New energy automobile charging power grid load prediction method based on crowdsourcing feature selection genetic algorithm

Abstract

The invention discloses a new energy automobile charging power grid load prediction method based on a crowdsourcing feature selection genetic algorithm, and the method comprises the following steps: initializing crowdsourcing worker nodes, and distributing a crowdsourcing data collection budget for each worker node; performing data cleaning and feature construction on the collected new energy charging load historical data; executing a feature selection optimization task by each worker node, constructing an objective function of the feature selection optimization task, and performing parallel search and fitness evaluation on a feature subset based on a multi-objective genetic algorithm; outputting a local optimal feature solution set after iterative optimization of a multi-target genetic algorithm; aggregating the local optimal feature solution set to obtain a solution set of the crowdsourcing task for load prediction of the new energy charging station; and performing repeated detection on the solution set of the crowdsourcing task, removing repeated solutions, and performing non-dominated sorting on all reserved solution sets to obtain an optimal feature subset in the new energy charging station load prediction task. According to the invention, the precision and generalization ability of the new energy load prediction model can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid load forecasting, and in particular to a new energy vehicle charging power grid load forecasting method based on a crowdsourcing feature selection genetic algorithm. Background Art

[0002] As more new energy vehicle charging stations are connected to the power grid, they will generate power load while providing charging services. This load will significantly impact the stability and dispatchability of the regional distribution network. Therefore, accurately predicting the short-term charging load of new energy vehicles is crucial for building smart energy systems and improving grid efficiency. Previous studies have shown that the charging load of new energy vehicles is influenced by a variety of factors, such as drivers' daily routines and travel patterns, weather conditions, traffic congestion, holiday distribution, and electricity pricing mechanisms. In actual modeling, it is necessary to integrate these complex factors and fully exploit the vast amount of time-series data from charging stations, the power grid system, and the external environment. However, this data is often high-dimensional, redundant, and noisy. Directly using it for model training not only increases the computational burden but also may affect prediction performance.

[0003] Feature selection (FS) is a process of extracting the most representative feature subset for prediction or classification tasks by removing irrelevant or redundant features from the original data. Therefore, effective feature selection and extracting the most representative and predictive features for load forecasting based on the characteristics of renewable energy loads in different regions have become core steps in the predictive modeling process. Among existing feature selection methods, the filter method is based on statistical indicators and independently evaluates each feature before model training. It has high computational efficiency but cannot fully capture the interactive relationship between features. The wrapper method regards the feature selection process as a search problem. By coupling it with the machine learning model, it iteratively evaluates the impact of different feature subsets on model performance, but the computational overhead is relatively high.

[0004] In summary, with the development of new energy technologies and smart grids, accurate prediction of charging load has become a key issue in improving the efficiency and stability of distribution network scheduling. However, due to factors such as the strong distribution of data, high real-time requirements, obvious regionality, and complex feature dimensions, traditional centralized modeling methods are difficult to meet the needs of actual applications. Therefore, how to effectively select features, distribute feature selection tasks to different nodes, and combine local optimization feedback to improve overall computing efficiency while protecting data privacy, and support large-scale parallel evolutionary optimization to provide a scalable technical path for solving high-dimensional complex prediction problems has become a technical problem that needs to be solved urgently. Summary of the Invention

[0005] In order to overcome the defects and shortcomings of the existing technology, the present invention provides a new energy vehicle charging grid load forecasting method based on a crowdsourcing feature selection genetic algorithm. The present invention distributes the feature selection task to multiple worker nodes with local data and computing capabilities. Each worker node performs multi-objective genetic optimization based on local data to generate a non-dominated feature subset that balances the prediction error and feature compression rate. The server-side synchronously aggregates the solution sets uploaded by the worker nodes, and forms a global feature fingerprint library through non-dominated sorting and congestion distance calculation. In the deaggregation stage, a duplicate solution detection and target fitness correction mechanism are introduced to improve the consistency and robustness of the feature selection results, thereby significantly improving the accuracy and generalization ability of the new energy load forecasting model.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] The present invention provides a new energy vehicle charging grid load forecasting method based on a crowdsourcing feature selection genetic algorithm, comprising the following steps:

[0008] Initialize crowdsourcing worker nodes and allocate crowdsourcing data collection budget to each worker node;

