A method and system for short-term load forecasting of charging stations

Abstract

This invention relates to the field of data reasoning technology and discloses a method and system for short-term load forecasting of charging stations. The method includes: acquiring historical load data and corresponding temperature data of the charging station; performing empirical mode decomposition on the historical load data to obtain the intrinsic mode components and residual components of the charging station; performing cosine similarity fusion on the intrinsic mode components to obtain the fused load component of the intrinsic mode components; initializing individual gray wolf individuals and setting the parameters of a long short-term memory neural network based on the gray wolf individuals to obtain an initial prediction model for the charging station; iteratively training the initial prediction model based on the fused load component, residual component, and temperature data to obtain a trained prediction model; and inputting real-time data of the target charging station into the trained prediction model to obtain the short-term load forecast value of the target charging station. This invention can improve the accuracy of short-term load forecasting of charging stations.

Description

Technical Field

[0001] This invention relates to the field of data reasoning technology, and in particular to a method and system for short-term load forecasting of charging stations. Background Technology

[0002] In the field of short-term load forecasting for charging stations, existing technologies often fail to effectively decompose and extract features from historical load data at multiple scales, making it difficult to accurately capture the complex fluctuation patterns and potential dynamics contained within the load data. This results in forecasting models failing to fully utilize the effective information in the data, leading to significant discrepancies between forecast results and actual load, and making it difficult to meet the requirements for high-precision forecasting.

[0003] Meanwhile, existing prediction methods have limitations in model construction and optimization. They often fail to fully integrate key environmental factors such as temperature, and the adjustment of model parameters lacks an adaptive mechanism, easily leading to slow convergence speed and weak generalization ability. In scenarios where charging station loads are affected by multiple dynamic factors, existing technologies struggle to quickly respond to load change trends, resulting in insufficient timeliness and reliability of short-term predictions, and failing to provide effective support for the operation and scheduling of charging stations. Summary of the Invention

[0004] This invention provides a method and system for short-term load forecasting of charging stations to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a short-term load forecasting method for charging stations, comprising:

[0006] S1. Obtain historical load data and corresponding temperature data for the charging station;

[0007] S2. Perform empirical mode decomposition on the historical load data to obtain the intrinsic mode components and residual components of the charging station;

[0008] S3. Perform cosine similarity fusion on the intrinsic mode components to obtain the fused load components of the intrinsic mode components;

[0009] S4. Initialize the gray wolf population individuals, and set the parameters of the long short-term memory neural network based on the gray wolf population individuals to obtain the initial prediction model of the charging station.

[0010] S5. Based on the fused load component, the residual component, and the temperature data, the initial prediction model is iteratively trained to obtain the trained prediction model.

[0011] S6. Input the real-time data of the target charging station into the trained prediction model to obtain the short-load prediction value of the target charging station.

[0012] In a preferred embodiment, the step of performing empirical mode decomposition on the historical load data to obtain the intrinsic mode components and residual components of the charging station includes:

[0013] Identify local extreme points in historical load data;

[0014] The upper and lower envelope curves of the charging station are fitted based on the local extreme points.

[0015] Determine the mean curves of the upper envelope curve and the lower envelope curve;

[0016] The difference between the historical load data and the mean curve is used as a temporary component. The historical load data is repeatedly filtered until the intrinsic mode component condition is met, thereby obtaining the intrinsic mode component and residual component of the charging station.

[0017] In a preferred embodiment, the step of performing cosine similarity fusion on the intrinsic mode components to obtain the fused load component of the intrinsic mode components includes:

[0018] Extract the fluctuation characteristics of the intrinsic mode components;

[0019] The directional consistency between any two feature vectors in the wave feature is quantified to obtain the directional consistency metric value of the wave feature.

[0020] The intrinsic modal components with similarity higher than a set threshold in the directional consistency metric are merged into a group of components of the same type.

[0021] The intrinsic mode components within the same component group are weighted and superimposed to generate a fused load component.

[0022] In a preferred embodiment, the initialization of individual gray wolf individuals and the setting of parameters for a long short-term memory neural network based on these individuals to obtain an initial prediction model for the charging station includes:

[0023] An initial population location vector set is constructed based on a chaotic sequence generator to obtain the individual gray wolf population.

[0024] The location coordinates of the individuals in the gray wolf population are mapped to the configuration values ​​of the number of neurons in the hidden layer of a long short-term memory neural network.

[0025] A subset of the position coordinate parameters is converted into initial values ​​for the weight matrix and the bias vector.

[0026] The neural network architecture is constructed based on the configured number of neurons, the initial weight matrix, and the initial bias vector to obtain the initial prediction model of the charging station.

[0027] In a preferred embodiment, the step of iteratively training the initial prediction model based on the fused load component, the residual component, and the temperature data to obtain the trained prediction model includes:

[0028] Based on the time-series alignment matrix of the fused load component, the residual component, and the temperature data, a multi-source input dataset for the charging station is constructed.

[0029] The multi-source input dataset is divided into a training subset, a validation subset, and a test subset;

[0030] A dynamic learning rate adjustment strategy is used to control the parameter update step size of the initial prediction model;

[0031] The initial prediction model is iteratively trained based on the training subset and the parameter update step size;

[0032] By monitoring the convergence status of the prediction error during iterative training using the validation subset, a training termination instruction for the initial prediction model can be obtained.

[0033] When the iterative training satisfies the training termination instruction, the network parameters of the initial prediction model are locked.

[0034] The locked initial prediction model is validated based on the test subset, and the validated network parameters are input into the initial prediction model to obtain the trained prediction model.

[0035] In a preferred embodiment, the iterative training of the initial prediction model based on the training subset and the parameter update step size includes:

[0036] Using the training subset as training data and the parameter update step size as the learning rate, the initial prediction model is iteratively trained.

[0037] Calculate the fitness value during iterative training, wherein the fitness value is calculated using the following formula:

[0038]

[0039] In the formula, The fitness value is... The length of the time series. As a time factor, The actual load value for iterative training. The predicted load value for iterative training;

[0040] The individual positions in the gray wolf population are updated according to the fitness value, and the updated individual positions are input into the long short-term memory neural network to obtain the updated long short-term memory neural network.

[0041] In a preferred embodiment, the formula for calculating the individual's location is as follows:

[0042]

[0043] In the formula, For the individual's location, Set a maximum value for the location of the individual. It is a natural exponential function. As a time factor, This represents the maximum value of the time factor.

