invention, based on the Time Series Analysis Network (TimesMLP), introduces deep reinforcement learning. By fusing TimesMLP and deep reinforcement learning, the control strategy of a virtual power plant is dynamically adjusted. TimesMLP can accurately capture complex time series variation characteristics, especially excelling in periodic and nonlinear changes, while deep reinforcement learning (DRL) enables real-time scheduling optimization, effectively balancing the scheduling of various distributed resources in the virtual power plant, improving system responsiveness and peak-shaving efficiency. By introducing reinforcement learning algorithms and multi-dimensional time series analysis, the dynamic operating status of the virtual power plant can be more accurately reflected. It can achieve dynamic adjustment and intelligent, efficient scheduling optimization in complex power market environments, thereby optimizing the operating efficiency and responsiveness of the virtual power plant. It is suitable for the collaborative optimization and real-time response of massive distributed flexible resources, making up for the shortcomings of existing virtual power plant technologies, such as a single feasible adjustment domain, limited response capability, and insufficient dynamic state monitoring. It can overcome the dynamic response problem in existing methods and significantly improve the operating accuracy and scheduling flexibility of virtual power plants.

[0056] This invention not only enables dynamic evaluation and optimization of the response performance of virtual power plants in peak-shaving ancillary services, but also effectively improves the market adaptability and scheduling accuracy of virtual power plants. Attached Figure Description

[0057] Figure 1 This is a flowchart of the method of the present invention;

[0058] Figure 2 This is a flowchart illustrating the implementation of the method of the present invention;

[0059] Figure 3 The TimesMLP model structure is as follows;

[0060] Figure 4 For forecasting the output of photovoltaic and wind power. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, other embodiments obtained by those skilled in the art without creative effort are all within the protection scope of this invention.

[0062] Embodiment 1 of this invention provides a dynamic control method for a virtual power plant based on TimesMLP, such as... Figure 1-2 As shown, it includes:

[0063] Step 1: Establish a database of typical distributed flexible resources within the virtual power plant, including:

[0064] Considering the diversity and differences of resources, historical output data of typical distributed flexible resources (photovoltaics, wind power, energy storage, and load) in the virtual power plant, as well as meteorological data (wind direction, wind speed, light intensity, temperature, etc.) and electricity market data (such as electricity price and dynamic changes in the electricity spot market environment) are collected to form a multi-dimensional dataset and establish a database of typical distributed flexible resources (DER), providing a data foundation for subsequent prediction and optimization.

[0065] Furthermore, based on the different regulation characteristics of different resource types (wind / solar / storage / load) under real-time electricity spot data, the typical distributed flexible resources in the multi-dimensional dataset of step 1.1 can be clustered to obtain a database of typical distributed flexible resources in the virtual power plant. This database can be used for real-time scheduling and optimization. Different resources can be classified and managed according to market price signals, so as to realize real-time scheduling and optimized allocation of resources in the dynamic electricity market.

[0066] Step 2: Classify and analyze the adjustment characteristics of typical distributed flexible resources;

[0067] More preferably, based on the collected historical characteristic data of DER, the differences in the operating characteristics of different types of distributed resources in the VPP are considered, especially their differences in adjustable power range, adjustment continuity, and operating constraints. According to the response method of each resource to frequency regulation (the way frequency regulation is completed), they are divided into three categories: positive power source, negative power source, and energy storage power source, as follows:

[0068] (1) Positive power supply: When the system frequency decreases, the positive power supply increases its output proportionally; when the system frequency increases, the positive power supply decreases its output proportionally.

[0069] Positive power supply regulation characteristics: Positive power supply refers to the resource that generates power externally. Its frequency regulation method is the same as that of conventional generator sets. When the system frequency decreases, its output power is increased, and when the system frequency increases, its output power is decreased.

[0070] In VPP, positive power sources include various distributed power sources, such as CHP units, small hydropower, biomass power generation, photovoltaic, wind power, and other new energy sources.

[0071] Positive power sources are typically continuously adjustable resources, and their droop regulation characteristics can be expressed as a mathematical model of equation (1). represents the operating power of various power sources (output is positive, absorption is negative).

[0072]

[0073] In the formula: P ps P0 is the expected output power of the positive power supply at the actual system frequency f; P0 is the output power of the positive power supply at the rated frequency f0; kps P is the frequency regulation coefficient of the positive power supply (unit: MW / Hz); max This is the maximum output power of the positive power supply.

[0074] (2) Negative power supply: For a continuously adjustable negative power supply, when the system frequency increases, the negative power supply increases power consumption proportionally, and when the system frequency decreases, the negative power supply decreases power consumption proportionally; For a stepped adjustable negative power supply, the segmented step function corresponding to the stepped adjustment is expressed as a continuous linear droop characteristic.

[0075] Negative power supply regulation characteristics: Controllable loads, also known as flexible loads, absorb power from the system, but because they can be controlled and adjusted within a certain range, they can be considered as negative power sources. The frequency regulation method of negative power supplies is similar to demand response; when the system frequency decreases, their power consumption decreases, and when the system frequency increases, their power consumption increases. Based on whether the regulation is continuous, they can be further divided into continuously adjustable negative power supplies and ranged adjustable negative power supplies.

[0076] 1) Continuously adjustable negative power supply

[0077] Continuously adjustable loads mainly refer to temperature-controlled loads, including central air conditioning and electric water heating systems. Since the temperature can change continuously within a certain range, the power consumption of a temperature-controlled load is a continuous function, and its droop regulation characteristic can be expressed as the mathematical model of equation (2):

[0078]

[0079] In the formula: P nc k represents the expected output power of the continuously adjustable negative power supply at the actual system frequency f. nc P is the frequency adjustment coefficient for the negative power supply. min This is the lower limit of the output power.

