Electric vehicle load characteristic global modeling method and system based on multiple scenes

Abstract

The invention is suitable for the technical field of electric vehicle load characteristic modeling, and provides an electric vehicle load characteristic global modeling method and system based on multiple scenes, and the method comprises the steps: obtaining electric vehicle load related data based on multiple channels, and carrying out the preprocessing of the obtained data; extracting initial vector features according to the preprocessed data, wherein the initial vector features comprise power features, time features, spatial features and behavior features; determining a first target vector sub-feature according to the power feature, the time feature and the spatial feature, and determining a second target vector sub-feature according to the behavior feature; and a multi-scene electric vehicle load model is constructed based on the first target vector sub-features and the second target vector sub-features, the constructed model is optimized by using meta learning, and an optimized model is obtained, the electric vehicle load characteristics are dynamically described from multiple dimensions, and an accurate and comprehensive data basis is provided for subsequent load prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of electric vehicle load characteristic modeling, and particularly relates to a global modeling method and system for electric vehicle load characteristics based on multiple scenarios. Background Art

[0002] In practical applications, the charging load of electric vehicles is affected not only by factors such as the number of charging piles, the layout of charging facilities, and user behavior, but also by spatio-temporal dynamic factors and the external environment, making the load characteristics of electric vehicles show complexity and variability. Currently, most existing electric vehicle load modeling methods focus on the load characteristic analysis under a single scenario, lacking globality and dynamics, and are prone to ignoring the correlation between different scenarios and the impact of spatio-temporal changes. In addition, most models adopt static models and do not consider the randomness and uncertainty of the behavior characteristics of users using electric vehicles, making it difficult to adaptively optimize and schedule under different scenarios and time conditions, resulting in the inability to accurately predict and effectively manage the charging load of electric vehicles.

[0003] In view of this, a global modeling method and system for electric vehicle load characteristics based on multiple scenarios are needed. Summary of the Invention

[0004] Embodiments of the present invention provide a global modeling method and system for electric vehicle load characteristics based on multiple scenarios, which are used to solve the problem of being unable to accurately predict and effectively manage the charging load of electric vehicles.

[0005] A first aspect of an embodiment of the present invention provides a global modeling method for electric vehicle load characteristics based on multiple scenarios, including:

[0006] Obtaining electric vehicle load-related data based on multiple channels and preprocessing the obtained data;

[0007] Extracting initial vector features according to the preprocessed data, where the initial vector features include power features, time features, space features, and behavior features;

[0008] Determining first target vector sub-features according to the power features, time features, and space features, and determining second target vector sub-features according to the behavior features;

[0009] Constructing a multi-scenario electric vehicle load model based on the first target vector sub-features and the second target vector sub-features, and optimizing the constructed model using meta-learning to obtain an optimized model.

[0010] Furthermore, the extracting initial vector features according to the preprocessed data, where the initial vector features include power features, time features, space features, and behavior features, includes:

[0011] Determine the power characteristics by calculating the average charging power, power fluctuation range, and power change rate;

[0012] Use the fast Fourier transform algorithm and autoregressive model to determine the periodic time characteristics and short-term change time characteristics respectively;

[0013] Determine the changes in electric vehicle loads at different geographical locations and the spatial relationship characteristics between regions by combining the geographically weighted regression algorithm and adjacency matrix;

[0014] Analyze the user behavior characteristics according to the machine learning algorithm.

[0015] Furthermore, determining the first target vector sub-characteristics according to the power characteristics, time characteristics, and spatial characteristics, and simultaneously determining the second target vector sub-characteristics according to the behavior characteristics, includes:

[0016] Normalize the data corresponding to the power characteristics, time characteristics, and spatial characteristics;

[0017] Use the trained random forest model to determine the key characteristics corresponding to the power characteristics, time characteristics, and spatial characteristics;

[0018] Perform feature fusion according to the key characteristics corresponding to the power characteristics, time characteristics, and spatial characteristics to obtain the fused first target vector sub-characteristics.

[0019] Furthermore, performing feature fusion according to the key characteristics corresponding to the power characteristics, time characteristics, and spatial characteristics to obtain the fused first target vector sub-characteristics, includes:

[0020]

[0021]

[0022] Where: w Pi 、w Tj and W Sl are the weights corresponding to the i-th power characteristic, the weight corresponding to the j-th time characteristic, and the weight corresponding to the l-th spatial characteristic respectively, and are the i-th standardized power characteristic value, the j-th standardized time characteristic value, and the l-th standardized spatial characteristic value respectively, P * 、T * and S * are the fused power characteristic information, time characteristic information, and spatial characteristic information respectively.

