A method for constructing a water-light virtual generator and light storage virtual energy storage aggregation model

Abstract

This invention discloses a method for constructing an aggregated model of a hydro-solar virtual generator and a photovoltaic-storage-charging virtual energy storage system, relating to the field of distributed flexible resource collaborative control technology. The method includes: acquiring the state of charge data and its distribution dispersion of the photovoltaic-storage-charging virtual energy storage, the real-time adjustability margin of the hydro-solar virtual generator, and collaborative task requirements; inputting the resource state data into a limit state perception network to obtain a resource limit state representation vector; inputting the vector and task requirement data into an attention mechanism network, calculating the correlation, and outputting a collaborative strategy weight distribution sequence; determining the target collaborative mode based on the weight sequence and generating a target operating model. This invention, through the collaboration of limit state perception and the attention mechanism network, provides early warning of resource physical limits and dynamically matches them with grid tasks, adaptively determining and switching operating modes, effectively avoiding system failures and equipment risks caused by resource exceeding limits, and significantly improving the system's safety, robustness, and reliability under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of distributed flexible resource aggregation and collaborative control technology, specifically to a method for constructing a water-solar virtual generator and a photovoltaic-storage-charging virtual energy storage aggregation model. Background Technology

[0002] With the rapid development of new power systems, aggregating massive distributed energy resources to participate in grid peak shaving, frequency regulation, and other ancillary services has become an important means of improving grid flexibility. Among them, the "photovoltaic-storage-charging virtual energy storage" system and the "hydro-photovoltaic virtual generator" system are two core flexible regulation resources in distributed flexible resource aggregation systems. In existing technologies, the collaborative control mechanisms for multiple types of virtual resources are mostly based on conventional normal potential assessment models for task allocation and scheduling decisions.

[0003] In actual engineering operations, the existing routine collaboration mechanisms have gradually revealed the following serious technical flaws:

[0004] Existing models lack a deep understanding of the electrochemical critical state of "virtual energy storage." The actual regulation capability of "photovoltaic-storage-charging virtual energy storage" is strictly constrained by its battery state of charge (SOC). After continuous participation in high-frequency peak shaving of the grid or multiple consecutive response commands, energy storage power stations are highly susceptible to entering critical states of "charge saturation (overcharge risk)" or "charge depletion (over-discharge risk)." More seriously, due to the physical degradation differences among a large number of individual cells at the underlying level, there is often a significant "weakest link effect" (i.e., high dispersion in SOC distribution) within the system. Existing scheduling models typically rely only on macroscopic average charge for coarse-grained dispatching, failing to identify the hidden risks under high dispersion. If high-power operation commands are continued to be issued according to the normal model in the critical state, it will not only lead to the failure of scheduling commands but also seriously damage the service life of battery equipment, and even cause physical safety accidents such as thermal runaway.

[0005] Existing models lack sufficient early warning capabilities regarding the natural physical boundaries of the "virtual generator." The output regulation of the "hydro-solar virtual generator" is rigidly constrained by natural conditions such as inflow, reservoir capacity limitations, and solar radiation. During the scheduling process, reservoir water levels may approach flood control limits or dead water levels at any time, while photovoltaic output is also limited by the confidence interval of meteorological forecasts. Existing coordination mechanisms often passively respond to these boundary conditions, lacking real-time early warning when the resources of both parties approach their limits.

[0006] Existing scheduling systems lack dynamic matching and adaptive switching mechanisms between "task difficulty" and "resource limits." When receiving complex collaborative task requests, existing distributed flexible resource aggregation systems often employ fixed global optimization algorithms (such as conventional heuristics) to solve problems. Even when the available capacity of "virtual energy storage" or "virtual generators" is nearing its limit, the system still attempts to achieve economic optimization or full-load output under normal conditions. Once a single resource's capacity is exhausted or hardware-level protection is triggered, the system often disconnects directly from the grid, lacking a proactive "resource capacity limit state awareness and seamless collaborative mode switching mechanism." This results in the system's inability to smoothly transition to a safe, degraded operating state when facing resource shortages or severe operating conditions. The collapse of a single resource can easily trigger a domino effect of failures in the overall collaborative control system, posing significant safety risks to the local power grid.

[0007] Therefore, there is an urgent need in this field for a model construction method that can sense the extreme state of water-light and light-storage-charging resources in real time, and can provide early warnings and automatically adjust collaborative strategies to degrade and protect equipment when resources approach physical limits, so as to significantly improve the robustness and reliability of distributed flexible resource aggregation systems. Summary of the Invention

[0008] This invention provides a method for constructing a virtual generator-energy storage and charging virtual energy storage aggregation model, which solves the technical problem that the existing distributed flexible resource collaborative scheduling mechanism lacks early warning and adaptive switching capabilities when resources approach physical limits, thus causing the system to fail as a whole and equipment to be damaged due to the depletion of local resources.

[0009] This invention is achieved through the following technical solution:

[0010] In a first aspect, this application provides a method for constructing a hybrid model of a hydro-solar virtual generator and a photovoltaic-storage-charging virtual energy storage system, characterized by the following steps:

[0011] Acquire average state of charge data and state of charge distribution dispersion data for the photovoltaic-storage-charging virtual energy storage; acquire real-time adjustability margin data and current collaborative task requirements data for the water-photovoltaic virtual generator.

[0012] Input the average state of charge data, the state of charge distribution dispersion data, and the real-time adjustable margin data into a preset limit state perception network to obtain a resource limit state representation vector.

[0013] Input the resource limit state representation vector and the current collaborative task requirement data into a preset attention mechanism network, calculate the correlation between the resource limit state representation vector and the current collaborative task requirement data through the attention mechanism network, and output the collaborative strategy weight distribution sequence.

