Power grid dispatching method and system

Abstract

The present invention relates to the field of smart grid dispatching technology, and in particular to a grid dispatching method and system, the method comprising: collecting environmental data, equipment health data, and economic and market data, and using edge computing devices to generate fused data; using a multi-model integrated deep learning model to predict power demand and supply, combining a reinforcement learning algorithm to simulate grid dispatching scenarios, and generating an optimal dispatching strategy; dynamically adjusting the output power of power stations and grid node load distribution, and updating dispatching strategies in real time. The present invention improves prediction accuracy through multi-model integration, optimizes scenario simulation strategies using reinforcement learning, and combines multi-level feedback analysis to achieve coordinated optimization of global and local states, significantly improving the stability and economy of grid operation, while reducing carbon emissions, and is suitable for complex scenarios of modern smart grid dispatching.

Description

Technical Field

[0001] The present invention relates to the technical field of smart grid dispatching, and in particular to a grid dispatching method and system. Background Art

[0002] Grid dispatching is the core link to ensure efficient operation and stable power supply of the power system. Its purpose is to optimize the resource allocation of power generation, transmission and distribution to meet the ever-changing power demand while reducing operating costs and environmental impact. However, the existing technology still has significant deficiencies in achieving these goals. For example, the existing technology (Chinese invention patent, publication number: CN118174278A, name: a grid dispatching method based on carbon emissions) optimizes the resource dispatching of low-carbon power generation by means of allocating carbon emission constraint values ​​in different zones and sorting the environmental protection coefficients of thermal power plants. However, since it mainly relies on the centralized planning calculation of the central server, it lacks the ability to respond efficiently to real-time dynamic changes. In addition, this method has deficiencies in the accuracy of wind power and photovoltaic power generation prediction and the flexibility of dispatching strategies. The detailed optimization of carbon emission constraints and the coordination of global dispatching are weak, which is easy to cause insufficient power generation or failure of dispatching strategies in emergencies. Summary of the invention

[0003] In view of the many problems existing in the above-mentioned prior art, the present invention provides a power grid dispatching method and system, which uses a multi-model integrated deep learning model to predict power demand and supply, combines a reinforcement learning algorithm to dynamically simulate power grid dispatching scenarios, and generates optimal dispatching strategy data. Combined with multi-level feedback analysis and real-time optimization dispatching, the present invention can realize dynamic adjustment of power stations and power grid nodes, ensure the stability of power grid operation and significantly improve energy utilization efficiency.

[0004] A power grid dispatching method comprises the following steps:

[0005] Collect environmental data, social activity data, equipment health data, and economic and market data, and pre-process and integrate the collected data through edge computing devices to generate fused data.

[0006] Based on the fused data, a multi-model integrated deep learning model is used to perform power demand forecasting and power supply forecasting to generate power demand forecast data and power supply forecast data. The multi-model integrated deep learning model includes: long short-term memory network model, Transformer model and gated recurrent unit model.

[0007] Based on electricity demand forecast data and electricity supply forecast data, the power grid dispatching scenario is simulated using reinforcement learning algorithm and multi-model integration algorithm to generate scenario simulation data and optimal dispatching strategy data.

[0008] Dynamically adjust the power generation and load of power stations and grid nodes according to the optimal dispatching strategy data, and update the grid dispatching strategy.

[0009] Preferably, the environmental data includes: temperature data, humidity data, wind speed data, light intensity data and air quality data; the social activity data includes: pedestrian flow data and traffic flow data; the equipment health data includes: equipment vibration data, equipment surface temperature data and infrared imaging data; the economic and market data includes: real-time energy market price data and carbon emission trading price data.

[0010] Preferably, the processing steps of the multi-model integrated deep learning model include:

[0011] The long short-term memory network model is used to extract time series features from the fused data, including: modeling the long-term trend of changes in electricity demand and identifying periodic fluctuation patterns.

[0012] The Transformer model is used to perform feature weighting processing on key nodes in time series features based on the self-attention mechanism, extract global correlation features, and generate intermediate prediction data.

[0013] The gated recurrent unit model is used to perform feature correction on the intermediate prediction data, including: eliminating abnormal data points and adjusting short-term fluctuations to generate the final power demand forecast data and power supply forecast data.

[0014] Preferably, when processing fused data, the long short-term memory network model uses input gates, forget gates, and output gates to dynamically adjust the input state and forget state of the time series. The Transformer model uses position encoding and multi-head self-attention mechanisms to analyze the temporal dependency and multi-dimensional correlation of time series features. The gated recurrent unit model controls the weight distribution between input data and historical states through gated units.

[0015] Preferably, the reinforcement learning algorithm and the multi-model integration algorithm simulate the power grid dispatching scenario including:

[0016] Initial scenario simulation data is constructed based on power demand forecast data and power supply forecast data, and training is performed through a scenario simulation environment.

[0017] The reinforcement learning algorithm is used to define the training process through state space, action space and reward function. The state space includes: the current load distribution state of the power grid and the operating state of the power station, the action space includes: the adjustment plan of the dispatching strategy, and the reward function is designed based on the stability of power grid operation and energy efficiency.

[0018] The output results of the reinforcement learning algorithm are optimized through a multi-model integration algorithm, including: using a resampling mechanism for model training, iteratively optimizing historical prediction errors, and generating optimal scheduling strategy data through model combination.

[0019] Preferably, the multi-model integration algorithm includes: a resampling integration algorithm, a weighted optimization algorithm and a model stacking algorithm, wherein the resampling integration algorithm generates optimized simulation data by independently training different samples of scenario simulation data and performing weighted averaging on the model output results; the weighted optimization algorithm iteratively optimizes the model parameters by adjusting the error distribution of the weak learner; the model stacking algorithm generates optimal scheduling strategy data by constructing the outputs of multiple basic models and generating a high-order prediction model.

[0020] Preferably, dynamically adjusting the power generation and load of the power station and the grid node according to the optimal dispatching strategy data includes:

[0021] Receive optimal dispatch strategy data and make real-time adjustments to the output power of power stations, including wind power stations, solar power stations and thermal power stations.

[0022] Optimize the load distribution strategy of power grid nodes, including: dynamic balancing based on load flow between nodes and load transfer to nodes that may be overloaded.

[0023] Generate dispatch execution data and apply the adjustment results to power plants and grid nodes in conjunction with the grid control system.

[0024] Preferably, updating the power grid dispatching strategy includes: performing multi-level analysis on real-time feedback data of the power grid operating status, the multi-level analysis including: generating dispatching optimization data based on analysis of local power grid node operating status and trend analysis of global power grid status, and using the dispatching optimization data to update the dispatching strategy.

[0025] A system for implementing the power grid dispatching method, comprising:

[0026] The data collection module is used to collect environmental data, social activity data, equipment health data, and economic and market data, and pre-process and integrate the environmental data, social activity data, equipment health data, and economic and market data to generate fused data.