[0023] Further, determining the first target vector sub-feature according to the power feature, time feature, and spatial feature, and simultaneously determining the second target vector sub-feature according to the behavior feature includes:

[0024] Using Bayesian inference to determine the charging behavior sub-feature and travel behavior sub-feature;

[0025] Constructing an association model between weather and charging behavior sub-feature and travel behavior sub-feature, analyzing the relationship between traffic flow data and charging behavior sub-feature, travel behavior sub-feature, and load characteristics, and analyzing the variation law of charging behavior sub-feature, travel behavior sub-feature, and load characteristics during holidays;

[0026] Screening the correlation coefficients between the charging behavior sub-feature and travel behavior sub-feature and the load characteristics, and linearly fusing the selected sub-features.

[0027] Further, the using Bayesian inference to determine the charging behavior sub-feature and travel behavior sub-feature includes:

[0028]

[0029] Where P(H∣D) represents the posterior probability of charging behavior or travel behavior H given the data D, P(D∣H) is the likelihood function, P(H) is the prior probability, and P(D) is the marginal likelihood of the data.

[0030] Further, the screening the correlation coefficients between the charging behavior sub-feature and travel behavior sub-feature and the load characteristics, and linearly fusing the selected sub-features includes:

[0031]

[0032]

[0033] Y = w1x1 + w2x2 +... + w n x n

[0034] Where: r1 is the correlation coefficient between the charging duration and the load characteristics, r2 is the correlation coefficient between the travel probability and the load characteristics, w1 and w2 are the corresponding weights respectively, n is the number of selected behavior sub-features, Y is the fused feature value, x1 is the charging duration feature value, and x2 is the travel frequency feature value.

[0035] Further, constructing a multi-scenario electric vehicle load model based on the first target vector sub-feature and the second target vector sub-feature, and optimizing the constructed model using meta-learning to obtain an optimized model includes:

[0036] Integrate the first target vector sub - feature and the second target vector sub - feature;

[0037] Divide the usage scenarios of electric vehicles according to different dimensions;

[0038] Select several target model architectures to construct an electric vehicle load model according to data characteristics and modeling purposes.

[0039] Furthermore, constructing a multi - scenario electric vehicle load model based on the first target vector sub - feature and the second target vector sub - feature, and optimizing the constructed model using meta - learning to obtain an optimized model, including:

[0040] Divide the electric vehicle load data under different scenarios into multiple tasks according to a preset meta - learning framework;

[0041] Update the initial parameters of the model by methods such as gradient descent according to the training set data of multiple tasks;

[0042] Test the optimized model with new task data until the model performance meets the requirements.

[0043] The second aspect of the embodiments of the present invention provides a global modeling system for the load characteristics of electric vehicles based on multiple scenarios, including:

[0044] A data acquisition and pre - processing unit, configured to acquire electric vehicle load - related data based on multiple channels and pre - process the acquired data;

[0045] An initial vector feature extraction unit, configured to extract initial vector features according to the pre - processed data, where the initial vector features include power features, time features, spatial features, and behavior features;

[0046] A first target vector sub - feature and second target vector sub - feature determination unit, configured to determine the first target vector sub - feature according to the power feature, time feature, and spatial feature, and simultaneously determine the second target vector sub - feature according to the behavior feature;

[0047] A multi - scenario electric vehicle load model construction unit, configured to construct a multi - scenario electric vehicle load model based on the first target vector sub - feature and the second target vector sub - feature, and optimize the constructed model using meta - learning to obtain an optimized model.

[0048] From the above technical solutions, it can be seen that the embodiments of the present invention have the following advantages:

[0049] Based on the acquired data and preprocessing, power, time, space, and behavior characteristics are extracted from different dimensions, comprehensively characterizing the characteristics of electric vehicle loads. Moreover, the extracted characteristics can reflect the key information during the charging and usage processes of electric vehicles, providing an accurate and comprehensive data basis for model construction. The constructed multi-scenario division enables the model to consider the impacts of factors such as different times, spaces, and user behaviors on electric vehicle loads, improving the adaptability and accuracy of the model and enabling a more realistic simulation of the electric vehicle load situation in actual scenarios. Using meta-learning allows the model to quickly adapt and optimize under different scenarios, enhancing the generalization ability and robustness of the model and avoiding problems such as overfitting or performance degradation of the model in new scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 FIG. is a schematic flowchart of an embodiment of a global modeling method for electric vehicle load characteristics based on multiple scenarios in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0052] The global modeling method for electric vehicle load characteristics based on multiple scenarios in this embodiment is used to realistically simulate the electric vehicle load situation in actual scenarios from different dimensions. The implementation method in this embodiment can be implemented in a system, on a server, or on a terminal, and no specific limitation is made.

