Optical storage and charging micro-grid scheduling method and system

Abstract

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

Description

Technical Field

[0001] This invention relates to the field of new energy control technology, and in particular to a distributed photovoltaic energy storage and charging control method and system. Background Technology

[0002] As a new type of energy management system, the photovoltaic-storage-charging microgrid system plays an important role in promoting the consumption of renewable energy and supporting the stable operation of the power grid by integrating photovoltaic power generation, energy storage equipment and electric vehicle charging stations.

[0003] Currently, common microgrid scheduling technologies mainly employ rule-based scheduling strategies or traditional optimization algorithms. Examples include time-of-use pricing-oriented energy storage charging and discharging control, or photovoltaic-energy storage joint optimization methods based on linear programming.

[0004] More advanced technologies employ deep learning methods to construct predictive models, using photovoltaic power generation forecasts, load forecasts, and charging demand forecasts as inputs, and training neural networks to derive scheduling strategies. This method can better handle the nonlinear characteristics of the system and improve prediction accuracy.

[0005] However, existing technologies have the following problems: 1) They use fixed neural network structures, which are difficult to adapt to the optimization needs of different scenarios; 2) When dealing with optimization problems involving mixed classification and continuous variables, it is difficult to fully consider the interaction between variables, which affects the optimization effect. Summary of the Invention

[0006] The purpose of this invention is to provide a scheduling method and system for a photovoltaic-storage-charging microgrid, in order to solve the technical problems in the prior art where the fixed neural network structure is difficult to adapt to the optimization needs of different scenarios, and where it is difficult to fully consider the interaction between variables when dealing with mixed classification-continuous variable optimization problems.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for scheduling a photovoltaic-storage-charging microgrid includes:

[0009] Acquire photovoltaic power generation data, energy storage status data, charging demand data, and grid operation data; clean and perform feature engineering on the photovoltaic power generation data, energy storage status data, charging demand data, and grid operation data to obtain standardized feature vectors.

[0010] Based on the standardized feature vectors, the neural network search space is constructed using architecture embedding and transition ranking methods to obtain the optimal network structure for different prediction tasks.

[0011] Based on the optimal network structure, the multi-head self-attention layer is used to model the dependency relationship between categorical and continuous variables to obtain a mixed variable optimization model.

[0012] Based on the aforementioned mixed-variable optimization model, the improved NSGA-III algorithm is used to perform multi-objective optimization solutions for energy balance constraints, grid support indicators, and charging demand satisfaction, thereby obtaining the scheduling strategy for the photovoltaic-storage-charging microgrid. The scheduling of the photovoltaic-storage-charging microgrid is then executed based on the scheduling strategy.

[0013] Optionally, the photovoltaic power generation data, energy storage status data, charging demand data, and grid operation data are cleaned and feature-engineered to obtain the standardized feature vector, including:

[0014] The data is preprocessed by using a sliding time window to detect outliers and handle missing values.

[0015] Temporal statistical features are extracted from the preprocessed data to obtain temporal feature data;

[0016] The time-series feature data is correlated with meteorological data and temperature data to obtain a correlation feature matrix;

[0017] The associated feature matrix is ​​subjected to Min-Max normalization to obtain the normalized feature vector.

[0018] Optionally, the neural network search space is constructed using the architecture embedding and transition ranking methods to obtain the optimal network structure for different prediction tasks, including:

[0019] The neural network search space is constructed based on a candidate operation set containing convolutional layers, recurrent units, and attention modules. The convolutional layers include input dimension and output dimension parameters, and the recurrent units include hidden state dimension parameters, thus obtaining the initial search space.

[0020] The network structure in the initial search space is converted into a directed acyclic graph representation to obtain graph structure data;

[0021] The graph structure data is encoded using a graph neural network to obtain a continuous vector of fixed dimensions.

[0022] Optionally, the method further includes:

[0023] A Siamese network is constructed using the fixed-dimensional continuous vectors, and trained on historical architecture evaluation results to obtain an architecture performance ranking model.

[0024] Based on the architecture performance ranking model, the policy gradient method is used to optimize the search space to obtain the optimal network structure for different prediction tasks.

[0025] Optionally, the multi-head self-attention layer is used to model the dependency relationship between the categorical variable and the continuous variable to obtain the mixed variable optimization model, including:

[0026] The categorical variables are one-hot encoded, and the continuous variables are normalized to obtain the encoded input variables.

[0027] The encoded input variables are projected onto the query space, key space, and value space to obtain the projected variables;

[0028] The attention weights of the projected variables are calculated using scaled dot-product attention to obtain the attention features;

[0029] The attention features are subjected to multi-layer cross-feature learning to obtain the mixed variable optimization model.

[0030] Optionally, multi-layer cross-feature learning is performed on the attention features to obtain the mixed-variable optimization model, including:

[0031] Construct a multi-layered cross-network structure, calculate the higher-order interaction features between variables, and obtain the interaction features;

[0032] A gating mechanism is used to dynamically adjust the importance of the interaction features to obtain optimized interaction features;

[0033] The optimized interaction features are fused with the attention features through residual connections to obtain fused features;

[0034] The fused features are subjected to layer normalization to obtain the hybrid variable optimization model.

[0035] Optionally, the improved NSGA-III algorithm is used to perform multi-objective optimization on the energy balance constraint, the grid support index, and the charging demand satisfaction, to obtain the scheduling strategy of the photovoltaic-storage-charging microgrid, including:

[0036] Construct an energy balance objective function that includes power balance constraints and energy storage capacity constraints, and take the device operation limitations and the network topology constraints as constraints to obtain a complete optimization problem;

[0037] Based on the aforementioned optimization problem, a Pareto front search is performed using a reference point mechanism and crowding ordering to obtain the Pareto solution set.

[0038] The scheduling policy is generated based on the Pareto solution set.

[0039] Optionally, generating the scheduling policy based on the Pareto solution set includes:

[0040] The priority of each optimization objective is determined by fuzzy hierarchical analysis, and the objective weights are obtained.

[0041] Based on the target weights, the TOPSIS method is used to evaluate and rank the solutions from the Pareto solution set to obtain the optimal solution.

[0042] Based on the optimal solution, a scheduling instruction that meets the actual operational requirements is generated, thus obtaining the scheduling strategy of the optical storage and charging microgrid.

[0043] Optionally, the standardized feature vector includes the mean feature extracted from the photovoltaic power generation data, the variance feature extracted from the energy storage status data, and the peak feature extracted from the charging demand data.

[0044] The power grid support indicators include voltage stability indicators and power factor indicators in the power grid operation data;

[0045] The charging demand satisfaction includes charging time indicators and power demand indicators.

[0046] A photovoltaic-storage-charging microgrid scheduling device, comprising:

[0047] The preprocessing module is used to acquire photovoltaic power generation data, energy storage status data, charging demand data and grid operation data, and to clean and perform feature engineering on the data to obtain standardized feature vectors.

[0048] An architecture search module is used to construct a neural network search space based on the standardized feature vectors using the architecture embedding and transition ranking methods, thereby obtaining the optimal network structure for different prediction tasks.

[0049] The optimization modeling module is used to model the dependency relationship between the categorical variables and the continuous variables using the multi-head self-attention layer based on the optimal network structure, so as to obtain the mixed variable optimization model.

[0050] The strategy generation module is used to perform multi-objective optimization on the energy balance constraint, the grid support index and the charging demand satisfaction based on the mixed variable optimization model and the improved NSGA-III algorithm, so as to obtain the scheduling strategy of the photovoltaic-storage-charging microgrid.

