Power grid dispatching method based on emergencies and graph structures

Abstract

The invention discloses a power grid dispatching method based on emergencies and a graph structure. The method comprises the following steps: establishing a baseline load prediction model by collecting historical electrical load data; recognizing emergencies influencing the power utilization demand in real time; and in combination with real-time power consumption behavior data, constructing a time-varying power consumption disturbance factor, and dynamically correcting load prediction. Geographic distribution and historical power generation data of new energy power stations such as wind power and photovoltaic power stations are acquired, a power generation graph structure model is constructed, the power generation amount is independently predicted by adopting a long-short-term memory network, and a prediction result is corrected in combination with a graph structure. And based on the corrected load and power generation prediction, considering the capacity and the energy state of the energy storage system, optimizing multi-time-sequence-scale energy scheduling, and formulating an energy storage charging and discharging and generator set operation scheme. According to the method, the dynamic influence of emergencies and the spatial relevance of new energy are fused, the load prediction precision and the new energy power generation prediction accuracy are remarkably improved, more reasonable power grid dispatching is achieved, and the safety and stability of a power grid are enhanced.

Description

Technical Field

[0001] This invention belongs to the field of power grid dispatching technology, specifically relating to a power grid dispatching method based on emergencies and graph structures. Background Technology

[0002] With the increasing proportion of renewable energy generation, the operation and dispatch of the power grid are facing more and more challenges. Renewable energy generation such as wind power and photovoltaic power are characterized by strong volatility, high uncertainty, and wide spatial distribution, which brings great difficulty to traditional power grid load forecasting and energy dispatch. At the same time, sudden events (such as natural disasters and large-scale public events) have a significant impact on electricity demand, leading to abrupt changes in load, which further exacerbates the complexity of load forecasting.

[0003] In the prior art, Chinese patent CN118971200A discloses a dynamic evaluation system and method for power supply capacity under new energy load conditions, relating to the field of smart grid technology. The system includes a power load prediction module that predicts power load and generates a load prediction time series; a power generation demand analysis module that analyzes power generation demand based on the load prediction time series and frequent power allocation losses, generating a preset power generation time series; a prediction model building module that establishes a hybrid power generation prediction model for the new energy power grid; an energy dispatch fitting module that, within a preset evaluation time zone, uses the preset power generation time series as the target and performs energy dispatch fitting based on the hybrid power generation prediction model to generate an energy dispatch time series; and a power supply capacity evaluation module that compares the energy dispatch time series with the preset power generation time series to generate a power supply capacity evaluation result.

[0004] However, current load forecasting technologies primarily rely on historical data and statistical models, making it difficult to accurately capture the dynamic impact of sudden events on electricity demand in real time, resulting in significant discrepancies between forecasts and actual loads. New energy power generation forecasts are mostly based on single-point time-series data, neglecting the spatial correlation between different power generation sites and the heterogeneity of meteorological conditions, thus failing to fully reflect complex power generation behavior. Existing scheduling methods typically fail to effectively integrate multi-source heterogeneous data and lack the ability to model the global spatiotemporal characteristics of the power grid, limiting the real-time optimization effectiveness of scheduling schemes. When dealing with complex systems involving multiple time scales and multivariate coupling, scheduling models suffer from high computational complexity and insufficient real-time scheduling capabilities, making it difficult to meet the requirements for safe and stable power grid operation under the background of high new energy penetration.

[0005] Therefore, existing technologies still have significant shortcomings in terms of the accuracy of load and new energy power generation forecasting, the ability to respond to emergencies, and the dynamic adaptability of dispatching strategies. There is an urgent need to study new methods to improve the intelligence level and operational efficiency of power grid dispatching. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a power grid dispatching method based on sudden events and graph structure.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] This invention provides a power grid dispatching method based on sudden events and graph structure, comprising the following steps:

[0009] Obtain historical electricity load data of the target power grid supply area within a preset continuous time period, and establish a baseline load prediction model based on the historical electricity load data;

[0010] By utilizing social media data, geolocation service trajectory data, and public event databases, we continuously monitor and identify emergencies that affect electricity demand;

[0011] Based on the dynamic identification and time-varying feature extraction of sudden events in a continuous time series, combined with real-time electricity consumption behavior data, a time-varying electricity consumption disturbance factor is constructed, and dynamically corrected in combination with the baseline load forecasting model to generate a disturbance-adjusted load forecasting time series.

[0012] Obtain geographical distribution information and historical power generation data of new energy power generation sites, including wind power and photovoltaic power generation sites, and construct a multi-node power generation graph structure model;

[0013] The power generation of each new energy power generation site is predicted independently using a long short-term memory network model. The power generation prediction results of each new energy power station are then corrected using the power generation graph structure model to obtain the corrected new energy power generation prediction time series.

