Distribution network energy storage strategy recommendation method based on artificial intelligence

Abstract

The invention provides a distribution network energy storage strategy recommendation method based on artificial intelligence, and relates to the technical field of energy storage optimization, and the method comprises the steps of data collection, model construction, model training, strategy optimization and the like. According to the method, the energy storage strategy optimization model is constructed, the weather data under the extreme weather condition is corrected, the wind, light and load prediction is carried out based on the corrected weather data, and the corresponding energy storage strategy is formulated for the energy storage system through the prediction result. The effect of quickly adjusting the energy storage strategy under multiple fluctuation characteristics of wind and light resource output and power grid load is achieved, and the stability of the power grid is improved.

Description

Technical Field

[0001] This application relates to the field of energy storage optimization technology, and in particular to an artificial intelligence-based method for recommending energy storage strategies for distribution networks. Background Art

[0002] In power distribution networks, energy storage systems play a crucial role. These systems must handle highly complex scenarios, including load regulation at battery charging and swapping stations, fluctuations in wind and solar power resources, and errors in load forecasting. Therefore, optimizing power grid energy storage across multiple application scenarios is particularly important. However, existing optimization methods often focus on specific scenarios, lacking comprehensiveness and flexibility.

[0003] Chinese invention patent application CN118281924A provides a method, system, and readable storage medium for optimizing a distribution network energy storage system across multiple application scenarios. This patent optimizes the time scale of battery charging and swapping stations based on a multi-scenario composite application management framework for energy storage systems. Specific optimization methods include day-ahead blue bar optimization, intraday rolling optimization, and intraday real-time optimization. It also optimizes the output sequence of the battery charging and swapping stations using a joint optimization scheduling model, which includes at least a day-ahead wind and solar power output model, an intraday wind and solar power output model, and a real-time wind and solar power output model. Finally, it analyzes the optimization results based on a scenario stochastic differential method, specifically setting three sets of comparative references to analyze and compare the optimization results.

[0004] While the aforementioned technical solutions involve multi-scenario applications of energy storage, they fail to fully consider the impact of extreme weather conditions on the energy storage system. In the event of extreme weather events such as typhoons, blizzards, and strong winds, both wind and solar power output and grid load will fluctuate dramatically. Given these multiple fluctuations in wind and solar power output and grid load, if the energy storage system cannot adjust quickly and accurately, grid stability will be threatened, impacting grid stability. Summary of the Invention

[0005] To enhance the dynamic response capability of energy storage systems under extreme weather events and improve grid stability, this application provides an artificial intelligence-based method for recommending distribution network energy storage strategies.

[0006] Firstly, this application provides an artificial intelligence-based method for recommending distribution network energy storage strategies, employing the following technical solution: An AI-based method for recommending distribution network energy storage strategies includes the following steps: Data acquisition: Collect real-time weather data, real-time energy storage data, and historical data; historical data includes historical weather data, historical energy storage data, and historical energy storage strategies; Model Construction: Construct an energy storage strategy optimization model, which includes a weather sub-model, a correction sub-model, a wind, solar and load sub-model, and an energy storage sub-model. Model training includes first recognition, first processing, second processing, and third processing; First identification: Input historical weather data into the weather sub-model to obtain predicted weather data, and determine whether the predicted weather data is extreme weather; If so, proceed with the first processing step; If not, the predicted weather data will be recorded as the first data, and the second processing step will be performed; First processing: Input the predicted weather data into the correction sub-model to obtain the correction factor, correct the predicted weather data based on the correction factor to obtain the corrected predicted weather data, update the corrected predicted weather data to the first data, and then execute the second processing steps. Second processing: Input the first data into the wind and solar load sub-model and output the second data; The third process involves training the energy storage sub-model using the first data, the second data, historical energy storage data, and historical energy storage strategies to obtain the trained energy storage strategy optimization model. Strategy optimization: Real-time weather data and real-time energy storage data are used as inputs to the trained energy storage strategy optimization model to obtain the real-time energy storage strategy.

[0007] By adopting the above technical solutions, based on the constructed energy storage strategy optimization model, weather data is predicted and extreme weather is identified. For identified extreme weather, the predicted weather data is corrected through a correction sub-model, improving the accuracy and reliability of weather prediction results under extreme weather conditions. This helps the energy storage system formulate corresponding energy storage strategies based on the corrected predicted weather data, enhancing the targeting and effectiveness of the energy storage system. Furthermore, based on the corrected predicted weather data, wind, solar, and load data are predicted. Combined with the prediction of wind, solar resources, and grid load under extreme weather conditions, the energy storage sub-model is trained. This helps the energy storage strategy optimization model learn the fluctuation patterns of wind, solar resources, and grid load under extreme conditions, improving the model's strategy optimization capability under extreme conditions. Simultaneously, comprehensively considering the impact of extreme weather conditions on the fluctuations of wind, solar resources, and grid load, the energy storage system can adjust its strategies in a timely manner based on the energy storage strategy optimization model's output under different extreme weather scenarios. This improves the timeliness, accuracy, and coping ability of the energy storage system in executing energy storage strategies under extreme weather conditions, thereby enhancing grid stability.

[0008] Optionally, after performing the model building step and before performing the model training step, the following steps are also included: Settings window: Defines a preset time window; First collection: collect meteorological data from each weather station within the preset time window; First setting: each weather station is regarded as the i-th level node of the tree structure, and the node features of the i-th level node include the meteorological data of each weather station; Node aggregation: Use the clustering algorithm to perform preliminary aggregation on the node features of the i-th layer nodes to obtain preliminary aggregation results, and determine the i+1-th layer nodes based on the preliminary aggregation results; Node determination: Determine if the number of nodes in the (i+1)th layer is 1. If so, output the tree structure, and then execute the first calculation step; If not, then the (i+1)th layer node is taken as the new i-th layer node, and the node aggregation step is performed; First calculation: Calculate the similarity of meteorological data between the i-th layer node and the (i+1)-th layer node, and use the calculated similarity as the connection strength between the i-th layer node and the (i+1)-th layer node; First construction: Obtain the neurons corresponding to the nodes at each layer of the tree structure in the weather sub-model, update the connection weights between the neurons based on the connection strength between the nodes, and obtain the updated weather sub-model; In the first identification step, historical weather data is input into the updated weather sub-model.

[0009] By adopting the above technical solutions and organizing meteorological data into a hierarchical tree structure, it is helpful to reflect the spatial relationships between different meteorological stations and regional climate differences. This allows meteorological data to be classified and managed according to geographical regions and climate models, providing a more efficient data access and processing method for weather sub-models. Simultaneously, it helps weather sub-models find a balance between local meteorological characteristics and overall climate trends, improving the accuracy of weather forecasts and enhancing their adaptability to meteorological changes in different regions and time periods. This is particularly suitable for high-frequency short-term weather forecasts and improves their adaptability to complex weather patterns in the face of variable extreme weather. Furthermore, by calculating the similarity of meteorological data between nodes as the node connection strength, weather sub-models can focus on data with similar meteorological conditions, avoiding interference from dissimilar data in forecasts. This improves the sensitivity and accuracy of weather sub-models to weather changes, enabling energy storage systems to adjust their energy storage strategies in a timely manner based on accurate forecasts of extreme weather. This enhances the timeliness, accuracy, and response capability of energy storage systems in executing energy storage strategies under extreme weather conditions, thereby strengthening the stability of the power grid.

[0010] Optionally, after executing the node judgment step and before executing the first calculation step, the method further includes: First determination: Calculate the change between the average meteorological data of each node in the current preset time window and the average meteorological data in the previous preset time window, and determine whether the change exceeds a preset change threshold. If so, proceed with the second judgment step; If not, no action will be taken; Second judgment: Calculate the similarity of meteorological data between the i-th layer node and the (i+1)-th layer node within the current preset time window, denoted as the first similarity, and determine whether the first similarity is lower than the preset similarity threshold: If so, then perform structural splitting on the i-th layer node, obtain the splitting result, update the splitting result to the i-th layer node, and execute the node aggregation step; If not, no action will be taken.