[0044] In a preferred embodiment, the step of monitoring the convergence state of the prediction error during iterative training through the validation subset to obtain a training termination instruction for the initial prediction model includes:

[0045] The mean absolute error and root mean square error of the initial prediction model are calculated based on the predicted values ​​of the initial prediction model during iterative training and the validation subset. The formula for calculating the mean absolute error is as follows:

[0046]

[0047] In the formula, The mean absolute error is... The total number of data points in the training subset. For the first The true values ​​corresponding to each training data point in the validation subset. For the first Predicted values ​​for each training data set The data ordinal number of the training subset;

[0048] The root mean square error is calculated using the following formula:

[0049]

[0050] In the formula, The root mean square error is... For the first The true values ​​corresponding to each training data point in the validation subset. For the first Predicted values ​​for each training data set The data ordinal number of the training subset;

[0051] When the mean absolute error and the root mean square error are at the error threshold for three consecutive window periods, the training termination instruction for the initial prediction model is obtained.

[0052] In a preferred embodiment, the step of inputting real-time data of the target charging station into the trained prediction model to obtain the short-load prediction value of the target charging station includes:

[0053] The current load data stream and ambient temperature monitoring value of the target charging station are collected to obtain the real-time data of the target charging station;

[0054] The real-time data is used to construct features to obtain the standardized input features of the target charging station;

[0055] The standardized input features are imported into the trained prediction model to obtain the original prediction result sequence of the target charging station;

[0056] A short-load forecast report for the target charging station is generated based on the original forecast result sequence.

[0057] To address the above problems, the present invention also provides a short-term load forecasting system for charging stations, the system comprising:

[0058] The data acquisition module is used to acquire historical load data and temperature data for the corresponding time period of the charging station;

[0059] The historical load data decomposition module is used to perform empirical mode decomposition on the historical load data to obtain the intrinsic mode components and residual components of the charging station.

[0060] The cosine similarity fusion module is used to perform cosine similarity fusion on the intrinsic mode components to obtain the fused load component of the intrinsic mode components;

[0061] The gray wolf population initialization module is used to initialize individual gray wolf individuals and set the parameters of the long short-term memory neural network based on the gray wolf individuals to obtain the initial prediction model of the charging station.

[0062] The model training module is used to iteratively train the initial prediction model based on the fused load component, the residual component, and the temperature data to obtain the trained prediction model.

[0063] The load prediction module is used to input real-time data of the target charging station into the trained prediction model to obtain the short-term load prediction value of the target charging station.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] 1. This invention obtains intrinsic mode components and residual components by performing empirical mode decomposition on historical load data, and combines it with cosine similarity fusion processing, which can effectively extract complex fluctuation characteristics in load data, improve the utilization rate of data information, provide more accurate basic data support for subsequent forecasting, and thus enhance the accuracy of short-term load forecasting.

[0066] 2. This invention uses gray wolf population optimization to set parameters of long short-term memory neural network, combines multi-source data for iterative training, and optimizes the model training process through dynamic learning rate adjustment strategy, so that the prediction model has better convergence performance and generalization ability, can efficiently handle dynamic changes in charging station load, significantly improve the efficiency and reliability of short-term load forecasting, and provide strong data support for the operation and management of charging stations. Attached Figure Description

[0067] Figure 1 This is a flowchart illustrating a short-term load forecasting method for charging stations provided in an embodiment of the present invention.

[0068] Figure 2 This is a diagram of the improved gray wolf optimization algorithm architecture of the charging station short-term load forecasting method provided in an embodiment of the present invention;

[0069] Figure 3 Here is a flowchart of the IGWO-LSTM short-term load forecasting method for charging stations provided in an embodiment of the present invention.

[0070] Figure 4 This is a functional block diagram of a short-term load forecasting system for charging stations provided in an embodiment of the present invention.

[0071] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0072] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0073] This application provides a method for short-term load forecasting of charging stations. The execution entity of the short-term load forecasting method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the short-term load forecasting method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0074] Reference Figure 1 The diagram shown is a flowchart illustrating a short-term load forecasting method for charging stations according to an embodiment of the present invention. In this embodiment, the short-term load forecasting method for charging stations includes:

[0075] S1. Obtain historical load data and corresponding temperature data for the charging station;

[0076] In this embodiment of the invention, the specific information of the target charging station is specified, including the accurate name of the charging station, its detailed geographical location (such as province, city, district, street and specific house number or landmark), and the daily operating hours of the charging station (such as operating all day from 00:00 to 24:00, or specific hours from 6:00 am to 10:00 pm, etc.). This information will serve as the basis for subsequent data acquisition.

[0077] Furthermore, contact the operator of the target charging station, explaining the need for historical load data and providing specific information about the confirmed charging station to verify its identity. After verifying the information, the operator will extract the load data for each time period (e.g., hourly, half-hourly) from its internal charging management system database, according to the required historical time period (e.g., the past month, a quarter, etc.). This data directly reflects the charging station's power consumption during that time period, and is presented as specific numerical records including the corresponding date and time information.

[0078] Furthermore, based on the detailed geographical location of the target charging station, access the official website of the local meteorological department or an authoritative meteorological data platform (such as China Weather Network), enter the geographical location information in the data query function, and select the same historical time period as the charging station's historical load data. The platform will filter out the temperature data for that geographical location within the corresponding time period based on the input information. This data includes the specific temperature value for each time period (consistent with the time period division of the load data, such as hourly temperature), as well as the corresponding date and time information, ensuring a one-to-one correspondence between the temperature data and the load data in time.

[0079] Furthermore, the historical load data obtained from the charging station operator and the corresponding time period temperature data obtained from the meteorological platform will be sorted together. The dates and times of the two sets of data will be checked to see if they match completely. If there are any missing or mismatched data for individual time periods, the corresponding data provider will be contacted in a timely manner to supplement or correct them. Finally, a complete and one-to-one combination of historical load data and corresponding time period temperature data of the charging station will be formed.

[0080] In summary, obtaining historical load data and corresponding temperature data for charging stations provides comprehensive and closely related foundational information for predictive models. Historical load data directly reflects the past electricity consumption patterns and fluctuation characteristics of charging stations, serving as the core basis for capturing load change trends; while temperature data for corresponding time periods, as a key environmental factor affecting charging demand, can effectively supplement external driving information on load changes.

[0081] In summary, the combination of the two allows the prediction model to simultaneously consider internal load patterns and the influence of the external environment, providing multi-dimensional data support for subsequent empirical mode decomposition, feature fusion, and model training. This helps improve the model's ability to capture complex load changes, thereby laying a solid data foundation for improving the accuracy and reliability of short-term load forecasting.