[0014] By combining the load forecast timing with the new energy power generation forecast timing, and based on the energy storage system capacity, charging and discharging efficiency and energy state parameters, the energy dispatching at multiple time scales within the preset continuous time period is optimized to generate energy storage charging and discharging strategies and generator start-up and output schemes.

[0015] Furthermore, the establishment of a baseline load forecasting model based on the historical electricity load data specifically includes:

[0016] Collect historical electricity load data of the target power grid supply area within a preset continuous time period, and construct a load time series dataset arranged in time series;

[0017] The historical electricity load data is cleaned and normalized to remove outliers and missing items.

[0018] The processed load time series dataset is divided into multiple sample sequences using a sliding time window approach. Each sample contains a fixed-length historical load input sequence and the corresponding predicted target value.

[0019] Based on the sample sequence, a Long Short-Term Memory (LSTM) network model is constructed. The historical load input sequence is used to predict future load values. During the training process, the mean square error between the predicted value and the actual load value is used as the loss function.

[0020] Divide the model into training and validation sets, optimize the model parameters through multiple rounds of iterative training, and use the validation set to evaluate the prediction accuracy to avoid overfitting.

[0021] After training is completed, a baseline load prediction model based on the Long Short-Term Memory (LSTM) network model is obtained.

[0022] Furthermore, the construction of a time-varying electricity consumption disturbance factor based on the dynamic identification and time-varying feature extraction of sudden events within a continuous time series, combined with real-time electricity consumption behavior data, specifically includes:

[0023] A training dataset containing historical incident samples was collected, including the incident occurrence time t. e Area of ​​influence R e Duration Δt e Event Type e Social media popularity H e Change in pedestrian flow intensity ΔP e and the load change ΔL in the target area before and after the event i (t);

[0024] Construct the input feature vector based on samples of sudden events:

[0025]

[0026] Among them, X e (t) represents the input feature vector of the sudden event e at time t, and norm(·) represents the normalization process for the variables. The area R affected by the emergency is indicated by the symbol e. e With the target power grid supply area A i Spatial overlap ratio;

[0027] The corresponding historical load change ΔL i (t) serves as the supervision label, and a perturbation prediction model is constructed. Where f θ The nonlinear model with trainable parameters is represented by a multilayer perceptron (MLP) and trained by minimizing the loss function of the perturbation prediction model.

[0028] After the model training is completed, in actual scheduling, for each sudden event identified within the continuous time series [t0, t0+ΔT], its input features are constructed in real time and input into the model to predict the time-varying power consumption disturbance factor at the current moment:

[0029]

[0030] in, Supplying power to the target power grid area A i The time-varying power disturbance factor in a continuous time series [t0, t0+ΔT] is the load change caused by a sudden event in the continuous time series [t0, t0+ΔT].

[0031] Furthermore, the loss function of the disturbance prediction model is:

[0032]

[0033] in, The loss function for the perturbation prediction model is defined by indices e, i, t, which iterate through all events, regions, and times. θ (X e (t) represents the load change predicted by the model, ΔL i (t) represents the actual load change.

[0034] Furthermore, the generation of the disturbance-adjusted load forecast time series specifically includes:

[0035] Load forecast sequence output from baseline load forecasting model As the initial predicted load, i represents the target grid supply area, and t represents a time in the time series;

[0036] By utilizing the disturbance prediction model obtained through training and combining it with the characteristics of sudden events identified within continuous time series, the time-varying power disturbance factor sequence of the target area can be predicted in real time. in This represents the predicted load disturbance caused by a sudden event at time t;

[0037] By weighted summation, the disturbance factor is added to the baseline load forecast sequence to obtain the disturbance-adjusted load forecast sequence:

[0038]

[0039] in, Let α be the load forecast value for the target power grid supply area i at time t, and α be a learnable weight parameter. The forecast value is trained and optimized by minimizing the error between the forecast value after disturbance adjustment and the actual load data.

[0040] Furthermore, the acquisition of geographical distribution information and historical power generation data of new energy power generation sites, including wind power and photovoltaic power generation sites, and the construction of a multi-node power generation map structure model, specifically includes:

[0041] Collect the geographical location information of the new energy power generation sites and their respective power grid areas, and construct a node set {v1, v2, ..., v N}, where each node v i For each new energy power generation site, the node attributes include the node's historical power generation data;

[0042] Based on the spatial distance d between nodes ij Historical power generation sequence correlation corr(P) i ,P j ) and the similarity of real-time meteorological conditions sin(C i C j Construct a weighted adjacency matrix W = [w ij ] N×N The weight calculation formula is:

[0043]

[0044] Among them, w ij Represents node v i With node v j Edge weights between; I(d) ij ≤d th ) is an indicator function, when d ij Less than or equal to the threshold d th The value is 1 if the distance is constant and 0 otherwise, controlling the connection range of nodes in the graph; σ is the Gaussian kernel bandwidth parameter, used to adjust the decay rate of the distance weight; corr(P i ,P j ) is node v i With node v j Pearson correlation coefficient of historical power generation time series, P i 、P j They are nodes v i With node v j Historical power generation time series, sin(C i C j ) is node v i With node v i The similarity of feature vectors of real-time meteorological conditions at the corresponding time; C i C j They represent nodes v respectively i With node v j The real-time meteorological condition feature vector; β1, β2, and β3 are learnable fitting coefficients, and N is the total number of new energy power generation sites.