[0009] Each worker node collects historical data on new energy charging loads to obtain a local data set;

[0010] Perform data cleaning and feature construction based on the collected historical data of new energy charging load;

[0011] Each worker node performs feature selection optimization tasks, constructs the objective function of the feature selection optimization task, and performs parallel search and fitness evaluation on feature subsets based on a multi-objective genetic algorithm;

[0012] After iterative optimization by multi-objective genetic algorithm, the local optimal feature solution set is output;

[0013] Aggregate the local optimal feature solution set to obtain the solution set for the crowdsourcing task of new energy charging station load forecasting;

[0014] The solution set of the crowdsourcing task is duplicated and removed, and all retained solution sets are non-dominated sorted to obtain the optimal feature subset in the new energy charging station load forecasting task.

[0015] As a preferred technical solution, the data cleaning includes missing value filling, outlier detection and elimination, and normalization and standardization processing.

[0016] As a preferred technical solution, the feature construction adopts a unified feature template and indicator format.

[0017] As a preferred technical solution, the objective function of the feature selection optimization task is constructed, which is specifically expressed as:

[0018] min F i (x)=(f i rat (x∣D i ),f i err (x∣D i ))

[0019]

[0020] st|D i |·c i ≤∈ i

[0021] Where x∈{0,1} p is the feature selection vector of individuals in the node population, p is the total number of features, c i is the collection budget consumed by workers to collect local raw data of the unit, f i rat (x|D i ) represents the feature selection rate, f i err (x|D i ) represents the mean absolute percentage error, which is used to measure the prediction error of the model under the corresponding feature subset, d i Represents the local dataset, x l represents the feature subset, P t is the actual load value at the t-th time node, is the predicted load value at the t-th time node, ∈ i represents the crowdsourcing data collection budget.

[0022] As a preferred technical solution, a multi-objective genetic algorithm is used to perform parallel search and fitness evaluation on feature subsets, specifically including:

[0023] Each worker node generates an initial population based on the local data set, and the initial population is initialized;

[0024] Execute a multi-objective genetic algorithm optimization iterative loop, specifically including:

[0025] Use non-dominated sorting to select non-dominated solutions from the current population as parents;

[0026] Perform single-point or multi-point crossover operations on parent individuals to generate multiple new offspring individuals;

[0027] Reverse the solution of the offspring individuals;

[0028] Update the fitness of the offspring population;

[0029] Merge the offspring population with the current population and select the best individual using non-dominated sorting;

[0030] After completing a round of local evolution operations, all candidate solutions in the current population are sorted non-dominated and the crowding distance is calculated to provide a better set of parent solutions for the next round of evolution and update the fitness of each individual in the objective function.

[0031] As a preferred technical solution, the initial population is initialized in a random manner based on Bernoulli distribution.

[0032] As a preferred technical solution, non-dominated sorting and crowding distance calculation are performed on all candidate solutions in the current population, specifically including:

[0033] All individuals in the population are non-dominated and graded according to their dual objective function values ​​to obtain several non-dominated grades, where the first grade is the Pareto frontier of the current round;

[0034] After the sorting is completed, the crowding degree of each individual in the target space is evaluated within each non-dominated level. For the j-th individual after sorting, the crowding distance calculation formula on the target m is:

[0035]

[0036] Among them, f i (m) (x) represents the evaluation value of the solution x on the target m, Represents two adjacent individuals, f i (m)max Indicates the maximum value of the target m in the current level, f i (m)min Indicates the minimum value of the target m within the current level;

[0037] The distances of each individual on the two targets are accumulated to obtain the total crowding distance of the individual.

[0038] The present invention also provides a new energy vehicle charging grid load forecasting system based on crowdsourcing feature selection genetic algorithm, which is used to implement the above-mentioned new energy vehicle charging grid load forecasting method based on crowdsourcing feature selection genetic algorithm, including: an initialization module, a data acquisition module, a data preprocessing module, a genetic optimization module, a local optimal feature solution set output module, an aggregation module, and an optimal feature subset output module;

[0039] The initialization module is used to initialize crowdsourcing worker nodes and allocate crowdsourcing data collection budget to each worker node;

[0040] The data acquisition module is used to collect historical data of new energy charging load based on each worker node to obtain a local data set;

[0041] The data preprocessing module is used to perform data cleaning and feature construction based on the collected new energy charging load historical data;

