Virtual power plant resource optimization scheduling system based on multi-head self-calibration tensor factorization

Abstract

The invention relates to the technical field of power system optimization scheduling, in particular to a virtual power plant resource optimization scheduling method based on multi-head self-calibration tensor factors. Comprising the following steps: S1, collecting virtual power plant data, preprocessing the virtual power plant data, and carrying out standardization, anomaly detection and interpolation on various heterogeneous resource data in a virtual power plant; s2, constructing a hierarchical gradient tensor network, executing multi-head self-calibration decomposition, and establishing a virtual power plant decision optimization model, and S3, designing a multi-head self-calibration decomposition mechanism according to the virtual power plant decision optimization model, and generating a virtual power plant resource optimization scheduling scheme. The technical problems that in existing virtual power plant large-scale heterogeneous resource scheduling, calculation complexity is high, memory occupation is large, and control precision and efficiency are difficult to balance can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of power system optimization and scheduling, and in particular to a virtual power plant resource optimization and scheduling method based on multi-head self-calibration tensor factorization. Background Art

[0002] Virtual power plants (VPPs), a new energy management model, are rapidly developing globally. Using information technology and advanced control methods, VPPs aggregate geographically dispersed and diverse distributed energy resources, such as distributed photovoltaics, wind power, energy storage, and controllable loads, into a unified, regulated entity that participates in power market transactions and system regulation.

[0003] Traditional virtual power plant resource scheduling methods are primarily based on linear or nonlinear programming models, such as mixed-integer linear programming (MILP) and quadratic programming (OP). These methods work well for small-scale systems, but as virtual power plants scale and resource types become increasingly diverse, they begin to exhibit numerous limitations. First, traditional optimization methods face the "curse of dimensionality." When dealing with tens or even hundreds of thousands of heterogeneous resources, computational complexity increases exponentially, making it difficult to solve within the time window required for real-time scheduling. For example, a virtual power plant with 10,000 heterogeneous resources typically takes several hours to solve using traditional MILP methods, failing to meet the 5-15 minute scheduling time requirement. Second, large-scale heterogeneous resources exhibit complex spatiotemporal coupling and nonlinear interactions, making traditional methods difficult to accurately model and efficiently solve. Third, the randomness and volatility of renewable energy significantly increase system uncertainty, reducing the stability of optimization results. Furthermore, traditional methods often treat all resources equally, failing to differentiate resources based on their contribution to the system, resulting in wasted computing resources.

[0004] With the development of artificial intelligence (AI) technology, researchers have begun exploring new data-driven approaches for virtual power plant resource scheduling. Machine learning methods such as support vector machines (SVM) and random forests (RF) are being used to predict renewable resource processing and electricity load, improving forecast accuracy. Deep learning methods such as deep neural networks (DNNs) and long short-term memory networks (LSTMs) are being applied to end-to-end scheduling policy generation. However, existing methods still have significant shortcomings. First, the "black box" nature of neural networks makes scheduling decisions lacking in interpretability, making it difficult to gain the trust of system operators. Second, the large number of model parameters and the significant computational resources required for training and inference make them difficult to deploy in edge computing environments. Furthermore, existing methods have limited processing capabilities for high-dimensional, heterogeneous data, making it difficult to effectively analyze the complex interactions between different types of resources.

[0005] Tensor decomposition, a high-dimensional data dimensionality reduction and analysis technique, has recently begun to be applied to power systems, demonstrating promising improvements in forecasting accuracy and reduced computational time. However, these methods primarily focus on data analysis and forecasting, with limited application to resource scheduling optimization. Furthermore, they employ static tensor decomposition models, lacking the ability to adapt to dynamic environments and struggling to cope with real-time changes in virtual power plant resource characteristics and system states. Furthermore, existing tensor methods fail to fully account for the physical constraints and specialized knowledge of power systems, resulting in solutions that may not meet actual operational requirements.

[0006] In summary, the existing virtual power plant resource scheduling has the following problems: 1. The computational complexity of the high-dimensional decision space caused by large-scale heterogeneous resources. 2. The irrational allocation of computing resources due to differences in resource importance. 3. The large number of model parameters leads to memory usage and deployment difficulties. 4. The lack of model adaptability due to dynamic changes in resource status. These issues seriously hinder the large-scale application and efficient operation of virtual power plants. Summary of the Invention

[0007] The technical problem solved by the present invention is to provide a The basic solution provided by the present invention is a virtual power plant resource optimization scheduling system based on multi-head self-calibration tensor factorization, which includes the following steps: S1. Collect and pre-process virtual power plant data, standardize, detect anomalies, and interpolate various heterogeneous resource data within the virtual power plant; S2, constructs a hierarchical gradient tensor network and performs multi-head self-calibration decomposition to establish a virtual power plant decision optimization model, S2 as Figure 3 The specific steps shown include: S21. Use tensor factorization method to decompose the high-dimensional control decision tensor into the product form of super core tensor and factor matrix:

[0008] in, represents the high-dimensional control decision tensor, Indicates the key information after dimensionality reduction stored in the super core tensor, represents the steamed dumpling gene matrix, represents the self-calibration factor matrix, represents the tensor-matrix product along the i-th mode, Represents the dimensionality reduction coefficient of each dimension, satisfying ; S22. Build a gradient tensor network, represent the tensor factorization process as a computational graph, and achieve end-to-end differentiable optimization S22, multi-head self-calibration decomposition includes constructing a self-calibration factor matrix:

[0009] in, represents the multi-head attention parameter matrix, Represents a sparse mask matrix based on resource characteristics, GELU represents the Gaussian error linear unit activation function, layerNorm represents the normalization operation, STE represents the pass-through estimator, Indicates that the parameter is The adaptive threshold function, represents the learnable weight matrix, represents the Hadamrd product.

