Method for predicting potential of electric vehicle demand response for large-scale events in park

Abstract

The application discloses a park large-scale activity electric vehicle demand response potential prediction method. It is related to the field of electric power system and intelligent transportation. The application collects charging behavior, built environment and activity characteristic data under the park large-scale activity, cleans and integrates into a fusion data set with time and space tags, divides the time and space scene and clusters to identify the charging behavior mode, combines XGBoost and TCN models to extract static and time sequence feature weights respectively, uses a dynamic fusion strategy to generate a charging power prediction curve, calculates the difference value with the baseline load to quantify the demand response potential. The application fully considers the characteristics of large-scale activities, collects relevant data, and proposes a fusion method for multi-source heterogeneous data, so as to combine the user energy use behavior characteristics and the activity characteristics, and more accurately predict the demand potential of the park.

Description

Technical Field

[0001] This invention relates to the fields of power systems and intelligent transportation, and in particular to a method and system for predicting the demand response potential of electric vehicles for large-scale events in industrial parks. Background Technology

[0002] During large-scale events held in the park, the demand response behavior of electric vehicle users has obvious spatiotemporal aggregation, intermittent load concentration, strong influence of event attributes on user behavior, and relatively fixed response time; while the park has high requirements for the reliability of the power system. The existing related patent technical solutions generally adopt a single research method, including: (1) constructing a database based on historical charging data (charging time period, power, user behavior), analyzing user behavior characteristics through K-means clustering or travel chain model to extract charging peak / valley patterns, and generating user samples with the help of Monte Carlo simulation. However, such methods are insufficient in modeling the coupling mechanism between exogenous variables such as the built environment and response behavior, resulting in limited model generalization ability; (2) using time series models such as LSTM and BP neural network to predict the base load curve, and quantifying the impact of user behavior on load regulation (such as delayed charging and response price strategy) through membership function, genetic algorithm or road network-power grid coupling model. In general, existing technologies typically compare conventional prediction curves with adjustment curves after user response to calculate adjustable potential capacity (such as peak shaving and valley filling space). Common technologies include scenario-based modeling and data-driven behavioral clustering. The core objective is to improve the scientific nature of power grid dispatching decisions and the reliability of response potential predictions. Common technological bottlenecks manifest in three aspects: lack of multi-timescale coordination (disconnect between minute-level control and annual-level planning), inefficient cross-domain data fusion, and inaccurate spatiotemporal scenario deconstruction. Regarding data, existing technologies typically use a single data source (such as historical load data of charging piles) or static scenario segmentation (such as classification only by weekdays / holidays). Regarding models, existing technologies typically predict charging load through linear regression or a single machine learning model (such as LSTM), resulting in poor model generalization. Regarding model input features, existing technologies do not incorporate park activity features into the model. Specifically: (1) Inaccurate recognition of spatiotemporal charging scenarios: Existing technologies lack the ability to fuse multi-source heterogeneous data and refine spatiotemporal scenario deconstruction, resulting in low accuracy of spatiotemporal scenario recognition. The models lack the ability to generalize to different spatiotemporal scenarios (such as before / after events, holidays / weekdays, and different geographical regions), and there is a risk of overfitting. (2) Response demand potential prediction models cannot simultaneously handle static and dynamic data: Spatiotemporal modeling is fragmented: Existing technologies usually use a single model (such as LSTM or traditional regression models), resulting in fragmented spatiotemporal data processing dimensions, making it difficult to simultaneously handle single scenario subset data (such as event features, response electricity prices, etc.) and high-dimensional time series data. At the same time, the prediction granularity is coarse and lacks a multi-scale feature extraction mechanism. Poor dynamic adaptability: Existing technologies usually use static hyperparameter optimization methods and lack an event scenario-driven prediction framework, resulting in low accuracy of the prediction model. (3) The input features of the existing model do not fully consider the characteristics of the park: usually only the historical load data of charging piles are considered, and multi-source information such as the built environment and park characteristics are not integrated, resulting in the failure of demand response prediction to associate the demand response behavior of electric vehicle users with the characteristics of the park, and lack of prediction of the demand response potential of the park. Summary of the Invention

[0003] This invention proposes a method for predicting the demand response potential of electric vehicles during large-scale events in a park. By finely identifying the energy consumption behavior patterns of electric vehicle users around the park and combining static features with dynamic time-series features for joint modeling, it achieves high-precision prediction of the dispatchable load potential of electric vehicles under different spatiotemporal scenarios. Specifically, it predicts the demand response potential of clustered electric vehicles generated under the scenario of large-scale events in the park, thus meeting the park's requirements for high reliability of the power system.

[0004] It is understood that the parks in this invention include, but are not limited to, industrial parks, cultural and tourism parks, and urban complex parks, referring to specific areas with independent geographical boundaries, supporting power systems, and the capacity to host large-scale events.

[0005] Another objective of this invention is to propose a system for predicting the demand response potential of electric vehicles for large-scale events in industrial parks.

[0006] To achieve the above objectives, this invention proposes a method for predicting the demand response potential of electric vehicles for large-scale events in industrial parks, comprising: A fused dataset is obtained by preprocessing multi-source heterogeneous data on electric vehicle energy consumption behavior during large-scale events held in the park. The fused dataset is divided into subsets with spatiotemporal feature labels according to time and space dimensions. Then, a clustering algorithm is used to analyze the subsets to identify data subsets with multiple spatiotemporal dimension cross combinations. The spatiotemporal feature labels include activity cycle rules, time period classification rules, spatial region division rules, and date type. An integrated prediction model is constructed and trained, including an XGBoost model, a TCN model, and a dynamic fusion module. A subset of data with cross-combined spatiotemporal dimensions is input into the XGBoost model to output static feature weights, and the temporal feature data in the subset of data with cross-combined spatiotemporal dimensions is input into the TCN model to output temporal feature weights. The dynamic fusion module is used to fuse the static feature weights and the temporal feature weights according to a preset dynamic adjustable ratio to predict the final demand response power curve and optimize the integrated prediction model. The baseline load curve is calculated based on historical data of inactive days corresponding to the data subsets with cross-combination of spatiotemporal dimensions. Real-time spatiotemporal scenario subset data is input into the optimized integrated prediction model to output the predicted load curve. The difference between the predicted load curve and the baseline load curve is calculated to quantify the demand response potential under different spatiotemporal scenario subsets.