[0080] 2) Adjustable negative power supply

[0081] Some industrial loads, lighting loads, etc., can have a portion of their load cut off under certain circumstances, and are called interruptible loads. However, these types of loads can only be switched between several levels, and the adjustment is not continuous. They belong to the level-adjustable negative power supply, and their droop adjustment characteristics can be expressed as the mathematical model of equation (3).

[0082]

[0083] In the formula: P ng The expected output power of the stepped adjustable negative power supply at the actual system frequency f; P1, P2, ..., P n These represent the output power of its 1st, 2nd, ..., nth gears; f1, f2, ..., f n-1Frequency boundaries for different load levels; f min and f max These represent the upper and lower limits of the system's frequency, respectively.

[0084] As can be seen, the droop adjustment characteristic of the adjustable negative power supply exhibits a piecewise step function. Similarly, connecting the midpoints of adjacent segments in the piecewise function yields the transition segment between two frequency ranges. The reciprocal of the slope of this segment is defined as its frequency adjustment coefficient. For ease of adjustment, the frequency adjustment coefficients for all the above intervals should be the same, that is, the slopes of all transition segments should be the same. Therefore, the line connecting the midpoints of all segments in the piecewise function should be a straight line. The output power at the rated frequency is the r-th level (0≤r≤n, f r-1 <f0<f r Then, the droop adjustment characteristic in equation (3) can be rewritten as follows:

[0085]

[0086] (3) Energy storage power supply: Set a frequency regulation dead zone to avoid frequent charging and discharging. It does not operate within the frequency regulation dead zone and stops frequency regulation when the SOC exceeds the limit;

[0087] Energy storage power supply regulation characteristics: Energy storage can also be viewed as a power source used for frequency regulation. It releases stored energy when the system frequency decreases and absorbs energy from the system when the system frequency increases. In virtual power plants, energy storage power supplies mainly include various types such as batteries and pumped hydro storage. To prevent the lifespan of energy storage devices from decreasing due to frequent output adjustments caused by system frequency fluctuations, a frequency dead zone needs to be set. Furthermore, energy storage power supplies also have special energy constraints, namely, the state of charge (SOC) needs to be maintained within a certain range.

[0088] Therefore, the droop regulation characteristics of energy storage power sources can be expressed as a mathematical model of equation (5).

[0089]

[0090] In the formula: P ss k represents the expected output power of the energy storage power source at the actual system frequency f. ss Its frequency adjustment coefficient; f ud and f dd These are the upper and lower limits for adjusting the dead zone, respectively; k SOC The state of charge of energy storage; k SOC,min and k SOC,max These represent the upper and lower limits of the state of charge of energy storage, respectively.

[0091] Step 3: Input the historical power output data and meteorological data of the positive source into the Time Series MLP prediction model to obtain the predicted power output value of the positive source;

[0092] More preferably, according to Figure 3 , Figure 4 A power prediction model for positive power sources such as distributed photovoltaic (PV) and wind power is established. The Time Series Analysis Network (TimeNet) is incorporated, which effectively represents intra-cycle and inter-weekly variations, addressing the modeling problem of complex time-varying power. By combining TimeNet with a Multilayer Perceptron (MLP), a multi-dimensional characteristic power output prediction model is constructed, utilizing historical characteristic data of distributed PV and wind power to predict short-term power output. The model will optimize PV and wind power output in real time based on a Maximum Power Point Tracking (MPPT) strategy, thereby improving the virtual power plant's adaptability to market fluctuations and scheduling accuracy.

[0093] In practical implementation, based on the multi-periodicity of time series, a modular time network structure is proposed to capture time patterns in different periods. By introducing Fast Fourier Transform (FFT) to perform frequency domain analysis on the time series data, the most significant periodic features are extracted, and a TimesNet backbone prediction neural network is constructed. Based on this, a bidirectional network structure is designed to better capture the long-term dependencies and periodic changes in the time series data.

[0094] The TimesMLP model comprises a Timesblock module, a prediction module, and a multi-scale, multi-dimensional feature interaction module. The Timesblock module includes a two-dimensional tensor decomposition module based on TimesNet and adaptive aggregation of feature tensors. The prediction module primarily predicts multi-scale, multi-dimensional features. The multi-scale, multi-dimensional feature interaction module uses multiple Timesblock modules in parallel to construct multi-scale, multi-dimensional features such as wind speed, wind direction, and power as input for parallel training, enabling the prediction of corresponding feature values. The obtained feature values ​​are then output as power predictions through a parallel multilayer perceptron. Increasing the feature dimension of the training data naturally improves the accuracy of the prediction model, and parallel training reduces the likelihood of overfitting and non-convergence. Considering that the TimeNet backbone network can represent changes within and between periods, thus solving the problem of modeling complex time variations, a TimesMLP multi-dimensional feature power prediction model based on TimesNet and a multilayer perceptron is established.

[0095] Specifically, the historical power output data and meteorological data of positive power sources in the database from step 1 are input into the TimesMLP prediction model to obtain predicted power output values ​​for positive power sources such as photovoltaic and wind power, including:

[0096] Step 3.1: In the TimesMLP prediction model, the multi-dimensional time data (such as wind direction, wind speed, light intensity, temperature, and power data) from the database in Step 1 is input as the raw time series data. Each dimension of the raw time series data is then processed... Perform a fast Fourier transform according to equation (6), filter the top k significant frequencies in the frequency domain, calculate the corresponding period length for each significant frequency, and reshape the one-dimensional time data of each period length into a two-dimensional tensor matrix. The reshaping process is to stack the components with k significant frequencies together to complete the reshaping. In this tensor matrix, rows represent changes during the period and columns represent changes within the period.