[0011] By employing the aforementioned technical solution, the change in meteorological data for each node within the current and previous time windows is calculated. This change is compared to a preset threshold, and the similarity between meteorological data points between nodes is compared to a preset similarity threshold. This process detects whether significant changes have occurred in the meteorological conditions of a node. This dynamic detection mechanism enhances the weather sub-model's ability to respond promptly to meteorological changes, especially in cases of sudden weather changes or large fluctuations. It helps the weather sub-model react quickly to changes, adjust its prediction structure, and improve its adaptability and prediction accuracy in the face of sudden weather events. Furthermore, based on this dynamic detection mechanism, when significant changes are detected at a node, the weather sub-model structure is promptly split and reconstructed. This helps the weather sub-model flexibly respond to regional or localized sudden meteorological changes. Especially when weather patterns in certain regions undergo drastic changes, splitting nodes maintains the representativeness of the node data, which helps improve the prediction accuracy of the weather sub-model under extreme weather conditions. Meanwhile, based on the aforementioned dynamic detection and adaptation mechanism and node splitting and aggregation mechanism, the weather sub-model's sensitivity and adaptability to weather changes are enhanced. In particular, when dealing with extreme weather and sudden meteorological events, the real-time performance and accuracy of the weather sub-model's predictions are improved. Moreover, based on the adaptive adjustment structure, the prediction accuracy of the weather sub-model is optimized, and the accuracy of local weather predictions is improved. This enables the model to not only make accurate predictions of large-scale weather changes, but also to respond promptly to local meteorological changes. This helps the energy storage system to adjust its energy storage strategy in a timely manner based on accurate predictions of extreme weather, improving the timeliness, accuracy, and response capability of the energy storage system in executing its energy storage strategy under extreme weather conditions, thereby enhancing the stability of the power grid.

[0012] Optionally, after performing the first construction step and before performing the model training step, the following steps are also included: Feature addition: Obtain the depth of all nodes in the tree structure; node features also include node depth. Graph structure construction: Treat the nodes in the tree structure as graph nodes, and the connections between pairs of nodes in the tree structure as edges between pairs of graph nodes; Set weights: use the connection strength between each pair of nodes as the meteorological feature weight of the edge, calculate the spatial distance between the positions of each pair of nodes, and use the reciprocal of the spatial distance as the spatial distance weight of the edge. Model update: Based on the edges between graph nodes and graph node pairs, construct a graph structure, update the connection weights between neurons in the weather sub-model based on the graph structure, and obtain the updated weather sub-model. In the first identification step, historical weather data is input into the updated weather sub-model.

[0013] By adopting the above technical solution and using the depth of each node as an additional feature, the model's ability to understand nodes at different levels is enhanced. This helps the model capture more geographical and meteorological information within the multi-level tree structure, enabling it to predict weather not only based on information at a single level but also on the hierarchical relationships and interactions between different layers of the tree structure. This improves the model's ability to understand and predict complex weather patterns. Updating the weather sub-model based on the constructed graph structure helps the model more accurately capture the geographical and meteorological data relationships between different weather stations, improving the accuracy of both local and global weather predictions. It also enhances the model's adaptability to sudden weather events and the accuracy of weather predictions. This helps energy storage systems adjust their energy storage strategies promptly based on accurate predictions of extreme weather, improving the timeliness, accuracy, and response capabilities of energy storage systems under extreme weather conditions, thereby enhancing the stability of the power grid.

[0014] Optionally, after performing the first identification step and before performing the first processing step, the method further includes: First extraction: The time when the predicted weather data is obtained is denoted as the first time. Second extraction: obtain the historical weather data at the first time, recorded as the first labeled data; Calculate the change value: calculate the difference between the predicted weather data and the first labeled data to obtain the deviation data set; Matrix construction: Calculate the correlation between each sample pair in the biased dataset, construct a matrix based on the calculated correlation, and obtain the shared matrix; Feature fusion: The shared matrix is ​​multiplied with the predicted weather data, and the result is recorded as the shared feature; Model optimization: Define a loss function, use a task-specific branch to process shared features, iteratively optimize the modified sub-model by minimizing the loss function, and use the optimized modified sub-model as the new modified sub-model.

[0015] By employing the aforementioned technical solutions, and based on a multi-task learning framework, the correlation between each sample pair in the deviation dataset is calculated to construct a shared matrix. This helps the model learn the error correlations between various features in the prediction data under different weather categories, based on the interrelationships between multiple features in the deviation dataset. Furthermore, multiplying the correlations between various features in the deviation dataset with the predicted weather data to construct shared features helps the model learn error patterns in the deviation dataset and feature patterns in the prediction data, enhancing the relationship patterns between various features in the error data of the prediction data under different weather categories. Simultaneously, model optimization based on the constructed shared features improves the flexibility and adaptability of the model in adjusting the predicted weather data according to different weather scenarios, enhancing the model's interpretability. Moreover, the correction factor based on the output improves the accuracy of the predicted weather data, enabling the energy storage system to adjust its energy storage strategy in a timely manner based on accurate predictions of extreme weather. This improves the timeliness, accuracy, and coping ability of the energy storage system in executing its energy storage strategy under extreme weather conditions, thereby enhancing the stability of the power grid.

[0016] Optionally, after performing the matrix construction step and before performing the feature fusion step, the following steps are also included: Feature filtering: The sample pairs whose relevance is greater than the preset feature relevance threshold are denoted as the first sample; Feature mapping: The kernel value of the first sample pair is calculated using a kernel function, and a kernel matrix is ​​constructed based on the calculation result; Feature unification: Based on the dimension of the shared matrix, adjust the dimension of the kernel matrix and use the adjusted kernel matrix as the new kernel matrix; Matrix update: Perform a product operation between the new kernel matrix and the shared matrix, and use the result as the new shared matrix.

[0017] By employing the aforementioned technical solutions, samples are filtered based on a preset feature correlation threshold, retaining those with strong correlations, reducing data dimensionality and complexity, and improving computational efficiency. Furthermore, mapping the first sample using a kernel function helps the model capture the nonlinear relationships between strongly correlated samples in the biased dataset. Simultaneously, updating the shared matrix based on the kernel matrix and the shared matrix helps combine the nonlinear feature information in the kernel matrix with the linear feature information in the shared matrix, effectively integrating the relationships between different features in the updated shared features and enhancing their expressive power. Simultaneously, model optimization based on the fused shared features helps improve the model's learning ability on multidimensional data and increases the accuracy of the model's prediction correction factor. Moreover, the improved accuracy of weather data prediction based on the output correction factor helps the energy storage system adjust its energy storage strategy promptly based on accurate forecasts of extreme weather, improving the timeliness, accuracy, and coping ability of the energy storage system in executing energy storage strategies under extreme weather conditions, thereby enhancing the stability of the power grid.

[0018] Optionally, after performing the model building step and before performing the model training step, the following steps are also included: First definition: Based on historical data, an optimization objective is defined, which includes minimizing redundancy, minimizing noise, and maximizing information content; Construct function: construct objective function based on optimization goal; First setting: Define the particles as a set of data in the historical data, use the particle swarm optimization algorithm to solve the objective function, calculate the velocity of the particles, and update the particle positions according to the particle velocities; Second setting: Calculate the difference between the fitness of the particle after the update and the fitness of the particle before the update, record the calculated difference as the first difference, and adjust the particle position again based on the probability acceptance mechanism of the simulated annealing algorithm. First optimization: Obtain the historical data corresponding to the adjusted particle position, record it as the first historical data, and update the first historical data to the historical data.

[0019] By adopting the above technical solutions, and comprehensively considering the optimization objectives of minimizing redundancy, minimizing noise, and maximizing information content, a multi-stage optimization technique is used. Particle swarm optimization and simulated annealing algorithms are introduced to calculate the fitness difference of particles and adjust their positions, thereby achieving the goal of dynamically adjusting the distribution of historical data and making the optimization of historical data adaptive. Simultaneously, the optimized data distribution better reflects the inherent structure of historical data, contributing to improved model accuracy. Furthermore, optimizing historical data through methods such as particle swarm optimization and simulated annealing increases the diversity and coverage of historical data, helping the model adapt to changes in different data types and improving the model's prediction accuracy and generalization ability.

[0020] Optionally, after performing the second setting step and before performing the first optimization step, the method further includes: Third setting: Obtain the random number generated when the particle position is updated again, and determine whether the random number is less than a preset probability threshold. If so, the krill swarm algorithm is used to adjust the particle positions a third time, performing the particle optimization steps: If not, proceed with the first optimization step; Particle optimization: Calculate the fitness of the particles before and after position adjustment, and determine whether the fitness after adjustment is greater than the fitness before adjustment: If so, the particle position after the third adjustment will be used as the new particle position, and the adjusted particle position will be updated using the new particle position. If not, the position of the particle is restored to the re-adjusted particle position.