[0082] S2. Perform empirical mode decomposition on the historical load data to obtain the intrinsic mode components and residual components of the charging station;

[0083] In this embodiment of the invention, the step of performing empirical mode decomposition on the historical load data to obtain the intrinsic mode components and residual components of the charging station includes:

[0084] Identify local extreme points in historical load data;

[0085] The upper and lower envelope curves of the charging station are fitted based on the local extreme points.

[0086] Determine the mean curves of the upper envelope curve and the lower envelope curve;

[0087] The difference between the historical load data and the mean curve is used as a temporary component. The historical load data is repeatedly filtered until the intrinsic mode component condition is met, thereby obtaining the intrinsic mode component and residual component of the charging station.

[0088] Specifically, examine the historical load data of the charging station. Starting from the first data point, observe the relationship between the values ​​of each data point and its two adjacent data points. If the value of a data point is greater than the values ​​of the preceding and following data points, then that data point is a local maximum. If the value of a data point is less than the values ​​of the preceding and following data points, then that data point is a local minimum. Find all data points that meet the above conditions. These data points together constitute the local extreme points of the historical load data.

[0089] Furthermore, the local maxima are extracted from all local extreme points. According to the time sequence of these points in the historical load data, a smooth curve is used to connect each local maxima in turn, so that the curve accurately passes through each local maxima. This curve is the upper envelope curve of the charging station. Then, the local minima are extracted from all local extreme points. Similarly, according to the time sequence, a smooth curve is used to connect each local minima in turn, so that the curve accurately passes through each local minima. This curve is the lower envelope curve of the charging station.

[0090] Furthermore, at each time point corresponding to the historical load data, the corresponding values ​​of the upper envelope curve and the lower envelope curve at that time point are found. The two values ​​are added together and then divided by 2. The result is the mean value at that time point. Connecting the mean values ​​of all time points in chronological order forms the mean curve of the upper envelope curve and the lower envelope curve.

[0091] Furthermore, in the historical load data, the data at each time point is subtracted from the corresponding mean of the mean curve at that time point. The result is used as a temporary component. This temporary component is used as a new processing object. The operations of identifying local extreme points, fitting the upper and lower envelope curves, determining the mean curve, and calculating the new temporary component are repeated until the obtained temporary component meets the conditions of the intrinsic mode component. That is, the mean of the upper and lower envelope curves of the temporary component is close to zero, and the number of its local extreme points is equal to or differs from the number of zero crossings by at most one. At this time, this temporary component is the intrinsic mode component of the charging station. After subtracting all the obtained intrinsic mode components from the original historical load data, the remaining part is the residual component.

[0092] In summary, performing empirical mode decomposition (EMD) on historical load data to obtain intrinsic mode components (IMCs) and residual components can decompose complex load data into multiple IMCs with different time scales and fluctuation characteristics, as well as residual components reflecting the overall trend. This decomposition method can accurately extract various fluctuation information contained in the load data, transforming the originally difficult-to-analyze raw data into a single feature component that is easier to process, effectively highlighting the inherent patterns and local characteristics of the load data.

[0093] In summary, the components obtained from the decomposition specifically present different levels of load change characteristics, providing a clear processing target for subsequent cosine similarity fusion. This helps improve the accuracy and effectiveness of feature fusion, thereby providing more accurate and representative data input for the prediction model and laying an important foundation for improving the accuracy of short-term load forecasting.

[0094] S3. Perform cosine similarity fusion on the intrinsic mode components to obtain the fused load components of the intrinsic mode components;

[0095] In this embodiment of the invention, the step of performing cosine similarity fusion on the intrinsic mode components to obtain the fused load component of the intrinsic mode components includes:

[0096] Extract the fluctuation characteristics of the intrinsic mode components;

[0097] The directional consistency between any two feature vectors in the wave feature is quantified to obtain the directional consistency metric value of the wave feature.

[0098] The intrinsic modal components with similarity higher than a set threshold in the directional consistency metric are merged into a group of components of the same type.

[0099] The intrinsic mode components within the same component group are weighted and superimposed to generate a fused load component.

[0100] Specifically, observe the value of each intrinsic mode component as time changes, record the time period when the value rises, the time period when it falls, the maximum value of the fluctuation, the minimum value of the fluctuation, and the time interval between two adjacent fluctuation peaks for each component. Integrate this information into the feature vector of each intrinsic mode component. Each feature vector contains all the specific information related to the fluctuation mentioned above, thereby completing the extraction of the fluctuation characteristics of the intrinsic mode components.

[0101] Furthermore, from all the eigenvectors of the fluctuation characteristics, any two eigenvectors are selected, and the rising and falling periods recorded in these two vectors are compared. The number of times the rising periods completely overlap and the number of times the falling periods completely overlap in the two vectors are counted. The sum of these two overlaps is divided by the total number of periods in the two vectors (the sum of the number of rising and falling periods). The result is the measure of the directional consistency between the two eigenvectors. The directional consistency measure between all any two eigenvectors is calculated in this way.

[0102] Furthermore, a fixed threshold, such as 0.6, is set, and the directional consistency metric value corresponding to each pair of intrinsic modal components is checked one by one. When the directional consistency metric value of a pair of components is greater than 0.6, the two components are grouped into the same group. This operation is repeated for all intrinsic modal components. If the directional consistency metric value of a component and multiple components in a group are all higher than 0.6, then the component is added to the group. Finally, multiple groups of components of the same type are formed, and the directional consistency metric value between the intrinsic modal components in each group is higher than the set 0.6.

[0103] Furthermore, for each group of similar components, the fluctuation amplitude of each intrinsic mode component within the group is calculated. The fluctuation amplitude is the difference between the maximum and minimum values ​​of that component. The fluctuation amplitude of each component is divided by the sum of the fluctuation amplitudes of all components in the group to obtain the weight of each component in the group. Then, at each time point, the value of each intrinsic mode component in the group is multiplied by its corresponding weight, and all products are added together. The result is the fusion load component at that time point. The results of all time points are sorted in chronological order to generate the complete fusion load component.

[0104] In summary, cosine similarity fusion of intrinsic mode components to obtain fused load components can accurately identify intrinsic mode components with similar fluctuation characteristics by quantifying the directional consistency between any two feature vectors. Merging components with similarity higher than a set threshold into similar component groups and performing weighted superposition can effectively integrate load characteristics with coordinated change patterns, reduce data redundancy, and enhance the representation capability of key fluctuation information.

[0105] In summary, this fusion approach retains the core features of each component while increasing the aggregation of data through merging similar components. This allows subsequent model training to be based on more concise and representative load features, which helps enhance the model's ability to capture load change trends and provides strong support for improving the accuracy and efficiency of short-term load forecasting.