[0045] Furthermore, the real-time meteorological condition feature vector includes wind speed, wind direction, solar radiation intensity, and cloud cover index.

[0046] Furthermore, the learnable fitting coefficients β1, β2, and β3 are jointly optimized using the following fitting coefficient objective function:

[0047]

[0048] Among them, P j (t) represents the actual power generation of node j at time t, and N(j) represents the set of neighboring nodes of node j.

[0049] Furthermore, the step of correcting the power generation prediction results of each new energy power station using the power generation diagram structure model to obtain the corrected new energy power generation prediction time series specifically includes:

[0050] For each new energy power generation site v j Based on its historical power generation time series and real-time meteorological data, a long short-term memory network model is used to independently predict its preliminary power generation sequence.

[0051] Based on the constructed power generation diagram structure model, the weighted adjacency matrix W = [w ij ] N×N compute node v j The formula for the power generation correction value at time t is:

[0052]

[0053] in, For new energy power generation sites v j Corrected power generation at time t;

[0054] Based on preliminary power generation sequence With corrected power generation The revised forecast time series for renewable energy generation has been obtained:

[0055]

[0056] in, For node v j The revised forecast of new energy power generation at time t, where γ is the fusion weighting coefficient, is determined through training or cross-validation.

[0057] Furthermore, the weighted adjacency matrix W = [w ij ] N×N After each dispatch, the data is dynamically updated based on the latest actual power generation data and meteorological observation data through the following steps:

[0058] Collect the actual power generation of each node within the latest time period and the corresponding real-time meteorological data C j (t);

[0059] Based on the collected actual power generation The fitting coefficients β1, β2, and β3 can be learned by jointly optimizing the objective function of minimizing the prediction error.

[0060] Using the updated fitting coefficients β1, β2, β3, the weighted calculation formula is applied to the actual power generation. and the corresponding real-time meteorological data C j (t) Update the weighted adjacency matrix W of the power generation graph structure model.

[0061] Compared with the prior art, the present invention has the following advantages:

[0062] (1) This invention integrates social media data, geographic trajectory data and public event database to dynamically identify and model emergencies, which solves the problem that the existing power grid load forecasting system lacks a real-time perception mechanism for emergencies and cannot effectively capture the impact of emergencies on electricity consumption behavior. It realizes the ability to discover and model the impact of abnormal events such as natural disasters and large-scale events on electricity load in real time. It has the advantages of diverse perception channels, strong data timeliness and good adaptability, which significantly improves the power grid dispatching response to emergencies.

[0063] (2) This invention constructs a time-varying power disturbance factor and uses an MLP model to predict load disturbances. This solves the technical difficulty that existing methods can only predict historical loads and cannot predict the short-term impact of sudden events on power loads at the time of occurrence. It realizes the quantitative modeling of the impact of different types of sudden events (such as high temperature, work stoppage, and holiday activities) on load in the spatial and temporal dimensions. It has the advantages of dynamic updating of prediction results, strong learning ability, and adaptability to different regions and different event types, providing a real-time intervention mechanism for load prediction.

[0064] (3) This invention generates a load prediction sequence after disturbance correction by weighted fusion of disturbance factor and baseline load prediction sequence. This solves the problems of existing methods such as the inability to adjust prediction curve, response lag and result distortion when facing sudden disturbances. It realizes the introduction of a dynamic disturbance correction mechanism into the prediction model, so that the prediction results can be updated in real time as the event changes. It has the advantages of significantly improved prediction accuracy, trainable and optimizable fusion mechanism and dynamic adaptation to multiple event superposition disturbances, thus improving the prediction flexibility and accuracy of the system.

[0065] (4) This invention constructs a multi-node graph structure model containing new energy power generation sites. The edge weights between nodes are combined with spatial distance, power generation correlation, and meteorological similarity. Based on a graph-structured weighted adjacency matrix, the prediction results are corrected and fused. This solves the problems of independent modeling of each site in existing new energy power generation predictions, neglecting spatial and meteorological correlations, which leads to amplified prediction errors. It also solves the problem of the original LSTM model being sensitive to local abnormal weather and prone to overfitting when predicting single points. It realizes an information propagation mechanism based on graph structure, enhances the complementary modeling between sites, integrates and corrects prediction results from multiple neighboring sites, and enhances the overall stability and accuracy of the prediction. It has the advantages of utilizing spatial and environmental information to enhance prediction expressiveness, improve model generalization, and have strong resistance to data silos. It also has the advantages of strong resistance to local abnormal fluctuations, smooth and natural prediction results, and fusion of information from strongly correlated regions.

