New energy access region power grid balance scheduling method based on artificial intelligence

Abstract

The invention discloses a new energy access region power grid balance scheduling method based on artificial intelligence, and relates to the technical field of power grid scheduling, and the method comprises the steps: data collection and preprocessing: obtaining and processing power grid multi-source data and micro-scale meteorological data; meteorological feature coding: extracting dynamic meteorological features by using a long short-term memory network; constructing a dynamic space-time hypergraph and embedding nodes, and generating node dynamic embedding in combination with a graph convolutional network; based on power grid state prediction and pre-fault analysis of causal intervention, accurate prediction and fault identification are realized; and generating and executing a pre-fault scheduling strategy, and generating and executing an optimization strategy through reinforcement learning, so that the method can realize the transformation of the power grid from response type recovery to prospective self-healing, improves the toughness, reliability and economy of the power grid in an extreme scene, and is suitable for the power grid balance scheduling of a new energy access region.

Description

Technical Field

[0001] This invention relates to the field of power grid dispatching technology, specifically to a power grid balancing dispatching method for areas with new energy access based on artificial intelligence. Background Technology

[0002] With the deepening implementation of the "dual-carbon" strategy, the penetration rate of intermittent new energy sources such as wind and solar power in the power system has significantly increased. It is estimated that by 2030, the global share of renewable energy power generation will approach 50%. The contradiction between the explosive growth of new energy installed capacity and the insufficient flexibility of the power system is becoming increasingly prominent. Its inherent randomness, volatility, and intermittency pose a huge challenge to the real-time power balance and safe and stable operation of the power grid. This problem is particularly prominent under extreme weather events (such as typhoons, hail, cold waves, etc.). Extreme weather not only causes drastic fluctuations in new energy output but may also cause grid equipment failures, further exacerbating the risks to grid operation. Building a new power system with a high proportion of new energy integration urgently requires breaking through the limitations of traditional dispatching modes and developing more intelligent, agile, and forward-looking balance dispatching technologies.

[0003] Currently, numerous studies both domestically and internationally have explored the application of artificial intelligence technology in power grid dispatching. For example, the invention patent "A Method for Coordinated Control of Source-Grid-Load-Storage under Extreme Weather" (Publication No. CN119401562A) applied for by Guangdong Power Grid Co., Ltd. includes the following technical solutions: acquiring system load, topology, power data, and data on adjustable generator units, new energy units, and energy storage units; constructing an objective function with the goal of minimizing the comprehensive costs of power generation, start-up and shutdown, wind and solar curtailment, and load shedding; and considering load constraints, capacity constraints, and node power balance constraints to generate an optimized dispatching strategy. This type of method focuses primarily on post-fault optimization dispatching, guiding dispatching decisions through an economic objective function, thus improving the economic efficiency of system operation to some extent. However, its essence remains a reactive dispatching model, lacking the ability to proactively anticipate and defend against external disturbances such as extreme weather, making it difficult to achieve dynamic coordinated optimization of the power grid topology and dispatching strategies, and exhibiting a lag in responding to sudden faults.

[0004] Therefore, the problem this invention aims to solve is: how to overcome the limitations of reactive recovery in existing technologies and achieve synergy between dynamic topology optimization and pre-fault scheduling of the power grid under extreme weather conditions, thereby completing the paradigm shift towards proactive self-healing. Specifically, this means how to utilize artificial intelligence technology to deeply integrate meteorological information with the real-time operating status of the power grid, accurately predict the potential impact of extreme weather on the power grid, and on this basis, collaboratively optimize power grid topology adjustment strategies and source-grid-load-storage scheduling commands to implement proactive defense before a fault occurs, fundamentally improving the resilience, reliability, and self-healing capabilities of high-proportion renewable energy power grids under extreme scenarios.

[0005] In summary, although existing technologies have made some progress in the field of renewable energy dispatch, there is still a lack of integrated solutions that combine accurate forecasting, dynamic topology optimization, and proactive fault-prevention dispatch to address extreme weather conditions. Developing a novel artificial intelligence dispatching method capable of achieving these functions is of paramount importance for ensuring the safe and stable operation of new power systems. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an artificial intelligence-based method for power grid balancing and dispatching in areas with new energy access. This method involves collecting multi-source power grid operation data and micro-scale meteorological data, preprocessing them, extracting dynamic meteorological feature vectors using a long short-term memory network, constructing a dynamic spatiotemporal hypergraph using power grid nodes, and generating dynamic node embedding representations using a graph convolutional network. Subsequently, a causal intervention mechanism is introduced, combined with a recurrent neural network and attention graph convolution to achieve power grid state prediction and pre-fault analysis. Finally, a pre-fault dispatching strategy is generated using a reinforcement learning agent. This enables coordinated dynamic topology optimization and pre-fault dispatching under extreme weather conditions, thereby transforming power grid dispatching from a post-fault reactive recovery to a pre-fault proactive self-healing approach.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On the one hand, a method for grid balancing and dispatching in areas with new energy access based on artificial intelligence, the method comprising the following steps:

[0008] S100. Data Acquisition and Preprocessing: Real-time acquisition of power output data from new energy power generation nodes, operation data from traditional power generation nodes, load data from power consumption nodes, power flow data from transmission nodes, and micro-scale meteorological data from meteorological departments, including temperature, humidity, wind speed, wind direction, air pressure, and precipitation.

[0009] Multi-source heterogeneous data is cleaned, aligned, and normalized to form a spatiotemporal sequence dataset with consistent timestamps;

[0010] S200, Meteorological Feature Encoding: The preprocessed microscale meteorological data is input into a meteorological encoder based on a long short-term memory network to extract the spatiotemporal features of the meteorological data and output a dynamic meteorological feature vector, which represents the potential impact of meteorological conditions on power grid nodes in a specific future time period.

[0011] S300, Dynamic Spatiotemporal Hypergraph Construction and Node Embedding: Using power generation, power consumption, and transmission nodes as hypergraph nodes, initial hyperedges are constructed based on the physical connection and electrical coupling relationship of the power grid.

[0012] By integrating dynamic meteorological feature vectors and historical time-series data of nodes, and dynamically calculating and updating hyperedge weights through a gated spatiotemporal convolution module, a dynamic spatiotemporal hypergraph that responds to real-time meteorological changes is formed.

[0013] By using graph convolutional networks to perform convolution operations on dynamic spatiotemporal hypergraphs, dynamic embedding representations of each power grid node are learned and output. These representations comprehensively reflect the state of the nodes under specific weather conditions and their dynamic interaction with surrounding nodes.