[0014] The target collaborative mode is determined based on the weight distribution sequence of the collaborative strategy, and the target operation model of the water-solar virtual generator and the photovoltaic-storage-charging virtual energy storage is generated based on the target collaborative mode.

[0015] A further optimized solution is that the acquisition of real-time adjustable margin data of the water-solar virtual generator includes:

[0016] Collect the current reservoir water level data of the water-solar virtual generator and read the preset water level limit data stored in the database;

[0017] The difference between the preset water level limit value data and the current reservoir water level data is used to obtain water level distance limit value feature data, and the water level distance limit value feature data is normalized to obtain a standard water level feature vector.

[0018] The photovoltaic output prediction confidence interval data of the water-solar virtual generator is obtained through the meteorological forecasting system, and the upper and lower bound values ​​of the photovoltaic output prediction confidence interval data are extracted to construct the photovoltaic output boundary matrix.

[0019] The standard water level feature vector and the photovoltaic power output boundary matrix are fused together to output the real-time adjustable margin data of the hydro-photovoltaic virtual generator.

[0020] A further optimization scheme involves inputting the average state of charge data, the state of charge distribution dispersion data, and the real-time adjustable margin data into a preset limit state perception network to obtain a resource limit state representation vector, including:

[0021] The average state of charge data and the state of charge distribution dispersion data are spliced ​​together to form a first state tensor, and the real-time adjustable margin data is converted into a second state tensor.

[0022] The first state tensor is input into the first multilayer perceptron branch of the limit state perception network to extract the energy storage limit feature vector.

[0023] The second state tensor is input into the second multilayer perceptron branch of the limit state perception network to extract the power generation limit feature vector.

[0024] The inner product of the energy storage limit feature vector and the power generation limit feature vector is calculated to obtain the coupled sensing matrix. A fully connected dimensionality reduction operation is then performed on the coupled sensing matrix to generate the resource limit state representation vector.

[0025] A further optimization scheme is as follows: the step of calculating the correlation between the resource limit state representation vector and the current collaborative task requirement data through the attention mechanism network, and outputting the collaborative strategy weight distribution sequence, includes:

[0026] The current collaborative task requirement data is mapped as a query vector, and the resource limit state representation vector is mapped as a key vector and a value vector, respectively.

[0027] Calculate the dot product between the query vector and the key vector, and divide the dot product result by a preset scaling factor to obtain the relevance score matrix;

[0028] Applying a normalized exponential function to the relevance score matrix, the attention weight matrix is ​​output.

[0029] The product of the attention weight matrix and the value vector is calculated to obtain the fused feature representation, which is then input into a linear classifier to output the collaborative strategy weight distribution sequence.

[0030] A further optimization scheme is that determining the target collaborative mode based on the collaborative strategy weight distribution sequence includes:

[0031] Extract the weight values ​​of each mode corresponding to a plurality of preset alternative collaborative modes from the collaborative strategy weight distribution sequence;

[0032] Select the largest mode weight value from among the various mode weight values;

[0033] Determine whether the maximum mode weight value is greater than a preset mode switching threshold;

[0034] When the maximum mode weight value is determined to be greater than the mode switching threshold, the candidate collaborative mode corresponding to the maximum mode weight value is directly locked as the target collaborative mode.

[0035] A further optimization is that the method further includes:

[0036] When the maximum mode weight value is determined to be less than or equal to the mode switching threshold, the most recent stable operating mode data in the historical collaborative mode library is obtained.

[0037] Analyze the data from the most recent stable operating mode and extract the upper limit parameter of energy storage output and the lower limit parameter of generator regulation.

[0038] The updated energy storage limit parameters are obtained by multiplying the preset attenuation coefficient by the upper limit parameter of the energy storage output, and the updated generator limit parameters are obtained by adding the preset compensation constant to the lower limit parameter of the generator adjustment.

[0039] By combining the updated energy storage limiting parameters with the updated generator limiting parameters, a degraded collaborative mode is constructed, and the degraded collaborative mode is set as the target collaborative mode.

[0040] A further optimized solution is that the step of generating the target operation model of the hydro-solar virtual generator and the photovoltaic-storage-charging virtual energy storage based on the target collaborative mode includes:

[0041] Obtain the initial set of control parameters for the target operating model and the preset global optimization objective function;

[0042] The initial control parameter set is initialized to the ant position coordinates of the ant colony algorithm, and multiple virtual ants are released in the parameter solution space of the target running model;

[0043] The fitness value of each virtual ant is calculated based on the global optimization objective function, and the pheromone concentration in the parameter solution space is updated based on the fitness value.

[0044] When the ant colony algorithm reaches the preset maximum number of iterations, it outputs the optimal solution of control parameters corresponding to the path with the highest pheromone concentration, and uses the optimal solution of control parameters to update the target running model.

[0045] A further optimized solution is that the acquisition of average state of charge data and state of charge distribution dispersion data for the photovoltaic-storage-charging virtual energy storage includes:

[0046] Receive the individual state of charge observation values ​​of multiple independent energy storage units that constitute the virtual energy storage of the photovoltaic energy storage;

[0047] The average state of charge (SOC) data of all the individual unit SOC observations is obtained by calculating the average SOC data.

[0048] Calculate the root mean square error between each individual state of charge observation and the average state of charge data;

[0049] The mean square error data is input into a preset dispersion evaluation function for smoothing and filtering to output the dispersion data of the state of charge distribution of the photovoltaic-storage-charging virtual energy storage.

[0050] A further optimized solution is that obtaining the current collaborative task requirement data includes:

[0051] The original collaborative task requirements were analyzed, and the active power command time series and reactive power command time series were separated.