[0066] (5) By dynamically updating the adjacency matrix W and learnable fitting coefficients of the power generation graph structure, this invention solves the problems of static non-updating and inability to adapt to changes in weather and power generation mode in existing graph structure modeling methods. It realizes real-time adaptive updating of the graph structure within the scheduling cycle, so that the prediction model can be continuously optimized with the environment and actual power generation performance. It has the advantages of strong system self-learning ability, high real-time prediction update and good long-term operation stability.

[0067] (6) This invention integrates the load forecast timing with disturbance correction and the new energy power generation forecast timing with graph structure optimization, and constructs a multi-time-scale scheduling optimization model by combining the operating parameters of the energy storage system. This solves the problem that the existing scheduling strategy is based only on static forecast results and lacks a linkage optimization mechanism for real-time forecast changes and load uncertainties. It realizes the global optimal joint scheduling of energy storage charging and discharging plan and generator start-up, shutdown and output arrangement. It has the advantages of strong dynamic adaptability of scheduling scheme, high energy utilization efficiency and enhanced grid stability. It can be widely applied to the smart grid operation management in new energy scenarios with high penetration rate. Attached Figure Description

[0068] Figure 1 This is a flowchart of the method of the present invention;

[0069] Figure 2 This is a system model diagram of the present invention. Detailed Implementation

[0070] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0071] Example 1:

[0072] This invention provides a power grid dispatching method based on sudden events and graph structures, such as... Figure 1 As shown, it includes the following steps:

[0073] Step S1: Obtain historical electricity load data of the target power grid supply area within a preset continuous time period, and establish a baseline load prediction model based on the historical electricity load data;

[0074] Step S1 specifically includes:

[0075] Collect historical electricity load data of the target power grid supply area within a preset continuous time period, and construct a load time series dataset arranged in time series;

[0076] Historical electricity load data is cleaned and normalized to remove outliers and missing items;

[0077] The processed load time series dataset is divided into multiple sample sequences using a sliding time window approach. Each sample contains a fixed-length historical load input sequence and the corresponding predicted target value.

[0078] A Long Short-Term Memory (LSTM) model is constructed based on sample sequences. It uses historical load input sequences to predict future load values. During training, the mean square error between the predicted value and the actual load value is used as the loss function.

[0079] Divide the model into training and validation sets, optimize the model parameters through multiple rounds of iterative training, and use the validation set to evaluate the prediction accuracy to avoid overfitting.

[0080] After training is completed, a baseline load prediction model based on the Long Short-Term Memory (LSTM) network model is obtained.

[0081] Step S2: Utilize social media data, geolocation service trajectory data, and public event databases to continuously monitor and identify emergencies that affect electricity demand;

[0082] In step S2 of this embodiment, continuous monitoring and identification of emergencies affecting electricity demand are achieved by integrating social media data, geolocation service trajectory data, and a public event database. Specific technical means include: First, based on social media text data, natural language processing methods (such as keyword extraction, named entity recognition, and sentiment analysis) are used to automatically identify potential emergencies and their time, location, and event type information, while simultaneously quantifying the event's popularity and impact intensity using text popularity indicators (such as posting frequency and number of reposts and comments); Second, geolocation service trajectory data is used to mine changes in group behavior patterns, and based on changes in trajectory density and population gathering, the affected area and population response behavior are determined; Third, structured public event database information is invoked to verify, match, and supplement the identification results, achieving structured extraction of the event's occurrence time, duration, spatial impact range, and type, and performing risk rating; Fourth, an event fusion model is constructed to unify and correlate multi-source event information, outputting a standardized event description vector.

[0083] This embodiment addresses the technical problems of traditional power grid dispatching methods, such as the inability to obtain real-time information on sudden events, reliance on post-event statistics, and lack of ability to model the impact of sudden events on electricity load. It enables accurate identification and location of events in their early stages, significantly improving the sensing capability and response speed of the load forecasting system. This method possesses strong timeliness and accuracy, not only sensing sudden events within a region in real time but also assessing their potential disturbance to the electricity consumption behavior of users in different areas, thus providing a reliable basis for subsequent load disturbance modeling and prediction. Compared to existing prediction models based solely on historical data, this invention significantly enhances the dynamic adaptability of the power load forecasting system to sudden risks, effectively reduces prediction errors, and improves the robustness and flexibility of power grid dispatching.

[0084] Step S3: Based on the dynamic identification and time-varying feature extraction of sudden events in a continuous time series, combined with real-time electricity consumption behavior data, a time-varying electricity consumption disturbance factor is constructed, and dynamically corrected by combining the baseline load forecasting model to generate a disturbance-adjusted load forecast time series.