[0027] The multi-model integrated deep learning module is used to perform power demand forecasting and power supply forecasting based on the fused data, and generate power demand forecasting data and power supply forecasting data, wherein the multi-model integrated deep learning module includes:

[0028] Long short-term memory network model, used to extract time series features and generate preliminary forecast data.

[0029] Transformer model, used to extract global features and optimize preliminary prediction data based on the self-attention mechanism.

[0030] The gated recurrent unit model is used to perform feature correction on the optimized prediction data to generate final power demand prediction data and power supply prediction data.

[0031] The reinforcement learning and multi-model integrated algorithm module is used to simulate power grid dispatching scenarios based on power demand forecast data and power supply forecast data, and generate scenario simulation data and optimal dispatching strategy data.

[0032] The dispatching execution module is used to dynamically adjust the power generation of the power station and the load of the power grid node according to the optimal dispatching strategy data, and update the power grid dispatching strategy in real time.

[0033] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0034] The present invention achieves high accuracy and robustness in electricity demand and supply forecasts through multi-model integrated deep learning model technology, greatly reducing the uncertainty of wind power and photovoltaic power generation forecasts.

[0035] The present invention realizes accurate optimization of complex power grid operation scenarios through dynamic scenario simulation technology that combines reinforcement learning with multi-model integration algorithm, and significantly improves the flexibility and adaptability of the dispatching strategy.

[0036] The present invention realizes dynamic coordinated optimization of global and local power grid states through real-time adjustment strategies and multi-level feedback analysis technical means, and solves the problem of insufficient response capability to emergencies in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the process of the present invention;

[0038] Figure 2 This is a schematic diagram of the multi-model integrated deep learning model architecture in the present invention;

[0039] Figure 3 It is a flow chart of reinforcement learning and scenario simulation in the present invention;

[0040] Figure 4 It is a schematic diagram of the multi-level feedback analysis architecture in the present invention;

[0041] Figure 5 It is a structural block diagram of the system of the present invention. DETAILED DESCRIPTION

[0042] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is obvious that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0043] like Figure 1 As shown, a power grid dispatching method comprises the following steps:

[0044] Collect environmental data, social activity data, equipment health data, and economic and market data, and pre-process and integrate the collected data through edge computing devices to generate fused data.

[0045] Preferably, the environmental data includes: temperature data, humidity data, wind speed data, light intensity data and air quality data; the social activity data includes: pedestrian flow data and traffic flow data; the equipment health data includes: equipment vibration data, equipment surface temperature data and infrared imaging data; the economic and market data includes: real-time energy market price data and carbon emission trading price data.

[0046] The first step in power grid dispatching is to generate fused data through the collection, preprocessing and integration of multi-dimensional data, including environmental data, social activity data, equipment health data and economic and market data. Environmental data reflects important external factors that affect power grid operation and renewable energy generation, such as temperature data, humidity data, wind speed data, light intensity data and air quality data. They are obtained in real time through meteorological sensors, ground monitoring stations and satellite remote sensing systems. In particular, in wind power and photovoltaic power generation, wind speed and light intensity are used to estimate wind power output and photovoltaic power generation respectively. Social activity data is directly related to power demand. By analyzing pedestrian flow data and traffic flow data, the peak hours of electricity consumption and load distribution can be accurately grasped. The data mainly comes from camera networks, mobile communication signal analysis systems and big data platforms. Equipment health data is an important guarantee for the reliable operation of the power grid, including equipment vibration data, equipment surface temperature data and infrared imaging data. These data are collected in real time by vibration sensors, thermal imagers and online monitoring modules installed at transformers and line nodes, and are used to evaluate the operating status of equipment and warn of potential failure risks. Economic and market data include real-time energy market price data and carbon emission trading price data. These data are obtained through interfaces with market platforms to provide economic and environmentally friendly constraints for optimizing scheduling strategies.

[0047] In order to improve data quality and consistency, the present invention uses edge computing devices to preprocess and integrate the above data. Edge computing devices are deployed at physical nodes close to the data collection end, and improve the integrity and availability of data through steps such as data cleaning, format conversion and noise reduction. For example, for wind speed data in environmental data, edge computing devices use wavelet transform technology to filter out high-frequency noise, and use interpolation algorithms to repair missing values ​​that occur during sensor acquisition; for traffic flow data in social activity data, images collected by the camera network are analyzed in real time to extract key features such as vehicle density; and for equipment health data, by analyzing the frequency change trend of vibration data and the abnormal temperature distribution pattern of infrared imaging, operational equipment status indicators are generated in real time. During the data integration process, edge computing devices use streaming processing technology to standardize multi-source heterogeneous data to ensure timestamp synchronization and consistent data structure, and at the same time achieve rapid storage and access to large-scale data through distributed databases.

[0048] Through this process, the fused data finally generated has high real-time performance, high consistency and high accuracy, providing a solid data foundation for subsequent power demand forecasting, power supply forecasting and scheduling optimization. In a practical application, such as a regional distributed power grid, edge computing devices integrate wind speed data, light intensity data and economic market data in real time to provide multi-dimensional input for power forecasting of wind power and photovoltaic power generation. At the same time, through the integration of equipment health data, the load distribution of high-load lines is dynamically adjusted to avoid equipment overload, thereby improving the stability and safety of power grid operation. In the above manner, the present invention ensures the comprehensiveness and flexibility of scheduling decisions, while significantly improving the efficiency of power grid operation.

[0049] like Figure 2 As shown, based on the fused data, a multi-model integrated deep learning model is used to perform power demand forecasting and power supply forecasting to generate power demand forecast data and power supply forecast data, wherein the multi-model integrated deep learning model includes: long short-term memory network model, Transformer model and gated recurrent unit model.

[0050] In this invention, based on the fusion data, a multi-model integrated deep learning model is used to predict power demand and power supply, aiming to provide high-precision load forecasting and power generation forecasting data for power grid dispatching, thereby providing support for subsequent dispatch optimization. The multi-model integrated deep learning model combines the characteristics of the long short-term memory network model, the Transformer model, and the gated recurrent unit model to make up for the limitations of a single model in a specific scenario, and realizes a comprehensive analysis and efficient prediction of time series characteristics.

[0051] The long short-term memory network model is used to extract time series features from power demand and supply data, and is particularly suitable for capturing long-term trends and cyclical fluctuations. It controls the weight distribution of input data, memory state, and output state through input gates, forget gates, and output gates, thereby effectively solving the gradient vanishing problem that is prone to occur in traditional recurrent neural networks in long time series. In the present invention, the long short-term memory network model models the historical power demand data and power generation data in the fusion data to generate preliminary power demand forecast data and power supply forecast data. These preliminary forecast results mainly reflect long-term trends and cyclical changes, such as seasonal demand for residential electricity or cyclical fluctuations in wind power and photovoltaic power generation.

