Multi-strategy operation method, device and equipment for optical storage and charging station based on deep learning

Abstract

The present invention discloses a multi-strategy operation method, device and equipment for a photovoltaic energy storage charging station based on deep learning. The method includes: obtaining operation data of the integrated photovoltaic energy storage charging station, which is multi-source heterogeneous; performing data fusion on the operation data based on an improved spatio-temporal heterogeneous graph fusion model and using an improved gated graph convolutional network to obtain the fused data; performing feature extraction operations on the fused data to obtain high-dimensional feature vectors; performing dimensionality reduction and key information extraction on the high-dimensional feature vectors through an improved sparse autoencoder to generate low-dimensional representations, and performing a global search on the low-dimensional representations through an improved hybrid evolutionary algorithm to generate an optimal operation strategy; and applying the optimal operation strategy to the operation of the integrated photovoltaic energy storage charging station in real time. The present invention can effectively improve the operation management efficiency and operation management effect of the integrated photovoltaic energy storage charging station.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular, to a multi-strategy operation method, device and equipment for a photovoltaic energy storage charging station based on deep learning. Background Art

[0002] In the operation scenario of a photovoltaic energy storage integrated energy station, the existing technologies generally face many challenges. On the one hand, when dealing with multi-source heterogeneous data, the existing systems often rely on traditional data analysis means and are difficult to accurately capture the complex coupling relationships among the data of photovoltaic panels, energy storage batteries and charging piles. For example, in different climate conditions and peak-valley periods of electricity consumption, the coordinated operation of photovoltaic power generation, energy storage allocation and charging demand is extremely complex, and the existing technologies cannot fully explore the deep associations hidden behind these data, resulting in limited energy comprehensive utilization efficiency. On the other hand, the current operation strategies are mostly based on experience or simple historical data statistics and lack the ability to adaptively adjust to real-time dynamic changes. When the grid electricity price fluctuates and the photovoltaic power generation changes greatly due to weather conditions, it is difficult to quickly make optimal charging pile pricing and energy storage charge-discharge decisions, resulting in high operation costs and poor user experience. Such operation strategies lacking dynamics and pertinence are difficult to adapt to the complex and changeable operating environment of the photovoltaic energy storage integrated energy station, leading to low energy utilization efficiency and high operation costs.

[0003] In short, the existing operation methods of photovoltaic energy storage integrated energy stations are difficult to effectively adapt to the complex and changeable operating environment of photovoltaic energy storage integrated energy stations, resulting in low operation management efficiency and unsatisfactory operation management effects. Summary of the Invention

[0004] Embodiments of the present invention provide a multi-strategy operation method, device and equipment for a photovoltaic energy storage charging station based on deep learning, which can effectively improve the operation management efficiency and operation management effects of a photovoltaic energy storage integrated energy station.

[0005] An embodiment of the present invention provides a multi-strategy operation method for a photovoltaic energy storage charging station based on deep learning, including:

[0006] Obtain the operation data of the photovoltaic energy storage integrated energy station as multi-source heterogeneous, where the operation data includes the state data of photovoltaic panels and energy storage batteries and the usage state data of charging piles;

[0007] Based on an improved spatio-temporal heterogeneous graph fusion model and using an improved gated graph convolutional network, perform data fusion on the operation data to obtain the fused data;

[0008] Perform feature extraction operations on the fused data to obtain a high-dimensional feature vector that comprehensively reflects the complex dependence relationships among the multi-source heterogeneous operation data;

[0009] Dimensionality reduction and key information extraction are performed on the high-dimensional feature vectors through an improved sparse autoencoder to generate a low-dimensional representation, and a global search is performed on the low-dimensional representation through an improved hybrid evolutionary algorithm to generate an optimal operation strategy; the optimal operation strategy includes a charging pile pricing strategy, a charge and discharge strategy for energy storage batteries, and a user discount push strategy;

[0010] The optimal operation strategy is applied to the operation of the integrated photovoltaic, energy storage and charging station in real time; among them, according to the charging pile pricing strategy, the pricing of the charging piles in the integrated photovoltaic, energy storage and charging station is dynamically adjusted through an intelligent electricity meter, discount information is pushed to the user terminal according to the user discount push strategy, and the charge and discharge strategy of the energy storage batteries in the integrated photovoltaic, energy storage and charging station is optimized through an energy storage management system according to the energy storage battery charge and discharge strategy.

[0011] As an improvement to the above solution, data fusion is performed on the operation data based on the improved spatio-temporal heterogeneous graph fusion model and using the improved gated graph convolutional network to obtain the fused data, including:

[0012] The operation data is normalized using a normalization algorithm that combines logarithmic transformation and linear scaling, so as to map the operation data to an adapted interval to obtain the preprocessed operation data;

[0013] According to the preprocessed operation data, the photovoltaic panels, energy storage batteries and charging piles of the integrated photovoltaic, energy storage and charging station are abstracted into nodes of a spatio-temporal heterogeneous graph and their attributes are extended. At the same time, the time edges and space edges of the spatio-temporal heterogeneous graph are defined, and the weights of each edge of the spatio-temporal heterogeneous graph are calculated to reflect the spatio-temporal correlation relationship between nodes, and a spatio-temporal heterogeneous graph is constructed.

[0014] Using the improved gated graph convolutional network, layer-by-layer operations of multiple layers of networks are performed on the constructed spatio-temporal heterogeneous graph, and the local features of the nodes are aggregated to obtain the fused data.

[0015] As an improvement to the above solution, feature extraction operations are performed on the fused data to obtain high-dimensional feature vectors that comprehensively reflect the complex dependence relationships between multi-source heterogeneous operation data, including:

[0016] The fused data is reshaped, and it is constructed into a data cube structure according to different data sources and time series. At the same time, windows are dynamically divided according to the data fluctuation conditions to generate adaptable dynamic window data;

[0017] Based on the obtained dynamic window data, the comprehensive importance of each feature is evaluated from the aspects of information entropy and feature correlation within each window, and important features are selected to form a selected feature subset;

[0018] For the selected feature subset, a deep residual attention network is built to solve the gradient problem by using residual connections and strengthen the expression of key features with the attention mechanism. After processing, global average pooling is performed to obtain a high-dimensional feature vector that comprehensively reflects the complex dependence relationships of multi-source heterogeneous operation data.

[0019] As an improvement to the above solution, the high-dimensional feature vector is dimension-reduced and key information is extracted through an improved sparse autoencoder to generate a low-dimensional representation, and a global search is performed on the low-dimensional representation through an improved hybrid evolutionary algorithm to generate an optimal operation strategy; the optimal operation strategy includes a charging pile pricing strategy, a charge and discharge strategy for energy storage batteries, and a user preference push strategy, including:

[0020] First, preprocessing based on the data distribution characteristics is performed on the high-dimensional feature vector. By constructing an information topology graph and using a unique probability activation function for encoding, and introducing a sparse constraint mechanism based on the information gain rate for training, a dimension-reduced low-dimensional representation is generated.

[0021] Based on the low-dimensional representation, a chaotic mapping is used to generate an initial population, and the population individuals correspond to different combinations of operation strategies.

[0022] A multi-objective fitness function that comprehensively considers the total revenue of the energy station, user satisfaction, and energy utilization rate is constructed to evaluate the fitness of the initial population.

[0023] Chaotic search is used to generate perturbation vectors in the neighborhood of the initial population individuals. By comparing the fitness of the new individuals with that of the original individuals, the population is updated to achieve global exploration.

[0024] On the basis of the population after chaotic search, a multi-objective game mechanism is introduced. Individuals generate new individuals in the neighborhood and play games according to the benefits of different operation objectives to adjust their own strategy parameters and promote the evolution of the population.

[0025] After multiple alternating iterations of chaotic search and multi-objective game, when the termination condition is met, the strategy corresponding to the individual with the highest fitness in the population is selected to generate an optimal operation strategy including charging pile pricing, charge and discharge of energy storage batteries, and user preference push.

[0026] Another embodiment of the present invention correspondingly provides a multi-strategy operation device for a photovoltaic-storage-charging station based on deep learning, including:

[0027] A data acquisition module for acquiring operation data of the photovoltaic-storage integrated energy station that is multi-source heterogeneous, where the operation data includes status data of photovoltaic panels and energy storage batteries and usage status data of charging piles.

[0028] A data fusion module, which is used to perform data fusion on the operation data based on an improved spatio-temporal heterogeneous graph fusion model and by using an improved gated graph convolutional network to obtain the fused data;

[0029] A feature extraction module, which is used to perform feature extraction operations on the fused data to obtain a high-dimensional feature vector that comprehensively reflects the complex dependency relationships among multi-source heterogeneous operation data;

[0030] A strategy generation module, which is used to reduce the dimension and extract key information from the high-dimensional feature vector through an improved sparse autoencoder to generate a low-dimensional representation, and perform a global search on the low-dimensional representation through an improved hybrid evolutionary algorithm to generate an optimal operation strategy; the optimal operation strategy includes a charging pile pricing strategy, a charge and discharge strategy for energy storage batteries, and a user preferential push strategy;

[0031] A strategy application module, which is used to apply the optimal operation strategy to the operation of the integrated photovoltaic, energy storage and charging station in real time; among them, according to the charging pile pricing strategy, the pricing of the charging piles in the integrated photovoltaic, energy storage and charging station is dynamically adjusted through an intelligent electricity meter, discount information is pushed to the user terminal according to the user preferential push strategy, and the charge and discharge strategy of the energy storage batteries in the integrated photovoltaic, energy storage and charging station is optimized through an energy storage management system according to the charge and discharge strategy of the energy storage batteries.

[0032] As an improvement to the above solution, the data fusion module is specifically used for:

[0033] Using a normalization algorithm that combines logarithmic transformation and linear scaling to perform normalization processing on the operation data, so as to map the operation data to an adapted interval to obtain the preprocessed operation data;

[0034] According to the preprocessed operation data, abstract the photovoltaic panels, energy storage batteries and charging piles of the integrated photovoltaic, energy storage and charging station into nodes of a spatio-temporal heterogeneous graph and expand their attributes, and at the same time define the time edges and space edges of the spatio-temporal heterogeneous graph, and calculate the weights of each edge of the spatio-temporal heterogeneous graph to reflect the spatio-temporal correlation relationship between nodes, and construct a spatio-temporal heterogeneous graph;

[0035] Using an improved gated graph convolutional network to perform layer-by-layer operations on the constructed spatio-temporal heterogeneous graph, and aggregate the local features of the nodes to obtain the fused data.