[0097] Specifically, based on the multi-period nature of time series, a modular time network structure is proposed to capture time patterns in different periods. For each period, in order to capture the corresponding intra-period and inter-period changes, this invention designs a TimesBlock in TimesNet (a model for time series analysis), which can convert one-dimensional time series into two-dimensional space and simultaneously model both types of changes through parameter-efficient initial blocks.

[0098] Each point in time involves two types of temporal changes simultaneously across different periods: intra-period changes and inter-period changes, both related to adjacent regions and in phase. However, this original one-dimensional time series structure can only represent changes between adjacent points in time. To overcome this limitation, this invention explores a two-dimensional temporal structure that can explicitly represent changes within and between time periods, thus offering greater representational power and facilitating subsequent representation learning.

[0099] Specifically, for a time series of length T and record variables C, its original 1D is constructed as follows: To represent the periodic variations, this invention requires first identifying the period. Technically, this invention utilizes the Fast Fourier Transform (FFT) to analyze the time series in the frequency domain as follows:

[0100]

[0101] This invention can transform a one-dimensional time series into a structured two-dimensional tensor, thus facilitating two-dimensional kernel processing. By performing the same reshaping operation on all variables of the time series, this invention can extend the above process to multivariate time series.

[0102] Here, FFT(·) and Amp(·) represent the calculation of FFT and amplitude values, respectively. The calculated amplitude for each frequency is represented by the average value calculated using Avg(·) over C dimensions. Note that the j-th value A represents the period length. The strength of the corresponding frequency j-period basis function.

[0103] Considering the sparsity of the frequency domain, and to avoid noise from meaningless high frequencies, this invention selects only the first k amplitudes, using unnormalized amplitude values. Obtain the most significant frequencies {f1,…,f k}, where k is a hyperparameter. These selected frequencies also correspond to k period lengths {p1,…,p}. k Due to the conjugacy of the frequency domain, this invention only considers... The frequency within. The equations of this invention are summarized as follows.

[0104] A,{f1,…,f k},{p1,…,p k} = Period(X 1D (7)

[0105] Based on the selected frequency {f1,…,f k} can reconstruct 1D time series and the corresponding period lengths {p1,…,p k The present invention decomposes into multiple two-dimensional tensors, where Padding(·) is used to expand the time series by 0 along the time dimension to make it more comparable. Where p i and f i Let represent the number of rows and columns of the transformed two-dimensional tensor, respectively. Indicates based on frequency f i The i-th reconstructed time series has columns and rows representing the corresponding period lengths p, respectively. i The changes within the cycle and the changes during the cycle. Ultimately, as... Figure 3 As shown, based on the selected frequency and the estimated period, this invention obtains a set of two-dimensional tensors. It represents k different two-dimensional time variations derived from different periods.

[0106] It is also worth noting that this transformation introduces two types of locality into the transformed two-dimensional tensor: locality between adjacent time points (columns, variations within a period) and locality between adjacent periods (rows, variations within a period). Therefore, two-dimensional temporal transformations can be easily handled by two-dimensional kernels.

[0107] Step 3.2: The TimesBlock module based on the parameter-sharing Inception block projects and reshapes the original time-series data in two dimensions. Combined with the two-dimensional tensor from Step 3.1, multi-period two-dimensional convolutional features are obtained. Then, through amplitude-weighted feature fusion, the fused time-series feature vector is obtained. This invention establishes TimesBlock using a ResNet approach. Specifically:

[0108] 1) The original time-series data is projected into a d-dimensional space through an embedding layer to obtain d-dimensional embedding features;

[0109] For a 1D input time series of length T This invention begins with an embedding layer Project the original input onto the depth features middle.

[0110] For the l-th level of TimesBlock, the input is For the l-th TimesBlock, the entire process consists of two consecutive parts: capturing 2D time variations and adaptively aggregating representations from different periods.

[0111] 2) Capture time 2d change: The d-dimensional embedded features are reshaped into a two-dimensional tensor matrix according to the period length obtained in step 3.1. The parameter-sharing Inception block is applied to it, and after periodically aligning with the two-dimensional tensor in step 3.1, multi-period two-dimensional convolutional features are obtained.

[0112] This invention can estimate depth features The period length Period(·) is transformed into a two-dimensional space to obtain a set of two-dimensional tensors. This invention can conveniently obtain information representation through an initial block with valid parameters, where X is the i-th transformed two-dimensional tensor.

[0113] After conversion, this invention processes the two-dimensional tensor into Inception(·) through a parameter-efficient Inception block, which involves a multi-scale two-dimensional kernel.

[0114] Then the invention will learn the 2D representation Convert back to 1D space In this invention, for aggregation, Trunc(·) is used to group elements of length (p) i ×f i The filled sequence is truncated to the original length T. Thanks to the transformation of the one-dimensional time series, the two-dimensional kernel in the initial block can simultaneously aggregate multi-scale intra-period variations (columns) and inter-period variations (rows), covering adjacent time points and adjacent time periods.

[0115] Furthermore, to improve parameter efficiency, this invention applies to two-dimensional tensors of different shapes. By using shared concept blocks, the model size remains unchanged regardless of the choice of hyperparameter k.