[0106] S4. Initialize the gray wolf population individuals, and set the parameters of the long short-term memory neural network based on the gray wolf population individuals to obtain the initial prediction model of the charging station.

[0107] In this embodiment of the invention, the initialization of individual gray wolf individuals and the setting of parameters for a long short-term memory neural network based on these individuals to obtain an initial prediction model for the charging station include:

[0108] An initial population location vector set is constructed based on a chaotic sequence generator to obtain the individual gray wolf population.

[0109] The location coordinates of the individuals in the gray wolf population are mapped to the configuration values ​​of the number of neurons in the hidden layer of a long short-term memory neural network.

[0110] A subset of the position coordinate parameters is converted into initial values ​​for the weight matrix and the bias vector.

[0111] The neural network architecture is constructed based on the configured number of neurons, the initial weight matrix, and the initial bias vector to obtain the initial prediction model of the charging station.

[0112] Specifically, a chaotic sequence generator is activated, which generates a series of numerical sequences with chaotic characteristics through a specific iterative method. The numerical values ​​in these sequences are arranged into multiple vectors according to a preset dimension. Each vector is a position vector. All these position vectors together constitute the initial population position vector set, and each position vector corresponds to a gray wolf population individual.

[0113] Furthermore, for the location coordinate parameters of each individual gray wolf in the population, the range of its values ​​is determined, and this range is proportionally mapped to the effective range of the number of neurons in the hidden layer of the long short-term memory neural network (e.g., between 10 and 100). Through numerical transformation, each value of the location coordinate parameter corresponds to an integer within this effective range, and this integer is the configuration value of the number of neurons in the hidden layer corresponding to that individual gray wolf in the population.

[0114] Furthermore, a subset of continuous values ​​is selected from the position coordinate parameters of individuals in the gray wolf population. The value range of this subset is determined, and this range is linearly transformed to a preset range for the initial values ​​of the weight matrix (e.g., -0.5 to 0.5) and the preset range for the initial values ​​of the bias vector (e.g., -0.1 to 0.1). The transformed values ​​are used as the initial values ​​of the weight matrix and the bias vector, respectively. The initial values ​​of the weight matrix are arranged according to the dimension of the inter-layer connections of the neural network, and the initial values ​​of the bias vector are arranged according to the order of the corresponding neurons.

[0115] Furthermore, based on the obtained configuration value of the number of hidden layer neurons, the number of neurons in the hidden layer of the long short-term memory neural network is set, the initial value of the converted weight matrix is ​​assigned to the connection weights between each layer of the neural network, the initial value of the bias vector is assigned to the bias terms of each layer of neurons, and the network structure is built in the order of input layer, hidden layer, and output layer to ensure that the signal transmission path between each layer is correct. The network constructed in this way is the initial prediction model of the charging station.

[0116] In summary, initializing individual gray wolf individuals and setting the parameters of the Long Short-Term Memory (LSTM) neural network based on these individuals to obtain an initial prediction model allows for diverse initial configurations of the neural network parameters with global search capabilities, utilizing the initial population location vector set constructed by a chaotic sequence generator. Mapping the location coordinates of individual individuals to the initial values ​​of the number of hidden layer neurons, weight matrix, and bias vector enables intelligent setting of key parameters of the neural network architecture, avoiding local optima traps that may result from manual parameter initialization.

[0117] In summary, this parameter initialization method based on gray wolf population optimization allows the initial prediction model to have a more reasonable architectural foundation and parameter distribution before training, providing a good starting point for subsequent iterative training. It helps to improve the convergence speed and optimization efficiency of the model, enabling the model to approach the optimal state more quickly, thus laying an important foundation for improving the overall performance of short-term load forecasting.

[0118] S5. Based on the fused load component, the residual component, and the temperature data, the initial prediction model is iteratively trained to obtain the trained prediction model.

[0119] In this embodiment of the invention, the step of iteratively training the initial prediction model based on the fused load component, the residual component, and the temperature data to obtain the trained prediction model includes:

[0120] Based on the time-series alignment matrix of the fused load component, the residual component, and the temperature data, a multi-source input dataset for the charging station is constructed.

[0121] The multi-source input dataset is divided into a training subset, a validation subset, and a test subset;

[0122] A dynamic learning rate adjustment strategy is used to control the parameter update step size of the initial prediction model;

[0123] The initial prediction model is iteratively trained based on the training subset and the parameter update step size;

[0124] By monitoring the convergence status of the prediction error during iterative training using the validation subset, a training termination instruction for the initial prediction model can be obtained.

[0125] When the iterative training satisfies the training termination instruction, the network parameters of the initial prediction model are locked.

[0126] The locked initial prediction model is validated based on the test subset, and the validated network parameters are input into the initial prediction model to obtain the trained prediction model.

[0127] The iterative training of the initial prediction model based on the training subset and the parameter update step size includes:

[0128] Using the training subset as training data and the parameter update step size as the learning rate, the initial prediction model is iteratively trained.

[0129] Calculate the fitness value during iterative training, wherein the fitness value is calculated using the following formula:

[0130]

[0131] In the formula, The fitness value is... The length of the time series. As a time factor, The actual load value for iterative training. The predicted load value for iterative training;

[0132] The individual positions in the gray wolf population are updated according to the fitness value, and the updated individual positions are input into the long short-term memory neural network to obtain the updated long short-term memory neural network.

[0133] The formula for calculating the individual's location is as follows:

[0134]

[0135] In the formula, For the individual's location, Set a maximum value for the location of the individual. It is a natural exponential function. As a time factor, This represents the maximum value of the time factor.

[0136] The step of monitoring the convergence state of the prediction error during iterative training through the validation subset to obtain the training termination instruction for the initial prediction model includes:

[0137] The mean absolute error and root mean square error of the initial prediction model are calculated based on the predicted values ​​of the initial prediction model during iterative training and the validation subset. The formula for calculating the mean absolute error is as follows:

[0138]

[0139] In the formula, The mean absolute error is... The total number of data points in the training subset. For the first The true values ​​corresponding to each training data point in the validation subset. For the first Predicted values ​​for each training data set The data ordinal number of the training subset;

[0140] The root mean square error is calculated using the following formula:

[0141]

[0142] In the formula, The root mean square error is... For the first The true values ​​corresponding to each training data point in the validation subset. For the first Predicted values ​​for each training data set The data ordinal number of the training subset;

[0143] When the mean absolute error and the root mean square error are at the error threshold for three consecutive window periods, the training termination instruction for the initial prediction model is obtained.

