Power grid response demand dynamic analysis method based on multi-source data fusion

Abstract

The invention discloses a power grid response demand dynamic analysis method based on multi-source data fusion, and particularly relates to the technical field of power system load prediction. Collecting power load, new energy output, weather, electricity price and user behavior data, and constructing a fusion feature vector group; extracting a response sensitive feature sequence by using a bidirectional gating recurrent neural network and an attention mechanism; constructing a state evolution diagram based on the high response features, identifying response mode sub-diagrams through diagram embedding and clustering, and extracting candidate response trigger factors; combining the factors with future time period prediction data, inputting the combined factors and future time period prediction data into a prediction model of an encoder-decoder structure, generating a power grid response demand change curve, and identifying a high-risk time period; and finally, outputting an executable load response scheduling scheme in combination with power grid scheduling resource constraints. According to the invention, dynamic identification and accurate prediction of the load response behavior are realized, and the safety and intelligent level of power grid operation are improved.

Description

Technical Field

[0001] This invention relates to the field of power system load forecasting technology, and specifically to a method for dynamic analysis of power grid response demand based on multi-source data fusion. Background Technology

[0002] In modern power systems, with the increasing proportion of renewable energy integration year by year, the load response capacity of traditional power grids is struggling to meet dynamically changing electricity demands. Particularly in urban areas with high proportions of renewable energy integration (such as wind and solar power) and high electric vehicle penetration rates, power grid load fluctuations exhibit stronger characteristics of short cycles, high frequencies, and uncertainty. This multi-source, heterogeneous disturbance pattern severely impacts the stability of the power grid and the real-time performance of dispatch strategies.

[0003] Existing power grid response demand analysis methods mostly rely on single data sources (such as historical electricity load data and electricity prices) to build predictive models, lacking the integration and utilization of multi-source data (such as weather data, user behavior data, and renewable energy output data), making it difficult to accurately perceive changes in response demand during actual power grid operation. Furthermore, current methods generally suffer from insufficient sensitivity to local fluctuations, response time delays, and poor robustness under conditions of missing or abnormal data.

[0004] Especially in scenarios where the power grid frequently experiences rapid switching between "source-load-storage", such as sudden weather changes in urban load centers, concentrated charging of regional electric vehicles, or sudden shutdown of distributed energy storage, traditional static response analysis methods cannot quickly and effectively locate the load-side response potential, affecting the accuracy and timeliness of dispatch decisions. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic analysis method for power grid response demand based on multi-source data fusion, so as to overcome the shortcomings of the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for dynamic analysis of power grid response demand based on multi-source data fusion, comprising: Collect multi-source data within the target area, including electricity load data, renewable energy output data, meteorological data, electricity price data, and user behavior data, to form an initial dataset; The initial dataset is preprocessed to obtain a preprocessed dataset. Based on the preprocessed dataset, a multi-source feature fusion vector group F={f1,f2,...,fi,...,fn} is constructed, where each fi represents the comprehensive state feature across all data sources at a given time. The multi-source feature fusion vector group F is input into the response demand identification model, and the output is the power grid response sensitive feature sequence of the target area at the current time. Based on the power grid response sensitive feature sequence, a state evolution diagram is constructed. The state evolution diagram takes the high response features in the power grid response sensitive feature sequence as key nodes, and mines their fluctuation patterns and triggering conditions in historical time series to obtain a set of candidate response triggering factors. The candidate response triggering factor set is input into the dynamic response prediction model. Combined with multi-source prediction data of the target area for future periods, the power grid response demand change curve for future periods is output, and high response risk periods are marked. Based on the power grid response demand change curve and the power grid dispatch resource constraints, a dispatch scheme is output.

[0007] Preferably, the comprehensive status characteristics include: load characteristics, new energy characteristics, meteorological characteristics, electricity price characteristics, and user behavior characteristics.

[0008] Preferably, the multi-source feature fusion vector set F is input into the response requirement identification model, including: Construct a response demand identification model based on bidirectional gated loop units, input multi-source feature fusion vector group F, and extract context-related state information for each time slice; The extracted temporal features are weighted to obtain the attention distribution of the current time slice across all historical states; Generate the power grid response sensitive feature vector of the target area at the current moment based on the weighted feature representation; The power grid response-sensitive feature vectors are constructed into a response-sensitive feature sequence in chronological order.

