Energy coordination optimization control method and system for virtual power plant

Abstract

The present invention provides a method and system for coordinated energy optimization control of a virtual power plant. Multiple virtual power plant node state data are encoded to obtain multiple first encoded feature representations; multiple first encoded feature representations are sequentially spliced ​​to obtain a first fused encoded feature representation and loaded into the first several domain adaptation networks; based on the response mode selector corresponding to each response mode, a second several domain adaptation network is determined from the first several domain adaptation networks, and the output of the determined second several domain adaptation networks is fused to obtain a second encoded feature representation corresponding to the response mode; the second encoded feature representation corresponding to the multiple response modes is loaded into the inference network corresponding to the multiple response modes, and the inferenced target virtual power plant node's response trigger confidence for the coordinated control strategy and the multiple response modes is obtained. The stability and efficiency of energy coordinated optimization control are increased.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to an energy coordination and optimization control method and system for a virtual power plant. Background Art

[0002] In the field of energy management, especially in the energy coordination and optimization control of virtual power plant systems, traditional technical solutions, for example, rely on public domain adaptation networks to handle prediction and decision-making tasks for various response modes. These public networks are designed to handle many different types of response modes, but due to their versatility, it is often difficult to provide optimal performance for a specific response mode. In order to improve processing efficiency and accuracy, some technical solutions have begun to adopt domain adaptation networks dedicated to response modes, that is, designing and training an independent network for each response mode. However, although this method improves performance to a certain extent, it also brings about a waste of computing resources and a substantial increase in computing power requirements, especially when there are a large number of domain adaptation networks.

[0003] On the other hand, with the rapid development of smart grid and distributed energy technology, the amount and complexity of data that virtual power plant systems need to process has increased dramatically. Traditional centralized control methods have been unable to meet the requirements of real-time and accuracy, so it is necessary to introduce more flexible and efficient distributed control strategies. In this context, how to effectively utilize computing resources and improve the stability and processing efficiency of energy coordinated optimization control has become an urgent problem to be solved.

[0004] In order to overcome the shortcomings of existing technologies, a new type of energy coordination optimization control method is needed. This method can dynamically determine the number of domain adaptation networks that need to be activated through selective activation strategies without increasing the global computing power fluctuations, thereby achieving accurate prediction and efficient processing of different response modes. Summary of the invention

[0005] The object of the present invention is to provide a method and system for energy coordination and optimization control of a virtual power plant.

[0006] The present invention is achieved in that:

[0007] In a first aspect, the present invention provides a method for coordinated and optimized energy control of a virtual power plant, comprising: encoding multiple virtual power plant node status data to obtain multiple first encoded feature representations, wherein the virtual power plant node status data is used to predict the response trigger confidence of a target virtual power plant node to a coordinated control strategy for multiple response modes; sequentially splicing multiple first encoded feature representations to obtain a first fused encoded feature representation; loading the first fused encoded feature representation into a first plurality of domain adaptation networks; based on a response mode selector corresponding to each of the response modes, determining a second plurality of the domain adaptation networks in the first plurality of the domain adaptation networks, and fusing the outputs of the determined second plurality of the domain adaptation networks to obtain a second encoded feature representation corresponding to the response mode, wherein the second number is smaller than the first number; loading the second encoded feature representation corresponding to the plurality of the response modes into an inference network corresponding to the plurality of the response modes to obtain the inferred response trigger confidence of the target virtual power plant node to the coordinated control strategy for the plurality of the response modes.

[0008] In a second aspect, the present invention provides a computer system comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method as described above is implemented.

[0009] In the embodiment of the present invention, the public domain adaptation network and the domain adaptation network dedicated to the response mode are not distinguished. Instead, they are executed based on the selective activation strategy, that is, for each response mode, only the second number of domain adaptation networks are activated in the first number of domain adaptation networks. Based on this, no matter how large the number of domain adaptation networks is, only the second number of domain adaptation networks are used, and the fluctuation of global computing power is not large. Under the premise of maintaining low computing power and demand, the number of domain adaptation networks is increased, the stability of energy coordination optimization control is increased, and the processing efficiency is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for describing the embodiments of the present invention are briefly introduced below.

[0011] Figure 1 It is a flow chart of an energy coordination and optimization control method of a virtual power plant provided by an embodiment of the present invention.

[0012] Figure 2 It is a schematic diagram of the composition of a computer system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0013] The following describes the embodiments of the present invention in conjunction with the accompanying drawings in the embodiments of the present invention. The terms used in the implementation method part of the embodiments of the present invention are only used to explain the specific embodiments of the present invention, and are not intended to limit the present invention.

[0014] The execution subject of the energy coordination optimization control method of the virtual power plant in the embodiment of the present invention is a computer system, including but not limited to a server, a personal computer, a laptop, a tablet computer, a smart phone, etc. Figure 1 As shown, the method includes:

[0015] Step S100: Encode multiple virtual power plant node status data to obtain multiple first encoding feature representations, wherein the virtual power plant node status data is used to predict the response trigger confidence of the target virtual power plant node to the coordinated control strategy for multiple response modes.

[0016] Virtual power plant nodes refer to those physical units that can participate in energy coordination and optimization control, which form the basis of the distributed charging network. These nodes include not only rechargeable smart light poles on both sides of the road, but may also cover other types of distributed energy resources, energy storage equipment and adjustable loads. Smart light poles are devices deployed on both sides of the road. They not only have lighting functions, but also integrate charging facilities, environmental monitoring sensors and communication modules. They can monitor their own power and environmental parameters (such as temperature, humidity, light intensity, etc.) in real time, and adjust their working status according to the coordinated control strategy (such as turning on energy-saving mode, participating in grid dispatch, etc.).

[0017] In addition to smart light poles, virtual power plant nodes may also include other types of distributed energy resources, such as solar photovoltaic panels, wind turbines, etc. These resources can generate electricity and participate in the supply and demand balance of the power grid through the virtual power plant system. Energy storage devices such as battery energy storage systems can store electricity when electricity demand is low and release electricity at peak times, thereby smoothing out grid load fluctuations. They are also an important part of virtual power plant nodes.

[0018] The virtual power plant node status data covers a variety of information related to the target virtual power plant node, including but not limited to the basic data of the light pole, environmental monitoring data and coordinated control strategy data. Taking the basic data of the light pole as an example, it may include basic information such as the geographical location, model, rated power, current power, etc. of the light pole. This information helps the computer system understand the physical properties and current status of the light pole. Environmental monitoring data may include environmental parameters such as temperature, humidity, wind speed, and light intensity, which have a direct impact on the energy consumption and performance of the light pole. The coordinated control strategy data refers to the specific information of the coordinated control strategy that the computer system intends to adopt, such as whether to turn on the energy-saving mode and whether to participate in power grid dispatch.

[0019] In order to integrate these diverse data into feature representations that can be used for machine learning, the computer system uses encoding technology to convert the original data into a format that is easier for the machine learning model to process. In an embodiment of the present invention, the encoding technology may include one-hot encoding, label encoding, embedding encoding, etc.

[0020] Taking one-hot encoding as an example, suppose that the virtual power plant node status data contains the model information of the lamp pole, and the lamp poles have multiple models. For each model, the computer system can assign it a unique integer identifier. Then, for each lamp pole, the computer system can create a binary vector equal to the number of models, in which only the bits corresponding to the lamp pole model are set to 1, and the remaining bits are set to 0. In this way, each lamp pole model is encoded into a unique binary vector. In addition to one-hot encoding, embedded coding can be used to learn the potential representation of variables by mapping discrete variables into high-dimensional continuous space. In an embodiment of the present invention, the computer system can learn an embedding vector for each virtual power plant node status data item (such as lamp pole model, environmental monitoring parameters, etc.). These embedded vectors can capture the complex relationships between data items and provide more valuable information for subsequent machine learning models.

[0021] During the encoding process, the computer system can also handle some missing values ​​or outliers. For missing values, the computer system can adopt a filling strategy, such as using the mean, median or mode for filling, or using an interpolation method for estimation. For outliers, the computer system can use an outlier detection algorithm to identify them, and choose to retain, correct or delete these outliers according to the actual situation. After completing the encoding, the computer system will obtain multiple first encoded feature representations. These feature representations are compact and informative representations of the original virtual power plant node state data. They will be used in subsequent machine learning models to predict the response trigger confidence of the target virtual power plant node to multiple response modes under different coordinated control strategies. For example, one response mode may be "turning on energy-saving mode", while another response mode may be "participating in power grid dispatching". Based on the first encoded feature representation, the computer system will use a machine learning model (such as a neural network, a decision tree, etc.) to calculate the confidence of the target virtual power plant node triggering these response modes under a given coordinated control strategy.

[0022] Step S200: sequentially concatenate multiple first coding feature representations to obtain a first fused coding feature representation.

[0023] In step S100, the computer system has encoded the virtual power plant node status data to obtain a plurality of first encoding feature representations.

[0024] For example, suppose there are three virtual power plant nodes: Node A, Node B, and Node C. The first coded feature representation of Node A may be a vector containing five elements [A1, A2, A3, A4, A5], where A1 represents the power of Node A, A2 represents the temperature of the environment where Node A is located, A3 represents the output power of the power generation equipment of Node A, A4 represents the power state of the energy storage equipment of Node A, and A5 represents the adjustable load power consumption of Node A. Similarly, the first coded feature representations of Node B and Node C are also vectors containing five elements [B1, B2, B3, B4, B5] and [C1, C2, C3, C4, C5] respectively.

[0025] In step S200, the task of the computer system is to concatenate the three first coded feature representations to form a new vector, namely the first fused coded feature representation. The concatenation process is simple, that is, to arrange the elements in all feature representations in order in a new vector. Continuing with the above example, the concatenated first fused coded feature representation will be a vector containing 15 elements [A1, A2, A3, A4, A5, B1, B2, B3, B4, B5, C1, C2, C3, C4, C5].

[0026] This new vector contains the status information of all virtual power plant nodes, providing a comprehensive and structured input for the subsequent machine learning model. Based on this input vector, the machine learning model will learn the complex relationship between the virtual power plant node status and the response trigger confidence, thereby optimizing the coordinated control strategy.

[0027] It is worth noting that the dimension (i.e., the number of elements) of the first fused encoded feature representation after splicing will depend on the number of virtual power plant nodes and the dimension of the first encoded feature representation of each node. In the embodiment of the present invention, there are three nodes, and the feature representation of each node has five dimensions, so the feature representation after splicing has 15 dimensions. In actual applications, the number of virtual power plant nodes and the feature dimension of each node may vary depending on the specific situation, resulting in different dimensions of the feature representation after splicing.

[0028] Step S300: Loading the first fused encoded feature representation into the first plurality of domain adaptation networks.

[0029] In step S200, the computer system has concatenated the first coded feature representations of multiple virtual power plant nodes to form a large vector containing all node status information. This vector, namely the first fused coded feature representation, is the input for subsequent processing steps. It contains information on multiple dimensions such as the power of the virtual power plant nodes, environmental parameters, power output of power generation equipment, power status of energy storage equipment, and power consumption of adjustable loads, providing comprehensive data support for the domain adaptation network.

