Urban road meteorological condition prediction method fusing vehicle-mounted and fixed meteorological stations

Abstract

The invention provides an urban road meteorological condition prediction method fusing vehicle-mounted and fixed meteorological stations, and the method comprises the following steps: 1, building a data sharing network of a mobile vehicle meteorological station and a fixed meteorological station through an Internet of Vehicles technology, and carrying out the collection of meteorological data through the data sharing network; and step 2, establishing a meteorological condition prediction module based on a meteorological station, performing data preprocessing and analysis, and completing prediction of urban road meteorological conditions based on a long short-term memory network and a graph convolutional neural network. The method aims at solving the problem that the prediction precision is not high due to the fact that the number of traditional fixed meteorological stations is insufficient, and can cover a wider urban road network. By combining the layout scheme of the mobile vehicle-mounted weather station and the fixed weather station, the method has high adaptability and expandability, and can flexibly meet different road traffic network requirements. In addition, deployment and adjustment can be carried out according to actual application conditions so as to meet the requirements of various road meteorological condition prediction tasks.

Description

Technical Field

[0001] The invention relates to a method for predicting urban road meteorological conditions by integrating vehicle-mounted and fixed meteorological stations. Background Art

[0002] Meteorological conditions directly affect the safety of urban roads. For example, in low-temperature and freezing weather, the road surface is prone to ice, increasing the risk of vehicle loss of control; in foggy weather, visibility is poor, increasing the risk of rear-end collisions; and inclement weather such as rain and snow may cause road closures, forcing drivers to slow down, which in turn causes traffic congestion on major roads and affects the smoothness and efficiency of the overall transportation network. Although current urban weather forecasting technology can predict the overall weather conditions in the next few days, it is still difficult to accurately predict the real-time conditions of urban roads, which poses a challenge to meeting the real-time weather forecasting needs of smart traffic management.

[0003] Affected by factors such as the cost and difficulty of building meteorological stations, the coverage of roadside traffic meteorological stations deployed on urban roads is limited. They can only cover some urban main roads and cannot completely cover the urban road network. Summary of the invention

[0004] Purpose of the invention: To solve the problem of forecasting short- and medium-term meteorological conditions for urban road traffic, the present invention proposes a method for forecasting urban road meteorological conditions that integrates vehicle-mounted and fixed meteorological stations. The method dynamically networks vehicle-mounted mobile meteorological stations and roadside fixed meteorological stations through vehicle-road collaborative technology to form a real-time meteorological data network. Each station is equipped with a meteorological condition forecasting agent, which uses deep neural networks for continuous self-learning and training to achieve iterative forecasting of road meteorological conditions. The forecasting agent can make rolling forecasts of meteorological conditions for the next 1-2 hours at 5-minute intervals, which can effectively meet the needs of forecasting short- and medium-term meteorological conditions for urban road networks.

[0005] The method of the present invention specifically comprises the following steps:

[0006] Step 1: Use the Internet of Vehicles technology to establish a data sharing network between mobile vehicle weather stations and fixed weather stations, and collect meteorological data through the data sharing network;

[0007] Step 2: Establish a meteorological condition prediction module based on meteorological stations, perform data preprocessing and analysis, and complete the prediction of urban road meteorological conditions based on long short-term memory networks and graph convolutional neural networks.

[0008] In step 1, the data sharing network includes a data input layer and a data communication layer based on the Internet of Vehicles;

[0009] The data input layer includes a meteorological data acquisition unit, which is used to dynamically network on-board mobile meteorological stations and roadside fixed meteorological stations to form a real-time meteorological data network. The meteorological data acquisition unit includes an on-board mobile meteorological station installed on an intelligent connected vehicle and a fixed meteorological station on the roadside.