[0021] By employing this technical solution, the random number generated when particle positions are updated is compared with a preset probability threshold to determine whether to use the krill swarm algorithm for optimization adjustments. This helps the optimization algorithm continue to explore new solution spaces under uncertainty, thereby breaking the limitation of local optimal solutions and increasing the diversity of global search. Furthermore, using the krill swarm algorithm for particle position adjustment further enhances the particle swarm's exploration capabilities and improves the model's adaptability to different data and optimization objectives. Furthermore, by dynamically selecting optimization strategies, the limitations of fixed optimization paths can be reduced, improving the model's overall optimization efficiency.

[0022] Optionally, the energy storage strategy optimization model further includes a charge / discharge load prediction sub-model; the method further includes: Second collection: collect real-time electric vehicle data; real-time electric vehicle data includes real-time charging pile charging data, real-time electricity prices and real-time holiday data; In the strategy optimization step, real-time electric vehicle data, real-time weather data, and real-time energy storage data are input into the trained energy storage strategy optimization model.

[0023] By adopting the above technical solution, a charging and discharging load prediction sub-model is constructed, which takes into account the changes in electric vehicle charging behavior under extreme weather conditions and the impact of electricity price changes on the grid load, and optimizes the grid load. This helps the model to comprehensively consider the impact of electric vehicle charging load under extreme weather conditions, energy price fluctuations, and holiday-specific demand fluctuations on the grid load, thereby improving the accuracy and rationality of the energy storage decisions output by the model. At the same time, it helps the model to flexibly adjust energy storage strategies in different scenarios, improve the overall operating efficiency of the energy storage system, and further enhance the stability of the grid.

[0024] Optionally, the method further includes: Third data collection: Collect input data, which includes historical charging data of charging piles marked with holiday or non-holiday labels, historical electricity prices, and first data. Charge and discharge prediction: Input the input data into the charge and discharge load prediction sub-model to obtain the third data; First concatenation: taking the sum of the third data and the second data as new second data.

[0025] By adopting the above technical solution and combining information such as holiday tags, historical charging pile data, and historical electricity prices, the charging and discharging load prediction sub-model accurately captures the dynamic changes of charging and discharging load, further improving the accuracy of grid load prediction. This helps the energy storage system to more accurately predict grid load demand under extreme weather conditions, and optimize energy storage strategies in a timely manner based on the prediction results, quickly responding to grid load fluctuations, improving the stability and economy of the energy storage system, and enhancing grid stability.

[0026] In summary, this application includes at least one of the following beneficial technical effects: 1. Based on the constructed energy storage strategy optimization model, weather data is predicted, and extreme weather events are identified. For identified extreme weather events, the predicted weather data is corrected through a correction sub-model, improving the accuracy and reliability of weather forecasts under extreme weather conditions. This helps the energy storage system formulate corresponding energy storage strategies based on the corrected predicted weather data, enhancing the system's targeting and effectiveness. Furthermore, based on the corrected predicted weather data, wind, solar, and load data are predicted. Combined with predictions of wind, solar resources, and grid load under extreme weather conditions, the energy storage sub-model is trained. This helps the energy storage strategy optimization model learn the fluctuation patterns of wind, solar resources, and grid load under extreme conditions, improving the model's strategy optimization capability under extreme conditions. Simultaneously, comprehensively considering the impact of extreme weather conditions on the fluctuations of wind, solar resources, and grid load helps the energy storage system adjust its strategies in a timely manner based on the energy storage strategy optimization model under different extreme weather scenarios. This improves the timeliness, accuracy, and coping ability of the energy storage system in executing energy storage strategies under extreme weather conditions, thereby enhancing grid stability.

[0027] 2. Organizing meteorological data into a hierarchical tree structure helps reflect the spatial relationships between different meteorological stations and regional climate differences. This allows for the classification and management of meteorological data based on geographical regions and climate patterns, providing a more efficient data access and processing method for weather sub-models. Simultaneously, it helps weather sub-models find a balance between local meteorological characteristics and overall climate trends, improving the accuracy of weather forecasts and enhancing their adaptability to meteorological changes in different regions and time periods. This is particularly suitable for high-frequency short-term weather forecasts and improves their adaptability to complex weather patterns in the face of variable extreme weather. Furthermore, calculating the similarity of meteorological data between nodes as the node connection strength helps weather sub-models focus on data with similar meteorological conditions, avoiding interference from dissimilar data. This improves the sensitivity and accuracy of weather changes, enabling energy storage systems to adjust their energy storage strategies promptly based on accurate extreme weather forecasts. This enhances the timeliness, accuracy, and resilience of energy storage systems in executing energy storage strategies under extreme weather conditions, thereby strengthening grid stability.

[0028] 3. Based on a multi-task learning framework, the correlation between each sample pair in the deviation dataset is calculated to construct a shared matrix. This helps the model learn the error correlations between various features in the prediction data under different weather categories, based on the interrelationships between multiple features in the deviation dataset. Furthermore, multiplying the correlations between features in the deviation dataset with the predicted weather data to construct shared features helps the model learn error patterns in the deviation dataset and feature patterns in the prediction data, enhancing the relationship patterns between features in the error data of the prediction data under different weather categories. Simultaneously, model optimization based on the constructed shared features improves the flexibility and adaptability of the model in adjusting predicted weather data based on different weather scenarios, enhancing the model's interpretability. Moreover, the correction factor based on the output improves the accuracy of the predicted weather data, enabling energy storage systems to adjust their energy storage strategies promptly based on accurate predictions of extreme weather. This improves the timeliness, accuracy, and coping ability of energy storage systems to execute energy storage strategies under extreme weather conditions, thereby enhancing the stability of the power grid.

[0029] 4. Taking into account the optimization goals of minimizing redundancy, minimizing noise, and maximizing information, we introduced particle swarm optimization and simulated annealing algorithms based on multi-stage optimization techniques to calculate the fitness difference of particles and adjust their positions. This dynamically adjusts the distribution of historical data, making the optimization of historical data adaptive. Furthermore, the optimized data distribution better reflects the inherent structure of historical data, helping to improve model accuracy. Furthermore, optimizing historical data through methods such as particle swarm optimization and simulated annealing increases the diversity and coverage of historical data, helping the model adapt to changes in different data types and improving its predictive accuracy and generalization capabilities. Attached Figure Description

[0030] Figure 1 This is a flowchart of Embodiment 1 of this application; Figure 2 This is a flowchart of the S21 weather sub-model optimization in Embodiment 2 of this application; Figure 3 This is a flowchart of S311 correction sub-model optimization in Example 3 of the present application; Figure 4 This is a flowchart of S22 historical data optimization in Embodiment 4 of this application. DETAILED DESCRIPTION

[0031] The following combination Figures 1 to 4 This application is described in further detail.

[0032] Example 1: This example discloses an artificial intelligence-based method for recommending distribution network energy storage strategies, such as... Figure 1 As shown, the method includes: collecting real-time weather data, real-time energy storage data and historical data; historical data includes historical weather data, historical energy storage data and historical energy storage strategies, constructing an energy storage strategy optimization model, the energy storage strategy optimization model includes a weather sub-model, a correction sub-model, a wind, light and load sub-model, and an energy storage sub-model; inputting historical weather data into the weather sub-model to obtain predicted weather data, judging whether the predicted weather data is extreme weather, if so, inputting the predicted weather data into the correction sub-model to obtain a correction factor, correcting the predicted weather data based on the correction factor to obtain corrected predicted weather data, and updating the corrected predicted weather data to the first data; if not, recording the predicted weather data as the first data; inputting the first data into the wind, light and load sub-model, outputting the second data, using the first data, the second data, the historical energy storage data and the historical energy storage strategy to train the energy storage sub-model to obtain a trained energy storage strategy optimization model, using the real-time weather data and the real-time energy storage data as inputs to the trained energy storage strategy optimization model to obtain a real-time energy storage strategy. This embodiment includes the following steps: S1 Data Acquisition: Collects real-time weather data, real-time energy storage data, and historical data. Historical data includes historical weather data, historical wind and solar load data, historical energy storage data, and historical energy storage strategies.

[0033] Weather data includes temperature, humidity, air pressure, precipitation, wind speed, wind direction, light intensity, and cloud cover. Among these, temperature, humidity, air pressure, precipitation, wind speed, wind direction, light intensity, and cloud cover are the weather parameters of the weather data.