[0030] A domain adaptation network is an expert network designed to process data in a specific domain or task. Compared with general machine learning models, domain adaptation networks have stronger domain adaptability and higher performance. In an embodiment of the present invention, a domain adaptation network is used to extract and transform useful information in the first fusion coding feature representation so that subsequent steps can more accurately predict the response trigger confidence of the virtual power plant node to multiple response modes under different coordinated control strategies.

[0031] In step S300, the computer system loads the first fused encoded feature representation into the first number of domain adaptation networks. The first number here refers to the number of domain adaptation networks, which depends on the specific application scenario and requirements. In actual operation, the computer system can determine the number of domain adaptation networks based on factors such as the scale of the virtual power plant, the number of nodes, and the dimension of the feature representation.

[0032] Taking neural networks as an example, each domain adaptation network can be regarded as a multi-layer perceptron (MLP), which contains an input layer, a hidden layer, and an output layer. The input layer receives the first fusion encoded feature representation as input; the hidden layer extracts and transforms useful information in the input features through a series of nonlinear transformations; and the output layer generates the output of the domain adaptation network. It should be noted that different domain adaptation networks may have different network structures and parameter settings to adapt to different processing requirements and task characteristics.

[0033] During the loading process, the computer system inputs the first fused encoded feature representation into each domain adaptation network. Each network processes the input features independently and generates corresponding outputs. These outputs may be high-dimensional feature vectors, probability distributions, or other forms of representations that contain useful information extracted from the input features.

[0034] For example, the domain adaptation network extracts and transforms useful information from input features through a series of calculations and nonlinear transformations. These calculations may include matrix multiplication, nonlinear activation functions (such as ReLU, sigmoid, etc.), pooling operations, etc. Through these calculations, the domain adaptation network is able to transform input features into higher-level representations, providing more valuable information for subsequent steps.

[0035] For example, suppose there are two domain adaptation networks: network A and network B. Network A may focus on processing features related to power and environmental parameters, while network B may pay more attention to the features of the output power of power generation equipment and the power state of energy storage equipment. When the first fusion encoded feature representation is loaded into these two networks, they will respectively extract and transform the feature information they are concerned about. Network A can generate a feature vector containing information related to power and environmental parameters as output, while network B can output a feature vector related to the state of power generation equipment and energy storage equipment.

[0036] These output feature vectors will then be used in subsequent steps, such as the selective activation and fusion operations in step S400. By combining the output feature vectors of multiple domain adaptation networks, the computer system can more comprehensively understand the state information of the virtual power plant nodes and make more accurate predictions and decisions.

[0037] Step S400: Based on the response mode selector corresponding to each response mode, a second number of domain adaptation networks are determined in the first number of domain adaptation networks, and the outputs of the determined second number of domain adaptation networks are fused to obtain a second encoded feature representation corresponding to the response mode, wherein the second number is smaller than the first number.

[0038] In an embodiment of the present invention, the response mode selector is designed to select the most relevant part of the network from multiple domain adaptation networks for activation according to the current task requirements (i.e., a specific response mode). This selective activation mechanism helps to reduce unnecessary computational overhead and improve the processing efficiency of the system. The response mode selector can be a simple gating node or a more complex decision network that makes a selection based on the input feature representation and the information of the response mode.

[0039] Specifically, the computer system obtains a selection signal from the response mode selector according to the current response mode. This selection signal may be a binary vector, in which each element corresponds to a domain adaptation network, and the value of the element (0 or 1) indicates whether the network is selected. For example, suppose there are 10 domain adaptation networks (i.e., the first number is 10), and the selection signal of the current response mode is [1, 0, 1, 0, 1, 0, 0, 1, 0, 0], which means that only the 1st, 3rd, 5th, and 8th networks are selected for activation (i.e., the second number is 4).

[0040] After determining the domain adaptation network to be activated, the computer system fuses the outputs of these networks. The purpose of the fusion operation is to integrate the output feature representations of multiple networks into a unified representation so that subsequent steps can make predictions and decisions based on this unified representation. In an embodiment of the present invention, the fusion operation can be performed in a weighted summation manner. Specifically, for each selected domain adaptation network, a weighted sum is calculated based on its output feature representation and the weight in the corresponding selection signal (1 in the embodiment of the present invention). In this way, the output feature representation of each network will contribute to the final fusion result according to its importance. It can be understood that weighted summation is not the only fusion method. In practical applications, the computer system can also select other fusion methods according to specific needs, such as maximum pooling, average pooling, attention mechanism, etc. These methods have their own advantages and disadvantages and are suitable for different application scenarios and task requirements.

[0041] During the fusion process, the computer system may also consider how to determine the weight of each selected network. In an embodiment of the present invention, the weight may be provided directly by the response mode selector or calculated by other means. For example, an attention mechanism may be used to calculate the attention weight of each network, which reflects the correlation between the network output feature representation and the current response mode. The calculation of the attention weight, for example, involves an additional attention network that outputs the attention score of each network based on the input feature representation and the response mode information. These scores may then be normalized (e.g., using a softmax function) to obtain the final weights.

[0042] After the fusion operation is completed, the computer system will obtain a second encoded feature representation corresponding to the current response mode. This feature representation contains the most relevant feature information extracted from multiple domain adaptation networks for a specific response mode. Compared with the first fused encoded feature representation, the second encoded feature representation is more focused on the needs of the current response mode, and therefore can provide more accurate and valuable information for subsequent prediction and decision-making processes.

[0043] Taking the actual scenario as an example, suppose the virtual power plant system needs to predict whether a certain energy storage device should be started in a specific time period to respond to the needs of the power grid. In this case, the response mode selector can select the most relevant domain adaptation network for activation based on the current load of the power grid, the power status of the energy storage device, and weather forecast information. For example, if the weather forecast shows that high temperature weather is about to occur, then the temperature-related domain adaptation network (such as a network that processes environmental parameters) can be selected; if the power grid load is high and the energy storage device has sufficient power, then the network related to the status of the power generation equipment and the energy storage device can be selected. Then, the computer system fuses the outputs of these selected networks to obtain a second encoded feature representation that comprehensively considers multiple factors. Based on this feature representation, the computer system can more accurately predict whether the energy storage device should be started to respond to the needs of the power grid.

[0044] Step S500: Load the second coded feature representation corresponding to the multiple response modes into the inference network corresponding to the multiple response modes, and obtain the inferred response trigger confidence of the target virtual power plant node to the coordinated control strategy for the multiple response modes.

[0045] In step S400, the computer system has extracted the feature information most relevant to the specific response mode from multiple domain adaptation networks by selective activation and feature fusion, and formed a second encoded feature representation. These feature representations are calculated independently for each response mode, and they contain feature information closely related to the specific response mode extracted from the original virtual power plant node state data.

[0046] Next, these second encoded feature representations are loaded into the inference network corresponding to the multiple response modes. The inference network is a machine learning model used to predict the value of the target variable based on the input feature representation. In an embodiment of the present invention, the task of the inference network is to predict the response trigger confidence of the target virtual power plant node for multiple response modes under a given coordinated control strategy based on the second encoded feature representation. The response trigger confidence is a value between 0 and 1, which represents the probability or confidence level that the virtual power plant node triggers a specific response mode under specific conditions.

[0047] The structure and parameters of the inference network are designed, for example, according to specific application scenarios and task requirements. In an embodiment of the present invention, the inference network is, for example, a multi-layer perceptron (MLP), a convolutional neural network (CNN), a recurrent neural network (RNN), or other types of neural networks. These networks can map the input second encoding feature representation to the output space, i.e., the distribution space of the response trigger confidence, through a series of nonlinear transformations and parameter learning.

[0048] Taking a multi-layer perceptron as an example, the inference network contains multiple hidden layers, each of which is composed of multiple neurons. Each neuron receives the output from the previous layer of neurons as input, and is linearly transformed by weights and biases, and then nonlinearly transformed by activation functions (such as ReLU, sigmoid, or tanh). In this way, the second encoded feature representation of the input is processed by multiple hidden layers, and the final output is the response trigger confidence.

[0049] In actual operation, the computer system loads the second encoded feature representation corresponding to each response mode into the corresponding inference network. Each inference network processes the input features independently and outputs the response trigger confidence for the response mode. In this way, for each response mode, the computer system will obtain a response trigger confidence value, which together constitutes the response behavior prediction result of the target virtual power plant node under a given coordinated control strategy.

[0050] In order to obtain more accurate response trigger confidence prediction results, the inference network needs to go through a lot of debugging and optimization. The debugging process includes the following steps: collect a large amount of historical data as training samples, which should contain the state information of the virtual power plant nodes, the coordinated control strategy and the corresponding response behavior labels; secondly, use these training samples to train the inference network, and optimize the network weights and bias parameters through the back propagation algorithm; finally, evaluate the performance of the inference network through the validation set and the test set, and adjust and optimize the network structure and parameters according to the evaluation results.

[0051] Taking a practical scenario as an example, suppose that the virtual power plant system needs to predict whether a certain energy storage device should be started within a specific time period to respond to the needs of the power grid. In this case, the computer system can define two response methods: starting the energy storage device and not starting the energy storage device. Then, it will extract the second encoded feature representations related to these two response methods through the domain adaptation network and the response method selector based on historical data and current state information. Next, these feature representations are loaded into the inference network corresponding to the response method to obtain the response trigger confidence for each response method. Finally, based on these confidence values ​​and other constraints and objective functions, an optimal coordinated control strategy is formulated, such as deciding whether to start the energy storage device and when to start it.

[0052] As an implementation manner, in step S400, determining to obtain a second plurality of domain adaptation networks from the first plurality of domain adaptation networks includes:

[0053] Step S410: Generate an influence coefficient vector according to the first fused coding feature representation and the first two-dimensional array, wherein the number of columns in the first two-dimensional array is the first number, the number of rows is the dimension of the first fused coding feature representation, and the influence coefficient vector is used to characterize the influence coefficient of the first number domain adaptation network;

[0054] Step S420: determining a second plurality of domain adaptation networks from the first plurality of domain adaptation networks according to the influence coefficient vector.

[0055] In step S410, the computer system generates an influence coefficient vector according to the first fused encoding feature representation and the first two-dimensional array. Each element in the vector represents an influence coefficient of a domain adaptation network, that is, the importance or weight of the network in the current task.

[0056] The first fused coded feature representation is obtained by concatenating multiple first coded feature representations in step S200, and contains comprehensive status information of the virtual power plant nodes. The first two-dimensional array is a predefined matrix whose number of columns is equal to the first number (i.e., the total number of domain adaptation networks) and the number of rows is equal to the dimension of the first fused coded feature representation. This matrix can be regarded as a set of parameters for interacting with the first fused coded feature representation to generate an influence coefficient vector.