[0009] Preferably, the state evolution diagram is constructed based on the power grid response sensitive feature sequence, including: The response intensity is calculated based on the magnitude of each feature vector in the power grid response sensitive feature sequence, and feature points exceeding a preset threshold are marked as high response features. Using high-response features as key nodes, directed connections are established based on their chronological order on the time axis and feature similarity to construct the initial graph structure of the state evolution graph. Cluster analysis is performed on the nodes in the state evolution graph to identify response pattern subgraphs with common evolution paths; By combining the contextual features of key nodes in the response pattern subgraph with the input vector of the corresponding time period, a set of candidate response triggering factors that cause significant changes in the power grid load response is extracted.

[0010] Preferably, the step of performing cluster analysis on the nodes in the state evolution graph to identify response pattern subgraphs with common evolution paths includes: For each key node in the state evolution graph, extract its structural adjacency information and the response-sensitive feature vector of the corresponding time period, and construct a node structure-attribute fusion representation vector; An embedding algorithm based on graph neural networks is used to perform low-dimensional embedding mapping on the fused representation vectors to obtain the set of representation vectors of each node in a unified space; Clustering of node embedding vectors based on density peak clustering method identifies node clusters with similar response feature evolution trends; Reconstruct the response pattern subgraph by arranging the nodes in the same clustering result according to their chronological order and their connection relationships in the graph.

[0011] Preferably, the output of the grid response demand change curve over future periods includes: Acquire multi-source forecast data for the target area in the future time period, including electricity load forecast, meteorological forecast, new energy output forecast and electricity price forecast data, and construct the input sequence for the future time period; The candidate response triggering factor set is feature-aligned with the input sequence of future time periods to form a highly correlated prediction input vector sequence; A long short-term memory neural network model based on an encoder-decoder structure is constructed as a dynamic response prediction model. The prediction input vector sequence is time-series modeled to output the power grid response demand change curve in the future cycle. Based on the gradient changes in the prediction curve, time periods exceeding a set threshold are identified and marked as high-risk response periods.

[0012] Preferably, the period identified as a high-risk response period includes: Time difference operation is performed on the predicted power grid response demand change curve to calculate the change gradient sequence between adjacent time points; Set the response change gradient threshold G0; Traverse the gradient sequence and mark the time intervals in which the gradient changes continuously and is greater than G0 as the initial high-risk intervals; By combining the contextual features before and after each initial high-risk segment, the boundary is expanded, and boundary points with significant response intensity are retained to finally determine the high-response-risk period.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention addresses the problems of existing technologies, such as reliance on single data sources, insufficient accuracy in response feature identification, and inability to dynamically predict high-risk periods, by introducing multi-source data fusion, response-sensitive feature modeling, and state evolution graph analysis. By constructing a power grid response-sensitive feature sequence and combining graph embedding and clustering analysis methods, it can accurately extract key factors affecting load response and identify typical response evolution paths, thereby improving the perception and prediction capabilities of power grid operating state changes.

[0014] 2. This invention constructs a time-series prediction model based on an encoder-decoder structure and combines it with the actual constraints of power grid scheduling resources to output an executable scheduling strategy, thereby realizing proactive intervention and control of the power grid during periods of high response risk. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] For examples, please refer to Figure 1 As shown in this embodiment, a method for dynamic analysis of power grid response demand based on multi-source data fusion includes: Collect multi-source data within the target area, including electricity load data, renewable energy output data, meteorological data, electricity price data, and user behavior data, to form an initial dataset.

[0019] In this embodiment, to achieve dynamic analysis of power grid response demands, it is first necessary to comprehensively collect multi-source data within the target area to form an initial dataset covering the power grid's operating status. Specifically, the multi-source data includes the following five categories: Collect historical electricity load data of users at all levels within the target area, including time-of-use load curves for residential, commercial, and industrial users, with a time resolution of no less than 15 minutes; and include parameters such as real-time power flow, load factor, and load forecast values ​​of distribution transformers.

[0020] Acquire historical and real-time power output data of renewable energy sources such as photovoltaic and wind power connected to the target area, specifically including installed capacity, power output, active / reactive power, and power output fluctuation rate. Data sources include photovoltaic inverters, wind turbine control systems, and new energy monitoring platforms.