[0007] The electric vehicle demand response potential prediction method for large-scale events in industrial parks according to embodiments of the present invention may also have the following additional technical features: In one embodiment of the present invention, the multi-source heterogeneous data includes charging behavior data, built environment data, and activity characteristic data; wherein, the charging behavior data includes battery state of charge, charging start time and end time, charging duration, real-time charging power, charging station location coordinates, battery state of charge start value and end value, time period marker, and user ID; the built environment data includes POI data and road network feature data; the activity characteristic data includes park data, activity type, activity time, activity duration, and number of participants.

[0008] In one embodiment of the present invention, multi-source heterogeneous data on electric vehicle energy consumption behavior during large-scale events in a park are preprocessed to obtain a fused dataset, including: Define activity time Matching charging behavior time By judging conditions The charging behavior data is correlated with the activity characteristic data to obtain the first correlation result; GeoHash is used for unified encoding, and the charging behavior data is associated with the built environment data and activity characteristic data according to spatial proximity by calculating the same or adjacent GeoHash codes to obtain the second association result. We uniformly adopted Beijing time and used Pandas to align the time series in minutes to obtain the time series alignment results. Based on the first association result, the second association result, and the time series alignment result, charging behavior data, built environment data, and activity characteristic data are integrated to construct a fused dataset.

[0009] In one embodiment of the present invention, a clustering algorithm is used to analyze the partitioned data subsets to identify data subsets with multiple spatiotemporal dimensions that are cross-combined, including: Calculate the statistical characteristics of each variable in each combined data set based on multi-source heterogeneous data, and then perform Z-Score standardization on the data of each variable: , and The first The first combination of data The mean and standard deviation of each variable, As input variables for clustering; Set neighborhood radius and minimum points And use Euclidean distance to measure the similarity between samples:

[0010] Iterate through all combinations of data for the sample set. any sample point Constructing based on sample points Centered on, with The set of neighborhood points of the neighborhood radius And determine the number of sample points contained in the neighborhood point set. Does it meet the requirements? If satisfied, then the sample points Taking this core point as the core, and starting from this core point, expanding to sample points with achievable density to form corresponding clusters; After clustering is completed, the silhouette coefficient is calculated for all combinations. As a standard for evaluating the practical utility and rationality of clustering structures, the total number of valid sample points participating in the silhouette coefficient calculation is... If the average profile coefficient Enter the training phase; otherwise, parameters need to be adjusted. and Re-clustering; After training, several spatiotemporal scene clusters are finally output, which are the identified spatiotemporal scene subsets. The subsets are then labeled according to their characteristics.

[0011] In one embodiment of the present invention, inputting a subset of data with cross-combined spatiotemporal dimensions into the static feature weights output by the XGBoost model includes: Data subsets that combine spatiotemporal dimensions Preprocessing is performed to remove outliers; Based on the preprocessed data, and according to the XGBoost principle, a regression tree is iteratively constructed to fit the data. The objective function is:

[0012] in, For predicted values Compared with the true value The loss function; For the number of trees, For the first Tree; , is a regularization term that controls model complexity. The number of leaf nodes. The weights of the leaf nodes are... and For regularization parameters; Based on the number of splits and gain metrics of the features in the tree model, obtain the feature weights: For N Each subset of scenes corresponds to a certain number of scenes. K 3D features, output weight matrix:

[0013] in, Indicates the first n The first scene subset k Weights of dimensional features.

[0014] In one embodiment of the present invention, the temporal feature data of a data subset with cross-combined spatiotemporal dimensions is input into the TCN model to output temporal feature weights, including: For time series feature data Perform preprocessing and remove outliers; N Each scene subset, Include L Dimensional time series features; set up For the input time series feature sequence at time step The value of , , The time series length is specified; TCN captures long-term dependencies through multiple layers of causal dilated convolution, i.e., for the th... Layer convolution performs the following operations: ;in, P The kernel size is [size]. For the first The expansion rate of the layer, The first convolution kernel Each weight, It is the ReLU activation function; By stacking multiple layers of causal dilated convolutional layers, dependencies at different time scales are captured: ;in, , For the final hidden layer output, and For the first Layer weights and biases; For each scene subset n Temporal feature weight vectors are generated through global average pooling and fully connected layers, i.e.: Indicates the first The temporal feature aggregation vector obtained by global average pooling of a subset of scenes is then processed. Perform feature aggregation; T The length of the time series; Indicates the first Temporal feature weight vectors corresponding to each scene subset This represents the bias vector during the weight mapping process. This represents the fully connected weight mapping matrix, used to aggregate temporal features into vectors. Mapped to weight scores corresponding to each time-series feature, through Perform weight mapping. Ensure the sum of the weights is 1; Combine the weight vectors of all scene subsets into a matrix. ,in For the first n The temporal feature weight vector of each scene subset, with dimension [ ]. L .

[0015] In one embodiment of the present invention, static feature weights are... With time series feature weights Fusion is performed using a dynamically adjustable ratio to predict the final demand response power curve; Fusion formula:

[0016] in, Adaptive weights for scene subsets ( The Sigmoid activation function introduces non-linear characteristics.

[0017] In one embodiment of the present invention, training two sub-models, XGBoost and TCN, includes: To minimize the performance of each sub-model in the corresponding scene subset With RMSE as the target, each sub-model is trained independently:

[0018]

[0019] in, Scene subset s The number of samples in the sample; Represents a subset of scenes s The Middle i The actual observed values ​​of each sample; and These represent the hyperparameters corresponding to the XGBoost and TCN models, respectively. and The predicted value; The hyperparameters used in training include: learning rate Maximum tree depth Subsample ratio ; The hyperparameters used in training include: kernel size. Expansion rate Network layers Dropout rate ( ); The model performance was validated and evaluated for each hyperparameter combination using five-fold cross-validation, including: Determine the hyperparameter spaces of XGBoost and TCN models; Using a Gaussian process as a surrogate model, the following operations are performed in each round of Bayesian optimization for both XGBoost and TCN models: Based on the current surrogate model, determine the next optimal combination of hyperparameters in the hyperparameter space; calculate the mean RMSE of the model performance using five-fold cross-validation as the objective function value; update the surrogate model and iterate until the preset number of optimizations or convergence; the five-fold cross-validation formula is:

[0020] in, Indicates the first The number of samples in folded cross-validation; Finally, the optimal hyperparameter combination for each sub-model is obtained. and .