[0052] The Transformer model further optimizes the preliminary prediction results. Through its self-attention mechanism, the Transformer model can effectively identify key features and global correlations in time series. Specifically, the Transformer model calculates the weighted matrix product of query, key, and value, focusing on extracting key factors affecting electricity demand and supply, such as sudden climate change or abnormal fluctuations in economic activities. In the present invention, the Transformer model optimizes the preliminary prediction data through global feature extraction to generate optimized electricity demand forecast data and electricity supply forecast data. This optimization process not only improves the accuracy of the prediction, but also enhances the model's adaptability to dynamic environmental changes.

[0053] The gated recurrent unit model further performs feature correction based on the optimized data. Compared with the long short-term memory network model, the gated recurrent unit model has a simpler structure, but also has good time series modeling capabilities. In the present invention, the gated recurrent unit model processes short-term fluctuations and abnormal data, such as eliminating sudden spikes in power demand or correcting errors caused by equipment failures in supply forecasts. By controlling the balance between short-term data and historical data through the gated unit, the gated recurrent unit model can generate more stable and reliable final power demand forecast data and power supply forecast data.

[0054] Overall, the multi-model integrated deep learning model achieves high-precision prediction of power demand and power supply by combining the time series trend analysis capability of the long short-term memory network model, the global feature extraction capability of the Transformer model, and the short-term correction capability of the gated recurrent unit model. In one embodiment, in a regional power grid where wind power and photovoltaic power generation account for a high proportion, the long short-term memory network model uses the power generation data of the past few years to capture the seasonal variation trend of wind power output; the Transformer model identifies the impact of recent wind speed mutations on power generation through the self-attention mechanism; the gated recurrent unit model eliminates data anomalies caused by wind speed sensor failures, and finally generates accurate power generation forecast data. Through this process, the present invention not only improves the accuracy of load forecasting and power generation forecasting, but also significantly enhances the power grid's ability to respond to dynamic changes, providing a reliable data basis for subsequent power dispatch optimization.

[0055] Preferably, the processing steps of the multi-model integrated deep learning model include:

[0056] The long short-term memory network model is used to extract time series features from the fused data, including: modeling the long-term trend of changes in electricity demand and identifying periodic fluctuation patterns.

[0057] The Transformer model is used to perform feature weighting processing on key nodes in time series features based on the self-attention mechanism, extract global correlation features, and generate intermediate prediction data.

[0058] The gated recurrent unit model is used to perform feature correction on the intermediate prediction data, including: eliminating abnormal data points and adjusting short-term fluctuations to generate the final power demand forecast data and power supply forecast data.

[0059] The present invention processes the fused data through a multi-model integrated deep learning model, and its core steps include: using a long short-term memory network model to extract time series features and identify long-term trends and periodic fluctuations in power demand and supply; using the self-attention mechanism of the Transformer model to perform global correlation analysis on features and generate intermediate prediction data; finally, using a gated recurrent unit model to perform feature correction on the intermediate prediction data, remove abnormal data points and adjust short-term fluctuations, thereby generating the final high-precision power demand prediction data and power supply prediction data. This process not only embodies the multi-dimensional analysis of complex time series features in theory, but also significantly improves the accuracy and stability of predictions in practice, providing key support for the optimization of power grid dispatching.

[0060] In the first step, the long short-term memory network model (LSTM) focuses on extracting time series features. Its core is to dynamically control the transmission and forgetting of information through memory units and gating mechanisms (input gate, forget gate, output gate), thereby effectively modeling the long-term trend and periodic fluctuation pattern of power demand changes. Taking the historical load data in the power grid as an example, LSTM uses the formula:

[0061]

[0062] in, Indicates time The hidden state output of is the activation value of the output gate, which controls the impact of the current state on the output; is the memory cell state, representing the accumulated time series information. Through this process, LSTM is able to extract the seasonal demand patterns of residential electricity consumption, such as the increase in air conditioning load in summer or the peak electricity consumption of heating equipment in winter.

[0063] Next, the Transformer model performs feature weighting on the key nodes in the time series features based on the self-attention mechanism, extracts global correlation features and generates intermediate prediction data. The Transformer self-attention mechanism calculates the query matrix , key matrix Sum Matrix The dot product of derives the attention weight, which is used to measure the importance of features at different time steps. The calculation formula is:

[0064]

[0065] in, , , are query, key, and value matrices respectively, represents the dimension of the key vector, used for normalization, represents the transpose operation of the matrix, represents the self-attention mechanism, In this way, the Transformer model can identify potential factors that lead to load surges, such as sudden increases in commercial power load during holidays or abnormal power demand caused by sudden changes in weather, providing support for the optimization of subsequent forecast data.

[0066] Finally, the gated recurrent unit model (GRU) performs feature correction on the intermediate prediction data, mainly for the removal of abnormal data points and the adjustment of short-term fluctuations. Compared with LSTM, GRU has a more compact structure, and controls the update and reset of hidden states through update gates and reset gates. Its core calculation formula is:

[0067]

[0068] in, It's time The hidden state output of It is the update gate, which controls the balance between historical information and current input; is a candidate hidden state, representing the information of the current time step. GRU is used in the power grid to remove abnormal sensor data, such as prediction deviation caused by wind speed sensor failure, and optimize short-term fluctuations such as local adjustments of peak power consumption to generate high-precision power demand forecast data and power supply forecast data.

[0069] In one embodiment, in a power grid area with a high proportion of wind power, LSTM identifies seasonal trends by modeling historical wind speed and output data; Transformer identifies the global correlation between recent wind speed mutations and power generation fluctuations; GRU eliminates data anomalies caused by anemometer failures and adjusts short-term power generation fluctuations, ultimately generating power generation forecasts with an error rate of less than 2%. This multi-model integration method significantly improves the accuracy and stability of power demand and supply forecasts by giving full play to the strengths of different models, providing efficient and reliable data support for power grid operations.

[0070] Preferably, when processing fused data, the long short-term memory network model uses input gates, forget gates, and output gates to dynamically adjust the input state and forget state of the time series. The Transformer model uses position encoding and multi-head self-attention mechanisms to analyze the temporal dependency and multi-dimensional correlation of time series features. The gated recurrent unit model controls the weight distribution between input data and historical states through gated units.

[0071] In the present invention, the application of multi-model integrated deep learning model in power grid dispatching analyzes time series characteristics through the combination of long short-term memory network model (LSTM), Transformer model and gated recurrent unit model (GRU), accurately extracts historical trends, time dependencies and multi-dimensional correlation features, generates high-quality prediction data, and provides strong support for the dynamic optimization of power demand and supply.