[0021] Collect real-time and forecast meteorological data corresponding to the target area, including meteorological elements such as temperature, humidity, wind speed, wind direction, light intensity, rainfall, and air pressure.

[0022] To obtain information such as real-time electricity prices, day-ahead prices, peak-valley prices, and time-of-use prices in the electricity market, so as to reflect the impact of electricity prices on user response behavior.

[0023] Collect data on the behavioral patterns of load-side users during energy consumption, including equipment operating time, start-stop frequency, electric vehicle charging behavior, air conditioning load adjustment behavior, and historical response records.

[0024] All collected data are timestamped and sliced ​​according to a preset data sampling period (e.g., 5 minutes, 15 minutes, or 1 hour) to finally construct a multi-dimensional, multi-time-scale initial dataset D0 for subsequent feature extraction and model analysis.

[0025] The initial dataset is preprocessed to obtain a preprocessed dataset. The specific data preprocessing process includes the following steps: Since the sampling periods of different data sources may be inconsistent (e.g., meteorological data is hourly and load data is 15-minute), this embodiment first resamples and aligns all data to a uniform time resolution (e.g., 15 minutes). For data with high time granularity, mean / sum downsampling is performed, and for data with low granularity, upsampling is performed through linear interpolation or sliding prediction to ensure that all data samples are aligned at the same timestamp.

[0026] The initial dataset may contain missing values ​​due to factors such as equipment failure or communication interruption. To improve data integrity and analytical accuracy, a hybrid strategy is used for missing value imputation: Linear interpolation is used to fill short-term gaps (less than 2 periods); For long-term missing data, historical window average values ​​or imputation strategies based on similar day patterns are used as replacements. For highly time-varying data such as new energy power output, intelligent filling can be achieved using a regression interpolation algorithm based on random forests.

[0027] Outlier data can severely interfere with modeling results. A combination of Z-Score detection and the Isolation Forest algorithm is used for outlier identification. For identified outliers, smoothing or replacement with local means is performed based on the trends of adjacent time periods to ensure the reasonableness and continuity of data fluctuations.

[0028] To eliminate the impact of differences in the dimensions of data, min-max normalization or standard deviation normalization is used to uniformly map all numerical features to [-1, 1] or a standard normal distribution with a mean of 0 and a variance of 1, which is beneficial to the convergence and stability of models such as neural networks.

[0029] To preserve the periodic influence of time factors on load and behavioral characteristics, this embodiment converts the timestamp of each time point into periodic time features (such as hour, day of the week, holiday markers, etc.) and introduces them into the dataset through one-hot encoding.

[0030] All processed data is organized into a unified structured multidimensional tensor format, where each record contains a complete set of feature fields, including: load features, meteorological features, electricity price features, behavioral features, and time features; and stored in the efficient Parquet or HDF5 format for efficient reading and training of subsequent models. This processing results in the preprocessed dataset D1.

[0031] Based on the preprocessed dataset, a multi-source feature fusion vector group F={f1,f2,...,fi,...,fn} is constructed, where each fi represents the comprehensive state feature across all data sources at a given time. The specific construction method is as follows: The features in the preprocessed dataset D1 are divided into the following five categories according to the data source: Load characteristics Li include: current total load, active / reactive load, load growth rate, and peak load percentage; New energy category characteristics Ri include wind / solar power output, prediction error, power output volatility, and the slope of the unit power output curve; Meteorological features Wi include temperature, humidity, wind speed, solar radiation intensity, and weather change indicators; Electricity price characteristics Pi include time-of-use pricing, real-time price fluctuations, and peak-valley price differences; User behavior features Bi include device start / stop frequency, air conditioning / charging load ratio, and response history markers.

[0032] This embodiment employs the following two fusion strategies to construct the comprehensive state vector fi at each time step: Concatenation: The above five types of features are concatenated at each time slice to obtain fi=[Li,Ri,Wi,Pi,Bi]; where each feature sub-vector has been normalized in the previous stage, and after concatenation, a comprehensive feature vector with consistent dimensions and uniform format is formed.