[0053] Embodiment 1

[0054] Please refer to Figure 1 , an embodiment of a global modeling method for electric vehicle load characteristics based on multiple scenarios in the present invention includes the following steps:

[0055] S11. Acquire electric vehicle load-related data based on multiple channels and preprocess the acquired data;

[0056] Specifically, the data obtained here includes real-time charging data of charging piles, including information such as charging power, charging time, charging electricity, and charging location; obtaining driving data of the vehicle from sensors and data recording systems built into the vehicle, such as driving mileage, speed, battery status, etc.; obtaining traffic flow data, road congestion information, etc.; and additionally including the distribution of electric vehicle charging piles in different scenarios. Through the geographic information system, the geographic coordinates of each charging pile and the specific location of the area where it is located can be obtained. Information data on external factors such as temperature, humidity, weather changes, traffic flow, and holidays, and information data on factors such as charging price, user behavior patterns, and regional economic levels.

[0057] The preprocessing includes standardization, cleaning, and noise processing. Data standardization is to normalize the load data collected under different scenarios. Assume that the electric vehicle load data is X = {x1, x2, …, x n} and the standardization formula is:

[0058]

[0059] where μ is the mean of the data and σ is the standard deviation of the data. After standardization, the mean of the data is 0 and the standard deviation is 1, ensuring that the data has a unified scale.

[0060] For noise processing, wavelet transform is used to remove noise in the data. Wavelet transform can decompose the signal and filter high-frequency noise, thus retaining the main features of the signal. Assume that the load signal is y(t), and the signal after wavelet transform is y wavelet (t), where:

[0061]

[0062] where W i is the wavelet coefficient and ψ i (t) is the wavelet basis function. Through wavelet transform, high-frequency noise in the signal can be removed.

[0063] S12. Extract initial vector features based on the preprocessed data. The initial vector features include power features, time features, spatial features, and behavior features;

[0064] 1. Determine power features by calculating the average charging power, power fluctuation range, and power change rate;

[0065] Through data statistical analysis methods, extract the average charging power from the preprocessed charging data, and calculate the ratio of the total charging electricity to the charging time within a certain period. Monitor the maximum power and minimum power during the charging process to determine the power fluctuation range. Use the differential calculation method to obtain the power change rate, and calculate the ratio of the difference in charging power between adjacent time points to the time interval to determine the power change trend.

[0066] 2. Use the fast Fourier transform algorithm and the autoregressive model respectively to determine the periodic time characteristics and short-term change time characteristics;

[0067] Specifically, the fast Fourier transform algorithm is used to analyze the periodic changes in the load data. Assuming the load sequence is y(t), FFT can convert it into frequency-domain data, thereby extracting the frequency characteristics of the load changes. The FFT formula is:

[0068]

[0069] where Y(f) is the frequency-domain data, f is the frequency, and T is the data length. By analyzing the frequency-domain data, the periodic characteristics of the load changes can be extracted, especially the load fluctuations related to factors such as daily charging peaks, seasonal changes, and weather changes.

[0070] The autoregressive model is used to capture the time-dependent characteristics of the load data. The mathematical formula is:

[0071]

[0072] where y t is the load value at the current time, φ i is the autoregressive coefficient, p is the model order, and ∈ t is the noise term. By fitting the historical load data, a time-dependent model of the load data can be established. In the present invention, the AR model not only captures the changes in the electric vehicle load over time, but also considers factors such as the charging demand fluctuations in different time periods and the uncertainties of user charging behaviors.

[0073] The above steps first use FFT for periodic feature extraction, and then use the AR model for short-term time-dependent modeling. FFT and the AR model are respectively used to capture different time characteristics of the electric vehicle load. Among them, FFT is used to extract periodic characteristics in a larger range and is suitable for load fluctuations on a longer time scale, such as the periodic changes in charging demands between day and night; while the AR model focuses on capturing the time dependence of the load in the short term, especially the charging demand fluctuations in different time periods and different scenarios.

