Wind field prediction system based on multi-scale image convolution and reinforcement learning

Abstract

The invention discloses a wind field prediction system based on multi-scale image convolution and reinforcement learning, and the system carries out the dynamic modeling through a multi-scale image convolution network, and comprehensively captures the spatial-temporal characteristics of a wind field. Through the reinforcement learning module, the model can optimize a strategy in real time in a complex and dynamically changing wind field environment so as to improve the path planning and obstacle avoidance capabilities of the aircraft. The reinforcement learning algorithm dynamically adjusts a strategy according to an accumulation reward mechanism, so that the intelligent agent can quickly adapt to wind field changes, and the flight safety is guaranteed. According to the invention, for the dynamic change of the wind field in the complex meteorological environment, the graph convolutional neural network and the reinforcement learning optimization strategy are combined, and the accuracy and real-time performance of wind field prediction are improved. The spatial-temporal characteristics of the wind field are modeled and analyzed through the multi-scale image convolutional network, and then the prediction model is optimized through reinforcement learning, so that the system can provide more accurate and more real-time wind field data support to ensure the flight safety and path optimization of an aircraft under complex meteorological conditions.

Description

Technical Field

[0001] The present invention belongs to the field of aircraft navigation technology, and in particular relates to a wind field prediction system based on multi-scale graph convolution and reinforcement learning. Background Art

[0002] Aircraft trajectory planning and flight safety assessment are closely related to accurate wind field prediction. Factors such as wind speed, wind direction, and turbulence directly affect the flight stability, range accuracy, and fuel consumption of aircraft. Especially under severe weather conditions, sudden changes and extreme changes in wind fields pose challenges to flight safety. Traditional wind field prediction methods, such as numerical weather prediction (NWP) models, can provide certain wind speed and wind direction predictions, but due to their limitations in small scale and real-time updates, they cannot meet the dynamic response requirements of aircraft during flight.

[0003] Existing wind field prediction technologies mainly rely on static numerical prediction models, which usually operate at fixed time and spatial resolutions and have difficulty capturing rapidly changing local wind fields, especially in predicting short-term, high-frequency changes in wind speed and direction. In addition, the time delay of traditional models also limits their ability to update in real time, making it impossible for aircraft to quickly adjust their tracks based on real-time wind field data. Especially under complex meteorological conditions, the prediction capabilities of traditional methods cannot meet the needs of dynamic environments.

[0004] Graph convolutional neural networks (GCNs) have become a powerful tool for wind farm forecasting because they can effectively process spatiotemporal graph data and capture spatial dependencies and time series changes between nodes. Unlike traditional convolutional neural networks, graph convolutional networks can better handle the spatial and temporal complexity of wind farm data, especially in establishing effective connections between multi-level data and data from aircraft at different flight altitudes. Through multi-scale modeling, GCNs can capture global and local changes in wind farms and improve the accuracy and adaptability of forecasts.

[0005] At the same time, reinforcement learning (RL), as a technology that can perform adaptive optimization in a dynamic environment, can further enhance the real-time and robustness of wind field forecasting. Reinforcement learning optimizes decisions through interaction with the environment and can adjust the forecasting strategy based on real-time feedback, so that wind field forecasting can dynamically respond to meteorological changes. The wind field forecasting method that combines graph convolutional networks and reinforcement learning can effectively make up for the shortcomings of traditional methods, provide high-precision and real-time updated wind field data support, and significantly improve the safety of aircraft and the accuracy of range forecasting. Summary of the invention

[0006] In view of this, the present invention aims to overcome the shortcomings of the above-mentioned problems in the prior art, and proposes a wind field prediction system based on multi-scale graph convolution and reinforcement learning, aiming to overcome the problems of insufficient accuracy, real-time and adaptability in the existing wind field prediction technology, and provide efficient and accurate wind field prediction data support for aircraft.