[0051] The beneficial effects of this invention are:

[0052] 1. This invention employs a multi-task neural architecture search method that combines architecture embedding and transition ranking, enabling automatic optimization of network structures for different prediction tasks and improving the model's adaptability and prediction accuracy.

[0053] 2. This invention designs a mixed variable optimization model based on an attention mechanism. Through a multi-head self-attention layer and a cross-feature network, it effectively handles the interaction between categorical and continuous variables, thereby improving the accuracy of variable modeling.

[0054] 3. This invention proposes a multi-objective joint optimization framework. Through the improved NSGA-III algorithm, it achieves coordinated optimization of energy balance, grid support and charging demand, thereby improving the overall operating efficiency of the microgrid system. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0056] Figure 1 The flowchart of the photovoltaic-storage-charging microgrid scheduling method provided by the present invention is shown below.

[0057] Figure 2 The data preprocessing and feature extraction flowchart provided by this invention;

[0058] Figure 3 The flowchart of the multi-task neural architecture search method provided by the present invention is shown below.

[0059] Figure 4 The structure diagram of the mixed variable optimization model provided by this invention;

[0060] Figure 5 The flowchart for solving the multi-objective joint optimization problem provided by this invention;

[0061] Figure 6 The diagram shows the structure of the photovoltaic-storage-charging microgrid scheduling system provided by this invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0063] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0064] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0065] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0067] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, features, advantages, or spirit of the various embodiments of the invention described below can be combined with each other without conflict.

[0068] like Figure 1 As shown, the present invention provides a method for scheduling a photovoltaic-storage-charging microgrid, comprising the following steps:

[0069] Step S1: Obtain photovoltaic power generation data, energy storage status data, charging demand data, and grid operation data; clean and perform feature engineering processing on the photovoltaic power generation data, energy storage status data, charging demand data, and grid operation data to obtain standardized feature vectors.

[0070] Step S1 primarily processes the various types of data required by the system. The system first acquires photovoltaic (PV) power generation data (such as real-time output power of PV panels and historical power generation), energy storage status data (such as the state of charge of battery packs and charge / discharge efficiency), charging demand data (such as charging requests from electric vehicles and estimated charging time), and grid operation data (such as grid voltage and frequency parameters). After acquiring this raw data, the system performs a series of data cleaning tasks, including removing outliers, filling in missing values, and eliminating noise. Next, feature engineering is performed to extract valuable features from the raw data, such as extracting intraday variation patterns from PV power generation data and charge / discharge cycle features from energy storage status data. Finally, these heterogeneous features are standardized to the same numerical range, generating standardized feature vectors that provide a foundation for subsequent modeling.

[0071] Step S2: Based on the standardized feature vectors, construct the neural network search space using architecture embedding and transition ranking methods to obtain the optimal network structure for different prediction tasks;

[0072] Step S2 focuses on automatically finding the optimal neural network structure. Based on the standardized feature vectors obtained in step S1, the system constructs the neural network search space using architecture embedding and transition ranking methods. Architecture embedding transforms different network structures into continuous vector representations, allowing optimization in a continuous space; the transition ranking method guides the search process by comparing the relative performance of different architectures, reducing the need for large amounts of computational resources. Through this method, the system can automatically search for and determine the most suitable neural network structure for different tasks such as photovoltaic power generation prediction, energy storage scheduling, and charging demand assessment, avoiding the limitations of manually designing network structures and improving the model's prediction accuracy and adaptability.

[0073] Step S3: Based on the optimal network structure, the multi-head self-attention layer is used to model the dependency relationship between categorical and continuous variables to obtain a mixed variable optimization model;

[0074] Based on the optimal network structure obtained in step S2, the system employs a multi-head self-attention mechanism to model different types of variables. In the microgrid scheduling problem, there are both categorical variables such as device on / off states and continuous variables such as power values, making it difficult for traditional methods to effectively handle the complex interactions between them. The multi-head self-attention layer allows the model to simultaneously focus on multiple interaction patterns between different types of variables, calculating the correlation weights between variables through a query-key-value mechanism. The system also utilizes a cross-feature network to further capture higher-order interaction features between variables, ultimately constructing a mixed-variable optimization model capable of accurately modeling the dependencies between categorical and continuous variables.

[0075] Step S4: Based on the hybrid variable optimization model, the improved NSGA-III algorithm is used to perform multi-objective optimization solutions for energy balance constraints, grid support indicators and charging demand satisfaction, so as to obtain the scheduling strategy of the photovoltaic-storage-charging microgrid, and the scheduling of the photovoltaic-storage-charging microgrid is executed based on the scheduling strategy.

[0076] Based on the mixed-variable optimization model constructed in the previous step, an improved NSGA-III algorithm is used to solve the multi-objective optimization problem. In microgrid scheduling, three potentially conflicting objectives need to be considered simultaneously: energy balance (ensuring a balance between power generation and consumption), grid support (ensuring stable grid operation), and charging demand satisfaction (meeting the charging needs of electric vehicle users). The improved NSGA-III algorithm, through a reference point mechanism and congestion ranking method, can effectively search for multiple equilibrium non-dominated solutions, forming a Pareto front. The system then selects the most suitable scheduling strategy from these solutions based on the decision-maker's preferences and translates it into specific scheduling instructions to guide the real-time coordinated operation of photovoltaic power generation equipment, energy storage systems, and charging stations, achieving efficient operation of the entire microgrid system.

[0077] By organically combining these four steps, this method realizes intelligent scheduling of photovoltaic-storage-charging microgrids, overcomes the limitations of traditional methods in handling complex nonlinear relationships and multi-objective optimization, and improves the overall operating efficiency of the system and the capacity for renewable energy absorption.

[0078] Exemplarily, in another embodiment,

[0079] In step S1, the system first acquires various operational data of the photovoltaic-storage-charging microgrid, including photovoltaic power generation data, energy storage status data, charging demand data, and grid operation data. Preprocessing this data forms the basis for subsequent modeling and optimization.

[0080] Specifically, the system employs a sliding time window technique to detect outliers and handle missing values ​​in the data. For example, for photovoltaic power generation data, the system sets a 24-hour sliding window to identify outliers that deviate from historical data for the same period by more than 30%; for energy storage status data, the system detects abrupt changes in charging and discharging states; and for missing values, it uses methods such as linear interpolation, mean filling, or previous value filling to repair them, depending on the data type, to obtain preprocessed data.

[0081] Next, the system extracts time-series statistical features from the preprocessed data. For photovoltaic power generation data, it extracts statistical features such as mean, variance, and maximum value at different time scales (hour, day, week); for energy storage status data, it extracts time-series features such as charge-discharge cycle characteristics and capacity change rate; and for charging demand data, it extracts features such as peak periods and duration, forming time-series feature data.

[0082] Then, the system correlates the time-series feature data with external environmental data (such as meteorological data and temperature data) to construct a correlation feature matrix. For example, it correlates photovoltaic power generation data with meteorological factors such as irradiance and cloud cover; and it correlates charging demand with environmental factors such as temperature and precipitation, thereby enhancing the expressive power of the data.

[0083] Finally, the correlation feature matrix is ​​subjected to Min-Max standardization to unify features of different dimensions into the [0,1] interval, resulting in standardized feature vectors, which lay the foundation for subsequent modeling and optimization.

[0084] In step S2, the system constructs a neural network search space based on standardized feature vectors using architecture embedding and transition ranking methods, thereby achieving automatic optimization of the network structure for different prediction tasks.