[0057] Specifically, the process of generating an influence coefficient vector includes, for example: the computer system first performs a matrix multiplication operation on the first two-dimensional array and the first fused coding feature representation. The result of this operation is a vector, each element of which is the dot product of a column of the first two-dimensional array and the first fused coding feature representation. This vector can be regarded as an initial influence coefficient vector, which reflects the preliminary correlation between each domain adaptation network and the current input feature. For example, suppose there are 5 domain adaptation networks (that is, the first number is 5), and the dimension of the first coding feature representation of each network is 10 (that is, the dimension of the first fused coding feature representation is 10). Then, the first two-dimensional array is a 5×10 matrix. If the first fused coding feature representation is a 10-dimensional vector [x1, x2, ..., x 10 ], then the result of the matrix multiplication is a 5-dimensional vector [y1, y2, ..., y5], where y i is the initial value of the influence coefficient of the i-th domain adaptation network.

[0058] In order to increase the robustness and generalization ability of the system, the computer system can introduce some perturbations. In this step, it uses another two-dimensional array (called the second two-dimensional array) to perform matrix multiplication with the first fusion encoding feature representation to generate a perturbation base array. This array has the same dimension as the first two-dimensional array, but it contains some random perturbations or noise, which help the system avoid overfitting during training and improve adaptability to different input features. Next, the computer system generates an influence coefficient adjustment array based on the dynamic adjustment array (a random perturbation array) and the perturbation base array. The elements in the dynamic adjustment array are random numbers that follow a Gaussian distribution or other distributions, which are used to further adjust the perturbation base array to generate the final influence coefficient adjustment array. This adjustment array reflects the dynamic response and adaptability of the system to the current input features. Finally, the computer system combines the initial influence coefficient vector with the influence coefficient adjustment array (such as addition or weighted summation) to generate the final influence coefficient vector. Each element in this vector represents the final influence coefficient or weight of a domain adaptation network in the current task.

[0059] In step S420, the computer system determines which domain adaptation networks will be activated based on the influence coefficient vector generated in step S410. Specifically, it selects the top N networks with the largest influence coefficients for activation based on the element values ​​in the influence coefficient vector, where N is the second number (i.e., the number of networks to be activated).

[0060] For example, the computer system sorts the elements in the influence coefficient vector from large to small. In this way, the first element represents the domain adaptation network with the largest influence coefficient. Next, the domain adaptation networks corresponding to the first N elements in the sorted influence coefficient vector are selected for activation. These networks are considered to be the most relevant to the current task and the most valuable information. Finally, the selected domain adaptation networks are marked as activated, and their output feature representations are passed to subsequent steps for further processing. The unselected networks remain inactive, and their outputs will not be considered or used in subsequent calculations.

[0061] Taking the actual scenario as an example, suppose that the virtual power plant system needs to predict whether a certain energy storage device should be started within a specific time period to respond to the needs of the power grid. In this case, the computer system can generate a first fusion coding feature representation based on historical data and current status information. Then, it will generate an influence coefficient vector through the process in step S410, which reflects the importance of different field adaptation networks (such as networks that process electricity data, networks that process environmental parameters, networks that process the status of power generation equipment, etc.) in the current task. Finally, in step S420, several networks with the largest influence coefficients are selected for activation. These networks will predict the response trigger confidence of the energy storage device startup based on the input features. In this way, the system can more accurately identify the feature information most relevant to the current task, thereby improving the accuracy and efficiency of the prediction.

[0062] As an implementation manner, step S410, generating an influence coefficient vector according to the first fusion coding feature representation and the first two-dimensional array, includes:

[0063] Step S411: multiply the first two-dimensional array and the first fusion coding feature representation to obtain an initial influence coefficient vector;

[0064] Step S412: The second two-dimensional array is multiplied by the first fused coding feature representation to obtain a perturbation base array, wherein the number of columns in the second two-dimensional array is the first number, and the number of rows is the dimension of the first fused coding feature representation;

[0065] Step S413: generating an influence coefficient adjustment array according to the dynamic adjustment array and the disturbance basic array;

[0066] Step S414: Generate an influence coefficient vector according to the initial influence coefficient vector and the influence coefficient adjustment array.

[0067] In step S411, the computer system performs a matrix multiplication operation on the first two-dimensional array and the first fused coding feature representation to obtain an initial influence coefficient vector. Each element in this vector represents an influence coefficient or weight of a domain adaptation network in an initial state.

[0068] The first two-dimensional array is a predefined matrix whose number of columns is equal to the total number of domain adaptation networks (i.e., the first number), and the number of rows is equal to the dimension of the first fused coding feature representation. This matrix can be regarded as a set of parameters for interacting with the input features to generate influence coefficients. The first fused coding feature representation is obtained by concatenating multiple first coding feature representations in step S200, which contains comprehensive status information of the virtual power plant nodes. The specific process of matrix multiplication operation is as follows: Assume that the first two-dimensional array is A, whose dimension is m×n (m is the total number of domain adaptation networks, and n is the dimension of the first fused coding feature representation); the first fused coding feature representation is x, whose dimension is n×1. Then, the result of the matrix multiplication y=Ax is an m×1 vector, that is, the initial influence coefficient vector. Each element y in this vector i (i=1,2,...,m) are the dot products of the i-th row of A and x, that is, y i =Σ(A ij *x j )(j=1,2,...,n), where A ij Represents the element in the i-th row and j-th column of matrix A.

[0069] Taking the actual scenario as an example, suppose there are 5 domain adaptation networks (m=5), and the first encoding feature representation dimension of each network is 10 (n=10). Then, the first two-dimensional array A is a 5×10 matrix, and the first fusion encoding feature representation x is a 10x1 vector. Through matrix multiplication, a 5×1 initial influence coefficient vector y will be obtained, in which each element represents the influence coefficient of a domain adaptation network in its initial state.

[0070] In step S412, the computer system performs a matrix multiplication operation on another two-dimensional array (i.e., the second two-dimensional array) and the first fused coding feature representation to obtain a perturbation base array. This array will be used to generate the influence coefficient adjustment array in the subsequent steps, thereby increasing the robustness and generalization ability of the system. The second two-dimensional array has the same dimension as the first two-dimensional array, i.e., m×n. However, the element values ​​in the second two-dimensional array may be generated randomly or according to a specific rule. These random or specific element values ​​will be used as perturbations to adjust the initial influence coefficient vector in the subsequent steps.

[0071] The process of matrix multiplication is similar to step S411. Assuming that the second two-dimensional array is B, the result of the matrix multiplication z=Bx is an m×1 vector, namely the perturbation basis array. Each element z in this vector i (i=1,2,...,m) are the dot products of the i-th row of B and x, i.e. z i =Σ(B ij* x j )(j=1,2,...,n), where B ij Represents the element in the i-th row and j-th column of matrix B.

[0072] Since the element values ​​in the second two-dimensional array may be random, the perturbation base array obtained each time step S412 is performed may be different. This randomness helps the system avoid overfitting during training and improves adaptability to different input features.

[0073] In step S413, the computer system will use a dynamic adjustment array (a random perturbation array) to interact with the perturbation base array obtained in step S412 to generate an influence coefficient adjustment array. This adjustment array will be used to further adjust the initial influence coefficient vector in subsequent steps.

[0074] The dynamic adjustment array is a randomly generated array with the same dimension as the perturbation base array, i.e., m×1. The element values ​​in this array follow a certain probability distribution (such as Gaussian distribution) to introduce additional randomness. By interacting with the perturbation base array (such as element-by-element multiplication or weighted summation), the dynamic adjustment array is able to generate an influence coefficient adjustment array that reflects the system's dynamic response and adaptability to the current input features.

[0075] The specific interaction method may vary depending on the implementation. For example, one possible method is to multiply the dynamic adjustment array by the perturbation base array element by element, that is, to obtain the adjustment array r = z * w, where w is the dynamic adjustment array and z is the perturbation base array. Another possible method is to use the dynamic adjustment array as a weight and perform a weighted summation on the perturbation base array, that is, to obtain the adjustment array r = Σ(w i * z i )(i=1,2,...,m), where w i is the i-th element of the dynamically adjusted array.

[0076] Regardless of the interaction method used, the purpose of the influence coefficient adjustment array is to introduce some random perturbations or dynamic adjustments based on the initial influence coefficient vector, thereby increasing the robustness and generalization ability of the system.

[0077] Step S414 is the last step of generating the influence coefficient vector. In this step, the computer system combines (such as adding or weighted summing) the initial influence coefficient vector obtained in step S411 with the influence coefficient adjustment array obtained in step S413 to generate the final influence coefficient vector.

[0078] The specific combination method may vary depending on the implementation method. For example, one possible method is to add the initial influence coefficient vector and the influence coefficient adjustment array element by element, that is, to obtain the final influence coefficient vector v=y+r, where y is the initial influence coefficient vector and r is the influence coefficient adjustment array. Another possible method is to use the influence coefficient adjustment array as a weight to perform weighted summation on the initial influence coefficient vector, that is, to obtain the final influence coefficient vector v=Σ(r i * y i )(i=1,2,...,m), where ri is the i-th element of the influence coefficient adjustment array, y i is the i-th element of the initial influence coefficient vector.

[0079] Regardless of the combination method, the final influence coefficient vector v will be used in the subsequent steps to determine which domain adaptation networks will be activated. Each element in this vector represents the final influence coefficient or weight of a domain adaptation network in the current task, which reflects the importance or contribution of the network to the current input feature.

[0080] Step S410 aims to improve the robustness and generalization ability of the system through matrix multiplication, random perturbation, dynamic adjustment, etc. Through the step-by-step processing of steps S411 to S414, the computer system can start from an initial, possibly inaccurate influence coefficient vector, and gradually adjust and optimize to obtain the final influence coefficient vector, thereby providing strong support for the selection of domain adaptation networks in subsequent steps.

[0081] As an implementation mode, step S413 generates an influence coefficient adjustment array according to the dynamic adjustment array and the disturbance basic array, including:

[0082] Step S4131: obtaining a set of variant values, wherein the variant values ​​in the set of variant values ​​follow a Gaussian distribution;

[0083] Step S4132: randomly selecting the first number of variant values ​​in the variant value set to construct a dynamic adjustment array;

[0084] Step S4133: perform a dot product calculation on the dynamic adjustment array and the disturbance basic array to obtain an influence coefficient adjustment array.

[0085] In step S4131, the computer system generates a set of variant values, each element of which is a random number following a Gaussian distribution, and is referred to as a variant value.

[0086] In actual operation, the computer system can use a random number generator to generate these variant values. In an embodiment of the present invention, the computer system will configure the random number generator to generate random numbers that follow a Gaussian distribution. The mean and standard deviation of these random numbers can be set according to specific needs to control the distribution range and concentration of the variant values.

[0087] For example, suppose you need to generate a set of 100 variant values ​​that follow a Gaussian distribution with a mean of 0 and a standard deviation of 1. Then, the computer system will call the random number generator to generate these 100 random numbers, each of which will independently follow the specified Gaussian distribution. In this way, you will get a set of 100 variant values ​​that will be used to construct a dynamically adjusted array in subsequent steps.

[0088] In step S4132, the computer system will arbitrarily select the first number of variation values ​​(i.e., the total number of domain adaptation networks) from the variation value set obtained in step S4131 to construct a dynamic adjustment array. This array will be used in subsequent steps to interact with the perturbation base array to generate an influence coefficient adjustment array.