[0021] In one embodiment of the present invention, based on the optimal hyperparameter combination and The objective function for optimizing the fusion layer is:

[0022]

[0023] in, Represents a subset of scenes The corresponding dynamic adaptive fusion weights, and , For scene subset The optimal solution for the lower dynamic fusion layer is obtained through objective function optimization; By minimizing Training Dynamic adaptive weights in ,in, For the first Each sample inputs a static feature vector to XGBoost. For the first The temporal feature sequence of each sample input TCN, and These are the XGBoost model and the TCN model, respectively.

[0024] In one embodiment of the present invention, a baseline load curve is calculated based on historical data of inactive days corresponding to a data subset with cross-combined spatiotemporal dimensions. Real-time spatiotemporal scenario subset data is then input into an optimized integrated prediction model to output a predicted load curve. The difference between the predicted load curve and the baseline load curve is calculated to quantify the demand response potential under different spatiotemporal scenario subsets, including: Calculate a specific scene subset s The baseline load curve :

[0025] in, This is a collection of historical load data for days with no activity or lowest load days. This represents the time series load data for day d in the set, representing a time series point. t=1, …,T ; Calculate the optimal adaptive weights for all sub-scenes Then, calculate and obtain Static feature weights in With time series feature weights That is, to obtain the optimized XGBoost-TCN ensemble model, call the optimized XGBoost-TCN ensemble model, and input... and Data, output predicted load curve ; Calculate the difference between the predicted load curve and the baseline load curve. Quantify the demand response potential, quantify the demand response potential under different spatiotemporal scenario subsets, and store them according to spatiotemporal scenario subsets.

[0026] To achieve the above objectives, another aspect of the present invention proposes a system for predicting the demand response potential of electric vehicles for large-scale events in industrial parks, comprising: The multi-source heterogeneous data mining module is used to preprocess multi-source heterogeneous data on the energy consumption behavior of electric vehicles during large-scale events held in the park to obtain a fused dataset. The spatiotemporal scene deconstruction and recognition module is used to divide the fused dataset according to time and space dimensions to obtain data subsets with spatiotemporal feature labels, and to use clustering algorithms to analyze the divided data subsets to identify data subsets with multiple spatiotemporal dimensions; wherein, the spatiotemporal feature labels include activity cycle rules, time period classification rules, spatial region division rules and date type; An integrated model building module for different spatiotemporal scenarios is used to build and train an integrated prediction model, including an XGBoost model, a TCN model, and a dynamic fusion module. The module inputs a subset of data with cross-combined spatiotemporal dimensions into the XGBoost model to output static feature weights, and inputs the temporal feature data from the subset of data with cross-combined spatiotemporal dimensions into the TCN model to output temporal feature weights. The dynamic fusion module uses a preset dynamic adjustable ratio to fuse the static feature weights and temporal feature weights to predict the final demand response power curve and optimize the integrated prediction model. The demand response potential prediction module is used to calculate the baseline load curve based on historical data of inactive days corresponding to data subsets with cross-combination of spatiotemporal dimensions. It also inputs real-time spatiotemporal scenario subset data into the optimized integrated prediction model to output the predicted load curve and calculates the difference between the predicted load curve and the baseline load curve to quantify the demand response potential under different spatiotemporal scenario subsets.

[0027] The electric vehicle demand response potential prediction method and system for large-scale events in industrial parks, as described in this invention, refines the spatiotemporal scenario to improve the accuracy of the prediction system. It integrates static and dynamic data within the same potential prediction system framework and proposes a dynamic model training strategy, thereby providing a prediction model with greater generalization capabilities. By fully considering the characteristics of large-scale events and collecting relevant data, a multi-source heterogeneous data fusion method is proposed, combining user energy consumption behavior characteristics with event characteristics, enabling more accurate prediction of the cluster electric vehicle demand response potential generated under the scenario of large-scale events in the industrial park.

[0028] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0029] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a method for predicting the electric vehicle demand response potential for large-scale events in a park, according to an embodiment of the present invention. Figure 2 This is a framework diagram of a method for predicting the demand response potential of electric vehicles for large-scale events in a park, according to an embodiment of the present invention. Figure 3 This is a structural diagram of an electric vehicle demand response potential prediction system for large-scale events in a park, according to an embodiment of the present invention. Detailed Implementation

[0030] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0031] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0032] The following describes, with reference to the accompanying drawings, a method and system for predicting the demand response potential of electric vehicles for large-scale events in industrial parks, according to embodiments of the present invention.

[0033] Figure 1 This is a flowchart of a method for predicting the electric vehicle demand response potential for large-scale events in a park, according to an embodiment of the present invention. Figure 1 and Figure 2 As shown, it includes: S1 is a fused dataset obtained by preprocessing multi-source heterogeneous data on the energy consumption behavior of electric vehicles during large-scale events held in the park.

[0034] Firstly, this embodiment of the invention collects data when a large-scale event is held in the park, including park and event data, reflecting the characteristics of large-scale events; charging behavior data, reflecting energy consumption behavior characteristics; and built environment data, etc.

[0035] The charging behavior data includes the following data collected from the public charging pile operation platform: battery state of charge (SOC), charging start and end times (accurate to the minute), charging duration (minutes), real-time charging power (kW), charging station location coordinates (latitude and longitude), battery state of charge (SOC) start value (%) and end value (%), time period markers (peak hours: 17:00-22:00, off-peak hours: other times), and user ID.

[0036] Among them, the built environment data includes: using the Gaode Map API to extract POI data within 3 kilometers of the park, including building types (commercial, residential, public facilities, etc.), the number and type of charging facilities, the number and capacity of parking lots, etc., dividing the space into spatial units according to a 500m×500m grid, and using the spatial raster technology (500m×500m grid) in ArcGIS software to divide the spatial units; and extracting the road network characteristics in the area, including the number of lanes, road grade, road network density and connectivity index, etc.