[0034] Real-time weather data includes real-time temperature, humidity, air pressure, precipitation, wind speed, wind direction, light intensity, and cloud cover. Real-time weather data can be obtained through official meteorological websites, local weather stations, online weather platforms, satellite remote sensing technology, and satellite cloud imagery. The geographical location of the real-time weather data is then used to label it with corresponding geographic location tags.

[0035] Real-time energy storage data includes the charging and discharging status of energy storage devices, energy storage capacity data, and power output data. This data can be collected in real-time through built-in sensors or monitoring systems within the energy storage system.

[0036] Historical weather data includes historical temperature, humidity, air pressure, precipitation, wind speed, wind direction, sunshine intensity, and cloud cover. Historical weather data can be obtained through official meteorological websites, local weather stations, online weather platforms, satellite remote sensing technology, and satellite cloud imagery. The geographical location of the historical weather data is then used to label it.

[0037] Historical landscape and load data include historical landscape resource data and historical load data.

[0038] Historical wind and solar power resource data includes the historical power generation of wind turbines and photovoltaic power generation systems. This data can be obtained through monitoring systems for wind turbines and photovoltaic power generation systems.

[0039] Historical load data can be obtained by measuring and collecting data from smart meters and calculating historical power grid load data, or by obtaining data from a load-side management platform.

[0040] Historical energy storage data includes the historical charge / discharge status of energy storage devices, historical energy storage capacity data, and historical power output data. This data can be obtained through built-in sensors or monitoring systems within the energy storage devices.

[0041] Historical energy storage strategy: The historical energy storage strategy executed by the energy storage system based on historical weather data, historical energy storage data and historical wind and solar resource data collected at the same time can be obtained through the local database of the energy storage system or the cloud platform.

[0042] In this embodiment, a set of historical data includes historical weather data, historical wind and solar load data, historical energy storage data, and historical energy storage strategies at the same point in time.

[0043] In this embodiment, the collected data needs to be preprocessed. The data preprocessing steps include data cleaning and data normalization.

[0044] Data cleaning: Performing operations such as noise reduction, handling missing values, and removing outliers on the collected data.

[0045] Data normalization: Normalize the collected data.

[0046] S2 builds the model: builds the energy storage strategy optimization model, which includes the weather sub-model, the correction sub-model, the wind, solar and load sub-model, and the energy storage sub-model.

[0047] The weather sub-model can use convolutional neural networks, recurrent neural networks, long short-term memory networks, or gated recurrent units as the base model. In this embodiment, the weather sub-model uses a long short-term memory network model as the base model.

[0048] The method for constructing the modified sub-model includes: selecting a Transformer model as the base model for the modified sub-model; therefore, using a portion of the Transformer encoder layers as shared layers in the modified sub-model; and using the remaining Transformer encoder layers and the fully connected layers in the Transformer model that perform specific tasks as task-specific layers in the modified sub-model. Based on weather parameters from the weather data, task-specific branches are created, including temperature, humidity, air pressure, precipitation, wind speed, wind direction, light intensity, and cloud cover tasks.

[0049] The wind-solar-load sub-model includes two sets of long short-term memory (LSTM) network models. The first set of LTM network models is used to predict wind and solar power generation data based on weather data, and is referred to as the first model. The second set of LTM network models is used to predict grid load based on weather data, and is referred to as the second model.

[0050] The energy storage sub-model can use convolutional neural networks, recurrent neural networks, or long short-term memory networks as its base model. In this embodiment, the energy storage sub-model only needs to be trained with historical data to have the function of outputting energy storage strategies.

[0051] S3 model training includes S31 first recognition, S32 first processing, S33 second processing, and S34 third processing.

[0052] S31 First Identification: Input historical weather data into the weather sub-model to obtain predicted weather data, and determine whether the predicted weather data is extreme weather.

[0053] If so, then execute S33, the first process.

[0054] If not, the predicted weather data will be recorded as the first data, and the second processing step S33 will be executed.

[0055] Methods for determining whether predicted weather data constitutes extreme weather include methods based on numerical thresholds and methods based on constructing extreme weather identification models.

[0056] The numerical threshold determination method includes: determining whether there are weather parameters in the predicted weather data that exceed the corresponding preset weather parameter threshold; if there are weather parameters that exceed the corresponding preset weather parameter threshold, determining whether the number of weather parameters that exceed the corresponding preset weather parameter threshold exceeds the preset quantity threshold; if so, determining that the predicted weather data is extreme weather; otherwise, determining that the predicted weather data is not extreme weather.

[0057] In the numerical threshold determination method, the preset weather parameter threshold and preset quantity threshold corresponding to each weather parameter can be set according to actual needs.

[0058] The method for constructing an extreme weather identification model includes: using a VAE-GAN model and a classification model to build the extreme weather identification model. In the VAE-GAN model, the VAE part is selected as a conditional variational autoencoder, which includes an encoder and a decoder. The GAN part in the VAE-GAN model is a generative adversarial network model, which includes a generator and a discriminator. The classification model can use a convolutional neural network, a recurrent neural network, or a Transformer model, etc. In this embodiment, the Transformer model is selected as the base model for the classification model.

[0059] The model training process of the extreme weather recognition model is as follows: mark the corresponding weather category label for the historical weather data, record it as the first recognition data, input the first recognition data into the encoder to obtain the latent variable, use the latent variable as the input of the generator to generate pseudo data; use the pseudo data as the input of the discriminator to obtain the discrimination result of the pseudo data. If the discrimination result is unqualified, optimize the parameters of the generator and repeat the above steps; if the discrimination result is qualified, use the pseudo data and the first recognition data as the input of the classification model to perform model inference, define the mean square error function as the loss function, calculate the loss value, use the model parameter gradient calculated by the backpropagation algorithm, and use the gradient descent method to update the model parameters until the preset number of iterative training is completed or the calculated loss value of the classification model no longer decreases, and the trained extreme weather recognition model is obtained.

[0060] The predicted weather data is input into the trained extreme weather recognition model, which outputs weather category labels for the predicted weather data. Based on the weather category labels, it is determined whether the predicted weather data is extreme weather.

[0061] In this embodiment, an extreme weather identification model is used to determine whether the predicted weather data is extreme weather.

[0062] S32: First Processing: The forecasted weather data is input into the correction sub-model to obtain a correction factor for each weather parameter in the forecasted weather data. The numerical value of each weather parameter in the forecasted weather data is added to the corresponding correction factor, and the result is used as the corrected forecasted weather data. The corrected forecasted weather data is then updated as the first data, and then the second processing S33 is executed.

[0063] S33 Second Processing: Input the first data into the first model to obtain predicted wind and solar power generation data, input the first data into the second model to obtain predicted grid load data, and use the obtained predicted wind and solar power generation data and predicted grid load data as the second data.

[0064] S34 third processing: using the first data, the second data, the historical energy storage data and the historical energy storage strategy to train the energy storage sub-model to obtain a trained energy storage strategy optimization model.

[0065] The first data, the second data, and historical energy storage data are input into the energy storage sub-model to obtain the predicted energy storage strategy.

[0066] The loss function of the energy storage strategy optimization model is defined as: ; ; ; ; ; .

[0067] in, This represents the loss function of the weather sub-model. represents the loss function of the first model, This represents the loss function of the second model. This represents the loss function for modifying the sub-model. This represents the loss function of the energy storage sub-model. The number of historical weather data points in the historical data. The first in the weather sub-model Real labels in historical weather data The first in the weather sub-model One weather forecast data, The quantity of historical landscape resource data in the historical data. For the first model in the middle A true label for historical landscape resource data. The first in the weather sub-model One predicted wind and solar power generation data, The number of historical load data in the historical data. For the first model A true label for historical load data, The first in the weather sub-model One predicted power grid load data, To correct the number of task-specific branches in the sub-model, For the The weight of a specific branch of a task. For the The loss function for a specific branch of a task. Input the number of historical weather data points into the correction submodel from the historical data. For the The first task-specific branch The true label of the error value between the predicted weather data and the corresponding historical weather data of each input correction sub-model. For the The first task-specific branch The correction factor for the predicted weather data of the input correction sub-model. The number of samples input into the energy storage sub-model from historical data. In the energy storage sub-model The true labels of the samples input to the energy storage sub-model. The first in the weather sub-model A predictive energy storage strategy.