[0089] In actual operation, the computer system can use a random sampling method to select variant values ​​from the variant value set. Random sampling is a statistical method used to randomly select a portion of samples from a population for inference or analysis. In an embodiment of the present invention, the computer system will randomly select variant values ​​from the variant value set that are equal to the total number of domain adaptation networks to construct a dynamic adjustment array.

[0090] For example, assuming that there are 5 domain adaptation networks (i.e., the first number is 5), and the set of variant values ​​generated in step S4131 contains 100 values. Then, the computer system will randomly select 5 values ​​from these 100 values ​​to construct a dynamic adjustment array. These 5 values ​​will correspond to 5 domain adaptation networks respectively, and will be used to adjust their influence coefficients in subsequent steps.

[0091] It should be noted that since the variation value is randomly generated, the dynamic adjustment array obtained each time step S4132 is executed may be different. This randomness helps the system avoid overfitting during training and improves adaptability to different input features.

[0092] In step S4133, the computer system performs a dot product calculation (also called a dot product calculation or an inner product calculation) on the dynamic adjustment array constructed in step S4132 and the disturbance basic array obtained in step S412 to generate an influence coefficient adjustment array.

[0093] The scalar product calculation is a linear algebraic operation used to calculate the sum of the products of the corresponding elements of two vectors. In the embodiment of the present invention, the dynamic adjustment array and the perturbation base array are both one-dimensional vectors, and their lengths are equal (i.e., the total number of domain adaptation networks). By calculating the scalar product of these two vectors, a scalar value can be obtained, which reflects the overall influence of the dynamic adjustment array on the perturbation base array.

[0094] In the embodiment of the present invention, instead of directly calculating the dot product of the dynamic adjustment array and the perturbation basic array to obtain a scalar value, it is desired to obtain an influence coefficient adjustment array with the same dimension as the perturbation basic array. To achieve this goal, the dynamic adjustment array can be regarded as a weight vector, and it is multiplied element by element (or weighted summed) with the perturbation basic array to obtain the influence coefficient adjustment array.

[0095] Specifically, suppose the dynamically adjusted array is w=[w1, w2, ..., w m ], the perturbation basis array is z=[z1, z2,..., z m ], where m is the total number of domain adaptation networks. Then, the influence coefficient adjustment array r can be calculated by the following formula:

[0096] r=w⊙z.

[0097] Among them, ⊙ represents the element-by-element multiplication operator. The influence coefficient adjustment array r=[r1, r2,..., r m ], where r i =w i * z i (i=1,2,...,m).

[0098] Since the elements in the dynamic adjustment array are randomly generated variant values, the influence coefficient adjustment array obtained each time step S4133 is executed may be different. This randomness enables the system to continuously explore different influence coefficient combinations during the training process to find the optimal solution.

[0099] In summary, step S413 successfully generates the influence coefficient adjustment array by obtaining the set of variant values, constructing the dynamic adjustment array, and performing scalar product calculation on the dynamic adjustment array and the disturbance basic array. This process not only introduces randomness and dynamic adjustment, but also enhances the robustness and generalization ability of the system. By continuously iterating and optimizing these steps, the computer system can gradually improve the accuracy and efficiency of energy coordinated optimization control.

[0100] As an implementation manner, step S420, determining a second plurality of domain adaptation networks from the first plurality of domain adaptation networks according to the influence coefficient vector, includes:

[0101] Step S421: determining the first second several vector units in the vector units in the influence coefficient vector arranged in descending order;

[0102] Step S422: The domain adaptation network corresponding to the first second number of vector units is used as the determined second number of domain adaptation networks.

[0103] In step S421, the computer system sorts each vector unit in the influence coefficient vector (ie, the influence coefficient of each domain adaptation network). The purpose of sorting is to identify which domain adaptation networks have the highest importance or contribution in the current task.

[0104] Specifically, the computer system arranges the vector units in the influence coefficient vector in descending order according to their values. Descending order means that the values ​​of the vector units are arranged in order from large to small, so that the vector unit in the front represents the domain adaptation network with the largest influence coefficient.

[0105] Assume that there are 5 domain adaptation networks, and the corresponding influence coefficient vector is [0.8, 0.6, 0.9, 0.5, 0.7]. In step S421, the computer system will sort this vector in descending order to obtain a new vector [0.9, 0.8, 0.7, 0.6, 0.5]. This new vector clearly shows the importance ranking of each domain adaptation network in the current task.

[0106] Next, the computer system determines the number of domain adaptation networks that need to be activated, which is the second number. This number can be set according to the actual application scenario and requirements. For example, if you want to activate only the two most important domain adaptation networks, then the second number is 2.

[0107] Based on the sorted influence coefficient vector and the determined second number, the computer system will select the first N (N equals the second number) vector units in the sorted vector. Continuing with the above example, if it is decided to activate two networks, the computer system will select the first two units in the sorted vector, that is, [0.9, 0.8].

[0108] In step S422, the computer system associates the first N vector units determined in step S421 with the domain adaptation network to determine which network will be activated.

[0109] Specifically, the position of each vector unit in the influence coefficient vector corresponds to a specific domain adaptation network. For example, in the sorted influence coefficient vector [0.9, 0.8, 0.7, 0.6, 0.5], the first vector unit 0.9 corresponds to the first domain adaptation network, the second vector unit 0.8 corresponds to the second domain adaptation network, and so on.

[0110] Therefore, in step S422, the computer system will find the corresponding domain adaptation network according to the position of the first N vector units determined in step S421, and mark these networks as activated. Continuing with the above example, if it is decided to activate two networks, the computer system will mark the first and second domain adaptation networks as activated.

[0111] In this way, through the combined effect of steps S421 and S422, the computer system successfully determines the domain adaptation networks that need to be activated in the current task. These activated networks will play an important role in the subsequent energy coordination optimization control process based on their high influence coefficients.

[0112] This network selection method based on the influence coefficient vector is highly flexible and adaptive. In practical applications, the second number (i.e., the number of networks to be activated) can be dynamically adjusted according to the task requirements and the system resource conditions to balance the computational efficiency and optimization effect. For example, when resources are sufficient, more networks can be activated to obtain more accurate optimization results; when resources are limited, fewer networks can be activated to save computing resources.

[0113] In addition, the generation process of the influence coefficient vector (as described in step S410) also ensures the robustness and generalization ability of this network selection method. By introducing random perturbations and dynamic adjustment mechanisms, the computer system can generate different influence coefficient vectors under different input features, thereby achieving flexible selection and activation of networks adapted to different fields. This mechanism helps to improve the adaptability and stability of the system in different scenarios.

[0114] As an implementation mode, in step S400, the outputs of the second plurality of domain adaptation networks determined are fused, including:

[0115] Step S430: replacing the remaining vector units except the second number of vector units arranged in descending order in the influence coefficient vector with infinitesimal values;

[0116] Step S440: performing a standardized exponential conversion on the influence coefficient vector to obtain a first influence coefficient of a second domain adaptation network;

[0117] Step S450: fusing the outputs of the second plurality of domain adaptation networks determined according to the first influence coefficient.

[0118] In step S430 , the computer system replaces the inactivated vector units in the influence coefficient vector (ie, the remaining vector units other than the second several vector units arranged in descending order) with infinitesimal values ​​(eg, a very small value that is very close to 0 but not equal to 0).

[0119] The purpose of this step is to reduce the influence of inactivated networks on the fusion results. Since inactivated networks are less important in the current task, their outputs may contain noise or irrelevant information, which can interfere with the fusion results. By setting the influence coefficients of inactivated networks to infinitesimal values, the computer system can effectively ignore the outputs of these networks during the fusion process, thereby ensuring the accuracy and reliability of the fusion results.

[0120] For example, assuming there are five domain adaptation networks, the corresponding influence coefficient vector is [0.9, 0.8, 0.7, 0.5, 0.4], and it is decided to activate the first two networks. In step S430, the computer system replaces the last three vector units in this vector (i.e., 0.5, 0.4 and the remaining vector units not shown) with infinitesimal values ​​to obtain a new vector [0.9, 0.8, ε, ε, ε], where ε represents an infinitesimal value.

[0121] Step S440 is the second step of the fusion process, and its purpose is to normalize the influence coefficient vector to obtain the first influence coefficient of the activated network. Normalized exponential transformation (also known as softmax function) is a normalization method that can convert a real number vector into a probability distribution vector so that the sum of all elements of the vector is 1 and each element is between 0 and 1. In this step, the computer system will apply the softmax function to the influence coefficient vector processed in step S430 to obtain the first influence coefficient of the activated network. The calculation formula of the softmax function is as follows:

[0122] ;

[0123] Among them, z i represents the i-th element in the influence coefficient vector, and m represents the length of the influence coefficient vector (i.e., the total number of domain adaptation networks). By applying the softmax function, the computer system is able to convert the influence coefficient vector into a probability distribution vector, in which each element represents the weight or influence of the corresponding domain adaptation network in the fusion process.

[0124] Continuing with the above example, assume that the influence coefficient vector has been processed to [0.9, 0.8, ε, ε, ε], where ε is an infinitesimal value very close to 0. In step S440, the computer system will apply the softmax function to this vector and obtain a new vector [0.525, 0.475, close to 0, close to 0, close to 0]. This new vector is the first influence coefficient vector of the activated network, which represents the weight or influence of each activated network in the fusion process.

[0125] The purpose of step S450 is to perform weighted summation on the output of the activated network according to the first influence coefficient to obtain the final fusion result. In this step, the computer system will first obtain the output feature representation of the activated network, and then perform weighted summation on these feature representations according to the first influence coefficient.

[0126] Specifically, assuming there are n activated networks (i.e., the second number is n), the output feature of each network is represented as h i (i=1,2,...,n), the corresponding first influence coefficient is w i (i=1,2,...,n). Then, the final fusion result H can be calculated by the following formula:

[0127] ;

[0128] Among them, w i represents the first influence coefficient of the i-th activated network, h i Represents the output feature representation of the i-th activated network. Through weighted summation, the computer system can integrate the output feature representations of multiple activated networks into a unified representation, which contains useful information from all activated networks, and the information from different networks is weighted according to their importance.

[0129] Continuing with the above example, assume that the first influence coefficient vectors [0.525, 0.475] of the two activation networks and their corresponding output feature representations h1 and h2 have been obtained. In step S450, the computer system will perform a weighted summation of the output feature representations according to these first influence coefficients to obtain the final fusion result H= This fusion result will be used as the input of the subsequent steps to predict the response trigger confidence of the target virtual power plant node to multiple response modes under a given coordinated control strategy.

[0130] As an implementation method, the domain adaptation network includes a two-dimensional array of network weights and an offset, the number of columns in the two-dimensional array of network weights is the dimension of the output of the domain adaptation network, the number of rows is the dimension represented by the first fused coding feature, and the dimension of the offset is equal to the dimension of the output of the domain adaptation network. Based on this, step S450, according to the first influence coefficient, the outputs of the second number of domain adaptation networks determined are fused, including:

[0131] Step S451: obtaining a multi-layer weight set, where the multi-layer weight set represents a two-dimensional array of network weights of the first domain adaptation network;

[0132] Step S452: obtaining an offset matrix, the offset matrix representing the offsets of the first several domain adaptation networks;

[0133] Step S453: in the multi-layer weight set, obtaining a two-dimensional array of network weights corresponding to the second number of first influence coefficients;

[0134] Step S454: obtaining the offsets corresponding to the second number of first influence coefficients in the offset matrix;

[0135] Step S455: Determine a fusion result according to the second number of first influence coefficients, the two-dimensional array of network weights corresponding to the second number of first influence coefficients, and the offsets corresponding to the second number of first influence coefficients.