[0033] This vector group F not only fully expresses the operating status of the power grid at each moment, but also preserves the dynamic correlation between the source data, which facilitates subsequent model response demand analysis and prediction based on sequence input.

[0034] The multi-source feature fusion vector group F is input into the response demand identification model, and the power grid response sensitive feature sequence of the target area at the current moment is output.

[0035] This embodiment constructs a response demand identification model based on a bidirectional gated recurrent unit (Bi-GRU) to model the context state of the fused vector group F.

[0036] Specifically, the fused vector set F is used as the input sequence and fed into a structure consisting of gated recurrent units in two directions. The forward GRU processes the sequence sequentially from time t=1 to t=n, extracting forward features; the backward GRU processes the sequence in reverse from time t=n to t=1, extracting backward features. The context-dependent state vector h for each time slice t is... t It is obtained by concatenating the positive output ht⁺ and the negative output ht⁻. The internal structure of the gated loop unit includes an update gate, a reset gate and a candidate state unit. By nonlinearly combining the previous state and the current input, it realizes dynamic memory and forgetting of historical features.

[0037] In this embodiment, the hidden state dimension of GRU is set to 128, the activation function is the hyperbolic tangent function (Tanh), and the training method is the backpropagation algorithm.

[0038] To enhance the model's ability to identify key moment features, this embodiment introduces attention-weighted calculation based on the temporal feature ht to generate the attention distribution weights for the current time slice t. Specifically, for each ht, its attention score et is first calculated using a fully connected neural network: ;in, Let represent the activation function, which is the hyperbolic tangent function. The weight matrix has a dimension of (64×256). We represents the weight matrix, and be represents the bias term, which is a vector of length 64. Then, the scores of each time slice are normalized using the Softmax function to obtain the attention weight αt of all historical time slices at the current analysis time t: ;in This represents the summation of the scores for all time slices k after performing an exponential operation.

[0039] Finally, based on the attention weights, the temporal features ht at all times are summed in a weighted manner to obtain the global context fusion representation vector ct for the current time slice: This vector ct represents the feature correlation and key focus area of ​​the current time point in the entire historical sequence.

[0040] Based on the weighted global context representation vector ct, this embodiment inputs it into a fully connected neural network to output the grid response-sensitive feature vector rt of the target region at the current moment. The specific structure is: rt = ReLU(We × ct + be); where ReLU represents the linear rectification function, used to introduce nonlinear feature representation, the weight matrix dimension is set to (32 × 256), and the output rt is a response-sensitive vector of length 32. The vector rt expresses the core influencing factors that cause rapid load fluctuations or potential responses at this moment, such as abnormally high temperatures, large-scale grid connection of electric vehicles, or amplified fluctuations in distributed power sources.

[0041] The response-sensitive feature vectors rt generated at all times from t=1 to t=n are combined in chronological order to form a complete response-sensitive feature sequence R={r1,r2,...,rn}. This sequence not only preserves the response intensity characteristics at each time point in the historical time series, but also provides a high-dimensional input foundation for subsequent modeling, used to predict future response trends or construct state evolution paths.

[0042] Based on the power grid response sensitive feature sequence, a state evolution diagram is constructed. The state evolution diagram takes the high response features in the power grid response sensitive feature sequence as key nodes, and mines their fluctuation patterns and triggering conditions in historical time series to obtain a set of candidate response triggering factors.

[0043] First, a preset threshold T0 for the response intensity is set. The value of T0 is based on the mean μ and standard deviation σ of the norm of the feature vectors in the training set, and is defined as: T0 = μ + 1.5 × σ; where μ is the mean of the Euclidean norm of all response-sensitive feature vectors, and σ is their standard deviation.

[0044] The response intensity St is calculated for the eigenvector rt of each time slice in the power grid response sensitive feature sequence, and is defined as: , where rti represents the value of the i-th dimension in the rt vector; if St≥T0, then the feature vector corresponding to the time slice t is marked as a high-response feature and added as a key node to the state evolution graph.

[0045] All labeled high-response features are sorted according to their chronological order on the time axis, and the similarity between any two high-response feature vectors is calculated using the cosine similarity formula: ;like If t1 > t2, then a directed edge is created from node t1 to t2 in the graph. The graph structure constructed in this way is the initial state evolution graph, which includes key response nodes and their temporal evolution paths.