[0036] As an improvement to the above solution, the feature extraction module is specifically used for:

[0037] Perform reshaping processing on the fused data, construct it into a data cube structure according to different data sources and time series, and at the same time dynamically divide windows according to the data fluctuation conditions to generate adaptable dynamic window data;

[0038] Based on the obtained dynamic window data, evaluate the comprehensive importance of each feature from the aspects of information entropy and feature correlation within each window, screen out important features, and form a screened feature subset;

[0039] For the screened feature subset, build a deep residual attention network, use residual connections to solve the gradient problem, strengthen the expression of key features with the help of the attention mechanism, and then perform global average pooling after processing to obtain a high-dimensional feature vector that comprehensively reflects the complex dependence relationship of multi-source heterogeneous operation data.

[0040] As an improvement to the above solution, the policy generation module is specifically used for:

[0041] First perform preprocessing on the high-dimensional feature vector based on the data distribution characteristics, encode it by constructing an information topology graph and using a unique probability activation function, and at the same time introduce a sparse constraint mechanism based on the information gain rate for training to generate a reduced-dimensional low-dimensional representation;

[0042] Based on the low-dimensional representation, use chaotic mapping to generate an initial population, and the population individuals correspond to different operation strategy combinations;

[0043] Construct a multi-objective fitness function that comprehensively considers the total revenue of the energy station, user satisfaction, and energy utilization rate, and evaluate the fitness of the initial population;

[0044] Use chaotic search to generate perturbation vectors in the neighborhood of the initial population individuals, update the population by comparing the fitness of the new individuals and the original individuals, and achieve global exploration;

[0045] On the basis of the population after chaotic search, introduce a multi-objective game mechanism. Individuals generate new individuals in the neighborhood, play games according to the benefits of different operation goals, adjust their own strategy parameters, and promote the evolution of the population;

[0046] After multiple alternating iterations of chaotic search and multi-objective game, when the termination condition is met, select the strategy corresponding to the individual with the highest fitness in the population to generate the optimal operation strategy including charging pile pricing, energy storage battery charging and discharging, and user preferential push.

[0047] Another embodiment of the present invention provides a multi-strategy operation device for a photovoltaic energy storage charging station based on deep learning, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the multi-strategy operation method for a photovoltaic energy storage charging station based on deep learning described in the above-mentioned embodiment of the invention.

[0048] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0049] First, obtain multi-source heterogeneous operation data covering the status of photovoltaic panels, energy storage batteries, and charging piles to provide a comprehensive information basis for subsequent analysis. An improved spatio-temporal heterogeneous graph fusion model and a gated graph convolutional network are used for data fusion. The improved model can accurately capture the complex correlations between different data sources in the time and space dimensions, overcome the defects of traditional methods that cannot adapt to dynamic changes, and enable the fused data to more accurately reflect the actual operating conditions of the energy station. Feature extraction is performed on the fused data to obtain high-dimensional feature vectors, fully mining the complex dependence relationships between multi-source heterogeneous data and providing a key basis for subsequent strategy formulation. An improved sparse autoencoder is used to reduce the dimension of the high-dimensional feature vectors and extract key information, avoiding data redundancy and improving computational efficiency; a global search is performed on the low-dimensional representation through an improved hybrid evolutionary algorithm, which can quickly find the optimal operation strategy in a complex solution space. This strategy covers charging pile pricing, energy storage battery charging and discharging, and user preferential push, and has strong pertinence and dynamic adaptability. Finally, the optimal strategy is applied to the operation of the energy station in real time, and precise regulation is achieved with the help of smart meters, user terminals, and energy storage management systems. Through the innovative data fusion, feature extraction, and strategy generation technologies in the embodiments of the present invention, the complex dependence relationships between multi-source heterogeneous operation data are deeply mined, the optimal operation strategy is accurately and real-time formulated and applied, and the energy utilization efficiency and intelligent operation level of the integrated photovoltaic energy storage charging station are significantly improved. Description of the Drawings

[0050] Figure 1 is a schematic flowchart of a multi-strategy operation method for a photovoltaic energy storage charging station based on deep learning provided by an embodiment of the present invention;

[0051] Figure 2 is a schematic structural diagram of a multi-strategy operation device for a photovoltaic energy storage charging station based on deep learning provided by an embodiment of the present invention;

[0052] Figure 3 is a schematic structural diagram of a multi-strategy operation device for a photovoltaic energy storage charging station based on deep learning provided by an embodiment of the present invention. Detailed Embodiments

[0053] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0054] See Figure 1 , which is a schematic flowchart of a multi-strategy operation method for a photovoltaic energy storage charging station based on deep learning provided by an embodiment of the present invention. The multi-strategy operation method for a photovoltaic energy storage charging station based on deep learning includes the following steps:

[0055] S10. Obtain the operation data of the integrated photovoltaic-storage-charging energy station, where the operation data includes the status data of photovoltaic panels and energy storage batteries and the usage status data of charging piles.

[0056] S11. Based on the improved spatio-temporal heterogeneous graph fusion model and using the improved gated graph convolutional network, perform data fusion on the operation data to obtain the fused data.

[0057] S12. Perform feature extraction operations on the fused data to obtain a high-dimensional feature vector that comprehensively reflects the complex dependence relationships among the operation data of multi-source heterogeneity.

[0058] S13. Through the improved sparse autoencoder, perform dimensionality reduction and key information extraction on the high-dimensional feature vector to generate a low-dimensional representation, and through the improved hybrid evolutionary algorithm, perform a global search on the low-dimensional representation to generate an optimal operation strategy; the optimal operation strategy includes a charging pile pricing strategy, an energy storage battery charge and discharge strategy, and a user discount push strategy.

[0059] S14. Apply the optimal operation strategy to the operation of the integrated photovoltaic-storage-charging energy station in real time; among them, according to the charging pile pricing strategy, dynamically adjust the pricing of the charging piles of the integrated photovoltaic-storage-charging energy station through an intelligent electricity meter, push discount information to the user terminal according to the user discount push strategy, and optimize the charge and discharge strategy of the energy storage batteries of the integrated photovoltaic-storage-charging energy station through an energy storage management system according to the energy storage battery charge and discharge strategy.

[0060] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0061] First, obtain multi-source heterogeneous operation data covering the status of photovoltaic panels, energy storage batteries, and charging piles, providing a comprehensive information basis for subsequent analysis. An improved spatio-temporal heterogeneous graph fusion model and a gated graph convolutional network are used for data fusion. The improved model can accurately capture the complex correlations between different data sources in the time and space dimensions, overcoming the defect that traditional methods cannot adapt to dynamic changes, enabling the fused data to more accurately reflect the actual operation status of the energy station. Feature extraction is performed on the fused data to obtain high-dimensional feature vectors, fully mining the complex dependence relationships between multi-source heterogeneous data and providing a key basis for subsequent strategy formulation. An improved sparse autoencoder is used to reduce the dimension of the high-dimensional feature vectors and extract key information, avoiding data redundancy and improving computational efficiency; a global search is performed on the low-dimensional representation through an improved hybrid evolutionary algorithm, which can quickly find the optimal operation strategy in a complex solution space. This strategy covers charging pile pricing, energy storage battery charging and discharging, and user preferential push, and has strong pertinence and dynamic adaptability. Finally, the optimal strategy is applied to the operation of the energy station in real time, and precise regulation is achieved with the help of smart meters, user terminals, and energy storage management systems. Through the innovative data fusion, feature extraction, and strategy generation technologies in the embodiments of the present invention, the complex dependence relationships between multi-source heterogeneous operation data are deeply mined, the optimal operation strategy is accurately and real-time formulated and applied, significantly improving the energy utilization efficiency and intelligent operation level of the integrated photovoltaic-energy storage-charging energy station.

[0062] As an example of the above solution, the process of obtaining the operation data of the integrated photovoltaic-energy storage-charging energy station, which is multi-source heterogeneous, and the operation data including the status data of photovoltaic panels, energy storage batteries, and the usage status data of charging piles includes the following steps:

[0063] Step 1: Determine the data types and collection points: Clearly define the types of operation data to be obtained, namely the status data of photovoltaic panels, energy storage batteries, and the usage status data of charging piles. At the same time, determine the specific collection points for each data type.

[0064] Collection points for photovoltaic panel status data: In the photovoltaic panel array, install sensors for each photovoltaic panel group (for example, every 10 photovoltaic panels as a group) to collect data. Install current sensors and voltage sensors at the output end of the photovoltaic panels to obtain the generated current and voltage data; install light intensity sensors and temperature sensors on the surface of the photovoltaic panels to measure the light intensity and the surface temperature of the photovoltaic panels respectively.

[0065] Collection points for energy storage battery status data: In the energy storage battery pack, equip sensors for each battery module. Install voltage sensors and current sensors at the positive and negative poles of the battery module to monitor the charging and discharging voltage and current of the battery; install temperature sensors inside the battery module to monitor the temperature change of the battery in real time; at the same time, obtain data such as the state of charge (SOC) and state of health (SOH) of the battery through the battery management system (BMS).

[0066] Data acquisition points for the charging pile usage status: Install corresponding sensors and monitoring devices on each charging pile. Install current sensors and voltage sensors at the charging interface of the charging pile to record the current and voltage during the charging process; in the control unit of the charging pile, obtain data such as charging time, charging power, and user information (such as user ID, charging start and end times).

[0067] Step 2: Select appropriate data acquisition devices: According to the determined data types and acquisition points, select appropriate data acquisition devices to ensure the accuracy and stability of data acquisition.

[0068] Sensors: For current and voltage sensors, select high-precision, wide-range Hall effect sensors that can accurately measure data within different current and voltage ranges. The light intensity sensor uses a photodiode-type sensor, which features fast response and high sensitivity. The temperature sensor is selected from thermocouple or thermistor sensors according to different measurement ranges and accuracy requirements.

[0069] Data collectors: Use industrial-grade data collectors with multiple analog input and digital input channels, capable of collecting data from multiple sensors simultaneously. The data collector supports multiple communication protocols, such as Modbus, CAN bus, etc., facilitating communication with sensors and subsequent data processing systems.