[0010] In step 1, the data communication layer (Dedicated Short Range Communication, DSRC) based on the Internet of Vehicles adopts a dedicated short-range communication network or a 5G mobile network to share meteorological data between different meteorological stations within a certain range d, and establish a real-time meteorological data network G = (V, E), where the node set V in the network represents each meteorological station, and the edge set E represents the communication relationship between each meteorological station;

[0011] The data communication layer is for each weather station node v i ∈V establishes the feature vector of real-time meteorological conditions where v i represents the i-th weather station node, timestamp t i is the data collection time of the ith weather station node; the weather station location s i represents the longitude and latitude location information of the data collected by the i-th weather station node; the meteorological condition data vector d i =(T i ,H i ,P i ,WS i ,F i ), including temperature T i 、Humidity i , rainfall P i , wind speed WS i 、Icy road conditions F i ;c i is the data source classification, used for the ith weather station, whether it is a fixed weather station or a mobile weather station.

[0012] In step 2, the meteorological condition prediction module includes a data preprocessing module and a neural network meteorological prediction module; the data preprocessing module is responsible for preprocessing the collected data, and the neural network meteorological prediction module uses the data processed by the data preprocessing module as input to predict meteorological conditions.

[0013] In step 2, the data preprocessing module records the meteorological station v i In the past M The historical meteorological data at the time is defined as the data vector in Indicates the weather station v i Meteorological data at time t;

[0014] The data preprocessing module processes the data vector x i Perform denoising and standardization processing, including:

[0015] Step 2-a1, data denoising: Calculate the data vector x i The mean of each component in mark The data component is abnormal data, σ is the standard deviation; for abnormal data The moving average method is used for correction, and the first k data are set as but The correction formula is:

[0016]

[0017] Among them, α is the dynamic weight coefficient, defined as

[0018] Step 2-a2, data standardization: Use the following formula to standardize the data:

[0019]

[0020] in, is the standardized data, min(x i ) and max(x i ) are the data vectors x i The minimum and maximum classification values ​​of .

[0021] In step 2, a neural network meteorological prediction module is set up at each meteorological station node, and the neural network meteorological prediction module first converts the standardized data Through the graph convolutional neural network, the meteorological condition correlation vector o of the road network layer is obtained i , and then associate the vector o i As input to the long short-term memory network to obtain weather forecast values.

[0022] In step 2, to obtain the meteorological condition correlation vector o of the road network layer i ,First, a multi-layer graph convolutional neural network is used to establish a meteorological condition information association model between different weather stations, including the following steps:

[0023] Step 2-b1: standardize the data Input graph convolutional neural network;

[0024] Step 2-b2, let the degree matrix of the real-time meteorological data network G be D, the diagonal matrix be I, and the adjacency matrix be A, then the hidden layer h of the graph convolutional neural network i for:

[0025]

[0026] Output the meteorological condition correlation vector o of the road network layer i :

[0027]

[0028] Among them, σ i is the ReLu activation function, W x is the input layer weight parameter vector, W h is the hidden layer weight parameter vector.

[0029] In step 2, the long short-term memory network performs the following steps:

[0030] Step 2-c1: associate the meteorological conditions of the road network layer with the vector o i Enter the Long Short-Term Memory Network;

[0031] Step 2-c2, calculate the predicted value y of the meteorological conditions i ,t:

[0032] y i，t =σ′ i (W L (h′ i,t-1 , o i )+b L )

[0033] Among them, σ′ i is the tanh activation function, W L and b L are weight parameter vector and bias parameter vector respectively, h′ i，t-1 is the hidden layer output value of the previous moment; the update formula of the hidden layer is:

[0034] h′ i，t =y i,t tanh(c i,t )

[0035] where h′ i，t is the hidden layer output value at the current moment, c i，t is the memory cell state parameter, which is updated through gate control logic and is expressed as:

[0036]

[0037] Among them, f t For the forget gate, i t is the input gate, c i,t-1 is the state of historical memory cells, is the candidate memory cell state, ⊙ is the element-by-element multiplication method, and the forget gate parameters, input gate parameters and candidate value parameters are updated through the back propagation algorithm to optimize the memory cell state.