[0068] Historical weather data, historical wind and solar load data, historical energy storage data, and historical energy storage strategies are input into the energy storage strategy optimization model for training. Based on the loss function of the above energy storage strategy optimization model, the energy storage strategy optimization model is iteratively trained by minimizing the loss function of the energy storage strategy optimization model to obtain the trained energy storage strategy optimization model.

[0069] S4 Strategy Optimization: Real-time weather data and real-time energy storage data are used as inputs to the trained energy storage strategy optimization model to obtain the real-time energy storage strategy.

[0070] Real-time weather data is input into the weather sub-model to obtain predicted real-time weather data. It is then determined whether the predicted real-time weather data represents extreme weather. If so, the predicted real-time weather data is input into the correction sub-model to obtain a correction factor for each weather parameter in the predicted real-time weather data. The value of each weather parameter in the predicted real-time weather data is added to the corresponding correction factor for each weather parameter, and the result is used as the corrected predicted real-time weather data, denoted as the first real-time data. If not, the obtained predicted real-time weather data is directly used as the first real-time data. The obtained first real-time data is input into the wind, solar, and load sub-model to obtain the second real-time data. The obtained first real-time data, second real-time data, and real-time energy storage data are input into the energy storage sub-model to obtain the real-time energy storage strategy.

[0071] In this embodiment, based on the constructed energy storage strategy optimization model, weather data is predicted, and extreme weather events are identified. For identified extreme weather events, the predicted weather data is corrected through a correction sub-model, improving the accuracy and reliability of weather prediction results under extreme weather conditions. This helps the energy storage system formulate corresponding energy storage strategies based on the corrected predicted weather data, enhancing the targeting and effectiveness of the energy storage system. Furthermore, based on the corrected predicted weather data, wind, solar, and load data are predicted. Combined with the predictions of wind and solar resources and grid load under extreme weather conditions, the energy storage sub-model is trained. This helps the energy storage strategy optimization model learn the fluctuation patterns of wind and solar resources and grid load under extreme conditions, improving the model's strategy optimization capability under extreme conditions.

[0072] Example 2: The difference from Example 1 is that: After executing S2 to build the model and before executing S3 to train the model, the S21 weather sub-model optimization is also included.

[0073] S21 Weather sub-model optimization includes: S211 Setting window, S212 First acquisition, S213 First setting, S214 Node aggregation, S215 Node judgment, S216 First judgment, S217 Second judgment, S218 First calculation, S219 First construction and S2110 Second construction, such as Figure 2 shown.

[0074] S211 Settings Window: Defines a preset time window.

[0075] The preset time window can be set to multiple collection time periods as needed.

[0076] S212 First Collection: Determine the location of the weather stations to be collected, and collect the weather data and location information of each weather station at a certain moment within a preset time window. The location information of the weather station can be obtained through a geographic information system.

[0077] Meteorological data from weather stations include: temperature, humidity, air pressure, precipitation, wind speed, wind direction, light intensity and cloud cover.

[0078] S213 First setting: each weather station is regarded as the i-th level node of the tree structure, and the node characteristics of the i-th level node include the weather data of each weather station. When i=1, it means that all weather stations are regarded as the bottom level nodes of the tree structure.

[0079] S214 Node Aggregation: Using the node features of the i-th layer nodes as node features, a clustering algorithm (such as K-means clustering, DBSCAN clustering, or hierarchical clustering) is used to initially aggregate the node features of the i-th layer nodes, obtaining a preliminary aggregation result. The nodes of the (i+1)-th layer are then determined based on this preliminary aggregation result. When determining the (i+1)-th layer nodes, the geometric center point of the i-th layer nodes under the current (i+1)-th layer node is calculated based on the position information of the i-th layer nodes in each cluster, and this calculated geometric center point is used as the current (i+1)-th layer node. After determining the (i+1)-th layer nodes, the Euclidean distance between each i-th layer node and its corresponding (i+1)-th layer node is calculated. The calculated Euclidean distance is then normalized, and the processed data is used as the weight of each i-th layer node under the current (i+1)-th layer node. A weighted sum is then performed based on the weights of the i-th layer nodes and their corresponding node features, and this weighted sum is used as the node feature of the current (i+1)-th layer node, where i≧1.

[0080] In this embodiment, except when i=1, all weather stations represented by the i-th layer nodes are actual weather stations, and other i-th layer nodes refer to virtual weather stations.

[0081] For example, if the preset time window is defined as 1 hour, data will be collected once every hour. At 14:00, weather stations A, B, C, D, and E respectively record the following weather data: A: temperature = 20°C, humidity = 70%, wind speed = 5.2m / s, coordinates = (100 o 30 o B: Temperature = 25℃, Humidity = 60%, Wind speed = 4.8m / s, Coordinates = (101) o 31 o C: Temperature = 30℃, Humidity = 55%, Wind speed = 6.0m / s, Coordinates = (102) o 32.5 o D: Temperature = 18℃, Humidity = 80%, Wind speed = 4.5m / s, Coordinates = (99 o 29.5 o E: Temperature = 22℃, Humidity = 65%, Wind speed = 5.5m / s, Coordinates = (100.5) o 31.5 o A tree structure is constructed based on the meteorological data from weather stations A, B, C, D, and E (for simplicity, only a portion of the collected weather parameters are used as examples of meteorological data here).

[0082] Considering these five weather stations as the first-level nodes of the tree, we cluster the meteorological data of these five weather stations using the K-means clustering algorithm. The clustering result is two clusters, indicating that we have two second-level nodes. The first-level nodes under second-level node 1 are nodes A, B, and E, and the first-level nodes under second-level node 2 are nodes C and D.

[0083] Based on the clustering results, calculate the geometric center of each second-layer node. The coordinates of the second-layer node 1 are (100.5 o 30.83 o The coordinates of node 1 in the second layer are (100.75). o 31 o ).

[0084] Calculate the Euclidean distance from each weather station to its corresponding second-layer node. The Euclidean distance from weather station A to second-layer node 1 is 0.96, from weather station B to second-layer node 1 is 0.52, from weather station E to second-layer node 1 is 0.67, from weather station C to second-layer node 2 is 2.28, and from weather station D to second-layer node 2 is 2.28. Normalize the weights of second-layer nodes 1 (weather stations A, B, E) and second-layer nodes 2 (weather stations C, D). The weights of weather station A and weather station B are 0.242, from weather station E to weather station C to weather station D to weather station D.

[0085] Meteorological data for the second-layer nodes are calculated based on weights and meteorological data from each meteorological station.

[0086] The temperature of node 1 in the second layer is T1 = (0.446 × 20) + (0.242 × 25) + (0.312 × 22) = 8.92 + 6.05 + 6.864 = 21.834 °C; The humidity of node 1 in the second layer is H1 = (0.446 × 70) + (0.242 × 60) + (0.312 × 65) = 31.22 + 14.52 + 20.28 = 66.02%; The wind speed at node 1 in the second layer is F1 = (0.446 × 5.2) + (0.242 × 4.8) + (0.312 × 5.5) = 2.315 + 1.162 + 1.716 = 5.193 m / s; The temperature of node 2 in the second layer is T2 = (0.5 × 30) + (0.5 × 18) = 15 + 9 = 24℃; The humidity of node 2 in the second layer is H2 = (0.5 × 55) + (0.5 × 80) = 27.5 + 40 = 67.5%; The wind speed at node 2 of the second layer is F2 = (0.5 × 6.0) + (0.5 × 4.5) = 3 + 2.25 = 5.25 m / s.

[0087] The second-layer node 1 has the following characteristics: temperature = 22℃, humidity = 66%, wind speed = 5.2m / s, and coordinates = (100.5). o 30.83 o ); Second layer node 2: Temperature = 24℃, Humidity = 67.5%, Wind speed = 5.3m / s, Coordinates = (100.75) o 31 o ).

[0088] S215 Node judgment: judge whether the number of nodes in the i+1th layer is 1.

[0089] If so, it indicates that the tree structure has been constructed, the tree structure is output, and then the first judgment in S216 is executed.

[0090] If not, then the (i+1)th layer node is taken as the new i-th layer node, and S214 node aggregation is performed based on the new i-th layer node.

[0091] S216 First Judgment: Calculate the average value of meteorological data for each node in the current preset time window and the average value of meteorological data in the previous preset time window. Based on the obtained average value of meteorological data for each node in the current preset time window and the average value of meteorological data in the previous preset time window, calculate the change between the average value of meteorological data for each node in the current preset time window and the average value of meteorological data in the previous preset time window, and determine whether the change exceeds a preset change threshold.