[0085] Specifically, the system first constructs a neural network search space based on a set of candidate operations including convolutional layers, recurrent units, and attention modules. For example, convolutional layers include operations such as 1D convolution (input dimension 32-128, output dimension 64-256) and 2D convolution; recurrent units include operations such as LSTM (hidden state dimension 64-256) and GRU; and attention modules include operations such as self-attention and cross-attention, forming the initial search space.

[0086] Next, the system transforms the network structure in the initial search space into a directed acyclic graph representation. Each node represents an operation (such as a convolutional layer, LSTM unit, etc.), and the edges represent the direction of data flow, forming a complete neural network computation graph, thus obtaining graph structure data.

[0087] Then, the system uses a graph neural network to encode the graph structure data. Graph Convolutional Networks (GCNs) or Graph Attention Networks (GATs) are used to aggregate the features of nodes and edges, encoding the entire network architecture into a fixed-dimensional continuous vector (e.g., 128-dimensional), preserving the topological information and operational features of the original architecture.

[0088] After obtaining the continuous vector representation of the architecture, the system constructs a Siamese network to train on the historical architecture evaluation results. This network takes the embedding vectors of two architectures, outputs their performance ranking relationship, and optimizes it through a triplet loss function to obtain a ranking model that can predict architecture performance.

[0089] Finally, based on the architecture performance ranking model, the system employs a policy gradient method to optimize within the search space. The architecture search problem is modeled as a Markov decision process, and Monte Carlo tree search is used to enhance the exploration effect, ultimately outputting the optimal network structure for different tasks such as photovoltaic power generation prediction, energy storage scheduling prediction, and charging demand prediction.

[0090] In step S3, based on the optimal network structure obtained in step S2, the system uses a multi-head self-attention layer to model the dependency relationship between categorical and continuous variables, and constructs a mixed variable optimization model.

[0091] Specifically, the system first preprocesses categorical and continuous variables. For categorical variables (such as device on / off status, operating mode, etc.), one-hot encoding is used to convert them into binary vectors; for continuous variables (such as power values, voltage, etc.), z-score standardization or min-max scaling is used for normalization to obtain the encoded input variables.

[0092] Next, the system projects the encoded input variables into the query space, key space, and value space. For the input variable X, the projected variables are obtained by calculating Q = XWq, K = XWk, and V = XWv using the three projection matrices Wq, Wk, and Wv, respectively.

[0093] Then, the system uses scaled dot-product attention to calculate the attention weights of the projected variables. The specific calculation formula is: Attention(Q,K,V)=softmax(QK) T / d k )V, where d k This refers to the dimension of the key vector. Through a multi-head mechanism (such as 8 heads), the system can simultaneously focus on different types of variable interaction patterns, thus obtaining attention features.

[0094] After obtaining the attention features, the system performs multi-layer cross-feature learning. First, a multi-layer cross-network structure is constructed, with each layer calculating second-order or higher-order interaction features between variables. For example, for variable x... i and x j Calculate their interaction features x i ⊙x j The interactive features are then transformed through a fully connected layer.

[0095] Next, the system employs a gating mechanism to dynamically adjust the importance of interaction features. Weights in the [0,1] interval are generated using the sigmoid activation function, and the influence of the interaction features is adjusted element-wise to obtain the optimized interaction features.

[0096] Then, the system fuses the optimized interaction features with the attention features through residual connections. The residual connection formula is: Y = X + F(X), where X is the original attention feature and F(X) is the feature learned through the cross-network to obtain the fused feature.

[0097] Finally, the system performs layer normalization on the fused features. By adjusting the feature distribution through normalization, the training stability and generalization ability of the model are improved, resulting in the final mixed variable optimization model.

[0098] In step S4, the system uses a mixed variable optimization model and an improved NSGA-III algorithm to perform multi-objective optimization solutions for energy balance constraints, grid support indicators, and charging demand satisfaction, thereby obtaining the scheduling strategy for the photovoltaic-storage-charging microgrid.

[0099] Specifically, the system first constructs an energy balance objective function that includes power balance constraints and energy storage capacity constraints. The power balance constraint ensures that the total power generation of the system equals the total load, which can be expressed as: P pv (t)+P ess (t)=P load (t)+P ev (t), where P pv For photovoltaic power generation, P ess For energy storage capacity, P load Based on the load, P ev For charging load; energy storage capacity constraints ensure that the state of charge (SOC) of energy storage devices is within a safe range. min ≤SOC(t)≤SOC max Meanwhile, by incorporating equipment operating limitations (such as ramp rate and maximum power) and network topology constraints, a complete optimization problem is obtained.

[0100] Next, based on the optimization problem, the system employs a reference point mechanism and crowding ordering for Pareto front search. The reference point mechanism guides population evolution by using reference points uniformly distributed in the target space, while crowding ordering ensures the diversity of the solution set. In each generation of evolution, the system selects non-dominated solutions to form a new population and applies crossover and mutation operations to generate offspring, gradually approaching the Pareto front and ultimately obtaining the Pareto solution set.

[0101] After obtaining the Pareto solution set, the system needs to select the final scheduling strategy based on actual needs. Fuzzy hierarchical analysis is used to determine the priority of each optimization objective. By constructing a judgment matrix and calculating eigenvalues ​​and eigenvectors, the weights of the three objectives—energy balance, grid support, and charging demand—are obtained.

[0102] Then, based on the target weights, the system uses the TOPSIS method to evaluate and rank the solutions from the Pareto solution set. The TOPSIS method calculates the distance between each solution and the ideal solution and the negative ideal solution, comprehensively evaluates the merits of each solution, and selects the solution with the best overall performance.

[0103] like Figure 2 As shown, in one embodiment of the present invention, the photovoltaic power generation data, energy storage status data, charging demand data, and grid operation data are cleaned and feature-engineered to obtain the standardized feature vector, including:

[0104] Step S1.1: Use a sliding time window to perform outlier detection and missing value processing on the data to obtain preprocessed data;

[0105] Step S1.2: Extract time-series statistical features from the preprocessed data to obtain time-series feature data;

[0106] Step S1.3: Perform feature association between the time-series feature data and meteorological data and temperature data to obtain the association feature matrix;

[0107] Step S1.4: Perform Min-Max standardization on the associated feature matrix to obtain the standardized feature vector.

[0108] In the data preprocessing stage of the photovoltaic-storage-charging microgrid, steps S1.1 to S1.4 constitute a complete data preparation process.

[0109] In step S1.1, the system uses a sliding time window technique to process the raw data. Specifically, for photovoltaic power generation data, the system sets a 24-hour time window that slides across the time series, comparing the current data point with historical data from the same period within the window. If the photovoltaic power generation value at a certain moment deviates from the historical average for the same period by more than 30% without any obvious weather change as a cause, it is marked as an outlier and replaced with the median within the window. For energy storage status data, the system detects jumps in the state of charge (SOC). If the rate of change of SOC between two adjacent time points exceeds the physically permissible range (usually no more than 3% per minute), it is determined to be an anomaly. For missing data, the system adopts different strategies based on the duration of the missing data: short-term missing data (less than 30 minutes) is filled using linear interpolation, while longer-term missing data is filled using a similar day pattern replacement method, i.e., searching for date data with similar weather conditions and seasonal characteristics in historical data.