[0038] In step 2, the long short-term memory network is trained through an iterative method to achieve rolling real-time optimal weather condition forecasting, which includes:

[0039] Step 2-d1, change the past t M The historical meteorological data with a data capacity of N per hour is used as the training set to calculate the predicted value y of the meteorological conditions. i，t , and the true value y of the training set i，labei For comparison, calculate the loss function:

[0040]

[0041] Step 2-d2, through the back propagation algorithm, updates all weight parameters of the neural network model to minimize the loss function value, including the following steps:

[0042] Step 2-d2-1, update the parameters of the long-short term memory (LSTM) layer, including the forget gate f t , input gate i t and candidate memory cell states Parameters for the forget gate f t , first calculate the error gradient df of the forget gate t :

[0043]

[0044] where dc i,t Indicates the memory cell state c i,t The error gradient, dh i,t represents the error gradient of the hidden layer;

[0045] Step 2-d2-2, based on the error gradient of the forget gate, calculate the gradient dW of the weight parameter of the forget gate respectively f and the gradient db of the bias parameter f , the formula is:

[0046] dW f =df t ⊙(h′ i，t-1 , o i )

[0047]

[0048] Step 2-d2-3, use gradient descent method to update the forget gate f t The weight W at time tf，t and bias b f，t , where η is the learning rate, the formula is:

[0049] W f，t =W f，t-1 -ηdW f

[0050] b f，t =b f，t-1 -ηdb f

[0051] Step 2-d2-4, update the input gate i in sequence t , candidate memory cell states Memory cell state parameter c i，t , the hidden layer output value h′ at the current time t i，t ;

[0052] Step 2-d2-5, based on the loss function L and the new parameters of the LSTM layer, the weight parameter W of the Graph Convolution Network (GCN) layer is updated using the standard back-propagation algorithm. x and W h ;

[0053] Step 2-d2-6, based on the new parameters of the LSTM layer and the GCN layer, recalculate steps 2-c1 to 2-c2 to update the predicted value y i,t , and recalculate the loss function; when the loss function value is less than the set threshold ζ1, the training process is completed;

[0054] Step 2-d2-7, use the trained long short-term memory network to predict the current weather conditions, and record the real-time detection data; when the deviation between the predicted data and the real-time detection data exceeds the set threshold ζ2, or the interval from the last training time is greater than t M Hours, repeat steps 2-d2-1 to 2-d2-7 to update the neural network training.

[0055] The present invention also provides an electronic device, comprising a processor and a memory, wherein the memory stores program code, and when the program code is executed by the processor, the processor executes the steps of the described method.

[0056] The present invention integrates the data collected by the vehicle-mounted mobile weather station and the fixed weather station equipment to predict the meteorological conditions of the urban road network. First, the vehicle networking technology is used to establish a data sharing network of mobile vehicle weather stations and fixed weather stations. On this basis, based on the Long Short Term Memory (LSTM) and Graph Convolutional Network (GCN) models, a prediction method for urban road meteorological conditions is established. The LSTM model can effectively capture the time series information in the historical meteorological data, which is used to realize the medium- and short-term prediction of road meteorological conditions; while the GCN model can capture the spatial correlation characteristics between meteorological sites, thereby improving the accuracy of the prediction.

[0057] The method for predicting urban road meteorological conditions by integrating the data of a vehicle-mounted mobile meteorological station and a fixed meteorological station proposed in the present invention is mainly used for accurately predicting urban road meteorological conditions in the medium and short term. The method can make up for the problem of insufficient accuracy of existing urban meteorological forecasts and meet the needs of urban traffic management departments to grasp future traffic meteorological conditions, make decisions and manage in real time. When an emergency occurs, the method can quickly provide accurate forecasts of road meteorological conditions, help management departments formulate and implement efficient response strategies, thereby improving traffic safety and operational efficiency. In addition, urban traffic management departments can also issue traffic control plans in advance based on meteorological forecasts, reduce accident risks, and promote the sustainable development of urban traffic safety management.

[0058] The present invention has the following beneficial effects:

[0059] (1) The present invention proposes a method for forecasting meteorological conditions that integrates mobile vehicle-mounted meteorological stations and fixed meteorological stations. The method aims to solve the problem of low forecasting accuracy due to the insufficient number of traditional fixed meteorological stations, and can cover a wider range of urban road networks. By combining the deployment schemes of mobile vehicle-mounted meteorological stations and fixed meteorological stations, the method is highly adaptable and scalable, and can flexibly respond to different road traffic network requirements. In addition, the deployment can be adjusted according to the actual application situation to meet the needs of various road meteorological condition forecasting tasks.