[0074] 3. Determine the changes in the electric vehicle load at different geographical locations and the spatial relationship characteristics between regions through the combination of the geographically weighted regression algorithm and the adjacency matrix;

[0075] Geographically weighted regression (GWR) is used to analyze the relationship between the electric vehicle load and the geographical location. The formula is:

[0076]

[0077] where y iis the load data for the i-th location, x ij is the j-th feature of the i-th location, β j is the regression coefficient, ∈ i is the error term. Through the GWR model, the spatial dependence of electric vehicle loads in different regions can be extracted, especially considering the influence of the distribution of charging piles, the power demand fluctuations in each region, and the correlations with external factors such as climate and traffic flow.

[0078] The adjacency matrix model represents the spatial relationship of charging piles by constructing a graph structure. The elements of the adjacency matrix represent the spatial connection strength between charging piles. Suppose A is the adjacency matrix, A ij represents the spatial correlation degree between the i-th charging pile and the j-th charging pile. By calculating the structural characteristics of the graph, the change patterns of electric vehicle loads in different regions can be further analyzed, especially when the charging demand of electric vehicles fluctuates greatly, considering the mutual dependence of loads between regions.

[0079] The combination of the above GWR and adjacency matrix enables the model to not only capture the load differences brought by geographical locations but also reveal the spatial dependence of charging demands between different regions. GWR analyzes the changes in electric vehicle loads at different geographical locations, while the adjacency matrix reveals the spatial relationships between regions through the graph structure, especially the mutual correlation of charging demands in different regions during peak hours or holidays and other special situations.

[0080] 4. Analyze user behavior characteristics according to machine learning algorithms.

[0081] Using classification and regression methods in machine learning, based on the charging history data and travel trajectory data of users, combined with the basic information of users, a user behavior model is constructed. Specific user behavior characteristics include charging behavior sub-characteristics and travel behavior sub-characteristics, which are specifically described in reference to step S13.

[0082] S13. Determine the first target vector sub-characteristics according to the power characteristics, time characteristics, and spatial characteristics, and at the same time determine the second target vector sub-characteristics according to the behavior characteristics;

[0083] 1. Standardize the data corresponding to the power characteristics, time characteristics, and spatial characteristics;

[0084] Extract the power characteristics, time characteristics, and spatial characteristics from the electric vehicle load characteristic dataset, form these characteristics into a feature matrix, where each row represents a sample and each column represents a feature; at the same time, determine the corresponding target variable, such as the load value.

[0085] Construct a random forest model, which consists of multiple decision trees. During the training process, randomly sample samples with replacement from the original dataset to form a new training subset. When splitting each node, randomly select features from all features, and select the optimal splitting feature and splitting point among these features to maximize the information gain of the split node.

[0086] 2. Use the trained random forest model to determine the key features corresponding to the power feature, time feature, and spatial feature;

[0087] After the random forest model is trained, calculate the feature importance to select the key features. Calculate the out-of-bag data error before and after splitting the out-of-bag data on each feature in each decision tree. According to the calculated feature importance, set a threshold and select the features with importance higher than the threshold as the key features.

[0088] 3. Perform feature fusion based on the key features corresponding to the power feature, time feature, and spatial feature to obtain the first target vector sub-feature after fusion.

[0089] Determine the corresponding weight vectors for the power feature, time feature, and spatial feature through the above random forest algorithm. The specific feature fusion formula is as follows:

[0090]

[0091] where: w Pi 、w Tj and w Sl are the weights corresponding to the i-th power feature, the weights corresponding to the j-th time feature, and the weights corresponding to the l-th spatial feature respectively, and are the i-th standardized power feature value, the j-th standardized time feature value, and the l-th standardized spatial feature value respectively, P * 、T * and S * are the power feature information, time feature information, and spatial feature information after fusion respectively.

[0092] In this embodiment, the following is also included:

[0093] 1. Use Bayesian inference to determine the charging behavior sub-feature and travel behavior sub-feature;

[0094] 2. Construct an association model between weather and charging behavior sub-feature and travel behavior sub-feature, analyze the relationship between traffic flow data and charging behavior sub-feature, travel behavior sub-feature, and load characteristics, and analyze the variation law of charging behavior sub-feature and travel behavior sub-feature and load characteristics during holidays;

[0095] Determine prior knowledge. For example, based on past data statistics, obtain the charging probability distributions of electric vehicles in different seasons and time periods, as well as the average charging duration and power range of different vehicle models, and use this as prior information. Collect real-time charging data, including observed values such as the start time, end time, and charging power of charging.