[0052] Perform a fast Fourier transform on the active power command time series to extract the task frequency domain feature vector;

[0053] Wavelet decomposition is performed on the reactive power command time series to obtain spatial domain feature vectors;

[0054] The task frequency domain feature vector and the spatial domain feature vector are fused to generate a multidimensional task feature matrix that characterizes the complexity of the collaborative task, and the multidimensional task feature matrix is ​​used as the current collaborative task requirement data.

[0055] Secondly, this application provides a system for constructing a virtual generator-photovoltaic-energy storage aggregation model that integrates photovoltaic-energy storage and charging, comprising:

[0056] The data acquisition module is used to acquire the average state of charge data of the photovoltaic-storage-charging virtual energy storage, the state of charge distribution dispersion data of the photovoltaic-storage-charging virtual energy storage, the real-time adjustable margin data of the hydro-photovoltaic virtual generator, and the current collaborative task requirement data.

[0057] The limit state perception module is communicatively connected to the data acquisition module and is used to receive the average state of charge data, the state of charge distribution dispersion data, and the real-time adjustable margin data, and input them into the preset limit state perception network to generate a resource limit state representation vector.

[0058] The strategy calculation module is communicatively connected to the limit state perception module and the data acquisition module. It is used to receive the resource limit state representation vector and the current collaborative task requirement data, and input them into a preset attention mechanism network. The attention mechanism network is used to calculate the correlation between the resource limit state representation vector and the current collaborative task requirement data to output a collaborative strategy weight distribution sequence.

[0059] The collaborative model generation module is communicatively connected to the strategy calculation module. It is used to determine the target collaborative mode according to the collaborative strategy weight distribution sequence, and generate the target operation model of the water-solar virtual generator and the photovoltaic-storage-charging virtual energy storage based on the target collaborative mode.

[0060] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0061] By using a limit state perception network and an attention mechanism network, the system achieves early perception of the physical limits of resources and dynamic matching of the difficulty of power grid tasks. This enables the system to adaptively determine and switch cooperative operation modes based on resource status and task requirements, avoiding cooperative failures and equipment risks caused by resource depletion. This significantly improves the system's operational safety and overall robustness under complex operating conditions. Attached Figure Description

[0062] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0063] Figure 1 A flowchart illustrating a method for constructing a virtual energy storage aggregation model combining hydro-solar virtual generator and photovoltaic-storage-charging virtual energy storage, provided in this application embodiment;

[0064] Figure 2 A flowchart illustrating the steps of obtaining real-time adjustable margin data of the water-solar virtual generator in step S1, as provided in the embodiments of this application;

[0065] Figure 3 A flowchart illustrating the method for generating the resource limit state representation vector in step S2 of this application embodiment;

[0066] Figure 4 A flowchart of the method for generating a collaborative strategy weight distribution sequence in step S3 of this application embodiment. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of this invention are only for explaining this invention and are not intended to limit this invention.

[0068] First, some of the technical terms used in this application will be explained to help those skilled in the art understand this application.

[0069] BMS: Battery Management System;

[0070] MLP: Multilayer Perceptron;

[0071] CNN: Convolutional Neural Network;

[0072] MSE: Mean Square Error;

[0073] FFT: Fast Fourier Transform;

[0074] ACO: Ant Colony Optimization;

[0075] PLC: Programmable Logic Controller;

[0076] MAE: Mean Absolute Error;

[0077] SGD: Stochastic Gradient Descent;

[0078] SCADA: Supervisory Control And Data Acquisition, a data acquisition and monitoring control system;

[0079] IEC: International Electrotechnical Commission.

[0080] This invention provides a method for constructing a virtual generator-energy storage and charging virtual energy storage aggregation model, which aims to solve the core problem of lack of early warning and adaptive switching capabilities in the collaborative scheduling of distributed flexible resources due to resources approaching physical limits through deep learning and attention mechanisms.

[0081] Firstly, such as Figure 1 As shown, this application provides a method for constructing a hybrid model of a hydro-solar virtual generator and a photovoltaic-storage-charging virtual energy storage system, including the following steps:

[0082] Step S1: Obtain multi-dimensional benchmark data; specifically, this includes the following steps:

[0083] Step S1A: Obtain the average state of charge data of the virtual energy storage system (PV-Storage-Charge).

[0084] Step S1B: Obtain the dispersion data of the state of charge distribution of the photovoltaic-storage-charging virtual energy storage;

[0085] Step S1C: Obtain real-time adjustable margin data of the water-solar virtual generator;

[0086] Step S1D: Obtain the current collaborative task requirement data.

[0087] Step S2: Input the average state of charge data, the state of charge distribution dispersion data, and the real-time adjustable margin data into the preset limit state perception network to obtain the resource limit state representation vector.

[0088] Step S3: Input the resource limit state representation vector and the current collaborative task requirement data into a preset attention mechanism network, calculate the correlation between the resource limit state representation vector and the current collaborative task requirement data through the attention mechanism network, and output the collaborative strategy weight distribution sequence.

[0089] Step S4: Determine the target collaborative mode according to the collaborative strategy weight distribution sequence, and generate the target operation model of the water-solar virtual generator and the photovoltaic-storage-charging virtual energy storage according to the target collaborative mode.

[0090] This embodiment provides comprehensive and accurate underlying data input for subsequent extreme state perception and strategy matching by simultaneously collecting the health status of the energy storage side, the regulation potential of the power generation side, and the task requirements of the grid side, breaking through the extensive mode of traditional scheduling that only relies on average power.

[0091] In one embodiment, step S1A: obtaining the average state of charge data of the photovoltaic-storage-charging virtual energy storage specifically includes the following steps:

[0092] Step S1A1: Receive in real-time observations of the individual state of charge (SOC) values ​​of the N independent energy storage units constituting the virtual energy storage system via the communication interface. .