[0060] (2) The urban road meteorological condition prediction method based on LSTM and GCN neural network proposed in the present invention can provide predictions with accuracy up to the minute level, covering rich professional information such as temperature, humidity, wind direction, rainfall, road icing conditions, etc. at important node locations of urban roads. Compared with the existing urban weather forecasts, the existing forecast content is relatively limited, including only basic information such as temperature and weather conditions, and can only provide single-day level forecasts. This method can not only meet the professional needs of road meteorological conditions, but also provide timely and accurate meteorological forecast data for urban traffic management departments at critical moments, which helps to optimize traffic management decisions and improve the safety and efficiency of urban transportation systems.

[0061] (3) The present invention utilizes a deep neural network model to ensure the real-time nature of the prediction and dynamically adjusts the model according to changes in meteorological conditions, thereby achieving high-precision predictions. By adopting the iterative update method of "training-prediction-actual situation comparison-optimization", the model parameters are dynamically adjusted and optimized according to the actual meteorological conditions to ensure the accuracy and timeliness of the prediction results. At the same time, the latest meteorological data is continuously incorporated into the database to ensure the continuity of model updates. In addition, the prediction results can be displayed through visualization tools such as road meteorological heat maps, so that traffic management departments and the public can intuitively understand the prediction situation. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.

[0063] Figure 1 It is a schematic diagram of the network structure for collecting weather condition information data and usage scenarios of the present invention.

[0064] Figure 2 It is a schematic diagram of the data flow and neural network model structure of the present invention.

[0065] Figure 3 It is a schematic diagram of the update timing of the prediction method of the present invention.

[0066] Figure 4 The present invention is a flowchart of a neural network model training for the meteorological condition prediction method. DETAILED DESCRIPTION

[0067] This embodiment provides a method for predicting urban road weather conditions by integrating vehicle-mounted and fixed weather stations, which includes a weather condition information data collection network composed of vehicle-mounted mobile weather stations and roadside fixed weather stations. The vehicle-mounted mobile weather station can move on urban roads and monitor road weather conditions in real time, especially providing additional data support in areas where fixed weather stations are sparsely distributed, thereby providing more detailed road weather information. In the embodiment of the present invention, the vehicle-mounted mobile weather station is defaulted to stop at a specific road area according to the needs of the meteorological management department, forming a temporary fixed weather station node. Any weather station v in the collection network i Collect for a period of time t M The meteorological data in the i , and then obtain the standardized data through the data preprocessing module In this embodiment, H is taken as 1 hour.

[0068] Fixed and mobile weather stations on the road are freely networked according to their distances. The weather station takes itself as the origin and establishes a regional communication network with a radius of 500 meters to share weather information data. Figure 1 As shown, the mobile weather station v move The topological structure of the regional communication network composed of the core is represented by G move =(V4,E3), that is, the network contains four weather station nodes and three communication links. Adjacent weather station nodes send their real-time weather data to the mobile weather station node through dedicated short-range communication or 5G mobile network, forming the regional road weather condition information at the current moment.

[0069] like Figure 2 As shown, the method of this embodiment first passes Figure 1 The fixed weather station and the vehicle-mounted mobile weather station collect real-time weather data information X, and then preprocess and normalize the data. The steps are as follows:

[0070] (1) Record each meteorological station v i In the past M = 1 hour of historical meteorological data, and define it as a data vector

[0071] (2) For the data vector x i Perform denoising and standardization. First calculate x i The mean of each component in mark The data component is abnormal data, σ is the standard deviation threshold, and the value in this example is 0.1;

[0072] (3) For abnormal data The moving average method is used for correction. In this example, the first 10 data of the abnormal data are set as but The correction formula is:

[0073]

[0074] Among them, α is the dynamic weight coefficient, defined as

[0075] (4) The data is standardized using the following formula:

[0076]

[0077] in, is the standardized data, min(x i ) and max(x i ) are the data vectors x i The minimum and maximum classification values ​​of .