[0085] Step S3 specifically includes:

[0086] Collect a training dataset containing historical incident samples, including the incident occurrence time t. e , representing the start time of the sudden event, is used to locate the impact window of the event in the time series, ensuring that the model can accurately capture the changing trend of electricity load before and after the event; the impact area R eThis represents the geographical extent of the event's impact, typically expressed as a spatial polygon or a set of geographic coordinates. It measures the spatial overlap between the event and the power grid's supply area, thus estimating the event's direct impact on the target area's electricity consumption. Duration Δt e Event Type e This is used to identify the specific type of emergency, such as natural disasters (e.g., typhoons, earthquakes), public safety incidents (e.g., fires, traffic accidents), and large-scale events (e.g., concerts, sporting events). Different types of events have different impact patterns on power load, which is an important classification basis for the model to learn disturbance patterns. Social media popularity H e Change in pedestrian flow intensity ΔP e This represents the magnitude of change in pedestrian density or movement frequency within the area where the event occurred, typically derived from LBS (Location-Based Services) trajectory data. It reflects the gathering or evacuation behavior caused by the event and has a significant impact on short-term load fluctuations; the load change ΔL in the target area before and after the event. i (t);

[0087] Construct the input feature vector based on samples of sudden events:

[0088]

[0089] Among them, X e (t) represents the input feature vector of the sudden event e at time t, and norm(·) represents the normalization process for the variables. The area R affected by the emergency is indicated by the symbol e. e With the target power grid supply area A i Spatial overlap ratio;

[0090] The corresponding historical load change ΔL i (t) serves as the supervision label, and a perturbation prediction model is constructed. Where f θ The nonlinear model with trainable parameters is represented by a multilayer perceptron (MLP) and trained by minimizing the loss function of the perturbation prediction model.

[0091] After the model training is completed, in actual scheduling, for each sudden event identified within the continuous time series [t0, t0+ΔT], its input features are constructed in real time and input into the model to predict the time-varying power consumption disturbance factor at the current moment:

[0092]

[0093] in, Supplying power to the target power grid area A iThe time-varying power disturbance factor in a continuous time series [t0, t0+ΔT] is the load change caused by a sudden event in the continuous time series [t0, t0+ΔT].

[0094] The loss function of the perturbation prediction model is:

[0095]

[0096] in, The loss function for the perturbation prediction model is defined by indices e, i, t, which iterate through all events, regions, and times. θ (X e (t) represents the load change predicted by the model, ΔL i (t) represents the actual load change.

[0097] Generate the disturbance-adjusted load forecast time series, specifically including:

[0098] Load forecast sequence output from baseline load forecasting model As the initial predicted load, i represents the target grid supply area, and t represents a time in the time series;

[0099] By utilizing the disturbance prediction model obtained through training and combining it with the characteristics of sudden events identified within continuous time series, the time-varying power disturbance factor sequence of the target area can be predicted in real time. in This represents the predicted load disturbance caused by a sudden event at time t;

[0100] By weighted summation, the disturbance factor is added to the baseline load forecast sequence to obtain the disturbance-adjusted load forecast sequence:

[0101]

[0102] in, Let α be the load forecast value for the target power grid supply area i at time t, and α be a learnable weight parameter. The forecast value is trained and optimized by minimizing the error between the forecast value after disturbance adjustment and the actual load data.

[0103] Step S4: Obtain the geographical distribution information of new energy power generation sites, including wind power and photovoltaic power generation sites, and their historical power generation data, and construct a multi-node power generation map structure model;

[0104] Step S4 specifically includes:

[0105] Collect the geographical location information of new energy power generation sites and their respective power grid areas, and construct a node set {v1, v2, ..., v N}, where each node v iFor each new energy power generation site, the node attributes include the node's historical power generation data;

[0106] Based on the spatial distance d between nodes ij Historical power generation sequence correlation corr(P) i ,P j ) and the similarity of real-time meteorological conditions sin(C i C j Construct a weighted adjacency matrix W = [w ij ] N×N The weight calculation formula is:

[0107]

[0108] Among them, w ij Represents node v i With node v j Edge weights between; I(d) ij ≤d th ) is an indicator function, when d ij Less than or equal to the threshold d th The value is 1 if the distance is constant and 0 otherwise, controlling the connection range of nodes in the graph; σ is the Gaussian kernel bandwidth parameter, used to adjust the decay rate of the distance weight; corr(P i ,P j ) is node v i With node v j Pearson correlation coefficient of historical power generation time series, P i 、P j They are nodes v i With node v j Historical power generation time series, sin(C i C j ) is node v i With node v j The similarity of feature vectors of real-time meteorological conditions at the corresponding time; C i C j They represent nodes v respectively i With node v j The real-time meteorological condition feature vector includes wind speed, wind direction, solar radiation intensity, and cloud cover index; β1, β2, and β3 are learnable fitting coefficients, and N is the total number of new energy power generation sites.