[0096] According to historical travel data, determine the prior probability distributions in aspects such as travel purpose, travel frequency, and travel time. Combine observed information such as the vehicle's GPS positioning data and travel time records, and use Bayesian inference to calculate the posterior probabilities of different travel behaviors. Calculate the probability that a user travels for commuting purposes during a specific time period on weekdays, and use this probability as one of the sub-features of travel behavior. The specific expression is as follows:

[0097]

[0098] Among them, P(H∣D) represents the posterior probability of the charging behavior or travel behavior H given the data D, P(D∣H) is the likelihood function, P(H) is the prior probability, and P(D) is the marginal likelihood of the data.

[0099] Establish an association model between weather and sub-features of behavior and load characteristics. Including that in hot weather, the use of electric vehicle air conditioners increases, affecting the cruising range, and thus affecting the charging demand and travel plans. Through data analysis, find out the sub-features of behavior that have a relatively high correlation with the load characteristics of electric vehicles under different weather conditions.

[0100] Analyze the relationship between traffic flow data and sub-features of behavior and load characteristics. Including that in traffic congestion, the vehicle speed is slow, the energy consumption increases, which may change the travel time and charging demand. Screen out the sub-features of behavior that have a significant correlation with the load characteristics under different traffic flow conditions. During peak congestion periods, the travel time is extended, and the travel time sub-feature of travel behavior is screened out; due to increased energy consumption caused by congestion, the charging behavior sub-feature of charging power demand is screened out.

[0101] Analyze the variation laws of sub-features of charging behavior and travel behavior and load characteristics during holidays. Including that people's travel purposes and travel frequencies during holidays are different from those on weekdays, which affect the charging behavior. During the Spring Festival holiday, long-distance travel increases, and the sub-feature of long-distance travel probability of travel behavior is screened out; on weekends, leisure shopping trips increase, and the sub-feature of leisure shopping travel probability of travel behavior is screened out; at the same time, due to the change in travel patterns, the sub-feature of charging location selection of charging behavior is screened out.

[0102] It should be noted that weather conditions directly affect users' travel needs and charging habits. Cold weather may prompt users to choose to charge at night, while warm weather may encourage them to go out for charging during the day. In the Bayesian inference model, the weather variable W is modeled as part of the conditional probability, and P(H∣W) describes the distribution of charging behavior under different weather conditions. Traffic flow affects users' choice of charging stations and charging times. During peak hours, traffic flow may make users more inclined to choose charging piles closer to home, while when traffic flow is low, users may choose a farther charging station. The traffic flow variable T is introduced into the model, and P(H∣T) quantifies the impact of traffic flow on charging decisions. Holidays may lead to more concentrated charging behaviors during certain periods. The holiday variable H j is introduced into the Bayesian model, and P(H∣H j ) characterizes the impact of holidays on charging habits.

[0103] Based on the above, the Bayesian inference model can derive the probability distribution of users' charging behaviors in different scenarios according to historical charging behavior data and external environmental factors, and further simulate users' charging habits and decision-making processes. The specific quantification formula is as follows:

[0104]

[0105] where W is the weather factor, T is the traffic flow, H j is the holiday effect, P(D∣H,W,T,H j ) is the likelihood function of charging behavior given these factors, and P(H∣W,T,H j ) is the prior probability of charging behavior.

[0106] 3. Screen the correlation coefficients between the sub-features of charging behavior and travel behavior and the load characteristics, and linearly fuse the selected sub-features. The specific calculation formula is as follows:

[0107]

[0108] Y = w1x1 + w2x2 +... + w n x n

[0109] where: r1 is the correlation coefficient between the charging duration and the load characteristics, r2 is the correlation coefficient between the travel probability and the load characteristics, w1 and w2 are the corresponding weights respectively, n is the number of selected behavior sub-features, Y is the fused feature value, x1 is the feature value of the charging duration, and x2 is the feature value of the travel frequency.

[0110] S14. Construct a multi-scenario electric vehicle load model based on the first target vector sub-feature and the second target vector sub-feature, and optimize the constructed model using meta-learning to obtain an optimized model.

[0111] 1. Integrate the first target vector sub-feature and the second target vector sub-feature;

[0112] 2. Divide the usage scenarios of electric vehicles according to different dimensions;

[0113] 3. Select several target model architectures to construct an electric vehicle load model according to data characteristics and modeling purposes.