[0092] If so, then execute the second judgment in S217.

[0093] If not, no action will be taken.

[0094] In this embodiment, during the first judgment in S216, the node can be judged starting from the bottom node (i.e., the first layer node) according to the node layer order to determine whether the change between the average value of the meteorological data of each node in the current preset time window and the average value of the meteorological data in the previous preset time window exceeds the preset change threshold.

[0095] S217 Second Judgment: Use a similarity algorithm (such as Euclidean distance, cosine similarity and Pearson correlation coefficient) to calculate the similarity of meteorological data between the i-th layer node and the i+1-th layer node within the current preset time window, denoted as the first similarity, and determine whether the first similarity is lower than the preset similarity threshold.

[0096] If so, then perform structural splitting on the i-th layer node, take the split i-th layer node as the unaggregated i-th layer node, update the unaggregated i-th layer node to the i-th layer node, and perform node aggregation in S214.

[0097] If not, no action will be taken.

[0098] In this process, the similarity between meteorological data, i.e., node features, between nodes in the i-th layer and nodes in the (i+1)-th layer is calculated based on a similarity algorithm. Therefore, node features include temperature features, humidity features, air pressure features, precipitation features, wind speed features, wind direction features, light intensity features, and cloud cover features. The calculated similarities of temperature, humidity, air pressure, precipitation, wind speed, wind direction, light intensity, and cloud cover are weighted and summed, and the weighted summation result is used as the first similarity.

[0099] S218 First Calculation: A similarity algorithm is used to calculate the similarity of meteorological data between all nodes in layer i and layer i+1, and this calculated similarity is used as the connection strength between the nodes in layer i and layer i+1. In this embodiment, the Pearson correlation coefficient can be used to calculate the similarity of meteorological data between all nodes in layer i and layer i+1.

[0100] S219 First Construction: Based on the number of layers in the tree structure and the number of nodes in each layer, each node in the tree structure is mapped to a neuron in the neural network. After mapping, each node in each layer corresponds to a neuron in the neural network. The connection strength between each node is calculated as the connection weight between each neuron to obtain the updated weather sub-model.

[0101] S2110 Second construction: includes S21101 feature addition, S21102 graph structure construction, S21103 weight setting and S21104 model update.

[0102] S21101 Feature Addition: Obtain the depth of all nodes in the tree structure and add a node depth feature to the node features of each node in the tree structure. The node depth represents the level of the node in the tree structure.

[0103] S21102 Graph Structure Construction: Treat the nodes in the tree structure as graph nodes, and treat the connections between pairs of nodes in the tree structure as edges between pairs of graph nodes.

[0104] S21103 Weight Setting: Based on the similarity value calculated in the first calculation in S218, the connection strength between each node pair is used as the meteorological feature weight of the edge. The spatial distance between the locations of each node pair (e.g., the Euclidean distance between the geographic coordinates of each node pair) is calculated, and the reciprocal of the spatial distance is used as the spatial distance weight of the edge. The closer the spatial distance between node pairs, the larger the spatial distance weight of the edge, indicating a stronger relationship between the node pairs. The obtained meteorological feature weights and spatial distance weights of the edges are weighted and summed, and the weighted sum is used as the total weight of the edge.

[0105] Model update S21104: Based on the edges between graph nodes and graph node pairs, a graph structure is constructed. Based on the graph structure, the connection weights between neurons in the weather sub-model are updated through graph convolution to obtain the updated weather sub-model.

[0106] In this embodiment, in the first identification at S31, the historical weather data is input into the weather sub-model that is updated again in the model update at S2184.

[0107] In this embodiment, organizing meteorological data into a hierarchical tree structure helps reflect the spatial relationships between different meteorological stations and regional climate differences. Furthermore, the implementation of dynamic detection and adaptation mechanisms and node splitting and aggregation mechanisms enhances the sensitivity and adaptability of the weather sub-model to weather changes, particularly in responding to extreme weather and sudden meteorological events. This improves the real-time performance and accuracy of the weather sub-model's predictions. Moreover, the adaptive adjustment structure helps optimize the prediction accuracy of the weather sub-model, improving the accuracy of local weather predictions. This allows the model to accurately predict not only large-scale weather changes but also respond promptly to local meteorological changes. Simultaneously, it helps the model more accurately capture the geographical and meteorological data relationships between different meteorological stations, improving the accuracy of both local and global weather predictions, as well as enhancing the model's adaptability to sudden weather events and overall weather prediction accuracy.

[0108] Example 3: The difference from Example 1 is that: After performing the first identification in S31 and before performing the first processing in S32, the process also includes S311, which involves optimizing the modified sub-model.

[0109] The S311 sub-model optimization includes: S3111 first extraction, S3112 second extraction, S3113 calculating change values, S3114 constructing the matrix, S3115 feature selection, S3116 feature mapping, S3117 feature unification, S3118 matrix update, S3119 feature fusion, and S31110 model optimization. Figure 3 shown.

[0110] S3111 First Extraction: The time of obtaining the predicted weather data obtained in S31 First Recognition is denoted as the first time. The first time here is not the time when the weather sub-model outputs the predicted weather data. Historical weather data is timeline data. When the historical weather data at a certain moment is input into the weather sub-model, the time label of the historical weather data is also input. Therefore, the first time here actually refers to the time label of the predicted weather data output by the weather sub-model.

[0111] S3112 Second Extraction: Based on the first time, obtain the real historical weather data at the first time from the historical weather data, and record it as the first labeled data.

[0112] S3113 Calculate the change value: Calculate the difference between the predicted weather data and the first labeled data to obtain the deviation dataset, and label the deviation dataset with weather category labels, which are the weather categories of each historical weather data.

[0113] For example, if the predicted weather data = {predicted temperature, predicted humidity, predicted air pressure, predicted precipitation, predicted wind speed, predicted wind direction, predicted light intensity, predicted cloud cover}, and the first labeled data = {first temperature, first humidity, first air pressure, first precipitation, first wind speed, first wind direction, first light intensity, first cloud cover}, then the bias dataset = {predicted temperature - first temperature, predicted humidity - first humidity, predicted air pressure - first air pressure, predicted precipitation - first precipitation, predicted wind speed - first wind speed, predicted wind direction - first wind direction, predicted light intensity - first light intensity, predicted cloud cover - first cloud cover} = {temperature bias, humidity bias, air pressure bias, precipitation bias, wind speed bias, wind direction bias, light intensity bias, cloud cover bias}.

[0114] S3114 Matrix Construction: The correlation between each sample pair in the biased dataset is calculated using the Pearson correlation coefficient. Based on the calculated correlation, a matrix is ​​constructed to obtain the shared matrix S. In this embodiment, the samples refer to temperature bias, humidity bias, air pressure bias, precipitation bias, wind speed bias, wind direction bias, light intensity bias, and cloud cover bias.

[0115] Build a The shared matrix S, the rows and columns of the shared matrix S correspond to the correlation of each sample pair.

[0116] .

[0117] S3115 Feature Filtering: Set a preset feature correlation threshold, and designate samples corresponding to sample pairs with a correlation greater than the preset feature correlation threshold as the first sample. In this embodiment, as long as the correlation of a sample pair is greater than the preset feature correlation threshold, both samples in the current sample pair are retained. For example, if the correlation between temperature deviation and humidity deviation is greater than the preset feature correlation threshold, then temperature deviation and humidity deviation are retained, and temperature deviation and humidity deviation are designated as the first sample.

[0118] S3116 Feature Mapping: Calculate the Euclidean distance of the first sample pair. Based on the calculated Euclidean distance of the first sample pair, use the Gaussian kernel function to calculate the kernel value of the first sample pair. Based on the calculation results, combine all the kernel values ​​to construct the kernel matrix L.

[0119] The Gaussian kernel function formula is: .

[0120] in, Let j represent the j-th first sample. represents the kth first sample, represents the Euclidean distance between the jth first sample and the kth first sample, Represents the bandwidth parameter of the Gaussian kernel function.

[0121] Elements in kernel matrix L This represents the kernel value between the j-th first sample and the k-th first sample.

[0122] S3117 Feature Unification: Based on the dimension of the shared matrix S, adjust the dimension of the kernel matrix L so that the dimension of the adjusted kernel matrix L is consistent with the dimension of the shared matrix S, and use the adjusted kernel matrix L as the new kernel matrix .