[0010] S3. Based on the virtual power plant decision optimization model, a multi-head self-calibration decomposition mechanism is designed to generate a virtual power plant resource optimization scheduling plan.

[0011] Furthermore, the S1 comprises the following steps: S11. Collect the status information of various heterogeneous resources in the virtual power plant, including the state vector of each resource node i , response characteristic vector , cost characteristic vector , distance feature vector , and energy availability index , the state vector Indicates the current working state of the resource, including power generation power, energy storage power and load power. The response characteristic vector Represents the response characteristics of resources to control instructions, including ramp rate, response time and control accuracy. The cost characteristic vector Represents the economic characteristics of resources, including power generation cost, control cost, start-up and shutdown cost. The distance characteristic vector Indicates the relative position relationship of resources in the electrical network, energy availability index Indicates the predictable handling of renewable energy or the adjustable capacity of controllable load; S12. Data Adaptive Normalization:

[0012] in, represents the standardized data, x represents the original data, represents the data mean, represents the standard deviation of the data, A small positive number used to prevent the denominator from being 0; S1.3 Constructing high-dimensional decision tensors ,in Indicates the resource quantity dimension, Represents the time dimension.

[0013] Furthermore, the step S22 includes the following steps: S221, forward Zhuang Bo calculation reconstruction tensor :

[0014] S222, back propagation calculates gradients and updates network parameters:

[0015]

[0016] in, represents the loss function, Represents a supercore tensor The expansion matrix on the i-th mode, represents the Kronecker product.

[0017] Furthermore, the step S3 includes the following steps: S31 builds self-calibration factor matrix :

[0018] in, 、 、 is the parameter matrix in the multi-head attention mechanism, is the dimension of the attention head, It is a sparse mask matrix based on resource characteristics, GELU represents the Gaussian error linear unit activation function, and LayerNorm represents the layer normalization operation. represents the Hadamard product, STE represents the straight-through estimation, is a parameter The adaptive threshold function, is a learnable weight matrix; S32. Design a multi-head attention mechanism to divide the self-calibration factor mean into H independent attention heads, each of which focuses on modeling the interaction of a specific type of resource:

[0019]

[0020] in, 、 、 is the parameter matrix of the h-th attention head, is the sparse mask matrix of the h-th attention head, is the output projection matrix; S33. Design mask sparse matrix , based on three different resource relationships to establish sparse connections, physical connection relationship mask Based on the topological connection relationship of resources in the power network, historical interaction mask Generate a time-dependent mask based on the frequency of interactions between resources in historical data Based on the resource response time design, the mask matrix is ​​finally obtained through weighted fusion:

[0021] in, 、 、 is a learnable weight coefficient that satisfies and .

[0022] Furthermore, the method further comprises the following steps: S4. Introduce a resource importance adaptive evaluation mechanism based on Fisher information to identify and prioritize key resources.

[0023] Further, the S4 is as follows Figure 4 The following steps are involved: S41. Design a resource importance evaluation function to calculate the importance weight of each resource node i :

[0024] in, is the Swish activation function, defined as , is a learnable parameter, The parameter is Multilayer perceptron, 、 、 、 、 They are the state vector, response characteristic vector, cost characteristic vector, distance characteristic vector, and energy availability index of resource i; S42. Calculate the Fisher information matrix of the resource parameters to quantify their importance to the model output:

[0025] in, is a parameter The Fisher information matrix of is the data distribution, is the model prediction probability, and the diagonal approximation is used to simplify the calculation:

[0026] S43, based on resource importance weight and Fisher information matrix , design resource layering processing strategies and set importance thresholds , classify resources into high importance , medium importance and low importance three floors; The high-importance resources are calculated using high-precision complete calculations, the medium-importance resources are calculated using simplified calculations, and the low-importance resources are calculated using warning calculations.

[0027] Furthermore, the method further comprises the following steps: S5 combines LoRA technology to implement low-rank adaptive factor optimization and expresses the self-calibration factor matrix as a pseudo-low-rank matrix of the identity matrix. Specifically, the following steps are included: S51. Optimize the self-calibration factor matrix using low-rank representation :

[0028] in, is the identity matrix, and is the low-order projection matrix, d is the low-order projection dimension, satisfying ; S52. Design an initialization strategy for the low-rank projection matrix. The matrix is ​​initialized using a Gaussian distribution with variance scaled to :

[0029] The A matrix is ​​initialized to a zero matrix, which is used to make the self-calibration factor matrix close to the identity matrix in the early stage of training; S53, introduce the conditional calculation dynamic activation mechanism, and selectively activate the low-rank air volume through the gating function:

[0030] in, is the optimized self-calibration factor matrix, is the gating function based on the input state x, k is the number of expert models, and is the low-rank projection matrix of the j-th expert model, and the gating function Using a lightweight multilayer perceptron:

[0031]

[0032] in, 、 、 and are the gating network parameters, is the input dimension, is a hidden dimension, is the temperature coefficient, which is used to adjust the smoothness of the activation probability distribution.