[0070] Communication devices: Select appropriate communication devices according to the distance between the acquisition point and the data processing center and the communication environment. For acquisition points with a relatively short distance, wired Ethernet communication can be used; for acquisition points with a relatively long distance or a scattered distribution, wireless communication methods such as GPRS, LoRa, etc. are adopted.

[0071] Step 3: Set the data acquisition frequency: Set a reasonable data acquisition frequency according to the change characteristics of different data types and actual operation requirements.

[0072] Photovoltaic panel status data: Since the light intensity and the temperature of the photovoltaic panel change rapidly within a day with time and weather conditions, the data acquisition frequency of light intensity and temperature data is set to once per minute to capture these changes. The data acquisition frequency of the generated current and voltage data is set to once every 5 minutes because their changes are relatively slow.

[0073] Energy storage battery status data: The charge and discharge current and voltage of the battery change significantly during the charge and discharge process, so the data acquisition frequency is set to once every 2 minutes. The temperature, remaining power, and health status of the battery change relatively slowly, and the data acquisition frequency is set to once every 10 minutes.

[0074] Charging pile usage status data: The charging current and voltage change in real time during the charging process. The acquisition frequency is set to once per second to accurately record the charging process. The charging time, charging power, and user information are acquired at the start and end of charging and do not require real-time acquisition.

[0075] Step 4: Data transmission and storage: Transmit the acquired data to the data processing center through a communication device and store it. Among them, data transmission: The data collector packs the acquired data according to the set communication protocol and transmits it to the data processing center through a wired or wireless communication device. During the transmission process, data encryption and verification technologies are used to ensure the security and integrity of the data. Data storage: After receiving the data, the data processing center stores it in a database. You can choose a relational database (such as MySQL) or a non-relational database (such as MongoDB) according to the data structure and query requirements. At the same time, to facilitate subsequent data processing and analysis, the stored data is classified and labeled, for example, classified according to data type, acquisition time, device number, etc.

[0076] For easy understanding, the following example is given here: Suppose a photovoltaic-storage-charging integrated energy station has 100 photovoltaic panels, divided into 10 photovoltaic panel groups; 5 groups of energy storage batteries, each group containing 10 battery modules; and 20 charging piles. During the operation on a certain day, data acquisition is carried out through the above steps.

[0077] Photovoltaic panel data: At 10 am, the light intensity sensor of the 3rd group of photovoltaic panels measures a light intensity of 8000 lux, the temperature sensor measures the surface temperature of the photovoltaic panel as 30 °C, the current sensor measures a generated current of 5 A, and the voltage sensor measures a generated voltage of 220 V. These data are acquired once every minute and transmitted to the data processing center through a wireless communication device and stored in the database.

[0078] Energy storage battery data: At 2 pm, the voltage sensor of the 5th battery module in the 2nd group of energy storage batteries measures a battery voltage of 3.7 V, the current sensor measures a charging current of 2 A, the temperature sensor measures the battery temperature as 25 °C, and the battery management system sends the remaining power of this battery module as 80% and the health status as good. These data are acquired every 2 minutes (current, voltage) or every 10 minutes (temperature, remaining power, health status) and transmitted to the data processing center for storage.

[0079] Charging pile data: At 7 p.m., charging pile No. 15 starts charging an electric vehicle. During the charging process, the current sensor collects the charging current once per second, and the voltage sensor collects the charging voltage once per second. After charging is completed, the control unit of the charging pile records the charging time as 1 hour, the charging power as 20 degrees, and the user ID as 12345. After these data are transmitted to the data processing center, they are stored in the database for subsequent data analysis and operation decision-making.

[0080] As an example of the above solution, the operation data is fused using the improved spatio-temporal heterogeneous graph fusion model and the improved gated graph convolutional network to obtain the fused data, including:

[0081] The operation data is normalized using a normalization algorithm that combines logarithmic transformation and linear scaling, thereby mapping the operation data to an appropriate interval to obtain the preprocessed operation data;

[0082] According to the preprocessed operation data, the photovoltaic panels, energy storage batteries, and charging piles of the integrated photovoltaic energy storage and charging station are abstracted into nodes of a spatio-temporal heterogeneous graph and their attributes are expanded. At the same time, the time edges and space edges of the spatio-temporal heterogeneous graph are defined, and the weights of each edge of the spatio-temporal heterogeneous graph are calculated to reflect the spatio-temporal correlation relationship between nodes, and a spatio-temporal heterogeneous graph is constructed;

[0083] Using the improved gated graph convolutional network, layer-by-layer operations of multiple layers of the constructed spatio-temporal heterogeneous graph are performed to aggregate the local features of the nodes, thereby obtaining the fused data.

[0084] In this embodiment, first, a unique normalization algorithm that combines logarithmic transformation and linear scaling is used to accurately adjust the scale of the operation data and map it to an appropriate interval, effectively solving the problem of difficult processing caused by scale differences in multi-source heterogeneous data and laying a good foundation for subsequent operations. Next, the equipment of the integrated photovoltaic energy storage and charging station is abstracted into nodes of a spatio-temporal heterogeneous graph and their attributes are expanded. At the same time, the time edges and space edges are defined in detail and the weights are calculated to fully explore the spatio-temporal correlation relationship between nodes, and a spatio-temporal heterogeneous graph that can comprehensively reflect the operation logic of the energy station is constructed. Finally, with the help of the improved gated graph convolutional network, multi-layer operations are performed on the spatio-temporal heterogeneous graph to realize the aggregation of node local features, thereby obtaining highly fused data that can reflect the complex dependence relationship between data. In summary, this embodiment can not only effectively integrate multi-source heterogeneous operation data, but also deeply explore the spatio-temporal correlation behind the data, providing high-quality data support for subsequent feature extraction and operation strategy formulation, and improving the management and operation level of the integrated photovoltaic energy storage and charging station.

[0085] Specifically, the working process of this embodiment is as follows:

[0086] (1) Data preprocessing and normalization:

[0087] The multi-source heterogeneous operation data obtained from step S10 have differences in data format, dimension, and value range due to different sources. Let the original data be , where represents the data type is the photovoltaic panel data, is the energy storage battery data, is the charging pile data, represents the th data sample. A normalization method that combines logarithmic transformation and linear scaling is adopted, and the formula is as follows:

[0088] ;

[0089] where is the offset constant set according to the data type to prevent negative numbers or zeros from appearing in the logarithmic operation; is the maximum value of the data type plus a small constant (such as 0.01) for scale adjustment of the logarithmic transformation; is the scaling factor, is the offset, which is set according to the subsequent model requirements and is used to map the normalized data to an appropriate interval. For example, for the light intensity data of the photovoltaic panel, its value range is mapped to the [0,1] interval through this formula for convenient subsequent processing.

[0090] (2) Construction of the improved spatio-temporal heterogeneous graph:

[0091] Node definition and attribute expansion: Abstract the photovoltaic panel, energy storage battery, and charging pile as nodes of the spatio-temporal heterogeneous graph. Each node not only has conventional attributes but also adds unique dynamic attributes. For example, in addition to the rated power, conversion efficiency, etc., the photovoltaic panel node also adds the future light intensity trend value predicted based on time series, and the occlusion coefficient considering the influence of surrounding occlusion, indicating the time step. The energy storage battery node adds the battery health index calculated based on the aging model, and the fluctuation coefficient of the power interaction with the power grid. The charging pile node includes the charging demand probability analyzed based on user behavior, and the competition coefficient of the competition relationship with surrounding charging piles.

[0092] Edge relationship modeling and weight calculation: The edges are divided into time edges and space edges. The time edges connect the data of the same node at different time steps, reflecting the time series characteristics; the space edges connect different types of nodes, reflecting the spatial correlation.

[0093] Temporal edge weight , where represents different time steps, and the calculation formula is as follows:

[0094] ;

[0095] where, and are the values of the -th attribute at time steps and respectively, and are the mean values of all attributes at time steps and respectively, is the number of attributes, is the time decay parameter, which controls the influence degree of time on the edge weight, is a very small constant to prevent the denominator from being zero. This formula reflects the stronger correlation for more recent time by calculating the covariance of data at different time steps and combining with the exponential decay function.

[0096] Spatial edge weight ( represents different types of nodes) is calculated using a method based on information gain and mutual information. First, calculate the mutual information between the attributes of node , and then combine it with the information gain . The formula is:

[0097] ;

[0098] where, is the maximum value of the mutual information between all possible node pairs, which is used to normalize the weight. The mutual information measures the dependence degree between the attributes of two nodes, and the information gain represents the reduction degree of the uncertainty of another node's attribute when the attribute of one node is known. In this way, it can more accurately reflect the association between spatial nodes.

[0099] (3) Improved gated graph convolutional network (GGCN) fusion:

[0100] Construct a multi-layer improved gated graph convolutional network for data fusion. In each layer of GGCN, the feature update formula of the node is as follows:

[0101] ;

[0102] where, represents the feature vector of node at the -th layer, is the node 's set of neighbor nodes, is the node and 's spatial edge weight, is the gating vector, used to control the inflow and outflow of information, and the calculation formula is:

[0103] ;

[0104] Here, is the learnable weight vector, used to adjust the influence degree of neighbor node information on the current node; is the bias term; is the improved activation function, such as the LeakyReLU function that dynamically adjusts the slope according to the data distribution; is the enhancement factor that is dynamically adjusted according to the importance of the node and the current network state, used to highlight key node information. Through the layer-by-layer calculation of the multi-layer GGCN, the local features of the nodes are gradually aggregated to obtain the fused data.