[0014] S400, Power Grid State Prediction and Pre-Fault Analysis Based on Causal Intervention: Dynamic meteorological feature vectors are dynamically embedded and represented by power grid nodes, and input into a prediction model based on recurrent neural networks and attention graph convolution;

[0015] S500, Pre-fault scheduling strategy generation and execution: Based on the prediction and analysis results of step S400, with the goal of minimizing system outage risk, frequency deviation and power generation cost, a pre-fault scheduling strategy is generated using a reinforcement learning agent, including adjusting the grid topology, allocating energy storage charging and discharging power and scheduling standby unit output.

[0016] Furthermore, in step S200:

[0017] The process of the meteorological encoder based on a long short-term memory network processing meteorological data is defined by the following formula:

[0018] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0019] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0020]

[0021]

[0022] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0023] h t =o t *tanh(C t )

[0024] Where, x t h represents the meteorological data vector input at time step t. t C represents the hidden state output at time step t. t W represents the cell state at time step t. fW i W C W o b are the weight matrices for the forget gate, input gate, candidate cell states, and output gate, respectively. f b i b C b o Let σ be the corresponding bias vector, σ represent the sigmoid activation function, * represent the Hadamard product, and h be the hidden state at the last time step. T That is, the dynamic meteorological feature vector H. weather .

[0025] Furthermore, in step S300:

[0026] The process of dynamically calculating and updating the hyperedge weights is defined by the following formula:

[0027]

[0028] Among them, w e (t) represents the dynamic weight of the hyperedge e at time step t, h v (t) represents the feature vector of node v belonging to hyperedge e at time step t, |e| represents the number of nodes connected by the hyperedge, and H weather (t) represents the dynamic meteorological feature vector at time step t, [·,·] represents the vector concatenation operation, and MLP represents a multilayer perceptron, which maps the concatenated high-dimensional vector to a scalar weight value.

[0029] Furthermore, in step S300:

[0030] The process of performing convolution operations on a dynamic spatiotemporal hypergraph using a graph convolutional network is defined by the following formula:

[0031]

[0032] Among them, Z (l) Z represents the node feature matrix of the l-th convolutional layer. (l+1) This represents the node feature matrix output by the (l+1)th convolutional layer. The incidence matrix of the hypergraph. This is a hyperedge-weighted diagonal matrix, whose diagonal elements are determined by w. e (t) constitutes, Let Θ be the degree matrix of the hypergraph. (l) Let H be the trainable weight parameter matrix of the l-th layer, σ be the non-linear activation function, and finally output the dynamic embedding representation matrix H of all nodes. grid .

[0033] Furthermore, in step S400:

[0034] The model introduces a causal intervention mechanism to separate the confounding factors between meteorological events and power grid status. It constructs a causal graph with meteorological events as the cause, power grid status as the effect, and dispatching strategies as the intervention means. It predicts the status of each node in the power grid under extreme weather conditions, including frequency deviation, voltage fluctuation and line overload risk, and identifies potential fault nodes and vulnerable lines.

[0035] The prediction model based on recurrent neural networks and attention map convolution is implemented through the following forward computation process:

[0036]

[0037]

[0038] in, This represents the hidden state of an RNN at time step t, which incorporates historical states. And the current grid node embedded H grid (t), Θ rnn Let α be the parameter of the RNN. ij Let represent the attention coefficient of node j to node i, and let a and W be the learnable parameter vector and matrix, respectively. Let h' represent the set of neighbors of node i, || denotes the vector concatenation operation, and h' i This is a new feature representation of node i after aggregating neighborhood information through a graph attention mechanism, which is ultimately used to predict the power grid state.

[0039] Furthermore, the specific steps of introducing the causal intervention mechanism include: constructing a structural causal model containing three core variables: meteorological event W, power grid state G, and dispatching strategy S, with the causal relationship being W, G, and S; intervening through do-calculus, i.e., executing do(G=g) to cut off the influence of meteorological event W on power grid state G, thereby calculating the distribution of dispatching strategy S under the condition of forced power grid state g, which is used to eliminate the confusion bias between meteorological event and power grid state, obtain a more accurate estimate of the net effect of meteorological event on power grid state, and provide an unbiased data basis for subsequent prediction.

[0040] Furthermore, in step S500:

[0041] The state space of the reinforcement learning agent is a dynamic embedded representation of power grid nodes, the action space is a pre-fault scheduling strategy, and the reward function integrates actual power outage duration, frequency stability, and scheduling economic indicators.

[0042] The optimized scheduling strategy is converted into control commands and sent to the energy management system or distributed controller for execution, so as to realize the forward-looking self-healing and balanced scheduling of the power grid in the event of extreme weather.

[0043] The reward function of the reinforcement learning agent is defined by the following formula:

[0044] R(s,a)=-(λ1·C outage (s,a)+λ2·|Δf(s,a)|+λ3·C gen (s,a))

[0045] Where s represents the current state, i.e., the dynamic embedded representation of the power grid node, a represents the current action, i.e., the pre-fault scheduling strategy adopted, and C outage (s,a) represents the estimated power outage cost in state s after action a, |Δf(s,a)| represents the absolute value of the system frequency deviation in state s after action a, and C gen (s,a) represents the total power generation cost in state s after action a is performed, and λ1, λ2, and λ3 are hyperparameters used to balance the weights of various penalty terms.

[0046] On the other hand, an electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a method.

[0047] Compared with existing technologies, this AI-based method for grid balancing and dispatching in areas with new energy access has the following advantages:

[0048] I. This invention integrates microscale meteorological data with multi-source power grid operation data. First, it uses a long short-term memory network to encode the meteorological data to extract dynamic meteorological feature vectors. Then, it constructs a dynamic spatiotemporal hypergraph based on power grid nodes and combines it with a graph convolutional network to generate dynamic node embedding representations. Subsequently, it introduces a causal intervention mechanism and combines a recurrent neural network with attention graph convolution to achieve accurate prediction of power grid status and pre-fault analysis. Finally, it generates a pre-fault scheduling strategy based on a reinforcement learning agent, which can realize the coordinated control of dynamic topology optimization and pre-fault scheduling of the power grid under extreme weather conditions. This transforms the power grid scheduling mode from the traditional post-fault response recovery to a pre-fault proactive self-healing mode. It solves the problems of lack of proactive prediction and active defense capabilities for extreme weather and the lag in responding to sudden faults in existing technologies, effectively improving the operational resilience and power supply reliability of power grids in areas with a high proportion of renewable energy access under extreme scenarios.

[0049] Second, this invention cleanses, aligns, and normalizes multi-source heterogeneous data from new energy power generation, traditional power generation, power consumption, and transmission nodes to construct a standardized spatiotemporal sequence dataset. It then dynamically updates the hypergraph and hyperedge weights using dynamic meteorological feature vectors to accurately characterize the dynamic interaction relationships between power grid nodes. Simultaneously, a causal intervention mechanism eliminates the confusion and bias between meteorological events and power grid states to improve prediction accuracy. Finally, with the goal of minimizing system outage risk, frequency deviation, and power generation costs, it generates optimized scheduling strategies through reinforcement learning. This achieves synergistic optimization of power grid scheduling accuracy and operational economy, thus avoiding the problem of low power grid operating efficiency caused by poor data quality, large prediction deviations, or a single scheduling objective. It solves the problem of balancing the safe and stable operation of the power grid with scheduling economy in existing technologies, ensuring the continuous and stable operation of the power grid under conditions of high-proportion new energy access.