[0093] Step S1A2: Calculate the arithmetic mean of all observations to obtain the average state of charge (SOC) data. The calculation formula is as follows:

[0094]

[0095] In the formula, The average state of charge data; This refers to the total number of the independent energy storage units; For the first Individual state of charge observations of each of the independent energy storage units.

[0096] In one embodiment, step S1B: obtaining the dispersion data of the state of charge distribution of the photovoltaic-storage-charging virtual energy storage specifically includes the following steps:

[0097] Step S1B1: Calculate the observation value for each individual unit. Compared with the average The mean squared error data D is calculated using the following formula:

[0098]

[0099] In the formula, This represents the mean squared error data. The larger the mean squared error data, the more abnormal battery cells are present in the system that are about to be depleted or saturated, indicating an extremely high risk of high-power charging and discharging.

[0100] Step S1B2: Input the mean square error data D into a preset dispersion evaluation function (such as a first-order hysteresis filter) for smoothing and filtering, and output the dispersion data of the state of charge distribution of the photovoltaic-storage-charging virtual energy storage.

[0101] In one embodiment, such as Figure 2 As shown, step S1C: Obtaining the real-time adjustable margin data of the water-solar virtual generator, specifically includes the following steps:

[0102] Step S1C1: Collect the current reservoir water level using water level sensors deployed at the reservoir nodes. Read the preset water level limit data (dead water level) stored in the database. and flood control limit water level ).

[0103] Step S1C2: Subtract the preset water level limit data from the current reservoir water level data to obtain water level distance limit feature data, and normalize this data to obtain a standard water level feature vector. The specific calculation formula for the normalization process is as follows:

[0104]

[0105] In the formula, These are elements in the standard water level feature vector; The current reservoir water level data; The dead water level in the preset water level limit data; The flood control limit water level is the preset water level limit data.

[0106] Step S1C3: Obtain the photovoltaic power output confidence interval data through the meteorological forecasting system, and extract its upper and lower bound values ​​to construct the photovoltaic power output boundary matrix.

[0107] Step S1C4: Perform feature fusion operation (such as tensor splicing) on ​​the standard water level feature vector and the photovoltaic power output boundary matrix to output the real-time adjustable margin data of the hydro-photovoltaic virtual generator.

[0108] In one embodiment, step S1D, obtaining the current collaborative task requirement data, specifically includes the following steps:

[0109] Step S1D1: Analyze the original collaborative task requirement data and separate the active power command time series and reactive power command time series.

[0110] Step S1D2: Perform a fast Fourier transform on the active power command time series to extract the task frequency domain feature vector.

[0111] Step S1D3: Perform wavelet decomposition on the reactive power command time series to obtain spatial domain feature vectors.

[0112] Step S1D4: Cross-fuse the task frequency domain feature vector with the spatial domain feature vector to generate a multi-dimensional task feature matrix representing the complexity of the collaborative task, which serves as the current collaborative task requirement data.

[0113] Based on step S1, this application utilizes a pre-trained deep neural network to perform nonlinear feature extraction and deep fusion on multi-source heterogeneous resource status data, mapping discrete observations into a high-dimensional, comprehensive representation vector that quantifies the degree to which the system is approaching its physical limits. This enables intelligent and early perception of critical states such as "power saturation", "power depletion", and "water level exceeding limits".

[0114] In one embodiment, such as Figure 3 As shown, step S2, inputting the average state of charge data, the state of charge distribution dispersion data, and the real-time adjustable margin data into a preset limit state perception network to obtain the resource limit state representation vector, specifically includes the following steps:

[0115] Step S21: Concatenate the average state of charge data and the state of charge distribution dispersion data into a first state tensor, and convert the real-time adjustable margin data into a second state tensor.

[0116] Step S22: Input the first state tensor into the first multilayer perceptron branch of the limit state perception network to extract the energy storage limit feature vector. .

[0117] Step S23: Input the second state tensor into the second multilayer perceptron branch of the limit state perception network to extract the power generation limit feature vector. .

[0118] Step S24: Perform an inner product operation on the energy storage limit eigenvector and the power generation limit eigenvector to obtain the coupled sensing matrix. Then, a fully connected dimensionality reduction operation is performed on the matrix to generate the resource limit state representation vector. The inner product operation is expressed as:

[0119]

[0120] In the formula, The coupling sensing matrix; This is the energy storage limit feature vector; This is the power generation limit feature vector; This represents an extension of the inner or outer product operation of vectors. The resulting coupled-aware matrix contains deep information about resource interactions.

[0121] Performing a fully connected dimensionality reduction operation uses linear mapping to filter out redundant features, and the compressed matrix is ​​a one-dimensional representation vector of the resource limit state, which greatly reduces the computational complexity of the subsequent attention network.

[0122] This embodiment introduces an attention mechanism to dynamically calculate the correlation between power grid task requirements and resource limit states, enabling the system to intelligently assess whether the current complex task matches the vulnerable links of resources. This mechanism achieves precise alignment between "task difficulty" and "resource limits," thereby outputting a set of probabilistic collaborative strategy weights. This provides a quantitative basis for adaptive decision-making and fundamentally avoids command failures or equipment risks caused by task-capability mismatches.

[0123] In one embodiment, such as Figure 4 As shown, step S3: Input the resource limit state representation vector and the current collaborative task requirement data into a preset attention mechanism network, calculate the correlation between the resource limit state representation vector and the current collaborative task requirement data through the attention mechanism network, and output the collaborative strategy weight distribution sequence, specifically including the following steps:

[0124] Step S31: Map the current collaborative task requirement data as a query vector. And respectively map the resource limit state representation vector as a key vector. And the value vector V.