[0042] The genetic optimization module is used to perform feature selection optimization tasks based on each worker node, construct an objective function for the feature selection optimization task, and perform parallel search and fitness evaluation on feature subsets based on a multi-objective genetic algorithm;

[0043] The local optimal feature solution set output module is used to output the local optimal feature solution set after iterative optimization by a multi-objective genetic algorithm;

[0044] The aggregation module is used to aggregate the local optimal feature solution set to obtain the solution set of the crowdsourcing task of new energy charging station load forecasting;

[0045] The optimal feature subset output module is used to perform duplicate detection on the solution set of the crowdsourcing task, remove duplicate solutions, perform non-dominated sorting on all retained solution sets, and obtain the optimal feature subset in the new energy charging station load forecasting task.

[0046] The present invention also provides a computer-readable storage medium storing a program, which, when executed by a processor, implements the above-mentioned new energy vehicle charging grid load forecasting method based on crowdsourcing feature selection genetic algorithm.

[0047] The present invention also provides a computer device, comprising a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the above-mentioned new energy vehicle charging grid load prediction method based on crowdsourcing feature selection genetic algorithm is implemented.

[0048] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0049] (1) The present invention is based on the task distribution strategy of the crowdsourcing platform, and stores the collected new energy charging data in distributed nodes. This strategy supports sending feature selection tasks to multiple data collection nodes, and each node completes feature construction and optimization locally, thereby avoiding the centralized uploading of original data. While ensuring data privacy, it significantly reduces the network transmission burden, improves the scalability and flexible scheduling capabilities of the system, and can dynamically allocate crowdsourcing budgets according to the scale or complexity of different tasks.

[0050] (2) In view of the fact that edge nodes have local computing capabilities but limited resources, the present invention constructs a distributed worker node local parallel evolutionary optimization mechanism. Each worker node runs a multi-objective genetic algorithm locally to perform parallel search and fitness evaluation on feature subsets. This not only effectively utilizes the computing resources of edge nodes, but also improves the overall optimization efficiency and task response speed.

[0051] (3) Aiming at the trade-off between the number of features and the prediction accuracy in the new energy load forecasting, the present invention generates multiple sets of mutually non-dominated feature selection solutions based on a multi-objective genetic optimization strategy. The non-dominated solution set is used as a candidate solution, which not only provides a variety of options for the user or the system's subsequent model training, but also can build a feature fingerprint library to improve the system's generalization ability and robustness in different scenarios.

[0052] (4) In order to solve the problem of inconsistent and redundant solution distribution caused by the parallel optimization of multiple worker nodes in the crowdsourcing scenario, the present invention constructs a server-side collaborative aggregation and correction mechanism. This mechanism realizes the integration of global optimization solutions by synchronously collecting the non-dominated solutions obtained by each worker node in the local optimization and performing unified non-dominated sorting and congestion distance evaluation on the server side. At the same time, the duplicate solutions uploaded by multiple nodes are detected and identified, and their objective function values ​​are weightedly corrected or averaged, thereby improving the robustness and consistency of the global solution set. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flow chart of a method for predicting the load of a new energy vehicle charging grid based on a crowdsourcing feature selection genetic algorithm according to the present invention;

[0054] Figure 2 Schematic diagram of the feature selection scenario architecture for crowdsourcing new energy load forecasting in the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, 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 intended to limit the present invention.

[0056] Example 1

[0057] like Figure 1 As shown, this embodiment provides a new energy vehicle charging grid load forecasting method based on crowdsourcing feature selection genetic algorithm, which is applied to the feature selection system architecture of crowdsourcing new energy load forecasting, such as Figure 2 As shown in the figure, in this architecture, participating nodes mainly include two types of roles:

[0058] (1) The crowdsourcing server S is responsible for coordinating the scheduling, control, and result aggregation of the entire crowdsourcing task. The server undertakes core functions such as task dispatching, parameter configuration, monitoring the status of each participant, final solution integration, and feature fingerprint library construction. It is the coordination and integration center of the entire system.