[0050] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0052] Figure 1 This is a flowchart illustrating the operation of the present invention.

[0053] Figure 2 This is a flowchart of the system data processing and feature generation of the present invention;

[0054] Figure 3 This is a schematic diagram of the causal intervention prediction and scheduling mechanism of the present invention. Detailed Implementation

[0055] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0056] Example 1

[0057] like Figures 1 to 3As shown, this embodiment aims to provide a specific implementation method for power grid balancing and dispatching in areas with high-proportion renewable energy access, addressing the challenges of power balance and poor safety and stability caused by the randomness, volatility, and extreme weather disturbances of renewable energy output. This embodiment collects and preprocesses multi-source power grid and meteorological data in real time, utilizes Long Short-Term Memory (LSTM) networks to encode meteorological spatiotemporal features, constructs a dynamic spatiotemporal hypergraph, and generates dynamic embeddings of power grid nodes through graph convolutional networks. Combined with a causal intervention mechanism, it achieves accurate prediction of power grid status and pre-fault analysis. Finally, it generates pre-fault dispatching strategies through reinforcement learning, realizing the transformation of the power grid from "response-based recovery" to "proactive self-healing," improving the resilience, reliability, and operational economy of the power grid in areas with renewable energy access. Simultaneously, this embodiment also provides electronic devices adapted to this method to ensure its real-time implementation.

[0058] The specific implementation of data acquisition and preprocessing is as follows:

[0059] The core purpose of this step is to obtain high-quality data to support subsequent AI model analysis. Since the output of new energy sources (wind power, solar power, etc.) is significantly affected by meteorological conditions and has strong intermittency and randomness, and the data of traditional power generation, power consumption and transmission nodes also have multi-source heterogeneous characteristics (such as differences in sampling frequency and data format), directly using the raw data will lead to a decrease in the accuracy of subsequent models. Therefore, it is necessary to construct a standardized spatiotemporal sequence dataset through the "acquisition-preprocessing" process.

[0060] The data collection targets and sources must cover all aspects of power grid operation and meteorological influencing factors, as detailed below:

[0061] Output data of new energy power generation nodes: The data collection targets are all grid-connected points of wind farms and photovoltaic power stations within the regional power grid. Data types include real-time active power, reactive power, and output prediction deviation. The collection frequency is set to once per minute (because the output of new energy fluctuates rapidly, high-frequency collection can capture instantaneous changes). The data is uploaded to the regional power grid dispatch center database in real time through the SCADA (Supervisory Control and Data Acquisition) system of the new energy power station.

[0062] Traditional power generation node operation data: The data collection targets are the power supply nodes of traditional power sources such as thermal power units and hydropower units. The data types include real-time output, start-up and shutdown status, reserve capacity, and fuel consumption rate of the units. The collection frequency is set to 5 minutes / time (the output of traditional units is relatively stable, so there is no need for excessively high frequency). The data is connected to the dispatch center through the power plant EMS (Energy Management System).

[0063] Electricity load data: The data collection targets are typical electricity load nodes such as industrial, residential and commercial enterprises in the region. The data types include real-time active load, reactive load and load characteristic classification (such as the continuity of industrial load and the time-of-use nature of residential load). The collection frequency is set to once per minute. The data is obtained through the power distribution network electricity information collection system.

[0064] Transmission node power flow data: The data collection targets are transmission equipment such as 220kV and above transmission lines and transformers within the regional power grid. Data types include real-time active power, reactive power, current, voltage amplitude and phase angle of the lines. The collection frequency is set to 1 minute / time. The data is generated in real time through the power flow calculation module of the power grid EMS.

[0065] Microscale meteorological data: The data is collected from meteorological monitoring stations within the power grid coverage area (including dedicated meteorological stations around wind farms and photovoltaic power stations). Data types include temperature, humidity, wind speed, wind direction, air pressure, and precipitation. The collection frequency is set to 10 minutes / time (meteorological data changes relatively slowly, and this frequency can balance accuracy and data volume). The data is pushed to the dispatch center in real time through the meteorological department's microscale meteorological monitoring network.

[0066] The preprocessing workflow includes three steps: cleaning, alignment, and normalization. Each step is designed to address the deficiencies of the original data, as detailed below:

[0067] Data cleaning aims to remove outliers, missing values, and noise from the raw data. For outliers (such as sudden drops in new energy output to 0 or far exceeding rated values, or voltage amplitudes exceeding permissible ranges), the 3σ criterion is used for identification and removal (i.e., if a data value exceeds the range of "mean ± 3 standard deviations", it is considered an outlier). For missing values ​​(such as data loss in some time steps due to communication interruptions), linear interpolation is used to fill in the gaps (linear fitting is performed based on valid data before and after the missing period to ensure data continuity). For noise (such as high-frequency interference caused by sensor fluctuations), a moving average filtering method is used for smoothing (the average of five adjacent time steps is taken as the current time step data to reduce random interference).

[0068] Data alignment aims to unify timestamps across multiple data sources. Since different data sources have varying acquisition frequencies (e.g., meteorological data every 10 minutes, power grid data every 1 minute), the minimum acquisition interval for power grid data (1 minute) is used as a benchmark. For low-frequency data (e.g., meteorological data), linear interpolation is employed to interpolate timestamps, ensuring complete consistency across all data (e.g., each hour 00 corresponds to one time step). This prevents feature misalignment in subsequent models due to inconsistent time dimensions.

[0069] Data normalization aims to eliminate the dimensional differences between data of different dimensions (e.g., power in MW, temperature in ℃, voltage in kV) and prevent data with large dimensional differences from having a dominant influence on model training. The Min-Max normalization method is used to map all data to the [0,1] interval. The calculation formula is as follows:

[0070]

[0071] Where χ represents the original data, x min x max These are the historical minimum and maximum values ​​for this dimension of data (obtained based on historical data from the past year).

[0072] After the above processing, a spatiotemporal sequence dataset with consistent timestamps, no abnormal noise, and unified dimensions is formed. This dataset contains three-dimensional information of "time-node-feature", providing a data foundation for subsequent meteorological feature coding and power grid status analysis.

[0073] The core objective of this step is to extract dynamic features with grid impact correlation from preprocessed microscale meteorological data. Meteorological data is time-series data (changing dynamically over time), and the impact of different meteorological elements (such as wind speed and precipitation) on the power grid has long-term and short-term dependencies (such as continuous strong winds not only affecting current wind power output but also potentially causing subsequent line icing). Traditional time-series feature extraction methods (such as sliding windows) are difficult to capture such dependencies. Therefore, an LSTM network is used to construct a meteorological encoder to achieve effective extraction of meteorological spatiotemporal features.