[0116] 3) Adaptive aggregation: Based on the frequency domain amplitude weighted fusion of multi-period two-dimensional convolution features, output the d-dimensional fused temporal feature vector.

[0117] This invention requires fusing k distinct 1D representations X for the next layer. Inspired by autocorrelation, the amplitude A can reflect the relative importance of the selected frequency and period, thus corresponding to the importance of each transformed 2D tensor. Since intra- and inter-period variations already involve multiple highly structured 2D tensors, TimesBlock can simultaneously and completely capture multi-scale temporal 2D variations. Therefore, TimesNet can achieve more efficient representation learning than directly from one-dimensional time series.

[0118] The versatility of 2D vision backbones: Thanks to converting one-dimensional time series into two-dimensional temporal variations, various computer vision backbones can be chosen to replace the inception block for representation learning, such as the widely used ResNet and ResNeXt, the advanced ConyNeXt, and attention-based models. Therefore, the two-dimensional temporal variation design of this invention also connects one-dimensional time series with the burgeoning 2D backbone, allowing time series analysis to fully leverage the developments in the computer vision community. Generally, the more robust the 2D backbone used for representation learning, the better the performance.

[0119] Step 3.3: Design a multilayer perceptron (MLP) to process the temporal feature vector after fusion of two-dimensional data, and predict the output of photovoltaic and wind power based on the temporal feature vector. After optimization, output the predicted values ​​of photovoltaic and wind power output. The hidden layer of the MLP adopts residual connection and Dropout regularization.

[0120] In addition to the input and output layers, a multilayer perceptron can have multiple hidden layers. The simplest MLP contains only one hidden layer, i.e., a three-layer structure.

[0121] In a multilayer perceptron, all layers are fully connected. The bottom layer of a multilayer perceptron is the input layer, the middle layers are the hidden layers, and the bottom layer is the output layer.

[0122] The neurons in the hidden layer are fully connected to the input layer. Assuming the input layer is represented by a vector X, the output of the hidden layer is f(W1X+b1), where W1 is the weight, b1 is the bias, and the activation function f is usually the sigmoid function or the tanh function.

[0123] This invention is based on Figure 3 , Figure 4 A multi-dimensional feature output prediction model based on TimesNet and multilayer perceptron was established using TimesMLP.

[0124] If a prediction model is used to directly predict future power output based on historical photovoltaic / wind power output, the number of features provided to the prediction model is relatively small. Even though TimesNet performs FFT decomposition on the data to increase the number of features and achieve more accurate predictions, the TimesMLP model includes a Timesblock module, a prediction module, and a multi-scale, multi-dimensional feature interaction module. The Timesblock module includes TimesNet's two-dimensional tensor decomposition module and adaptive aggregation of feature tensors. The prediction module is mainly used for multi-scale, multi-dimensional feature prediction. The multi-scale, multi-dimensional feature interaction module uses multiple Timesblocks in parallel to construct multi-scale, multi-dimensional features such as wind speed, wind direction, and power as input for parallel training, enabling the prediction of corresponding feature values. The obtained feature values ​​are then output as power prediction values ​​through a multilayer perceptron in parallel. Increasing the feature dimension of the training data naturally improves the accuracy of the prediction model, and parallel training is less prone to overfitting and non-convergence problems.

[0125] Step 4: Calculate the comprehensive regulation capability of the positive power source, negative power source and energy storage power source in the virtual power plant based on the predicted output value of the positive power source and the power market data. Analyze the response effect of the virtual power plant based on the comprehensive regulation capability of the virtual power plant and dynamically adjust the control strategy of the virtual power plant.

[0126] Further preferably, considering typical distributed flexible resource clustering, a static comprehensive regulation characteristic model for VPPs is established based on the regulation characteristics of positive power sources, negative power sources, and energy storage power sources. Combining the predicted output of positive power sources such as distributed photovoltaic / wind power, the response effect of VPPs is predicted, and their control strategies are further adjusted to complete the control tasks, achieving dynamic evaluation and prediction of the operating power of massive distributed flexible resource aggregation virtual power plants. On this basis, deep reinforcement learning technology is introduced to further optimize the control strategy, achieving collaborative optimization of resources within the VPP. This ensures optimal resource scheduling and load balancing in response to dynamic changes and emergencies in the electricity market, enabling intelligent scheduling of various resources within the virtual power plant (VPP), allowing it to efficiently respond to real-time demands and provide peak-shaving ancillary services in the electricity market. A cross-VPP cooperation mechanism is utilized to achieve resource sharing and joint scheduling among multiple virtual power plants, thereby improving the regulation capacity and operational stability of the entire power grid system. A new dynamic collaborative scheduling mechanism is introduced to ensure that various resources can complement each other and work synergistically during resource aggregation, thereby maximizing the overall performance of the system.

[0127] In step 4, the comprehensive regulation capacity of the positive power source, negative power source, and energy storage power source within the virtual power plant is calculated based on the predicted output value of the positive power source and electricity market data. The response effect of the virtual power plant is analyzed based on its comprehensive regulation capacity to dynamically adjust the virtual power plant control strategy, including:

[0128] Step 4.1: Based on the characteristics of the positive power source, negative power source, and energy storage power source, establish a static integrated regulation characteristic model of the virtual power plant for calculating the integrated regulation capacity;

[0129] When power sources, loads, and energy storage nodes within a region or substation are aggregated into a single Virtual Power Controller (VPP), it is inevitable that a VPP will contain multiple DERs (Regulators and Dependents). In this case, various types of DERs may be used simultaneously for frequency regulation. Due to the significant differences in the droop control characteristics of different resources, the VPP will have multiple possible operating states. For each operating state, the comprehensive regulation characteristics of the VPP can be obtained by combining the droop control characteristic functions of the resources involved in regulation. This is a static regulation characteristic, indicating the active power output that the VPP can provide at the current frequency.