[0033] Furthermore, the method further comprises the following steps: S6. Design a complete algorithm training and inference process, including phased training strategy, mixed precision training, and dynamic precision quantization. Specifically, the following steps are included: S61. Design a phased training strategy. Phase 1: Freeze the self-calibration factor matrix. , only train the supercore tensor and the factor matrix , learning basic data representation; Phase 2: Unfreeze the self-calibration factor matrix But freeze the super core tensor and the factor matrix , optimize the self-calibration capability; the third stage: unfreeze all parameters and perform end-to-end adjustment; S62. Design a loss function that combines multiple objectives, including reconstruction error, dilution constraint, energy balance, and cost optimization:

[0034] in, is the reconstruction loss, which measures the difference between the original tensor and the reconstructed tensor using the Frobenius norm:

[0035] is the sparsity loss, and L1 regularization is used to promote the sparsification of the self-calibration factor matrix:

[0036] is the energy balance loss, which is used to ensure the balance of energy supply and demand in the virtual power plant:

[0037] in, represents all control decisions of the i-th resource, 1 represents a full 1 vector, is the demand curve, is the cost optimization loss, which is used to minimize the total cost of resource scheduling:

[0038] in, is the cost coefficient for resource i to perform control k at time j, 、 、 and is the weight coefficient, which is used to control the relative importance of each loss item; Design an adaptive weight adjustment strategy:

[0039] in, is the weight coefficient of the t-th round of training, is the initial weight, is the decay rate, is the minimum weight; S63, adopt hybrid progress training to accelerate the calculation process, use FP half-precision floating point numbers for forward and backward propagation calculations, use FP32 single-precision floating point numbers to store main parameters and perform parameter updates, and design a dynamic loss scaling strategy to prevent gradient underflow:

[0040]

[0041] in, is the scaled loss, s is the scaling factor, and are the growth and reduction coefficients, set to 2 and 4 respectively.

[0042] Furthermore, the method further comprises the following steps: S7. Improve the computational efficiency of the inference phase through dynamic precision quantization and sparse computing optimization. This includes the following steps: S71. Design dynamic precision quantification strategy based on resource importance weights Dynamically select calculation accuracy, high-precision resources Using FP32 progress, moderately important resources Use FP16 precision, low critical resources Use INT8 precision; design a quantization lookup table, pre-calculate the quantization results of commonly used values ​​to accelerate the quantization process, and use a symmetric quantization method for INT8 quantization:

[0043] Where q is the quantized value, x is the original floating-point value, s is the quantization scale factor, round indicates the rounding function, and the quantization scale factor S is adaptively calculated based on the data range:

[0044] S72. Using the self-calibration factor matrix The structured sparsity of the tensor is used to accelerate tensor operations by setting the threshold , set the elements close to zero to zero and increase the matrix dilution:

[0045] Then the diluted matrix is ​​stored in a compressed diluted row format to reduce memory usage; S73 combines multiple control instructions for batch processing to improve computing throughput.

[0046] The principle and effect of the present invention are: By combining hierarchical gradient tensor networks with multi-head self-calibration decomposition technology, we construct an efficient dimensionality reduction framework for high-dimensional decision spaces, enabling optimal scheduling of virtual power plant resources. The core of this invention lies in establishing an innovative decomposition representation of the high-dimensional control decision tensor: T≈G×1(U1P1)×2(U2P2)×3(U3P3). This introduces a self-calibration factor matrix P1 for refined adjustments, significantly improving the expressiveness and adaptability of traditional tensor decomposition methods.

[0047] Building on this foundation, this paper employs a multi-head self-calibration decomposition mechanism. By designing an innovative self-calibration factor matrix construction method: P1 = GELU(LayerNorm(Q1K1ᵀ / √d1)⊙M1)V1⊙STE(h_ϕ(W1)), this method introduces a multi-head sparse attention and entropy regularization self-calibration mechanism to achieve differentiated modeling and precise coordination between different resource types. Compared to traditional methods, this mechanism can adaptively capture the characteristics and interaction patterns of different resource types in a virtual power plant, significantly improving the model's expressiveness.

[0048] This paper also introduces an adaptive resource importance assessment mechanism based on Fisher information. Using w1=σ(f_ϕ(s1,r1,c1,d1,e1)), this mechanism ranks resources by importance, accurately identifying and prioritizing key resources. This avoids the waste of computational resources that results from traditional approaches that treat all resources equally. This mechanism, combined with a neural structure distillation network, can extract key influencing factors from multidimensional features, providing a basis for subsequent differentiated processing.

[0049] To further optimize computational efficiency, this paper combines LoRA technology to implement low-rank adaptive factor optimization. The self-calibration factor matrix is ​​expressed as P1=I_{r1}+BA, where the low-rank projection matrices B and A significantly reduce the number of parameters and memory usage. This design reduces the computational complexity from O(r1²) to O(r1·d), where d ≪ r1, while maintaining high accuracy, achieving significant computational speedup.