[0007] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0008] A wind field prediction system based on multi-scale graph convolution and reinforcement learning, comprising a data input module, a graph convolution module, a reinforcement learning strategy optimization module, and a cumulative reward calculation and result output module connected in sequence;

[0009] The data input module obtains aircraft communication addressing and reporting system data from the ground station and wind speed, wind direction, temperature and air pressure data from the global forecast system, and integrates key data of the environmental wind field based on the aircraft's spatiotemporal position;

[0010] The graph convolution module performs multi-scale graph structure modeling on the input wind field data, extracts the spatial and spatiotemporal characteristics of the wind field through the graph convolution network, and generates accurate wind field prediction results;

[0011] The reinforcement learning strategy optimization module uses the spatiotemporal features generated by the graph convolution module to optimize the prediction accuracy of the wind field prediction model through a reinforcement learning algorithm;

[0012] The cumulative reward calculation and result output module calculates the wind field prediction error according to the cumulative reward, outputs the optimized wind field prediction result, improves the accuracy and stability of the wind field prediction, and ensures the safety and flight efficiency of the aircraft in a complex meteorological environment.

[0013] Furthermore, the data input module specifically includes:

[0014] The ground station data acquisition unit is used to receive and decode ACARS data in real time, capture the position information of the aircraft at a specific time point, and provide a spatial basis for subsequent space-time graph modeling;

[0015] The GFS data acquisition unit is used to access the global forecast system data in real time, collect wind speed, wind direction, temperature and air pressure data with matching spatiotemporal resolution, combine it with the aircraft position information to generate a spatiotemporal input data graph, and normalize it to meet the input requirements of the graph convolutional network.

[0016] Furthermore, the graph convolution module specifically includes:

[0017] Adjacency matrix construction unit: Generates the graph structure adjacency matrix of the wind farm based on the geographical location and time point of the spatiotemporal data, defines the weights between nodes based on the similarity of location and wind farm parameters, and accurately captures the spatial distribution and change characteristics of the wind farm;

[0018] Multi-scale graph convolution unit: extracts wind field features at local and global scales through multi-scale graph convolution operations, adapts to the change patterns of different scales in the wind field, and thus obtains multi-level wind field information;

[0019] Activation unit: The nonlinear activation function is used to improve the feature representation ability after graph convolution and enhance the model's expression effect on complex wind field data, so as to better reflect the impact of high-variability areas in the wind field.

[0020] Furthermore, the reinforcement learning strategy optimization module includes:

[0021] State estimation unit: The spatiotemporal graph embedding features generated by the graph convolution module are used as the state input of the current wind farm. The accuracy of state estimation is ensured by selecting a state representation that adapts to the dynamic changes of the wind farm.

[0022] Policy optimization unit: Apply the policy gradient and Q learning algorithms in reinforcement learning, use the current state and accumulated rewards to update the strategy, and further optimize the prediction accuracy of the wind farm prediction model.

[0023] Furthermore, the cumulative reward calculation and result output module includes:

[0024] Cumulative reward calculation unit: Calculates cumulative rewards based on the model's prediction results, actual wind field measurement data, and flight safety parameters, covering prediction accuracy and wind field stability, so as to provide effective feedback and adjustments to the wind field prediction strategy;

[0025] The result output unit generates and outputs optimized wind field prediction results, ensuring that the flight control system can obtain accurate wind field data in a timely manner, thereby improving the accuracy of navigation decisions and flight safety.

[0026] Furthermore, the implementation process of the wind field prediction system includes:

[0027] Step 1: Collect and process flight data and wind field data from ACARS and GFS through the data input module to generate spatiotemporal graph data input;

[0028] Step 2: Use the graph convolution module to process the input wind field map data, extract the multi-scale features of the wind field and generate spatiotemporal feature embedding;

[0029] Step 3: Dynamically adjust the wind farm prediction strategy according to the accumulated reward value through the reinforcement learning strategy optimization module to improve the accuracy of wind farm prediction;

[0030] Step 4: Calculate and output the optimized wind field prediction results in the cumulative reward calculation and result output module to provide real-time wind field prediction data support for the flight control system.