[0105] To facilitate the understanding of this embodiment, the following example is given here: For example, in a comprehensive optical storage and charging energy station located in an industrial park, the power generation data of photovoltaic panels, the charge and discharge data of energy storage batteries, and the usage data of charging piles constitute multi-source heterogeneous operation data. Since the dimensions and value ranges of these data vary greatly, for example, the power generation power of photovoltaic panels may be between dozens of kilowatts and hundreds of kilowatts, while the usage duration of charging piles may be between dozens of minutes and several hours. By fusing the normalization algorithms of logarithmic transformation and linear scaling, these data are uniformly mapped to a suitable interval for subsequent processing. Then, the photovoltaic panels, energy storage batteries, and charging piles are abstracted as nodes of a spatio-temporal heterogeneous graph. Attributes such as light intensity and temperature at different time periods are added to the photovoltaic panel nodes, attributes such as remaining battery power and charge and discharge efficiency are added to the energy storage battery nodes, and attributes such as charging power and the number of waiting vehicles are added to the charging pile nodes. At the same time, time edges are defined to reflect the change correlation of each device's data over time, such as the change relationship of the power generation power of photovoltaic panels at different times; spatial edges are defined to reflect the spatial correlation between devices, such as the energy transmission relationship between photovoltaic panels and energy storage batteries. After calculating the weights of each edge, a spatio-temporal heterogeneous graph is constructed. Finally, the improved gated graph convolutional network is used to operate on this graph, aggregating the local features of each device node to obtain the fused data. For example, the fused data can clearly reflect complex spatio-temporal correlation situations such as the increase in the power generation power of photovoltaic panels, the start of charging of energy storage batteries, and the increase in the usage demand of charging piles in the afternoon with sufficient sunlight, providing a data basis for the operation and management of the energy station.

[0106] As one example of the above solution, the feature extraction operation is performed on the fused data to obtain a high-dimensional feature vector that comprehensively reflects the complex dependency relationship between multi-source heterogeneous operation data, including:

[0107] Reshape the fused data and build it into a data cube structure according to different data sources and time series. At the same time, dynamically divide the windows according to the data fluctuation conditions to generate adaptive dynamic window data.

[0108] Based on the obtained dynamic window data, the comprehensive importance of each feature is evaluated in each window from the information entropy and feature correlation level, and the important features are screened out to form a screened feature subset;

[0109] For the filtered feature subset, a deep residual attention network is built, and the residual connection is used to solve the gradient problem. The attention mechanism is used to strengthen the expression of key features. After processing, global average pooling is performed to obtain a high-dimensional feature vector that comprehensively reflects the complex dependencies between multi-source heterogeneous operation data.

[0110] In this embodiment, first, the fused data is reshaped to construct a unique data cube structure, with different data sources and time series as dimensions, breaking through the traditional data organization form, so that the multi-source attributes and time dynamics of the data can be intuitively presented. At the same time, the dynamic window division mechanism is very innovative, abandoning the fixed window mode, adjusting the window size in real time according to the data fluctuation, ensuring that the generated dynamic window data is closely fitted to the actual data, and effectively capturing the data features under different degrees of fluctuation. Then, the importance of features is evaluated from the dual dimensions of information entropy and feature correlation. This multivariate evaluation method is more comprehensive and scientific than single indicator screening, and can deeply mine the features that play a key role in reflecting the complex dependencies of data, and then screen out important feature subsets. Then, a deep residual attention network is built, and residual connections are used to overcome the gradient problem of deep networks, so as to realize the learning of deep features, focus on key features with the help of attention mechanisms, and highlight important information. Finally, the processed features are integrated through global average pooling to form a high-dimensional feature vector. In summary, this embodiment can not only efficiently process multi-source heterogeneous data and deeply mine the complex connections between data, but also greatly improve the accuracy and effectiveness of feature extraction through innovative window division, multi-dimensional evaluation and deep network structure, providing high-quality data support for subsequent dimensionality reduction and strategy generation.

[0111] Specifically, the working process of this embodiment is as follows:

[0112] (1) Data structure reshaping and dynamic window division:

[0113] Get the fused data After that, due to its multi-source heterogeneous characteristics, the internal dependency relationships in the data vary across different time scales and spatial dimensions. To capture these complex relationships more precisely, we first reshape the data structure. The data is arranged in three dimensions according to different data sources (photovoltaic panels, energy storage batteries, charging piles) and time series to form a data cube , whose dimensions are the number of samples , the number of features , and the time step . Then, a dynamic window partitioning mechanism is introduced. Traditional window partitioning methods often have a fixed window size and are difficult to adapt to the dynamic changes in the data. Here, we dynamically adjust the window size according to the data fluctuations. Define a fluctuation index . For each time step , calculate the sum of the standard deviations of the data within the previous and next time steps as the fluctuation index:

[0114] ;

[0115] where std represents the standard deviation calculation function. According to the size of the fluctuation index , dynamically determine the window size :

[0116] ;

[0117] where is the base window size, is the adjustment coefficient, represents floor function. In this way, a larger window is used in the area with large data fluctuations to capture a wider range of dependency relationships; a smaller window is used in the area with stable data to improve the calculation efficiency.

[0118] (2) Feature selection based on information entropy and correlation:

[0119] Within each dynamic window, feature selection is performed on the data. Traditional feature selection methods are mostly based on a single statistical indicator and are difficult to comprehensively reflect the complex relationships among multi-source heterogeneous data. Here, we combine information entropy and the correlation between features for selection. For each feature within the window, calculate its information entropy :

[0120] ;

[0121] where is the feature at the The probability of taking values in a sample. The information entropy reflects the uncertainty of a feature. The greater the information entropy, the richer the information contained in the feature. At the same time, calculate the mutual information between features to measure the correlation between features:

[0122] ;

[0123] in is the feature at the th sample and the feature at the th sample. Define a comprehensive score to evaluate the importance of each feature:

[0124] ;

[0125] where and are weight coefficients, and . According to the comprehensive score , sort the features, and select the top features with higher scores to form a filtered feature subset .

[0126] (3) Feature extraction based on the deep residual attention network:

[0127] To extract high-dimensional feature vectors that comprehensively reflect the complex dependence relationships of multi-source heterogeneous operation data from the filtered feature subset , construct a deep residual attention network (DRAN). DRAN is stacked by multiple residual attention blocks (RABs). Each RAB contains a residual connection and an attention mechanism.

[0128] Residual connection: For the input feature , after passing through the convolutional layer and the activation function , we get , and then pass through the convolutional layer to get . Through the residual connection, the output , which can alleviate the problem of gradient disappearance and enable the network to learn deeper features. Attention mechanism: On the basis of the residual connection, introduce an attention mechanism to enhance the expression of important features. For the input feature , calculate the attention weight :

[0129] ;

[0130] where and are learnable weight matrices. Multiply the attention weights element-wise with the input features to obtain the attention-enhanced features , where represents element-wise multiplication. After processing by multiple RABs, perform global average pooling on the output of the last RAB to obtain the final high-dimensional feature vector . This high-dimensional feature vector comprehensively reflects the complex dependency relationships among multi-source heterogeneous operation data and can be used for subsequent tasks such as dimensionality reduction and policy generation.

[0131] To facilitate understanding of this embodiment, the following example is given here: For example, in a comprehensive photovoltaic energy storage charging station located in an industrial park, the fused data covers multi-source information such as the power generation efficiency of photovoltaic panels at different times, the state of charge of energy storage batteries, and the busyness of charging piles. After constructing these data into a data cube structure, taking time as the clue and devices such as photovoltaic panels, energy storage batteries, and charging piles as the data source dimensions, the internal hierarchy of the data is clearly presented. When the industrial park enters the production peak period and the electricity demand surges, the power generation efficiency of photovoltaic panels may fluctuate greatly. At this time, the dynamic window partitioning mechanism will quickly adjust the window size according to the data fluctuation situation, expanding the window when the fluctuation is large to capture more information related to the energy supply and demand changes. Within each dynamic window, from the perspective of information entropy analysis, if the information entropy of the state of charge of the energy storage battery is high at a certain moment, it means that this feature has a large uncertainty and may have an important impact on the stable power supply of the energy station; from the aspect of feature correlation, it will be found that there is a certain correlation between the usage frequency of charging piles and the power generation efficiency of photovoltaic panels. Combining the evaluations of these two dimensions, important features such as the power generation efficiency of photovoltaic panels, the state of charge of energy storage batteries, and the usage frequency of charging piles are selected to form a feature subset. Subsequently, these feature subsets are input into the deep residual attention network. The residual connection helps the network learn more complex feature relationships among various parts of the energy station, and the attention mechanism focuses on key features such as a sharp drop in the power generation efficiency of photovoltaic panels to strengthen their expression. Finally, through global average pooling, a high-dimensional feature vector is obtained. This vector comprehensively reflects the complex dependency relationships among photovoltaic panels, energy storage batteries, and charging piles during the production peak period of the industrial park, such as how the energy storage battery supplements power when the photovoltaic panel generates insufficient power and how the usage situation of charging piles will change, providing a core basis for formulating scientific and reasonable operation strategies for the energy station.

[0132] As an example of the above solution, the high-dimensional feature vector is reduced in dimension and key information is extracted by an improved sparse autoencoder to generate a low-dimensional representation, and an improved hybrid evolutionary algorithm is used to perform a global search on the low-dimensional representation to generate an optimal operation strategy; the optimal operation strategy includes a charging pile pricing strategy, a charge and discharge strategy for energy storage batteries, and a user preferential push strategy, including:

[0133] Perform preprocessing on the high-dimensional feature vector based on the data distribution characteristics, encode it by constructing an information topology graph and using a unique probability activation function, and at the same time introduce a sparse constraint mechanism based on the information gain rate for training to generate a reduced-dimensional low-dimensional representation;

[0134] Based on the low-dimensional representation, use chaotic mapping to generate an initial population, and the population individuals correspond to different operation strategy combinations;

[0135] Construct a multi-objective fitness function that comprehensively considers the total revenue of the energy station, user satisfaction, and energy utilization rate, and evaluate the fitness of the initial population;

[0136] Use chaotic search to generate a perturbation vector in the neighborhood of the initial population individuals, update the population by comparing the fitness of the new individuals and the original individuals, and achieve global exploration;

[0137] On the basis of the population after chaotic search, introduce a multi-objective game mechanism. Individuals generate new individuals in the neighborhood, play games according to the benefits of different operation objectives, adjust their own strategy parameters, and promote the evolution of the population;

[0138] After multiple alternating iterations of chaotic search and multi-objective game, when the termination condition is met, select the strategy corresponding to the individual with the highest fitness in the population to generate an optimal operation strategy including charging pile pricing, charge and discharge of energy storage batteries, and user preferential push.