[0144] Specifically, 70% of the data is randomly selected from the multi-source input dataset and used as the training subset for learning the parameters of the initial prediction model. Then, 15% of the remaining data is randomly selected as the validation subset to monitor the performance changes during model training. Finally, the remaining 15% of the data is used as the test subset to evaluate the final prediction ability of the model after training.

[0145] Furthermore, the dynamic learning rate adjustment strategy is as follows: In the first 10 rounds of training the initial prediction model, the learning rate is kept at a fixed value to allow the model parameters to be updated quickly; starting from the 11th round, if the prediction error of the validation subset decreases by less than 0.01 for three consecutive rounds, the learning rate is reduced to 0.5 times the original value to slow down the parameter update speed and allow the model to converge more stably; if the error does not decrease significantly after the learning rate is reduced to 0.01 times the initial value, the learning rate is maintained until training terminates.

[0146] Furthermore, the input data of the training subset is sequentially input into the initial prediction model in chronological order. The model calculates the output prediction value based on the current weight matrix and bias vector. The prediction value is compared with the corresponding actual value in the training subset to obtain the prediction error. The step size is updated based on the error and the current parameters, and the model's weight matrix and bias vector are adjusted to complete one parameter update. The above process is repeated until the preset maximum number of training rounds is reached or a training termination instruction is received.

[0147] Furthermore, after each iteration of the initial prediction model training, the input data of the validation subset is input into the model to obtain the validation prediction value. The error between the validation prediction value and the actual value in the validation subset is calculated and recorded. The error change is continuously monitored for 10 rounds. If the average error of these 10 rounds decreases by less than 0.005 compared with the average error of the previous 10 rounds, and the fluctuation range of the error value in the last 3 rounds is less than 0.003, then the model is determined to have converged, and a training termination instruction is generated.

[0148] Furthermore, when a training termination instruction is received during iterative training, the parameter updates of the initial prediction model are immediately stopped, and the weight matrix values ​​and bias vector values ​​in the model at this time are fixed and saved without any further modification. These fixed parameters are the optimal parameters of the model in the current training state.

[0149] Furthermore, the input data of the test subset is completely input into the initial prediction model with locked parameters. The model outputs test prediction values, and the average error between the test prediction values ​​and the actual values ​​in the test subset is calculated. If the average error is less than the preset threshold of 0.05, the verification is deemed successful, and the locked network parameters are reloaded into the initial prediction model. At this point, the model is the prediction model that has been trained.

[0150] Specifically, all data from the training subset are input into the initial prediction model one by one in chronological order. The model calculates and outputs the corresponding predicted value based on the current weight matrix and bias vector. Each predicted value is compared with the corresponding actual load value in the training subset to obtain the prediction error for each data point. Then, according to the parameter update step size, the weight matrix values ​​connecting each layer of neurons and the bias vector values ​​of each neuron in the model are adjusted to reduce the prediction error. After completing one parameter adjustment, the data from the training subset is input into the model again to repeat the above process until the preset number of iterations for this round of training is completed.

[0151] Furthermore, the fitness value is calculated based on all prediction errors generated during the iterative training process. Specifically, the sum of prediction errors for all data points in the training subset is calculated first, and then the reciprocal of this sum is taken. The result is the fitness value for this round of iterative training. The larger the fitness value, the better the prediction performance of the model in this round of training.

[0152] Furthermore, the fitness values ​​of all individuals in the gray wolf population are compared, and the individual with the highest fitness value is identified as the optimal individual. Its position vector is recorded, and other individuals move towards the position of the optimal individual. The distance they move is determined by the difference between their own fitness value and the fitness value of the optimal individual. The smaller the difference, the shorter the movement distance; the larger the difference, the longer the movement distance. After updating the positions of all individuals, the updated position coordinate parameters of each individual are converted into the configuration values ​​of the number of hidden layer neurons, the initial values ​​of the weight matrix, and the initial values ​​of the bias vector of the Long Short-Term Memory Neural Network according to the previous mapping method. The network constructed in this way is the updated Long Short-Term Memory Neural Network.

[0153] Specifically, in the formula for calculating the fitness value, It is the length of the time series, and its value comes from the total number of time points contained in the training subset. Each time point corresponds to a time factor. , The value starts from 1 and increases sequentially to 1. This covers all time records in the training subset. These are the actual load values ​​from iterative training, derived from the charging stations recorded at various time points in the training subset. The actual load data are real observations collected and processed in the early stages. These are the load prediction values ​​from iterative training, derived from the initial prediction model during iterative training at each time point. The result is output after predicting the load of the charging station.

[0154] Furthermore, the significance of this formula lies in calculating all time points. Actual load value Compared with load forecast The fitness value is obtained by summing the squares of the differences between the values ​​and taking the negative of this sum. This is used to measure the prediction performance of the initial prediction model during iterative training. The fitness value directly reflects how close the model's prediction results are to the true values.

[0155] Furthermore, the trend of the formula is as follows: when the load forecast value Compared with the actual load value As they get closer, each time point corresponding The smaller the square value, the smaller the sum of squares at all time points, and the fitness value obtained after taking a negative number. The larger the load forecast, the greater the value. Compared with the actual load value The greater the difference, the more important each time point... corresponding The larger the square value, the larger the sum of squares at all time points; the fitness value obtained after taking a negative number is... The smaller the fitness value, the higher the fitness value will be; that is, the fitness value will increase as the model's prediction accuracy improves and decrease as the prediction accuracy decreases.

[0156] Specifically, in the formula for calculating the individual's position, It represents the individual's location, and its value is derived from the position coordinates of the individual in the gray wolf population during the iteration process. It is a specific location value calculated using this formula. This is the maximum value set for an individual's location. This value is preset by the user based on the maximum possible range of an individual's location in the actual problem, and is used to limit the upper limit of the individual's location. It is a time factor, the value of which corresponds to the current step number in the iterative training, increasing sequentially from 1 at the start of the iteration until it reaches a certain value. . It is the maximum value of the time factor, which is preset by the user and represents the total number of steps in iterative training. This is the natural exponential function, which performs natural exponential operations on the value within the parentheses, that is, calculates the result with the natural constant as the base and the value within the parentheses as the exponent.

[0157] Furthermore, this formula is used to calculate the position of an individual in a gray wolf population at the current time factor t. By using time factors With the set maximum value By combining these methods and utilizing the properties of the natural exponential function, individual positions can be determined. During the iteration process, with the time factor... It dynamically adjusts according to changes, thereby achieving reasonable updates of individual positions.