[0109]

[0110] Among them, P i (t), P j (t) represent the historical power generation values ​​of node i and node j at time t, respectively. Let represent the mean of the historical power generation sequence of node i and node j, respectively. Let i and j represent the standard deviations of their historical power generation sequences, respectively. This is the expectation operator.

[0111]

[0112] Where ||·|| represents the Euclidean norm.

[0113] The learnable fitting coefficients β1, β2, and β3 are jointly optimized using the following fitting coefficient objective function:

[0114]

[0115] Among them, P j (t) represents the actual power generation of node j at time t, and N(j) represents the set of neighboring nodes of node j.

[0116] Step S5: Use the Long Short-Term Memory Network Model to predict the power generation of each new energy power generation site independently, and use the power generation graph structure model to correct the power generation prediction results of each new energy power station to obtain the corrected new energy power generation prediction time series.

[0117] First, renewable energy generation such as wind and solar power is highly dependent on natural meteorological conditions, which are continuous and correlated in both space and time. In other words, neighboring geographical areas are often affected by similar meteorological factors such as wind speed, wind direction, and solar radiation. Since meteorological conditions are the main factor driving the power generation of renewable energy, the power generation of different power generation sites exhibits a certain degree of synchronicity and correlation.

[0118] Specifically, geographically proximate wind farms or photovoltaic power plants often experience similar weather systems, such as the same storm cloud cluster or clear skies, causing their power generation to exhibit similar fluctuation patterns over time. This spatial meteorological continuity is the physical basis for their power generation.

[0119] For this fundamental reason, this invention employs a weighted adjacency matrix, which integrates spatial distance, correlation of historical power generation sequences, and similarity of real-time meteorological conditions into edge weights, quantitatively reflecting the "influence intensity" or "interdependence" between power plants. The closer the distance between stations, the more similar their historical power generation sequences, and the more similar their real-time meteorological conditions, the greater the weight, indicating a closer connection between their power generation changes.

[0120] Furthermore, due to the uncertainties and prediction errors in the actual meteorological data and the single-point power generation predicted by the LSTM model, relying solely on the prediction of a single site may lead to biased results. By fusing power generation predictions from neighboring nodes and meteorological information, and using a weighted adjacency matrix to correct the predicted values, errors can be effectively reduced, and the overall accuracy and stability of the prediction can be improved. This allows for a better reflection of the spatial dynamic correlation between new energy power plants, enabling more refined power generation prediction and scheduling optimization.

[0121] By correcting the power generation forecast results of each new energy power plant in the power generation diagram structure model, the corrected new energy power generation forecast time series is obtained, which specifically includes:

[0122] For each new energy power generation site v j Based on its historical power generation time series and real-time meteorological data, a long short-term memory network model is used to independently predict its preliminary power generation sequence.

[0123] Based on the constructed power generation diagram structure model, the weighted adjacency matrix W = [w ij ] N×N compute node v j The formula for the power generation correction value at time t is:

[0124]

[0125] in, For new energy power generation sites v j Corrected power generation at time t;

[0126] Based on preliminary power generation sequence With corrected power generation The revised forecast time series for renewable energy generation has been obtained:

[0127]

[0128] in, For node v j The revised forecast of new energy power generation at time t, where γ is the fusion weighting coefficient, is determined through training or cross-validation.

[0129] Weighted adjacency matrix W = [w ij ] N×N After each dispatch, the data is dynamically updated based on the latest actual power generation data and meteorological observation data through the following steps:

[0130] Collect the actual power generation of each node within the latest time period and the corresponding real-time meteorological data C j (t);

[0131] Based on the collected actual power generation The fitting coefficients β1, β2, and β3 can be learned by jointly optimizing the objective function of minimizing the prediction error.

[0132] Using the updated fitting coefficients β1, β2, β3, the weighted calculation formula is applied to the actual power generation. and the corresponding real-time meteorological data C j (t) Update the weighted adjacency matrix W of the power generation graph structure model.

[0133] Step S6: Combining the load forecast time series with disturbance correction and the new energy power generation forecast time series, and based on the energy storage system capacity, charging and discharging efficiency and energy state parameters, optimize the energy dispatching at multiple time scales within a preset continuous time period to generate energy storage charging and discharging strategies and generator start-up and output schemes.

[0134] Example 2:

[0135] This embodiment describes a power grid dispatching system based on sudden events and a graph structure, such as... Figure 2 Shown, including:

[0136] The data acquisition module is used to acquire historical electricity load data, social media data, geographic location service trajectory data and public event database information of the target power grid supply area, and continuously monitor and identify emergencies that affect electricity demand;

[0137] The load forecasting module establishes a baseline load forecasting model based on historical electricity load data, and combines real-time identified sudden event characteristics and electricity consumption behavior data to construct a time-varying electricity disturbance factor, thereby generating a time-series load forecast after disturbance adjustment.