[0139] In this embodiment, an intelligent decision-making system is constructed from high-dimensional feature vectors to optimal operation strategies to solve the problems of multi-source heterogeneous data processing and complex strategy optimization in the operation of integrated energy stations for photovoltaic energy storage and charging. Among them, for the operation of integrated energy stations for photovoltaic energy storage and charging, starting from high-dimensional feature vectors, the optimal operation strategies are generated through dimensionality reduction and global search. The innovation is reflected in multiple key links. During dimensionality reduction, based on the data distribution characteristics, the high-dimensional feature vectors are preprocessed. Constructing an information topology graph can clearly present the data association structure. A unique probability activation function can accurately capture the probability distribution of data features. The sparse constraint mechanism based on the information gain rate removes redundant information, making the generated low-dimensional representation more representative and effectively reducing the computational complexity. In terms of generating the optimal operation strategies, a chaotic mapping is used to generate the initial population, breaking the limitations of traditional random initialization and making the population distribution more global and diverse. A multi-objective fitness function is constructed by comprehensively considering the total revenue of the energy station, user satisfaction, and energy utilization rate to comprehensively measure the advantages and disadvantages of operation strategies. Chaotic search generates perturbation vectors in the neighborhood to update the population, realizing global exploration and avoiding falling into local optima. The multi-objective game mechanism simulates the competition and cooperation among different operation goals, promoting the continuous evolution and optimization of the population. In summary, through the innovative dimensionality reduction and global search methods, this embodiment efficiently extracts key information from complex high-dimensional data. The generated optimal operation strategies can comprehensively balance the interests of multiple aspects of the energy station, improve the operation efficiency and economic benefits, and enhance the adaptability of the energy station in different scenarios.

[0140] Specifically, the following technical improvements have been made in this embodiment compared with the prior art:

[0141] The following technical improvements have been made in the dimensionality reduction stage:

[0142] Preprocessing based on data distribution characteristics: Traditional data preprocessing methods often adopt unified standardization or normalization methods without fully considering the actual distribution of data. In this embodiment, the distribution characteristics of data, such as normal distribution and skewed distribution, are deeply analyzed, and specific preprocessing means are adopted for different distribution types. For example, for data with skewed distribution, through methods such as logarithmic transformation or Box-Cox transformation, it is converted into a form closer to normal distribution, making the data have better stability and comparability in subsequent processing and avoiding feature extraction deviation caused by uneven data distribution.

[0143] Constructing an information topology graph: In the past, data processing rarely analyzed the relationships between data from the perspective of topological structure. In this embodiment, an information topology graph is innovatively constructed, where each feature in the high-dimensional feature vector is regarded as a node in the graph, and the association strength between features is represented by the weight of the edge. This topological structure can intuitively display the complex dependence relationships between data, providing richer information for subsequent coding and feature extraction. For example, by analyzing the information topology graph, potential associations between the power generation of photovoltaic panels and weather factors (such as light intensity, temperature) and the usage frequency of charging piles can be discovered, which helps to capture key information in the data more accurately.

[0144] Unique probability activation function: Traditional activation functions (such as Sigmoid, ReLU, etc.) are deterministic and cannot effectively handle the uncertainty in data. In this embodiment, a unique probability activation function is introduced. This function is based on a probability model and can dynamically adjust the activation state of neurons according to the probability distribution of the input data. For example, when processing photovoltaic panel power generation data, considering the uncertainty of weather, the probability activation function can determine the activation degree of neurons according to the power generation probability under different weather conditions, thereby more accurately reflecting the characteristics and laws of the data and improving the accuracy and robustness of coding.

[0145] Sparse constraint mechanism based on information gain ratio: Traditional sparse constraint mechanisms mainly focus on the sparsity of data while ignoring the information value of features. In this embodiment, sparse constraint is performed based on the information gain ratio, which measures the contribution degree of each feature to the overall data information. During the training process, features with a low information gain ratio are constrained to remove redundant information and retain the most critical features for generating operation strategies. This can not only reduce the dimension of the data, reduce the computational amount, but also improve the quality of the low-dimensional representation, making the subsequent strategy search more efficient.

[0146] The following technical improvements are made in the stage of generating the optimal operation strategy:

[0147] Generating the initial population using chaotic mapping: Traditional genetic algorithms or evolutionary algorithms usually generate the initial population in a random initialization manner, which is prone to uneven population distribution and getting trapped in local optimal solutions. In this embodiment, the randomness, ergodicity, and sensitivity to initial conditions of chaotic mapping are utilized to generate a more extensive and evenly distributed initial population. For example, through chaotic systems such as Logistic mapping or Tent mapping, a set of initial solutions are randomly generated in the search space, and these solutions can cover a wider area, increasing the possibility of finding the global optimal solution.

[0148] Constructing a multi-objective fitness function: Traditional optimization of operation strategies often only considers a single objective, such as the economic benefits or energy utilization rate of an energy station, while ignoring other important factors. In this embodiment, a multi-objective fitness function that comprehensively considers the total revenue of the energy station, user satisfaction, and energy utilization rate is constructed. By reasonably setting the weights of each objective, the advantages and disadvantages of each operation strategy can be comprehensively evaluated. For example, during the peak tourist season, the energy station may pay more attention to user satisfaction to attract more users; while during the energy shortage period, energy utilization rate may become a more important objective. The multi-objective fitness function can dynamically adjust the optimization direction of the strategy according to different scenarios and requirements.

[0149] Chaotic search: Traditional search algorithms are prone to falling into local optimal solutions during the search process and cannot find the global optimal solution. In this embodiment, chaotic search is introduced. By generating a perturbation vector in the neighborhood of the initial population individuals, a small perturbation is made to the current solution, thereby jumping out of the local optimal trap and achieving global exploration. For example, when searching for the charging pile pricing strategy, chaotic search can make a small adjustment based on the current price and try different price combinations to find the optimal price that can both increase revenue and meet user needs.

[0150] Multi-objective game mechanism: The operation of an energy station in reality involves multiple stakeholders and multiple operation objectives, and there are often conflicts and competitions among these objectives. In this embodiment, a multi-objective game mechanism is introduced to simulate the game process among different operation objectives. After each individual generates a new individual in its neighborhood, it conducts a game based on the revenues of different operation objectives and dynamically adjusts its own strategy parameters. For example, when formulating the charge and discharge strategy of a energy storage battery, there may be a conflict between the two objectives of maximizing revenue and maximizing energy utilization rate. Through the multi-objective game mechanism, the individual can find a balance between the two, promoting the continuous evolution of the population and generating better operation strategies.

[0151] In summary, in the dimensionality reduction stage, the generated low-dimensional representation can accurately retain the key information in the high-dimensional feature vector, while reducing the dimensionality of the data, decreasing the computational complexity, and improving the efficiency of subsequent strategy search. In the stage of generating the optimal operation strategy, it can comprehensively balance the multiple interests of the energy station. The generated strategy not only considers the economic benefits of the energy station but also takes into account user satisfaction and energy utilization rate, improving the operation efficiency and competitiveness of the energy station, enabling it to better adapt to different market environments and user needs.

[0152] Specifically, the working process of this embodiment is as follows:

[0153] (1) Dimensionality reduction based on an improved sparse autoencoder with information topology and probabilistic activation:

[0154] Step 1: Data preprocessing and distribution adaptation: Obtain the high-dimensional feature vector output in step S12 , these vectors integrate the complex dependencies of multi-source heterogeneous data in the integrated energy station of photovoltaic energy storage and charging. To address the scale differences and distribution characteristics between different features, first calculate each feature 's skewness and kurtosis . Skewness is used to measure the degree of asymmetry of the data distribution, and kurtosis reflects the steepness of the data distribution. The mean and standard deviation of each feature are obtained through statistical calculations. According to the formula , each feature is standardized. This formula can not only unify the data scale but also fine-tune the standardization process according to the data distribution characteristics. For example, for features with a large skewness, the standardization result is corrected through this term to make it more suitable for subsequent coding operations, thus obtaining the preprocessed feature vector .

[0155] Step 2: Construct the information topology graph and calculate the encoding layer:

[0156] Based on the preprocessed feature vector , construct an information topology graph. The nodes in the graph correspond to each feature, and the weights of the edges are determined according to the mutual information between the features. Mutual information is used to measure the degree of dependence between two features. By calculating the mutual information between different feature pairs, the association strength between features can be clarified. During the encoding process, for the -layer encoding layer, the output of the neuron is calculated according to the formula . Among them, is the weight from the -th neuron to the -th neuron in the -th layer, is the number of neurons in the -th layer, (the number of neurons in the input layer); is the bias of the -th neuron in the -th layer; is a newly designed probability activation function, defined as , is a random number obeying the standard normal distribution, is a parameter controlling the degree of random perturbation (for example ). represents the features and The mutual information between them. In this way, the encoding layer can dynamically adjust the weight of information transmission according to the correlation strength between features. At the same time, the randomness introduced by the probability activation function helps the model jump out of the local optimal solution, so as to obtain the output of the encoding layer of the th layer

[0157] Step 3: Strengthen sparse constraint and training:

[0158] To highlight key information, a constraint term based on the information gain rate is introduced on the basis of the traditional sparse constraint. For the th neuron in the th layer, calculate its information gain rate . The information gain rate is obtained by calculating the contribution degree of this neuron to the reduction of the reconstruction error under different input data. Minimize the objective function: for model training. Among them, is the number of training samples, is the th input sample, is the output sample after decoding; is the weight of the sparse penalty term (e.g., ), which is used to balance the reconstruction error and sparsity; is the expected information gain rate value, which is set according to the data complexity and the key information extraction target. By continuously adjusting the weight parameters of the encoding layer and the decoding layer, the value of the objective function is gradually reduced, prompting the neuron to focus on the key information that has the greatest impact on the reconstruction error while ensuring the reconstruction accuracy. After multiple rounds of training, the output of the innermost hidden layer is the low-dimensional representation after dimensionality reduction.

[0159] (2) Improved hybrid evolutionary algorithm generation strategy based on chaotic search and multi-objective game:

[0160] Step 1: Initial population generation and strategy mapping:

[0161] Generate an initial population based on the low-dimensional representation after dimensionality reduction, and set the population size to (e.g., ). Each individual corresponds to a combination of operation strategies. For the charging pile pricing strategy, the individual encoding includes the proportional factors of price adjustment in different time periods (such as peak, flat peak, and valley); the encoding of the energy storage battery charge and discharge strategy covers the charge and discharge decisions under different electricity thresholds; the encoding of the user preferential push strategy includes information such as the type of preference (such as discount, free charging duration), the intensity of preference, and the push timing. Use chaotic mapping to generate initial values to ensure the diversity of the initial population. For example, use the Logistic chaotic mapping , where Take 4 (in a chaotic state). Generate a series of chaotic sequences through multiple iterations and map them to the value range of the corresponding policy parameters. Assume that the value range of the price adjustment ratio factor during peak hours of the charging pile is [0.8, 1.2], and linearly transform the chaotic sequence values (where , is the chaotic sequence value) to map it to this range to obtain the individuals of the initial population and form the initial population .