[0078] After data preprocessing and normalization, GCN is used to extract spatial feature information between meteorological stations. Each mobile meteorological station first calculates G move The degree matrix D and adjacency matrix A of the network are then used to calculate the information correlation vector o of the network. move In this embodiment, the degree matrix Adjacency Matrix The information matrix X, D, A, the identity matrix I, and the weight parameter matrix W of GCN x and W h Substituting into the calculation formula of meteorological information correlation vector, we can calculate v move Relevance information:

[0079]

[0080] Relevance information i The current meteorological data of this meteorological station and surrounding meteorological stations are integrated to provide high-precision data support for LSTM to accurately predict the future meteorological conditions in the area.

[0081] LSTM neural network for mobile weather stations i To input information, continuously forecast the weather conditions data for the next hour at five-minute intervals. The steps are:

[0082] (4-1) Record the hidden layer state h of the LSTM neural network at the current moment i i ,t , memory cell state c i,t , forget gate f tand input gate i t ;

[0083] (4-2) Update memory cell state c i,t , the formula is:

[0084]

[0085] where c i,t-1 is the cell state at the previous moment, Based on the current input o i And the candidate memory state of the hidden layer state at the previous moment. Calculate c i,t The steps include:

[0086] (4-2-1) Calculate the forget gate vector f t =σ i (W f [h′ i,t-1 ,o i ]+b f ). Among them, W f and b f are the weight parameters and bias values ​​of the forget gate, σ i is the Relu activation function;

[0087] (4-2-2) Calculate the input gate vector i t =σ i (W i [h′ i,t-1 ,o i ]+b i ), where W i and b i They are the weight parameters and bias values ​​of the input gate respectively;

[0088] (4-2-3) Calculate the candidate value vector Where W c and b c are the weight parameters and bias values ​​of the alternative values ​​respectively;

[0089] (4-2-4) Using element-by-element multiplication, f t ,i t , and c i,t-1 Multiply the elements in the vector one by one to get c i,t .

[0090] (4-3) Update the hidden layer state h′ i，t , the formula is:

[0091] h′ i，t =y i,t tanh(c i，t )

[0092] (4-4) Calculate the predicted value y of meteorological conditions i，t , the formula is:

[0093] y i，t =σ′ i (W L (h′ i,t-1 , o i )+b L )

[0094] (4-5) Forecast the weather conditions in the next hour at 5-minute intervals. The method is to convert the predicted value y i，t As the input value for the next moment, repeat steps (4-1) to (4-4) until y is calculated. i,t+1 , ..., y i,t+N , until t+N≥60.

[0095] like Figure 3 As shown, the neural network model parameters are updated based on the latest collected meteorological information data, and the temperature forecast value shown in the figure is obtained based on the neural network model. At the same time, the error between the real values ​​collected by the meteorological station within 1 hour is checked. If the mean error is greater than the threshold ζ2, enter Figure 4 In the embodiment, ζ2 takes the value of 0.1, which means that when the average temperature difference between the predicted temperature and the actual temperature exceeds 10% of the actual temperature value, a new round of training will be carried out to reduce the prediction error.

[0096] like Figure 4 As shown in the figure, the training process of the neural network module of the mobile weather station is:

[0097] (5-1) Determine the training conditions. The training process is triggered when any of the following conditions is met: (1) There is a change in the network topology due to the movement of the mobile weather station in the regional network; (2) The time interval from the last training exceeds the preset interval value t K In this embodiment, t K Take 1 hour; (3) a period of time t N The mean error between the predicted value and the measured value is greater than the threshold ζ2, and the calculation formula of ζ2 is:

[0098]

[0099] Where ||·|| is the modulus of the vector. In this embodiment, t N The value is 0.5 hours.