[0136] In step S450, the computer system determines how to fuse the output of the activated network according to the first influence coefficient (i.e., the first influence coefficient of the activated network obtained in step S440). This process involves the use of the two-dimensional array of network weights and offsets, and their interaction with the first influence coefficient.

[0137] The purpose of step S451 is to obtain a multi-layer weight set representing the two-dimensional array of network weights of all domain adaptation networks. The multi-layer weight set is a three-dimensional array (or tensor), whose dimensions are represented as [m, n, p], for example, where m represents the total number of domain adaptation networks, n represents the number of rows of the two-dimensional array of network weights (i.e., the dimension represented by the first fusion encoding feature), and p represents the number of columns of the two-dimensional array of network weights (i.e., the dimension of the domain adaptation network output).

[0138] In practice, a computer system can retrieve a collection of multi-layer weights from the model’s parameter storage. These parameters are learned during the training of the domain adaptation network and define how the network transforms input features into output features.

[0139] For example, suppose there are five domain adaptation networks, and the two-dimensional array of network weights for each network is a 10×5 matrix (i.e., the input feature dimension is 10 and the output feature dimension is 5). Then, the multi-layer weight set is a 5×10×5 three-dimensional array that contains the network weights of all networks.

[0140] The purpose of step S452 is to obtain an offset matrix representing the offsets of all domain adaptation networks. The offset matrix is ​​a two-dimensional array, and its dimension is represented as [m, p], for example, where m represents the total number of domain adaptation networks and p represents the dimension of the offset (i.e., the dimension of the domain adaptation network output). The offset plays a role in adjusting the output features in the network, which enables the network to adapt to different input features more flexibly.

[0141] Similar to the multi-layer weight set, the offset matrix is ​​also retrieved from the parameter storage of the model. In actual operation, the computer system can store the offset matrix separately from the multi-layer weight set so that they can be retrieved separately when needed.

[0142] Continuing with the above example, assume that the offset of each domain adaptation network is a 5-dimensional vector. Then, the offset matrix is ​​a 5×5 2D array containing the offsets of all networks.

[0143] The purpose of step S453 is to retrieve the two-dimensional array of network weights corresponding to the activated network from the multi-layer weight set. This step depends on the first influence coefficient obtained in step S440 and the activated network determined in step S420.

[0144] Specifically, the computer system determines which domain adaptation networks are activated based on the size of the first influence coefficient, and finds the two-dimensional arrays of network weights of these networks in the multi-layer weight set. Since the first influence coefficients have been normalized by the softmax function, they can be directly used as indices or weights to select the two-dimensional array of network weights.

[0145] For example, assuming that the first two domain adaptation networks have been determined to be activated and their first influence coefficients are 0.6 and 0.4 respectively, then the computer system will select the first two two-dimensional arrays of network weights (i.e., the first two "slices" of the multi-layer weight set) in the multi-layer weight set for subsequent processing.

[0146] The purpose of step S454 is to retrieve the offset corresponding to the activated network from the offset matrix. This step is similar to step S453, and also determines the offset to be retrieved based on the first influence coefficient and the activated network.

[0147] Specifically, the computer system finds the corresponding offset vector in the offset matrix based on the information of the activated network. Since the dimension of the offset matrix matches the total number of domain adaptation networks, the required offset can be easily retrieved by indexing.

[0148] Continuing with the above example, assuming that the first two domain adaptation networks have been determined to be activated, the computer system will select the first two offset vectors (i.e., the first two rows of the offset matrix) in the offset matrix for subsequent processing.

[0149] The purpose of step S455 is to determine the final fusion result based on the two-dimensional array of network weights of the activated network, the offset, and the first influence coefficient. Specifically, the computer system uses the two-dimensional array of network weights of the activated network and the first fusion coding feature representation to perform matrix multiplication operations to obtain a preliminary output feature representation. Then, it adds the preliminary output feature representation to the corresponding offset to adjust the value of the output feature. Finally, it performs weighted summation on the adjusted output feature representation according to the first influence coefficient to obtain the final fusion result. For example, assume that the two-dimensional arrays of network weights W1 and W2, offsets b1 and b2, and their first influence coefficients of 0.6 and 0.4 of the first two activated networks have been obtained. Assume that the first fusion coding feature representation is x. Then, the fusion process can be expressed as the following steps:

[0150] Compute the preliminary output feature representation: and .

[0151] The preliminary output feature representation is weighted summed according to the first influence coefficient: .

[0152] The final H is the fusion result, which contains the useful information of all activated networks and is weighted according to their importance. This fusion result will be used as the input of the subsequent steps to predict the response trigger confidence of the target virtual power plant node to multiple response modes under a given coordinated control strategy.

[0153] As an implementation mode, step S455 determines the fusion result according to the second number of first influence coefficients, the two-dimensional array of network weights corresponding to the second number of first influence coefficients, and the offset corresponding to the second number of first influence coefficients, including:

[0154] Step S4551: multiplying the two-dimensional array of network weights corresponding to each of the second plurality of domain adaptation networks by the first fused encoding feature representation loaded into the domain adaptation network to obtain a transition feature representation;

[0155] Step S4552: Add the transition feature representation and the offset corresponding to the domain adaptation network to obtain an added feature representation;

[0156] Step S4553: performing nonlinear transformation on the added feature representation to obtain the output of the domain adaptation network;

[0157] Step S4554: Obtain a fusion result based on the outputs of the second plurality of domain adaptation networks and the first influence coefficient corresponding to the domain adaptation network.

[0158] The purpose of step S4551 is to calculate the transition feature representation of each activated domain adaptation network. The transition feature representation is obtained by performing a matrix multiplication operation on the two-dimensional array of network weights and the first fused encoding feature representation loaded into the domain adaptation network.

[0159] Specifically, suppose there are n activated domain adaptation networks (i.e., the second number is n), and the two-dimensional array of network weights of each network is W i (i=1,2,...,n), the first fusion encoding feature is represented as x. Then, for the i-th activated network, its transition feature is represented as h i It can be calculated by the following formula:

[0160] Here, · represents a matrix multiplication operation. The result of this operation is a vector whose dimension is the same as the number of columns of the two-dimensional array of network weights (i.e., the output dimension of the domain adaptation network). The transition feature representation h i Contains the feature information related to the i-th activated network extracted from the first fused encoded feature representation x.

[0161] For example, suppose there are two activated domain adaptation networks, and their network weight 2D arrays are , the first fusion encoding feature is expressed as (m is the dimension of the first fusion encoding feature representation). Then, for the first activated network, its transition feature representation h1 can be calculated by For the second activated network, its transition feature representation h2 can be calculated by get.

[0162] The purpose of step S4552 is to add the transition feature representation to the corresponding offset to obtain the added feature representation. The offset is a vector whose dimension is the same as the output dimension of the domain adaptation network, which is used to adjust the value of the transition feature representation to make it more suitable for subsequent nonlinear transformations.

[0163] Specifically, for the i-th activated network, its added feature representation h' i It can be calculated by the following formula: .

[0164] Among them, h i is the transition feature representation of the i-th activated network, b i is the corresponding offset. The added feature represents h' i is the transition characteristic h i It is an adjustment that enables the network to adapt more flexibly to different input features.

[0165] Continuing with the above example, assuming that the transition feature representations h1 and h2 of the two activated networks and their corresponding offsets b1 and b2 have been obtained, then for the first activated network, its added feature representation h'1 can be obtained by calculating h1+b1; for the second activated network, its added feature representation h'2 can be obtained by calculating h2+b2.

[0166] The purpose of step S4553 is to perform a nonlinear transformation on the added feature representation to obtain the output of the domain adaptation network. Nonlinear transformation can map the input features to a higher dimensional space and introduce nonlinear relationships, thereby enhancing the expressive power of the model. In the domain adaptation network, nonlinear transformation is achieved, for example, through an activation function. An activation function is a function that maps input to output, which can introduce nonlinear factors so that the network can learn complex patterns. Common activation functions include ReLU, sigmoid, tanh, etc.

[0167] Specifically, for the i-th activated network, its output o i It can be calculated by the following formula:

[0168] o i =f(h' i );

[0169] Among them, h' i is the additive feature representation of the i-th activated network, and f is the activation function. Output o i is a vector with the same dimension as the output dimension of the domain adaptation network.

[0170] Continuing with the above example, assuming that ReLU is selected as the activation function, then for the first activated network, its output o1 can be obtained by calculating ReLU(h'1); for the second activated network, its output o2 can be obtained by calculating ReLU(h'2).

[0171] The purpose of step S4554 is to calculate the final fusion result according to the output of the activated network and the corresponding first influence coefficient. This step is implemented by weighted summation, where the weight is the first influence coefficient of the activated network.

[0172] Specifically, the final fusion result H can be calculated by the following formula:

[0173] ;

[0174] Where n is the number of activated networks, w i is the first influence coefficient of the i-th activated network, o i is the output of the i-th activated network. The fusion result H is a vector whose dimension is the same as the output dimension of the domain adaptation network. It contains the useful information of all activated networks and is weighted according to their importance.

[0175] Continuing with the above example, assume that we have obtained the outputs o1 and o2 of the two activated networks, as well as their corresponding first influence coefficients w1 and w2. Then, the final fusion result H can be calculated by This fusion result will be used as the input of the subsequent steps to predict the response trigger confidence of the target virtual power plant node to multiple response modes under a given coordinated control strategy.

[0176] In summary, step S455 achieves effective fusion of the outputs of multiple activation domain adaptation networks by calculating transition feature representation, adding feature representation, performing nonlinear transformation, and weighted summation. This process not only takes into account the influence of network weights and offsets, but also weights the output of the activation network through the first influence coefficient, thereby ensuring the accuracy and reliability of the fusion result. By continuously iterating and optimizing this process, the computer system can gradually improve the accuracy and efficiency of energy coordination optimization control.

[0177] As an implementation method, the process of determining the second number includes:

[0178] Step S1: Obtain the response method type of the response method;

[0179] Step S2: obtaining the inference target accuracy of the response mode;

[0180] Step S3: Determine a second number based on the response mode type and the inference target accuracy.

[0181] The purpose of step S1 is to obtain the type of the current response mode. The response mode type is a classification or identification of different response modes, which reflects the basic attributes and characteristics of the response mode. In the virtual power plant system, the response mode may include adjusting the output power of the power generation equipment, controlling the charging and discharging behavior of the energy storage equipment, adjusting the power consumption of the adjustable load, and other types.

[0182] In actual operation, the computer system can retrieve the type information of the current response method from a preset response method type library. This type library is a database or data structure containing multiple response methods and their corresponding types, which allows the computer system to quickly find the type of the response method according to its name or identifier.

[0183] For example, if there is a response method called "starting energy storage equipment", the computer system can search for this response method in the response method type library and determine that its type is "energy storage control". This type information will provide an important reference for subsequent steps.