[0031] Compared with the prior art, the wind field prediction system based on multi-scale graph convolution and reinforcement learning described in the present invention has the following advantages:

[0032] This paper organically combines graph convolutional networks with reinforcement learning technology, and innovatively proposes an overall architecture of multi-scale graph convolutional wind field modeling and reinforcement learning strategy optimization, which has the following advantages:

[0033] High-precision wind field modeling: The wind field modeling method based on multi-scale graph convolution can capture spatial and temporal characteristics of different scales, significantly improving the accuracy of wind field prediction;

[0034] Real-time path optimization: Through the reinforcement learning algorithm, the strategy is optimized based on the graph convolution wind field characteristics to provide real-time and dynamic path planning solutions for the aircraft;

[0035] Adaptability and robustness: The policy optimization mechanism based on cumulative rewards can actively adjust the path planning decision when the wind field changes, improving the safety and stability of the aircraft in complex wind field environments;

[0036] Aviation safety assurance: The present invention can provide wind field characteristics and optimized path information in real time, helping aircraft avoid high-risk areas under adverse weather conditions, and effectively improving the overall safety of civil aviation flights.

[0037] The present invention is not only applicable to the path optimization of civil aircraft, but can also be extended to the navigation and obstacle avoidance application scenarios of drones and other aircraft, providing an effective and innovative solution for flight safety technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:

[0039] Figure 1 This is a schematic diagram of the structure of a wind field prediction system based on multi-scale graph convolution and reinforcement learning of the present invention. DETAILED DESCRIPTION

[0040] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0041] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and the like are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0042] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood by specific circumstances.

[0043] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0044] like Figure 1 As shown, the present invention provides a wind field prediction system based on multi-scale graph convolution and reinforcement learning, including a data input module, a graph convolution module, a reinforcement learning strategy optimization module, and a cumulative reward calculation and result output module connected in sequence;

[0045] The data input module obtains aircraft communication addressing and reporting system data from the ground station and wind speed, wind direction, temperature and air pressure data from the global forecast system, and integrates key data of the environmental wind field based on the aircraft's spatiotemporal position;

[0046] The graph convolution module performs multi-scale graph structure modeling on the input wind field data, extracts the spatial and spatiotemporal characteristics of the wind field through the graph convolution network, and generates accurate wind field prediction results;

[0047] The reinforcement learning strategy optimization module uses the spatiotemporal features generated by the graph convolution module to optimize the prediction accuracy of the wind field prediction model through a reinforcement learning algorithm;

[0048] The cumulative reward calculation and result output module calculates the wind field prediction error according to the cumulative reward, outputs the optimized wind field prediction result, improves the accuracy and stability of the wind field prediction, and ensures the safety and flight efficiency of the aircraft in a complex meteorological environment.

[0049] Specifically, the data input module specifically includes:

[0050] The ground station data acquisition unit is used to receive and decode ACARS data in real time, capture the position information of the aircraft at a specific time point, and provide a spatial basis for subsequent space-time graph modeling;

[0051] The GFS data acquisition unit is used to access the global forecast system data in real time, collect wind speed, wind direction, temperature and air pressure data with matching spatiotemporal resolution, combine it with the aircraft position information to generate a spatiotemporal input data graph, and normalize it to meet the input requirements of the graph convolutional network.

[0052] Specifically, the graph convolution module specifically includes:

[0053] Adjacency matrix construction unit: Generates the graph structure adjacency matrix of the wind farm based on the geographical location and time point of the spatiotemporal data, defines the weights between nodes based on the similarity of location and wind farm parameters, and accurately captures the spatial distribution and change characteristics of the wind farm;

[0054] Multi-scale graph convolution unit: extracts wind field features at local and global scales through multi-scale graph convolution operations, adapts to the change patterns of different scales in the wind field, and thus obtains multi-level wind field information;

[0055] Activation unit: The nonlinear activation function is used to improve the feature representation ability after graph convolution and enhance the model's expression effect on complex wind field data, so as to better reflect the impact of high-variability areas in the wind field.

[0056] Specifically, the reinforcement learning strategy optimization module includes:

[0057] State estimation unit: The spatiotemporal graph embedding features generated by the graph convolution module are used as the state input of the current wind farm. The accuracy of state estimation is ensured by selecting a state representation that adapts to the dynamic changes of the wind farm.

[0058] Policy optimization unit: Apply the policy gradient and Q learning algorithms in reinforcement learning, use the current state and accumulated rewards to update the strategy, and further optimize the prediction accuracy of the wind farm prediction model.