[0046] For each key node in the state evolution graph, extract its set of neighboring nodes and the response-sensitive feature vector corresponding to the current node. Concatenate the feature vector of the current node with the average feature vector of its neighboring nodes to construct a structure-attribute fusion representation vector vn: vn = connect(rn, average neighbor vector); where the connect operation represents vector-level concatenation, and the average neighbor vector is the mean of the response feature vectors of the node's direct neighbors. This fusion vector comprehensively expresses the node's own response characteristics and its local evolution relationship in the graph structure.

[0047] The aforementioned fused representation vector set is input into the graph neural network model for embedding training. The graph neural network uses a graph convolutional network, whose core operation is: hn = activation function (weight matrix × normalized adjacency matrix × vn); where the weight matrix is ​​a learnable parameter, the activation function is a rectified linear unit (ReLU), and the embedding dimension is set to 64.

[0048] After multi-layer graph convolution operations, the embedding vector hn′ of each node is output, representing the graph topology and feature fusion representation in low-dimensional space.

[0049] Subsequently, density peaks clustering was used to perform unsupervised clustering on all embedded vectors. This method automatically identifies the centroids and divides them into clusters based on local density and minimum distance. Each cluster represents a set of nodes with similar response evolution trends.

[0050] Nodes belonging to the same clustering result are rearranged in chronological order in the original state evolution graph, while retaining their original connections, to reconstruct a complete response pattern subgraph.

[0051] Finally, for each key node in the response pattern subgraph, its contextual features (response feature vectors of multiple time slices before and after) and the input vector corresponding to that time slice (i.e., the fused feature vector ft) are extracted. An information gain-based feature evaluation method is used to quantify and rank the correlation between various input features and high-response events.

[0052] Input features whose information gain exceeds a set threshold T1 (e.g., 0.05) are marked as candidate response triggering factors, forming a set of candidate response triggering factors for subsequent prediction modeling.

[0053] The candidate response triggering factor set is input into the dynamic response prediction model M2. Combined with multi-source prediction data of the target area for future periods, the power grid response demand change curve for future periods is output, and high response risk periods are marked.

[0054] First, collect various forecast data related to the power grid operation status of the target area within the forecast period (e.g., the next 24 hours), specifically including: Electricity load forecast data: Time series forecast results constructed based on historical load curves and meteorological variables; Meteorological forecast data: including hourly forecasts of temperature, humidity, wind speed, and light intensity; New energy output forecast data: Wind power and photovoltaic output forecast values ​​are obtained by using a forecasting method based on physical modeling and time series fusion; Electricity price forecast data: Future time-of-use or real-time electricity price forecasts are estimated based on the current market clearing results and time-series trends.

[0055] The above-mentioned prediction data are organized according to a uniform time granularity (such as 15 minutes or 1 hour), and a fusion vector containing all the above features is constructed for each future time slice to form the future time period input sequence X′={x′1,x′2,...,x′t}.

[0056] Based on the extracted set of candidate response triggering factors T={f1,f2,...,fm}, feature fields that are consistent with the feature types in T are selected from the input sequence in the future time period.

[0057] By employing field matching and position mapping, feature vectors highly correlated with triggering factors in the predicted sequence are retained, and a highly correlated predicted input sequence X″={x″1,x″2,...,x″t} is constructed for inputting into the model. The dimension of each vector x″t is consistent with the set of triggering factors, ensuring the relevance of the input and the modeling effect.

[0058] This embodiment uses a Long Short-Term Memory (LSTM) neural network based on an encoder-decoder structure as the dynamic response prediction model. The specific model structure includes: The encoder part consists of two stacked LSTM networks. It receives X″ as the input sequence, performs sequence compression on the historical input features, and extracts the context-dependent hidden states. Decoder section: Using the last hidden state output by the encoder as the initial state, it progressively generates the grid response demand forecast for each future time slice, forming a response demand change forecast sequence. .

[0059] The number of hidden layer units in the LSTM is set to 128, the activation function is the hyperbolic tangent function, the mean squared error is used as the loss function during training, and the Adam optimizer is used as the optimization method.