[0130] It is worth noting that EVs are a special type of distributed resource, and their power attributes are closely related to their operating modes. When an EV is in demand mode (i.e., the period when the power in the EV load curve is positive), it is equivalent to a continuously adjustable negative power source; when it is in free mode (i.e., the period when the power in the EV load curve is zero), it is considered an energy storage power source and cannot be dispatched by the VPP when in travel mode.

[0131] Consider the following two types of heterogeneous configurations for distributed resources: 1) Combinations of resources of the same type (i.e., combinations of two positive power sources / continuously adjustable negative power sources / tiered adjustable negative power sources / energy storage power sources, totaling 4 heterogeneous configurations); 2) Combinations of resources of different types (combinations of positive power sources and continuously adjustable negative power sources, positive power sources and energy storage power sources, tiered adjustable negative power sources and energy storage power sources, etc., totaling 11 heterogeneous configurations). In the naming of heterogeneous configurations, P, NC, NG, and S represent positive power sources, continuously adjustable negative power sources, tiered adjustable negative power sources, and energy storage power sources, respectively. ao P ao1 P ao2 P represents the rated operating power of each positive power source. bo P bo1 P bo2 P represents the rated operating power of each continuously adjustable negative power supply. co P co1 P co2 The rated operating power of each adjustable negative power source is given, assuming that the rated operating power of each energy storage power source is zero.

[0132] Taking the "P-NG-S" heterogeneous approach as an example, this paper illustrates the derivation and generation process of the static integrated regulation characteristics of VPP. Wherein, P... a P b2 These are the operating power of power sources A and B at their rated frequencies (taking the rated operating power of the energy storage power source P0 = 0 as an example); P a,max P represents the maximum output power of positive power supply A; negative power supply B is a three-level adjustable load.b1 and P b3 These are the other two power levels; P c,min and P c,max Here, f1 and f6 represent the maximum charging and discharging power of energy storage power source C; f3 and f7 represent the inflection points of the positive power supply droop characteristic; f2, f4, f5, and f8 represent the inflection points of the energy storage power supply droop characteristic (f4 and f5 are the upper and lower limits of the adjustment dead zone). When the above three types of power sources are combined and operated, let their frequencies be f1, f2, f3, f4, f5, and f8 respectively. A f B and f C The output power is P A P B and P C Then the frequency of the virtual power plant is f = f A =f B =f C Total output power P VPP =P A +P B +P C Based on different droop characteristic parameters, several other operating conditions of VPP can be obtained, and their comprehensive droop characteristics can be derived using the same method.

[0133] The output of a positive power source mainly depends on external climatic conditions (such as sunlight, wind speed, etc.). Its regulation capability model is as follows:

[0134] P adjustable,gen (t)=α·P gen (t)

[0135] Where α is the adjustment coefficient, which typically depends on the power supply type and output stability. P gen (t) represents the output power of the positive power supply.

[0136] Negative power sources participate in grid load balancing by regulating electricity demand. Their demand response model can be defined as:

[0137] D i (t)=D i0 -β i ·△P price(t)

[0138] Where D i0 For the user's basic needs, β i For user response sensitivity, ΔP price(t) This refers to demand adjustments caused by price changes.

[0139] Regulation capability model: The regulation capability of a negative power source can be defined as:

[0140] P adjustable,dem (t)=i∑βi ·△P price(t)

[0141] Where, ∑β i It is the sum of the demand response coefficients of all users.

[0142] Modeling the characteristics of energy storage power sources (such as battery energy storage systems): The role of energy storage power sources in virtual power plants is to balance instantaneous supply and demand fluctuations and provide rapid response. Charge and discharge model: The charging and discharging process of an energy storage system can be represented by power balance equations.

[0143] P adjustable,storage (t)=min(P discharge_max ,max(-P charge_max ,P storage(t) ))

[0144] Among them, P storage(t) It is the net power output of the energy storage system, P charge and P discharge These represent charging and discharging power, respectively.

[0145] Regulation capability model: The regulation capability of energy storage power sources is limited by their capacity and charge / discharge capabilities, and can be defined as:

[0146] P adjustable(t) =min(P discharge_max ,max(-P charge_max ,P storage(t) ))

[0147] Among them, P charge_max and P discharge_max These are the maximum power for charging and discharging, respectively.

[0148] A comprehensive regulation capacity model for virtual power plants:

[0149] Combining the regulation capabilities of all power sources, the overall regulation capability of the virtual power plant can be expressed as:

[0150] P VPP_adjustable (t)=P adjustable,gen (t)+P adjustable,dem (t)+P adjustable,storage (t)

[0151] Among them, P adjustable,gen (t) represents the regulation capability of the positive power supply, P adjustable,dem (t) represents the regulation capability of the negative power supply, P adjustable,storage (t) represents the regulation capability of the energy storage power source.

[0152] Step 4.2: The static integrated regulation characteristic model of the virtual power plant established in Step 4.1, combined with the positive power output prediction results in Step 3 and the power market data, is used to obtain the integrated regulation capability of the positive power source, negative power source and energy storage power source in the virtual power plant, and then predict the response effect of the virtual power plant. Based on the response effect of the virtual power plant, a VPP regulation algorithm based on multi-agent deep reinforcement learning is designed to adjust the virtual power plant control strategy.