[0050] Another innovation of this invention is the introduction of a dynamic activation mechanism for conditional computation: P̃1=I_{r1}+∑ⱼ₌1ᵏgⱼ(x)·BⱼAⱼ. This mechanism dynamically activates the most important low-rank components based on the input state, enabling on-demand allocation of computing resources. This mechanism avoids the redundant computations found in traditional methods, dynamically matching the computational effort with the current resource state and scheduling requirements, and further improving system responsiveness.

[0051] In terms of algorithm training and inference, this paper designs a comprehensive optimization strategy, including phased training, mixed-precision training, dynamic precision quantization, sparse computing optimization, batch processing and pre-computation techniques, and model distillation, forming a complete and efficient computing framework. The combined application of these strategies enables the method to maintain excellent computing efficiency and control accuracy in large-scale virtual power plant scenarios.

[0052] The present invention is also specially optimized for heterogeneous resource scenarios and large-scale scenarios, designs a resource type-specific factor matrix structure and hierarchical computing framework, and introduces a dynamic environment adaptation mechanism to ensure that the system maintains stable and efficient performance in complex and changeable actual operating environments.

[0053] Through the above technical solutions, the present invention has achieved a number of technological breakthroughs in large-scale heterogeneous resource scheduling of virtual power plants: first, the multi-head self-calibration tensor factorization technology makes the expression and calculation of high-dimensional decision space efficient and feasible; second, the adaptive evaluation mechanism based on resource importance realizes the reasonable allocation of computing resources; third, low-rank adaptive factor optimization and conditional calculation dynamic activation greatly reduce memory usage and computational complexity; finally, comprehensive training and inference optimization strategies further improve system performance.

[0054] Experimental results demonstrate that, in a large-scale virtual power plant test scenario involving 50,000 heterogeneous nodes, the proposed method reduces computation time by 99.2%, memory usage by 89.4%, control accuracy by 97.9%, and resource utilization by 95.2% compared to existing technologies. Furthermore, the proposed method demonstrates excellent scalability, growing logarithmically with the number of nodes rather than linearly or exponentially, demonstrating superlinear scalability. It is also highly adaptable to environmental changes and interference, enabling rapid model adjustments to new operating conditions. This provides an innovative solution for the efficient coordinated scheduling of large-scale virtual power plants.

[0055] This invention, through the integration of multi-head self-calibration tensor factorization and a series of innovative optimization techniques, successfully addresses key technical challenges in the coordinated scheduling of large-scale heterogeneous resources in virtual power plants, achieving breakthroughs in both computational efficiency and control accuracy. This provides key support for the widespread adoption of virtual power plant technology and the efficient use of renewable energy. Compared to traditional methods, this invention achieves significant improvements in computational performance, memory usage, control accuracy, and adaptability, demonstrating outstanding technical innovation and practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a schematic diagram of the system architecture of an embodiment of a virtual power plant resource optimization and scheduling system based on multi-head self-calibration tensor factorization according to the present invention; Figure 2 This is a resource optimization scheduling flow chart of an embodiment of a virtual power plant resource optimization scheduling system based on multi-head self-calibration tensor factorization of the present invention; Figure 3 A schematic diagram of multi-head self-calibration tensor factorization of an embodiment of a virtual power plant resource optimization and scheduling system based on multi-head self-calibration tensor factorization according to the present invention; Figure 4 A schematic diagram of a resource importance adaptive evaluation mechanism according to an embodiment of a virtual power plant resource optimization and scheduling system based on multi-head self-calibration tensor factorization according to the present invention; Figure 5 A schematic diagram of a low-rank adaptive factor optimization and conditional calculation dynamic activation mechanism of an embodiment of a virtual power plant resource optimization and scheduling system based on multi-head self-calibration tensor factorization of the present invention; Figure 6 This is a performance comparison diagram of a virtual power plant resource optimization scheduling method according to an embodiment of a virtual power plant resource optimization scheduling system based on multi-head self-calibration tensor factorization of the present invention. DETAILED DESCRIPTION

[0057] The following is further described in detail through specific implementation methods: The embodiment is basically as shown in the attached Figure 1 As shown: A virtual power plant resource optimization scheduling system based on multi-head self-calibration tensor factorization includes the following steps: S1. Collect and pre-process virtual power plant data, standardize, detect anomalies, and interpolate various heterogeneous resource data within the virtual power plant; S2: Construct a hierarchical gradient tensor network and perform multi-head self-calibration decomposition to establish a virtual power plant decision optimization model. S2 specifically includes the following steps: S21. Use tensor factorization method to decompose the high-dimensional control decision tensor into the product form of super core tensor and factor matrix:

[0058] in, represents the high-dimensional control decision tensor, Indicates the key information after dimensionality reduction stored in the super core tensor, represents the steamed dumpling gene matrix, represents the self-calibration factor matrix, represents the tensor-matrix product along the i-th mode, Represents the dimensionality reduction coefficient of each dimension, satisfying ; S22. Build a gradient tensor network, represent the tensor factorization process as a computational graph, and achieve end-to-end differentiable optimization S22, multi-head self-calibration decomposition includes constructing a self-calibration factor matrix:

[0059] in, represents the multi-head attention parameter matrix, Represents a sparse mask matrix based on resource characteristics, GELU represents the Gaussian error linear unit activation function, layerNorm represents the normalization operation, STE represents the pass-through estimator, Indicates that the parameter is The adaptive threshold function, represents the learnable weight matrix, represents the Hadamrd product.