[0158] Furthermore, the trend of the formula is as follows: when the time factor... When gradually increasing from 1, The value gradually changed from close to 0 to -1. The value of then gradually changes from close to 1 to close to the reciprocal of the natural constant (approximately 0.3679). The value gradually changed from close to 0.5 to close to 0.730, therefore The value will be close to The larger values ​​subtracted from 0.5 are gradually decreased until the time factor is reached. achieve hour, The value becomes close to Subtract the smaller value of 0.730, which is the individual location. It will change with the time factor It gradually decreases as it increases.

[0159] Specifically, in each round of iterative training, all the input data contained in the validation subset are input into the initial prediction model in chronological order. The model outputs the load prediction value corresponding to each time point, and at the same time, the actual load data corresponding to these time points are extracted from the validation subset as the true values, ensuring that the predicted values ​​and the true values ​​correspond one-to-one in time, forming pairs of numerical groups.

[0160] Furthermore, when calculating the mean absolute error, each pair of predicted and true values ​​is processed one by one. The predicted value at each time point is subtracted from the true value at that time point to obtain the error at each time point. Then, the absolute value of each error is calculated. The absolute values ​​of the errors at all time points are added together, and the sum is divided by the total number of time points in the validation subset. The result is the mean absolute error of the initial prediction model in this round of iterative training.

[0161] Furthermore, when calculating the root mean square error, for each pair of predicted and true values, the difference between the predicted and true values ​​at each time point is first calculated, and then each difference is squared to obtain the squared value of the error at each time point. The squared values ​​of the errors at all time points are added together, and the sum is divided by the total number of time points in the validation subset to obtain the average value of the squared errors. Then, the square root of this average value is taken, and the result is the root mean square error of the initial prediction model in this round of iterative training.

[0162] Furthermore, a mean absolute error threshold and a root mean square error threshold are preset. The mean absolute error threshold is determined based on the fluctuation range of the historical load data of the charging station. For example, if the maximum fluctuation value of the historical load data is 100, the mean absolute error threshold is set to 5. The root mean square error threshold is also set based on the characteristics of historical data, for example, it is set to 8. These two thresholds are used to determine whether the model has achieved sufficient prediction accuracy.

[0163] Furthermore, after each round of iterative training, the calculated mean absolute error is compared with the set mean absolute error threshold, and the root mean square error is also compared with the set root mean square error threshold. If the mean absolute error is less than or equal to the mean absolute error threshold, and the root mean square error is less than or equal to the root mean square error threshold, and this situation occurs consecutively for three rounds, then the initial prediction model is determined to have converged, and a training termination instruction is generated; if the above conditions are not met, the next round of iterative training continues.

[0164] Specifically, In This represents the total number of data points in the training subset. Its value comes from the total number of data records contained in the training subset, i.e., how many data points are in the training subset. Take only the amount you need. It is the first The true value corresponding to the training data in the validation subset, which is derived from the data in the validation subset and the training data in the validation subset. The actual observations corresponding to each training data time or feature are real data that has been collected and organized in the early stage. It is the first The predicted value for the training data is derived from the initial prediction model's prediction of the first training data. The output result after predicting from the training data. It is the ordinal number of the training subset, which starts from 1 and increases sequentially to 1. This corresponds to the sequential number of each data point in the training subset. (RMSE) , , , The origin and The values ​​in the dataset are completely consistent, each corresponding to the total number of data in the training subset, the true value in the validation subset, the model's predicted value, and the ordinal number of the data.

[0165] Furthermore, The significance is to calculate the sum of the absolute values ​​of the differences between the true and predicted values ​​of all data in the training subset, and then divide this sum by the total number of data points. The results are used to measure the mean absolute deviation between the predicted values ​​of the initial prediction model and the actual values. The smaller the value, the smaller the average deviation of the model's predictions, and the more stable the prediction performance. The significance is to first calculate the sum of the squares of the differences between the true values ​​and the predicted values ​​of all data in the training subset, divide the sum by the total number of data N to obtain the mean squared deviation, and then take the square root of the mean squared deviation. The result is used to measure the overall deviation between the model's predicted values ​​and the true values, and is especially sensitive to large deviations. The smaller the RMSE, the smaller the overall deviation of the model's prediction, especially the fewer cases of large deviations.

[0166] Furthermore, The trend is as follows: when the predicted value y_i differs from the actual value The closer they get, the more each The smaller the value, the more... The smaller the sum, the more likely it is to be divided by The result obtained later The smaller the predicted value, the better; Compared with the true value The greater the gap, the more each The larger the value, the larger the sum, and the larger the MAE. That is, MAE decreases as the predicted value gets closer to the actual value, and increases as the deviation increases. The trend of RMSE is as follows: when the predicted value... Compared with the true value The closer they get, the more each The smaller the value, the more... The smaller the sum, the more we divide by The smaller the mean square deviation, the better the result after taking the square root. The smaller the value, the better; when there is a large deviation between the predicted value and the actual value, The sum will be amplified due to the squaring operation, resulting in a significant increase in the total. This also increases significantly, meaning that RMSE decreases as prediction accuracy improves, and its response to large biases is more pronounced. More sensitive.

[0167] Specifically, the improved Grey Wolf optimization algorithm endows the initial population with chaotic ergodicity through chaotic mapping, breaking the distribution limitations of random initialization and avoiding early local optima; nonlinear control of the convergence rhythm precisely balances the "global exploration" and "local development" stages, improving search efficiency. The dual-strategy synergy enhances the global optimization capability of the improved Grey Wolf optimization algorithm, such as... Figure 2 The improved Grey Wolf optimization algorithm architecture for short-term load forecasting of charging stations is shown in the diagram.

[0168] In summary, iterative training of the initial prediction model based on fused load, residual, and temperature data allows for the collaborative consideration of different characteristic components of the load data with key environmental factors through the construction of multi-source input datasets, providing a more comprehensive training basis for the model. Dividing the dataset into training, validation, and testing subsets allows for real-time monitoring of the prediction error convergence status during training via the validation subset. Combined with a dynamic learning rate adjustment strategy, this precisely controls the parameter update step size, ensuring the accuracy and efficiency of the model training direction.

[0169] In summary, this training method not only makes full use of the complementary information of multi-dimensional data, but also avoids model overfitting through a hierarchical validation mechanism. At the same time, it achieves dynamic optimization of parameters by using the fitness calculation of gray wolf population optimization, enabling the trained model to more accurately capture the load change pattern, and has stronger generalization ability and prediction stability, providing key support for improving the accuracy and reliability of short-term load forecasting for charging stations.

[0170] S6. Input the real-time data of the target charging station into the trained prediction model to obtain the short-load prediction value of the target charging station.