[0123] Determine whether the dimensions of the shared matrix S and the kernel matrix L are consistent. If they are, no adjustment is needed. If they are inconsistent, further determine whether the dimension of the kernel matrix L is greater than that of the shared matrix S. If so, truncate the kernel matrix L to remove redundant rows and columns to make its dimension consistent with that of the shared matrix S. If not, adjust the dimension of the kernel matrix L using interpolation to make it consistent with that of the shared matrix S.

[0124] S3118 Matrix Update: Update the new kernel matrix Perform matrix product operation with shared matrix S and use the result as the new shared matrix .

[0125] ; .

[0126] Shared matrix The calculation formula is: .

[0127] S3119 Feature Fusion: Sharing Matrix Multiply the data with the predicted weather data and record the result as the shared feature Z.

[0128] For example: the obtained forecast weather data is {12.3℃, 78%, 1012.5Pa, 0mm, 2.1m / s, 45...} o , 350 lux, 20%}. In this embodiment, the numerical setting of wind direction is set with north as 0 degrees and increases in clockwise direction. For the convenience of calculation, the predicted weather data is converted into a matrix form, which is recorded as the predicted weather matrix V.

[0129] .

[0130] The calculation formula of shared feature Z is: .

[0131] S31110 Optimization Model: Define the mean squared error as the loss function, use a task-specific branch to process the shared features, and use the gradient descent method to minimize the loss function to iteratively optimize the modified sub-model until the loss function reaches its minimum value or reaches the preset number of iterations, thus completing the model optimization. The optimized modified sub-model is then used as the new modified sub-model.

[0132] In this embodiment, in the first processing step S32, the predicted weather data is input into the new modified sub-model in the optimization model in S31110.

[0133] In this embodiment, based on a multi-task learning framework, the correlation between each sample pair in the deviation dataset is calculated to construct a shared matrix. This helps the model learn the error correlation between various features of the prediction data under different weather categories, based on the interrelationships between multiple features in the deviation dataset. Furthermore, multiplying the correlations between various features in the deviation dataset with the predicted weather data to construct shared features helps the model learn error patterns in the deviation dataset and feature patterns in the prediction data, enhancing the relationship patterns between various features in the error data of the prediction data under different weather categories. Simultaneously, model optimization based on the constructed shared features helps improve the model's flexibility and adaptability in adjusting the predicted weather data according to different weather scenarios, improving the model's interpretability. Moreover, the correction factor based on the output improves the accuracy of the predicted weather data.

[0134] Example 4: The difference from Example 1 is that: After executing S2 to build the model and before executing S3 to train the model, it also includes S22 historical data optimization.

[0135] S22 Historical data optimization includes: S221 first definition, S222 construction function, S223 first setting, S224 second setting, S225 third setting, S226 particle optimization and S227 first optimization, such as Figure 4 shown.

[0136] S221 First Definition: Based on historical data, define optimization objectives, including minimizing redundancy, minimizing noise, and maximizing information content.

[0137] S222 Build function: Build the objective function based on the optimization goal.

[0138] The objective function is: .

[0139] in, Describe the objective function. Represents historical data. represents the optimization objective of minimizing redundancy, This represents the optimization objective of minimizing noise. This represents the goal of maximizing information content optimization. 、 and This represents the weight of each optimization objective. .

[0140] S223 First Setting: Define particles as a set of historical data, including historical weather data, historical wind and solar load data, historical energy storage data, and historical energy storage strategies. A particle swarm optimization algorithm is used to solve the objective function, calculate the particle velocity, and update the particle position based on the particle velocity.

[0141] The particle velocity update formula is: .

[0142] The particle position update formula is: .

[0143] in, Represents particles exist The speed of time, The inertial weight of a particle. Indicates individual learning factors. Indicates the random number generated when the individual's optimal position changes, Represents particles The individual's optimal position Represents particles exist The current position at any given moment. represents the social learning factor, This represents a random number generated when the global optimal position changes. This indicates the globally optimal position.

[0144] S224 Second setting: Calculate the difference between the particle's fitness after the update and its fitness before the update. Record the calculated difference as the first difference. If the first difference is greater than 0, it indicates that the particle accepts the updated position. If the first difference is not greater than 0, the probability acceptance mechanism of the simulated annealing algorithm is used to calculate the probability of the particle accepting the updated position. If the calculated probability is greater than the random number generated after the particle update, it indicates that the particle accepts the updated position, and the particle position is adjusted to the updated position. If the calculated probability is not greater than the random number generated after the particle update, the particle position is adjusted to the position before the update.

[0145] The formula for calculating the first difference is: .

[0146] in, Represents particles exist Adaptability at any time Represents particles exist Adaptability at any given moment.

[0147] The formula for adjusting the particle position by the probability acceptance mechanism is: .

[0148] in, This represents the probability of accepting the new position of the particle. This represents the temperature parameter.

[0149] S225 The third setting: obtaining a random number in the interval [0, 1] generated when the particle position is updated again, and determining whether the random number is less than a preset probability threshold.

[0150] If so, the krill swarm algorithm is used to perform a local search near the particle's current position to explore the particle's possible best position. Then, a global search is used to calculate the relative positional relationship of all particles, and the particle with the best fitness is selected. After the local search and global search are completed, the particle's current position is adjusted for the third time, and S226 particle optimization is performed.

[0151] If not, then execute S227, the first optimization.

[0152] S226 Particle Optimization: Calculate the fitness of the particle before and after position adjustment, and determine whether the fitness after position adjustment is greater than the fitness before adjustment.

[0153] If so, the particle position after the third adjustment is used as the new particle position, and the new particle position is used to update the particle position after the third adjustment.

[0154] If not, the position of the particle is restored to the re-adjusted particle position.

[0155] S227 First Optimization: Obtain the historical data corresponding to the particle position after re-adjustment, record it as the first historical data, and update the first historical data as the historical data.

[0156] In this embodiment, during the S3 model training, the updated historical data is input into the energy storage strategy optimization model.

[0157] Example 5: The difference from Example 1 is that: The energy storage strategy optimization model also includes a charging and discharging load prediction sub-model.

[0158] S11 Second Data Acquisition: Acquire real-time electric vehicle data; real-time electric vehicle data includes real-time charging pile charging data, real-time electricity prices, and real-time holiday data.

[0159] Real-time charging data includes the real-time charging status, real-time charging power, and corresponding real-time charging duration.

[0160] Holiday data includes both weekdays and non-working days. Non-working days include statutory holidays (including Spring Festival, Qingming Festival, Labor Day, Dragon Boat Festival, Mid-Autumn Festival, and National Day), mobile holidays (including New Year's Day and Children's Day), and Saturdays and Sundays. Working days generally refer to Monday through Friday. In some cases, holidays overlap with each other, with Saturdays and Sundays, and with weekdays. In these cases, all overlapping dates are marked as the corresponding holidays. For example, the Dragon Boat Festival includes Friday, Saturday, and Sunday, collectively referred to as the Dragon Boat Festival. Real-time holiday data is imported through an electronic calendar system.

[0161] S12 Third Collection: Collect input data, which includes historical charging data of charging piles marked with holiday or non-holiday labels, historical electricity prices, and first data.

[0162] Historical charging data includes historical charging station status, historical charging power, and corresponding historical charging duration.

[0163] S35 Charge / Discharge Prediction: Input the input data into the charge / discharge load prediction sub-model to obtain the third data.

[0164] S36 First concatenation: The sum of the third data and the second data is used as the new second data, and then S34 Third processing is performed.

[0165] In the S4 strategy optimization, real-time electric vehicle data, real-time weather data, real-time energy storage data, and real-time weather data and real-time energy storage data are used as inputs to the trained energy storage strategy optimization model to obtain the real-time energy storage strategy.

[0166] In this embodiment, a charging and discharging load prediction sub-model is constructed, which considers the changes in electric vehicle charging behavior under extreme weather conditions and the impact of electricity price changes on grid load. This optimizes the grid load and helps the model to comprehensively consider the impact of electric vehicle charging load under extreme weather conditions, energy price fluctuations, and demand fluctuations unique to holidays on grid load fluctuations. This improves the accuracy and rationality of the energy storage decision output by the model. At the same time, it helps the model to flexibly adjust energy storage strategies under different scenarios, improves the overall operating efficiency of the energy storage system, and further enhances the stability of the power grid.