[0059] (2) Crowdsourcing worker nodes, usually refers to edge computing nodes deployed at each new energy charging station. Each worker node has a certain data collection capability and local computing capability, can collect local charging load data in real time, and after receiving the task parameters issued by the server, independently complete local data cleaning, feature construction and distributed feature selection optimization. The server S and each worker node are connected through a set of stable network topology structures, and a star-shaped communication mechanism can be adopted. Network transmission only involves information such as intermediate model parameters or feature subset encoding, and does not involve original data, thereby ensuring the local privacy of each node data and the communication efficiency of the system.

[0060] The specific steps include:

[0061] S1: Initialize crowdsourcing workers: The server publishes the crowdsourcing task of feature selection in the new energy charging station network where it is located, and waits for the charging stations to participate and respond. These new energy charging stations participating in the optimization task are called crowdsourcing workers. The worker set is defined as Where N is the number of crowdsourcing workers; at the same time, each worker is assigned a crowdsourcing data collection budget ∈ i ,This budget determines the number of data samples collected by the worker;

[0062] S2: Workers collect local data separately: according to the crowdsourcing data collection budget ∈ i ,Each worker node independently collects historical load data related to the local new energy charging station according to its geographical location, including but not limited to charging power, load curve, timestamp, voltage, current, weather conditions, holiday marks and user charging behavior, etc. The local data set of worker i is recorded as D i , the sample size is |D i |.

[0063] S3: Local Data Cleaning and Feature Construction: After completing local data collection, each worker cleans and structures the raw data. This method allows for some heterogeneity among participating nodes due to factors such as charging station type, equipment configuration, and geographic environment. However, it stipulates that all nodes must follow a unified feature template when completing feature construction to ensure that they search within the same feature space. This feature template can be defined and distributed by a central server, ensuring that features constructed by different workers have a consistent search space.

[0064] In this embodiment, based on the task distribution strategy of the crowdsourcing platform, the collected new energy charging data is stored in distributed nodes. This strategy supports the distribution of feature selection tasks to multiple data collection nodes, and each node completes feature construction and optimization locally, thus avoiding the centralized upload of raw data. While ensuring data privacy, it significantly reduces the network transmission burden, improves the scalability and flexible scheduling capabilities of the system, and can dynamically allocate crowdsourcing budgets according to the scale or complexity of different tasks.

[0065] In this embodiment, to ensure the effectiveness of feature selection and prediction model training, all crowdsourced worker nodes need to perform systematic data cleaning on the locally collected historical data of new energy charging loads, including missing value filling, outlier detection and removal, and normalization. Specifically, the following steps are required:

[0066] (1) Missing value filling: During the load data collection process, data at certain time points may be missing due to sensor failure or communication interruption. For time series data with null values, the present invention uses linear interpolation to fill in the missing values. Let the load value P at time t be t Missing, and its adjacent known value is P t-1 and P t+1 , the following formula is used for interpolation estimation:

[0067]

[0068] (2) Outlier elimination: Considering the random volatility of load data and the possible sudden abnormal points in practice, this embodiment adopts a sliding window statistical detection method to identify and eliminate outliers. For each time point t, the window width is set to u, and the mean μ of the load values ​​at the previous u moments is calculated. t and standard deviation σ t , if the current value P t If the following conditions are met, the data will be identified as outliers and removed or replaced:

[0069] |P t -μ t |>γ·σ t

[0070] Among them, γ is the outlier identification threshold, which is generally in the range of 2 to 3. For moments judged to be abnormal, local averaging or linear interpolation can be used for repair:

[0071]

[0072] This method can effectively remove mutation error data while keeping the overall sequence trend undistorted;

[0073] (3) Normalization: Since meteorological characteristics (such as temperature, wind speed, humidity, etc.) and load characteristics have significant differences in numerical scale and physical units, if they are directly input into the model, it is easy to cause numerical imbalance problems, affecting training stability and convergence speed. Therefore, the minimum-maximum normalization (Min-Max Scaling) is uniformly used to pre-process all numerical features, setting the original value to q and the maximum value to q max , the minimum value is q min , then the normalized value q ′ Defined as:

[0074]

[0075] After normalization, all eigenvalues ​​are mapped to the interval [0, 1], thereby improving the consistency and convergence efficiency of model training;

[0076] In this embodiment, different charging stations exhibit significant heterogeneity in terms of device type, data granularity, sampling frequency, and environmental context. To enable subsequent collaborative feature selection and global feature evaluation, all worker nodes must locally construct features according to a unified feature template and indicator format, ensuring that all nodes operate within the same feature space. This unified feature template can be preset by the server or generated through a shared consensus mechanism within the system. This ensures structural comparability and semantic consistency of data across distributed nodes, providing a foundation for cross-site feature selection consensus and feature fingerprint library aggregation.