[0153] Considering the predicted output of distributed photovoltaic and wind power, a certain prediction of the VPP response effect is made, realizing a dynamic assessment and prediction method for the operating power of virtual power plants aggregating massive distributed flexible resources. The control strategy is then adjusted accordingly to complete the control task.

[0154] Based on historical information on distributed photovoltaic (PV) and wind power (Wind) power output, the output is predicted over a long period. This predicted output, along with positive power sources such as CHP units, small hydropower, and biomass power generation, participates in the designated day-ahead planning. Short-term PV and Wind power output predictions forecast future changes in output, which in turn affect the virtual power plant's integrated regulation characteristic curve and system operating status. Based on these PV and Wind power output predictions, the virtual power plant participates in intraday grid regulation, providing peak-shaving ancillary services to the grid.

[0155] During the real-time operation of the VPP, each DER within it needs to sample its operating status. Let the sampling period be Ts, then the DER power and energy state information P(kT) at time kTs can be obtained. s )E(kT s This allows for further depiction of the VPP's active power-frequency (Pf) integrated regulation characteristic curve and the current operating state's position on the characteristic curve (i.e., the state interval and operating point). Within a scheduling cycle, the VPP's integrated regulation characteristic remains unchanged, and the aforementioned operating interval will move parallel to each other in a three-dimensional space with time t as the z-axis, power P as the y-axis, and frequency f as the z-axis, forming a state plane. As time increases, each DER will adjust its operating state according to the day-ahead economic plan curve, and the VPP aggregator will also perform frequency regulation power scheduling on the DERs (typically within a scheduling cycle T). D It should include several sampling periods Ts, meaning the scheduling timescale is larger than the state sampling timescale. Therefore, the VPP will transition on multiple state planes, which together constitute the state space of the VPP's dynamic operation, and the lines connecting the points on the planes constitute the VPP's state trajectory. The above process is the dynamic aggregation process of the VPP.

[0156] This invention analyzes the response effect of a virtual power plant based on its comprehensive regulation capability, and dynamically adjusts the virtual power plant control strategy accordingly, including:

[0157] If the overall regulation capability of the virtual power plant is lower than the set value, the response effect of the virtual power plant is considered to be below expectations, and the virtual power plant control strategy will be dynamically adjusted in the following way:

[0158] (1) A robust optimization algorithm is introduced to address extreme weather or equipment failures and improve system stability. Considering the operational stability of the virtual power plant in the face of uncertainties such as extreme weather and equipment failures, a robust optimization algorithm is proposed. In this algorithm, the virtual power plant can adjust its control strategy in real time to address potential disturbances, such as fluctuations in photovoltaic / wind power, ensuring stable operation even when the grid load fluctuates significantly. This method can effectively cope with various external uncertainties and enhance the reliability and security of the virtual power plant.

[0159] (2) Based on the response effect, the synergy between photovoltaic, wind power, energy storage, and load is optimized through a multi-level dynamic interaction model. This model is used to optimize the dynamic synergy among various distributed resources within the virtual power plant (VPP). By considering the dynamic interaction of resources such as photovoltaic, wind power, energy storage batteries, and electric vehicles, the model establishes a multi-level feedback mechanism, thereby enabling the virtual power plant to automatically adjust to fluctuations in electricity demand and resource changes. Through real-time data acquisition and analysis, the model dynamically adjusts the power output of various resources to ensure the stability of the power grid and the maximum benefit of the virtual power plant.

[0160] (3) When the assessment reveals insufficient local resources (i.e., the overall regulation capacity of the virtual power plant is lower than the set value, and the response effect of the virtual power plant does not meet expectations), resources and information are shared through a cross-virtual power plant cooperation mechanism. The cooperation mechanism between virtual power plants proposes a collaborative scheduling strategy among different virtual power plants, achieving joint scheduling of multiple virtual power plants in the power grid through information and resource sharing. Under this mechanism, each virtual power plant not only schedules according to its own resources but also shares and optimizes resource allocation based on the operating status of other virtual power plants. This cross-VPP cooperation mechanism will effectively improve the flexibility of the power grid, especially ensuring stable operation during peak electricity demand periods.

[0161] (4) Using the evaluation results as reward signals, the output of energy storage and new energy sources is dynamically adjusted based on deep reinforcement learning to optimize market response. A reinforcement learning-based scheduling optimization model is introduced to optimize the scheduling strategy of the virtual power plant. In this model, the virtual power plant adaptively adjusts the output of resources such as energy storage, wind power, and photovoltaics based on real-time environmental feedback (such as grid frequency and market electricity price fluctuations). Through reinforcement learning, the virtual power plant can autonomously learn how to dynamically adjust power allocation in a volatile electricity market, thereby optimizing operating efficiency.

[0162] Step 3 establishes a multi-dimensional feature-based power output prediction model using TimesNet and TimesMLP. This model predicts short-term photovoltaic / wind power output through multi-level, multi-dimensional historical feature data. It combines feature fusion from multiple data sources and further improves prediction accuracy through parallel training. Simultaneously, the reinforcement learning algorithm from Step 4 is used to dynamically adjust and optimize the model, enhancing the real-time performance and accuracy of the predictions.