[0072] When processing fused data, the long short-term memory network model (LSTM) uses input gates, forget gates, and output gates to dynamically adjust the input state, forget state, and output state of time series data to capture long-term trends and periodic characteristics. Its core is to achieve selective transmission of information through the following calculation formula:

[0073] Input gate activation value:

[0074]

[0075] in, Indicates time The input gate state controls the current input Impact on memory cells; and are weight matrix and bias respectively; Represents the hidden state of the previous moment and captures historical features.

[0076] Activation value of forget gate:

[0077]

[0078] in, Indicates the state of the forget gate, which controls whether historical information is retained or forgotten at the current moment.

[0079] Memory unit status:

[0080]

[0081] in, It is the current state of the memory unit, which combines the historical state and the influence of the current input features.

[0082] Output gate activation and hidden state:

[0083]

[0084] in, Controlling the memory cell state to the hidden state for the final output.

[0085] In the present invention, LSTM is used to extract the long-term variation trend of historical load data and supply data of power grid operation, such as capturing the daily power consumption cycle during the peak of air conditioning load in summer, and identifying seasonal fluctuation patterns through years of data.

[0086] The Transformer model further analyzes the multi-dimensional associations and time dependencies in time series features. The core lies in weighting key features through a multi-head self-attention mechanism. Position encoding is used to introduce the position information of the time series and solve the time dependency problem of sequence features in a non-recursive structure. The specific implementation is:

[0087] Self-attention calculation formula:

[0088] Transformer uses the self-attention mechanism to capture the correlation characteristics of historical load data and real-time environmental data, such as the correlation between abnormal wind speed and wind power output fluctuations.

[0089] The Gated Recurrent Unit (GRU) model corrects the intermediate prediction data by updating the gate and resetting the gate, mainly dealing with short-term fluctuations and abnormal data points. Its core calculation formula is as follows:

[0090] Update Gate:

[0091]

[0092] Controlling hidden states The update ratio of historical information to current input.

[0093] Reset the gate:

[0094]

[0095] Determines the extent to which the current input affects the hidden state.

[0096] Current hidden status:

[0097]

[0098] Combine historical status and current input to remove abnormal data points and smooth short-term fluctuations.

[0099] In power grid applications, GRU generates stable forecast results by eliminating abnormal data (such as instantaneous load spikes) caused by sensor errors and smoothing short-term fluctuations in load forecasts.

[0100] In one embodiment, through the integrated application of the above models, the present invention realizes multi-dimensional analysis, trend capture and short-term correction of time series characteristics. Taking a regional power grid as an example, LSTM identifies the seasonal trend of power load; Transformer analyzes the global correlation between economic activities and power demand; GRU eliminates the noise anomalies of load sensors and adjusts short-term fluctuations, and finally generates power demand forecast data with an error rate of less than 2%, providing an accurate and reliable data basis for power grid dispatch optimization. This method significantly improves the prediction accuracy and robustness, and improves the economy and stability of power grid operation.

[0101] Based on electricity demand forecast data and electricity supply forecast data, the power grid dispatching scenario is simulated using reinforcement learning algorithm and multi-model integration algorithm to generate scenario simulation data and optimal dispatching strategy data.

[0102] The present invention simulates the power grid dispatching scenarios generated based on power demand forecast data and power supply forecast data by combining reinforcement learning algorithm and multi-model integration algorithm to generate high-quality scenario simulation data and optimal dispatching strategy data, providing optimized decision support for power grid operation. This process gradually realizes the dynamic simulation and optimization of power grid dispatching scenarios from state decision optimization of reinforcement learning to prediction result enhancement of multi-model integration algorithm.

[0103] The application of reinforcement learning algorithms in power grid dispatching is centered on the construction of state space, action space, and reward function. The state space includes elements such as the power grid load distribution, power plant operating status, and transmission line load conditions reflected in the power demand forecast data and power supply forecast data. The action space defines the adjustment plan of the power grid dispatching strategy, such as dynamically adjusting the output power of the power plant, optimizing the load distribution, and adjusting the charging and discharging plan of the energy storage equipment. The reward function design is based on multiple objectives of power grid operation, usually including minimizing the cost of power generation, reducing carbon emissions, improving power supply reliability, and reducing line losses. Through iterative training of the reinforcement learning algorithm, the system can continuously try different dispatching strategies in a simulated environment and continuously optimize the choice of strategies based on reward feedback.

[0104] The multi-model integration algorithm further enhances the accuracy and stability of scheduling scenario simulation. By integrating the Bagging algorithm, the Boosting algorithm and the Stacking algorithm, the multi-model integration algorithm can combine the advantages of multiple basic models to improve the robustness and prediction accuracy of scenario simulation. The Bagging algorithm reduces the random error of a single model by resampling the data to train multiple models and averaging the output results; the Boosting algorithm makes the final prediction result more accurate by gradually optimizing the error of the previous model; the Stacking algorithm uses the outputs of multiple basic models as input data to train high-order models to achieve collaborative optimization between models. In the present invention, the multi-model integration algorithm not only corrects the preliminary scenario simulation data generated by the reinforcement learning algorithm, but also comprehensively evaluates the simulation results of different scheduling strategies to generate the optimal scheduling strategy data.

[0105] By combining reinforcement learning algorithms with multi-model integration algorithms, the system can simulate the execution effect of dispatch strategies in complex power grid operating environments, identify potential problems and optimize decisions. In one embodiment, in a regional power grid where wind power accounts for a high proportion, the reinforcement learning algorithm designs a dynamic dispatch strategy with real-time wind speed and load demand as state inputs based on the uncertainty of wind power output. The Bagging algorithm resamples historical simulation data to generate dispatch simulation results of different models; the Boosting algorithm optimizes the weight distribution of the dispatch strategy based on actual operation feedback; the Stacking algorithm combines the basic dispatch model and optimization strategy to generate comprehensive optimal dispatch strategy data. Ultimately, this strategy significantly reduces the wind abandonment rate in the region while ensuring the dynamic balance of the load.

[0106] Preferably, Figure 3 As shown in the figure, the reinforcement learning algorithm and the multi-model integration algorithm simulate the power grid dispatching scenario including:

[0107] Initial scenario simulation data is constructed based on power demand forecast data and power supply forecast data, and training is performed through a scenario simulation environment.

[0108] The reinforcement learning algorithm is used to define the training process through state space, action space and reward function. The state space includes: the current load distribution state of the power grid and the operating state of the power station, the action space includes: the adjustment plan of the dispatching strategy, and the reward function is designed based on the stability of power grid operation and energy efficiency.

[0109] The output results of the reinforcement learning algorithm are optimized through a multi-model integration algorithm, including: using a resampling mechanism for model training, iteratively optimizing historical prediction errors, and generating optimal scheduling strategy data through model combination.