[0060] Perform time-difference operation on the predicted result sequence Ŷ to obtain the response change gradient sequence ΔY={Δy2,Δy3,...,Δyt} between every two adjacent time slices, where: .

[0061] A response change gradient threshold G0 is set to distinguish between normal fluctuations and abnormal response growth. The threshold G0 is calculated based on the rate of change distribution of historical response events in the training data, using the 85th percentile as the set value, for example: G0 = Percentile85(ΔY_train). All values ​​in ΔY are iterated through, and segments where the gradient values ​​at multiple consecutive time points are greater than G0 are identified and marked as initial high-risk response segments. Further, for each initial high-risk segment, the trend of contextual response intensity changes is extracted within its preceding and following time ranges, and it is determined whether there are second-highest values ​​or potential secondary fluctuation characteristics. If so, the boundary is extended forward or backward to ensure that all relevant high-response periods are included. Finally, a set of high-response risk periods RH={[t1,t2],[t3,t4],...} containing time index ranges is output to support subsequent early warning issuance and load scheduling strategy formulation.

[0062] Based on the power grid response demand change curve and the power grid dispatch resource constraints, a dispatch scheme is output.

[0063] Before formulating a scheduling plan, it is necessary to clarify the available response resources and their constraints in the target area during the scheduling period, which mainly include the following categories: Adjustable load resource constraints: These include residential air conditioners, electric water heaters, and interruptible loads in industrial and commercial sectors. Each type of load has a maximum adjustable power P_max, a minimum continuous response time T_min, and a daily response frequency limit F_max. Constraints of controllable distributed energy sources: such as distributed energy storage devices (batteries) and controllable photovoltaic output, are limited by energy storage capacity, maximum charge and discharge rate, and photovoltaic output reduction tolerance. Electricity price incentives and constraints: Based on demand response market rules, the level of price guidance in different time periods is limited, and the response dispatch must not exceed the acceptable level under economic boundary conditions; Response execution latency constraints: Some devices have startup delays or execution response duration constraints, and the real-time reachability of scheduling execution needs to be considered.

[0064] The objective function of the dispatching scheme is: to minimize the load peak during high response risk periods and control user-side dispatching costs while ensuring stable grid operation. Its optimization objective can be described as: minimizing... ;in This represents the adjusted predicted load during the high-risk period t. St is the risk-weighted coefficient, with weights set according to risk levels, such as 1 for high-risk periods, 0.7 for medium-risk periods, and 0.5 for low-risk periods. Simultaneously, constraints ensure that: the scheduling power and time for each type of resource do not exceed their capacity boundaries; each scheduling operation meets response delay and scheduling interval requirements; and the overall scheduling cost (such as economic compensation) is controlled within the budget.

[0065] After constructing the objective function and constraints described above, this embodiment employs a genetic algorithm-based optimization method for scheduling. In the genetic algorithm, chromosome encoding represents the activation status and response power value of each type of resource in each time period; the fitness function is the inverse value of the objective function, along with a constraint penalty factor.

[0066] After iteration, the output optimal solution includes the scheduling resource type, participation ratio, response time, and expected load adjustment value for each high response risk period.

[0067] The following is an example of a scheduling scheme output: Time period Adjustable resource types Response power (kW) Response time (min) Startup delay (min) Scheduling priority 14:30-15:00 Commercial lighting load 1200 30 5 high 18:00-18:30 Residential air conditioning group control system 2500 20 3 middle 19:00-20:00 Distributed energy storage discharge 3000 60 0 high The above scheduling scheme can be directly used as the execution command input for the load aggregation platform or the power distribution dispatch center, while retaining the ability to link with the response demand forecast results to achieve closed-loop control of prediction-analysis-execution.

[0068] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for dynamic analysis of power grid response demand based on multi-source data fusion, characterized in that: include: Collect multi-source data within the target area, including electricity load data, renewable energy output data, meteorological data, electricity price data, and user behavior data, to form an initial dataset; The initial dataset is preprocessed to obtain a preprocessed dataset. Based on the preprocessed dataset, a multi-source feature fusion vector group F={f1,f2,...,fi,...,fn} is constructed, where each fi represents the comprehensive state feature across all data sources at a given time. The multi-source feature fusion vector group F is input into the response demand identification model, and the output is the power grid response sensitive feature sequence of the target area at the current time. Based on the power grid response sensitive feature sequence, a state evolution diagram is constructed. The state evolution diagram takes the high response features in the power grid response sensitive feature sequence as key nodes, and mines their fluctuation patterns and triggering conditions in historical time series to obtain a set of candidate response triggering factors. The candidate response triggering factor set is input into the dynamic response prediction model. Combined with multi-source prediction data of the target area for future periods, the power grid response demand change curve for future periods is output, and high response risk periods are marked. Based on the power grid response demand change curve and the power grid dispatch resource constraints, a dispatch scheme is output.