[0110] In step S1.2, the system extracts time-series statistical features from the preprocessed data. This step reveals the time-series patterns and variation laws of the data by calculating statistics at different time scales. For example, for photovoltaic power generation data, the system extracts hourly averages (reflecting intraday variations), daily averages (reflecting weather influences), weekly averages (reflecting seasonal variations), and corresponding features such as variance, extreme values, and peak-to-valley ratio. For energy storage data, the system calculates statistical indicators such as charge-discharge cycle frequency, average depth, and capacity decay rate. For charging demand data, the system extracts features such as peak period distribution, average duration, mean power demand, and standard deviation. These time-series statistical features can characterize the operating modes and behavioral characteristics of each component of the system.

[0111] Step S1.3 correlates time-series feature data with environmental data. The system introduces meteorological data (such as irradiance, cloud cover, and wind speed) and temperature data to establish correlations with the original time-series features. For example, it performs correlation analysis between photovoltaic power generation and solar irradiance and cloud cover during the same period to calculate a correlation coefficient matrix; it quantifies the relationship between energy storage charging and discharging efficiency and ambient temperature into a temperature sensitivity coefficient; and it encodes the correlation patterns between charging demand and environmental factors such as temperature and precipitation into feature vectors. This correlation analysis enhances the model's adaptability to changes in environmental conditions and improves the accuracy of predictions.

[0112] Step S1.4 standardizes the associated feature matrix. The system uses the Min-Max standardization method to map each feature value to the interval [0,1]: x norm =(xx) min ) / (x max -x minThis process eliminates dimensional differences between different features, preventing certain features with larger values ​​from dominating the model training process. For example, unifying photovoltaic power (which may be in the kilowatt range) and temperature (usually in degrees Celsius) to the same numerical range ensures that the subsequent neural network can fairly learn the importance of each feature.

[0113] like Figure 3 As shown, in one embodiment of the present invention, the neural network search space is constructed using the architecture embedding and transition ranking method to obtain the optimal network structure for different prediction tasks, including:

[0114] Step S2.1: Construct a neural network search space based on a candidate operation set containing convolutional layers, recurrent units, and attention modules. The convolutional layers include input dimension and output dimension parameters, and the recurrent units include hidden state dimension parameters, thus obtaining the initial search space.

[0115] Step S2.2: Convert the network structure in the initial search space into a directed acyclic graph representation to obtain graph structure data;

[0116] Step S2.3: Encode the graph structure data using a graph neural network to obtain a continuous vector of fixed dimensions.

[0117] In one embodiment of the present invention, the method further includes:

[0118] Step S2.4: Construct a Siamese network using the fixed-dimensional continuous vectors, train it on the historical architecture evaluation results, and obtain an architecture performance ranking model;

[0119] Step S2.5: Based on the architecture performance ranking model, the policy gradient method is used to optimize the search space to obtain the optimal network structure for different prediction tasks.

[0120] In the neural architecture search phase, steps S2.1 to S2.5 constitute the complete process of automatically discovering the optimal network structure.

[0121] Specifically, in step S2.1, the system defines a candidate operation set containing various deep learning components as the basic units for building the network. Specifically, convolutional layer operations include one-dimensional convolution (for temporal data processing) and two-dimensional convolution (for spatial feature extraction), where the input dimension of the one-dimensional convolution ranges from 32 to 128, the output dimension ranges from 64 to 256, and the kernel size can be selected from 3, 5, or 7. Recurrent units include LSTM and GRU options, with hidden state dimensions ranging from 64 to 256 and the number of layers selectable from 1 to 3. The attention module includes self-attention mechanisms (input dimension 128, head count selectable from 4 or 8) and cross-attention mechanisms. In addition, options such as activation functions (ReLU, Sigmoid, Tanh), normalization layers (BatchNorm, LayerNorm), and connection methods (concatenation, residual) are defined. The search space composed of these operations allows the system to explore up to 10... 10 Different combinations of network structures.

[0122] Step S2.2 converts the network structure in the initial search space into a directed acyclic graph representation. Each candidate network is represented as a graph G = (V, E), where nodes V represent operations (such as convolutional layers, LSTM units), and edges E represent data flow. For example, a network structure containing an input layer, two convolutional layers, one LSTM layer, and an output layer is encoded as a directed graph with 5 nodes. Node attributes include operation type and parameter information (such as kernel size, number of channels, and number of hidden units), and edge attributes represent tensor shape transformation information. This graph representation preserves the network's topology and computation order.

[0123] Step S2.3 encodes the graph structure data using a graph neural network. The system uses a Graph Attention Network (GAT) to process the network structure graph obtained in the previous step. The GAT contains three graph attention layers, each with eight attention heads and a hidden dimension of 128. For each node, the representation is updated by aggregating information from its neighboring nodes: Where α ij k W is the attention coefficient, representing the importance of node j to node i. k This is the transformation matrix. After multi-layer graph attention computation, the system performs pooling operations on the node representations of the entire graph to obtain a continuous vector of fixed dimensions (e.g., 128 dimensions). This vector encodes the overall information of the original network architecture.

[0124] Step S2.4 constructs a Siamese network based on fixed-dimensional continuous vectors and trains the architecture performance ranking model. The Siamese network consists of two sub-networks sharing weights, each containing three fully connected layers (with 256, 128, and 64 hidden units respectively). During training, the system inputs two embedding vectors of different architectures (a...i ,a j The network outputs a relative performance score s. ij , indicating architecture a i Compared to a j The probability of better performance. Optimization is achieved using the triplet loss function: L = max(0, margin-s) ik +s ij ), where architecture a i Its performance is better than a k But worse than a j This ranking learning method does not require precise performance evaluation values, only relative ranking relationships, which greatly reduces evaluation costs. In actual training, the system randomly selects 200 architectures for evaluation (such as accuracy on the validation set), generates approximately 10,000 training triples, and trains using the Adam optimizer with a learning rate of 0.001 for 50 epochs.

[0125] Step S2.5 uses the policy gradient method to find the optimal network structure based on the architecture performance ranking model. The system models the architecture search problem as a reinforcement learning task and uses a policy network (a recurrent neural network containing two layers of GRU with 256 hidden units) to generate a sequence of architecture descriptions. Each generated architecture a is evaluated for its performance score r(a) using the ranking model trained in step S2.4, and then the policy network parameters are updated using the policy gradient algorithm REINFORCE. To enhance exploration efficiency, the system employs Monte Carlo Tree Search (MCTS) technology, performing 1000 simulations for each candidate operation and selecting the operation with the highest expected return. In this way, the system can automatically discover the optimal neural network structure for various tasks such as photovoltaic power generation forecasting, energy storage scheduling, and charging demand forecasting. Experiments show that compared to manually designed network structures, the automatically searched structure improves prediction accuracy by more than 15% and exhibits better generalization ability.

[0126] like Figure 4 As shown, in one embodiment of the present invention, the multi-head self-attention layer is used to model the dependency relationship between the categorical variable and the continuous variable to obtain the mixed variable optimization model, including:

[0127] Step S3.1: Perform one-hot encoding on the categorical variables and normalize the continuous variables to obtain the encoded input variables;

[0128] Step S3.2: Project the encoded input variables into the query space, key space, and value space to obtain the projected variables;

[0129] Step S3.3: Calculate the attention weights of the projected variables using scaled dot-product attention to obtain the attention features;

[0130] Step S3.4: Perform multi-layer cross-feature learning on the attention features to obtain the mixed variable optimization model.