[0138] The new energy power generation prediction module collects the geographical distribution information and historical power generation data of new energy power generation sites such as wind power and photovoltaic power generation sites, constructs a multi-node power generation graph structure model, uses a long short-term memory network (LSTM) to independently predict the power generation of each new energy power station, and corrects the prediction results based on the graph structure to obtain a more accurate new energy power generation prediction time series.

[0139] The energy dispatch optimization module combines the load forecast timing with the corrected new energy power generation timing with parameters such as energy storage system capacity, charging and discharging efficiency and energy status to optimize energy dispatch at multiple time scales within a preset continuous time period, generating energy storage charging and discharging strategies and generator start-up and output schemes.

[0140] The real-time update module dynamically updates the weighted adjacency matrix in the power generation graph structure model. Based on the latest actual power generation and meteorological data, it jointly optimizes the graph model parameters to improve the timeliness and accuracy of the scheduling strategy.

[0141] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0142] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A power grid dispatching method based on sudden events and graph structure, characterized in that, Includes the following steps: Obtain historical electricity load data of the target power grid supply area within a preset continuous time period, and establish a baseline load prediction model based on the historical electricity load data; By utilizing social media data, geolocation service trajectory data, and public event databases, we continuously monitor and identify emergencies that affect electricity demand; Based on the dynamic identification and time-varying feature extraction of sudden events in a continuous time series, combined with real-time electricity consumption behavior data, a time-varying electricity consumption disturbance factor is constructed, and dynamically corrected in combination with the baseline load forecasting model to generate a disturbance-adjusted load forecasting time series. Obtain geographical distribution information and historical power generation data of new energy power generation sites, including wind power and photovoltaic power generation sites, and construct a multi-node power generation graph structure model; The power generation of each new energy power generation site is predicted independently using a long short-term memory network model. The power generation prediction results of each new energy power station are then corrected using the power generation graph structure model to obtain the corrected new energy power generation prediction time series. By combining the load forecast timing with the new energy power generation forecast timing, and based on the energy storage system capacity, charging and discharging efficiency and energy state parameters, the energy dispatching at multiple time scales within the preset continuous time period is optimized to generate energy storage charging and discharging strategies and generator start-up and output schemes.

2. The power grid dispatching method based on sudden events and graph structure according to claim 1, characterized in that, The establishment of a baseline load forecasting model based on the historical electricity load data specifically includes: Collect historical electricity load data of the target power grid supply area within a preset continuous time period, and construct a load time series dataset arranged in time series; The historical electricity load data is cleaned and normalized to remove outliers and missing items. The processed load time series dataset is divided into multiple sample sequences using a sliding time window approach. Each sample contains a fixed-length historical load input sequence and the corresponding predicted target value. Based on the sample sequence, a Long Short-Term Memory (LSTM) network model is constructed. The historical load input sequence is used to predict future load values. During the training process, the mean square error between the predicted value and the actual load value is used as the loss function. Divide the model into training and validation sets, optimize the model parameters through multiple rounds of iterative training, and use the validation set to evaluate the prediction accuracy to avoid overfitting. After training is completed, a baseline load prediction model based on the Long Short-Term Memory (LSTM) network model is obtained.

3. The power grid dispatching method based on sudden events and graph structure according to claim 1, characterized in that, The method for constructing a time-varying electricity consumption disturbance factor based on the dynamic identification and time-varying feature extraction of sudden events within a continuous time series, combined with real-time electricity consumption behavior data, specifically includes: A training dataset containing historical incident samples was collected, including the incident occurrence time t. e Area of ​​influence R e Duration Δt e Event Type e Social media popularity H e Change in pedestrian flow intensity ΔP e and the load change ΔL in the target area before and after the event i (t); Construct the input feature vector based on samples of sudden events: Among them, X e (t) represents the input feature vector of the sudden event e at time t, and norm(·) represents the normalization process for the variables. The area R affected by the emergency is indicated by the symbol e. e With the target power grid supply area A i Spatial overlap ratio; The corresponding historical load change ΔL i (t) serves as the supervision label, and a perturbation prediction model is constructed. Where f θ The nonlinear model with trainable parameters is represented by a multilayer perceptron (MLP) and trained by minimizing the loss function of the perturbation prediction model. After the model training is completed, in actual scheduling, for each sudden event identified within the continuous time series [t0, t0+ΔT], its input features are constructed in real time and input into the model to predict the time-varying power consumption disturbance factor at the current moment: in, Supplying power to the target power grid area A i The time-varying power disturbance factor in a continuous time series [t0, t0+ΔT] is the load change caused by a sudden event in the continuous time series [t0, t0+ΔT].