[0162] Step 2: Construction and evaluation of the multi-objective fitness function:

[0163] Design a fitness function to evaluate the pros and cons of each individual, considering multiple operation objectives of the energy station comprehensively. These include the total revenue of the energy station , user satisfaction and energy utilization rate .

[0164] The total revenue is obtained by calculating the charging pile revenue, photovoltaic power generation sales revenue, charge-discharge price difference of energy storage batteries, etc. For example, the charging pile revenue is calculated based on the charging prices and charging volumes at different times, the photovoltaic power generation sales revenue is calculated based on the power generation volume and the grid purchase price, and the charge-discharge price difference of the energy storage battery is calculated based on the charge-discharge power and the price difference. User satisfaction is quantitatively evaluated based on factors such as the user's charging waiting time and acceptance of charging prices. It can be through questionnaire surveys or statistical analysis of user behavior data, converting the user's feedback on waiting time and price into corresponding satisfaction scores. The energy utilization rate is measured by indicators such as the photovoltaic power consumption rate and the charge-discharge efficiency of the energy storage battery. The photovoltaic power consumption rate is the ratio of the power of photovoltaic power generation used in the station and transmitted to the grid to the total power generation volume, and the charge-discharge efficiency of the energy storage battery is the ratio of the discharge power to the charge power.

[0165] The fitness function is expressed as: , , are the minimum and maximum values of each indicator in the training samples respectively; is the weight coefficient (for example ), which is adjusted according to the actual operation focus of the energy station to balance the importance of different indicators in the optimization process. Calculate the fitness value for each individual in the initial population to obtain the fitness evaluation result.

[0166] Step 3: Chaotic search stage

[0167] At the beginning of the algorithm, chaotic search is used for global exploration. For the current individual , a set of chaotic perturbation vectors are generated through chaotic mapping . For example, the Tent chaotic mapping (where ) is used to generate a series of chaotic values. These chaotic values are mapped according to the variation range of the policy parameters to obtain the chaotic perturbation vectors . Suppose the variation range of the price adjustment ratio factor during the off-peak period in the charging pile pricing strategy is [-0.1, 0.1]. The chaotic values are mapped to this range through a linear transformation (where , is the chaotic value) to obtain the perturbation values corresponding to the policy parameters, and then the chaotic perturbation vectors are formed.

[0168] A new individual is obtained, and the fitness of is calculated . If , then the individual is updated to . By continuously performing chaotic search and updating on the individuals in the population, the algorithm can quickly traverse the solution space in the initial stage, search for better regions, and obtain the population updated through chaotic search.

[0169] Step Four: Multi-objective game optimization stage:

[0170] After chaotic search, a multi-objective game mechanism is introduced. The population individuals are regarded as game participants, and different operation objectives are regarded as different strategies of the game. For each individual in the population updated through chaotic search, a series of new individuals are generated within its neighborhood. The generation of the neighborhood can be achieved by performing small-range random perturbations on the policy parameters of the individual . For example, for the charging pile pricing policy parameters, new values are randomly generated within the range near their current values to obtain new individuals .

[0171] Individuals play games by comparing their payoffs under different objectives. For example, under the total revenue objective, individuals and compare and ; under the user satisfaction objective, compare and etc. According to the game results, individuals adjust their own policy parameters and evolve towards a better direction. For example, if individual Superior to under the total revenue target , but inferior to under the user satisfaction target , then according to the gap between the two and the weight coefficient, for the charging pile pricing strategy and the parameters of the user discount push strategy are adjusted. Assume the total revenue weight , the user satisfaction weight , if , , then according to the formula calculate the adjustment amount of the charging pile pricing strategy parameters (assuming that this parameter has a linear relationship with revenue and satisfaction), for the relevant parameters are adjusted to obtain the population optimized by multi-objective game.

[0172] Step 5: Iteration and Strategy Determination After multiple alternating iterations of chaotic search and multi-objective game, that is, repeat steps three and four in the process of "Improved Hybrid Evolutionary Algorithm for Generating Strategies Based on Chaotic Search and Multi-Objective Game". Each iteration can make the individuals in the population evolve continuously in a better direction. As the number of iterations increases, the population gradually converges. When a certain termination condition is met (for example, the fitness value of the optimal individual in the population has not improved significantly for several consecutive iterations), the operation strategy corresponding to the individual with the highest fitness in the population is the generated optimal operation strategy. These strategies cover the charging pile pricing strategy, the charge and discharge strategy of energy storage batteries, and the user discount push strategy, etc., to achieve the efficient operation of the integrated energy station for photovoltaic energy storage and charging.

[0173] To facilitate the understanding of this embodiment, the following examples are given: Suppose there is a comprehensive energy station for photovoltaic energy storage and charging located in an industrial center. This area has a large flow of people, significant fluctuations in the charging demand for electric vehicles, and large differences in grid electricity prices at different times. After obtaining the high-dimensional feature vector reflecting the operation of the energy station, preprocessing is carried out based on the characteristics of data distribution. For example, it is found that the usage frequency of charging piles shows an obvious skewed distribution during the morning and evening rush hours on weekdays, and it is transformed into data closer to a normal distribution through logarithmic transformation for subsequent processing. When constructing the information topology graph, features such as the power generation of photovoltaic panels, the battery level of energy storage, the usage frequency of charging piles, and the grid electricity price are used as nodes. Through analysis, it is found that there is a strong correlation between the power generation of photovoltaic panels and the light intensity and time, and there is also a certain correlation between the usage frequency of charging piles and the grid electricity price and time. Strong edges are established between the corresponding nodes in the information topology graph. A unique probability activation function coding is adopted. Considering the impact of weather changes on the power generation of photovoltaic panels, neurons are dynamically activated according to the power generation probability under different weather conditions to more accurately reflect the data characteristics. A sparse constraint mechanism based on information gain rate is introduced to remove some features that have less impact on the operation strategy, such as occasional equipment failure information, and generate a low-dimensional representation. Based on the low-dimensional representation, an initial population is generated using a chaotic mapping. Each individual represents a combination of operation strategies, including the charging pile pricing strategy, the charge and discharge strategy of energy storage batteries, and the user preferential push strategy. A multi-objective fitness function is constructed, and weights of 0.5, 0.3, and 0.2 are set for the total revenue of the energy station, user satisfaction, and energy utilization rate respectively. In the chaotic search stage, perturbations are made to the neighborhood of each individual, such as fine-tuning the charging pile pricing or the charge and discharge time of energy storage batteries. For example, during the low grid electricity price period, the charging amount of energy storage batteries is appropriately increased; during the peak period, the charging pile pricing is increased and some preferential activities are launched to balance revenue and user satisfaction. By comparing the fitness of the new individual with that of the original individual, the population is updated. In the multi-objective game mechanism, the strategies represented by different individuals play games to achieve their respective goals. For example, the strategy represented by one individual pays more attention to the revenue of the energy station, and the strategy represented by another individual pays more attention to user satisfaction. During the game process, they will dynamically adjust their strategy parameters according to the revenue of different operation goals. After multiple alternating iterations of chaotic search and multi-objective game, when the termination condition is met, the strategy corresponding to the individual with the highest fitness in the population is selected. The finally generated optimal operation strategy may be to purchase electricity from the grid at a lower price to charge the energy storage battery during the low grid electricity price period; during the peak period, increase the charging pile pricing, but at the same time launch preferential activities such as cashback for recharge to attract users; the energy storage battery supplies power to the charging pile in cooperation with the photovoltaic panel during the peak period to improve energy utilization rate and achieve the best balance of the energy station's revenue, user satisfaction, and energy utilization rate.

[0174] As an example of the above solution, the optimal operation strategy is applied to the operation of the integrated photovoltaic energy storage charging station in real time; among them, according to the charging pile pricing strategy and through the smart meter, the pricing of the charging piles in the integrated photovoltaic energy storage charging station is dynamically adjusted, discount information is pushed to the user terminal according to the user preference push strategy, and according to the charge and discharge strategy of the energy storage battery and through the energy storage management system, the charge and discharge strategy of the energy storage battery in the integrated photovoltaic energy storage charging station is optimized, including the following steps:

[0175] (1) Dynamically adjust the charging pile pricing by the smart meter:

[0176] Real-time data collection: The smart meter continuously collects multi-source data, including the real-time grid electricity price, which is obtained in real time from the grid operator through a dedicated communication interface and can accurately reflect the changes in the grid-side power cost at different times; the power generation power of the photovoltaic panel is monitored in real time, and the power generation data per second is obtained through the data line connected to the photovoltaic panel control system to understand the real-time supply of photovoltaic energy; and the real-time battery level of the energy storage battery is obtained through communication with the energy storage management system to clearly know the amount of energy that can be allocated by the energy storage; at the same time, the real-time power consumption load of the charging pile is recorded, and the real-time power data is obtained from the power monitoring module inside the charging pile to know the current charging demand intensity.

[0177] Pricing strategy matching: The collected data is matched and analyzed with the charging pile pricing strategy generated by the improved hybrid evolutionary algorithm. For example, if the pricing strategy stipulates that the charging pile price is increased when the grid electricity price is high and the photovoltaic power generation is insufficient, the smart meter judges that the current grid electricity price is higher than the set threshold, and the photovoltaic power can only meet 30% of the in-station electricity demand, then the price increase mechanism is triggered. If the strategy is to reduce the price during the low electricity consumption period and when the energy storage battery level is sufficient, when the smart meter detects that the grid electricity price is in the low valley period at night and the energy storage battery level reaches more than 80% at the same time, the price reduction operation is executed.

[0178] Price update and issuance: The smart meter calculates the adjusted charging pile price according to the analysis result and sends the new price information to the control unit of the charging pile through a standard communication protocol (such as ModbusTCP). After receiving the instruction, the control unit quickly updates the price information on the operation interface and display screen of the charging pile to ensure that the user can see the latest pricing when accessing the charging device.