[0131] In step S3.1, the system first performs encoding preprocessing on different types of variables. For categorical variables, such as the operating mode of energy storage devices (charging / discharging / standby) and the microgrid connection status (grid-connected / off-grid), the system uses one-hot encoding to convert them into binary vectors. For example, an operating mode variable with three states is encoded as [1,0,0], [0,1,0], or [0,0,1]. This encoding method allows categorical variables to be effectively processed by the neural network. For continuous variables, such as numerical variables like photovoltaic output power, energy storage state of charge, and charging power, the system uses normalization. Based on the distribution characteristics of the variables, the system chooses z-score standardization (for variables following a normal distribution) or min-max scaling (for variables with clear upper and lower bounds) to unify continuous variables of different magnitudes into a similar numerical range. This preprocessing ensures that different types of variables can participate fairly in the calculation in subsequent modeling.

[0132] Step S3.2 implements the core projection operation of the multi-head self-attention mechanism. The system projects the encoded input variables onto three different feature spaces: query space, key space, and value space. This step is achieved through three different linear transformations, namely, multiplying the input variables by three different weight matrices. These projection operations allow the model to learn the relationships between variables from different perspectives, enhancing its expressive power. Specifically, the system sets an independent projection matrix for each attention head, enabling different attention heads to focus on different types of variable interaction patterns. For example, in a system with eight attention heads, some heads may focus on learning the relationship between photovoltaic power generation and environmental conditions, while others may focus more on the interaction between energy storage state and charging demand.

[0133] Step S3.3 calculates attention weights using scaled dot-product attention. This attention mechanism is central to modeling dependencies between variables. Scaled dot-product attention works by calculating the similarity between the query vector and all key vectors, and then weighting the value vectors based on these similarities. Specifically, it first calculates the dot product of the query vector and key vectors to represent their similarity; then it scales the result by dividing it by the square root of the key vector dimension—this prevents the dot product from becoming too large and causing the softmax function's gradient to vanish; next, it applies the softmax function to the scaled result to obtain normalized attention weights; finally, it uses these weights to weightedly sum the value vectors to obtain the attention output. This mechanism allows the model to dynamically determine the importance relationships between variables, such as identifying a high correlation between photovoltaic power generation and current cloud cover, or the dependency between energy storage scheduling and grid load.

[0134] Step S3.4 performs multi-layer cross-feature learning on the attention features, which is elaborated in steps S3.4.1 to S3.4.4.

[0135] In one embodiment of the present invention, multi-layer cross-feature learning is performed on the attention features to obtain the mixed-variable optimization model, including:

[0136] Step S3.4.1: Construct a multi-layer cross-network structure, calculate the higher-order interaction features between variables, and obtain the interaction features;

[0137] In step S3.4.1, the system constructs a multi-layer cross-network structure to calculate higher-order interaction features between variables. Unlike simple linear combinations, cross-networks can capture nonlinear interaction relationships between variables. For example, photovoltaic power generation and energy storage charging may have a synergistic effect; that is, a strategy of increasing energy storage charging when photovoltaic power generation is sufficient is better than considering the two factors individually. The cross-network explicitly models these interaction relationships, calculates the product features between variables, and extracts higher-order interaction patterns through multi-layer transformations. The system designs a 3-layer cross-network, with each layer calculating the interaction between the features of the previous layer and the original features, constructing interaction features from second-order to higher-order layer by layer.

[0138] Step S3.4.2: Use a gating mechanism to dynamically adjust the importance of the interaction features to obtain optimized interaction features;

[0139] In step S3.4.2, the system employs a gating mechanism to dynamically adjust the importance of interaction features. Not all variable interactions are equally important to the current task; the gating mechanism allows the model to dynamically determine the importance of each interaction feature based on the current input. The system calculates a weight coefficient between 0 and 1 for each interaction feature, multiplying it by the original feature to obtain the adjusted feature. This mechanism is similar to attention allocation in human decision-making, selectively focusing on key factors according to different scenarios. For example, on a sunny day with stable photovoltaic output, the system may focus more on the interaction between charging demand and energy storage status; while on a cloudy or rainy day with fluctuating photovoltaic output, it may focus more on the coordination between photovoltaic forecasting and energy storage scheduling.

[0140] Step S3.4.3: Fuse the optimized interaction features with the attention features through residual connections to obtain fused features;

[0141] Step S3.4.3 fuses the optimized interaction features with the attention features through residual connections. Residual connections are an important network design technique that effectively alleviates the vanishing gradient problem in deep network training by directly adding the input to the output. In this system, residual connections ensure that the information of the original attention features is not lost during multi-layer interaction learning, while allowing the model to adaptively decide whether to utilize higher-order interaction features. In this way, the system retains the expressive power of the original features while enhancing its ability to model complex variable relationships.

[0142] Step S3.4.4: Perform layer normalization on the fused features to obtain the hybrid variable optimization model.

[0143] The system performs layer normalization on the fused features. Layer normalization is a common technique in deep learning, which standardizes features by calculating the mean and variance of each sample along its feature dimension. This operation stabilizes the training process of deep networks, accelerates convergence, and improves the model's generalization ability. In this system, layer normalization ensures that features from different sources (original attention features and interaction features) can be fused within similar numerical ranges, preventing certain features from dominating the model's output. After this series of processes, the system ultimately obtains a mixed-variable optimization model that can accurately model the complex dependencies between categorical and continuous variables.

[0144] For example, the system first performs encoding preprocessing on categorical and continuous variables. Assume the photovoltaic-storage-charging microgrid system contains the following variables: energy storage operating mode (charging / discharging / standby) as a categorical variable, and photovoltaic output power (0-500kW) and energy storage state of charge (SOC) (20%-95%) as continuous variables. For the categorical variable of energy storage operating mode, the system uses one-hot encoding to convert it into a three-dimensional binary vector: charging mode is represented as [1,0,0], discharging mode as [0,1,0], and standby mode as [0,0,1]. For photovoltaic output power, the system uses min-max normalization to map it to the [0,1] interval; for example, 380kW photovoltaic power is normalized to (380-0) / (500-0) = 0.76. For energy storage SOC, min-max normalization is also performed; for example, 60% SOC is normalized to (60-20) / (95-20) = 0.53. In this way, the original mixed-type variables are converted into a uniform numerical vector [1,0,0,0.76,0.53], which is convenient for subsequent neural network processing.

[0145] The system projects the encoded input variables into three different feature spaces. Assume we set the hidden dimension of the attention mechanism to 64 and use 8 attention heads. For the encoded input vector [1,0,0,0.76,0.53], the system first inputs it into a 5×64 linear transformation layer, expanding the 5-dimensional input into 64-dimensional features. Then, the system creates three sets of projection matrices for each attention head, projecting the 64-dimensional features into the query space, key space, and value space. Specifically, the system defines three 8×64-dimensional weight matrices Wq1, Wk1, and Wv1 for the first attention head (note that the output dimension of each head is the total hidden dimension divided by the number of heads, i.e., 64 / 8 = 8). Through matrix multiplication, the system obtains the query vector Q1, key vector K1, and value vector V1 for this attention head, each with a dimension of 8. The other 7 attention heads undergo similar projection operations, ultimately resulting in 8 sets of projected variables.