[0163] The beneficial effects of this invention are compared with those of the prior art:

[0164] This invention, for the first time, proposes a joint design of TimesBlock and Multilayer Perceptron (MLP) to construct a Time Series Analysis Network (TimeMLP) prediction model. Utilizing the TimesNet architecture, time series data is transformed from one-dimensional to two-dimensional, effectively improving the ability to capture periodic and cross-period changes. By combining Fast Fourier Transform (FFT), significant periodic features are extracted and efficiently modeled in two-dimensional space, thereby enhancing the analysis and prediction capabilities of complex time series patterns. This improves prediction accuracy while reducing model size for real-time prediction. It addresses the problem that existing virtual power plant regulation feasible regions or response capabilities only focus on boundary values ​​under constraints, neglecting intermediate states and actual operating conditions. Furthermore, existing response strategy research lacks consideration of distributed energy resources (DER) types in control methods, failing to effectively monitor and predict changes in the operating state of virtual power plants during dynamic processes. Current mainstream output prediction models lack both prediction accuracy and speed. The TimesMLP prediction model of this invention, through joint training of TimesNet and MLP, can accurately predict multi-dimensional feature data, effectively improving the model's adaptability to complex input data and thus enhancing prediction accuracy.

[0165] This invention unifies the modeling of the heterogeneous characteristics of various distributed resources and constructs a static integrated regulation characteristic model to calculate the integrated regulation capability of a virtual power plant, overcoming the limitations of existing methods in adapting to the regulation of heterogeneous resources. Based on this model, this invention calculates the integrated regulation capability of positive power sources, negative power sources, and energy storage power sources within the virtual power plant. It analyzes the response effect of the virtual power plant based on its integrated regulation capability, dynamically adjusting the virtual power plant control strategy to achieve dynamic adjustment of the virtual power plant's operation strategy, significantly improving the virtual power plant's rapid response capability to grid demand. This invention effectively solves the problems of numerous resource types, large quantities, and diverse regulation characteristics encountered in the dynamic assessment and prediction of operating power of virtual power plants with massive distributed flexible resources. By introducing deep reinforcement learning, more efficient resource coordination and optimized scheduling can be achieved in real-time dynamic regulation, thereby enhancing the dynamic response capability and market adaptability of the virtual power plant. This addresses the limitations of existing technologies where the dynamic response capability of virtual power plants is restricted by traditional static models, failing to fully capture complex time series characteristics and lacking real-time collaborative optimization mechanisms. Furthermore, existing methods cannot effectively balance economy and stability when facing scenarios such as power market fluctuations and sudden changes in the state of distributed resources.

[0166] This invention, based on the Time Series Analysis Network (TimesMLP), introduces deep reinforcement learning. By fusing the Time Series Analysis Network and deep reinforcement learning, the control strategy of the virtual power plant is dynamically adjusted. The Time Series Analysis Network can accurately capture complex time series variation characteristics, especially excelling in periodic and nonlinear changes. Meanwhile, deep reinforcement learning (DRL) can achieve real-time scheduling optimization, effectively balancing the scheduling of various distributed resources in the virtual power plant, improving system response capability and peak-shaving efficiency. By introducing reinforcement learning algorithms and multi-dimensional time series analysis, the dynamic operating status of the virtual power plant can be more accurately reflected. It can achieve dynamic adjustment and intelligent, efficient scheduling optimization in complex power market environments, thereby optimizing the operating efficiency and response capability of the virtual power plant. It is suitable for the collaborative optimization and real-time response of massive distributed flexible resources, making up for the shortcomings of existing virtual power plant technologies, such as a single feasible adjustment domain, limited response capability, and insufficient dynamic state monitoring. It can overcome the dynamic response problem in existing methods and significantly improve the operating accuracy and scheduling flexibility of the virtual power plant.

[0167] This invention not only enables dynamic evaluation and optimization of the response performance of virtual power plants in peak-shaving ancillary services, but also effectively improves the market adaptability and scheduling accuracy of virtual power plants.

[0168] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0169] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0170] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0171] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), may execute the computer-readable program instructions by utilizing state information from the computer-readable program instructions.

[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A dynamic control method for a virtual power plant based on TimesMLP, characterized in that, include: Step 1: Establish a database of typical distributed flexible resources within the virtual power plant, including historical output data, meteorological data, and electricity market data of typical distributed flexible resources; Step 2: Analyze the regulation characteristics of typical distributed flexible resources in the virtual power plant and classify them into positive power sources, negative power sources, and energy storage power sources. Step 3: Input the historical power output data and meteorological data of the positive source into the Time Series MLP prediction model to obtain the predicted power output value of the positive source; Step 4: Calculate the comprehensive regulation capability of the positive power source, negative power source and energy storage power source in the virtual power plant based on the predicted output value of the positive power source and the power market data. Analyze the response effect of the virtual power plant based on the comprehensive regulation capability of the virtual power plant in order to dynamically adjust the virtual power plant control strategy.

2. The method for dynamic control of a virtual power plant based on TimesMLP according to claim 1, characterized in that: The typical distributed flexible resources include photovoltaics, wind power, energy storage, and load; the meteorological data includes wind direction, wind speed, light intensity, and temperature; and the electricity market data includes electricity prices.

3. The method for dynamic control of a virtual power plant based on TimesMLP according to claim 1, characterized in that: In step 2, the adjustment characteristics of typical distributed flexible resources are analyzed. Based on the response mode of each resource to frequency adjustment, the resources are divided into three categories: (1) Positive power supply: When the system frequency decreases, the positive power supply increases its output proportionally; when the system frequency increases, the positive power supply decreases its output proportionally. (2) Negative power supply: For a continuously adjustable negative power supply, when the system frequency increases, the negative power supply increases power consumption proportionally, and when the system frequency decreases, the negative power supply decreases power consumption proportionally; For a stepped adjustable negative power supply, the segmented step function corresponding to the stepped adjustment is expressed as a continuous linear droop characteristic. (3) Energy storage power supply: It does not operate within the frequency regulation dead zone, and stops frequency regulation when the SOC exceeds the limit.