[0179] (2) Push discount information to the user terminal:

[0180] User Portrait Construction: The energy station operation platform integrates data such as users' registration information and historical charging records on the platform. Through data mining techniques, it analyzes users' charging time patterns. For example, some users are accustomed to charging before going to work on weekday mornings, while others often charge on weekend evenings; it counts the charging frequency to distinguish high-frequency users (such as those who charge more than 3 times a week) and low-frequency users; it calculates the charging power to understand the scale of different users' electricity consumption demands. Combining this information, it constructs a detailed portrait for each user.

[0181] Discount Plan Formulation: Generate personalized discount information based on the user portrait and the established user preferential push strategy. For high-frequency and high-charging users, it may provide coupons such as "minus a certain amount when reaching a certain amount", like "minus 50 yuan when reaching 200 yuan"; for users who often charge during off-peak electricity consumption periods, it gives a discount rate preference, such as "20% off for charging during off-peak periods"; for newly registered users, it launches an activity of reducing a certain amount for the first charge, such as "new users get an immediate 10 yuan discount for the first charge".

[0182] Information Push Implementation: Through the push service of the official mobile application of the energy station, the generated discount information is accurately pushed to the corresponding users. For users who have not installed the application, the discount code and usage instructions are sent via text message. After receiving the information, when users charge next time, they can enter the discount code in the application or on the charging pile operation interface to enjoy the corresponding discount.

[0183] (3) The energy storage management system optimizes the charge and discharge strategies of energy storage batteries:

[0184] Battery Status Monitoring: The energy storage management system real-time monitors multiple key parameters of the energy storage battery. Through built-in sensors and communication modules, it obtains the remaining battery capacity (SOC), accurate to two decimal places, to reflect the energy storage level of the battery in real time; it monitors the battery voltage to ensure it is within the normal working range and prevent overvoltage or undervoltage from damaging the battery; it measures the battery temperature because temperature has a significant impact on battery performance and life, and obtains the average temperature through temperature sensors distributed at different positions of the battery pack; it monitors the charge and discharge current and controls the current size to ensure the safe and stable operation of the battery.

[0185] Charge and Discharge Strategy Execution: Based on the charge and discharge strategies generated by the improved algorithm, the energy storage management system formulates specific implementation plans. When the photovoltaic power generation is higher than the in-station electricity demand and the battery power is lower than 80%, the system controls the battery to charge with an appropriate charging current, such as setting it to 0.2C (C is the battery capacity) of the battery rated capacity, which can not only ensure the charging efficiency but also avoid damage to the battery caused by excessive current. When the grid electricity price is high and the photovoltaic power generation is insufficient, it preferentially uses the energy storage battery to supply power to the charging pile and adjusts the discharge current according to the real-time electricity demand. For example, during peak electricity consumption periods, if the total power demand of the charging pile is 50kW, the energy storage system, based on the current state of the battery, outputs a stable current to meet this power demand.

[0186] Strategic dynamic adjustment: The energy storage management system continuously evaluates the execution effect of the charge and discharge strategy. If it is found during the charging process that the battery temperature rises too fast and exceeds the set safety temperature threshold (such as 40 °C), the system automatically reduces the charging current to prevent the battery from overheating; if the power demand of the energy station suddenly increases and exceeds the current available power of the energy storage battery, the system quickly adjusts the discharge strategy and tries to supplement part of the power from the power grid on the premise of ensuring battery safety to ensure stable power supply for the charging piles.

[0187] See Figure 2 , which is a schematic structural diagram of a multi-strategy operation device for a photovoltaic energy storage charging station based on deep learning provided by an embodiment of the present invention. The multi-strategy operation device for a photovoltaic energy storage charging station based on deep learning includes:

[0188] A data acquisition module 10, configured to acquire operation data of a multi-source heterogeneous photovoltaic energy storage charging station, where the operation data includes status data of photovoltaic panels and energy storage batteries and usage status data of charging piles;

[0189] A data fusion module 11, configured to perform data fusion on the operation data based on an improved spatio-temporal heterogeneous graph fusion model and using an improved gated graph convolutional network to obtain fused data;

[0190] A feature extraction module 12, configured to perform feature extraction operations on the fused data to obtain a high-dimensional feature vector that comprehensively reflects the complex dependence relationship between the operation data of multi-source heterogeneity;

[0191] A strategy generation module 13, configured to reduce the dimension and extract key information from the high-dimensional feature vector through an improved sparse autoencoder to generate a low-dimensional representation, and perform a global search on the low-dimensional representation through an improved hybrid evolutionary algorithm to generate an optimal operation strategy; the optimal operation strategy includes a charging pile pricing strategy, an energy storage battery charge and discharge strategy, and a user discount push strategy;

[0192] A strategy application module 14, configured to apply the optimal operation strategy to the operation of the photovoltaic energy storage charging station in real time; wherein, according to the charging pile pricing strategy, the pricing of the charging piles of the photovoltaic energy storage charging station is dynamically adjusted through an intelligent electricity meter, discount information is pushed to the user terminal according to the user discount push strategy, and the charge and discharge strategy of the energy storage battery of the photovoltaic energy storage charging station is optimized according to the energy storage battery charge and discharge strategy through an energy storage management system.

[0193] Compared with the prior art, the embodiment of the present invention has the following beneficial effects:

[0194] Embodiments of the present invention achieve in-depth mining of the complex dependency relationships among multi-source heterogeneous operation data through innovative data fusion, feature extraction, and strategy generation technologies, accurately and real-time formulate and apply optimal operation strategies, and significantly improve the energy utilization efficiency and intelligent operation level of the integrated photovoltaic energy storage charging station.

[0195] As an example of the above solution, the data fusion module is specifically used for:

[0196] Perform normalization processing on the operation data using a normalization algorithm that combines logarithmic transformation and linear scaling, thereby mapping the operation data to an adapted interval to obtain preprocessed operation data;

[0197] According to the preprocessed operation data, abstract the photovoltaic panels, energy storage batteries, and charging piles of the integrated photovoltaic energy storage charging station into nodes of a spatio-temporal heterogeneous graph and expand their attributes. At the same time, define the time edges and space edges of the spatio-temporal heterogeneous graph, and calculate the weights of each edge of the spatio-temporal heterogeneous graph to reflect the spatio-temporal correlation relationship between nodes, and construct a spatio-temporal heterogeneous graph through graphing;

[0198] Use an improved gated graph convolutional network to perform layer-by-layer operations on the constructed spatio-temporal heterogeneous graph, aggregate the local features of the nodes, and thus obtain fused data.

[0199] As an example of the above solution, the feature extraction module is specifically used for:

[0200] Perform reshaping processing on the fused data, construct it into a data cube structure according to different data sources and time series, and dynamically divide windows according to the data fluctuation conditions to generate adaptable dynamic window data;

[0201] Based on the obtained dynamic window data, evaluate the comprehensive importance of each feature from the aspects of information entropy and feature correlation within each window, screen out important features, and form a screened feature subset;

[0202] For the screened feature subset, build a deep residual attention network, use residual connections to solve the gradient problem, strengthen the expression of key features with the help of the attention mechanism, and then perform global average pooling after processing to obtain a high-dimensional feature vector that comprehensively reflects the complex dependency relationships of multi-source heterogeneous operation data.

[0203] As an example of the above solution, the strategy generation module is specifically used for:

[0204] Perform preprocessing on the high-dimensional feature vector based on the data distribution characteristics, encode it by constructing an information topology graph and using a unique probability activation function, and at the same time introduce a sparse constraint mechanism based on the information gain rate for training to generate a reduced-dimensional low-dimensional representation;

[0205] Based on the low-dimensional representation, use a chaotic map to generate an initial population, where the individuals in the population correspond to different combinations of operation strategies;

[0206] Construct a multi-objective fitness function that comprehensively considers the total revenue of the energy station, user satisfaction, and energy utilization rate, and evaluate the fitness of the initial population;

[0207] Use chaotic search to generate perturbation vectors in the neighborhood of the individuals in the initial population, and update the population by comparing the fitness of the new individuals with that of the original individuals to achieve global exploration;

[0208] On the basis of the population after chaotic search, introduce a multi-objective game mechanism. Individuals generate new individuals in the neighborhood, play games based on the benefits of different operation objectives, adjust their own strategy parameters, and promote the evolution of the population;

[0209] After multiple alternating iterations of chaotic search and multi-objective game, when the termination condition is met, select the strategy corresponding to the individual with the highest fitness in the population to generate the optimal operation strategy including charging pile pricing, energy storage battery charging and discharging, and user preferential push.

[0210] See Figure 3 , which is a schematic diagram of a multi-strategy operation device for a photovoltaic energy storage charging station based on deep learning provided by an embodiment of the present invention. The multi-strategy operation device for a photovoltaic energy storage charging station based on deep learning in this embodiment includes: a processor 100, a memory 101, and a computer program stored in the memory 101 and executable on the processor 100, such as a multi-strategy operation program for a photovoltaic energy storage charging station based on deep learning. When the processor 100 executes the computer program, the steps in the above-mentioned various method embodiments of the multi-strategy operation for a photovoltaic energy storage charging station based on deep learning are implemented. Alternatively, when the processor 100 executes the computer program, the functions of each module / unit in the above-mentioned various device embodiments are implemented.

[0211] Exemplarily, the computer program can be divided into one or more modules / units. The one or more modules / units are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing specific functions, and these instruction segments are used to describe the execution process of the computer program in the multi-strategy operation device for a photovoltaic energy storage charging station based on deep learning.

[0212] The multi-strategy operation device of the optical storage and charging station based on deep learning may be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The multi-strategy operation device of the optical storage and charging station based on deep learning may include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the schematic diagram is only an example of the multi-strategy operation device of the optical storage and charging station based on deep learning, and does not constitute a limitation on the multi-strategy operation device of the optical storage and charging station based on deep learning. It may include more or fewer components than those shown in the figure, or combine some components, or different components. For example, the multi-strategy operation device of the optical storage and charging station based on deep learning may also include input / output devices, network access devices, buses, etc.

[0213] The so-called processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The processor is the control center of the multi-strategy operation device of the optical storage and charging station based on deep learning, and connects various parts of the entire multi-strategy operation device of the optical storage and charging station based on deep learning through various interfaces and lines.