[0059] Specifically, the cumulative reward calculation and result output module includes:

[0060] Cumulative reward calculation unit: Calculates cumulative rewards based on the model's prediction results, actual wind field measurement data, and flight safety parameters, covering prediction accuracy and wind field stability, so as to provide effective feedback and adjustments to the wind field prediction strategy;

[0061] The result output unit generates and outputs optimized wind field prediction results, ensuring that the flight control system can obtain accurate wind field data in a timely manner, thereby improving the accuracy of navigation decisions and flight safety.

[0062] The implementation process of the wind field prediction system of the present invention includes:

[0063] Step 1: Collect and process flight data and wind field data from ACARS and GFS through the data input module to generate spatiotemporal graph data input;

[0064] Step 1.1: Decode the ACARS data through the ARINC 618 protocol to obtain the flight altitude, speed, heading, and surrounding wind speed, temperature, and air pressure data, and obtain the wind speed, wind direction, temperature, and air pressure data at the corresponding time from the GFS meteorological database;

[0065] Step 1.2: For the values ​​in the data collected in step 1 that deviate from the normal range, the outlier detection method of adversarial training is used to generate adversarial samples to enhance the robustness of the model to abnormal data. This method is particularly suitable for meteorological data with strong temporal and spatial characteristics such as wind speed data. By generating and identifying adversarial samples through adversarial training, the model can learn more detailed abnormal data features, thereby improving the outlier detection effect in wind speed prediction;

[0066] Step 1.3: For the missing parts and outliers in the data collected in step 1 and step 1.1, linear interpolation is used to fill them. The missing values ​​are linearly calculated based on the values ​​of the two previous and next time points. is the estimate of the missing value, t i is the time of missing values, y i+1 and i-1 is the value of the previous and next time points, and the formula is:

[0067]

[0068] Step 1.4: Standardize the data processed in step 1.3 and use the processed data as the data set for model training;

[0069] Step 1.5: Use the obtained data to construct a wind field prediction training set and test set, input the test set into the graph convolution reinforcement learning model, and the training set is D train , the test set is D text , each node x i(t) A wind field feature vector, including wind speed, wind direction, temperature, and air pressure. The feature vector of each node is x i (t) = [v i (t),θ i (t),T i (t),p i (t)] T , v i (t) represents the wind speed of node i at time t, θ i (t) represents the wind direction of node i at time t, T i (t) represents the temperature of node i at time t, p i (t) represents the air pressure of node i at time t.

[0070] For time t, the node feature matrix of the graph is represented as X t :

[0071]

[0072] Among them, N represents the number of nodes, indicating the position in the wind field, and D represents the feature dimension, indicating the number of observation features contained in each node.

[0073] Step 2: Use the graph convolution module to process the input wind field map data, extract the multi-scale features of the wind field and generate spatiotemporal feature embedding;

[0074] Step 2.1: The spatial structure of the wind field depends on the adjacency matrix A t , the adjacency matrix reflects the spatial connectivity between the measurement points in the wind field and determines which nodes have information transmission. In the spatiotemporal graph data, each element A of the adjacency matrix ij , represents the connection strength between nodes i and j, and calculates the similarity of node features to quantify the degree of association between nodes to construct A t :

[0075]

[0076] In this way, the adjacency matrix A t The similarity between nodes can be reflected based on input data such as wind speed, wind direction, temperature and air pressure.

[0077] Step 2.2: In order to prevent the node degree from having too great an impact on feature propagation, the present invention uses a standardized adjacency matrix to adjacency matrix. t The formula is:

[0078]

[0079] element a ijrepresents the relationship weight between nodes i and j. This weight represents the spatial correlation or similarity between wind field nodes at different measurement points. If the wind speed, wind direction, temperature, air pressure and other data of nodes i and j are highly similar or affect each other's wind field characteristics, then a ij The value will be higher.

[0080] Step 2.3: Perform normalization to obtain a standardized adjacency matrix

[0081]

[0082] Step 2.4: Aggregate the features of each node through the adjacency matrix A t And the normalized adjacency matrix Get the feature matrix H t :The feature matrix of the kth layer is

[0083]

[0084] is the output feature matrix of the k-1th layer, with k = 0 at the initial moment. W (k) is the weight matrix of the kth layer, which is used to perform linear transformation on the features, and σ is the ReLU activation function, which is used to capture the nonlinear characteristic changes of the wind field.