[0114] Divide multi-scenarios according to dimensions such as time, space, and user behavior. Here, taking the load modeling in two dimensions of time and space as an example, a modeling method combining spatio-temporal graph neural network (ST-GNN) and improved LSTM (Long Short-Term Memory Network) is adopted, as follows:

[0115] In ST-GNN, nodes represent charging piles, and edges represent the spatial relationship between charging piles. By combining the historical load data of electric vehicles and the spatial locations of charging piles, ST-GNN simulates the spatio-temporal fluctuations of the load, thereby improving the accuracy of load prediction. The update formula of the spatio-temporal graph neural network is:

[0116]

[0117] Among them, represents the feature of the l-th layer, N(i) is the adjacent node of node i, d i and d j are the degrees of node i and node j, σ is the activation function, γ t is the weight of the time feature, T t is the time embedding vector, indicating the influence of the time feature on the node. Through multiple layers of graph neural networks, ST-GNN can effectively handle the interaction of spatio-temporal data and capture the spatio-temporal dependence of electric vehicle loads.

[0118] In the spatio-temporal graph neural network, spatio-temporal features are fused through the adjacency matrix and the graph convolutional neural network (GCN), where time features and space features are learned simultaneously through different hierarchical structures. The model is trained using the gradient descent method, and the weight matrix in the spatio-temporal graph neural network is optimized by minimizing the loss function, enabling it to adapt to the changes in electric vehicle loads in different scenarios.

[0119] To handle the long-term dependence of load data, an improved LSTM network is adopted to effectively capture the long-term dependence in time series data through forget gates, input gates, and output gates. Assume the input sequence is X = {x1, x2, …, x T}, the output of the LSTM is h T , and the output after being weighted by the attention mechanism is:

[0120]

[0121] Among them, α t is the attention coefficient, e t is the attention score, h t and c t are the hidden state and cell state of the LSTM respectively. In this way, the LSTM can dynamically adjust the attention to different time periods according to the charging demand characteristics of each time step, especially for the peak and trough periods with large fluctuations in charging demand.

[0122] 4. Divide the electric vehicle load data in different scenarios into multiple tasks according to the preset meta - learning framework;

[0123] 5. Update the initial parameters of the model by methods such as gradient descent according to the training set data of multiple tasks

[0124] 6. Test the optimized model with new task data until the model performance meets the requirements.

[0125] Meta - learning optimization regards different scenarios as different tasks and uses task - level training to optimize the model parameters, enabling the model to quickly adapt to new scenarios with a small number of samples. Set a set of task sets T = {T1, T2, …, T m}, where each task T i represents a scenario and contains different input data and labels. The goal is to learn a shared parameter set θ by minimizing the loss function of each task, so that the model can quickly adapt to new tasks. The specific optimization process is as follows:

[0126]

[0127] Among them, L(T i ; θ) is the loss function on task T i , α is the learning rate, represents the gradient of the loss function with respect to the parameter θ, and θ * is the shared model parameter optimized by meta - learning.

[0128] Optimization is carried out through the gradient optimization algorithm in task - level training. Each task T i needs to calculate the gradient based on the existing data and update the model parameters in order to make better predictions in new tasks (i.e., new scenarios). During the training process, meta - learning first selects a batch of tasks T from the task set T i, and then train the model through an update method based on gradient descent. The specific gradient optimization process can be achieved through the following steps:

[0129] 1) Calculate the loss function L(T i ; θ) of the current task to obtain the gradient of the current task;

[0130] 2) Update the model parameters using the gradient descent method:

[0131]

[0132] where η is the learning rate, is the gradient of the loss function with respect to the parameter θ.

[0133] 3) Repeat the above steps until the model parameters are optimized for all tasks.

[0134] Through multiple rounds of training, the model can learn the prediction rules in different scenarios and quickly adjust its parameters in a new scenario for fast prediction. After training, various evaluation metrics are used to test the performance of the model, and further optimization is carried out according to the evaluation results. Here, the mean absolute percentage error (MAPE) and the root mean square error (RMSE) are adopted.

[0135] The calculation formula of MAPE is:

[0136]

[0137] where y i is the i-th actual value, is the i-th predicted value, and n is the total number of data. The smaller the MAPE value, the more accurate the prediction of the model.

[0138] The calculation formula of RMSE is:

[0139]

[0140] where y i and respectively represent the actual value and the predicted value, and n is the total number of data. The smaller the RMSE, the smaller the prediction error of the model.