[0184] The purpose of step S2 is to obtain the inference target accuracy of the current response mode. The inference target accuracy refers to the accuracy requirement for the result obtained by inference, which reflects the system's expectation and tolerance for the accuracy of the optimization result. In a virtual power plant system, different response modes may have different requirements for the accuracy of the inference results. For example, for the control response of key equipment, the system may require very high accuracy to ensure the safe operation of the equipment; while for some non-critical responses, the system may have relatively low accuracy requirements.

[0185] In actual operation, the computer system can retrieve the accuracy information of the current response method from a preset inference target accuracy library. This accuracy library is a database or data structure containing multiple response methods and their corresponding accuracy, which allows the computer system to quickly find its accuracy requirement according to the name or identifier of the response method.

[0186] For example, if we consider the response method of "starting the energy storage device" again, the computer system can look up this response method in the inference target accuracy library and determine that its accuracy requirement is 95%. This means that the system expects the inference result to be correct or close to correct with a probability of 95%.

[0187] In step S3, the computer system will comprehensively consider the two factors of response mode type and inference target accuracy to determine the number of domain adaptation networks that need to be activated (i.e., the second number). This process involves, for example, a series of judgments and calculations to ensure that the number of networks selected can meet the requirements of inference accuracy while minimizing computing costs and resource consumption.

[0188] In actual operation, the computer system can use a method based on scoring and mapping to determine the second number. Specifically, the system searches for the corresponding response method impact score from a scoring table based on the response method type. This scoring table is a database or data structure containing multiple response method types and their corresponding scores, which reflects the degree of impact of different response methods on the overall performance of the system. The higher the score, the greater the impact of the response method on the system performance, so more networks are needed to process it.

[0189] Next, the corresponding inference accuracy score is found from another scoring table based on the inference target accuracy. This scoring table is a database or data structure that contains multiple accuracy requirements and their corresponding scores, which reflects the degree of demand for system resources for different accuracy requirements. The higher the score, the higher the requirement for the accuracy of the inference result, so more networks are needed to improve the accuracy.

[0190] Then, the mode impact score and the inference accuracy score are combined to obtain a comprehensive score. This comprehensive score reflects the degree of demand for system performance and resources of the current response mode. Finally, according to the comprehensive score, the corresponding number of networks (i.e., the second number) is found from a score-number mapping relationship. This mapping relationship is a database or data structure containing multiple score intervals and their corresponding network numbers, which allows the system to determine the number of networks that need to be activated based on the comprehensive score. For example, the higher the score, the more corresponding networks are to ensure that the system performance and resource requirements can be met.

[0191] For example, let's consider the response mode "starting the energy storage device" again, with a mode impact score of 8 (indicating a large impact on system performance) and a reasoning accuracy score of 9 (indicating a high requirement for the reasoning result). The system can combine these two scores to get a comprehensive score of 17. Then, the system will look for the number of networks corresponding to a score of 17 in the score-number mapping relationship, assuming it is 3. Therefore, the system will eventually determine that 3 domain adaptation networks need to be activated to handle this response mode.

[0192] In summary, the process of determining the second number is a complex decision-making process that comprehensively considers the response mode type and the accuracy of the inference target. Through this process, the computer system can flexibly select the number of networks to be activated according to the characteristics and requirements of the current response mode to ensure the accuracy and reliability of the optimization results while minimizing the computing cost and resource consumption.

[0193] As an implementation mode, step S3, determining the second number according to the response mode type and the inference target accuracy, includes:

[0194] Step S31: Determine the method impact score according to the response method type;

[0195] Step S32: Determine the reasoning accuracy score according to the reasoning target accuracy;

[0196] Step S33: determining a combined score according to the mode impact score and the reasoning accuracy score;

[0197] Step S34: Determine a second number based on the combined score.

[0198] The purpose of step S31 is to evaluate the impact of the response mode on the overall system performance according to the response mode type and give a mode impact score. This score reflects the importance and complexity of the response mode in the optimization control process.

[0199] In actual operation, the computer system maintains a table of response mode types and mode impact scores. This table is a preset data structure that contains multiple response mode types and their corresponding mode impact scores. The scores are obtained based on the knowledge and experience of domain experts and simulation analysis of system performance.

[0200] For example, suppose there are the following response method types and their corresponding method impact scores:

[0201] Adjustment of power generation equipment output power: 8 points (indicating a significant impact on system performance);

[0202] Controlling the charging and discharging of energy storage equipment: 7 points (indicating a certain impact on system performance);

[0203] Adjusting the power consumption of adjustable loads: 6 points (indicating a small impact on system performance);

[0204] Now, assuming that the current response method is "adjust the output power of the power generation equipment", the computer system will look up the response method type in the table and determine that its method impact score is 8 points.

[0205] The purpose of step S32 is to evaluate the system's expectations and requirements for the accuracy of the inference results based on the inference target accuracy and give an inference accuracy score, which reflects the system's tolerance and expectation level for the accuracy of the optimization results.

[0206] Similar to step S31, the computer system also maintains a corresponding table of inference target accuracy and inference accuracy score. This table contains multiple inference target accuracy requirements and their corresponding inference accuracy scores. The scores are also based on the knowledge and experience of domain experts and the demand analysis of system performance.

[0207] For example, suppose there are the following inference target accuracy requirements and their corresponding inference accuracy scores:

[0208] High accuracy (≥95%): 9 points (indicating very high requirements for the inference results);

[0209] Medium accuracy (90%-95%): 7 points (indicating medium requirements for the inference results);

[0210] Low accuracy (≤90%): 5 points (indicating lower requirements for inference results);

[0211] Now, assuming that the inference target accuracy requirement for the current response method is ≥95%, the computer system will look up the accuracy requirement in the table and determine that its inference accuracy score is 9 points.

[0212] The purpose of step S33 is to combine the mode impact score and the reasoning accuracy score to obtain a combined score that comprehensively reflects the importance of the response mode and the reasoning accuracy requirement. This score will provide a basis for the subsequent determination of the number of activated networks.

[0213] There are many ways to combine the scores. One method is to weight the mode impact score and the reasoning accuracy score. The weight can be set according to the actual situation to reflect the relative importance of the two scores in the merging process.

[0214] For example, suppose the weight of the setting method impact score is 0.6, and the weight of the reasoning accuracy score is 0.4 (this means that the impact of the response method on system performance is more important, but there are also certain requirements for the accuracy of the reasoning results). Then, the calculation formula for the combined score is:

[0215] Combined score = 0.6*method impact score + 0.4*inference accuracy score.

[0216] Continuing with the above example, assuming the method impact score is 8 points and the reasoning accuracy score is 9 points, the combined score is:

[0217] Combined score = 0.6 * 8 + 0.4 * 9 = 4.8 + 3.6 = 8.4 points.

[0218] The purpose of step S34 is to determine the number of domain adaptation networks to be activated according to the combined score. This process, for example, involves a score-number mapping relationship, which associates different combined scores with the corresponding number of networks.

[0219] In actual operation, the computer system maintains a table or data structure of score-number mapping. This table contains multiple score intervals and their corresponding network numbers. When the combined score is calculated, the computer system will look up the interval to which the score belongs in the table and determine the corresponding number of networks.

[0220] For example, suppose there is the following score-count mapping relationship:

[0221] Score ≥ 9: 4 networks activated;

[0222] 8≤score<9: activate 3 networks;

[0223] 7≤score<8: 2 networks activated;

[0224] Score <7: 1 network activated;

[0225] Now, suppose the calculated combined score is 8.4, which falls into the interval 8 ≤ score < 9. Therefore, the computer system will determine that 3 domain adaptation networks need to be activated to process the current response mode.

[0226] As another implementation, in step S400, determining to obtain a second plurality of domain adaptation networks from the first plurality of domain adaptation networks includes:

[0227] Step S401: Acquire the response method type of the response method;

[0228] Step S402: determining a third plurality of domain adaptation networks corresponding to the response mode type in the first plurality of domain adaptation networks;

[0229] Step S403: determining to obtain a second number of domain adaptation networks from a third number of domain adaptation networks, wherein the first number is greater than the third number, and the third number is greater than the second number.

[0230] The purpose of step S401 is to obtain the type of the current response mode. The response mode type is a classification or identification of different response modes, which reflects the basic attributes and characteristics of the response mode. In a virtual power plant system, the response mode may include adjusting the output power of the power generation equipment, controlling the charging and discharging behavior of the energy storage equipment, adjusting the power consumption of the adjustable load, and other types.

[0231] In actual operation, the computer system can retrieve the type information of the current response method from a preset response method type library. This type library is a database or data structure containing multiple response methods and their corresponding types, which allows the computer system to quickly find the type of the response method according to its name or identifier.

[0232] For example, if there is a response method called "starting energy storage equipment to respond to peak grid demand", the computer system can look up this response method in the response method type library and determine that its type is "energy storage control". This type information will provide an important reference for subsequent steps.

[0233] The purpose of step S402 is to select the third plurality of networks that are most relevant to the current response mode type from the first plurality of domain adaptation networks. The purpose of this step is to narrow the search scope, reduce unnecessary computational overhead, and improve the processing efficiency of the system.

[0234] To achieve this goal, the computer system can use a method based on domain knowledge or expert experience to predefine the correspondence between each response type and the domain adaptation network. Specifically, the system can maintain a table or data structure of the correspondence between the response type and the domain adaptation network, which contains multiple response types and their corresponding domain adaptation network sets.

[0235] For example, it is assumed that the response mode of the type "energy storage control" is highly correlated with domain adaptation networks related to processing energy storage device status, grid load forecasting, electricity price signals, etc. Then, in step S402, the computer system can select a third number of networks closely related to these topics from the first number of domain adaptation networks for further consideration.

[0236] Assume that there are 10 domain adaptation networks in the first number, and based on domain knowledge or expert experience, it is determined that the third number of networks most relevant to the "energy storage control" type response mode is 5. Then, in step S402, the computer system will select the 5 networks most relevant to "energy storage control" from these 10 networks as candidate networks.

[0237] The purpose of step S403 is to further select the second plurality of networks for activation from the third plurality of candidate networks. The purpose of this step is to dynamically determine which networks are most important for the current task based on the current input features (i.e., the first fused encoding feature representation) and the specific requirements of the response mode.

[0238] To achieve this goal, the computer system can adopt a method based on a selective activation strategy. Specifically, the system can calculate the influence coefficient or weight of each candidate network based on the first fusion encoding feature representation and the two-dimensional array of network weights of the candidate network. Then, based on the size of these influence coefficients, the system selects the second number of networks with the largest influence coefficients for activation.

[0239] The calculation of the influence coefficient may involve a variety of factors, such as the degree of match between the network weights of the candidate network and the input features, the performance of the candidate network on historical data, etc. In actual operation, the computer system can use a complex algorithm or model to calculate these influence coefficients, such as a neural network model based on the attention mechanism.