[0074] The LSTM weather encoder consists of an input layer, an LSTM hidden layer (containing a forget gate, an input gate, a cell state gate, and an output gate), and an output layer. The number of neurons in the hidden layer is determined according to the dimensions of the meteorological data (e.g., if the meteorological data contains 6 elements, the number of neurons is set to 64 to ensure sufficient feature representation ability). The dataset for network training is the "microscale meteorological data - power grid output / fault correlation data" from the past 5 years (by labeling the actual impact of meteorological conditions on the power grid through historical data, a supervised training sample is constructed).

[0075] LSTM controls the forgetting and updating of information through a gating mechanism. The specific process is implemented through the following formulas, and the physical meaning and parameter descriptions of each formula are as follows:

[0076] Forget Gate Calculation:

[0077] Formula: f t =σ(W f ·[h t-1 ,x t ]+b f )

[0078] Function: Determines the cell state at the previous moment (C) t-1 Determining which information needs to be retained (or forgotten) is the core module of LSTM for capturing long-term dependencies.

[0079] Parameter description:

[0080] σ: sigmoid activation function, with an output value range of [0,1]. The closer the output is to 1, the more information is retained; the closer it is to 0, the more information is forgotten.

[0081] W f The weight matrix of the forgetting gate (dimension is "number of hidden layer neurons × (number of hidden layer neurons + meteorological data dimension)") is obtained through training and represents the contribution of "historical hidden state and current meteorological data" to the forgetting decision.

[0082] h t-1 The hidden state of the LSTM at the previous time step (t-1) contains the meteorological time series features up to time step t-1;

[0083] x t : The meteorological data vector at the current time (t) (the dimension is the number of meteorological elements, such as 6 dimensions);

[0084] [h t-1 ,x t ]: Concatenate historical hidden states with current meteorological data to form an input vector;

[0085] b f The bias vector of the forget gate (with the same dimension as the number of neurons in the hidden layer) is used to adjust the baseline value of the input to avoid gating failure due to the input being too small.

[0086] Input gate and cell state candidate value calculation:

[0087] Input gate formula:

[0088] Cell state candidate value formula:

[0089] Function: The input gate determines which information in the current meteorological data needs to be updated to the cell state, and the candidate cell state values ​​generate new information to be updated;

[0090] Parameter description:

[0091] W i b i The weight matrix and bias vector of the input gate, acting on W f b f Similarly, control the filtering of input information;

[0092] tanh: Hyperbolic tangent activation function, with an output value range of [-1, 1], maps candidate values ​​to a reasonable range to avoid numerical overflow;

[0093] W C b C The weight matrix and bias vector of cell state candidate values ​​are used to learn the feature representation of each element in the current meteorological data.

[0094] The candidate values ​​for the cell state at the current moment contain new features of the current meteorological data.

[0095] Cell status update:

[0096] formula:

[0097] Function: It integrates historical cell states with current candidate values ​​to form the cell state at the current moment, and is the "memory unit" for LSTM to store time-series information;

[0098] Symbol explanation: * represents Hadamard product (element-wise multiplication), that is, multiplying corresponding elements, realizing gating element-wise control of information — f t *C t-1 Preserve useful historical information. Add the latest information.

[0099] Output gate and hidden state update:

[0100] Output gate formula: o t =σ(W o ·[h t-1 ,x t ]+b o )

[0101] Hidden state formula: h t =o t *tanh(C t )

[0102] Function: The output gate determines which information from the cell state needs to be passed to the current hidden state h. t The hidden state is then used as the output of the LSTM for subsequent feature propagation.

[0103] Parameter description:

[0104] W o b o The weight matrix and bias vector of the output gate control the filtering of output information;

[0105] tanh(C t ): Maps the cell state to [-1,1] to avoid excessively large values ​​affecting subsequent calculations;

[0106] h t The hidden state at the current moment contains meteorological time series characteristics up to time t, and is updated iteratively with each time step.

[0107] The LSTM weather encoder performs iterative calculations step-by-step on a continuous period of weather data (e.g., the next 6 hours, with 10-minute intervals for a total of 36 time steps), and then takes the hidden state h of the last time step. T (T is the total time step) is used as the dynamic meteorological feature vector H weather The dimension of this vector is consistent with the number of neurons in the LSTM hidden layer (e.g., 64 dimensions), and its physical meaning is "the potential impact characteristics of meteorological conditions on power grid nodes within a specific future time period"—through training, H weather The relationship between meteorological elements and power grid status is implicitly included (such as the characteristic of a sudden increase in wind power output corresponding to high wind speed, and the characteristic of a decrease in line insulation corresponding to heavy precipitation), which can be directly used for subsequent supermap construction and status prediction.

[0108] The specific implementation of dynamic spatiotemporal hypergraph construction and node embedding is as follows:

[0109] The core objective of this step is to construct a topology that can characterize the dynamic relationships between power grid nodes and learn the comprehensive state characteristics of the nodes. Traditional power grid topology analysis uses ordinary graphs (which only represent pairwise connections between two nodes), but the connections between power generation, power consumption, and transmission nodes in the power grid have multi-node coupling relationships (such as a transmission line connecting multiple wind farms and load centers), and ordinary graphs cannot characterize such relationships. At the same time, changes in meteorological conditions will cause dynamic changes in the correlation strength between nodes (such as power transmission weights). Therefore, it is necessary to construct a "dynamic spatiotemporal hypergraph" and generate dynamic node embeddings through graph convolutional networks.

[0110] The construction process of a dynamic spatiotemporal hypergraph is as follows:

[0111] The supergraph nodes are new energy power generation nodes (such as wind farms and photovoltaic power stations), traditional power generation nodes (such as thermal power units), power consumption nodes (such as industrial load centers and residential transformer substations), and transmission nodes (such as transmission lines and transformers) within the regional power grid. The initial characteristics of each node are the corresponding node data (such as the output data of the power generation node and the load data of the power consumption node) after preprocessing in S100.

[0112] The initial hyperedge is constructed based on the physical connectivity of the power grid (such as the actual connection nodes of transmission lines) and electrical coupling (such as the power flow intensity and voltage correlation between nodes):

[0113] Physical connection superedge: If a 220kV transmission line connects 2 generating nodes and 3 consuming nodes, then the line is used as the superedge to connect the above 5 nodes;

[0114] Electrically coupled hyperedge: If the load fluctuations of 3 power-consuming nodes in a certain area are highly correlated (increase and decrease synchronously), and each is powered by 2 power-generating nodes, then a hyperedge covering the 3 power-consuming nodes and 2 power-generating nodes is constructed to characterize their electrical coupling relationship.