[0146] The system utilizes scaled dot-product attention to calculate attention weights. Taking the first attention head as an example, the system first calculates the dot product between the query vector Q1 and the key vector K1. This operation measures the correlation between different variables. For example, a higher dot product between the part of the query vector related to energy storage mode and the part of the key vector related to SOC indicates a strong correlation between them. Then, the system divides the dot product result by a scaling factor √8 (the square root of the dimension of the key vector). This scaling operation avoids the dot product value becoming too large in high-dimensional space, which could lead to the vanishing softmax gradient. Next, the softmax function is applied to the scaled result to obtain normalized attention weights, such as [0.15, 0.25, 0.05, 0.35, 0.20], representing the relative importance among the five original variables. Finally, the system uses these weights to perform a weighted summation on the value vector V1 to obtain the output feature of this attention head. This process is repeated for all eight attention heads, and their outputs are then concatenated to obtain a comprehensive attention feature vector with a dimension of 64 (8 heads × 8 dimensions per head).

[0147] The system performs multi-layer cross-feature learning on attention features to construct a mixed-variable optimization model. The system designs a three-layer cross-network, where each layer calculates the interaction between the original feature and the output of the previous layer. Taking the first layer as an example, assuming the attention feature vector is [f1, f2, ..., f64], the system calculates the second-order interaction between each pair of features, such as f1×f2, f1×f3, etc., capturing the nonlinear relationships between variables. These interaction features reflect more complex variable dependencies; for example, the interaction between photovoltaic power and energy storage mode may indicate that the advantage of choosing the charging mode is more pronounced when photovoltaic power is high. Then, the system uses a gating mechanism to assign weights to each interaction feature. For example, the interaction feature between photovoltaic power and SOC is assigned a weight of 0.8, while the interaction feature between standby mode and photovoltaic power is assigned only a weight of 0.3, indicating that the former is more important in the current decision. After applying the gating weights, the system adds the adjusted interaction feature to the original attention feature through residual connections, for example, new feature = original feature + 0.8 × interaction feature. This design ensures the transmission of original information in the deep network. Finally, the fused features are subjected to layer normalization to adjust the feature distribution and obtain the final output of the mixed variable optimization model.

[0148] Through this series of processes, the system successfully constructed a mixed-variable optimization model capable of accurately capturing the complex interactions between categorical and continuous variables. This model can automatically learn the importance relationships between different variables, such as identifying the strong correlation between energy storage discharge modes and grid support under cloudy conditions (low photovoltaic power); or the key interaction between photovoltaic power and charging station allocation strategies during peak charging demand periods. In practical applications, this model outperforms traditional linear programming or simple neural network models, reducing prediction errors by 25% while improving the system's adaptability to complex scenarios. For example, in a real-world test case, the traditional method had an average error of 15% when handling energy storage mode switching and charging strategy coordination, while this model's error was only 4.2%, significantly improving the accuracy and reliability of scheduling decisions.

[0149] This hybrid variable optimization model, based on multi-head self-attention and cross-feature learning, can effectively handle the complex variable relationships in the scheduling of photovoltaic-storage-charging microgrids, providing a precise decision-making basis for subsequent multi-objective optimization. Compared with traditional methods, this model has significant advantages in handling mixed optimization problems involving categorical and continuous variables, capturing richer variable interaction information and improving the overall optimization effect of the system.

[0150] like Figure 5 As shown, in one embodiment of the present invention, the improved NSGA-III algorithm is used to perform multi-objective optimization on the energy balance constraint, the grid support index, and the charging demand satisfaction, to obtain the scheduling strategy of the photovoltaic-storage-charging microgrid, including:

[0151] Step S4.1: Construct an energy balance objective function that includes power balance constraints and energy storage capacity constraints, and take the device operation limitations and the network topology constraints as constraints to obtain the complete optimization problem;

[0152] Step S4.1 first constructs a complete optimization problem with multiple constraints. The system defines an energy balance objective function, the core of which is the power balance constraint and the energy storage capacity constraint. The power balance constraint ensures that the total power generation of the microgrid system is equal to the total power consumption at any given time, which can be expressed as the sum of photovoltaic power generation and energy storage discharge power equals the sum of the base load, charging station load, and grid-connected power. This constraint is the fundamental guarantee for the safe and stable operation of the microgrid. The energy storage capacity constraint ensures that the state of charge of the energy storage devices is kept within a safe range, typically requiring a state of charge of no less than 20% (to prevent over-discharge) and no more than 95% (to prevent over-charge), while also considering the protection of battery cycle life. In addition to the energy balance objective, the system also includes equipment operation limits and network topology constraints as constraints. Equipment operation limits include the maximum power limit of photovoltaics, the charging and discharging power limit of energy storage, and the maximum power limit of charging piles; network topology constraints consider the physical connection relationship of the microgrid, such as line capacity limits and the allowable range of node voltages. These constraints together form a complex optimization problem that requires an efficient algorithm to solve.

[0153] Step S4.2: Based on the optimization problem, Pareto front search is performed using the reference point mechanism and crowding order to obtain the Pareto solution set;

[0154] In multi-objective optimization, there is usually no single solution that simultaneously optimizes all objectives; instead, there exists a set of non-dominated solutions, known as the Pareto solution set. The system employs an improved NSGA-III algorithm, an evolutionary algorithm specifically designed for multi-objective optimization problems. The reference point mechanism is a core innovation of NSGA-III. The system uniformly distributes a series of reference points in the objective space, guiding the population to evolve in different directions and ensuring a diverse Pareto solution set. For the three optimization objectives (energy balance, grid support, and charging demand satisfaction), the system typically sets 15 to 21 reference points. In each generation of evolution, the system selects elite individuals based on the non-dominated ranking principle and uses crowding ranking to ensure the diversity of the solution set. Crowding is an indicator of the density of solutions distributed in the objective space; solutions with higher crowding are more likely to be retained, thus preventing the solution set from becoming overly concentrated in certain regions. Furthermore, the system employs adaptive crossover and mutation operators to enhance search efficiency, such as dynamically adjusting crossover probabilities and mutation strengths based on population diversity. Through hundreds of generations of evolutionary iterations, the system ultimately obtains a balanced set of non-dominated solutions, forming the Pareto front and providing multiple alternatives for decision-making.

[0155] Step S4.3: Generate the scheduling policy based on the Pareto solution set.

[0156] Since the Pareto solution set contains multiple non-dominated solutions, the system needs to select the most suitable solution as the final strategy according to the actual application requirements. This process is elaborated in steps S4.3.1 to S4.3.3.

[0157] In one embodiment of the present invention, generating the scheduling policy based on the Pareto set includes:

[0158] Step S4.3.1: Use fuzzy hierarchical analysis to determine the priority of each optimization objective and obtain the objective weights;

[0159] Step S4.3.2: Based on the target weights, use the TOPSIS method to evaluate and rank the solutions from the Pareto solution set to obtain the optimal solution;

[0160] Step S4.3.3: Generate scheduling instructions that meet the actual operation requirements based on the optimal solution, and obtain the scheduling strategy of the optical storage and charging microgrid.

[0161] In step S4.3.1, the system uses fuzzy hierarchical analysis (AHP) to determine the priority of each optimization objective. Fuzzy hierarchical analysis is a multi-criteria decision-making method capable of handling uncertainty and fuzziness in the decision-making process. The system first establishes a judgment matrix to represent the relative importance of each objective in pairwise comparisons; then, it calculates eigenvalues ​​and eigenvectors to obtain the weight of each objective. For example, in some application scenarios, energy balance might be assigned a 50% weight, grid support a 30% weight, and charging demand satisfaction a 20% weight; while in other scenarios, these weights may differ, reflecting different operational priorities.