[0125] Step S32: Calculate the dot product of the query vector and the key vector, and divide the dot product result by a preset scaling factor. The correlation score matrix is ​​obtained. :

[0126]

[0127] In the formula, This is the correlation score matrix; The query vector; This is the transpose of the key vector; This is the preset scaling factor, where is the dimension of the key vector.

[0128] Step S33: Apply the normalized exponential function (Softmax function) to the correlation score matrix to output the attention weight matrix.

[0129] Step S34: Perform matrix multiplication on the attention weight matrix and the value vector V to obtain the fused feature representation, and input the fused feature representation into a linear classifier (such as a fully connected layer with a Softmax activation function) for classification mapping, and output the collaborative strategy weight distribution sequence.

[0130] This embodiment adaptively selects the optimal or safest cooperative mode based on the confidence level of the strategy weights and uses an optimization algorithm to generate a specific control model. When the confidence level is insufficient, a degradation safety fallback mechanism based on historical steady-state parameters is automatically triggered, forcing the system to switch to an absolutely safe operating boundary. This ensures that the system can maintain safe and controllable operation under any conditions (including when the algorithm is uncertain or resources are extremely scarce), significantly improving overall robustness and eliminating cascading failures caused by inaccurate control logic.

[0131] In one embodiment, step S4, which determines the target cooperative mode based on the cooperative strategy weight distribution sequence, specifically includes the following steps:

[0132] Step S4A1: Extract the weight values ​​of each mode corresponding to the preset multiple alternative collaborative modes from the collaborative strategy weight distribution sequence.

[0133] Step S4A2: Select the largest mode weight value from the various mode weight values.

[0134] Step S4A3: Determine whether the maximum mode weight value is greater than a preset mode switching threshold (e.g., 0.85).

[0135] Step S4A4: When it is determined that the maximum mode weight value is greater than the mode switching threshold, the candidate collaborative mode corresponding to the maximum weight value is directly locked as the target collaborative mode.

[0136] Step S4A5: When the maximum mode weight value is determined to be less than or equal to the mode switching threshold, a security degradation mechanism is triggered, specifically including:

[0137] Step S4A5a: Obtain the most recent stable operating mode data from the historical collaborative mode library.

[0138] Step S4A5b: Analyze the most recent stable operation mode data and extract the upper limit parameter of energy storage output and the lower limit parameter of generator regulation.

[0139] Step S4A5c: Multiply the upper limit parameter of energy storage output by a preset attenuation coefficient (e.g., 0.8) to obtain the updated energy storage limit parameter, and add the preset compensation constant (e.g., 5% of rated power) to the lower limit parameter of generator adjustment to obtain the updated generator limit parameter.

[0140] Step S4A5d: Combine the updated energy storage limiting parameters with the updated generator limiting parameters to construct a degraded collaborative mode, and set this degraded collaborative mode as the target collaborative mode.

[0141] In one embodiment, step S4, which generates the target operation model of the hydro-solar virtual generator and the photovoltaic-storage-charging virtual energy storage according to the target collaborative mode, specifically includes the following steps:

[0142] Step S4B1: Obtain the initial control parameter set corresponding to the target cooperative mode and the system's preset global optimization objective function.

[0143] Step S4B2: Initialize the initial control parameter set to the ant position coordinates of the ant colony algorithm, and release multiple virtual ants in the parameter solution space.

[0144] Step S4B3: Calculate the fitness value of each virtual ant according to the global optimization objective function, and update the pheromone concentration in the solution space based on the fitness value.

[0145] Step S4B4: When the ant colony algorithm reaches the preset maximum number of iterations, it outputs the optimal solution of the control parameters corresponding to the path with the highest pheromone concentration, and uses the optimal solution to update and generate the final target operation model of the water-solar virtual generator and the photovoltaic-storage-charging virtual energy storage.

[0146] To ensure the effectiveness of the pre-defined limit state awareness network and attention mechanism network, and to enable them to accurately learn the resource limit states and task-policy mapping relationship, both need to undergo sufficient offline training before deployment. This training is based on a large amount of historical running data, aiming to provide the model with a foundation for learning from experience.

[0147] Specifically, the training process of the limit state perception network is as follows:

[0148] First, extract data of the same type as described in steps S1A, S1B, and S1C from historical SCADA monitoring data. This includes historical average state of charge (SOC), SOC distribution dispersion, and real-time adjustable margin data, which will serve as input features for the model. Simultaneously, each input sample needs to be labeled with its corresponding ground truth label, representing the "remaining adjustable margin" of the system at that historical moment, indicating the distance from triggering hardware protection actions such as overcharging, over-discharging, or exceeding water level limits. This margin needs to be normalized to form a ground truth label vector. In the forward propagation and loss calculation phase of the network, the constructed input features are input into the initialized network. After processing and coupled calculation by its multilayer perceptron branch, the network outputs the predicted resource limit state representation vector. To measure the accuracy of network predictions, mean squared error (MSE) is used as the loss function L-sensitivity to quantify the predicted values. With real labels Loss values ​​between The formula for calculating its mean square error is:

[0149]

[0150] In the formula, This represents the number of samples in the training batch (BatchSize).

[0151] The goal of training is to minimize the loss function L_aware. To achieve this, the backpropagation algorithm combined with an optimizer (such as Adam) is used to calculate the gradient of the loss function with respect to all network weight parameters, and the parameters are iteratively updated according to the gradient direction. This process is repeated until the loss value converges to below a preset threshold, indicating that the network has been able to effectively extract and quantify the state of resources approaching physical limits from the input features. At this point, training is complete, and the network weight parameters are fixed.

[0152] The training of the attention mechanism network focuses on learning the dynamic matching relationship between "power grid task requirements" and "resource limit states" to output the optimal collaborative strategy. The construction of its training samples is more complex, requiring the establishment of a "command-state-decision result" triple sample library.