[0240] For example, in step S403, it is assumed that 5 candidate networks that are most relevant to the "energy storage control" type response mode have been screened out from 10 domain adaptation networks. Now, the computer system will calculate the influence coefficient of each candidate network based on the first fusion coding feature representation and the two-dimensional array of network weights of these 5 candidate networks. Assume that the calculated influence coefficients are 0.8, 0.7, 0.6, 0.5 and 0.4, respectively, and it is decided to activate the first two networks. Then, in step S403, the computer system will select the first two networks with the largest influence coefficients (i.e., the two networks with influence coefficients of 0.8 and 0.7) for activation.

[0241] The determination of the second number may not only depend on the size of the influence coefficient, but may also be limited by other factors, such as the system's computing resources, real-time requirements, etc. In practical applications, the computer system can dynamically adjust the value of the second number according to specific circumstances to ensure that the system's processing efficiency and accuracy are improved as much as possible while meeting performance requirements.

[0242] In summary, step S400 implements the process of dynamically selecting the second plurality of networks from the first plurality of domain adaptation networks for activation by acquiring the response mode type, determining the candidate network set corresponding to the response mode type, and determining the activated network based on the selective activation strategy. This process not only takes into account the characteristics and requirements of the response mode, but also makes full use of the advantages of domain knowledge, expert experience, and the selective activation strategy, thereby improving the processing efficiency and accuracy of the system.

[0243] As an implementation mode, step S100, encoding multiple virtual power plant node state data to obtain multiple first encoding feature representations, includes:

[0244] Step S110: Acquire the basic data of the lamp pole of the target virtual power plant node;

[0245] Step S120: Acquire environmental monitoring data of the environment where the target virtual power plant node is located;

[0246] Step S130: selecting a selected coordinated control strategy set from the set of coordinated control strategies to be selected, and taking the strategy information of each selected coordinated control strategy in the set of coordinated control strategies as coordinated control strategy data in turn;

[0247] Step S140: determining the lamp pole basic data, environmental monitoring data, and coordinated control strategy data as multiple virtual power plant node status data;

[0248] Step S150: Encode the determined multiple virtual power plant node status data to obtain multiple first encoding feature representations.

[0249] The purpose of step S110 is to obtain the basic data of the lamp pole of the target virtual power plant node. The basic data of the lamp pole refers to the basic information related to the smart lamp pole in the virtual power plant node, which includes key information such as the physical properties, technical parameters and operating status of the lamp pole.

[0250] In actual operation, the computer system can obtain these basic data of light poles by accessing the database or sensor network of the virtual power plant node. For example, the basic data of light poles may include the geographical location of the light pole (such as longitude and latitude coordinates), model (such as LED light pole, solar light pole, etc.), rated power, current power, working hours, etc. These data are crucial for understanding the basic status and operating characteristics of the light pole.

[0251] The purpose of step S120 is to obtain environmental monitoring data of the environment where the target virtual power plant node is located. Environmental monitoring data refers to real-time information related to the environment around the virtual power plant node, which reflects the current state and change trend of the environment.

[0252] Environmental monitoring data may include various types of parameters, such as temperature, humidity, light intensity, wind speed, wind direction, precipitation, etc. These data are of great significance for evaluating the impact of the environment on the operation of virtual power plant nodes and formulating corresponding coordinated control strategies. Computer systems can obtain these data through environmental monitoring sensors deployed around virtual power plant nodes.

[0253] The purpose of step S130 is to select a part of the coordinated control strategy set to be selected to form a selected coordinated control strategy set. The coordinated control strategy refers to a series of rules and algorithms used to guide virtual power plant nodes on how to respond to grid demand or environmental changes.

[0254] In actual operation, the computer system can dynamically select a part of the strategies from the preset coordinated control strategy set as the selected coordinated control strategy set that needs to be evaluated according to factors such as the current grid demand, the operating status of the virtual power plant node, and environmental monitoring data. For example, if the current grid demand is peak and the ambient light is sufficient, the system can select those strategies that aim to improve power generation efficiency and reduce energy consumption for evaluation.

[0255] Each strategy in the selected coordinated control strategy set contains a series of strategy information, such as strategy name, parameter settings, execution conditions, etc. This information will be used as coordinated control strategy data in subsequent steps.

[0256] The purpose of step S140 is to integrate the lamp pole basic data obtained in step S110, the environmental monitoring data obtained in step S120, and the coordinated control strategy data selected in step S130 to form virtual power plant node status data.

[0257] The virtual power plant node status data is a comprehensive data set containing multiple types of information, which fully reflects the current status of the virtual power plant node, the environment it is in, and the possible coordinated control strategies. These data will provide important input information for the subsequent coding and prediction process.

[0258] For example, suppose there is a virtual power plant node, whose basic data of light poles include light pole model (LED light pole), current power (80%), rated power (100W), etc.; environmental monitoring data include temperature (25℃), humidity (60%), light intensity (10,000 lux), etc.; selected coordination control strategy data include strategy name (strategy to improve power generation efficiency), parameter settings (adjust light sensor sensitivity), etc. Then, these data will be integrated together to form the status data of the virtual power plant node.

[0259] The purpose of step S150 is to encode the virtual power plant node state data determined in step S140 to obtain a first encoded feature representation. The encoding process is the process of converting raw data into a format that can be processed by a machine learning model, which helps to extract useful information from the data and reduce the dimension of the data.

[0260] In actual operation, the computer system can use a variety of encoding technologies to encode the virtual power plant node status data. For example, for discrete variables in the basic data of lamp poles (such as lamp pole models), the One-Hot Encoding technology can be used to convert them into binary vectors; for continuous variables in environmental monitoring data (such as temperature and humidity), normalization or standardization technology can be used to convert them into values ​​within a specific range; for text information in the coordinated control strategy data (such as strategy names), word embedding technology can be used to convert them into high-dimensional vector representations.

[0261] The encoded virtual power plant node state data will form multiple first encoded feature representations, each of which corresponds to the state information of a virtual power plant node. These first encoded feature representations will be used as inputs to subsequent machine learning models to predict the response trigger confidence of the target virtual power plant node to multiple response modes under a given coordinated control strategy.

[0262] In summary, step S100 obtains basic data of lamp poles, environmental monitoring data and selected coordinated control strategy data, integrates these data into virtual power plant node status data, and finally encodes these data to obtain the first encoded feature representation. This process provides important data support for subsequent predictions and decisions, ensuring the accuracy and effectiveness of the coordinated optimization control of virtual power plants.

[0263] As an implementation mode, the plurality of virtual power plant node state data are derived from a debugging learning sample in a debugging learning database, the debugging learning sample comprising a plurality of priori marks for a plurality of response modes, the priori marks indicating whether a response mode is triggered for the debugging learning sample;

[0264] The first several domain adaptation networks, response mode selectors, and reasoning networks are collaboratively tuned based on the following steps:

[0265] Step S10: for a response mode, determining a branch loss function of the response mode according to a priori mark and a response trigger confidence for the response mode;

[0266] Step S20: determining the loss function of the debugging learning example according to the branch loss function;

[0267] Step S30: Co-debugging the first several domain adaptation networks, the response mode selector and the inference network according to the loss function.

[0268] Each debugging learning example in the debugging learning database is a training sample, and each contains multiple virtual power plant node status data and corresponding prior tags. These prior tags indicate whether different response methods will be triggered under given conditions. For example, a debugging learning example may contain information about the power consumption, ambient temperature, grid load, etc. of a virtual power plant node in a specific time period, as well as prior tags for response methods such as whether to start energy storage equipment and adjust the output power of power generation equipment in response to these conditions.

[0269] At the beginning of network training, the computer system selects one or more debugging learning examples from the debugging learning database as training data. Then, for each response mode in each debugging learning example, the branch loss function of the response mode is determined based on the prior label and the response trigger confidence output by the inference network.

[0270] The branch loss function is a function used to measure the difference between the model's prediction results and the actual results. In the coordinated optimization control of virtual power plant energy, the branch loss function may take the form of cross entropy loss, mean square error loss, etc. The specific choice depends on the requirements of the task and the characteristics of the model. For example, if the response trigger confidence is expressed in a probabilistic form, then the cross entropy loss may be a suitable choice; if the response trigger confidence is expressed in a numerical form, then the mean square error loss may be more appropriate.

[0271] After determining the branch loss function for each response, the computer system integrates these loss functions to obtain a loss function that represents the loss of the entire debugging learning example. This step involves, for example, weighted summation of the branch loss functions or other forms of fusion strategies. Weighted summation is a simple and effective method that allows the computer system to adjust the weights of different loss functions based on the importance or complexity of the response.

[0272] For example, suppose there are two response methods: starting the energy storage device and adjusting the output power of the power generation device, and their corresponding branch loss functions are L1 and L2 respectively. If starting the energy storage device is considered more important, then a larger weight (such as 0.7) can be assigned to L1, and a smaller weight (such as 0.3) can be assigned to L2. In this way, the loss function of the entire debugging learning example can be expressed as 0.7L1+0.3L2.

[0273] After obtaining the loss function of the debugging learning example, the computer system can use this loss function to collaboratively debug the domain adaptation network, the response mode selector, and the inference network. Collaborative debugging is an iterative process that involves the following steps:

[0274] The computer system inputs the virtual power plant node state data in the debugging learning sample into the domain adaptation network to obtain a preliminary feature representation. These feature representations are then passed to the response mode selector to determine which domain adaptation network outputs will be used in the subsequent reasoning process. Finally, the selected domain adaptation network output is input into the reasoning network to obtain the response trigger confidence for each response mode.

[0275] Based on the response trigger confidence output by the inference network and the prior labels in the debugging learning example, the computer system calculates the branch loss function value of each response mode, and integrates these values ​​to obtain the loss function value of the entire debugging learning example.

[0276] Using optimization algorithms such as gradient descent, the computer system updates the parameters of the domain adaptation network, response mode selector, and inference network according to the loss function value to reduce the loss function value and improve the prediction accuracy of the model.

[0277] Repeat the above steps until the loss function value converges to a stable level or reaches the preset training rounds.

[0278] In the process of co-tuning, the domain adaptation network, response mode selector and reasoning network are interrelated, and their parameter updates will affect each other. Therefore, co-tuning can ensure that these three models can work closely together when processing actual tasks and jointly improve the overall performance of the system.

[0279] It should be noted that the network training process can involve a lot of calculation and optimization work. In order to ensure the efficiency and accuracy of training, the computer system can adopt some advanced technical means, such as batch processing, data enhancement, model pruning, etc. In addition, due to the complexity and uncertainty of the virtual power plant system, network training may also require multiple iterations and adjustments to gradually optimize the performance of the model.

[0280] In summary, network training is an important part of the implementation of the virtual power plant energy coordination optimization control method. Through the debugging learning examples in the debugging learning database, the computer system can collaboratively debug the domain adaptation network, the response mode selector, and the reasoning network, thereby improving their accuracy and efficiency in handling actual tasks. This process not only relies on advanced machine learning technology and optimization algorithms, but also requires a deep understanding and accurate modeling of the virtual power plant system.

[0281] As an implementation mode, step S20, determining the loss function of the debugging learning example according to the branch loss function, includes:

[0282] Step S21: obtaining response mode influence coefficients of multiple response modes;

[0283] Step S22: According to the response mode influence coefficient, the branch loss functions of multiple response modes are merged to obtain a loss function.