[0060] S3. Based on the virtual power plant decision optimization model, a multi-head self-calibration decomposition mechanism is designed to generate a virtual power plant resource optimization scheduling plan.

[0061] Said S1 comprises the following steps: S11. Collect the status information of various heterogeneous resources in the virtual power plant, including the state vector of each resource node i , response characteristic vector , cost characteristic vector , distance feature vector , and energy availability index , the state vector Indicates the current working state of the resource, including power generation power, energy storage power and load power. The response characteristic vector Represents the response characteristics of resources to control instructions, including ramp rate, response time and control accuracy. The cost characteristic vector Represents the economic characteristics of resources, including power generation cost, control cost, start-up and shutdown cost. The distance characteristic vector Indicates the relative position relationship of resources in the electrical network, energy availability index Indicates the predictable handling of renewable energy or the adjustable capacity of controllable load; S12. Data Adaptive Normalization:

[0062] in, represents the standardized data, x represents the original data, represents the data mean, represents the standard deviation of the data, A small positive number used to prevent the denominator from being 0; S1.3 Constructing high-dimensional decision tensors ,in Indicates the resource quantity dimension, Represents the time dimension.

[0063] Furthermore, the step S22 includes the following steps: S221, forward Zhuang Bo calculation reconstruction tensor :

[0064] S222, back propagation calculates gradients and updates network parameters:

[0065]

[0066] in, represents the loss function, Represents a supercore tensor The expansion matrix on the i-th mode, represents the Kronecker product.

[0067] The S3 includes the following steps: S31 builds self-calibration factor matrix :

[0068] in, 、 、 is the parameter matrix in the multi-head attention mechanism, is the dimension of the attention head, It is a sparse mask matrix based on resource characteristics, GELU represents the Gaussian error linear unit activation function, and LayerNorm represents the layer normalization operation. represents the Hadamard product, STE represents the straight-through estimation, is a parameter The adaptive threshold function, is a learnable weight matrix; S32. Design a multi-head attention mechanism to divide the self-calibration factor mean into H independent attention heads, each of which focuses on modeling the interaction of a specific type of resource:

[0069]

[0070] in, 、 、 is the parameter matrix of the h-th attention head, is the sparse mask matrix of the h-th attention head, is the output projection matrix; S33. Design mask sparse matrix , based on three different resource relationships to establish sparse connections, physical connection relationship mask Based on the topological connection relationship of resources in the power network, historical interaction mask Generate a time-dependent mask based on the frequency of interactions between resources in historical data Based on the resource response time design, the mask matrix is ​​finally obtained through weighted fusion:

[0071] in, 、 、 is a learnable weight coefficient that satisfies and .

[0072] Furthermore, the method further comprises the following steps: S4. Introduce a resource importance adaptive evaluation mechanism based on Fisher information to identify and prioritize key resources.

[0073] The S4 includes the following S41. Design a resource importance evaluation function to calculate the importance weight of each resource node i :

[0074] in, is the Swish activation function, defined as , is a learnable parameter, The parameter is Multilayer perceptron, 、 、 、 、 They are the state vector, response characteristic vector, cost characteristic vector, distance characteristic vector, and energy availability index of resource i; S42. Calculate the Fisher information matrix of the resource parameters to quantify their importance to the model output:

[0075] in, is a parameter The Fisher information matrix of is the data distribution, is the model prediction probability, and the diagonal approximation is used to simplify the calculation:

[0076] S43, based on resource importance weight and Fisher information matrix , design resource layering processing strategies and set importance thresholds , classify resources into high importance , medium importance and low importance three floors; The high-importance resources are calculated using high-precision complete calculations, the medium-importance resources are calculated using simplified calculations, and the low-importance resources are calculated using warning calculations.

[0077] like Figure 5 As shown, the following steps are also included: S5 combines LoRA technology to implement low-rank adaptive factor optimization and expresses the self-calibration factor matrix as a pseudo-low-rank matrix of the identity matrix. Specifically, the following steps are included: S51. Optimize the self-calibration factor matrix using low-rank representation :

[0078] in, is the identity matrix, and is the low-order projection matrix, d is the low-order projection dimension, satisfying ; S52. Design an initialization strategy for the low-rank projection matrix. The matrix is ​​initialized using a Gaussian distribution with variance scaled to :

[0079] The A matrix is ​​initialized to a zero matrix, which is used to make the self-calibration factor matrix close to the identity matrix in the early stage of training; S53, introduce the conditional calculation dynamic activation mechanism, and selectively activate the low-rank air volume through the gating function:

[0080] in, is the optimized self-calibration factor matrix, is the gating function based on the input state x, k is the number of expert models, and is the low-rank projection matrix of the j-th expert model, and the gating function Using a lightweight multilayer perceptron:

[0081]

[0082] in, 、 、 and are the gating network parameters, is the input dimension, is a hidden dimension, is the temperature coefficient, which is used to adjust the smoothness of the activation probability distribution.