[0100] (5-2) The past t M The historical data set with a data capacity of N per hour is used as the training set to calculate the predicted value y of the meteorological conditions.i，t , and the true value y of the training set i,label For comparison, calculate the loss function:

[0101]

[0102] (5-3) Through the back propagation method of weight gradients in the neural network model, all weight parameters in the GCN and LSTM neural network models are updated to minimize the loss function value. The main steps include:

[0103] (5-3-1) Update the forget gate f in the LSTM layer t , input gate i t and alternative values The parameters of the vector. Because the three structures are similar, taking the update of the forget gate matrix as an example, the update formula is:

[0104] (5-3-2) Calculate the error gradient of the forget gate:

[0105]

[0106] (5-3-3) Calculate the weight parameter gradient of the forget gate:

[0107] dW f =df t ⊙(h′ i,t-1 , o i )

[0108] (5-3-4) Calculate the bias parameter gradient of the forget gate:

[0109]

[0110] (5-3-5) Use gradient descent method to update the forget gate moment f t The weight W at time t f，t and bias b f，t , where η is the learning rate:

[0111] W f，t =W f，t-1 -ηdW f

[0112] b f，t =b f，t-1 -ηdb f

[0113] (5-3-6) Based on the loss function L and the updated parameters of the LSTM layer, the weight parameter W of the GCN layer is updated using the standard back propagation algorithm x and W h ; In this embodiment, the weight W of LSTM f , Wi and W c , and the weight W of the GCN layer x and W h The number of parameters is set to 256, and the bias parameter b of LSTM is f , b i , b c The number of is set to 1.

[0114] (5-4) When the loss function value is less than the set threshold ζ1=0.05, the training process is completed.

[0115] (5-5) Use the trained neural network to predict the current weather conditions and record real-time weather data. When the deviation between the predicted data and the real-time data exceeds the set threshold ζ2, or the interval from the last training time t K If the training time is greater than 1 hour, repeat the iterative training.

[0116] Through the above detailed training and deployment process, the method proposed in this embodiment can accurately predict meteorological conditions such as temperature, humidity, wind speed and other meteorological conditions of urban roads based on historical data and real-time meteorological data, provide decision support for meteorological departments and traffic management departments, and improve the safety and efficiency of road traffic.

[0117] The above is an example of data collection and prediction methods for mobile weather stations. For fixed weather stations, the process is exactly the same.

[0118] Result output and application: The forecast results are displayed in a visual form to the meteorological management department, traffic management department and the public, providing them with decision support and travel reference. The forecast results are displayed through visualization tools, such as heat maps of road network temperature and wind speed, so that traffic management departments and the public can understand them intuitively. Decision support is provided, such as traffic control suggestions and recommendations for the best travel time.

[0119] The urban road traffic meteorological condition prediction method of the present invention realizes short-term accurate prediction of urban road meteorological conditions by comprehensively considering meteorological factors and urban road data. The method has the characteristics of real-time, accuracy and flexibility, which helps to improve the level of urban meteorological management, and can serve urban traffic management departments to reduce the risk of traffic accidents under severe meteorological conditions.

[0120] The present invention provides a method for predicting urban road weather conditions by integrating vehicle-mounted and fixed weather stations. There are many methods and ways to implement the technical solution. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims

1. A method for predicting urban road weather conditions by integrating vehicle-mounted and fixed weather stations, characterized in that: The following steps are involved: Step 1: Use the Internet of Vehicles technology to establish a data sharing network between mobile vehicle weather stations and fixed weather stations, and collect meteorological data through the data sharing network; Step 2: Establish a meteorological condition prediction module based on meteorological stations, perform data preprocessing and analysis, and complete the prediction of urban road meteorological conditions based on long short-term memory networks and graph convolutional neural networks.

2. The method according to claim 1, characterized in that: In step 1, the data sharing network includes a data input layer and a data communication layer based on the Internet of Vehicles; The data input layer includes a meteorological data acquisition unit, which is used to dynamically network on-board mobile meteorological stations and roadside fixed meteorological stations to form a real-time meteorological data network. The meteorological data acquisition unit includes an on-board mobile meteorological station installed on an intelligent connected vehicle and a fixed meteorological station on the roadside.

