CN-122026423-A - Optical storage charging micro-grid scheduling method and system

CN122026423ACN 122026423 ACN122026423 ACN 122026423ACN-122026423-A

Abstract

The invention discloses a dispatching method and a dispatching system for an optical storage and charging micro-grid, wherein the method comprises the steps of obtaining photovoltaic power generation data, energy storage state data, charging demand data and power grid operation data, and performing data cleaning and feature engineering processing to obtain standardized feature vectors; the method comprises the steps of constructing a neural network search space by utilizing a framework embedding and transferring ranking method to obtain an optimal network structure aiming at different prediction tasks, modeling the dependency relationship between a classification variable and a continuous variable by adopting a multi-head self-attention layer to obtain a mixed variable optimization model, and carrying out multi-objective optimization solution by adopting an improved NSGA-III algorithm to obtain a scheduling strategy of an optical storage and filling micro-grid. According to the invention, the prediction precision is improved through automatic optimization of the framework, the interaction among the mixed variables is effectively processed, the collaborative optimization of energy balance, power grid support and charging requirements is realized, and the method has good practicability and economic benefit.

Inventors

LI BINGYAN
XU XIAODONG
YUAN YU
JI YANG

Assignees

国网江苏省电力有限公司盐城供电分公司
国网江苏省电力有限公司射阳县供电分公司
射阳县电气实业有限公司

Dates

Publication Date: 20260512
Application Date: 20260313

Claims (10)

1. The method for dispatching the optical storage and filling micro-grid is characterized by comprising the following steps of: acquiring photovoltaic power generation data, energy storage state data, charging demand data and power grid operation data, and cleaning and feature engineering processing the photovoltaic power generation data, the energy storage state data, the charging demand data and the power grid operation data to obtain standardized feature vectors; Based on the standardized feature vector, constructing a neural network search space by utilizing a framework embedding and transfer ranking method to obtain an optimal network structure aiming at different prediction tasks; modeling the dependency relationship between the classification variable and the continuous variable by adopting the multi-head self-attention layer according to the optimal network structure to obtain a mixed variable optimization model; And based on the mixed variable optimization model, performing multi-objective optimization solution on energy balance constraint, power grid support indexes and charging demand satisfaction by adopting an improved NSGA-III algorithm to obtain a scheduling strategy of the optical storage charging micro-grid, and performing optical storage charging micro-grid scheduling based on the scheduling strategy.
2. The method of claim 1, wherein cleaning and feature engineering the photovoltaic power generation data, energy storage state data, charging demand data, and grid operation data to obtain the normalized feature vector comprises: Performing outlier detection and missing value processing on the data by adopting a sliding time window to obtain preprocessed data; extracting time sequence statistical characteristics of the preprocessed data to obtain time sequence characteristic data; Carrying out characteristic association on the time sequence characteristic data and the meteorological data and the temperature data to obtain an association characteristic matrix; and carrying out Min-Max standardization processing on the association characteristic matrix to obtain the standardized characteristic vector.
3. The method of claim 1, wherein constructing a neural network search space using the architecture embedding and transfer ranking method results in the optimal network structure for different prediction tasks, comprising: constructing a neural network search space based on a candidate operation set comprising a convolution layer, a circulation unit and an attention module, wherein the convolution layer comprises input dimension parameters and output dimension parameters, and the circulation unit comprises hidden state dimension parameters to obtain an initial search space; Converting the network structure in the initial search space into directed acyclic graph representation to obtain graph structure data; and coding the graph structure data by using a graph neural network to obtain continuous vectors with fixed dimensions.
4. A method according to claim 3, characterized in that the method further comprises: constructing a Siamese network by using the continuous vectors with fixed dimensionality, and training a historical architecture evaluation result to obtain an architecture performance ordering model; And optimizing in the search space by adopting a strategy gradient method based on the architecture performance sequencing model to obtain the optimal network structure aiming at different prediction tasks.
5. The method of claim 1, wherein modeling the dependency relationship between the classification variable and the continuous variable using the multi-headed self-attention layer to obtain the hybrid variable optimization model comprises: Performing single-heat coding on the classified variables, and performing normalization processing on the continuous variables to obtain coded input variables; projecting the encoded input variable to a query space, a key space and a value space to obtain a projection variable; Calculating the attention weight of the projection variable by using scaled dot-product attention to obtain the attention characteristic; and carrying out multi-layer cross feature learning on the attention features to obtain the mixed variable optimization model.
6. The method of claim 5, wherein performing multi-layer cross feature learning on the attention features to obtain the hybrid variable optimization model comprises: constructing a multi-layer cross network structure, and calculating high-order interaction characteristics among variables to obtain the interaction characteristics; dynamically adjusting the importance of the interaction characteristics by adopting a gating mechanism to obtain optimized interaction characteristics; fusing the optimized interaction characteristics with the attention characteristics through residual connection to obtain fusion characteristics; And carrying out layer normalization processing on the fusion characteristics to obtain the mixed variable optimization model.
7. The method of claim 1, wherein performing a multi-objective optimization solution to the energy balance constraint, the grid support index, and the charging demand satisfaction using the modified NSGA-III algorithm to obtain a scheduling policy for the optical storage and charging micro-grid comprises: Constructing an energy balance objective function comprising power balance constraint and energy storage capacity constraint, and taking the equipment operation limit and the network topology constraint as constraint conditions to obtain a complete optimization problem; Based on the optimization problem, a reference point mechanism and crowding degree sequencing are adopted to conduct Pareto front search, and a Pareto solution set is obtained; And generating the scheduling strategy based on the Pareto solution set.
8. The method of claim 7, wherein generating the scheduling policy based on the Pareto solution set comprises: determining the priority of each optimization target by adopting a fuzzy analytic hierarchy process to obtain target weight; Based on the target weight, performing scheme evaluation and sequencing from the Pareto solution set by using a TOPSIS method to obtain an optimal solution; And generating a scheduling instruction meeting actual operation requirements according to the optimal solution to obtain a scheduling strategy of the optical storage and filling micro-grid.
9. The method according to claim 1, characterized in that: the standardized feature vector comprises a mean feature extracted from the photovoltaic power generation data, a variance feature extracted from the energy storage state data and a peak feature extracted from the charging demand data; The power grid support index comprises a voltage stability index and a power factor index in the power grid operation data; the charging demand satisfaction includes a charging time index and a power demand index.
10. An optical storage and filling micro-grid scheduling device, which is characterized by comprising: The preprocessing module is used for acquiring photovoltaic power generation data, energy storage state data, charging demand data and power grid operation data, and cleaning and characteristic engineering processing the data to obtain standardized characteristic vectors; The architecture searching module is used for constructing a neural network searching space by utilizing the architecture embedding and transferring ranking method based on the standardized feature vector to obtain the optimal network structure aiming at different prediction tasks; the optimization modeling module is used for modeling the dependency relationship between the classification variable and the continuous variable by adopting the multi-head self-attention layer according to the optimal network structure to obtain the mixed variable optimization model; And the strategy generation module is used for carrying out multi-objective optimization solution on the energy balance constraint, the power grid support index and the charging demand satisfaction degree by adopting the improved NSGA-III algorithm based on the mixed variable optimization model to obtain a scheduling strategy of the optical storage and charging micro-grid.