[0110] The present invention combines reinforcement learning algorithm and multi-model integration algorithm to simulate the power grid dispatching scenario, generate high-precision scenario simulation data and optimal dispatching strategy data, so as to optimize the operation status of complex power grids. This process consists of the construction of scenario simulation data, the training process of reinforcement learning and the optimization of multi-model integration algorithm, covering the comprehensive modeling and strategy optimization of the power grid status, and ultimately providing a scientific basis for the dynamic dispatching of the power grid.

[0111] Based on the power demand forecast data and power supply forecast data, the initial scenario simulation data is constructed. The data includes the load distribution status of the power grid, the operating status of the power station, and the load status of the transmission line. For example, in a regional power grid that includes wind power, photovoltaic power, and thermal power generation, the initial scenario simulation data may include the current wind power and photovoltaic power generation forecast data, the power generation of each power station, and the power flow distribution status of the transmission line. The construction of these data relies on the combination of historical operation data and real-time forecast data. Through integrated analysis, a multi-dimensional description of the current state of the power grid is generated, providing basic input for scenario simulation.

[0112] The reinforcement learning algorithm trains the dispatching strategy on this basis. The training process is based on the definition of state space, action space and reward function. The state space includes the current load distribution state of the power grid and the operating state of the power station, such as wind power generation, photovoltaic power generation and the load conditions of each node; the action space covers the optional dispatching strategy adjustment schemes, such as dynamically adjusting the output power of the power station, optimizing the load distribution of the line, and changing the charging and discharging strategy of the energy storage device. The reward function is designed to balance the multiple objectives of power grid operation, such as improving economy by maximizing energy efficiency and minimizing line losses, and improving power grid stability by reducing frequency fluctuations and voltage deviations. During the training process, the reinforcement learning algorithm continuously interacts with the scenario simulation environment and iteratively optimizes the dispatching strategy, so that each decision can develop in the direction of global optimization.

[0113] The multi-model integration algorithm further optimizes the output results of the reinforcement learning algorithm and significantly improves the accuracy and stability of the scheduling strategy. The key steps of multi-model integration include the resampling mechanism, iterative optimization of historical prediction errors, and the comprehensive application of model combination methods. In the resampling mechanism, multiple models are trained by multiple samplings of scenario simulation data to effectively reduce the impact of data noise on the scheduling strategy; the iterative optimization of historical prediction errors gradually improves the model performance through multiple rounds of training, such as adjusting the model weights according to actual operation feedback; finally, the output results of multiple basic models are fused through model combination methods (such as the Stacking algorithm) to generate more robust optimal scheduling strategy data.

[0114] In one embodiment, in a regional power grid, initial scenario simulation data is constructed based on power demand forecast data and supply forecast data, including the uncertainty of load distribution state and wind power output. The reinforcement learning algorithm takes these states as input and outputs dynamic scheduling strategies through training, such as increasing the power of thermal power plants when wind power generation is lower than expected, and charging energy storage devices to balance the load. The multi-model integration algorithm further optimizes these preliminary strategies, eliminates errors caused by abnormal wind speed prediction through a resampling mechanism, and uses the Stacking algorithm to combine multiple rounds of optimization results to generate the final scheduling strategy. The results show that the scheduling strategy significantly reduces line losses in practical applications, reduces wind and solar power abandonment, and improves energy efficiency.

[0115] Through this process, the reinforcement learning algorithm of the present invention is combined with the multi-model integration algorithm to form an efficient power grid dispatch optimization method. In theory, this method achieves accurate description and optimization decision-making of the complex operation state of the power grid through multi-dimensional data modeling and strategy optimization; in practical applications, the reliability of the dispatch strategy is improved through multi-model integration, providing an intelligent solution for modern power grid dispatch.

[0116] Preferably, the multi-model integration algorithm includes: a resampling integration algorithm, a weighted optimization algorithm and a model stacking algorithm, wherein the resampling integration algorithm generates optimized simulation data by independently training different samples of scenario simulation data and performing weighted averaging on the model output results; the weighted optimization algorithm iteratively optimizes the model parameters by adjusting the error distribution of the weak learner; the model stacking algorithm generates optimal scheduling strategy data by constructing the outputs of multiple basic models and generating a high-order prediction model.

[0117] In the present invention, the multi-model integration algorithm optimizes the scenario simulation data of power grid dispatching by combining the resampling integration algorithm, the weighted optimization algorithm and the model stacking algorithm to generate high-precision optimization simulation data and optimal dispatching strategy data. This integrated algorithm design not only gives full play to the characteristics of each basic model, but also improves the robustness of scenario simulation and the reliability of dispatching strategy through multi-level optimization, thereby providing strong support for the intelligent dispatching of power grid operation.

[0118] In the initial stage, the resampling integration algorithm independently trains different samples of scenario simulation data, and performs weighted averaging on the output results of each model to generate optimized simulation data. The core idea of ​​the algorithm is to construct multiple data subsets through random resampling technology, and train an independent model on each subset, and finally weight the output results of multiple models to reduce the random error of a single model. In the power grid dispatch scenario, the resampling integration algorithm can capture the performance of the dispatch strategy under different operating conditions by independently training multiple scenario simulation data (such as load distribution in different time periods and changes in wind power output). For example, when simulating the impact of wind power fluctuations on power grid dispatch, the algorithm trains models for high wind speed, low wind speed and medium wind speed scenarios respectively, and performs weighted averaging on the results to obtain comprehensive optimized simulation data, which significantly improves the adaptability to wind power fluctuations.

[0119] Secondly, the weighted optimization algorithm iteratively optimizes the model parameters by adjusting the error distribution of weak learners, thereby improving the overall prediction accuracy of the model. Weak learners refer to situations where a single model cannot efficiently fit complex scenario data in the initial stage. The weighted optimization algorithm uses the error feedback mechanism to dynamically adjust the sample weights according to the error distribution in the previous round of model training, so that the next round of training focuses more on samples with larger errors. For example, in power grid load forecasting, for high-load nodes that may have prediction errors due to data scarcity, the weighted optimization algorithm will assign higher weights to high-load node samples in the next round of training, thereby improving prediction performance. This process uses multiple rounds of iterative optimization to enable the model to more accurately reflect the scheduling needs under different load scenarios.

[0120] The model stacking algorithm further improves the accuracy and robustness of the scheduling strategy by constructing the outputs of multiple basic models and generating a high-order prediction model. The core of the model stacking algorithm is to use the outputs of different basic models (such as models based on resampling and models based on weighted optimization) as input features to train a high-order meta-model, thereby achieving collaborative optimization between models. In one embodiment, the high-order prediction model can comprehensively consider the prediction results of each basic model, such as combining the output of the resampling model based on wind power prediction with the output of the weighted optimization model based on load fluctuations to generate more comprehensive scheduling strategy data. Taking a regional power grid as an example, when wind power and photovoltaic power generation fluctuate at the same time, the stacking model can integrate the outputs of different basic models, comprehensively adjust the load distribution of thermal power plants and energy storage equipment, and finally generate the optimal scheduling strategy to ensure the stability and economy of power grid operation.