[0077] After data cleaning and feature construction, each worker’s dataset D i The feature set in is represented as: θ1,θ2,…,θ p , where p is the number of features;

[0078] S4: Workers perform local genetic optimization. Each worker node is based on the local dataset D i Perform independent feature selection optimization tasks, using genetic algorithms as local multi-objective optimization strategies.

[0079] In this embodiment, a distributed worker node local parallel evolutionary optimization mechanism is constructed to address the characteristics of edge nodes, which have local computing capabilities but limited resources. Each worker node runs a multi-objective genetic algorithm locally to perform parallel search and fitness evaluation on feature subsets. This not only effectively utilizes the computing resources of edge nodes, but also improves overall optimization efficiency and task response speed.

[0080] In this embodiment, the optimization objectives of the crowdsourcing feature selection task are constructed, specifically including:

[0081] For each worker node participating in the crowdsourcing task, its local feature selection process can be modeled as a multi-objective optimization problem with resource constraints. Each node has its own local dataset D i , on the premise of meeting the crowdsourcing data collection budget, two objective functions are optimized: on the one hand, the number of selected features is minimized; on the other hand, the classification error rate of the constructed model is minimized to improve the prediction performance.

[0082] Specifically, workers w i The optimization objective function is expressed as follows:

[0083] min F i (x)=(f i rat (x∣D i ),f i err (x∣D i ))

[0084] st|D i |·c i ≤∈ i

[0085] Where x∈{0,1} p is the feature selection vector of individuals in the node population, p is the total number of features, c i is the collection budget consumed by workers to collect local raw data of the unit. The two optimization objectives can be expressed as:

[0086]

[0087]

[0088] Respectively, the first objective is the feature selection rate, and the second objective is the mean absolute percentage error (MAPE), which is used to measure the prediction error of the model under this feature subset. t is the actual load value at the t-th time node, is the predicted value, NT is the number of predicted sample points involved in the evaluation, and the predicted value is determined by the worker based on the selected feature subset x i and local dataset D i The trained prediction model π i The generated ones are:

[0089]

[0090] Due to the differences in data environments, load characteristics, and resource conditions faced by different worker nodes, the local objective functions of each node are structurally heterogeneous. Therefore, a server-side aggregation mechanism is needed to integrate the optimization results of each node and further extract a representative feature subset with global consistency to improve the stability and generalization capabilities of the overall new energy load forecasting model.

[0091] First, worker i updates the optimization beacon Δ i =1, and generate an initial population Here, M is the population size of local solutions, and each individual It represents the selection status of the feature subset of the j-th solution of worker i during the t-round iteration process.

[0092] In this embodiment, each worker node w i Based on its local dataset D i , generating an initial population to start the local optimization process of the genetic algorithm. To enhance the diversity of the population and avoid early convergence to the local optimal solution, the initial population is initialized using a random method based on the Bernoulli distribution. Specifically, the initialization process satisfies the following Bernoulli distribution:

[0093]

[0094] Through this initialization method, the probability of each feature being selected in the initial solution is 0.5, which ensures the uniform distribution of feature subsets in the solution space and the structural diversity of the initial population.

[0095] Then, for each individual, a machine learning prediction model π is trained using the local dataset i , evaluate the local population and update the fitness of each individual on the two objectives, recorded as They correspond to the feature rate of feature subset selection and the error rate on the classification model. Afterwards, each worker performs a local genetic algorithm optimization iteration cycle. The optimization process in the tth round of the cycle is as follows:

[0096] 1. Select the parent: Use non-dominated sorting to select the parent from the current population Select the non-dominated solution as the parent;

[0097] 2. Crossover: Perform single-point or multi-point crossover operations on the parent individuals to generate M new offspring individuals;

[0098] 3. Mutation: Reverse the solution of the offspring individual with a certain probability;

[0099] 4. Evaluation: Using Machine Learning to Predict Model π i Update the fitness of the offspring population;

[0100] 5. Update: The offspring population With the current population Merge and use non-dominated sorting to select the best N individuals to enter the next round of evolution;