[0115] The hyperedge weight reflects the strength of the relationship between nodes and needs to be updated in real time, combining dynamic meteorological characteristics and historical time-series data of the nodes, to avoid distortion of the relationship characterization due to meteorological changes. The update process is implemented through the following formula:

[0116] formula:

[0117] Function: The feature mean of the nodes connected by the hyperedge e is concatenated with dynamic meteorological features and mapped to the dynamic weight of the hyperedge through a multilayer perceptron (MLP), so as to realize the real-time adjustment of the weight according to the meteorological and node status.

[0118] Parameter and symbol explanation:

[0119] w e (t): The dynamic weight (scalar) of hyperedge e at time step t. The larger the weight, the stronger the association between the nodes connected by the hyperedge.

[0120] |e|: The number of nodes connected by the superedge e. Let h be the mean of the eigenvectors of all nodes of hyperedge e at time step t. v (t) is the historical time-series feature vector of node v, used to balance the feature contributions of hyperedges with different numbers of nodes;

[0121] H weather (t): Dynamic meteorological feature vector at time step t (from S200), used to incorporate the influence of meteorology on node association (e.g., under strong winds, the association weight between wind power nodes and transmission nodes increases).

[0122] [·,·]: Vector concatenation operation, which concatenates the node feature mean (e.g., 10-dimensional) with the meteorological feature vector (e.g., 64-dimensional) into a high-dimensional vector (e.g., 74-dimensional);

[0123] MLP: Multilayer Perceptron (containing 1 input layer, 1 hidden layer, and 1 output layer). The hidden layer activation function is ReLU, and the output layer is a linear activation function. Its function is to map the high-dimensional concatenated vector into scalar weights (because the hyperedge weights need to be single-valued to facilitate subsequent graph convolution calculations). The parameters of the MLP are obtained by training with historical "node features - meteorological features - hyperedge association strength" data.

[0124] The core function of Graph Convolutional Networks (GCNs) is to generate dynamic embedding representations that comprehensively reflect "the node's own state, the interaction relationships with surrounding nodes, and meteorological influences" by aggregating the neighborhood information of nodes in the hypergraph. The specific convolution operation is implemented using the following formula:

[0125] formula:

[0126] Function: Perform convolution transformation on the node feature matrix of the hypergraph to gradually improve the abstraction level of the features, and finally output the node dynamic embedding matrix;

[0127] Parameter and symbol explanation:

[0128] Z (l) : The node feature matrix of the l-th convolutional layer (dimension is "number of nodes × l-th layer feature dimension"), initial layer Z (0) This is the initial feature matrix of the hypergraph nodes;

[0129] Z (l+1) The node feature matrix output by the (l+1)th convolutional layer, with feature dimensions determined by Θ. (l) The dimensions determine this;

[0130] The hypergraph's incidence matrix (dimension is "number of nodes × number of hyperedges") contains elements. This indicates that node v belongs to hyperedge e, otherwise it is 0, used to characterize the connection relationship between a node and a hyperedge;

[0131] A hyperedge weight diagonal matrix (dimension is "number of hyperedges × number of hyperedges"), where the diagonal elements are the dynamic weights w of each hyperedge. e (t), where the off-diagonal elements are 0, are used to incorporate the superedge weights into the convolution process;

[0132] The degree matrix of a hypergraph (dimension is "number of nodes × number of nodes"), diagonal elements (i.e., the sum of the weights of all hyperedges to which node v belongs), with off-diagonal elements being 0. Its function is to normalize the node features and avoid feature shifts caused by differences in the number of hyperedges to which a node belongs.

[0133] The inverse square root of the degree matrix is ​​used to achieve symmetric normalization. This ensures the numerical stability of the features after convolution;

[0134] Θ (l) The trainable weight parameter matrix of the l-th convolutional layer (with dimensions of "l-th layer feature dimension × l+1-th layer feature dimension") learns the aggregation rules of node features through training.

[0135] σ: Non-linear activation function (using ReLU), used to introduce non-linear feature mapping to improve the model's ability to express complex relationships.

[0136] After 2-3 layers of graph convolution operations, the dynamic embedding representation matrix H of all power grid nodes is output. grid (The dimension is "number of nodes × embedding dimension", such as setting the embedding dimension to 128). Each row vector in this matrix corresponds to the embedding representation of a power grid node. Its physical meaning is "the node's own state (such as output / load level) under the current weather conditions and its dynamic interaction with surrounding nodes (such as power transmission dependence)", providing core feature inputs for subsequent power grid state prediction.

[0137] The specific implementation of power grid state prediction and pre-fault analysis based on causal intervention is as follows:

[0138] The core objective of this step is to accurately predict the power grid status under extreme weather conditions and identify potential faults. It addresses the prediction bias caused by the "confusion factors between meteorological events and power grid status" in traditional prediction models (such as seasonal factors affecting both wind speed and load, leading to the model misjudging the net impact of weather on the power grid). Therefore, a causal intervention mechanism is introduced, and a prediction model is constructed by combining recurrent neural networks (RNN) and attention map convolution to achieve unbiased prediction and pre-fault analysis.

[0139] First, a structured causal model is constructed, comprising three core variables: "meteorological event (W), power grid state (G), and dispatching strategy (S)," to clarify the causal relationships between the variables.

[0140] W→G: Meteorological events directly affect the state of the power grid (such as strong winds causing a sudden increase in wind power output, and heavy rain causing line faults);

[0141] W→S: Meteorological events indirectly affect dispatch strategies (e.g., when a typhoon is predicted, dispatchers adjust standby units in advance);

[0142] S→G: Dispatch strategies directly affect the grid status (e.g., adjusting the output of standby units can mitigate frequency deviations).

[0143] Traditional forecasting models do not consider the confounding effect of W on G between direct and indirect effects (through S), which makes it impossible to accurately separate the "net effect of weather on the power grid." Therefore, causal intervention is needed to eliminate the confounding bias.

[0144] The do-calculus is used to intervene in the power grid state G. Specifically, do(G=g) (forces the power grid state to the preset value g) to cut off the direct causal path W→G of the meteorological event W to the power grid state G, and only retain the indirect path W→S→G.

[0145] By using do(G=g), the probability distribution P(S|do(G=g),W) of the dispatch strategy S can be calculated under the premise of "excluding the direct impact of meteorology on the power grid". This distribution can accurately reflect the net effect of "meteorological events affecting the power grid status through dispatch strategies", avoid prediction bias caused by confounding factors, and provide an unbiased data basis for subsequent prediction models.