[0085] Step 3: Dynamically adjust the wind farm prediction strategy according to the cumulative reward value through the reinforcement learning strategy optimization module to improve the accuracy of wind farm prediction; the goal of the reinforcement learning strategy optimization module is to continuously optimize the strategy π(s t ,a t |θ) to maximize the accuracy of wind field prediction. Specifically include:

[0086] Step 3.1: State space s t Represents the wind field characteristics at the current time t, which includes the wind speed, wind direction, temperature and air pressure of the node. This information is extracted and encoded into a state vector through the graph convolutional network. The state expression of the wind field is:

[0087]

[0088] Step 3.2: Policy function π(s t ,a t |θ) is learned through a neural network, based on the current state of the wind field s t Probability distribution of output actions:

[0089]

[0090] Where Q(st ,a t ; θ) is the action value function, which means taking action a under the current wind field state t The corresponding value, Q(s t ,a′;θ) represents the current state s t The expected cumulative reward that can be obtained by taking an action a′.

[0091] Step 3.3: In order to evaluate the contribution of a specific action to the cumulative reward, we introduce the advantage function A(s t ,a t ), which is expressed as:

[0092] A(s t ,a t )=Q(s t ,a t )-V(s t );

[0093] Among them, Q(s t ,a t ) means in state s t Select action a t The expected reward after s is used to help the model identify the optimal decision in a specific wind farm scenario. The value function V(s t ) is the estimated wind field state s t The expected cumulative reward is expressed as:

[0094]

[0095] E π [] represents the expectation of a random variable under a given policy π. γ k Represents the discount weight multiplied by the reward after time step k. t+k Represents the immediate reward obtained at time step t+k.

[0096] Step 3.4: To further enhance the ability of policy optimization, we combine the Q-learning method with the value function. Q-learning calculates the value Q(s) of each state-action pair. t ,a t ) to evaluate the effect of taking a certain action. The update formula of the Q-function is:

[0097]

[0098] where r t+1 is the reward at time t+1, γ is the discount factor, and this update mechanism enables the model to gradually optimize the wind field forecast. α is the learning rate, which controls the update speed of new information.

[0099] Step 3.5: Update the strategy through the policy gradient method so that the selected action a t In the current state t The expected reward can be maximized by:

[0100]

[0101] E π [] represents the expectation of a random variable under a given policy π.

[0102] The update formula of θ is:

[0103]

[0104] Where η is the learning rate of the strategy, which updates the strategy parameters to maximize the cumulative reward in wind field changes and improve the accuracy of wind field prediction.

[0105] Step 3.6: Use the cumulative reward J(θ) as the evaluation indicator to reflect the effectiveness of the model's strategy under complex wind field conditions: The expression of J(θ) is:

[0106]

[0107] By maximizing the objective function J(θ), we can find the optimal policy π(s t ,a t |θ), thereby improving the accuracy of wind field prediction.

[0108] Step 4: Calculate and output the optimized wind field prediction results in the cumulative reward calculation and result output module to provide real-time wind field prediction data support for the flight control system.

[0109] Step 4.1: After graph convolution operation and reinforcement learning strategy optimization, the final model outputs the wind farm prediction results The predicted value is outputted by the graph convolutional network H L Perform full connection layer mapping and predict the results The expression is:

[0110]

[0111] W f is the weight matrix of the fully connected layer, which is responsible for mapping the output of the graph convolutional network to the prediction result space, and b f is a bias term. The model achieves high-precision prediction of dynamic wind field changes by continuously optimizing the reinforcement learning strategy.

[0112] Step 4.2: The loss function uses mean square error (MSE) to measure the gap between the predicted value and the true value to improve the accuracy of the prediction.

[0113]

[0114] Where T is the total number of time steps, is the predicted output of the model at time step t, Y t is the corresponding real wind field state.