[0141] After the model evaluation, if the results of the evaluation metrics are not satisfactory, the Bayesian optimization method is used to adjust the hyperparameters to further improve the prediction ability of the model. The basic process of Bayesian optimization is as follows:

[0142] 1) Select a surrogate model: Select a suitable surrogate model to describe the probability distribution of the objective function.

[0143] 2) Obtain sample data: Select some sample points in the parameter space and calculate the objective function values.

[0144] 3) Update the surrogate model: Update the surrogate model based on the known sample points and objective function values.

[0145] 4) Select the next query point: Select the next hyperparameter combination to query by obtaining the expected improvement value (EI, Expected Improvement) of the sampling points.

[0146] 5) Repeat the above steps until the best hyperparameter configuration is found.

[0147] The update formula for Bayesian optimization is:

[0148]

[0149] where \(x\) is a point in the hyperparameter space and \(EI(x)\) is the expected improvement value at that point.

[0150] Through Bayesian optimization, it is possible to efficiently search the hyperparameter space to find the optimal hyperparameter configuration, thereby further improving the prediction ability of the model.

[0151] The above embodiments provide effective data support for subsequent applications to actual scenarios, namely power grid regulation, charging pile planning, and user behavior guidance, by introducing meta-learning optimization, multi-scenario modeling of electric vehicle loads, spatio-temporal dynamic modeling, and comprehensively considering multiple factors such as power, time, space, and user behavior.

[0152] Embodiment 2

[0153] An embodiment of the global modeling system for electric vehicle load characteristics based on multiple scenarios in the present invention includes the following steps:

[0154] A data acquisition and preprocessing unit for acquiring electric vehicle load-related data based on multiple channels and preprocessing the acquired data;

[0155] An initial vector feature extraction unit for extracting initial vector features according to the preprocessed data, and the initial vector features include power features, time features, space features, and behavior features;

[0156] A first target vector sub-feature and second target vector sub-feature determination unit for determining the first target vector sub-feature according to power features, time features, and space features, and simultaneously determining the second target vector sub-feature according to behavior features;

[0157] The multi-scenario electric vehicle load model construction unit is used to construct a multi-scenario electric vehicle load model based on the first target vector sub-feature and the second target vector sub-feature, and optimize the constructed model using meta-learning to obtain an optimized model.

[0158] Those of ordinary skill in the art can realize that the units of each example described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the components of each example have been generally described according to their functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0159] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored, etc. In addition, the functional units in each embodiment of the present invention can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0160] If the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, read-only memories (ROMs), random access memories (RAMs), mobile hard disks, magnetic disks, or optical discs that can store program codes.

[0161] It can be understood that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the various embodiments of the present invention, and they should all be covered within the scope of the claims and the description of the present invention.

Claims

1. A global modeling method for electric vehicle load characteristics based on multi-scenario, characterized in that: include: Obtain electric vehicle load-related data based on multiple channels and pre-process the acquired data; Extracting initial vector features according to the preprocessed data, wherein the initial vector features include power features, time features, space features and behavior features; Determine a first target vector quantum feature according to the power feature, the time feature and the space feature, and determine a second target vector quantum feature according to the behavior feature; A multi-scenario electric vehicle load model is constructed based on the first target vector quantum features and the second target vector quantum features, and the constructed model is optimized by using meta-learning to obtain an optimized model.

2. The global modeling method of electric vehicle load characteristics based on multi-scenario according to claim 1 is characterized in that: The initial vector features are extracted according to the preprocessed data, and the initial vector features include power features, time features, space features and behavior features, including: Determine the power characteristics by calculating the average charging power, power fluctuation range, and power change rate; The fast Fourier transform algorithm and the autoregressive model are used to determine the periodic time characteristics and the short-term variation time characteristics respectively; The geographically weighted regression algorithm and adjacency matrix are combined to determine the changes in electric vehicle load in different geographical locations and the spatial relationship characteristics between regions; Analyze user behavior characteristics based on machine learning algorithms.

3. The global modeling method of electric vehicle load characteristics based on multi-scenario according to claim 1 is characterized in that: The step of determining the first target vector quantum feature according to the power feature, the time feature, and the space feature, and determining the second target vector quantum feature according to the behavior feature, comprises: Standardizing the data corresponding to the power characteristics, time characteristics and spatial characteristics; Determine the key features corresponding to the power feature, time feature and spatial feature using the trained random forest model; Feature fusion is performed according to key features corresponding to the power feature, time feature and space feature to obtain a fused first target vector feature.