[0214] The memory can be used to store the computer programs and / or modules. The processor realizes various functions of the multi-strategy operation device of the optical storage and charging station based on deep learning by running or executing the computer programs and / or modules stored in the memory, and by calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. Among them, the program storage area may store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phone book, etc.), etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0215] Among them, if the module / unit integrated with the multi-strategy operation device of the photovoltaic energy storage charging station based on deep learning is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-mentioned embodiment methods of the present invention, it can also be completed by a computer program instructing relevant hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-mentioned various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice within the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0216] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the attached drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0217] The above is the preferred implementation manner of the present invention. It should be pointed out that for those of ordinary skill in the art in the technical field, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A multi-strategy operation method for a photovoltaic storage and charging station based on deep learning, characterized in that: include: Acquire the multi-source heterogeneous operation data of the photovoltaic storage and charging integrated energy station, which includes the status data of photovoltaic panels, energy storage batteries and the usage status data of charging piles; Based on an improved spatiotemporal heterogeneous graph fusion model and using an improved gated graph convolutional network, data fusion is performed on the operation data to obtain fused data; Perform feature extraction on the fused data to obtain a high-dimensional feature vector that comprehensively reflects the complex dependencies between multi-source heterogeneous operational data; The high-dimensional feature vector is reduced in dimension and key information is extracted by an improved sparse autoencoder to generate a low-dimensional representation, and the low-dimensional representation is globally searched by an improved hybrid evolutionary algorithm to generate an optimal operation strategy; the optimal operation strategy includes a charging pile pricing strategy, an energy storage battery charging and discharging strategy, and a user discount push strategy; Apply the optimal operation strategy to the operation of the photovoltaic, storage and charging integrated energy station in real time; wherein, according to the charging pile pricing strategy and through the smart meter, the pricing of the charging piles of the photovoltaic, storage and charging integrated energy station is dynamically adjusted, according to the user preferential push strategy, discount information is pushed to the user terminal, and according to the energy storage battery charging and discharging strategy, the charging and discharging strategy of the energy storage battery of the photovoltaic, storage and charging integrated energy station is optimized through the energy storage management system; Wherein, the operation data is fused based on the improved spatiotemporal heterogeneous graph fusion model and using the improved gated graph convolutional network to obtain fused data, including: The operation data is normalized using a normalization algorithm that combines logarithmic transformation and linear scaling, thereby mapping the operation data to the adaptation interval to obtain the preprocessed operation data; According to the preprocessed operation data, the photovoltaic panels, energy storage batteries and charging piles of the photovoltaic storage and charging integrated energy station are abstracted into nodes of a spatiotemporal heterogeneous graph and their attributes are expanded. At the same time, the time edges and space edges of the spatiotemporal heterogeneous graph are defined, and the weights of each edge of the spatiotemporal heterogeneous graph are calculated to reflect the spatiotemporal correlation between the nodes, and the spatiotemporal heterogeneous graph is obtained by composing the graph; The improved gated graph convolutional network is used to perform layer-by-layer operations on the constructed spatiotemporal heterogeneous graph, and the local features of the nodes are aggregated to obtain fused data; wherein, in each layer of the improved gated graph convolutional network, the feature update formula of the node is as follows: ; in, Representation Node In the The feature vector of the layer, Is a node The set of neighbor nodes of Is a node and The spatial edge weights between is the gating vector, used to control the inflow and outflow of information; Based on the node The importance of and the enhancement factor of dynamic adjustment of the current network status; The feature extraction operation is performed on the fused data to obtain a high-dimensional feature vector that comprehensively reflects the complex dependency relationship between multi-source heterogeneous operation data, including: Reshape the fused data and build it into a data cube structure according to different data sources and time series. At the same time, dynamically divide the windows according to the data fluctuation conditions to generate adaptive dynamic window data. Based on the obtained dynamic window data, the comprehensive importance of each feature is evaluated in each window from the information entropy and feature correlation level, and the important features are screened out to form a screened feature subset; For the filtered feature subset, a deep residual attention network is built, residual connections are used to solve the gradient problem, and the key feature expression is strengthened with the help of the attention mechanism. After processing, global average pooling is performed to obtain a high-dimensional feature vector that comprehensively reflects the complex dependencies of multi-source heterogeneous operation data; The improved sparse autoencoder is used to reduce the dimension of the high-dimensional feature vector and extract key information to generate a low-dimensional representation, and the low-dimensional representation is globally searched by an improved hybrid evolutionary algorithm to generate an optimal operation strategy; the optimal operation strategy includes a charging pile pricing strategy, an energy storage battery charging and discharging strategy, and a user discount push strategy, including: The high-dimensional feature vector is first preprocessed based on data distribution characteristics, and a low-dimensional representation after dimensionality reduction is generated by constructing an information topology map and encoding it using a unique probability activation function, while introducing a sparse constraint mechanism based on information gain rate for training; Based on the low-dimensional representation, an initial population is generated using chaotic mapping, and individuals in the population correspond to different combinations of operation strategies; Construct a multi-objective fitness function that comprehensively considers the total revenue of the energy station, user satisfaction and energy utilization rate, and evaluate the fitness of the initial population; Use chaos search to generate disturbance vectors in the neighborhood of the initial population individuals, and update the population by comparing the fitness of the new individuals with the original individuals to achieve global exploration; Based on the population after chaotic search, a multi-objective game mechanism is introduced. Individuals generate new individuals in the neighborhood, play games based on different operating target benefits, adjust their own strategy parameters, and promote population evolution. After multiple iterations of chaotic search and multi-objective game, when the termination condition is met, the strategy corresponding to the individual with the highest fitness in the population is selected to generate the optimal operation strategy including charging pile pricing, energy storage battery charging and discharging, and user discount push.

2. A multi-strategy operation device for a photovoltaic storage and charging station based on deep learning, characterized in that: include: The data acquisition module is used to obtain the multi-source heterogeneous operation data of the photovoltaic storage and charging integrated energy station, and the operation data includes the status data of photovoltaic panels, energy storage batteries and the usage status data of charging piles; A data fusion module, used for fusing the operation data based on an improved spatiotemporal heterogeneous graph fusion model and using an improved gated graph convolutional network to obtain fused data; The feature extraction module is used to perform feature extraction operations on the fused data to obtain a high-dimensional feature vector that comprehensively reflects the complex dependencies between multi-source heterogeneous operational data; A strategy generation module, which is used to reduce the dimension and extract key information of the high-dimensional feature vector by an improved sparse autoencoder to generate a low-dimensional representation, and to perform a global search on the low-dimensional representation by an improved hybrid evolutionary algorithm to generate an optimal operation strategy; the optimal operation strategy includes a charging pile pricing strategy, an energy storage battery charging and discharging strategy, and a user discount push strategy; A strategy application module, for applying the optimal operation strategy to the operation of the photovoltaic, storage and charging integrated energy station in real time; wherein, the pricing of the charging piles of the photovoltaic, storage and charging integrated energy station is dynamically adjusted through smart meters according to the charging pile pricing strategy, discount information is pushed to user terminals according to the user preferential push strategy, and the charging and discharging strategy of the energy storage battery of the photovoltaic, storage and charging integrated energy station is optimized through an energy storage management system according to the energy storage battery charging and discharging strategy; Wherein, the data fusion module is specifically used for: The operation data is normalized using a normalization algorithm that combines logarithmic transformation and linear scaling, thereby mapping the operation data to the adaptation interval to obtain the preprocessed operation data; According to the preprocessed operation data, the photovoltaic panels, energy storage batteries and charging piles of the photovoltaic storage and charging integrated energy station are abstracted into nodes of a spatiotemporal heterogeneous graph and their attributes are expanded. At the same time, the time edges and space edges of the spatiotemporal heterogeneous graph are defined, and the weights of each edge of the spatiotemporal heterogeneous graph are calculated to reflect the spatiotemporal correlation between the nodes, and the spatiotemporal heterogeneous graph is obtained by composing the graph; The improved gated graph convolutional network is used to perform layer-by-layer operations on the constructed spatiotemporal heterogeneous graph, and the local features of the nodes are aggregated to obtain fused data; wherein, in each layer of the improved gated graph convolutional network, the feature update formula of the node is as follows: ； in, Representation Node In the The feature vector of the layer, Is a node The set of neighbor nodes of Is a node and The spatial edge weights between is the gating vector, used to control the inflow and outflow of information; Based on the node The importance of and the enhancement factor of dynamic adjustment of the current network status; The feature extraction module is specifically used for: Reshape the fused data and build it into a data cube structure according to different data sources and time series. At the same time, dynamically divide the windows according to the data fluctuation conditions to generate adaptive dynamic window data. Based on the obtained dynamic window data, the comprehensive importance of each feature is evaluated in each window from the information entropy and feature correlation level, and the important features are screened out to form a screened feature subset; For the filtered feature subset, a deep residual attention network is built, residual connections are used to solve the gradient problem, and the key feature expression is strengthened with the help of the attention mechanism. After processing, global average pooling is performed to obtain a high-dimensional feature vector that comprehensively reflects the complex dependencies of multi-source heterogeneous operation data; The strategy generation module is specifically used for: The high-dimensional feature vector is first preprocessed based on data distribution characteristics, and a low-dimensional representation after dimensionality reduction is generated by constructing an information topology map and encoding it using a unique probability activation function, while introducing a sparse constraint mechanism based on information gain rate for training; Based on the low-dimensional representation, an initial population is generated using chaotic mapping, and individuals in the population correspond to different combinations of operation strategies; Construct a multi-objective fitness function that comprehensively considers the total revenue of the energy station, user satisfaction and energy utilization rate, and evaluate the fitness of the initial population; Use chaos search to generate disturbance vectors in the neighborhood of the initial population individuals, and update the population by comparing the fitness of the new individuals with the original individuals to achieve global exploration; Based on the population after chaotic search, a multi-objective game mechanism is introduced. Individuals generate new individuals in the neighborhood, play games based on different operating target benefits, adjust their own strategy parameters, and promote population evolution. After multiple iterations of chaotic search and multi-objective game, when the termination condition is met, the strategy corresponding to the individual with the highest fitness in the population is selected to generate the optimal operation strategy including charging pile pricing, energy storage battery charging and discharging, and user discount push.

3. A multi-strategy operation device for a photovoltaic storage and charging station based on deep learning, characterized in that: It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the multi-strategy operation method of the photovoltaic storage and charging station based on deep learning as described in claim 1.

Images