[0171] In this embodiment of the invention, the step of inputting real-time data of the target charging station into the trained prediction model to obtain the short-load prediction value of the target charging station includes:

[0172] The current load data stream and ambient temperature monitoring value of the target charging station are collected to obtain the real-time data of the target charging station;

[0173] The real-time data is used to construct features to obtain the standardized input features of the target charging station;

[0174] The standardized input features are imported into the trained prediction model to obtain the original prediction result sequence of the target charging station;

[0175] A short-load forecast report for the target charging station is generated based on the original forecast result sequence.

[0176] Specifically, load monitoring sensors are installed in the power distribution system of the target charging station to collect current load data every 5 minutes, forming a continuous load data stream; at the same time, temperature sensors are set up in the charging station site to record the ambient temperature every 5 minutes, obtaining the corresponding ambient temperature monitoring value. The load data and temperature data at the same time point are paired and combined to form the real-time data of the target charging station.

[0177] Furthermore, the time information corresponding to each data point is extracted from the real-time data, including the number of hours and whether it is a working day, as a time feature; the load difference between two adjacent time points is calculated to obtain the load change rate feature; the load data stream and the ambient temperature monitoring value are linearly transformed according to their historical maximum and minimum values ​​to make the values ​​fall between 0 and 1, thus completing the standardization process. The time feature, the load change rate feature, and the standardized load and temperature data are integrated to form the standardized input features of the target charging station.

[0178] Furthermore, according to the data format required by the trained prediction model, the standardized input features are input into the model one by one. The model processes the input features through its internal network structure and outputs the load prediction values ​​every 5 minutes for the next hour. These prediction values ​​are arranged in chronological order to form the original prediction result sequence of the target charging station.

[0179] Furthermore, the original prediction result sequence is organized, the start and end times of the prediction are clearly marked, the predicted value at each time point is matched with the corresponding time information, the average value and maximum fluctuation range of the predicted value are calculated, the specific predicted values ​​are presented in tabular form, the load change trend within the prediction period is described in text, and these contents are integrated to form a short-load prediction report for the target charging station.

[0180] Specifically, the short-term load forecasting of charging stations constructs a closed loop of "decomposition-fusion-modeling-optimization": First, historical load data is decomposed into multi-scale intrinsic mode components (IMF1~IMF) through empirical mode decomposition (EMD). mThe algorithm first extracts the load component and residual components to remove complex fluctuation features. Then, it aggregates IMF components of similar fluctuation patterns using cosine similarity to eliminate redundancy and enhance feature synergy. Subsequently, it inputs environmental data such as fused load components, residuals, and temperature into a Long Short-Term Memory (LSTM) network to train the model using its ability to capture long and short-term dependencies. Generalization is ensured through training / validation / test sets. Simultaneously, an improved Grey Wolf Optimization (IGWO) algorithm is introduced. First, the wolf pack is initialized using a Tent chaotic mapping to obtain α, β, δ, and ω. Then, α, β, δ, and ω in the individual positions of the population are mapped to key LSTM parameters (such as weights and hidden layer structure). Here, α represents the leader of the wolf pack, the current optimal solution, guiding the pack towards prey during the search process. Its position information plays a crucial guiding role in updating the positions of other wolves. β is the second-highest level in the pack, a suboptimal solution, assisting α in decision-making. When α's decision has biases or uncertainties, β's opinion can supplement and correct it, assisting α in guiding the position updates of other wolves. δ belongs to the third level, the third optimal solution, and follows α. The commands of α, β, and δ play a crucial role in the social structure of the wolf pack, serving as a link between higher and lower wolves and a reference factor guiding the position updates of other wolves. ω represents an ordinary member of the pack, exploring new areas in the search space and updating its own position based on the positional information of α, β, and δ, thus expanding the search range and helping the algorithm avoid getting trapped in local optima and increasing its global search capability. The prediction error is used as the fitness function to iteratively optimize parameters, breaking through the local optimum trap. Finally, the trained model is output after iterative convergence. Through cross-technology collaboration of signal processing, deep learning, and intelligent optimization, the accuracy and stability of load forecasting are systematically improved, providing a highly reliable decision-making basis for power scheduling and resource allocation at charging stations. Figure 3 The flowchart of the IGWO-LSTM short-term load forecasting method for charging stations is shown.

[0181] In summary, inputting real-time data from the target charging station into the trained prediction model to obtain short-term load forecasts ensures consistency between the input data and the multi-source data features used during model training by collecting current load data streams and ambient temperature monitoring values, providing real-time evidence for accurate predictions. Constructing features from the real-time data and generating standardized input features ensures a high degree of matching between the data format and model input requirements, reducing information loss during data conversion and guaranteeing the model's effective interpretation of real-time data.

[0182] In summary, this forecasting method relies on the accurate pattern-capturing ability of the trained model and responds promptly to the current load status through dynamic input of real-time data. The generated short-term load forecast report can accurately reflect the load change trend of the target charging station, providing timely and reliable decision support for the operation scheduling and power resource allocation of the charging station, effectively improving the practicality and timeliness of short-term load forecasting.

[0183] like Figure 4 The diagram shown is a functional block diagram of a short-term load forecasting system for charging stations provided in an embodiment of the present invention.

[0184] The charging station short-term load forecasting system 100 of this invention can be installed in an electronic device. Depending on the functions implemented, the charging station short-term load forecasting system 100 may include a data acquisition module 101, a historical load data decomposition module 102, a cosine similarity fusion module 103, a gray wolf population initialization module 104, a model training module 105, and a load forecasting module 106. The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and are stored in the memory of the electronic device.

[0185] In this embodiment, the functions of each module / unit are as follows:

[0186] The data acquisition module 101 is used to acquire historical load data and temperature data for the corresponding time period of the charging station.

[0187] The historical load data decomposition module 102 is used to perform empirical mode decomposition on the historical load data to obtain the intrinsic mode components and residual components of the charging station.

[0188] The cosine similarity fusion module 103 is used to perform cosine similarity fusion on the intrinsic mode components to obtain the fused load component of the intrinsic mode components.

[0189] The initialization gray wolf population module 104 is used to initialize gray wolf population individuals and set the parameters of the long short-term memory neural network based on the gray wolf population individuals to obtain the initial prediction model of the charging station.

[0190] The model training module 105 is used to iteratively train the initial prediction model based on the fused load component, the residual component and the temperature data to obtain the trained prediction model.

[0191] The load prediction module 106 is used to input the real-time data of the target charging station into the trained prediction model to obtain the short-term load prediction value of the target charging station.