[0146] The prediction model uses "dynamic meteorological feature vectors" Dynamic Embedding Matrix H” of Power Grid Nodes grid The model takes "predicted state values ​​of each node in the power grid" (including frequency deviation, voltage fluctuation, and line overload risk) and "identification results of potential fault nodes / vulnerable lines" as inputs and outputs as outputs. The model structure and calculation process are as follows:

[0147] The RNN module is used to capture the temporal changes in the power grid state (such as the cumulative trend of frequency deviation over time). The calculation process is implemented through the following formula:

[0148] formula:

[0149] Function: To capture the temporal dependencies of the power grid state by fusing the hidden state of the RNN in the previous time step with the node embedding in the current time step;

[0150] Parameter description:

[0151] The hidden state of the RNN at time step t (dimension 128) contains the power grid time series features up to time t;

[0152] The hidden state of the RNN at the previous moment;

[0153] H grid (t): Dynamic embedding matrix of power grid nodes at time step t;

[0154] Θ rnn The trainable parameters of the RNN (including the weight matrix and bias vector) are obtained by training with historical "node embedding-power grid state" data;

[0155] [·,·]: Concatenates the historical hidden states of the RNN with the current node embedding to form the input vector.

[0156] The attention map convolution module focuses on nodes that are more important to the prediction result (such as transmission nodes with high overload risk), improving prediction accuracy. The calculation process includes two steps: attention coefficient calculation and feature aggregation.

[0157] Attention coefficient calculation:

[0158] formula:

[0159]

[0160] Function: Calculate the attention coefficient of node j to node i, which represents the contribution of the information of node j to the state prediction of node i;

[0161] Parameter and symbol explanation:

[0162] α ij : The attention coefficient of node j to node i, with a value range of [0,1], and the summation is 1 (normalized);

[0163] LeakyReLU: An activation function, expressed as LeakyReLU(x) = max(0.01x,x), which solves the gradient vanishing problem of ReLU on the negative half-axis and retains more node feature information;

[0164] a: Trainable attention parameter vector (dimension 256);

[0165] W: A trainable feature transformation matrix (with dimensions of "embedding dimension × embedding dimension", such as 128 × 128), used to embed nodes h. i h j Perform linear transformations to improve feature discrimination.

[0166] [·‖·]: Vector concatenation operation, concatenating Wh... i With Wh j Concatenate them into a 256-dimensional vector;

[0167] N i : The set of neighboring nodes of node i (determined based on the connection relationships of the hypergraph).

[0168] Node feature aggregation:

[0169] formula:

[0170] Function: Aggregate the features of neighboring nodes based on the attention coefficient to generate the final feature representation of node $i$;

[0171] Parameter description:

[0172] h' i The aggregated feature vector of node i (dimension 128) contains the association features between itself and its important neighbors;

[0173] σ: ReLU activation function, which introduces nonlinearity.

[0174] Hide the RNN state With the aggregated node features h' i The system is assembled, inputting a fully connected layer and outputting the predicted state values ​​for each node in the power grid.

[0175] Frequency deviation prediction: Outputs the absolute value of the frequency deviation of each node from the rated value (e.g., 50Hz), in Hz;

[0176] Voltage fluctuation prediction: Outputs the range of voltage amplitude variation at each node, in kV;

[0177] Line overload risk prediction: Output the ratio of the actual power to the rated power of each transmission node (line). The closer the ratio is to 1, the higher the overload risk.

[0178] Pre-fault analysis based on prediction results:

[0179] Potential fault node identification: Nodes with frequency deviations exceeding 0.5Hz or voltage fluctuations exceeding 5% are identified as "potential fault nodes";

[0180] Vulnerable line identification: Lines with an overload risk ratio exceeding 0.9 are identified as "vulnerable lines" and sorted in descending order of risk ratio to provide a priority basis for subsequent scheduling strategies.

[0181] The generation and execution of pre-fault scheduling strategies and their implementation in electronic devices are as follows:

[0182] The core objective of this step is to generate and execute a pre-fault scheduling strategy that minimizes grid risks and costs, while providing electronic devices adapted to this method to ensure the real-time implementation of the strategy. Since grid scheduling is a dynamic multi-objective optimization problem (which requires minimizing power outage risk, frequency deviation, and generation costs simultaneously), traditional optimization algorithms (such as linear programming) are difficult to cope with dynamic environments. Therefore, reinforcement learning is used to construct an intelligent agent to achieve autonomous optimization of the strategy.

[0183] Design of reinforcement learning agents

[0184] The core elements of a reinforcement learning agent include a state space, an action space, and a reward function, which are designed as follows:

[0185] The state space is defined as "dynamic embedding matrix of power grid nodes". "Power Grid State Prediction", Dimensions and H grid And the sum of the dimensions of the predicted values ​​must be consistent (e.g., if the number of nodes is 100, H). grid The dimension is 100×128, the predicted value dimension is 100×3, and the total dimension of the state space is 100×131. This state space can comprehensively represent the "interaction relationship between power grid nodes - current predicted state", providing the agent with accurate environmental perception.

[0186] The action space is defined as a set of pre-fault scheduling strategies, containing three core types of actions:

[0187] 1. Power grid topology adjustment: such as disconnecting vulnerable lines with high overload risk and putting them into backup lines to avoid line failures;

[0188] 2. Energy storage charging and discharging power allocation: such as instructing the wind power-supporting energy storage system to discharge, to mitigate the power gap caused by a sudden drop in wind power output;

[0189] 3. Standby unit output scheduling: If instructed, thermal power units can increase their output to the upper limit of standby capacity to supplement the fluctuations in new energy output.

[0190] The range of values ​​for each action is determined based on the actual parameters of the power grid (e.g., the energy storage charging and discharging power does not exceed the rated power of the energy storage system, and the output of the standby unit does not exceed the standby capacity).

[0191] The reward function is used to evaluate the merits of actions, with the objective of "minimizing power outage risk, frequency deviation, and generation cost," and is designed as a negative weighted sum:

[0192] Formula: R(s,a)=-(λ1·C outage (s,a)+λ2·|Δf(s,a)|+λ3·C gen (s,a))

[0193] Function: The lower the cost of an action, the higher the reward, guiding the agent to learn the optimal strategy;

[0194] Parameter and symbol explanation:

[0195] s: Current state (from the state space);

[0196] a: Current action (from the action space);

[0197] C outage (s,a): Estimated power outage cost (in ten thousand yuan) under state s after action a is performed, calculated based on the load size of the potential fault node and the duration of the power outage;

[0198] |Δf(s,a)|: The absolute value of the system frequency deviation in state s after action a (in Hz);

[0199] C gen (s,a): The total cost of power generation in state s after action a (in ten thousand yuan), including the fuel cost of traditional generators and the charging and discharging cost of energy storage;

[0200] λ1, λ2, λ3: Weight hyperparameters (within the range of [0,1], and λ1+λ2+λ3=1), used to balance the three optimization objectives—for example, under extreme weather conditions, λ1 can be set to 0.5 (prioritizing the reduction of power outage risk), λ2 to 0.3 (secondary priority to ensure frequency stability), and λ3 to 0.2 (last consideration to consider economic efficiency).