[0162] Step S4.3.2: Based on the objective weights, the TOPSIS method is used to evaluate and rank the solutions from the Pareto solution set. TOPSIS (Technique for Order Preference by Similarity to an Ideal Solution) is a multi-attribute decision-making method based on ideal solutions. The system first constructs a weighted decision matrix according to the objective weights; then it determines the positive ideal solution (where all objectives are optimal) and the negative ideal solution (where all objectives are worst-case); next, it calculates the distance between each Pareto solution and the positive and negative ideal solutions; finally, it calculates the relative proximity, i.e., how close the solution is to the positive ideal solution, and selects the solution with the highest proximity as the optimal solution. This method considers the weight differences of each objective and can find the optimal balance point among multiple objectives.

[0163] Step S4.3.3 generates scheduling instructions that meet actual operational requirements based on the selected optimal solution. The system transforms the abstract optimization results into specific operational instructions, including photovoltaic power limits, energy storage charging and discharging plans, and charging station load allocation schemes. These instructions need to consider the actual operating characteristics and limitations of the equipment, such as the ramp-up rate limit of energy storage equipment (usually 5% / minute of rated power) and the minimum charging power requirement of charging piles. In addition, the system also generates a time-series scheduling plan, specifying detailed operating strategies for each time period within the next 24 hours to address the time-varying characteristics of photovoltaic power generation and load demand. The final scheduling strategy satisfies both theoretical optimization requirements and considers the operability of the actual system, directly guiding the real-time operation of the photovoltaic-storage-charging microgrid.

[0164] Through this series of steps, the system successfully applied the previously established mixed-variable optimization model to the actual decision-making process, solving the balance problem among energy balance, grid support, and charging demand, and realizing the coordinated optimization scheduling of photovoltaic-storage-charging microgrids. Compared with traditional single-objective optimization or manual scheduling methods, the multi-objective optimization framework of this invention can more comprehensively consider all aspects of system operation, providing a more balanced and efficient scheduling strategy, effectively improving the overall operational efficiency of the microgrid system. Practical applications show that using this method can increase the renewable energy absorption rate by more than 15%, reduce the cycle degradation of energy storage equipment by about 20%, and simultaneously meet more than 90% of charging demand, achieving a win-win situation for both economic and environmental benefits.

[0165] Consider a typical photovoltaic-storage-charging microgrid system, comprising a 500kW photovoltaic power generation unit, a 200kWh / 100kW energy storage system, and 10 charging piles (each with a rated power of 7kW). In step S4.1, the system first constructs an energy balance objective function that includes power balance constraints and energy storage capacity constraints. Taking 14:00 on a certain day as an example, the predicted photovoltaic power generation is 380kW, the base load is 220kW, and 5 electric vehicles request charging, with an estimated total charging power demand of 35kW. At this time, the power balance constraint can be expressed as: 380kW (photovoltaic) + P ess (Energy storage, positive value indicates discharging, negative value indicates charging) = 220kW (base load) + 35kW (charging load) + P grid(Power exchange in the grid; positive values ​​represent output, negative values ​​represent absorption). Energy storage capacity constraints require the State of Charge (SOC) of the energy storage system to remain within a safe range: 20% ≤ SOC ≤ 95%. Assuming the current SOC is 60%, charging power should not cause the SOC to exceed 95%, and discharging power should not cause the SOC to fall below 20%. Furthermore, the system also considers equipment operation limitations, such as energy storage charging and discharging power not exceeding 100kW and a ramp rate limit of 5kW / minute; and network topology constraints, such as distribution line capacity limited to 400kW and node voltage deviation not exceeding ±5%. These constraints together constitute a complete optimization problem framework.

[0166] In step S4.2, the system performs Pareto front search based on the constructed optimization problem, employing a reference point mechanism and congestion ranking. For the three optimization objectives—energy balance, grid support, and charging demand satisfaction—15 reference points are evenly distributed in the target space. Using the NSGA-III algorithm with a population size of 100, a set of random solutions is initialized, and the iterative optimization process begins. In each generation, the system uses simulated binary crossover (crossover probability 0.9, distribution exponent 20) and multinomial mutation (mutation probability 1 / n, distribution exponent 20) to generate new solutions. Then, selection is made based on non-dominated ranking and reference point correlation, retaining elite individuals. For example, in the 50th generation of evolution, the system might obtain a set of solutions as follows: Solution A prioritizes energy balance (energy storage power is -125kW, i.e., charging power is 125kW, grid power is 0); Solution B prioritizes grid support (energy storage power is -100kW, grid power is -25kW, i.e., absorbing 25kW from the grid); Solution C prioritizes meeting charging demand (energy storage power is -110kW, grid power is -15kW, while increasing charging priority). After 300 generations of evolution, the system eventually obtains a set of approximately 20 non-dominated solutions, forming the Pareto front, providing multiple alternatives for decision-making.

[0167] In step S4.3, the system needs to select the most suitable solution from the Pareto solution set as the final scheduling strategy. Assume that, based on current operational needs, the manager decides to prioritize grid stability while also considering energy efficiency and user charging experience. Using fuzzy hierarchical analysis, the system assigns a 45% weight to grid support indicators, a 35% weight to energy balance targets, and a 20% weight to charging demand satisfaction. Then, the system uses the TOPSIS method to calculate the relative proximity of each Pareto solution. For example, solution A has a proximity of 0.68, solution B has a proximity of 0.82, and solution C has a proximity of 0.75. Since solution B has the highest proximity, the system selects it as the optimal solution. Based on this, the system generates specific scheduling instructions: the photovoltaic power limit is set to 380kW (full acceptance), the energy storage system charging power is set to 100kW, 25kW of power is absorbed from the grid, and the charging station allocates 35kW of power to five electric vehicles. In addition, the system generates a time-series scheduling plan for the next four hours, expecting energy storage charging to be completed before 16:00, at which point it will switch to discharge mode to support the evening peak load. This scheduling strategy not only ensures the current power balance but also prepares for future high-load periods, while improving the system's stability and reliability by making appropriate use of grid resources.

[0168] This practical example demonstrates that the multi-objective optimization framework of this invention can comprehensively consider all aspects of the operation of a photovoltaic-storage-charging microgrid. While meeting physical constraints, it balances different operational objectives and provides a scientific and reasonable scheduling strategy. Practical application data shows that, compared to traditional scheduling methods, this method can increase the renewable energy absorption rate from 75% to 92%, reduce grid fluctuations by more than 15%, and maintain a charging service satisfaction rate of over 95%, significantly improving the overall operational efficiency and sustainability of the microgrid system.

[0169] In one embodiment of the present invention, the standardized feature vector includes the mean feature extracted from the photovoltaic power generation data, the variance feature extracted from the energy storage status data, and the peak feature extracted from the charging demand data; the grid support indicators include the voltage stability indicator and the power factor indicator in the grid operation data; and the charging demand satisfaction includes the charging time indicator and the power demand indicator.

[0170] like Figure 6 As shown, the present invention also provides a photovoltaic-storage-charging microgrid scheduling system, comprising:

[0171] The preprocessing module is used to acquire photovoltaic power generation data, energy storage status data, charging demand data and grid operation data, and to clean and perform feature engineering on the data to obtain standardized feature vectors.

[0172] An architecture search module is used to construct a neural network search space based on the standardized feature vectors using the architecture embedding and transition ranking methods, thereby obtaining the optimal network structure for different prediction tasks.

[0173] The optimization modeling module is used to model the dependency relationship between the categorical variables and the continuous variables using the multi-head self-attention layer based on the optimal network structure, so as to obtain the mixed variable optimization model.