4. The global modeling method of electric vehicle load characteristics based on multi-scenario according to claim 3 is characterized in that: The performing feature fusion according to the key features corresponding to the power feature, the time feature and the space feature to obtain the fused first target vector quantum feature includes: Where: w Pi 、w Tj and w Sl are the weight corresponding to the i-th power feature, the weight corresponding to the j-th time feature, and the weight corresponding to the l-th spatial feature, respectively. and are the i-th standardized power eigenvalue, the j-th standardized time eigenvalue, and the l-th standardized space eigenvalue, respectively. * 、T * and S * They are respectively the fused power feature information, time feature information and space feature information.

5. The global modeling method of electric vehicle load characteristics based on multi-scenario according to claim 1 is characterized in that: The step of determining the first target vector quantum feature according to the power feature, the time feature, and the space feature, and determining the second target vector quantum feature according to the behavior feature, comprises: Using Bayesian reasoning to determine charging behavior sub-characteristics and travel behavior sub-characteristics; Construct a correlation model between weather and charging behavior sub-characteristics and travel behavior sub-characteristics, analyze the relationship between traffic flow data and charging behavior sub-characteristics, travel behavior sub-characteristics and load characteristics, and analyze the changing patterns of charging behavior sub-characteristics, travel behavior sub-characteristics and load characteristics during holidays; The correlation coefficients between the charging behavior sub-features and the travel behavior sub-features and the load characteristics are screened, and the screened sub-features are linearly fused.

6. The global modeling method of electric vehicle load characteristics based on multi-scenario according to claim 5 is characterized in that: The method of using Bayesian reasoning to determine the charging behavior sub-characteristics and the travel behavior sub-characteristics includes: Among them, P(H|D) represents the posterior probability of charging behavior or travel behavior H given the data D, P(D|H) is the likelihood function, P(H) is the prior probability, and P(D) is the marginal likelihood of the data.

7. The global modeling method of electric vehicle load characteristics based on multi-scenario according to claim 5 is characterized in that: The screening of correlation coefficients between the charging behavior sub-features and the travel behavior sub-features and the load characteristics, and linearly fusing the screened sub-features, includes: Y=w1x1+w2x2+...+w n x n Among them: r1 is the correlation coefficient between charging time and load characteristics, r2 is the correlation coefficient between travel probability and load characteristics, w1 and w2 are the corresponding weights, n is the number of filtered behavioral sub-features, Y is the fused feature value, x1 is the charging time feature value, and x2 is the travel frequency feature value.

8. The global modeling method of electric vehicle load characteristics based on multi-scenario according to claim 1 is characterized in that: The multi-scenario electric vehicle load model is constructed based on the first target vector quantum feature and the second target vector quantum feature, and the constructed model is optimized by using meta-learning to obtain the optimized model, including: Integrating the first target vector quantum feature with the second target vector quantum feature; Divide the usage scenarios of electric vehicles according to different dimensions; According to the data characteristics and modeling purpose, several target model architectures are selected to build the electric vehicle load model.

9. The global modeling method of electric vehicle load characteristics based on multi-scenario according to claim 8 is characterized in that: The multi-scenario electric vehicle load model is constructed based on the first target vector quantum feature and the second target vector quantum feature, and the constructed model is optimized by using meta-learning to obtain the optimized model, including: The electric vehicle load data in different scenarios is divided into multiple tasks according to the preset meta-learning framework; Based on the training set data of multiple tasks, the initial parameters of the model are updated through methods such as gradient descent; Use new task data to test the optimized model until the model performance meets the requirements.

10. A global modeling system for electric vehicle load characteristics based on multiple scenarios, characterized in that: include: A data acquisition and preprocessing unit, used to acquire electric vehicle load-related data based on multiple channels and preprocess the acquired data; An initial vector feature extraction unit, used to extract initial vector features according to the preprocessed data, wherein the initial vector features include power features, time features, space features and behavior features; A first target vector quantum feature and a second target vector quantum feature determining unit, configured to determine the first target vector quantum feature according to the power feature, the time feature and the space feature, and to determine the second target vector quantum feature according to the behavior feature; A multi-scenario electric vehicle load model construction unit is used to construct a multi-scenario electric vehicle load model based on the first target vector quantum feature and the second target vector quantum feature, and use meta-learning to optimize the constructed model to obtain an optimized model.

Images