[0192] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0193] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0194] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0195] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0196] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0197] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for short-term load forecasting of charging stations, characterized in that, The method includes: S1. Obtain historical load data and corresponding temperature data for the charging station; S2. Perform empirical mode decomposition on the historical load data to obtain the intrinsic mode components and residual components of the charging station; S3. Perform cosine similarity fusion on the intrinsic mode components to obtain the fused load components of the intrinsic mode components; S4. Initialize the gray wolf population individuals, and set the parameters of the long short-term memory neural network based on the gray wolf population individuals to obtain the initial prediction model of the charging station. S5. Based on the fused load component, the residual component, and the temperature data, the initial prediction model is iteratively trained to obtain the trained prediction model. S6. Input the real-time data of the target charging station into the trained prediction model to obtain the short-load prediction value of the target charging station.

2. The short-term load forecasting method for charging stations as described in claim 1, characterized in that, The step of performing empirical mode decomposition on the historical load data to obtain the intrinsic mode components and residual components of the charging station includes: Identify local extreme points in historical load data; The upper and lower envelope curves of the charging station are fitted based on the local extreme points. Determine the mean curves of the upper envelope curve and the lower envelope curve; The difference between the historical load data and the mean curve is used as a temporary component. The historical load data is repeatedly filtered until the intrinsic mode component condition is met, thereby obtaining the intrinsic mode component and residual component of the charging station.

3. The short-term load forecasting method for charging stations as described in claim 1, characterized in that, The step of performing cosine similarity fusion on the intrinsic mode components to obtain the fused load component of the intrinsic mode components includes: Extract the fluctuation characteristics of the intrinsic mode components; The directional consistency between any two feature vectors in the wave feature is quantified to obtain the directional consistency metric value of the wave feature. The intrinsic modal components with similarity higher than a set threshold in the directional consistency metric are merged into a group of components of the same type. The intrinsic mode components within the same component group are weighted and superimposed to generate a fused load component.

4. The short-term load forecasting method for charging stations as described in claim 1, characterized in that, The process of initializing individual gray wolf individuals and setting parameters for a long short-term memory neural network based on these individuals to obtain an initial prediction model for the charging station includes: An initial population location vector set is constructed based on a chaotic sequence generator to obtain the individual gray wolf population. The location coordinates of the individuals in the gray wolf population are mapped to the configuration values ​​of the number of neurons in the hidden layer of a long short-term memory neural network. A subset of the position coordinate parameters is converted into initial values ​​for the weight matrix and the bias vector. The neural network architecture is constructed based on the configured number of neurons, the initial weight matrix, and the initial bias vector to obtain the initial prediction model of the charging station.

5. The short-term load forecasting method for charging stations as described in claim 4, characterized in that, The step of iteratively training the initial prediction model based on the fused load component, the residual component, and the temperature data to obtain the trained prediction model includes: Based on the time-series alignment matrix of the fused load component, the residual component, and the temperature data, a multi-source input dataset for the charging station is constructed. The multi-source input dataset is divided into a training subset, a validation subset, and a test subset; A dynamic learning rate adjustment strategy is used to control the parameter update step size of the initial prediction model; The initial prediction model is iteratively trained based on the training subset and the parameter update step size; By monitoring the convergence status of the prediction error during iterative training using the validation subset, a training termination instruction for the initial prediction model can be obtained. When the iterative training satisfies the training termination instruction, the network parameters of the initial prediction model are locked. The locked initial prediction model is validated based on the test subset, and the validated network parameters are input into the initial prediction model to obtain the trained prediction model.

6. The short-term load forecasting method for charging stations as described in claim 5, characterized in that, The iterative training of the initial prediction model based on the training subset and the parameter update step size includes: Using the training subset as training data and the parameter update step size as the learning rate, the initial prediction model is iteratively trained. Calculate the fitness value during iterative training, wherein the fitness value is calculated using the following formula: ， In the formula, The fitness value is... The length of the time series. As a time factor, The actual load value for iterative training. The predicted load value for iterative training; The individual positions in the gray wolf population are updated according to the fitness value, and the updated individual positions are input into the long short-term memory neural network to obtain the updated long short-term memory neural network.

7. The short-term load forecasting method for charging stations as described in claim 6, characterized in that, The formula for calculating the individual's location is as follows: ， In the formula, For the individual's location, Set a maximum value for the location of the individual. It is a natural exponential function. As a time factor, This represents the maximum value of the time factor.

8. The short-term load forecasting method for charging stations as described in claim 7, characterized in that, The step of monitoring the convergence state of the prediction error during iterative training through the validation subset to obtain the training termination instruction for the initial prediction model includes: The mean absolute error and root mean square error of the initial prediction model are calculated based on the predicted values ​​of the initial prediction model during iterative training and the validation subset. The formula for calculating the mean absolute error is as follows: ， In the formula, The mean absolute error is... The total number of data points in the training subset. For the first The true values ​​corresponding to each training data point in the validation subset. For the first Predicted values ​​for each training data set The data ordinal number of the training subset; The root mean square error is calculated using the following formula: , In the formula, The root mean square error is... For the first The true values ​​corresponding to each training data point in the validation subset. For the first Predicted values ​​for each training data set The data ordinal number of the training subset; When the mean absolute error and the root mean square error are at the error threshold for three consecutive window periods, the training termination instruction for the initial prediction model is obtained.

9. The short-term load forecasting method for charging stations as described in claim 1, characterized in that, The step of inputting real-time data from the target charging station into the trained prediction model to obtain the short-load prediction value of the target charging station includes: The current load data stream and ambient temperature monitoring value of the target charging station are collected to obtain the real-time data of the target charging station; The real-time data is used to construct features to obtain the standardized input features of the target charging station; The standardized input features are imported into the trained prediction model to obtain the original prediction result sequence of the target charging station; A short-load forecast report for the target charging station is generated based on the original forecast result sequence.

10. A short-term load forecasting system for charging stations, characterized in that, The system includes: The data acquisition module is used to acquire historical load data and temperature data for the corresponding time period of the charging station; The historical load data decomposition module is used to perform empirical mode decomposition on the historical load data to obtain the intrinsic mode components and residual components of the charging station. The cosine similarity fusion module is used to perform cosine similarity fusion on the intrinsic mode components to obtain the fused load component of the intrinsic mode components; The gray wolf population initialization module is used to initialize individual gray wolf individuals and set the parameters of the long short-term memory neural network based on the gray wolf individuals to obtain the initial prediction model of the charging station. The model training module is used to iteratively train the initial prediction model based on the fused load component, the residual component, and the temperature data to obtain the trained prediction model. The load prediction module is used to input real-time data of the target charging station into the trained prediction model to obtain the short-term load prediction value of the target charging station.

Images