[0201] The agent is trained using the Deep Deterministic Policy Gradient (DDPG) algorithm. The training dataset consists of historical data from the past three years on "grid status - dispatch actions - actual costs". The training process is as follows:

[0202] 1. Initialize the policy network and value network parameters of the agent;

[0203] 2. Randomly sample "state-action-reward-next state" samples from the experience replay pool (which stores historical training samples);

[0204] 3. Calculate the value of actions using a value network and update the policy network parameters to maximize expected reward;

[0205] 4. Iterate training until the reward function converges (e.g., the reward value fluctuates by less than 5% over 1000 consecutive training steps).

[0206] After training, the agent can adjust its state based on the real-time input state s(H). grid (Based on the predicted value), the optimal pre-fault scheduling strategy a is generated autonomously.

[0207] Convert the pre-fault scheduling strategy generated by the agent into standardized control instructions:

[0208] Power grid topology adjustment instructions: Issued to the SCADA system of the regional power grid dispatch center to control the opening and closing of line circuit breakers;

[0209] Energy storage charging and discharging commands: These are sent to the local controller of the energy storage system, specifying the charging and discharging power and duration.

[0210] Standby unit output command: Issued to the EMS system of traditional generation nodes to adjust the unit output setpoint.

[0211] The execution cycle of control commands is consistent with the data acquisition cycle (e.g., 1 minute / time) to ensure that the strategy can respond in real time to the dynamic changes of the power grid and the weather.

[0212] This electronic device is used to perform the calculations and issue instructions for the aforementioned scheduling method. Its hardware components and functions are as follows:

[0213] Storage: High-capacity solid-state drives (SSD, capacity ≥ 1TB) and random access memory (RAM, capacity ≥ 64GB) are used. SSD is used to store computer programs (code to implement steps S100-S500) and historical datasets, while RAM is used to store data during real-time calculations (such as node embedding matrices and prediction results) to ensure data read and write speed.

[0214] Processor: Employs a multi-core central processing unit (CPU, such as Intel Xeon Gold 6348) and a graphics processing unit (GPU, such as NVIDIA A100). The CPU is responsible for data preprocessing and instruction issuance, while the GPU is responsible for parallel computation of models such as LSTM, graph convolution, and reinforcement learning, thereby improving the efficiency of the method (such as reducing the model computation time from seconds to milliseconds to meet real-time scheduling requirements).

[0215] Communication module: Adopting industrial Ethernet (transmission rate ≥1000Mbps), it enables data interaction with new energy power plants, traditional power plants, power consumption nodes, and meteorological departments, ensuring the reliability of real-time data acquisition and command issuance.

[0216] The software system of the electronic device is developed based on the Linux operating system, integrating Python (for model development, relying on the TensorFlow / PyTorch framework) and C++ (for real-time data processing). The software modules include a data acquisition module, a model calculation module, a policy generation module, and an instruction issuance module. These modules work together to ensure the complete execution of the scheduling method.

[0217] In summary, this embodiment fully implements an AI-based grid balancing and dispatching method for new energy access areas through seven core steps: First, a high-quality spatiotemporal sequence dataset is constructed in S100 to lay the foundation for subsequent analysis; in S200, LSTM is used to extract dynamic meteorological features and establish the correlation between meteorology and the power grid; in S300, a dynamic spatiotemporal hypergraph is constructed and node embeddings are generated to accurately characterize the dynamic correlation of power grid nodes; in S400, a causal intervention mechanism is introduced, combining RNN and attention graph convolution to achieve unbiased prediction of power grid status and pre-fault identification; in S500, the optimal pre-fault dispatching strategy is generated and executed through reinforcement learning; finally, the real-time implementation of the electronic equipment guarantee method is achieved.

[0218] Example 2

[0219] like Figure 1 As shown in Example 1, this example elaborates on the specific steps of the artificial intelligence-based power grid balancing and dispatching method for new energy access areas during operation. The specific steps are as follows:

[0220] 1. Data Acquisition and Preprocessing

[0221] Real-time data collection includes output data from new energy power generation nodes (wind farms, photovoltaic power stations), operation data from traditional power generation nodes (thermal power, hydropower), load data from power consumption nodes, power flow data from transmission nodes, and microscale meteorological data (including temperature, humidity, wind speed, wind direction, air pressure, and precipitation) provided by the meteorological department.

[0222] The collected multi-source heterogeneous data is cleaned, outliers and missing values ​​are removed, and timestamp alignment and normalization are performed to form a unified spatiotemporal sequence dataset.

[0223] 2. Meteorological Feature Coding

[0224] The preprocessed meteorological data is input into a meteorological encoder based on LSTM (Long Short-Term Memory) network to extract the spatiotemporal features of the meteorological data.

[0225] Output dynamic meteorological feature vectors to characterize the potential impact of meteorological conditions on various nodes of the power grid over a future period.

[0226] 3. Dynamic Spatiotemporal Hypergraph Construction and Node Embedding

[0227] Using power generation, power consumption, and transmission nodes as hypergraph nodes, initial hyperedges are constructed based on the physical connections and electrical coupling relationships of the power grid.

[0228] By combining dynamic meteorological characteristics and historical node data, a dynamic spatiotemporal hypergraph that responds to meteorological changes is constructed by dynamically calculating and updating hyperedge weights through a gated spatiotemporal convolution module.

[0229] By using Graph Convolutional Networks (GCNs) to perform convolution operations on the hypergraph, dynamic embedding representations of each node are learned and output, reflecting the state of the node under specific weather conditions and its interaction with surrounding nodes.

[0230] 4. Power grid state prediction and pre-fault analysis based on causal intervention

[0231] Dynamic meteorological feature vectors and nodes are embedded into the prediction model based on RNN and attention map convolution.

[0232] By introducing a causal intervention mechanism, a causal graph of meteorological events, power grid status, and dispatching strategies is constructed to separate confounding factors and predict the status of each node in the power grid under extreme weather conditions (such as frequency deviation, voltage fluctuation, and line overload risk).

[0233] Identify potential fault nodes and vulnerable lines to provide a basis for pre-fault scheduling.

[0234] 5. Generation and execution of pre-fault scheduling strategies

[0235] Based on the prediction results, with the goal of minimizing power outage risk, frequency deviation and generation cost, a reinforcement learning agent is used to generate a pre-fault scheduling strategy, including adjusting the grid topology, allocating energy storage charging and discharging power, and scheduling the output of standby units.

[0236] The generated scheduling strategy is converted into control commands, which are then issued and executed through the energy management system or distributed controller, enabling the power grid to proactively defend against and self-heal in the face of extreme weather.