[0284] The purpose of step S21 is to obtain the response mode influence coefficient of each response mode. The response mode influence coefficient is a quantitative indicator that reflects the importance or contribution of different response modes in the coordinated optimization control of virtual power plants. These coefficients are determined, for example, based on domain knowledge, expert experience or historical data, and are used to weight the branch losses of different response modes in the loss function calculation process.

[0285] In actual operation, the computer system can maintain a response mode influence coefficient table, which contains all possible response modes and their corresponding influence coefficients. These coefficients can be pre-set fixed values ​​or dynamically adjusted according to specific circumstances. For example, for those response modes that have a greater impact on the stability of the power grid (such as emergency load shedding), their influence coefficients can be set to higher values ​​to emphasize the importance of these response modes in the training process.

[0286] Assume that there are three response modes: starting the energy storage device (mode A), adjusting the output power of the power generation equipment (mode B), and adjusting the power consumption of the adjustable load (mode C). Based on domain knowledge and expert experience, the impact coefficients of these response modes may be set to 0.6, 0.3, and 0.1, respectively, indicating that mode A has the greatest impact on system performance, followed by mode B, and mode C has the least impact.

[0287] In step S22, the computer system will fuse the branch loss functions of multiple response modes according to the response mode influence coefficient obtained in step S21 to obtain a unified loss function. This loss function will be used in the subsequent network training process to guide the optimization direction of the model parameters.

[0288] The fusion process involves, for example, weighted summation of branch loss functions. Specifically, for each response mode, the computer system calculates its branch loss function value and weights the loss value according to the influence coefficient of the response mode. Then, all weighted loss values ​​are added together to obtain the loss function value of the entire debugging learning example.

[0289] Assume there are three response modes A, B and C, and their corresponding branch loss functions are L A , L B and L C , the influence coefficients are 0.6, 0.3 and 0.1 respectively. Then, the loss function L of the entire debugging learning example can be calculated by the following formula:

[0290] L=0.6·L A +0.3 L B +0.1 L C .

[0291] This formula shows that the branch loss function value of each response method is weighted according to its influence coefficient, and the proportion of the loss value of different response methods in the final loss function is proportional to its influence coefficient. In this way, the computer system can pay more attention to those response methods that have a greater impact on system performance during training, thereby improving the optimization efficiency and prediction accuracy of the model.

[0292] In summary, step S20 obtains a unified loss function by obtaining the response mode influence coefficient and fusing the branch loss functions of multiple response modes. This process not only takes into account the importance differences of different response modes in the coordinated optimization control of virtual power plant energy, but also realizes the effective integration of multiple loss functions by weighted summation. This loss function will provide a clear optimization target for subsequent network training and guide the continuous adjustment and optimization of model parameters.

[0293] An embodiment of the present invention provides a computer system, such as Figure 2As shown, the computer system 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the computer system 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the computer system 100 does not constitute a limitation on the embodiments of the present invention.

[0294] An embodiment of the present invention provides a computer system. The computer system in the embodiment of the present invention includes: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more programs are executed by the processor, the above method is implemented.

Claims

1. A method for energy coordination and optimization control of a virtual power plant, characterized in that: include: Encoding a plurality of virtual power plant node state data to obtain a plurality of first encoding feature representations, wherein the virtual power plant node state data is used to predict a response trigger confidence of a target virtual power plant node to a coordinated control strategy for a plurality of response modes; sequentially concatenating the plurality of the first coding feature representations to obtain a first fused coding feature representation; Loading the first fused encoded feature representation into a first several domain adaptation network; Based on the response mode selector corresponding to each of the response modes, a second number of the domain adaptation networks is determined from the first number of the domain adaptation networks, and the outputs of the determined second number of the domain adaptation networks are fused to obtain a second encoding feature representation corresponding to the response mode, wherein the second number is smaller than the first number; The second coded feature representation corresponding to the multiple response modes is loaded into the inference network corresponding to the multiple response modes, and the response trigger confidence of the target virtual power plant node to the coordinated control strategy for the multiple response modes obtained by inference is obtained.

2. The energy coordination optimization control method of a virtual power plant according to claim 1 is characterized in that: The determining of obtaining a second plurality of the domain adaptation networks from the first plurality of the domain adaptation networks comprises: Generate an influence coefficient vector according to the first fused coding feature representation and the first two-dimensional array, wherein the number of columns in the first two-dimensional array is the first number, the number of rows is the dimension of the first fused coding feature representation, and the influence coefficient vector is used to characterize the influence coefficient of the first number of the domain adaptation network; According to the influence coefficient vector, a second number of the domain adaptation networks are determined from the first number of the domain adaptation networks.

3. The energy coordination optimization control method of a virtual power plant according to claim 2 is characterized in that: The generating the influence coefficient vector according to the first fusion coding feature representation and the first two-dimensional array includes: Taking the product of the first two-dimensional array and the first fused coding feature representation to obtain an initial influence coefficient vector; A second two-dimensional array is multiplied by the first fused coding feature representation to obtain a perturbation base array, wherein the number of columns in the second two-dimensional array is the first number, the number of rows is the dimension of the first fused coding feature representation, and the element values ​​in the second two-dimensional array are randomly generated; Generate an influence coefficient adjustment array according to the disturbance basic array of the dynamic adjustment array, wherein the dynamic adjustment array is a random disturbance array, and the elements in the dynamic adjustment array are random numbers following Gaussian distribution or other distributions; The influence coefficient vector is generated according to the initial influence coefficient vector and the influence coefficient adjustment array.

4. The energy coordination optimization control method of a virtual power plant according to claim 3 is characterized in that: The generating of the influence coefficient adjustment array according to the dynamic adjustment array and the disturbance basic array includes: Obtaining a set of variant values, wherein the variant values ​​in the set of variant values ​​follow a Gaussian distribution; Randomly select the first number of the variation values ​​in the variation value set to construct the dynamic adjustment array; Performing a dot product calculation on the dynamic adjustment array and the disturbance basic array to obtain the influence coefficient adjustment array; Determining a second plurality of the domain adaptation networks from the first plurality of the domain adaptation networks according to the influence coefficient vector comprises: Determine the first second several vector units in the vector units in the influence coefficient vector arranged in descending order; The domain adaptation networks corresponding to the first second number of vector units are used as the determined second number of domain adaptation networks; The step of fusing the outputs of the second plurality of domain adaptation networks determined includes: Replacing the remaining vector units except the first second number of vector units arranged in descending order in the influence coefficient vector with infinitesimal values; Performing a standardized exponential transformation on the influence coefficient vector to obtain a second number of first influence coefficients of the domain adaptation network; The outputs of the second plurality of domain adaptation networks determined are fused according to the first influence coefficient.

5. The energy coordination optimization control method of a virtual power plant according to claim 4 is characterized in that: The domain adaptation network includes a two-dimensional array of network weights and an offset, the number of columns in the two-dimensional array of network weights is the dimension of the output of the domain adaptation network, the number of rows is the dimension represented by the first fused coding feature, and the dimension of the offset is equal to the dimension of the output of the domain adaptation network; The step of fusing the outputs of the second plurality of domain adaptation networks determined according to the first influence coefficient comprises: Obtain a multi-layer weight set, wherein the multi-layer weight set represents a two-dimensional array of network weights of a first plurality of the domain adaptation networks; Obtaining an offset matrix, wherein the offset matrix represents the offsets of a first number of the domain adaptation networks; In the multi-layer weight set, obtaining the two-dimensional array of network weights corresponding to a second number of the first influence coefficients; Obtaining the offsets corresponding to a second number of the first influence coefficients in the offset matrix; A fusion result is determined based on a second number of the first influence coefficients, a second number of the two-dimensional arrays of network weights corresponding to the first influence coefficients, and a second number of the offsets corresponding to the first influence coefficients.

6. The energy coordination optimization control method of a virtual power plant according to claim 5 is characterized in that: Determining a fusion result according to a second number of the first influence coefficients, a second number of the network weight two-dimensional arrays corresponding to the first influence coefficients, and a second number of the offsets corresponding to the first influence coefficients includes: Multiplying the two-dimensional array of network weights corresponding to each of the second plurality of domain adaptation networks by the first fused encoding feature representation loaded into the domain adaptation network to obtain a transition feature representation; Adding the transition feature representation to the offset corresponding to the domain adaptation network to obtain an added feature representation; Performing a nonlinear transformation on the added feature representation to obtain an output of the domain adaptation network; The fusion result is obtained according to the outputs of the second plurality of the domain adaptation networks and the first influence coefficient corresponding to the domain adaptation network.

7. The energy coordination optimization control method of a virtual power plant according to claim 1, characterized in that: The process of determining the second number includes: Obtaining a response method type of the response method; obtaining an inference target accuracy of the response mode; Determining the second number according to the response mode type and the inference target accuracy; Determining a response mode impact score based on the response mode type; Determining a reasoning accuracy score according to the reasoning target accuracy; Determining a combined score based on the mode impact score and the reasoning accuracy score; The second number is determined according to the combined score.

8. The energy coordination optimization control method of a virtual power plant according to claim 1, characterized in that: The determining of obtaining a second plurality of the domain adaptation networks from the first plurality of the domain adaptation networks comprises: Obtaining a response method type of the response method; Determining, among the first plurality of the domain adaptation networks, a third plurality of the domain adaptation networks corresponding to the response mode type; Determining a second number of the domain adaptation networks from the third number of the domain adaptation networks, wherein the first number is greater than the third number, and the third number is greater than the second number; The step of encoding the plurality of virtual power plant node state data to obtain a plurality of first encoding feature representations includes: Obtaining basic data of lamp poles of the target virtual power plant node; Acquire environmental monitoring data of an environment in which the target virtual power plant node is located; Selecting a selected coordinated control strategy set from the set of coordinated control strategies to be selected, and using the strategy information of each selected coordinated control strategy in the selected coordinated control strategy set as coordinated control strategy data in turn; Determine the lamp pole basic data, the environmental monitoring data, and the coordinated control strategy data as a plurality of virtual power plant node status data; Encoding the determined plurality of virtual power plant node status data to obtain a plurality of first encoding feature representations; Wherein, the plurality of virtual power plant node status data originate from a debugging learning sample in a debugging learning database, the debugging learning sample comprises a plurality of priori marks for the plurality of response modes, the priori marks indicating whether the response mode is triggered for the debugging learning sample; The first several domain adaptation networks, the response mode selectors, and the reasoning networks are collaboratively debugged based on the following steps: For one of the response modes, determining a branch loss function of the response mode according to the priori mark and the response trigger confidence for the response mode; Determining a loss function of the debugging learning example according to the branch loss function; The first plurality of the domain adaptation networks, the response mode selectors, and the inference network are collaboratively debugged according to the loss function.

9. The energy coordination optimization control method of a virtual power plant according to claim 8, characterized in that: Determining the loss function of the debugging learning example according to the branch loss function includes: Obtaining response mode influence coefficients of a plurality of response modes; According to the response mode influence coefficient, the branch loss functions of multiple response modes are merged to obtain the loss function.

10. A computer system, characterized in that: include: one or more processors; Memory; one or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 9 is implemented.

Images