[0121] Through the above process, the multi-model integration algorithm of the present invention theoretically realizes multi-level optimization of scenario simulation data, significantly improving the accuracy of simulation data and the applicability of scheduling strategies. In practical applications, the resampling integration algorithm reduces random errors in simulation data, the weighted optimization algorithm focuses on the precise fitting of key samples, and the model stacking algorithm further enhances the stability of prediction results through collaborative optimization. Taking the actual operation scenario of a certain power grid as an example, the combination of the three algorithms reduces the prediction error of the scheduling strategy to less than 2%, significantly improving the execution effect of the scheduling strategy, while reducing power generation costs and line losses, fully demonstrating the efficiency and reliability of the present invention in power grid scheduling optimization.

[0122] Dynamically adjust the power generation and load of power stations and grid nodes according to the optimal dispatching strategy data, and update the grid dispatching strategy.

[0123] In the present invention, the process of dynamically adjusting the power generation and load of power stations and grid nodes according to the optimal dispatching strategy data and updating the grid dispatching strategy is the core link to achieve efficient operation and dynamic optimization of the grid. By real-time collection and analysis of the grid operation status, combined with the guidance of the optimal dispatching strategy data, the system can achieve dynamic response in the power generation, transmission and distribution links, thereby maintaining the stability and economy of the grid in a complex and changeable operating environment.

[0124] Dynamically adjust the power generation of power stations according to the optimal dispatching strategy data. Power stations usually include wind power, photovoltaic, thermal power and energy storage, and each type has different output characteristics and regulation capabilities. During the dynamic adjustment process, the system optimizes the operating status of each power station in real time based on the optimal dispatching strategy data. For example, when the wind power output is insufficient, the thermal power generation is increased, and the energy storage equipment is discharged to make up for the load gap caused by insufficient wind power. For renewable energy power stations (such as wind power and photovoltaic), since their power generation is greatly affected by external conditions such as weather, the system will give priority to combining environmental data (such as wind speed and light intensity) and load forecast data to adjust their power generation plans in real time to ensure maximum energy utilization efficiency. For example, when the wind speed suddenly increases and the wind power output exceeds the load demand, the system will reduce the output of the thermal power unit according to the optimal dispatching strategy data to avoid wasting wind power resources and reduce the fuel consumption of thermal power.

[0125] In the load adjustment of power grid nodes, the system optimizes the power distribution plan by analyzing the real-time load distribution of each node and the power flow status of the transmission line to avoid grid instability caused by local node overload or line overload. In practical applications, this is usually achieved through intelligent load transfer. For example, when the load of a node is close to its upper limit, the system will transfer part of the load to other nodes by changing the load distribution ratio of adjacent nodes, and adjust the power flow of the corresponding lines to reduce the risk of overload. In addition, for power grids with distributed power sources and energy storage devices, the system will further optimize the load distribution of nodes by coordinating the charging and discharging plans of energy storage devices, thereby balancing the supply and demand relationship of the regional power grid.

[0126] The process of dynamically adjusting power generation and load will in turn affect the operating status of the power grid, and the system will update the power grid dispatch strategy in real time according to the new operating status. This update is based on the idea of ​​cyclic optimization. By continuously monitoring the operating status and iteratively optimizing the dispatch strategy, the system can quickly respond to environmental changes and emergencies. For example, in the case of a sudden increase in power grid load (such as peak power consumption in summer or sudden equipment failure), the system will update the operation plan of power stations and nodes in real time based on the optimal dispatch strategy data to ensure the stable operation of the power grid. At the same time, after the load returns to a stable state, the system will automatically adjust the strategy to restore the economic operation of the power grid.

[0127] In one embodiment, in a regional power grid dominated by wind power, a sudden decrease in wind speed leads to insufficient wind power output. The system quickly increases the power generation of thermal power units according to the optimal scheduling strategy data, and coordinates the energy storage equipment to release electricity. At the same time, since the load is mainly concentrated in a certain city node, the system adjusts the power flow of the transmission line to transfer part of the load to the backup node in the adjacent area, avoiding the occurrence of overload. After a period of time, the wind speed returns to normal, and the system reduces the thermal power output again and charges the energy storage equipment, while updating the scheduling strategy to adapt to the new operating status. In the end, this dynamic adjustment process significantly reduces the wind abandonment rate, improves energy utilization efficiency, and ensures the power supply security of the power grid.

[0128] Preferably, dynamically adjusting the power generation and load of the power station and the grid node according to the optimal dispatching strategy data includes:

[0129] Receive optimal dispatch strategy data and make real-time adjustments to the output power of power stations, including wind power stations, solar power stations and thermal power stations.

[0130] Optimize the load distribution strategy of power grid nodes, including: dynamic balancing based on load flow between nodes and load transfer to nodes that may be overloaded.

[0131] Generate dispatch execution data and apply the adjustment results to power plants and grid nodes in conjunction with the grid control system.

[0132] The present invention realizes real-time optimization of the power grid operation status by dynamically adjusting the power generation and load of the power station and the grid node according to the optimal dispatching strategy data. The adjustment process includes receiving the optimal dispatching strategy data and adjusting the output power of the power station in real time, optimizing the load distribution strategy of the grid node, and generating dispatching execution data and applying the adjustment results to the power station and the grid node in combination with the grid control system. This process combines multiple power generation methods with grid node management to ensure the economy and stability of the grid operation, while improving the responsiveness to dynamic demand.

[0133] In the first step, the system receives the optimal dispatching strategy data and adjusts the output power of the power station according to the real-time grid demand. Power stations include wind power stations, solar power stations and thermal power stations, which have significant differences in output characteristics and regulation capabilities. For wind power stations and solar power stations, since their output is limited by environmental factors such as wind speed and light intensity, the system will dynamically adjust them in combination with real-time environmental data. For example, when the wind speed increases, the system guides the wind power station to gradually increase its power output through the optimal dispatching strategy data, while reducing the output of the thermal power station, thereby reducing fuel consumption and improving the utilization efficiency of renewable energy. For thermal power stations, due to their slow adjustment speed and economic limitations, the system usually uses them as a means of regulating the basic load, and only adjusts the output of thermal power generation when renewable energy fluctuates greatly. For example, when photovoltaic power generation drops sharply in a short period of time due to sudden cloud cover, the system quickly increases the output of thermal power generation according to the optimal dispatching strategy data to avoid load gaps.