4. The method for dynamic control of a virtual power plant based on TimesMLP according to claim 1, characterized in that: In step 3, the historical power output data and meteorological data of the positive source are input into the TimesMLP prediction model to obtain the predicted power output value of the positive source, including: Step 3.1: Using the multidimensional time data composed of historical power output data and meteorological data of the positive power source as the original time series data, perform fast Fourier transform analysis on the original time series data to extract two-dimensional tensors; Step 3.2: Based on the parameter-sharing Inception block, the TimesBlock module projects and reshapes the original time series data in two dimensions. Combined with the two-dimensional tensor from Step 3.1, multi-period two-dimensional convolutional features are obtained. The fused time series feature vector is obtained through amplitude-weighted feature fusion. Step 3.3: The Multilayer Perceptron (MLP) performs output prediction on the fused temporal feature vector, and outputs the predicted output value after optimization by the loss function.

5. The method for dynamic control of a virtual power plant based on TimesMLP according to claim 4, characterized in that: Step 3.1 specifically includes: In the TimesMLP forecasting model, the historical power output data and meteorological data of the input positive power source constitute multidimensional time data as the raw time series data. Each dimension of the raw time series data X is then analyzed. lD Perform a fast Fourier transform according to equation (6) and filter the top k significant frequencies in the frequency domain, then calculate the corresponding period length for each significant frequency. One-dimensional time data for each period length is reconstructed into a two-dimensional tensor, which includes variations during the period and variations within the period. Where FFT(·) and Amp(·) represent the calculation of FFT and amplitude values, respectively; Avg(·) is the averaging function. A Represents the calculated amplitude at each frequency; {f1,…,f k } represents the top k significant frequencies in the selected frequency domain, {p1,…,p k } represents the length of the k periods corresponding to the first k significant frequencies in the frequency domain, and T represents the time series length.

6. The method for dynamic control of a virtual power plant based on TimesMLP according to claim 5, characterized in that: Step 3.2 specifically includes: 1) The TimesBlock module projects the original time series data into a d-dimensional space through an embedding layer to obtain d-dimensional embedding features; 2) Reshape the d-dimensional embedding features into a two-dimensional tensor according to the period length obtained in step 3.1, apply the parameter-sharing Inception block to it, and align it with the two-dimensional tensor in step 3.1 period by period to obtain multi-period two-dimensional convolutional features; 3) Based on the frequency domain amplitude weighted fusion of multi-period two-dimensional convolution features, output the d-dimensional fused temporal feature vector.

7. The method for dynamic control of a virtual power plant based on TimesMLP according to claim 1, characterized in that: In step 4, the formula for calculating the comprehensive regulation capacity of the positive power source, negative power source, and energy storage power source within the virtual power plant is: P VPP_adjustable (t)=P adjustable,gen (t)+P adjustable,dem (t)+P adjustable,storage (t) P adjustable,gen (t)=α·P gen (t) P adjustable,dem (t)=i∑β i ·△P price(t) P adjustable,storage (t)=min(P discharge_max ,max(-P charge_max ,P storage(t) )) Among them, P adjustable,gen (t) represents the regulation capability of the positive power supply, P adjustable,dem (t) represents the regulation capability of the negative power supply, P adjustable,storage (t) represents the regulation capability of the energy storage power source; α is the adjustment coefficient, P gen (t) represents the predicted output of the positive power source at time t; β i It is the demand response coefficient of user i, ΔP price(t) The demand adjustment caused by the price change at time t; P charge_max and P discharge_max P represents the maximum power for charging and discharging the energy storage power supply, respectively. storage(t) Let t be the net power output of the stored energy.

8. The method for dynamic control of a virtual power plant based on TimesMLP according to claim 1, characterized in that: In step 4, the response effect of the virtual power plant is analyzed based on its comprehensive regulation capability, so as to dynamically adjust the virtual power plant control strategy, including: If the overall regulation capability of the virtual power plant is lower than the set value, the response effect of the virtual power plant is considered to be unsatisfactory, and the virtual power plant control strategy will be dynamically adjusted in the following way: Optimize the synergy between photovoltaics, wind power, energy storage, and load through a multi-level dynamic interaction model; Resources and information are shared through a cross-virtual power plant collaboration mechanism; Deep reinforcement learning is used to dynamically adjust energy storage and renewable energy output in order to optimize market response to dynamic changes and emergencies in the electricity market.

9. A virtual power plant dynamic control system based on TimesMLP, used to run the method described in any one of claims 1-8, characterized in that, The system includes: The database construction module is used to build a database of typical distributed flexible resources within the virtual power plant, including historical output data, meteorological data, and electricity market data of typical distributed flexible resources. The resource classification and analysis module is used to analyze the regulation characteristics of typical distributed flexible resources in the virtual power plant and classify them into positive power sources, negative power sources and energy storage power sources. The power output prediction module is used to input the historical power output data and meteorological data of the positive power source into the Time Series MLP prediction model to obtain the predicted power output value of the positive power source. The dynamic control module is used to calculate the comprehensive regulation capability of the positive power source, negative power source and energy storage power source in the virtual power plant based on the predicted output value of the positive power source and the power market data. Based on the comprehensive regulation capability of the virtual power plant, the response effect of the virtual power plant is analyzed so as to dynamically adjust the control strategy of the virtual power plant.

10. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1-8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1-8.

Images