[0037] Among them, the event characteristic data includes: information such as the name, type, and maximum capacity of large-scale events are crawled from the official website of the park or third-party public platforms using web crawling technology; in one embodiment, data can be collected from official event release platforms (e.g., when holding a concert, relevant event information is crawled from platforms such as Damai.cn and Maoyan Performance) using web crawling technology, mainly including: event type (large conference, concert, sports event, exhibition, etc.), event time (start / end time), event duration (minutes), and the number of participants officially announced (large > 10,000 people, medium 5,000-10,000 people, small < 5,000 people); among them, the park data is obtained through public information from the park management department or geographic information system (GIS), including core information such as the park's location coordinates, land area, surrounding road layout, internal charging facility distribution, and power grid access capacity.

[0038] Secondly, the collected data is preprocessed to form a dataset.

[0039] Data denoising method: Define the Z-Score normalization formula for abnormal charging power data as follows: ,in, For data values, This is the average of the data. Set a threshold for the standard deviation of the data (e.g., Identify abnormal data.

[0040] Missing value handling: Linear interpolation is used to fill in missing values.

[0041] Data association and fusion methods: Time association method: Activity time Matching charging behavior time By judging conditions The charging data is correlated with the activity characteristic data to obtain the first correlation result; Spatial association method: GeoHash is used for unified encoding. By calculating the same or adjacent GeoHash codes, the charging pile data is associated with POI and park data according to spatial proximity (within 3 kilometers) to obtain the second association result. Time unification method: Beijing time (UTC+8) is used uniformly, and time series alignment is achieved by using Pandas to obtain the time series alignment result in minutes; Data fusion: Based on the above results, charging behavior, environmental attributes, and activity characteristics data are fused to construct a dataset.

[0042] S2, the fused dataset is divided according to time and space dimensions to obtain data subsets with spatiotemporal feature labels, and a clustering algorithm is used to analyze the divided data subsets to identify data subsets with multiple spatiotemporal dimension cross combinations; wherein, the spatiotemporal feature labels include activity cycle rules, time period classification rules, spatial region division rules and date type.

[0043] Understandably, the spatiotemporal dimension division is based on the preprocessed fusion dataset and defines the following spatiotemporal dimension division rules to structure and organize data and focus on the core spatiotemporal features of large-scale event scenarios.

[0044] Spatiotemporal dimension division and preprocessing dataset represent the collaborative relationship between the analysis framework and the data object: the preprocessing dataset integrates multi-source data such as charging behavior (including minute-level time, latitude and longitude), activity attributes (such as stage, number of people), and environmental characteristics (such as POI, road network) to form analysis objects with spatiotemporal labels.

[0045] The spatiotemporal dimension is divided by defining rules such as activity stage (before / during / after), time period type (peak / off-peak), and spatial region (core / periphery) to classify, label, group, and aggregate data, and construct activity-spatiotemporal analysis units (such as charging data set of the core area during peak hours after a concert). This provides a structured analysis basis for subsequent feature extraction (such as regional charging power statistics) and pattern recognition (such as charging peak after a concert), and directly serves the patent objectives (such as charging demand prediction and facility scheduling optimization).

[0046] The time dimension is divided into the following categories: The event period is finely divided as follows: Pre-event period: 3 hours before the start of the event (T-3~T); Event period: the time from the start to the end of the event (T~T+X, where X is the event duration); Post-event period: 3 hours after the end of the event (T+X~T+X+3h). Time period classification: Peak hours: 17:00-22:00; Off-peak hours: excluding peak hours; Date types: Categorize weekdays, weekends, and holidays using the calendar API via Python; Spatial dimension division: Spatial area identification is performed using ArcGIS, including: Core area identification: charging stations within a 1-kilometer radius of the park; Outer area identification: charging stations within a 1-3 kilometer radius of the park.

[0047] The embodiments of the present invention integrate multi-source heterogeneous data, which can correlate the spatiotemporal characteristics of large-scale events with the heterogeneity of energy consumption behavior of electric vehicle users, thereby more comprehensively depicting the charging demand characteristics of the park.

[0048] Furthermore, the classification dimensions (activity type, activity cycle, time period classification, spatial region, date type) and their specific values ​​are entirely based on and derived from the data and rules processed or defined in the preceding steps (multi-source heterogeneous data mining methods and spatiotemporal dimension partitioning). This logical continuity ensures the objectivity and repeatability of scene subset identification.

[0049] First, based on the fused dataset (containing fields such as charging behavior, activity characteristics, and environmental attributes) formed by the aforementioned preprocessing steps, and the activity cycle rules, time period classification rules, spatial region division rules, and date type judgment logic defined in the spatiotemporal dimension segmentation step, this method first automatically generates the dataset in Python using a rule engine. i Data subsets that combine multiple spatiotemporal dimensions, such as the combination of "Activity Type [Concert] - Activity Status [In Progress] - Time Period [Peak Time] - Space [Core Area] - Time [Weekend]", each combination includes user charging behavior data, built environment data, and activity characteristic data. Among them, the activity type is directly taken from the activity characteristic data, and the activity cycle, time period classification, and date type are calculated based on spatiotemporal division rules combined with the recorded timestamp / date. The spatial area is calculated based on spatiotemporal division rules combined with the location of the charging station and the park.

[0050] Secondly, based on the combined data, the DBSCAN clustering algorithm was further selected for cluster analysis. This method is robust to noise and does not require a preset number of clusters. Specifically: (1) Calculate the statistical characteristics (mean and standard deviation) of each variable in each combination of data, and perform Z-Score standardization on the data of each variable: , and The first The first combination of data The mean and standard deviation of each variable, As input variables for clustering; (2) Set the neighborhood radius and minimum points And use Euclidean distance to measure the similarity between samples:

[0051] Iterate through all combinations of data for the sample set. any sample point Constructing based on sample points Centered on, with The set of neighborhood points of the neighborhood radius And determine the number of sample points contained in the neighborhood point set. Does it meet the requirements? If satisfied, then the sample points Taking this core point as the core, and starting from this core point, expanding to sample points with achievable density to form corresponding clusters; (3) After clustering is completed, calculate the silhouette coefficient for all combinations. As a standard for evaluating the practical utility and rationality of the clustering structure, the total number of valid sample points participating in the silhouette coefficient calculation is: If the average profile coefficient If the clustering is successful, the model can proceed to the training phase; otherwise, parameters need to be adjusted. and Re-clustering; (4) Train the model and finally output several spatiotemporal scene clusters, that is, the spatiotemporal scene subsets that have been identified, and label the subsets according to their features.