[0083] The following steps are also included: S6. Design a complete algorithm training and inference process, including phased training strategy, mixed precision training, and dynamic precision quantization. Specifically, the following steps are included: S61. Design a phased training strategy. Phase 1: Freeze the self-calibration factor matrix. , only train the supercore tensor and the factor matrix , learning basic data representation; Phase 2: Unfreeze the self-calibration factor matrix But freeze the super core tensor and the factor matrix , optimize the self-calibration capability; the third stage: unfreeze all parameters and perform end-to-end adjustment; S62. Design a loss function that combines multiple objectives, including reconstruction error, dilution constraint, energy balance, and cost optimization:

[0084] in, is the reconstruction loss, which measures the difference between the original tensor and the reconstructed tensor using the Frobenius norm:

[0085] is the sparsity loss, and L1 regularization is used to promote the sparsification of the self-calibration factor matrix:

[0086] is the energy balance loss, which is used to ensure the balance of energy supply and demand in the virtual power plant:

[0087] in, represents all control decisions of the i-th resource, 1 represents a full 1 vector, is the demand curve, is the cost optimization loss, which is used to minimize the total cost of resource scheduling:

[0088] in, is the cost coefficient for resource i to perform control k at time j, 、 、 and is the weight coefficient, which is used to control the relative importance of each loss item; Design an adaptive weight adjustment strategy:

[0089] in, is the weight coefficient of the t-th round of training, is the initial weight, is the decay rate, is the minimum weight; S63, adopt hybrid progress training to accelerate the calculation process, use FP half-precision floating point numbers for forward and backward propagation calculations, use FP32 single-precision floating point numbers to store main parameters and perform parameter updates, and design a dynamic loss scaling strategy to prevent gradient underflow:

[0090]

[0091] in, is the scaled loss, s is the scaling factor, and are the growth and reduction coefficients, set to 2 and 4 respectively.

[0092] The following steps are also included: S7. Improve the computational efficiency of the inference phase through dynamic precision quantization and sparse computing optimization. This includes the following steps: S71. Design dynamic precision quantification strategy based on resource importance weights Dynamically select calculation accuracy, high-precision resources Using FP32 progress, moderately important resources Use FP16 precision, low critical resources Use INT8 precision; design a quantization lookup table, pre-calculate the quantization results of commonly used values ​​to accelerate the quantization process, and use a symmetric quantization method for INT8 quantization:

[0093] Where q is the quantized value, x is the original floating-point value, s is the quantization scale factor, round indicates the rounding function, and the quantization scale factor S is adaptively calculated based on the data range:

[0094] S72. Using the self-calibration factor matrix The structured sparsity of the tensor is used to accelerate tensor operations by setting the threshold , set the elements close to zero to zero and increase the matrix dilution:

[0095] Then the diluted matrix is ​​stored in a compressed diluted row format to reduce memory usage; S73 combines multiple control instructions for batch processing to improve computing throughput.

[0096] S8: Optimize for heterogeneous resource scenarios and large-scale scenarios, design resource-type-specific factor matrix structures and hierarchical computing frameworks to improve system performance in complex environments.

[0097] S8.1: Design resource-specific factor matrix structures. Categorize virtual power plant resources into four categories: renewable energy (e.g., photovoltaic and wind power), energy storage, controllable loads, and traditional generators. Develop a customized factor matrix structure for each resource type. For renewable energy, the factor matrix emphasizes forecast error modeling; for energy storage, the factor matrix focuses on state-of-charge constraints; for controllable loads, the factor matrix focuses on user comfort and response speed; and for traditional generators, the factor matrix emphasizes ramping limits and minimum start / stop time constraints. Through specific structural design, improve the modeling accuracy of different resource types.

[0098] S8.2: Implement a hierarchical computing framework to handle resource clusters of varying sizes in layers. Large-scale virtual power plant resources are partitioned into multiple regions. Within each region, resources are modeled using local tensor networks, and regions are connected via a global coordination layer. Local tensor networks utilize high-precision computational details, while the global coordination layer focuses on key interactions to ensure optimal allocation of computing resources. Design an adaptive region partitioning algorithm to dynamically adjust the region partitioning based on factors such as resource location, electrical connectivity, and interaction frequency to optimize computational efficiency.

[0099] S8.3: Introduce dynamic environmental adaptation mechanisms to improve the system's adaptability in complex and changing environments. Design an online model update strategy to adaptively adjust the update frequency based on the degree of environmental change. Introduce incremental learning methods to retain historical experience while adapting to new environmental characteristics. Design an abnormal situation handling mechanism to trigger an emergency response mode when abnormal environmental changes are detected, and adopt a conservative control strategy to ensure system safety.

[0100] S9: Combined with Figure 6 The performance comparison results shown demonstrate the significant advantages of the proposed method over existing technologies in key metrics such as computation time, memory usage, control accuracy, and resource utilization, as well as its superlinear scalability with increasing node size. Tests were conducted in a large-scale virtual power plant scenario with 50,000 heterogeneous nodes. The results show that the proposed method reduced computation time from 135 seconds to 1.08 seconds (a 99.2% reduction), reduced memory usage from 64GB to 6.8GB (an 89.4% reduction), improved control accuracy from 92.1% to 99.4% (a 97.9% improvement), and increased resource utilization from 89.5% to 95.2%. Furthermore, the proposed method exhibits superlinear scalability, with computation time growing logarithmically with the number of nodes, rather than the linear or exponential growth of traditional methods. It is also highly adaptable to environmental changes and disturbances, capable of adapting the model to new operating conditions within 10 seconds after environmental changes.