[0101] In this embodiment, after completing a round of local evolution operations, the worker node needs to perform non-dominated sorting and crowding distance calculation on all candidate solutions in the current population to screen out individuals with good performance and distribution characteristics, and provide a high-quality parent solution set for the next round of evolution. This embodiment uses a method improved from NSGA-II to sort and evaluate the distance of the population, which is suitable for dual-objective feature selection tasks. First, all individuals in the population are non-dominated and graded according to their dual-objective function values ​​to obtain several non-dominated levels, where the first level is the Pareto front of the current round. After the sorting is completed, within each non-dominated level, the degree of crowding of each individual in the target space needs to be further evaluated to retain diversity. Specifically, for the jth individual after sorting, its crowding distance calculation formula on the target m is as follows:

[0102]

[0103] Among them, f i (m) (x) represents the evaluation value of solution x on target m. The denominator is the difference between the maximum and minimum values ​​of the target in the current level, which is used for normalization. Subsequently, the distance of each individual on the two targets is accumulated to obtain the total crowding distance of the individual:

[0104]

[0105] In the non-dominated solution sorting, when selecting the next generation of individuals, if multiple solutions are at the same non-dominated level, individuals with larger crowding distance values ​​are retained first, thereby effectively improving the distribution diversity of the solution set.

[0106] The local optimization process is iterated independently on each worker node for T rounds, each round is denoted as t = 1, 2, ..., T; when the non-dominated frontier solution set obtained after three consecutive evolutionary iterations If no update occurs, the population is considered to have converged to a stable Pareto frontier, and the worker-local genetic optimization process is terminated.

[0107] In this embodiment, in order to balance the number of features and prediction accuracy in the new energy load forecast, multiple sets of mutually non-dominated feature selection solutions are generated based on a multi-objective genetic optimization strategy. This non-dominated solution set serves as a candidate solution, which not only provides diverse options for subsequent model training for users or systems, but also builds a feature fingerprint library to improve the system's generalization ability and robustness in different scenarios.

[0108] S5: The worker uploads the local optimal feature solution set to the server. When the local genetic optimization process is completed, the worker Together with the fitness corresponding to the solution, it is uploaded to the server and the optimization beacon is updated to Δ i =0;

[0109] S6: The server aggregates the feature solution sets uploaded by the workers to form the solution set of the crowdsourcing task of new energy charging station load forecasting in this network. The server continuously monitors the optimization status of each worker. When it detects that all crowdsourcing workers have When the optimization beacons of are all set to 0, it indicates that the optimization process of the worker side is over and the server enters the solution aggregation phase. First, the crowdsourcing server will upload the solution of each worker. Perform duplicate detection. If the individual nodes uploaded by multiple workers are exactly the same in the selection of features, they are considered duplicate solutions. For the processing of duplicate solutions, the server will calculate the fitness value f of the corresponding classification error rate. 2 Perform the average correction as follows:

[0110]

[0111] in, represents the fitness value of the adjusted repeated solution on the second target, K represents the number of repeated solutions in the solution set uploaded by the worker, It represents the fitness of the solution uploaded on the kth node optimized on the second objective of feature selection;

[0112] Based on the second goal, the server adjusts the fitness of all uploaded duplicate solutions and removes duplicate solutions. After completing this step, the server performs non-dominated sorting on all retained solution sets to form the Pareto front. This sorting process uses the feature subset dimension selection rate and model prediction error as two optimization goals to select the final non-dominated solution set. As the optimal solution set for the current crowdsourcing task. The server stores the fingerprint as part of the feature fingerprint library. This fingerprint library records the optimal feature subsets in the load forecasting task of new energy charging stations in different regions of the current network, which can be used as prior information for subsequent modeling.

[0113] In this embodiment, a server-side collaborative aggregation and correction mechanism is constructed to address the inconsistent and redundant solution distribution problems caused by the parallel optimization of multiple worker nodes in the crowdsourcing scenario. This mechanism synchronously collects the non-dominated solutions obtained by each worker node in the local optimization, and performs unified non-dominated sorting and congestion distance evaluation on the server side to achieve the integration of the global optimization solution. At the same time, the duplicate solutions uploaded by multiple nodes are detected and identified, and their objective function values ​​are weightedly corrected or averaged, which improves the robustness and consistency of the global solution set and promotes the system to form a stable optimization consensus in a distributed environment.