4. The power grid dispatching method based on sudden events and graph structure according to claim 1, characterized in that, The loss function of the perturbation prediction model is: in, The loss function for the perturbation prediction model is defined by indices e, i, t, which iterate through all events, regions, and times. θ (X e (t) represents the load change predicted by the model, ΔL i (t) represents the actual load change.

5. The power grid dispatching method based on sudden events and graph structure according to claim 1, characterized in that, The generation of the perturbation-adjusted load forecast time series specifically includes: Load forecast sequence output from baseline load forecasting model As the initial predicted load, i represents the target grid supply area, and t represents a time in the time series; By utilizing the disturbance prediction model obtained through training and combining it with the characteristics of sudden events identified within continuous time series, the time-varying power disturbance factor sequence of the target area can be predicted in real time. in This represents the predicted load disturbance caused by a sudden event at time t; By weighted summation, the disturbance factor is added to the baseline load forecast sequence to obtain the disturbance-adjusted load forecast sequence: in, Let α be the load forecast value for the target power grid supply area i at time t, and α be a learnable weight parameter. The forecast value is trained and optimized by minimizing the error between the forecast value after disturbance adjustment and the actual load data.

6. The power grid dispatching method based on sudden events and graph structure according to claim 1, characterized in that, The acquisition of geographical distribution information and historical power generation data of new energy power generation sites, including wind power and photovoltaic power generation sites, and the construction of a multi-node power generation map structure model, specifically includes: Collect the geographical location information of the new energy power generation sites and their respective power grid areas, and construct a node set {v1, v2, ..., v N }, where each node v i For each new energy power generation site, the node attributes include the node's historical power generation data; Based on the spatial distance d between nodes ij Historical power generation sequence correlation corr(P) i ,P j ) and the similarity of real-time meteorological conditions sin(C i C j Construct a weighted adjacency matrix W = [w ij ] N×N The weight calculation formula is: Among them, w ij Represents node v i With node v j Edge weights between; I(d) ij ≤d th ) is an indicator function, when d ij Less than or equal to the threshold d th The value is 1 if the distance is constant and 0 otherwise, controlling the connection range of nodes in the graph; σ is the Gaussian kernel bandwidth parameter, used to adjust the decay rate of the distance weight; corr(P i ,p j ) is node v i With node v j Pearson correlation coefficient of historical power generation time series, p i P j They are nodes v i With node v j Historical power generation time series, sin(C i C j ) is node v i With node v j The similarity of feature vectors of real-time meteorological conditions at the corresponding time; C i C j Representing node v respectively i With node v j The real-time meteorological condition feature vector; β1, β2, and β3 are learnable fitting coefficients, and N is the total number of new energy power generation sites.

7. A power grid dispatching method based on sudden events and graph structure according to claim 6, characterized in that, The real-time meteorological condition feature vector includes wind speed, wind direction, solar radiation intensity, and cloud cover index.

8. The power grid dispatching method based on sudden events and graph structure according to claim 1, characterized in that, The learnable fitting coefficients β1, β2, and β3 are jointly optimized using the following fitting coefficient objective function: Among them, P j (t) represents the actual power generation of node j at time t, and N(j) represents the set of neighboring nodes of node j.

9. A power grid dispatching method based on sudden events and graph structure according to claim 1, characterized in that, The step of correcting the power generation prediction results of each new energy power station using the power generation diagram structure model to obtain the corrected new energy power generation prediction time series specifically includes: For each new energy power generation site v j Based on its historical power generation time series and real-time meteorological data, a long short-term memory network model is used to independently predict its preliminary power generation sequence. Based on the constructed power generation diagram structure model, the weighted adjacency matrix W = [w ij ] N×N compute node v j The formula for the power generation correction value at time t is: in, For new energy power generation sites v j Corrected power generation at time t; Based on preliminary power generation sequence With corrected power generation The revised forecast time series for renewable energy generation has been obtained: in, For node v j The revised forecast of new energy power generation at time t, where γ is the fusion weighting coefficient, is determined through training or cross-validation.

10. A power grid dispatching method based on sudden events and graph structure according to claim 9, characterized in that, The weighted adjacency matrix W = [w ij ] N×N After each dispatch, the data is dynamically updated based on the latest actual power generation data and meteorological observation data through the following steps: Collect the actual power generation of each node within the latest time period and the corresponding real-time meteorological data C j (t); Based on the collected actual power generation The fitting coefficients β1, β2, and β3 can be learned by jointly optimizing the objective function of minimizing the prediction error. Using the updated fitting coefficients β1, β2, β3, the weighted calculation formula is applied to the actual power generation. and the corresponding real-time meteorological data C j (t) Update the weighted adjacency matrix W of the power generation graph structure model.

Images