[0174] The strategy generation module is used to perform multi-objective optimization on the energy balance constraint, the grid support index and the charging demand satisfaction based on the mixed variable optimization model and the improved NSGA-III algorithm, so as to obtain the scheduling strategy of the photovoltaic-storage-charging microgrid.

[0175] Finally, the system generates scheduling instructions that meet actual operational requirements based on the optimal solution. These instructions include photovoltaic power predictions, energy storage charging and discharging plans, and charging station load allocation schemes, forming a complete photovoltaic-storage-charging microgrid scheduling strategy, which is then issued for execution.

[0176] This invention employs a multi-task neural architecture search method involving architecture embedding and transition ranking to automatically optimize network structures for different prediction tasks, thereby improving model adaptability and prediction accuracy. By designing a mixed-variable optimization model based on an attention mechanism, it effectively handles the interaction between categorical and continuous variables. Furthermore, it proposes a multi-objective joint optimization framework to achieve coordinated optimization of energy balance, grid support, and charging demand. This invention is applicable to the scheduling optimization of distributed photovoltaic-storage-charging microgrid systems and has promising application prospects.

[0177] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall fall within the protection scope of the present invention.

[0178] This disclosure also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program performs the steps of a microgrid scheduling method for optical storage and charging described in the above-described method embodiments. The storage medium can be either volatile or non-volatile computer-readable storage.

[0179] Furthermore, this disclosure also provides a computer program product, which stores a computer program. When the computer program is run by a processor, it executes the steps of a microgrid scheduling method for optical storage and charging provided in any of the above embodiments of this disclosure. For details, please refer to the above method embodiments, which will not be repeated here.

[0180] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium, which can be a volatile or non-volatile computer-readable storage medium. In another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0181] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices and apparatuses described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this disclosure, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms.

[0182] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0183] In addition, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0184] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0185] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A method for scheduling a photovoltaic-storage-charging microgrid, characterized in that, include: Acquire photovoltaic power generation data, energy storage status data, charging demand data, and grid operation data; clean and perform feature engineering on the photovoltaic power generation data, energy storage status data, charging demand data, and grid operation data to obtain standardized feature vectors. Based on the standardized feature vectors, the neural network search space is constructed using architecture embedding and transition ranking methods to obtain the optimal network structure for different prediction tasks. Based on the optimal network structure, the multi-head self-attention layer is used to model the dependency relationship between categorical and continuous variables to obtain a mixed variable optimization model. Based on the aforementioned mixed-variable optimization model, the improved NSGA-III algorithm is used to perform multi-objective optimization solutions for energy balance constraints, grid support indicators, and charging demand satisfaction, thereby obtaining the scheduling strategy for the photovoltaic-storage-charging microgrid. The scheduling of the photovoltaic-storage-charging microgrid is then executed based on the scheduling strategy.

2. The method according to claim 1, characterized in that, The photovoltaic power generation data, energy storage status data, charging demand data, and grid operation data are cleaned and feature-engineered to obtain the standardized feature vector, including: The data is preprocessed by using a sliding time window to detect outliers and handle missing values. Temporal statistical features are extracted from the preprocessed data to obtain temporal feature data; The time-series feature data is correlated with meteorological data and temperature data to obtain a correlation feature matrix; The associated feature matrix is ​​subjected to Min-Max normalization to obtain the normalized feature vector.

3. The method according to claim 1, characterized in that, The neural network search space is constructed using the aforementioned architecture embedding and transition ranking methods to obtain the optimal network structure for different prediction tasks, including: The neural network search space is constructed based on a candidate operation set containing convolutional layers, recurrent units, and attention modules. The convolutional layers include input dimension and output dimension parameters, and the recurrent units include hidden state dimension parameters, thus obtaining the initial search space. The network structure in the initial search space is converted into a directed acyclic graph representation to obtain graph structure data; The graph structure data is encoded using a graph neural network to obtain a continuous vector of fixed dimensions.

4. The method according to claim 3, characterized in that, The method further includes: A Siamese network is constructed using the fixed-dimensional continuous vectors, and trained on historical architecture evaluation results to obtain an architecture performance ranking model. Based on the architecture performance ranking model, the policy gradient method is used to optimize the search space to obtain the optimal network structure for different prediction tasks.

5. The method according to claim 1, characterized in that, The multi-head self-attention layer is used to model the dependency relationship between the categorical variable and the continuous variable to obtain the mixed variable optimization model, which includes: The categorical variables are one-hot encoded, and the continuous variables are normalized to obtain the encoded input variables. The encoded input variables are projected onto the query space, key space, and value space to obtain the projected variables; The attention weights of the projected variables are calculated using scaled dot-product attention to obtain the attention features; The attention features are subjected to multi-layer cross-feature learning to obtain the mixed variable optimization model.

6. The method according to claim 5, characterized in that, Multi-layer cross-feature learning is performed on the attention features to obtain the mixed-variable optimization model, including: Construct a multi-layered cross-network structure, calculate the higher-order interaction features between variables, and obtain the interaction features; A gating mechanism is used to dynamically adjust the importance of the interaction features to obtain optimized interaction features; The optimized interaction features are fused with the attention features through residual connections to obtain fused features; The fused features are subjected to layer normalization to obtain the hybrid variable optimization model.

7. The method according to claim 1, characterized in that, The improved NSGA-III algorithm is used to perform multi-objective optimization on the energy balance constraint, the grid support index, and the charging demand satisfaction, resulting in the scheduling strategy for the photovoltaic-storage-charging microgrid, including: Construct an energy balance objective function that includes power balance constraints and energy storage capacity constraints, and take the device operation limitations and the network topology constraints as constraints to obtain a complete optimization problem; Based on the aforementioned optimization problem, a Pareto front search is performed using a reference point mechanism and crowding ordering to obtain the Pareto solution set. The scheduling policy is generated based on the Pareto solution set.

8. The method according to claim 7, characterized in that, The scheduling policy is generated based on the Pareto solution set, including: The priority of each optimization objective is determined by fuzzy hierarchical analysis, and the objective weights are obtained. Based on the target weights, the TOPSIS method is used to evaluate and rank the solutions from the Pareto solution set to obtain the optimal solution. Based on the optimal solution, a scheduling instruction that meets the actual operational requirements is generated, thus obtaining the scheduling strategy of the optical storage and charging microgrid.

9. The method according to claim 1, characterized in that: The standardized feature vector includes the mean feature extracted from the photovoltaic power generation data, the variance feature extracted from the energy storage status data, and the peak feature extracted from the charging demand data. The power grid support indicators include voltage stability indicators and power factor indicators in the power grid operation data; The charging demand satisfaction includes charging time indicators and power demand indicators.

10. A photovoltaic-storage-charging microgrid scheduling device, characterized in that, include: The preprocessing module is used to acquire photovoltaic power generation data, energy storage status data, charging demand data and grid operation data, and to clean and perform feature engineering on the data to obtain standardized feature vectors. An architecture search module is used to construct a neural network search space based on the standardized feature vectors using the architecture embedding and transition ranking methods, thereby obtaining the optimal network structure for different prediction tasks. The optimization modeling module is used to model the dependency relationship between the categorical variables and the continuous variables using the multi-head self-attention layer based on the optimal network structure, so as to obtain the mixed variable optimization model. The strategy generation module is used to perform multi-objective optimization on the energy balance constraint, the grid support index and the charging demand satisfaction based on the mixed variable optimization model and the improved NSGA-III algorithm, so as to obtain the scheduling strategy of the photovoltaic-storage-charging microgrid.

Images