2. The method for dynamic analysis of power grid response demand based on multi-source data fusion according to claim 1, characterized in that: The comprehensive status characteristics include: load characteristics, new energy characteristics, meteorological characteristics, electricity price characteristics, and user behavior characteristics.

3. The method for dynamic analysis of power grid response demand based on multi-source data fusion according to claim 1, characterized in that: The multi-source feature fusion vector set F is input into the response requirement identification model, including: Construct a response demand identification model based on bidirectional gated loop units, input multi-source feature fusion vector group F, and extract context-related state information for each time slice; The extracted temporal features are weighted to obtain the attention distribution of the current time slice across all historical states; Generate the power grid response sensitive feature vector of the target area at the current moment based on the weighted feature representation; The power grid response-sensitive feature vectors are constructed into a response-sensitive feature sequence in chronological order.

4. The method for dynamic analysis of power grid response demand based on multi-source data fusion according to claim 1, characterized in that: Based on the power grid response-sensitive feature sequence, a state evolution diagram is constructed, including: The response intensity is calculated based on the magnitude of each feature vector in the power grid response sensitive feature sequence, and feature points exceeding a preset threshold are marked as high response features. Using high-response features as key nodes, directed connections are established based on their chronological order on the time axis and feature similarity to construct the initial graph structure of the state evolution graph. Cluster analysis is performed on the nodes in the state evolution graph to identify response pattern subgraphs with common evolution paths; By combining the contextual features of key nodes in the response pattern subgraph with the input vector of the corresponding time period, a set of candidate response triggering factors that cause significant changes in the power grid load response is extracted.

5. The method for dynamic analysis of power grid response demand based on multi-source data fusion according to claim 4, characterized in that: The clustering analysis of nodes in the state evolution graph to identify response pattern subgraphs with common evolution paths includes: For each key node in the state evolution graph, extract its structural adjacency information and the response-sensitive feature vector of the corresponding time period, and construct a node structure-attribute fusion representation vector; An embedding algorithm based on graph neural networks is used to perform low-dimensional embedding mapping on the fused representation vectors to obtain the set of representation vectors of each node in a unified space; Clustering of node embedding vectors based on density peak clustering method identifies node clusters with similar response feature evolution trends; Reconstruct the response pattern subgraph by arranging the nodes in the same clustering result according to their temporal order and their connection relationships in the graph.

6. The method for dynamic analysis of power grid response demand based on multi-source data fusion according to claim 1, characterized in that: The output curve showing the change in grid response demand over future periods includes: Acquire multi-source forecast data for the target area in the future time period, including electricity load forecast, meteorological forecast, new energy output forecast and electricity price forecast data, and construct the input sequence for the future time period; The candidate response triggering factor set is feature-aligned with the input sequence of future time periods to form a highly correlated prediction input vector sequence; A long short-term memory neural network model based on an encoder-decoder structure is constructed as a dynamic response prediction model. The prediction input vector sequence is time-series modeled to output the power grid response demand change curve in the future cycle. Based on the gradient changes in the prediction curve, time periods exceeding a set threshold are identified and marked as high-risk response periods.

7. The method for dynamic analysis of power grid response demand based on multi-source data fusion according to claim 6, characterized in that: The periods identified as high-risk response periods include: Time difference operation is performed on the predicted power grid response demand change curve to calculate the change gradient sequence between adjacent time points; Set the response change gradient threshold G0; Traverse the gradient sequence and mark the time intervals in which the gradient changes continuously and is greater than G0 as the initial high-risk intervals; By combining the contextual features before and after each initial high-risk segment, the boundary is expanded, and boundary points with significant response intensity are retained to finally determine the high-response-risk period.

Images