[0052] This invention proposes a clustering-based spatiotemporal scene recognition method, which can deconstruct spatiotemporal scenes in a more refined manner.

[0053] S3. Construct and train an integrated prediction model, including an XGBoost model, a TCN model, and a dynamic fusion module. Input the data subset with cross-combined spatiotemporal dimensions into the XGBoost model to output static feature weights, and input the time-series feature data from the data subset with cross-combined spatiotemporal dimensions into the TCN model to output time-series feature weights. Use the dynamic fusion module to fuse the static feature weights and time-series feature weights according to a preset dynamic adjustable ratio, and predict the final demand response power curve to optimize the integrated prediction model.

[0054] It is understood that the prediction object of this invention is the charging power demand curve of charging piles (core area + outer area) within 3 kilometers around the park under different spatiotemporal scenarios, and the target output is: the charging power sequence for the next 48 hours, with a time granularity of 15 minutes.

[0055] We construct a weighted XGBoost + TCN ensemble model. The object of this predictive ensemble model is a subset of data generated by the aforementioned rule engine, which is a cross-combination of various spatiotemporal dimensions (e.g., the dataset corresponding to “concert-event-peak-core-weekend”).

[0056] The XGBoost module processes static subsets of data that combine spatiotemporal dimensions. Each data subset, which is a cross-combination of spatiotemporal dimensions, is processed separately. The focus is on extracting the static features of the data within that subset (including scene-specific attribute labels and aggregated statistical features of the data within that subset) and outputting a scene subset feature weight matrix. (N is the scene subset label number obtained in the previous step, and K is the feature dimension). The "data subset with cross-combination of spatiotemporal dimensions" refers to each independent data subset with a specific spatiotemporal scene label (such as "concert-event-peak-core area-weekend") generated by the "clustering-based scene subset identification method" step in this invention. This step clarifies how the data is preprocessed, divided, and finally organized into units with clear business meaning for the model to learn.

[0057] Specifically, the input to the XGBoost module is a static feature matrix. T regression trees are iteratively trained using the XGBoost method to optimize the objective function. At the same time, through regularization terms Avoid overfitting; Feature weight matrix Each element in the matrix Quantified the first n The first scene subset k The contribution of dimensional features to charging power prediction.

[0058] The data processing procedure using the XGBoost module in this embodiment of the invention is as follows: a. First, the data subset with static spatiotemporal dimensions cross-combined. Perform preprocessing to remove outliers.

[0059] b. Secondly, based on the XGBoost principle, a regression tree is iteratively constructed to fit the data, with the objective function being:

[0060] in, For predicted values Compared with the true value The loss function (i.e., mean squared error); For the number of trees, For the first Tree; , is a regularization term that controls model complexity. The number of leaf nodes. The weights of the leaf nodes are... and This is the regularization parameter.

[0061] c. Finally, in this module, feature weights are obtained based on the number of splits and gain metrics of the features in the tree model. For N Each subset of scenes corresponds to a certain number of scenes. K 3D features, output weight matrix: Indicates the first n The first scene subset k The weights of the features are dimensional, and the magnitude of the weights reflects the importance of the features to the prediction (e.g., the degree of influence of activity type on charging power).

[0062] TCN module: Processes time-series feature data for different subsets of scenarios. , including (charging power sequence) SOC time series change rate The algorithm extracts long-term dependencies through causal dilated convolution and outputs a temporal feature weight matrix. ( L The temporal feature dimension refers to the sequence of specific observations (such as charging power sequences, SOC time-series change rate sequences, etc.) contained within the "data subset with cross-combined spatiotemporal dimensions" in this method, arranged in chronological order. Its relationship with the data acquired above is that of component to whole, specific manifestation to the carrying entity. The TCN module utilizes these finely segmented and well-organized temporal series data to learn the dynamic evolution patterns of specific scenarios.

[0063] Specifically, the TCN module is input with time-series feature data. ,correspond N A subset of scenes T Each time step L Temporal dependencies are extracted through causal dilated convolutional layers, temporal dimension information is aggregated through global pooling, and normalized weight vectors are generated through fully connected layers and softmax; the output is a temporal feature weight matrix. Each element in the matrix Indicates the first n The first scene subset l The importance of time-series features.

[0064] The data processing procedure of the TCN module in this embodiment of the invention is as follows: a. First, regarding time-series characteristic data Preprocessing is performed to remove outliers, where N The number of scene subsets, each scene subset Include L 3D time-series characteristics, such as the charging power sequence from the past 24 time steps to the present. wait.

[0065] b. Secondly, TCN captures long-term dependencies through multiple layers of causal dilated convolution, i.e., for the th... Layer convolution performs the following operations: .in, P The kernel size is [size]. For the first The expansion rate of the layer; The first convolution kernel Each weight, This is the ReLU activation function.

[0066] c. Let For the input time series feature sequence at time step The value of , , The time series length is given; then, by stacking multiple layers of causal dilated convolutional layers, dependencies at different time scales are captured: .in, , For the final hidden layer output, and For the first Layer weights and biases.

[0067] d. Finally, for each scene subset n Temporal feature weight vectors are generated through global average pooling and fully connected layers, i.e.: Indicates the first The temporal feature aggregation vector obtained by global average pooling of a subset of scenes is then processed. Perform feature aggregation (T is the length of the time series, such as 100 time points); Indicates the first Temporal feature weight vectors corresponding to each scene subset This represents the bias vector during the weight mapping process. This represents the fully connected weight mapping matrix, used to aggregate temporal features into vectors. Mapped to weight scores corresponding to each time-series feature, through Perform weight mapping. Ensure the sum of the weights is 1; Combine the weight vectors of all scene subsets into a matrix. ,in For the first n The temporal feature weight vector of a subset of scenes, with dimension L.

[0068] Furthermore, the dynamic fusion module incorporates static feature weights. With time series feature weights The final demand response power curve is predicted by fusing the components using a dynamically adjustable ratio. The fusing formula is as follows:

[0069] in, Adaptive weights for scene subsets ; The Sigmoid activation function introduces non-linear characteristics.