[0134] In the second step, the system optimizes the load distribution strategy of the grid nodes to ensure efficient transmission of electricity and safe operation of the nodes. The optimization process is based on the dynamic balance of load flows between nodes and the load transfer strategy for nodes that may be overloaded. The dynamic balance of load flows between nodes optimizes the flow of power by monitoring the operating status of the loads and transmission lines of each node in real time, and combining the topology of the grid to optimize the allocation of power. For example, when the power demand of a load node increases and causes its transmission line to be fully loaded, the system adjusts the load distribution of other nodes through the optimal scheduling strategy data to reduce the transmission pressure on the line, thereby achieving dynamic balance of the load of the entire network. At the same time, for nodes that may be overloaded, the system will distribute part of the load to adjacent nodes through the load transfer strategy. For example, when industrial loads are concentrated on a certain node, the system diverts the load to other nodes by changing the power flow path to avoid node tripping and power outages caused by overload.

[0135] Finally, the system generates dispatch execution data and applies the adjustment results to power stations and grid nodes in combination with the power grid control system. The generation of dispatch execution data is based on the optimized power station output power plan and node load distribution plan to ensure the operability and accuracy of the adjustment. For example, the dispatch execution data includes the target output power of the wind power station, the load change curve of the thermal power generating unit, and the distribution plan of the load at each node. The generated dispatch execution data is sent to each power station and node execution terminal through the power grid control system. The power station adjusts the output power through the automatic control module, and the node changes the load distribution through the intelligent load regulator. The entire execution process is monitored in real time by the power grid control system to ensure that the adjustment plan can be implemented accurately and efficiently.

[0136] In one embodiment, in a regional power grid actual operation scenario, when a sudden increase in wind speed causes wind power output to increase by 40%, the system reduces the output of thermal power generation by 30% through the optimal scheduling strategy data, and at the same time guides the excess wind power to the neighboring nodes with lower loads through load distribution optimization to avoid line overload. In this process, the scheduling execution data guides the wind power station to gradually increase power output, and realizes dynamic power balance between nodes through the intelligent load regulator. The results show that this adjustment scheme not only avoids the phenomenon of wind abandonment, but also reduces coal consumption by 15%, significantly improving the economy and environmental protection.

[0137] Preferably, Figure 4 As shown, updating the grid dispatching strategy includes: performing multi-level analysis on real-time feedback data of the grid operating status, wherein the multi-level analysis includes: generating dispatching optimization data based on analysis of the operating status of local grid nodes and trend analysis of the global grid status, and using the dispatching optimization data to update the dispatching strategy.

[0138] In the present invention, the process of updating the grid dispatching strategy is based on real-time feedback data, and dispatching optimization data is generated through multi-level analysis. The dispatching strategy is iteratively updated in combination with the optimization results, thereby achieving accurate adaptation and dynamic optimization of the grid operation status. Specifically, the multi-level analysis includes analysis based on the operation status of local grid nodes and trend analysis of the global grid status. This hierarchical design can not only capture the instant changes of local nodes, but also optimize the overall operation of the grid from a global perspective.

[0139] Collect and preprocess the real-time feedback data of the power grid operation status. The real-time feedback data mainly includes parameters such as voltage, current, load conditions, frequency deviation, and line power flow of each node. These data are collected by sensors and monitoring equipment deployed in the power grid, and are cleaned, denoised, and formatted by edge computing devices to ensure the accuracy and real-time nature of the data. For example, in a regional power grid, real-time feedback data may show that the load of a certain node is close to overload, and the power flow of some lines is abnormally increased, which provides basic input for subsequent multi-level analysis.

[0140] Based on the analysis of the operating status of local grid nodes, the system independently evaluates the operating status of each node. Local node analysis focuses on load distribution, voltage deviation, and power exchange between nodes. For example, when the load of a node is continuously higher than the set threshold, the system marks the node as a high-risk node and formulates a load transfer plan based on its connection relationship with adjacent nodes. In addition, for nodes with large voltage deviations, the system proposes adjustment measures based on the operating status of the power factor correction equipment to ensure voltage stability. In practical applications, this local analysis can quickly identify and handle local fault hazards that may occur in the power grid. For example, when the power demand of a load node in an industrial park suddenly increases, the system detects that the node is overloaded through local analysis and recommends transferring part of the load to adjacent nodes to avoid node tripping and power outages.

[0141] The trend analysis of the global power grid status evaluates the operation trend and load change pattern of the power grid as a whole, and deduces the future state of the power grid by combining historical data and forecast data. The focus of the global analysis includes the overall balance of power generation and load, the power flow distribution of the trunk line, and the volatility of renewable energy power generation. For example, in a power grid with a high proportion of wind power, the global analysis can identify the short-term and drastic fluctuation trend of wind power output, and adjust the operation plan of thermal power plants and energy storage equipment in advance based on the overall network dispatch strategy to cope with the impact of wind power fluctuations on the overall network balance. This trend analysis can not only predict the operational challenges that the power grid may face, but also provide global guidance for the optimization of dispatch strategies.

[0142] The dispatch optimization data generated by multi-level analysis includes adjustment suggestions at the local and global levels, which are further used to update the grid dispatch strategy. The dispatch strategy is updated in an iterative optimization manner, combining real-time feedback data, historical operation records, and prediction results to dynamically adjust the dispatch strategy. For example, the updated dispatch strategy may include load transfer plans for high-load nodes, energy storage equipment charging and discharging plans for wind power fluctuations, and line operation optimization plans for abnormal power flow. These strategies are integrated into the dispatch execution system and sent to each device for implementation through the grid control system, forming a closed-loop optimization process from data collection to decision execution.

[0143] In one embodiment, when a regional power grid experiences overload at some nodes during the peak period of electricity consumption in summer, the system discovers through real-time feedback data that the load at the node has exceeded the safety threshold. Local analysis suggests short-term relief by adjusting the load distribution of adjacent nodes and reducing non-critical loads; global analysis combines load forecast data for the next few hours with wind power output forecast data, proposes a plan to increase load scheduling on the trunk line, and instructs thermal power plants to gradually increase output power to support the balance of the entire network. Ultimately, the updated scheduling strategy was implemented through the system, successfully avoiding node tripping and line overload, while ensuring the continuity and economy of power supply.

[0144] like Figure 5 As shown, a system for implementing the power grid dispatching method comprises:

[0145] The data collection module is used to collect environmental data, social activity data, equipment health data, and economic and market data, and pre-process and integrate the environmental data, social activity data, equipment health data, and economic and market data to generate fused data; through the data collection module, the system collects environmental data, social activity data, equipment health data, and economic and market data in real time. These data are pre-processed and integrated by edge computing devices to generate fused data containing multi-dimensional information, providing high-quality input for subsequent deep learning models.

[0146] The multi-model integrated deep learning module is used to perform power demand forecasting and power supply forecasting based on fused data, and generate power demand forecasting data and power supply forecasting data. The multi-model integrated deep learning module includes: a long short-term memory network model, which is used to extract time series features and generate preliminary forecasting data.

[0147] Transformer model, used to extract global features and optimize preliminary prediction data based on the self-attention mechanism.