[0101] The above are only embodiments of the present invention. Common knowledge such as the known specific structures and characteristics in the scheme are not described in detail here. Ordinary technicians in the field are aware of all common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all existing technologies in the field, and have the ability to apply conventional experimental means before that date. Ordinary technicians in the field can improve and implement this scheme in combination with their own abilities under the inspiration given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A virtual power plant resource optimization and scheduling system based on multi-head self-calibration tensor factorization, characterized by: The following steps are involved: S1. Collect and pre-process virtual power plant data, standardize, detect anomalies, and interpolate various heterogeneous resource data within the virtual power plant; S2: Construct a hierarchical gradient tensor network and perform multi-head self-calibration decomposition to establish a virtual power plant decision optimization model. S2 specifically includes the following steps: S21. Use tensor factorization method to decompose the high-dimensional control decision tensor into the product form of super core tensor and factor matrix: in, represents the high-dimensional control decision tensor, Indicates the key information after dimensionality reduction stored in the super core tensor, represents the steamed dumpling gene matrix, represents the self-calibration factor matrix, represents the tensor-matrix product along the i-th mode, Represents the dimensionality reduction coefficient of each dimension, satisfying ; S22. Build a gradient tensor network, represent the tensor factorization process as a computational graph, and achieve end-to-end differentiable optimization S22, multi-head self-calibration decomposition includes constructing a self-calibration factor matrix: in, represents the multi-head attention parameter matrix, Represents a sparse mask matrix based on resource characteristics, GELU represents the Gaussian error linear unit activation function, layerNorm represents the normalization operation, STE represents the pass-through estimator, Indicates that the parameter is The adaptive threshold function, represents the learnable weight matrix, represents the Hadamrd product. S3. Based on the virtual power plant decision optimization model, a multi-head self-calibration decomposition mechanism is designed to generate a virtual power plant resource optimization scheduling plan.

2. The virtual power plant resource optimization and scheduling system based on multi-head self-calibration tensor factorization according to claim 1 is characterized by: Said S1 comprises the following steps: S11. Collect the status information of various heterogeneous resources in the virtual power plant, including the state vector of each resource node i , response characteristic vector , cost characteristic vector , distance feature vector , and energy availability index , the state vector Indicates the current working state of the resource, including power generation power, energy storage power and load power. The response characteristic vector Represents the response characteristics of resources to control instructions, including ramp rate, response time and control accuracy. The cost characteristic vector Represents the economic characteristics of resources, including power generation cost, control cost, start-up and shutdown cost. The distance characteristic vector Indicates the relative position relationship of resources in the electrical network, energy availability index Indicates the predictable handling of renewable energy or the adjustable capacity of controllable load; S12. Data Adaptive Normalization: in, represents the standardized data, x represents the original data, represents the data mean, represents the standard deviation of the data, A small positive number used to prevent the denominator from being 0; S1.3 Constructing high-dimensional decision tensors ,in Indicates the resource quantity dimension, Represents the time dimension.

3. The virtual power plant resource optimization and scheduling system based on multi-head self-calibration tensor factorization according to claim 1 is characterized by: The S22 includes the following steps: S221, forward Zhuang Bo calculation reconstruction tensor : S222, back propagation calculates gradients and updates network parameters: in, represents the loss function, Represents a supercore tensor The expansion matrix on the i-th mode, represents the Kronecker product.

4. The virtual power plant resource optimization and scheduling system based on multi-head self-calibration tensor factorization according to claim 2 is characterized by: The S3 includes the following steps: S31 builds self-calibration factor matrix : in, 、 、 is the parameter matrix in the multi-head attention mechanism, is the dimension of the attention head, It is a sparse mask matrix based on resource characteristics, GELU represents the Gaussian error linear unit activation function, and LayerNorm represents the layer normalization operation. represents the Hadamard product, STE represents the straight-through estimation, is a parameter The adaptive threshold function, is a learnable weight matrix; S32. Design a multi-head attention mechanism to divide the self-calibration factor mean into H independent attention heads, each of which focuses on modeling the interaction of a specific type of resource: in, 、 、 is the parameter matrix of the h-th attention head, is the sparse mask matrix of the h-th attention head, is the output projection matrix; S33. Design mask sparse matrix , based on three different resource relationships to establish sparse connections, physical connection relationship mask Based on the topological connection relationship of resources in the power network, historical interaction mask Generate a time-dependent mask based on the frequency of interactions between resources in historical data Based on the resource response time design, the mask matrix is ​​finally obtained through weighted fusion: in, 、 、 is a learnable weight coefficient that satisfies and .

5. The virtual power plant resource optimization and scheduling system based on multi-head self-calibration tensor factorization according to claim 1 is characterized in that: The following steps are also included: S4. Introduce a resource importance adaptive evaluation mechanism based on Fisher information to identify and prioritize key resources.