[0070] This invention combines the XGBoost model and the TCN module model to form a new predictive ensemble model that can simultaneously process static single-scene data and dynamic time-series data.

[0071] The following is the training process of the ensemble prediction model according to an embodiment of the present invention: Each submodule of the model employs Bayesian optimization and five-fold cross-validation. The hyperparameters of the XGBoost and TCN models for each scene subset are trained independently, allowing each model to achieve its optimal performance. The evaluation criterion is minimizing the RMSE (Root Mean Squared Error). XGBoost optimization parameters include learning rate, maximum tree depth, and subsample ratio; TCN optimization parameters include kernel size, dilation rate, number of network layers, and dropout rate. The performance of each submodule is evaluated using the minimum RMSE. The specific training process is as follows: A. For the two sub-models XGBoost and TCN, minimize the time allotted for each sub-model within its corresponding scene subset. With RMSE as the target, each sub-model is trained independently:

[0072]

[0073] in, The number of samples in the scene subset s; Represents a subset of scenes s The Middle i The actual observed values ​​of each sample; and These represent the hyperparameters corresponding to the XGBoost and TCN models, respectively. and The predicted value; The hyperparameters used in training include: learning rate Maximum tree depth Subsample ratio wait; The hyperparameters used in training include: kernel size. Expansion rate Network layers Dropout rate ( )wait.

[0074] B. The optimization process is based on the Bayesian optimization algorithm. Five-fold cross-validation is used to verify and evaluate the model performance of each hyperparameter combination. The specific steps are as follows: First, define the hyperparameter space (i.e. the possible range of values ​​for the hyperparameters) of the XGBoost and TCN models. Next, using a Gaussian process (GP) as a surrogate model, we performed Bayesian optimization on both the XGBoost and TCN models in each round: a. Based on the current surrogate model, determine the next possible optimal combination of hyperparameters in the hyperparameter space; b. The mean RMSE of the model performance was calculated using five-fold cross-validation and used as the objective function value; c. Update the proxy model and iterate until the preset number of optimizations or convergence.

[0075] The formula for 50% cross-validation is:

[0076] in, Indicates the first The number of samples in folded cross-validation.

[0077] Finally, the optimal hyperparameter combination for each sub-model is obtained. and This is used for training and ensemble prediction of the subsequent dynamic fusion module.

[0078] Furthermore, the parameters of the dynamic fusion layer are trained to further adjust the fusion weights. We continue to optimize the gradient descent using RMSE as the model evaluation metric. The specific training process is as follows: the dynamic fusion layer parameters refer to the adaptive weights. That is, obtaining the optimal hyperparameters of the sub-model and : The optimal hyperparameter combination was obtained in the independent training module of the sub-model. and In the current fusion module, the outputs of these two sub-models are fixed and no longer participate in training; they are only used as input to the fusion layer. The optimization objective function of the fusion layer is (still using RMSE as the metric):

[0079]

[0080] By minimizing Training Dynamic adaptive weights in ,in, For the first Each sample inputs a static feature vector to XGBoost. For the first The temporal feature sequence of each sample input TCN, and These are the XGBoost model and the TCN model, respectively. After obtaining the optimal adaptive weights for all sub-scenes, the following calculations can be performed. Static feature weights in With time series feature weights .

[0081] When the RMSE of the ensemble model on the validation set is lower than the set threshold of 1.5kW, proceed to the next step.

[0082] This invention proposes a new dynamic training strategy for the XGBoost + TCN integrated model, which improves the predictive adaptability and accuracy of the model to charging and discharging demand response behavior under different spatiotemporal scenarios through dynamic fusion.

[0083] S4 calculates the baseline load curve based on historical data of inactive days corresponding to the data subsets of cross-combination of spatiotemporal dimensions, and inputs the real-time spatiotemporal scenario subset data into the optimized integrated prediction model to output the predicted load curve, and calculates the difference between the predicted load curve and the baseline load curve to quantify the demand response potential under different spatiotemporal scenario subsets.

[0084] The baseline load calculation module extracts historical data for each scenario subset on inactive days (or days with the lowest load) and calculates the baseline load curve. It also extracts the historical data corresponding to each single spatiotemporal scenario subset on inactive days (or days with the lowest load). Each scenario subset, in the aforementioned "Scenario Subset Identification Method Based on Clustering" step of this invention, is a specific "single spatiotemporal scenario subset" (e.g., "Concert - Event - Peak Hour - Core Area - Weekend") automatically generated by the rule engine. The calculations of baseline load, predicted load, and demand response potential are all based on these subdivided scenarios.

[0085] Calculate the baseline load curve for this specific subset s of scenarios. :

[0086] in, This is a collection of historical load data for days with no activity or lowest load days. This represents the time series load data for day d in the set, representing a time series point. t=1,…,T.

[0087] The predicted load module calculates the optimal adaptive weights for all sub-scenes. After that, it can be calculated and obtained Static feature weights in With time series feature weights This involves obtaining the optimized XGBoost-TCN ensemble model; then, by calling the optimized XGBoost-TCN ensemble model, inputting... and Data, output predicted load curve That is, to obtain the fusion formula ; The adjustable load potential prediction module calculates the difference between the predicted load and the baseline load for a subset of scenarios s. Quantify the demand response potential, quantify the demand response potential under different spatiotemporal scenario subsets, and store them according to spatiotemporal scenario subsets.

[0088] Therefore, by integrating multi-source data, identifying spatiotemporal scenarios, extracting features, training integrated models, and predicting demand response potential, a complete closed loop is formed, enabling refined prediction and control of electric vehicle charging and discharging loads across multiple scenarios and time periods.

[0089] The electric vehicle demand response potential prediction method for large-scale events in industrial parks according to embodiments of the present invention refines the spatiotemporal scenario to improve the accuracy of the prediction system. It integrates static and dynamic data within the same potential prediction system framework and proposes a dynamic model training strategy, thereby providing a prediction model with greater generalization capabilities. Taking full account of the characteristics of large-scale events, it collects relevant data and proposes a multi-source heterogeneous data fusion method, thereby combining user energy consumption behavior characteristics with event characteristics to predict the cluster electric vehicle demand response potential generated under the scenario of large-scale events in the industrial park.