[0237] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A power grid balancing and dispatching method for new energy access areas based on artificial intelligence, characterized in that, The method includes the following steps: S100, Data Acquisition and Preprocessing: Real-time acquisition of power output data from new energy power generation nodes, operation data from traditional power generation nodes, load data from power consumption nodes, power flow data from transmission nodes, and micro-scale meteorological data from meteorological departments in the regional power grid; Multi-source heterogeneous data is cleaned, aligned, and normalized to form a spatiotemporal sequence dataset with consistent timestamps; S200, Meteorological Feature Encoding: The preprocessed microscale meteorological data is input into a meteorological encoder based on a long short-term memory network to extract the spatiotemporal features of the meteorological data and output a dynamic meteorological feature vector, which represents the potential impact of meteorological conditions on power grid nodes in a specific future time period. S300, Dynamic Spatiotemporal Hypergraph Construction and Node Embedding: Using power generation, power consumption, and transmission nodes as hypergraph nodes, initial hyperedges are constructed based on the physical connection and electrical coupling relationship of the power grid. By integrating dynamic meteorological feature vectors and historical time-series data of nodes, and dynamically calculating and updating hyperedge weights through a gated spatiotemporal convolution module, a dynamic spatiotemporal hypergraph that responds to real-time meteorological changes is formed. By using graph convolutional networks to perform convolution operations on dynamic spatiotemporal hypergraphs, dynamic embedding representations of each power grid node are learned and output. These representations comprehensively reflect the state of the nodes under specific weather conditions and their dynamic interaction with surrounding nodes. S400, Power Grid State Prediction and Pre-Fault Analysis Based on Causal Intervention: Dynamic meteorological feature vectors are dynamically embedded and represented by power grid nodes, and input into a prediction model based on recurrent neural networks and attention graph convolution; S500, Pre-failure scheduling strategy generation and execution: Based on the prediction and analysis results of step S400, a pre-failure scheduling strategy is generated using a reinforcement learning agent with the goal of minimizing system power outage risk, frequency deviation and power generation cost.

2. The method for power grid balancing and dispatching in areas with new energy access based on artificial intelligence according to claim 1, characterized in that, In step S200: The process of the meteorological encoder based on a long short-term memory network processing meteorological data is defined by the following formula: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) the t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t *fishy(C) t ) Where, x t h represents the meteorological data vector input at time step t. t C represents the hidden state output at time step t. t W represents the cell state at time step t. f W i W C W o b are the weight matrices for the forget gate, input gate, candidate cell states, and output gate, respectively. f b i b C b o Let σ be the corresponding bias vector, σ represent the sigmoid activation function, * represent the Hadamard product, and h be the hidden state at the last time step. T That is, the dynamic meteorological feature vector H. weather .

3. The method for power grid balancing and dispatching in areas with new energy access based on artificial intelligence according to claim 1, characterized in that, In step S300: The process of dynamically calculating and updating the hyperedge weights is defined by the following formula: Among them, w e (t) represents the dynamic weight of the hyperedge e at time step t, h v (t) represents the feature vector of node v belonging to hyperedge e at time step t, |e| represents the number of nodes connected by the hyperedge, and H weather (t) represents the dynamic meteorological feature vector at time step t, [·,·] represents the vector concatenation operation, and MLP represents a multilayer perceptron, which maps the concatenated high-dimensional vector to a scalar weight value.

4. The method for power grid balancing and dispatching in areas with new energy access based on artificial intelligence according to claim 1, characterized in that, In step S300: The process of performing convolution operations on a dynamic spatiotemporal hypergraph using a graph convolutional network is defined by the following formula: Among them, Z (l) Z represents the node feature matrix of the l-th convolutional layer. (l+1) This represents the node feature matrix output by the (l+1)th convolutional layer. The incidence matrix of the hypergraph. This is a hyperedge-weighted diagonal matrix, whose diagonal elements are determined by w. e (t) constitutes, Let Θ be the degree matrix of the hypergraph. (l) Let H be the trainable weight parameter matrix of the l-th layer, σ be the non-linear activation function, and finally output the dynamic embedding representation matrix H of all nodes. grid .

5. The method for power grid balancing and dispatching in areas with new energy access based on artificial intelligence according to claim 1, characterized in that, In step S400: The model introduces a causal intervention mechanism to separate the confounding factors between meteorological events and power grid status, constructs a causal graph with meteorological events as the cause, power grid status as the effect, and dispatching strategies as the intervention means, predicts the status of each node in the power grid under extreme weather conditions, and identifies potential fault nodes and vulnerable lines. The prediction model based on recurrent neural networks and attention map convolution is implemented through the following forward computation process: in, This represents the hidden state of an RNN at time step t, which incorporates historical states. And the current grid node embedded H grid (t), Θ rnn Let α be the parameter of the RNN. ij Let represent the attention coefficient of node j to node i, and let a and W be the learnable parameter vector and matrix, respectively. Let h' represent the set of neighbors of node i, || denotes the vector concatenation operation, and h' i This is a new feature representation of node i after aggregating neighborhood information through a graph attention mechanism, which is ultimately used to predict the power grid state.

6. The method for power grid balancing and dispatching in areas with new energy access based on artificial intelligence according to claim 5, characterized in that, The specific steps of introducing the causal intervention mechanism include: constructing a structural causal model, which contains three core variables: meteorological event W, power grid state G, and dispatching strategy S, with the causal relationship being W, G, and S; intervening through do-calculus, i.e., executing do(G=g) to cut off the influence of meteorological event W on power grid state G, thereby calculating the distribution of dispatching strategy S under the condition of forced power grid state g, which is used to eliminate the confusion bias between meteorological event and power grid state.

7. The method for power grid balancing and dispatching in areas with new energy access based on artificial intelligence according to claim 1, characterized in that, In step S500: The state space of the reinforcement learning agent is a dynamic embedded representation of power grid nodes, the action space is a pre-fault scheduling strategy, and the reward function integrates actual power outage duration, frequency stability, and scheduling economic indicators. The optimized scheduling strategy is converted into control commands and sent to the energy management system or distributed controller for execution, so as to realize the forward-looking self-healing and balanced scheduling of the power grid in the event of extreme weather. The reward function of the reinforcement learning agent is defined by the following formula: R(s,a)=-(λ1·C outage (s,a)+λ2·|Δf(s,a)|+λ3·C gen (s,a)) Where s represents the current state, i.e., the dynamic embedded representation of the power grid node, a represents the current action, i.e., the pre-fault scheduling strategy adopted, and C outage (s,a) represents the estimated power outage cost in state s after action a, |Δf(s,a)| represents the absolute value of the system frequency deviation in state s after action a, and C gen (s,a) represents the total power generation cost in state s after action a is performed, and λ1, λ2, and λ3 are hyperparameters used to balance the weights of various penalty terms.

8. An electronic device, applicable to the artificial intelligence-based power grid balancing and dispatching method for new energy access areas as described in any one of claims 1-7, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor.

Images