6. The virtual power plant resource optimization and scheduling system based on multi-head self-calibration tensor factorization according to claim 5 is characterized by: The S4 comprises the following steps: S41. Design a resource importance evaluation function to calculate the importance weight of each resource node i : in, is the Swish activation function, defined as , is a learnable parameter, The parameter is Multilayer perceptron, 、 、 、 、 They are the state vector, response characteristic vector, cost characteristic vector, distance characteristic vector, and energy availability index of resource i; S42. Calculate the Fisher information matrix of the resource parameters to quantify their importance to the model output: in, is a parameter The Fisher information matrix of is the data distribution, is the model prediction probability, and the diagonal approximation is used to simplify the calculation: S43, based on resource importance weight and Fisher information matrix , design resource layering processing strategies and set importance thresholds , classify resources into high importance , medium importance and low importance three floors; The high-importance resources are calculated using high-precision complete calculations, the medium-importance resources are calculated using simplified calculations, and the low-importance resources are calculated using warning calculations.

7. The virtual power plant resource optimization and scheduling system based on multi-head self-calibration tensor factorization according to claim 6 is characterized by: The following steps are also included: S5 combines LoRA technology to implement low-rank adaptive factor optimization and expresses the self-calibration factor matrix as a pseudo-low-rank matrix of the identity matrix. Specifically, the following steps are included: S51. Optimize the self-calibration factor matrix using low-rank representation : in, is the identity matrix, and is the low-order projection matrix, d is the low-order projection dimension, satisfying ; S52. Design an initialization strategy for the low-rank projection matrix. The matrix is ​​initialized using a Gaussian distribution with variance scaled to : The A matrix is ​​initialized to a zero matrix, which is used to make the self-calibration factor matrix close to the identity matrix in the early stage of training; S53, introduce the conditional calculation dynamic activation mechanism, and selectively activate the low-rank air volume through the gating function: in, is the optimized self-calibration factor matrix, is the gating function based on the input state x, k is the number of expert models, and is the low-rank projection matrix of the j-th expert model, and the gating function Using a lightweight multilayer perceptron: in, 、 、 and are the gating network parameters, is the input dimension, is a hidden dimension, is the temperature coefficient, which is used to adjust the smoothness of the activation probability distribution.

8. The virtual power plant resource optimization and scheduling system based on multi-head self-calibration tensor factorization according to claim 7 is characterized by: The following steps are also included: S6. Design a complete algorithm training and inference process, including phased training strategy, mixed precision training, and dynamic precision quantization. Specifically, the following steps are included: S61. Design a phased training strategy. Phase 1: Freeze the self-calibration factor matrix. , only train the supercore tensor and the factor matrix , learning basic data representation; Phase 2: Unfreeze the self-calibration factor matrix But freeze the super core tensor and the factor matrix , optimize the self-calibration capability; the third stage: unfreeze all parameters and perform end-to-end adjustment; S62. Design a loss function that combines multiple objectives, including reconstruction error, dilution constraint, energy balance, and cost optimization: in, is the reconstruction loss, which measures the difference between the original tensor and the reconstructed tensor using the Frobenius norm: is the sparsity loss, and L1 regularization is used to promote the sparsification of the self-calibration factor matrix: is the energy balance loss, which is used to ensure the balance of energy supply and demand in the virtual power plant: in, represents all control decisions of the i-th resource, 1 represents a full 1 vector, is the demand curve, is the cost optimization loss, which is used to minimize the total cost of resource scheduling: in, is the cost coefficient for resource i to perform control k at time j, 、 、 and is the weight coefficient, which is used to control the relative importance of each loss item; Design an adaptive weight adjustment strategy: in, is the weight coefficient of the t-th round of training, is the initial weight, is the decay rate, is the minimum weight; S63, adopt hybrid progress training to accelerate the calculation process, use FP half-precision floating point numbers for forward and backward propagation calculations, use FP32 single-precision floating point numbers to store main parameters and perform parameter updates, and design a dynamic loss scaling strategy to prevent gradient underflow: in, is the scaled loss, s is the scaling factor, and are the growth and reduction coefficients, set to 2 and 4 respectively.

9. The virtual power plant resource optimization and scheduling system based on multi-head self-calibration tensor factorization according to claim 8 is characterized by: The following steps are also included: S7. Improve the computational efficiency of the inference phase through dynamic precision quantization and sparse computing optimization. This includes the following steps: S71. Design dynamic precision quantification strategy based on resource importance weights Dynamically select calculation accuracy, high-precision resources Using FP32 progress, moderately important resources Use FP16 precision, low critical resources Use INT8 precision; design a quantization lookup table, pre-calculate the quantization results of commonly used values ​​to accelerate the quantization process, and use a symmetric quantization method for INT8 quantization: Where q is the quantized value, x is the original floating-point value, s is the quantization scale factor, round indicates the rounding function, and the quantization scale factor S is adaptively calculated based on the data range: S72. Using the self-calibration factor matrix The structured sparsity of the tensor is used to accelerate tensor operations by setting the threshold , set the elements close to zero to zero and increase the matrix dilution: Then the diluted matrix is ​​stored in a compressed diluted row format to reduce memory usage; S73 combines multiple control instructions for batch processing to improve computing throughput.

Images