Description

Optical storage charging micro-grid scheduling method and system Technical Field The invention relates to the technical field of new energy control, in particular to a distributed optical storage and charging control method and system. Background The light storage and charging micro-grid system is used as a novel energy management system, and plays an important role in promoting renewable energy consumption and supporting stable operation of a power grid by integrating photovoltaic power generation, energy storage equipment and an electric vehicle charging station. The conventional micro-grid scheduling technology mainly adopts a rule-based scheduling strategy or a traditional optimization algorithm. For example, a time-of-use electricity price guided energy storage charge and discharge control method or a photovoltaic-energy storage combined optimization method based on linear programming is adopted. The advanced technology adopts a deep learning method to construct a prediction model, takes photovoltaic power generation prediction, load prediction and charging demand prediction results as input, and obtains a scheduling strategy through neural network training. The method can better process the nonlinear characteristics of the system and improve the prediction accuracy. However, the prior art has the problems that 1) a fixed neural network structure is adopted, so that the optimization requirements under different scenes are difficult to adapt, and 2) when the optimization problem of mixed classification-continuous variables is processed, interaction among the variables is difficult to fully consider, and the optimization effect is influenced. Disclosure of Invention The invention aims to provide an optical storage and filling micro-grid scheduling method and system, which are used for solving the technical problems that the neural network structure is fixed and is difficult to adapt to different scene optimization requirements in the prior art, and the interaction between variables is difficult to fully consider when the mixed classification-continuous variable optimization problem is processed. In order to achieve the above purpose, the present invention provides the following technical solutions: an optical storage charging micro-grid scheduling method comprises the following steps: acquiring photovoltaic power generation data, energy storage state data, charging demand data and power grid operation data, and cleaning and feature engineering processing the photovoltaic power generation data, the energy storage state data, the charging demand data and the power grid operation data to obtain standardized feature vectors; Based on the standardized feature vector, constructing a neural network search space by utilizing a framework embedding and transfer ranking method to obtain an optimal network structure aiming at different prediction tasks; modeling the dependency relationship between the classification variable and the continuous variable by adopting the multi-head self-attention layer according to the optimal network structure to obtain a mixed variable optimization model; And based on the mixed variable optimization model, performing multi-objective optimization solution on energy balance constraint, power grid support indexes and charging demand satisfaction by adopting an improved NSGA-III algorithm to obtain a scheduling strategy of the optical storage charging micro-grid, and performing optical storage charging micro-grid scheduling based on the scheduling strategy. Optionally, cleaning and feature engineering processing are performed on the photovoltaic power generation data, the energy storage state data, the charging demand data and the power grid operation data to obtain the standardized feature vector, including: Performing outlier detection and missing value processing on the data by adopting a sliding time window to obtain preprocessed data; extracting time sequence statistical characteristics of the preprocessed data to obtain time sequence characteristic data; Carrying out characteristic association on the time sequence characteristic data and the meteorological data and the temperature data to obtain an association characteristic matrix; and carrying out Min-Max standardization processing on the association characteristic matrix to obtain the standardized characteristic vector. Optionally, constructing a neural network search space by using the architecture embedding and transferring ranking method to obtain the optimal network structure for different prediction tasks, including: constructing a neural network search space based on a candidate operation set comprising a convolution layer, a circulation unit and an attention module, wherein the convolution layer comprises input dimension parameters and output dimension parameters, and the circulation unit comprises hidden state dimension parameters to obtain an initial search space; Converting the network structure in the initial search s