[0153] Specifically, the multi-dimensional task feature matrix obtained after processing historical scheduling tasks in step S1D is used as the input source for the query vector Q. Simultaneously, the resource limit state representation vector output by the previously trained limit state awareness network at the same historical moment is used as the input to the key vector K and value vector V, respectively. The true label Ptrue of the training samples is crucial to the training quality. It is determined by an expert system or a post-evaluation mechanism based on explicit safety rules. That is, based on the actual execution result of the historical task (whether it is safe and stable), it identifies which of the multiple pre-selected cooperative modes is the "historically optimal safe mode" at that time and converts it into a one-hot encoded form. During the forward propagation of training, the network takes Q, K, and V as inputs and performs the correlation calculation, attention weighting, and feature fusion operations as described in step S3. Finally, it outputs the predicted cooperative strategy probability distribution through a linear classifier. To guide the network to output the correct policy distribution, the cross-entropy loss function L-policy is used to measure the predicted distribution. Distribution of real labels The strategic loss value between The calculation formula is as follows:

[0154]

[0155] In the formula, This represents the total number of preset alternative collaboration modes.

[0156] The entire training process revolves around minimizing the cross-entropy loss L strategy, continuously adjusting the weight matrices mapping Q, K, and V, as well as the classifier parameters, through gradient descent. As training iterates, the network gradually learns to assign the highest probability weights to the safest and most suitable collaborative mode when faced with different combinations of resource states and task requirements. Training is complete when the model achieves the preset classification accuracy on an independent validation dataset, and the model is then available for use by an online scheduling system.

[0157] Furthermore, this embodiment of the invention also includes step S5: an online model fine-tuning mechanism, as an effective supplement to offline training to address potential device characteristic drift or environmental changes that may occur during long-term system operation. This mechanism constitutes a closed-loop online feedback optimization circuit.

[0158] After the aggregation model is officially put into operation, the system will establish a continuous data monitoring and comparison loop. Specifically, the system will synchronously collect actual operating status parameters of the hydro-solar virtual generator and the photovoltaic-storage-charging virtual energy storage, such as real-time power output, battery state of charge, and reservoir water level. Simultaneously, the system will use the expected baseline operating parameters (such as the expected power curve and state trajectory) calculated by the model based on the current scheduling instructions as a reference standard. By comparing the real-time collected parameters with this expected baseline on a time-by-time, multi-dimensional basis, the system can calculate the deviation between the two.

[0159] The calculated result of this deviation is quantified into a specific numerical indicator, namely the actual operational deviation loss value. This loss value intuitively reflects the degree of agreement between the current aggregated model's predicted output and the actual physical system response, and is a key indicator for measuring the model's online performance and adaptability. To proactively manage model performance degradation, the system internally presets a perceptual network fine-tuning threshold. This threshold serves as a trigger criterion to determine whether the model's prediction accuracy has decreased to a level requiring intervention.

[0160] When the system detects that the actual operational deviation loss value continuously exceeds the preset fine-tuning threshold over a period of time, an early warning is triggered. This indicates that the extreme state perception network, which was pre-trained using offline historical data, has become insufficient in its ability to represent the current actual operating conditions (which may be caused by equipment aging, component performance degradation, or changes in external environmental conditions), and the mapping relationship learned internally has become mismatched with the new operating conditions.

[0161] Once the update condition is triggered, the system will automatically enter the online fine-tuning process. This process first extracts the internal weight parameters of the current limit state perception network online. Next, the system uses the latest loss value, calculated from actual operational deviations, as the supervised learning signal. Using gradient descent, with this loss value as the target, it performs a small-step backpropagation and iterative update of the network's weight parameters. Essentially, this process uses the latest field data to "retrain" or "fine-tune" the network, fine-tuning its weights to better adapt to the current operating conditions.

[0162] By constructing the aforementioned closed loop of "monitoring-comparison-judgment-update," the system achieves continuous model optimization based on real-time operational feedback. This mechanism transforms the limit state awareness network—and indeed the entire aggregated model—from a static system fixed after deployment. Instead, it gains the ability to dynamically adapt to new operating conditions such as slow equipment aging, gradual performance degradation, and continuous evolution of the external environment. This continuous self-evolution and adaptive optimization significantly enhances the long-term robustness and decision-making accuracy of the entire aggregated model throughout its lifecycle, ensuring its reliable performance in complex, time-varying real-world power system environments.

[0163] In one embodiment, step S5 specifically includes the following steps:

[0164] Step S5A: Collect the actual operating status parameters of the system under the target operating model, compare them with the preset benchmark operating parameters, and calculate the actual operating deviation loss value.

[0165] Step S5B: Determine whether the actual operating deviation loss value exceeds the preset sensing network fine-tuning threshold.

[0166] Step S5C: When the threshold is exceeded, extract the internal network weight parameters of the limit state perception network.

[0167] Step S5D: Using the gradient descent method and combining the actual operational deviation loss data, the internal network weight parameters are updated iteratively in reverse.

[0168] Secondly, this application provides a system for constructing a virtual generator-photovoltaic-energy storage aggregation model that integrates photovoltaic-energy storage and charging, comprising:

[0169] The data acquisition module is used to acquire the average state of charge data of the photovoltaic-storage-charging virtual energy storage, the state of charge distribution dispersion data of the photovoltaic-storage-charging virtual energy storage, the real-time adjustable margin data of the hydro-photovoltaic virtual generator, and the current collaborative task requirement data.

[0170] The limit state perception module is communicatively connected to the data acquisition module and is used to receive the average state of charge data, the state of charge distribution dispersion data, and the real-time adjustable margin data, and input them into the preset limit state perception network to generate a resource limit state representation vector.