[0090] The beneficial effects of this invention are as follows: By predicting the charging and discharging behavior of EVs during large-scale events, grid operators can more accurately dispatch power resources, balance loads, reduce grid instability caused by load fluctuations, and improve the stability of the park's power system and the level of event services. Based on the prediction of response potential under different time and space scenarios, relevant departments can rationally plan and deploy charging facilities to avoid resource waste and improve the utilization rate and service capabilities of charging facilities. Accurate potential forecasting provides data support for grid dispatching, helps operators to rationally plan the layout of charging facilities, optimize V2G two-way interaction strategies, fully tap the energy storage potential of electric vehicles, and reduce grid upgrade costs.

[0091] To achieve the above embodiments, such as Figure 3 As shown, this embodiment also provides an electric vehicle demand response potential prediction system 10 for large-scale events in the park, including: The multi-source heterogeneous data mining module 100 is used to preprocess multi-source heterogeneous data during large-scale events held in the park to obtain a fused dataset. The spatiotemporal scene deconstruction and recognition module 200 is used to divide the fused dataset according to time and space dimensions to obtain data subsets with spatiotemporal feature labels, and to use a clustering algorithm to analyze the divided data subsets to identify data subsets with multiple spatiotemporal dimensions; wherein, the spatiotemporal feature labels include activity cycle rules, time period classification rules, spatial region division rules and date type; An integrated model building module 300 for different spatiotemporal scenarios is used to build and train an integrated prediction model, including an XGBoost model, a TCN model and a dynamic fusion module. The module inputs a subset of data with cross-combined spatiotemporal dimensions into the XGBoost model to output static feature weights, and inputs the temporal feature data from the subset of data with cross-combined spatiotemporal dimensions into the TCN model to output temporal feature weights. The dynamic fusion module uses the static feature weights and temporal feature weights to fuse them according to a preset dynamic adjustable ratio, and predicts the final demand response power curve to optimize the integrated prediction model. The demand response potential prediction module 400 is used to calculate the baseline load curve based on the historical data of inactive days corresponding to the data subsets of cross-combination of spatiotemporal dimensions, and input the real-time spatiotemporal scenario subset data into the optimized integrated prediction model to output the predicted load curve, and calculate the difference between the predicted load curve and the baseline load curve to quantify the demand response potential under different spatiotemporal scenario subsets.

[0092] The electric vehicle demand response potential prediction system for large-scale events in a park, according to embodiments of the present invention, refines the spatiotemporal scenario to improve the accuracy of the prediction system. It integrates static and dynamic data within the same potential prediction system framework and proposes a dynamic model training strategy, thereby providing a prediction model with greater generalization capabilities. By fully considering the characteristics of large-scale events, relevant data is collected, and a multi-source heterogeneous data fusion method is proposed, combining user energy consumption behavior characteristics with event characteristics, enabling the prediction of clustered electric vehicle demand response potential generated under the scenario of large-scale events in the park.

[0093] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0094] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for predicting the demand response potential of electric vehicles for large-scale events in a park, characterized in that, include: A fused dataset is obtained by preprocessing multi-source heterogeneous data on electric vehicle energy consumption behavior during large-scale events held in the park. The fused dataset is divided into subsets with spatiotemporal feature labels according to time and space dimensions. Then, a clustering algorithm is used to analyze the subsets to identify data subsets with multiple spatiotemporal dimension cross combinations. The spatiotemporal feature labels include activity cycle rules, time period classification rules, spatial region division rules, and date type. An integrated prediction model is constructed and trained, including an XGBoost model, a TCN model, and a dynamic fusion module. A subset of data with cross-combined spatiotemporal dimensions is input into the XGBoost model to output static feature weights, and the temporal feature data in the subset of data with cross-combined spatiotemporal dimensions is input into the TCN model to output temporal feature weights. The dynamic fusion module is used to fuse the static feature weights and the temporal feature weights according to a preset dynamic adjustable ratio to predict the final demand response power curve and optimize the integrated prediction model. The baseline load curve is calculated based on historical data of inactive days corresponding to the data subsets with cross-combination of spatiotemporal dimensions. Real-time spatiotemporal scenario subset data is input into the optimized integrated prediction model to output the predicted load curve. The difference between the predicted load curve and the baseline load curve is calculated to quantify the demand response potential under different spatiotemporal scenario subsets.

2. The method for predicting the electric vehicle demand response potential for large-scale events in a park according to claim 1, characterized in that, Multi-source heterogeneous data includes charging behavior data, built environment data, and activity characteristic data. Among them, charging behavior data includes battery state of charge, charging start and end times, charging duration, real-time charging power, charging station location coordinates, battery state of charge start and end values, time period markers, and user IDs. Built environment data includes POI data and road network characteristic data. Activity characteristic data includes activity type, activity time, activity duration, and number of participants.

3. The method for predicting the electric vehicle demand response potential for large-scale events in a park according to claim 2, characterized in that, A fused dataset was obtained by preprocessing multi-source heterogeneous data on electric vehicle energy consumption behavior during large-scale events held in the park, including: Define activity time Matching charging behavior time By judging conditions The charging behavior data is correlated with the activity characteristic data to obtain the first correlation result; GeoHash is used for unified encoding, and the charging behavior data is associated with the built environment data and activity characteristic data according to spatial proximity by calculating the same or adjacent GeoHash codes to obtain the second association result. We uniformly adopted Beijing time and used Pandas to align the time series in minutes to obtain the time series alignment results. Based on the first association result, the second association result, and the time series alignment result, charging behavior data, built environment data, and activity characteristic data are integrated to construct a fused dataset.

4. The method for predicting the electric vehicle demand response potential for large-scale events in industrial parks according to claim 1, characterized in that, Clustering algorithms are used to analyze the partitioned data subsets to identify data subsets with multiple spatiotemporal dimensions that intersect and combine, including: Calculate the statistical characteristics of each variable in each combined data set based on multi-source heterogeneous data, and then perform Z-Score standardization on the data of each variable: , and The first The first combination of data The mean and standard deviation of each variable, As input variables for clustering; Set neighborhood radius and minimum points And use Euclidean distance to measure the similarity between samples: Iterate through all combinations of data for the sample set. any sample point Constructing based on sample points Centered on, with The set of neighborhood points of the neighborhood radius And determine the number of sample points contained in the neighborhood point set. Does it meet the requirements? If satisfied, then the sample points Taking this core point as the core, and starting from this core point, expanding to sample points with achievable density to form corresponding clusters; After clustering is completed, the silhouette coefficient is calculated for all combinations. As a standard for evaluating the practical utility and rationality of clustering structures, the total number of valid sample points participating in the silhouette coefficient calculation is... If the average profile coefficient Enter the training phase; otherwise, parameters need to be adjusted. and Re-clustering; After training, several spatiotemporal scene clusters are finally output, which are the identified spatiotemporal scene subsets. The subsets are then labeled according to their characteristics.