[0167] The above are all preferred embodiments of the present application, and are not intended to limit the scope of protection of the present application. Therefore, any equivalent changes made based on the structure, shape, and principle of the present application should be included in the scope of protection of the present application.

Claims

1. A distribution network energy storage strategy recommendation method based on artificial intelligence, characterized in that: include: Data collection: collect real-time weather data, real-time energy storage data and historical data; historical data includes historical weather data, historical energy storage data and historical energy storage strategies; Model building: Building an energy storage strategy optimization model, which includes a weather sub-model, a correction sub-model, a wind and solar power and load sub-model, and an energy storage sub-model; Model training: including first recognition, first processing, second processing and third processing; First identification: input historical weather data into the weather sub-model to obtain predicted weather data, and determine whether the predicted weather data is extreme weather; If yes, then execute the first processing step; If not, the predicted weather data is recorded as the first data, and the second processing step is performed; First processing: inputting the forecast weather data into the correction sub-model to obtain a correction factor, correcting the forecast weather data based on the correction factor to obtain corrected forecast weather data, updating the corrected forecast weather data to the first data, and executing the second processing step; Second processing: inputting the first data into the wind, solar and load sub-model and outputting the second data; The third process: using the first data, the second data, the historical energy storage data and the historical energy storage strategy to train the energy storage sub-model, and obtain a trained energy storage strategy optimization model; Strategy optimization: Real-time weather data and real-time energy storage data are used as inputs to the trained energy storage strategy optimization model to obtain the real-time energy storage strategy.

2. The method for recommending distribution network energy storage strategies based on artificial intelligence according to claim 1, characterized in that: After executing the step of building the model and before executing the step of model training, it also includes: Set Window: Define preset time window; First collection: collect meteorological data from each meteorological station within a preset time window; The first setting is: each weather station is regarded as the i-th level node of the tree structure, and the node characteristics of the i-th level node include the meteorological data of each weather station; Node aggregation: Use the clustering algorithm to perform preliminary aggregation on the node features of the i-th layer nodes to obtain preliminary aggregation results, and determine the i+1-th layer nodes based on the preliminary aggregation results; Node judgment: judge whether the number of nodes in the i+1th layer is 1: If yes, the tree structure is output, and then the first calculation step is performed; If not, the i+1th layer node is used as the new i-th layer node and the node aggregation step is performed; First calculation: calculate the similarity of meteorological data between the nodes in the i-th layer and the nodes in the i+1-th layer, and use the calculated similarity as the connection strength between the nodes in the i-th layer and the nodes in the i+1-th layer; First construction: obtain the neurons corresponding to the nodes at each layer of the tree structure in the weather sub-model, update the connection weights between the neurons according to the connection strength between the nodes, and obtain the updated weather sub-model; In the first identification step, historical weather data is input into the updated weather sub-model.

3. The method for recommending distribution network energy storage strategies based on artificial intelligence according to claim 2, characterized in that: After the step of executing node judgment and before the step of executing the first calculation, the method further includes: First judgment: Calculate the change between the average value of the meteorological data of each node in the current preset time window and the average value of the meteorological data in the previous preset time window, and judge whether the change exceeds the preset change threshold: If yes, then execute the second judgment step; If not, no action will be taken; Second judgment: Calculate the similarity of meteorological data between the i-th layer node and the i+1-th layer node in the current preset time window, record it as the first similarity, and judge whether the first similarity is lower than the preset similarity threshold: If yes, then perform structural splitting on the i-th layer nodes, obtain the splitting results, update the splitting results to the i-th layer nodes, and perform the node aggregation step; If not, no action will be taken.

4. The method for recommending distribution network energy storage strategies based on artificial intelligence according to claim 3 is characterized in that: After executing the first construction step and before executing the model training step, the method further includes: Feature addition: Get all node depths in the tree structure. Node features also include node depths. Graph structure construction: The nodes in the tree structure are used as graph nodes, and the connection relationships between each pair of nodes in the tree structure are used as edges between pairs of graph nodes; Set weights: Use the connection strength between each node pair as the meteorological feature weight of the edge, calculate the spatial distance between the positions of each node pair, and use the inverse of the spatial distance as the spatial distance weight of the edge; Model update: Based on the edges between graph nodes and graph node pairs, a graph structure is constructed. Based on the graph structure, the connection weights between neurons in the weather sub-model are updated to obtain an updated weather sub-model. In the first identification step, historical weather data is input into the updated weather sub-model.

5. The method for recommending distribution network energy storage strategies based on artificial intelligence according to claim 1, characterized in that: After performing the first identification step and before performing the first processing step, the method further includes: First extraction: the time when the forecast weather data is obtained, recorded as the first time; Second extraction: obtaining historical weather data at the first time, recorded as the first labeled data; Calculate the change value: calculate the difference between the predicted weather data and the first labeled data to obtain the deviation data set; Constructing a matrix: Calculate the correlation between each sample pair in the deviation data set, construct a matrix based on the calculated correlation, and obtain a shared matrix; Feature fusion: Multiply the shared matrix with the predicted weather data, and record the result as the shared feature; Optimize the model: define the loss function, use task-specific branches to process the shared features, iteratively optimize the corrected sub-model by minimizing the loss function, and use the optimized corrected sub-model as the new corrected sub-model.

6. The method for recommending distribution network energy storage strategies based on artificial intelligence according to claim 5 is characterized in that: After executing the step of constructing the matrix and before executing the step of feature fusion, it also includes: Feature screening: The sample corresponding to the sample pair whose correlation is greater than the preset feature correlation threshold is recorded as the first sample; Feature mapping: using the kernel function to calculate the kernel value of the first sample pair, and constructing the kernel matrix based on the calculation result; Feature unification: Based on the dimension of the shared matrix, adjust the dimension of the kernel matrix and use the adjusted kernel matrix as the new kernel matrix; Matrix update: Perform product operation on the new kernel matrix and the shared matrix, and use the result as the new shared matrix.

7. The method for recommending distribution network energy storage strategies based on artificial intelligence according to claim 1, characterized in that: After executing the step of building the model and before executing the step of model training, it also includes: First definition: Based on historical data, define optimization objectives, which include minimizing redundancy, minimizing noise, and maximizing information content; Construct function: construct the objective function based on the optimization goal; First setting: define particles as a group of data in historical data, use particle swarm optimization algorithm to solve the objective function, calculate the particle speed, and update the particle position according to the particle speed; The second setting: calculate the difference between the fitness of the particle after the update and the fitness of the particle before the update, record the calculated difference as the first difference, and adjust the particle position again based on the first difference using the probability acceptance mechanism of the simulated annealing algorithm; First optimization: obtaining the historical data corresponding to the particle position after re-adjustment, recording it as the first historical data, and updating the first historical data as the historical data.

8. The method for recommending distribution network energy storage strategies based on artificial intelligence according to claim 7, characterized in that: After performing the second setting step and before performing the first optimization step, the method further includes: The third setting: Get the random number generated when the particle position is updated again, and determine whether the random number is less than the preset probability threshold: If yes, the krill swarm algorithm is used to adjust the particle position for the third time, and the particle optimization steps are performed: If not, then execute the first optimization step; Particle optimization: Calculate the fitness before and after the particle position is adjusted, and determine whether the fitness after the particle position is adjusted is greater than the fitness before the particle is adjusted: If yes, the particle position adjusted for the third time is used as the new particle position, and the new particle position is used to update the particle position adjusted again; If not, the position of the particle is restored to the re-adjusted particle position.

9. The method for recommending distribution network energy storage strategies based on artificial intelligence according to claim 1, characterized in that: The energy storage strategy optimization model also includes a charge and discharge load prediction sub-model; the method also includes: Second collection: collect real-time electric vehicle data; real-time electric vehicle data includes real-time charging pile charging data, real-time electricity prices and real-time holiday data; In the strategy optimization step, real-time electric vehicle data, real-time weather data, and real-time energy storage data are input into the trained energy storage strategy optimization model.

10. The method for recommending distribution network energy storage strategy based on artificial intelligence according to claim 9, characterized in that: The method further comprises: Third collection: collecting input data, the input data includes historical charging pile charging data marked with holiday labels or non-holiday labels, historical electricity prices and the first data; Charge and discharge prediction: Input data into the charge and discharge load prediction sub-model to obtain the third data; First concatenation: taking the sum of the third data and the second data as new second data.

Images