[0171] The strategy calculation module is communicatively connected to the limit state perception module and the data acquisition module. It is used to receive the resource limit state representation vector and the current collaborative task requirement data, and input them into a preset attention mechanism network. The attention mechanism network is used to calculate the correlation between the resource limit state representation vector and the current collaborative task requirement data to output a collaborative strategy weight distribution sequence.

[0172] The collaborative model generation module is communicatively connected to the strategy calculation module. It is used to determine the target collaborative mode according to the collaborative strategy weight distribution sequence, and generate the target operation model of the water-solar virtual generator and the photovoltaic-storage-charging virtual energy storage based on the target collaborative mode.

[0173] The functions of each module in the above-mentioned system for constructing a virtual generator and a virtual energy storage aggregation model for photovoltaic and energy storage are corresponding to the steps in the above-mentioned method for constructing a virtual generator and a virtual energy storage aggregation model for photovoltaic and energy storage. Their functions and implementation processes will not be described in detail here.

[0174] Thirdly, this application provides a device for constructing a virtual generator and a virtual energy storage aggregation model of water-solar virtual generator and solar energy storage-charging virtual energy storage. This device can be a personal computer (PC), a laptop, a server, or other device with data processing capabilities.

[0175] In this embodiment of the application, a device for constructing a virtual energy storage aggregation model of a water-solar virtual generator and a solar-storage-charging virtual energy storage model may include a processor, a memory, a communication interface, and a communication bus.

[0176] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0177] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces. These interfaces are used to interconnect devices within the photovoltaic-hydropower virtual generator and photovoltaic-storage-charging virtual energy storage aggregation model construction device, as well as interfaces to interconnect the device with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0178] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0179] The processor can be a general-purpose processor, which can call a program for constructing a virtual hydro-solar generator and a virtual energy storage aggregation model stored in memory, and execute the method for constructing a virtual hydro-solar generator and a virtual energy storage aggregation model provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the program for constructing a virtual hydro-solar generator and a virtual energy storage aggregation model is called can refer to the various embodiments of the method for constructing a virtual hydro-solar generator and a virtual energy storage aggregation model provided in this application, and will not be repeated here.

[0180] Fourthly, embodiments of this application also provide a readable storage medium.

[0181] This application stores a program for constructing a virtual energy storage aggregation model of a water-solar virtual generator and a virtual energy storage and charging virtual energy storage on a readable storage medium. When the program is executed by a processor, it implements the steps of the above-described method for constructing a virtual energy storage aggregation model of a water-solar virtual generator and a virtual energy storage and charging virtual energy storage.

[0182] The method implemented when the program for constructing a virtual generator and a virtual energy storage aggregation model of a water-solar power generation and a solar-storage-charging power generation is executed can be referred to in various embodiments of the method for constructing a virtual generator and a virtual energy storage aggregation model of a water-solar power generation and a solar-storage-charging power generation in this application, and will not be repeated here.

[0183] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a hybrid model of a hydro-solar virtual generator and a photovoltaic-storage-charging virtual energy storage system, characterized in that, Includes the following steps: Acquire average state of charge data and state of charge distribution dispersion data for the photovoltaic-storage-charging virtual energy storage; acquire real-time adjustability margin data and current collaborative task requirements data for the water-photovoltaic virtual generator. Input the average state of charge data, the state of charge distribution dispersion data, and the real-time adjustable margin data into a preset limit state perception network to obtain a resource limit state representation vector. Input the resource limit state representation vector and the current collaborative task requirement data into a preset attention mechanism network, calculate the correlation between the resource limit state representation vector and the current collaborative task requirement data through the attention mechanism network, and output the collaborative strategy weight distribution sequence. The target collaborative mode is determined based on the weight distribution sequence of the collaborative strategy, and the target operation model of the water-solar virtual generator and the photovoltaic-storage-charging virtual energy storage is generated based on the target collaborative mode.

2. The method for constructing a hybrid model of hydro-solar virtual generator and photovoltaic-storage-charging virtual energy storage according to claim 1, characterized in that, The acquisition of real-time adjustable margin data for the water-solar virtual generator includes: Collect the current reservoir water level data of the water-solar virtual generator and read the preset water level limit data stored in the database; The difference between the preset water level limit value data and the current reservoir water level data is used to obtain water level distance limit value feature data, and the water level distance limit value feature data is normalized to obtain a standard water level feature vector. The photovoltaic output prediction confidence interval data of the water-solar virtual generator is obtained through the meteorological forecasting system, and the upper and lower bound values ​​of the photovoltaic output prediction confidence interval data are extracted to construct the photovoltaic output boundary matrix. The standard water level feature vector and the photovoltaic power output boundary matrix are fused together to output the real-time adjustable margin data of the hydro-photovoltaic virtual generator.

3. The method for constructing a hybrid model of hydro-solar virtual generator and photovoltaic-storage-charging virtual energy storage according to claim 1, characterized in that, The process of inputting the average state of charge data, the state of charge distribution dispersion data, and the real-time adjustable margin data into a preset limit state perception network to obtain a resource limit state representation vector includes: The average state of charge data and the state of charge distribution dispersion data are spliced ​​together to form a first state tensor, and the real-time adjustable margin data is converted into a second state tensor. The first state tensor is input into the first multilayer perceptron branch of the limit state perception network to extract the energy storage limit feature vector. The second state tensor is input into the second multilayer perceptron branch of the limit state perception network to extract the power generation limit feature vector. The inner product of the energy storage limit feature vector and the power generation limit feature vector is calculated to obtain the coupled sensing matrix. A fully connected dimensionality reduction operation is then performed on the coupled sensing matrix to generate the resource limit state representation vector.