[0114] Example 2

[0115] This embodiment provides a new energy vehicle charging grid load forecasting system based on a crowdsourcing feature selection genetic algorithm, which is used to implement the new energy vehicle charging grid load forecasting method based on a crowdsourcing feature selection genetic algorithm in Example 1. The system includes: an initialization module, a data acquisition module, a data preprocessing module, a genetic optimization module, a local optimal feature solution set output module, an aggregation module, and an optimal feature subset output module;

[0116] In this embodiment, the initialization module is used to initialize the crowdsourcing worker nodes and allocate a crowdsourcing data collection budget to each worker node;

[0117] In this embodiment, the data collection module is used to collect historical data of new energy charging load based on each worker node to obtain a local data set;

[0118] In this embodiment, the data preprocessing module is used to perform data cleaning and feature construction based on the collected new energy charging load historical data;

[0119] In this embodiment, the genetic optimization module is used to perform feature selection optimization tasks based on each worker node, construct an objective function for the feature selection optimization task, and perform parallel search and fitness evaluation on feature subsets based on a multi-objective genetic algorithm;

[0120] In this embodiment, the local optimal feature solution set output module is used to output the local optimal feature solution set after iterative optimization by a multi-objective genetic algorithm;

[0121] In this embodiment, the aggregation module is used to aggregate the local optimal feature solution set to obtain the solution set of the crowdsourcing task of new energy charging station load forecasting;

[0122] In this embodiment, the optimal feature subset output module is used to perform duplicate detection on the solution set of the crowdsourcing task, remove duplicate solutions, perform non-dominated sorting on all retained solution sets, and obtain the optimal feature subset in the new energy charging station load forecasting task.

[0123] Example 3

[0124] This embodiment provides a storage medium, which can be a storage medium such as ROM, RAM, disk, or CD. The storage medium stores one or more programs. When the program is executed by the processor, the new energy vehicle charging grid load prediction method based on crowdsourcing feature selection genetic algorithm of Example 1 is implemented.

[0125] Example 4

[0126] This embodiment provides a computing device, which can be a desktop computer, a laptop computer, a smart phone, a PDA handheld terminal, a tablet computer or other terminal device with a display function. The computing device includes a processor and a memory, and the memory stores one or more programs. When the processor executes the program stored in the memory, the new energy vehicle charging grid load forecasting method based on the crowdsourcing feature selection genetic algorithm of Example 1 is implemented.

[0127] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A new energy vehicle charging grid load forecasting method based on crowdsourcing feature selection genetic algorithm, characterized in that: The steps include: Initialize crowdsourcing worker nodes and allocate crowdsourcing data collection budget to each worker node; Each worker node collects historical data on new energy charging loads to obtain a local data set; Perform data cleaning and feature construction based on the collected historical data of new energy charging load; Each worker node performs feature selection optimization tasks, constructs the objective function of the feature selection optimization task, and performs parallel search and fitness evaluation on feature subsets based on a multi-objective genetic algorithm; After iterative optimization by multi-objective genetic algorithm, the local optimal feature solution set is output; Aggregate the local optimal feature solution set to obtain the solution set for the crowdsourcing task of new energy charging station load forecasting; The solution set of the crowdsourcing task is duplicated and removed, and all retained solution sets are non-dominated sorted to obtain the optimal feature subset in the new energy charging station load forecasting task.

2. The new energy vehicle charging grid load forecasting method based on crowdsourcing feature selection genetic algorithm according to claim 1 is characterized in that: The data cleaning includes missing value filling, outlier detection and removal, and normalization processing.

3. The new energy vehicle charging grid load forecasting method based on crowdsourcing feature selection genetic algorithm according to claim 1 is characterized in that: The feature construction adopts a unified feature template and indicator format.

4. The new energy vehicle charging grid load forecasting method based on crowdsourcing feature selection genetic algorithm according to claim 1 is characterized in that: Construct the objective function of the feature selection optimization task, which is specifically expressed as: min F i (x)=(f i rat (x∣D i ),f i err (x∣D i )) s.t.|D i |·c i ≤∈ i Where s∈{0,1} p is the feature selection vector of individuals in the node population, p is the total number of features, c i is the collection budget consumed by workers to collect local raw data of the unit, f i rat (x|D i ) represents the feature selection rate, f i err (x|D i ) represents the mean absolute percentage error, which is used to measure the prediction error of the model under the corresponding feature subset, D i Represents the local dataset, x l represents the feature subset, P t is the actual load value at the t-th time node, is the predicted load value at the t-th time node, ∈ i represents the crowdsourcing data collection budget.