[0115] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A wind field prediction system based on multi-scale graph convolution and reinforcement learning, characterized by: It includes a data input module, a graph convolution module, a reinforcement learning strategy optimization module, and a cumulative reward calculation and result output module that are connected in sequence; The data input module obtains aircraft communication addressing and reporting system data from the ground station and wind speed, wind direction, temperature and air pressure data from the global forecast system, and integrates key data of the environmental wind field based on the aircraft's spatiotemporal position; The graph convolution module performs multi-scale graph structure modeling on the input wind field data, extracts the spatial and spatiotemporal characteristics of the wind field through the graph convolution network, and generates accurate wind field prediction results; The reinforcement learning strategy optimization module uses the spatiotemporal features generated by the graph convolution module to optimize the prediction accuracy of the wind field prediction model through a reinforcement learning algorithm; The cumulative reward calculation and result output module calculates the wind field prediction error according to the cumulative reward, outputs the optimized wind field prediction result, improves the accuracy and stability of the wind field prediction, and ensures the safety and flight efficiency of the aircraft in a complex meteorological environment.

2. The wind field prediction system based on multi-scale graph convolution and reinforcement learning according to claim 1, characterized in that: The data input module specifically includes: The ground station data acquisition unit is used to receive and decode ACARS data in real time, capture the position information of the aircraft at a specific time point, and provide a spatial basis for subsequent space-time graph modeling; The GFS data acquisition unit is used to access the global forecast system data in real time, collect wind speed, wind direction, temperature and air pressure data with matching spatiotemporal resolution, combine it with the aircraft position information to generate a spatiotemporal input data graph, and normalize it to meet the input requirements of the graph convolutional network.

3. A wind field prediction system based on multi-scale graph convolution and reinforcement learning according to claim 1, characterized in that: The graph convolution module specifically includes: Adjacency matrix construction unit: Generates the graph structure adjacency matrix of the wind farm based on the geographical location and time point of the spatiotemporal data, defines the weights between nodes based on the similarity of location and wind farm parameters, and accurately captures the spatial distribution and change characteristics of the wind farm; Multi-scale graph convolution unit: extracts wind field features at local and global scales through multi-scale graph convolution operations, adapts to the change patterns of different scales in the wind field, and thus obtains multi-level wind field information; Activation unit: The nonlinear activation function is used to improve the feature representation ability after graph convolution and enhance the model's expression effect on complex wind field data, so as to better reflect the impact of high-variability areas in the wind field.

4. The wind field prediction system based on multi-scale graph convolution and reinforcement learning according to claim 1, characterized in that: The reinforcement learning strategy optimization module includes: State estimation unit: The spatiotemporal graph embedding features generated by the graph convolution module are used as the state input of the current wind farm. The accuracy of state estimation is ensured by selecting a state representation that adapts to the dynamic changes of the wind farm. Policy optimization unit: Apply the policy gradient and Q learning algorithms in reinforcement learning, use the current state and accumulated rewards to update the strategy, and further optimize the prediction accuracy of the wind farm prediction model.

5. The wind field prediction system based on multi-scale graph convolution and reinforcement learning according to claim 1, characterized in that: The cumulative reward calculation and result output module includes: Cumulative reward calculation unit: Calculates cumulative rewards based on the model's prediction results, actual wind field measurement data, and flight safety parameters, covering prediction accuracy and wind field stability, so as to provide effective feedback and adjustments to the wind field prediction strategy; The result output unit generates and outputs optimized wind field prediction results, ensuring that the flight control system can obtain accurate wind field data in a timely manner, thereby improving the accuracy of navigation decisions and flight safety.

6. The wind field prediction system based on multi-scale graph convolution and reinforcement learning according to claim 1, characterized in that: The implementation process of the wind field prediction system includes: Step 1: Collect and process flight data and wind field data from ACARS and GFS through the data input module to generate spatiotemporal graph data input; Step 2: Use the graph convolution module to process the input wind field map data, extract the multi-scale features of the wind field and generate spatiotemporal feature embedding; Step 3: Dynamically adjust the wind farm prediction strategy according to the accumulated reward value through the reinforcement learning strategy optimization module to improve the accuracy of wind farm prediction; Step 4: Calculate and output the optimized wind field prediction results in the cumulative reward calculation and result output module to provide real-time wind field prediction data support for the flight control system.

Images