5. The method for predicting the electric vehicle demand response potential for large-scale events in a park according to claim 1, characterized in that, A subset of data with cross-combined spatiotemporal dimensions is input into the static feature weights output by the XGBoost model, including: Data subsets that combine spatiotemporal dimensions Preprocessing is performed to remove outliers; Based on the preprocessed data, and according to the XGBoost principle, a regression tree is iteratively constructed to fit the data. The objective function is: in, For predicted values Compared with the true value The loss function; For the number of trees, For the first Tree; , is a regularization term that controls model complexity. The number of leaf nodes. The weights of the leaf nodes are... and For regularization parameters; Based on the number of splits and gain metrics of the features in the tree model, obtain the feature weights: For N Each subset of scenes corresponds to a certain number of scenes. K 3D features, output weight matrix: in, Indicates the first The first scene subset Weights of dimensional features.

6. The method for predicting the electric vehicle demand response potential for large-scale events in a park according to claim 1, characterized in that, Temporal feature data from a subset of data with cross-combined spatiotemporal dimensions are input into the TCN model to output temporal feature weights, including: For time series feature data Perform preprocessing and remove outliers; Each scene subset, Include L Dimensional time series features; set up For the input time series feature sequence at time step The value of , , The time series length is specified; TCN captures long-term dependencies through multiple layers of causal dilated convolution, i.e., for the th... Layer convolution performs the following operations: ;in, P The kernel size is [size]. For the first The expansion rate of the layer, The first convolution kernel Each weight, It is the ReLU activation function; By stacking multiple layers of causal dilated convolutional layers, dependencies at different time scales are captured: ;in, , For the final hidden layer output, and For the first Layer weights and biases; For each scene subset n Temporal feature weight vectors are generated through global average pooling and fully connected layers, i.e.: Indicates the first The temporal feature aggregation vector obtained by global average pooling of a subset of scenes is then processed. Perform feature aggregation; T The length of the time series; Indicates the first Temporal feature weight vectors corresponding to each scene subset This represents the bias vector during the weight mapping process. This represents the fully connected weight mapping matrix, used to aggregate temporal features into vectors. Mapped to weight scores corresponding to each time-series feature, through Perform weight mapping, Ensure the sum of the weights is 1; Combine the weight vectors of all scene subsets into a matrix. ,in For the first n The temporal feature weight vector of each scene subset, with dimension [ ]. L .

7. The method for predicting the electric vehicle demand response potential for large-scale events in a park according to claim 1, characterized in that, static feature weights With time series feature weights Fusion is performed using a dynamically adjustable ratio to predict the final demand response power curve; Fusion formula: in, Adaptive weights for scene subsets ; The Sigmoid activation function introduces non-linear characteristics.

8. The method for predicting the electric vehicle demand response potential for large-scale events in a park according to claim 1, characterized in that, Training two sub-models, XGBoost and TCN, includes: To minimize the performance of each sub-model in the corresponding scene subset With RMSE as the target, each sub-model is trained independently: in, Scene subset s The number of samples in the sample; Represents a subset of scenes s The Middle The actual observed values ​​of each sample; and These represent the hyperparameters corresponding to the XGBoost and TCN models, respectively. and The predicted value; The hyperparameters used in training include: learning rate Maximum tree depth Subsample ratio ; The hyperparameters used in training include: kernel size. Expansion rate Network layers Dropout rate ; The model performance was validated and evaluated for each hyperparameter combination using five-fold cross-validation, including: Determine the hyperparameter spaces of XGBoost and TCN models; Using a Gaussian process as a surrogate model, the following operations are performed in each round of Bayesian optimization for both XGBoost and TCN models: Based on the current surrogate model, determine the next optimal combination of hyperparameters in the hyperparameter space; calculate the mean RMSE of the model performance using five-fold cross-validation as the objective function value; update the surrogate model and iterate until the preset number of optimizations or convergence; the five-fold cross-validation formula is: in, Indicates the first The number of samples in folded cross-validation; Finally, the optimal hyperparameter combination for each sub-model is obtained. and .

9. The method for predicting the electric vehicle demand response potential for large-scale events in a park according to claim 8, characterized in that, Based on optimal hyperparameter combination and The objective function for optimizing the fusion layer is: in, Represents a subset of scenes The corresponding dynamic adaptive fusion weights, and , For scene subset The optimal solution for the lower dynamic fusion layer is obtained through objective function optimization; By minimizing Training Dynamic adaptive weights in ,in, For the first Each sample inputs a static feature vector to XGBoost. For the first The temporal feature sequence of each sample input TCN, and These are the XGBoost model and the TCN model, respectively.

10. The method for predicting the electric vehicle demand response potential for large-scale events in a park according to claim 1, characterized in that, A baseline load curve is calculated based on historical data of inactive days corresponding to a data subset with cross-combined spatiotemporal dimensions. Real-time spatiotemporal scenario subset data is then input into the optimized ensemble forecasting model to output a predicted load curve. The difference between the predicted load curve and the baseline load curve is calculated to quantify the demand response potential under different spatiotemporal scenario subsets, including: Calculate a specific subset of scenes - s The baseline load curve : in, This is a collection of historical load data for days with no activity or lowest load. Represents the set of elements. d Daily load data, time series points t=1, …,T ; Calculate the optimal adaptive weights for all sub-scenes Then, calculate and obtain Static feature weights in With time series feature weights That is, to obtain the optimized XGBoost-TCN ensemble model, call the optimized XGBoost-TCN ensemble model, and input... and Data, output predicted load curve ; Calculate the difference between the predicted load curve and the baseline load curve. Quantify the demand response potential, quantify the demand response potential under different spatiotemporal scenario subsets, and store them according to spatiotemporal scenario subsets.

Images