4. The method for constructing a converged model of hydro-solar virtual generator and photovoltaic-storage-charging virtual energy storage according to claim 1, characterized in that, The step of calculating the correlation between the resource limit state representation vector and the current collaborative task requirement data through the attention mechanism network, and outputting the collaborative strategy weight distribution sequence, includes: The current collaborative task requirement data is mapped as a query vector, and the resource limit state representation vector is mapped as a key vector and a value vector, respectively. Calculate the dot product between the query vector and the key vector, and divide the dot product result by a preset scaling factor to obtain the relevance score matrix; Applying a normalized exponential function to the relevance score matrix, the attention weight matrix is ​​output. The product of the attention weight matrix and the value vector is calculated to obtain the fused feature representation, which is then input into a linear classifier to output the collaborative strategy weight distribution sequence.

5. The method for constructing a converged model of hydro-solar virtual generator and photovoltaic-storage-charging virtual energy storage according to claim 1, characterized in that, The step of determining the target collaborative mode based on the collaborative strategy weight distribution sequence includes: Extract the weight values ​​of each mode corresponding to a plurality of preset alternative collaborative modes from the collaborative strategy weight distribution sequence; Select the largest mode weight value from among the various mode weight values; Determine whether the maximum mode weight value is greater than a preset mode switching threshold; When the maximum mode weight value is determined to be greater than the mode switching threshold, the candidate collaborative mode corresponding to the maximum mode weight value is directly locked as the target collaborative mode.

6. The method for constructing a hybrid model of hydro-solar virtual generator and photovoltaic-storage-charging virtual energy storage according to claim 5, characterized in that, The method further includes: When the maximum mode weight value is determined to be less than or equal to the mode switching threshold, the most recent stable operating mode data in the historical collaborative mode library is obtained. Analyze the data from the most recent stable operating mode and extract the upper limit parameter of energy storage output and the lower limit parameter of generator regulation. The updated energy storage limit parameters are obtained by multiplying the preset attenuation coefficient by the upper limit parameter of the energy storage output, and the updated generator limit parameters are obtained by adding the preset compensation constant to the lower limit parameter of the generator adjustment. By combining the updated energy storage limiting parameters with the updated generator limiting parameters, a degraded collaborative mode is constructed, and the degraded collaborative mode is set as the target collaborative mode.

7. The method for constructing a converged model of hydro-solar virtual generator and photovoltaic-storage-charging virtual energy storage according to claim 1, characterized in that, The step of generating the target operation model of the hydro-solar virtual generator and the photovoltaic-storage-charging virtual energy storage according to the target collaborative mode includes: Obtain the initial set of control parameters for the target operating model and the preset global optimization objective function; The initial control parameter set is initialized to the ant position coordinates of the ant colony algorithm, and multiple virtual ants are released in the parameter solution space of the target running model; The fitness value of each virtual ant is calculated based on the global optimization objective function, and the pheromone concentration in the parameter solution space is updated based on the fitness value. When the ant colony algorithm reaches the preset maximum number of iterations, it outputs the optimal solution of control parameters corresponding to the path with the highest pheromone concentration, and uses the optimal solution of control parameters to update the target running model.

8. The method for constructing a hybrid model of hydro-solar virtual generator and photovoltaic-storage-charging virtual energy storage according to claim 1, characterized in that, The acquisition of average state of charge (SOC) data and SOC distribution dispersion data for virtual energy storage (PV-Storage-Charge) includes: Receive the individual state of charge observation values ​​of multiple independent energy storage units that constitute the virtual energy storage of the photovoltaic energy storage; The average state of charge (SOC) data of all the individual unit SOC observations is obtained by calculating the average SOC data. Calculate the root mean square error between each individual state of charge observation and the average state of charge data; The mean square error data is input into a preset dispersion evaluation function for smoothing and filtering to output the dispersion data of the state of charge distribution of the photovoltaic-storage-charging virtual energy storage.

9. The method for constructing a hybrid model of hydro-solar virtual generator and photovoltaic-storage-charging virtual energy storage according to claim 1, characterized in that, The process of obtaining the current collaborative task requirement data includes: The original collaborative task requirements were analyzed, and the active power command time series and reactive power command time series were separated. Perform a fast Fourier transform on the active power command time series to extract the task frequency domain feature vector; Wavelet decomposition is performed on the reactive power command time series to obtain spatial domain feature vectors; The task frequency domain feature vector and the spatial domain feature vector are fused to generate a multidimensional task feature matrix that characterizes the complexity of the collaborative task, and the multidimensional task feature matrix is ​​used as the current collaborative task requirement data.

10. A system for constructing a virtual generator-photovoltaic-energy storage aggregation model and a photovoltaic-energy storage-charging virtual energy storage model, characterized in that, include: The data acquisition module is used to acquire the average state of charge data of the photovoltaic-storage-charging virtual energy storage, the state of charge distribution dispersion data of the photovoltaic-storage-charging virtual energy storage, the real-time adjustable margin data of the hydro-photovoltaic virtual generator, and the current collaborative task requirement data. The limit state perception module is communicatively connected to the data acquisition module and is used to receive the average state of charge data, the state of charge distribution dispersion data, and the real-time adjustable margin data, and input them into the preset limit state perception network to generate a resource limit state representation vector. The strategy calculation module is communicatively connected to the limit state perception module and the data acquisition module. It is used to receive the resource limit state representation vector and the current collaborative task requirement data, and input them into a preset attention mechanism network. The attention mechanism network is used to calculate the correlation between the resource limit state representation vector and the current collaborative task requirement data to output a collaborative strategy weight distribution sequence. The collaborative model generation module is communicatively connected to the strategy calculation module. It is used to determine the target collaborative mode according to the collaborative strategy weight distribution sequence, and generate the target operation model of the water-solar virtual generator and the photovoltaic-storage-charging virtual energy storage based on the target collaborative mode.

Images