5. The new energy vehicle charging grid load forecasting method based on crowdsourcing feature selection genetic algorithm according to claim 1 is characterized in that: Parallel search and fitness evaluation of feature subsets based on multi-objective genetic algorithm, including: Each worker node generates an initial population based on the local data set, and the initial population is initialized; Execute a multi-objective genetic algorithm optimization iterative loop, specifically including: Use non-dominated sorting to select non-dominated solutions from the current population as parents; Perform single-point or multi-point crossover operations on parent individuals to generate multiple new offspring individuals; Reverse the solution of the offspring individuals; Update the fitness of the offspring population; Merge the offspring population with the current population and select the best individual using non-dominated sorting; After completing a round of local evolution operations, all candidate solutions in the current population are sorted non-dominated and the crowding distance is calculated to provide a better set of parent solutions for the next round of evolution and update the fitness of each individual in the objective function.

6. The new energy vehicle charging grid load forecasting method based on crowdsourcing feature selection genetic algorithm according to claim 5 is characterized in that: The initial population is initialized randomly using Bernoulli distribution.

7. The method for predicting the load of a new energy vehicle charging grid based on a crowdsourcing feature selection genetic algorithm according to claim 5, characterized in that: Perform non-dominated sorting and crowding distance calculation on all candidate solutions in the current population, including: All individuals in the population are non-dominated and graded according to their dual objective function values ​​to obtain several non-dominated grades, where the first grade is the Pareto frontier of the current round; After the sorting is completed, the crowding degree of each individual in the target space is evaluated within each non-dominated level. For the j-th individual after sorting, the crowding distance calculation formula on the target m is: Among them, f i (m) (x) represents the evaluation value of the solution x on the target m, Represents two adjacent individuals, f i (m)max Indicates the maximum value of the target m in the current level, f i (m)min Indicates the minimum value of target m within the current level; The distances of each individual on the two targets are accumulated to obtain the total crowding distance of the individual.

8. A new energy vehicle charging grid load forecasting system based on crowdsourcing feature selection genetic algorithm, characterized in that: A method for predicting the load of a new energy vehicle charging grid based on a crowdsourced feature selection genetic algorithm according to any one of claims 1 to 7, comprising: an initialization module, a data acquisition module, a data preprocessing module, a genetic optimization module, a local optimal feature solution set output module, an aggregation module, and an optimal feature subset output module; The initialization module is used to initialize crowdsourcing worker nodes and allocate crowdsourcing data collection budget to each worker node; The data acquisition module is used to collect historical data of new energy charging load based on each worker node to obtain a local data set; The data preprocessing module is used to perform data cleaning and feature construction based on the collected new energy charging load historical data; The genetic optimization module is used to perform feature selection optimization tasks based on each worker node, construct an objective function for the feature selection optimization task, and perform parallel search and fitness evaluation on feature subsets based on a multi-objective genetic algorithm; The local optimal feature solution set output module is used to output the local optimal feature solution set after iterative optimization by a multi-objective genetic algorithm; The aggregation module is used to aggregate the local optimal feature solution set to obtain the solution set of the crowdsourcing task of new energy charging station load forecasting; The optimal feature subset output module is used to perform duplicate detection on the solution set of the crowdsourcing task, remove duplicate solutions, perform non-dominated sorting on all retained solution sets, and obtain the optimal feature subset in the new energy charging station load forecasting task.

9. A computer-readable storage medium storing a program, characterized in that: When the program is executed by a processor, the method for predicting the load of a new energy vehicle charging grid based on a crowdsourcing feature selection genetic algorithm as described in any one of claims 1 to 7 is implemented.

10. A computer device comprising a processor and a memory for storing a program executable by the processor, characterized in that: When the processor executes the program stored in the memory, it implements the new energy vehicle charging grid load forecasting method based on crowdsourcing feature selection genetic algorithm as described in any one of claims 1 to 7.

Images