[0148] The gated recurrent unit model is used to perform feature correction on the optimized prediction data to generate final power demand prediction data and power supply prediction data.

[0149] Based on fused data, this module extracts time series features through the long short-term memory network model (LSTM) to identify the long-term trends and periodic fluctuations in electricity demand and supply; combines the self-attention mechanism of the Transformer model to extract global features and optimize the preliminary prediction results; uses the gated recurrent unit model (GRU) to further correct short-term fluctuations and abnormal data to generate high-precision electricity demand forecast data and electricity supply forecast data.

[0150] The reinforcement learning and multi-model integrated algorithm module is used to simulate power grid dispatching scenarios based on power demand forecast data and power supply forecast data, and generate scenario simulation data and optimal dispatching strategy data; it uses the reinforcement learning algorithm to dynamically optimize the power grid dispatching strategy in the scenario simulation environment, and generates scenario simulation data and optimal dispatching strategy data through the collaborative design of state space (such as node load, power plant status), action space (such as load distribution, power generation adjustment) and reward function.

[0151] The dispatch execution module is used to dynamically adjust the power generation of power plants and the load of grid nodes according to the optimal dispatch strategy data, and update the grid dispatch strategy in real time. According to the optimal dispatch strategy data, the power output of wind power, photovoltaic and thermal power plants, as well as the load distribution of grid nodes are dynamically adjusted. At the same time, the grid operation status is monitored in real time and the dispatch strategy is updated to achieve closed-loop optimization.

[0152] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A power grid dispatching method, characterized in that: The following steps are involved: Collect environmental data, social activity data, equipment health data, and economic and market data, and use edge computing devices to pre-process and integrate the collected data to generate fused data; Based on the fusion data, a multi-model integrated deep learning model is used to perform power demand forecasting and power supply forecasting to generate power demand forecasting data and power supply forecasting data. The multi-model integrated deep learning model includes: long short-term memory network model, Transformer model and gated recurrent unit model; the processing steps of the multi-model integrated deep learning model include: The long short-term memory network model is used to extract time series features from the fused data, including: modeling the long-term trend of electricity demand changes and identifying periodic fluctuation patterns; Use the Transformer model to perform feature weighting on key nodes in time series features based on the self-attention mechanism, extract global correlation features, and generate intermediate prediction data; Use the gated recurrent unit model to perform feature correction on the intermediate forecast data, including: removing abnormal data points and adjusting short-term fluctuations to generate the final power demand forecast data and power supply forecast data; Based on the power demand forecast data and power supply forecast data, the power grid dispatch scenario is simulated using the reinforcement learning algorithm and the multi-model integration algorithm to generate scenario simulation data and optimal dispatch strategy data; Dynamically adjust the power generation and load of power stations and grid nodes according to the optimal dispatching strategy data, and update the grid dispatching strategy.

2. The power grid dispatching method according to claim 1, characterized in that: The environmental data includes: temperature data, humidity data, wind speed data, light intensity data and air quality data; the social activity data includes: pedestrian flow data and traffic flow data; the equipment health data includes: equipment vibration data, equipment surface temperature data and infrared imaging data; the economic and market data includes: real-time energy market price data and carbon emission trading price data.

3. The power grid dispatching method according to claim 1, characterized in that: When processing fused data, the long short-term memory network model uses input gates, forget gates, and output gates to dynamically adjust the input state and forget state of the time series. The Transformer model uses position encoding and multi-head self-attention mechanisms to analyze the temporal dependency and multi-dimensional correlation of time series features. The gated recurrent unit model controls the weight distribution between input data and historical states through gated units.

4. The power grid dispatching method according to claim 1, characterized in that: Reinforcement learning algorithms and multi-model integration algorithms are used to simulate power grid dispatch scenarios, including: Construct initial scenario simulation data based on power demand forecast data and power supply forecast data, and train through scenario simulation environment; The training process is defined by using a reinforcement learning algorithm through a state space, an action space, and a reward function. The state space includes: the current load distribution state of the power grid and the operating state of the power station; the action space includes: the adjustment plan of the dispatching strategy; and the reward function is designed based on the stability of the power grid operation and energy efficiency. The output results of the reinforcement learning algorithm are optimized through a multi-model integration algorithm, including: using a resampling mechanism for model training, iteratively optimizing historical prediction errors, and generating optimal scheduling strategy data through model combination.

5. The power grid dispatching method according to claim 4, characterized in that: Multi-model integration algorithms include: resampling integration algorithm, weighted optimization algorithm and model stacking algorithm. Among them, the resampling integration algorithm generates optimized simulation data by independently training different samples of scenario simulation data and performing weighted averaging of the model output results; the weighted optimization algorithm iteratively optimizes the model parameters by adjusting the error distribution of the weak learner; the model stacking algorithm generates optimal scheduling strategy data by constructing the outputs of multiple basic models and generating a high-order prediction model.

6. The power grid dispatching method according to claim 1, characterized in that: Dynamically adjust the power generation and load of power stations and grid nodes based on the optimal dispatch strategy data, including: Receive optimal dispatch strategy data and make real-time adjustments to the output power of power stations, including wind power stations, solar power stations, and thermal power stations; Optimize the load distribution strategy of power grid nodes, including: dynamic balancing of load flows between nodes and load transfer to nodes that may be overloaded; Generate dispatch execution data and apply the adjustment results to power plants and grid nodes in conjunction with the grid control system.

7. The power grid dispatching method according to claim 1, characterized in that: Updating the grid dispatching strategy includes: performing multi-level analysis on real-time feedback data of the grid operation status, wherein the multi-level analysis includes: generating dispatching optimization data based on analysis of the operation status of local grid nodes and trend analysis of the global grid status, and using the dispatching optimization data to update the dispatching strategy.

8. A system for implementing the power grid dispatching method according to any one of claims 1 to 7, characterized in that: include: A data collection module, used to collect environmental data, social activity data, equipment health data, and economic and market data, and pre-process and integrate the environmental data, social activity data, equipment health data, and economic and market data to generate fused data; The multi-model integrated deep learning module is used to perform power demand forecasting and power supply forecasting based on the fused data, and generate power demand forecasting data and power supply forecasting data, wherein the multi-model integrated deep learning module includes: Long short-term memory network model, used to extract time series features and generate preliminary forecast data; Transformer model, which is used to extract global features and optimize preliminary prediction data based on the self-attention mechanism; A gated recurrent unit model is used to perform feature correction on the optimized forecast data to generate final power demand forecast data and power supply forecast data; Reinforcement learning and multi-model integration algorithm module, which is used to simulate power grid dispatch scenarios based on power demand forecast data and power supply forecast data, and generate scenario simulation data and optimal dispatch strategy data; The dispatching execution module is used to dynamically adjust the power generation of the power station and the load of the grid node according to the optimal dispatching strategy data, and update the grid dispatching strategy in real time.

Images