3. The method according to claim 2, characterized in that In step 1, the data communication layer based on the Internet of Vehicles adopts a dedicated short-range communication network or a 5G mobile network to share meteorological data between different meteorological stations within a range d, and establish a real-time meteorological data network G = (V, E), where the node set V in the network represents each meteorological station, and the edge set E represents the communication relationship between each meteorological station; The data communication layer is for each weather station node v i ∈y establishes the feature vector of real-time meteorological conditions where v i represents the i-th weather station node, timestamp t i is the data collection time of the ith weather station node; the weather station location s i Represents the longitude and latitude location information of data collection of the i-th weather station node; Meteorological condition data vector d i =(T i , H i , P i , WS i , F i ), including temperature T i 、Humidity i , rainfall P i , wind speed WS i 、Icy road conditions F i ;c i is the data source classification, used for the ith weather station, whether it is a fixed weather station or a mobile weather station.

4. The method according to claim 3, characterized in that In step 2, the meteorological condition prediction module includes a data preprocessing module and a neural network meteorological prediction module; the data preprocessing module is responsible for preprocessing the collected data, and the neural network meteorological prediction module uses the data processed by the data preprocessing module as input to predict meteorological conditions.

5. The method according to claim 4, characterized in that In step 2, the data preprocessing module records the meteorological station v i In the past M The historical meteorological data at the time is defined as the data vector in Indicates the weather station v i Meteorological data at time t; The data preprocessing module processes the data vector x i Perform denoising and standardization processing, including: Step 2-a1, data denoising: Calculate the data vector x i The mean of each component in mark The data component is abnormal data, σ is the standard deviation; for abnormal data The moving average method is used for correction, and the first k data are set as but The correction formula is: Among them, α is the dynamic weight coefficient, defined as Step 2-a2, data standardization: Use the following formula to standardize the data: in, is the standardized data, min(x i ) and max(x i ) are the data vectors x i The minimum and maximum classification values ​​of .

6. The method according to claim 5, characterized in that In step 2, a neural network meteorological prediction module is set up at each meteorological station node, and the neural network meteorological prediction module first converts the standardized data Through the graph convolutional neural network, the meteorological condition correlation vector o of the road network layer is obtained i , and then associate the vector o i As input to the long short-term memory network to obtain weather forecast values.

7. The method according to claim 6, characterized in that In step 2, to obtain the meteorological condition correlation vector o of the road network layer i ,First, a multi-layer graph convolutional neural network is used to establish a meteorological condition information association model between different weather stations, including the following steps: Step 2-b1: standardize the data Input graph convolutional neural network; Step 2-b2, let the degree matrix of the real-time meteorological data network G be D, the diagonal matrix be I, and the adjacency matrix be A, then the hidden layer h of the graph convolutional neural network i for: Output the meteorological condition correlation vector o of the road network layer i : Among them, σ i is the ReLu activation function, W x is the input layer weight parameter vector, W h is the hidden layer weight parameter vector.

8. The method according to claim 7, characterized in that In step 2, the long short-term memory network performs the following steps: Step 2-c1: associate the meteorological conditions of the road network layer with the vector o i Enter the Long Short-Term Memory Network; Step 2-c2, calculate the predicted value y of the meteorological conditions i,t : y i，t =σ′ i (W L (h′ i,t-1 ,the i )+b L ) Among them, σ′ i is the tanh activation function, W L and b L are weight parameter vector and bias parameter vector respectively, h′ i，t-1 is the hidden layer output value of the previous moment; the update formula of the hidden layer is: h' i，t =y i,t fishy i,t ) where h′ i，t is the hidden layer output value at the current moment, c i，t is the memory cell state parameter, which is updated through gate control logic and is expressed as: Among them, f t For the forget gate, i t is the input gate, c i,t-1 is the state of historical memory cells, is the candidate memory cell state, ⊙ is the element-by-element multiplication method, and the forget gate parameters, input gate parameters and candidate value parameters are updated through the back propagation algorithm to optimize the memory cell state.

9. The method according to claim 8, characterized in that In step 2, the long short-term memory network is trained through an iterative method to achieve rolling real-time optimal weather condition forecasting, which includes: Step 2-d1, change the past t M The data capacity in the hour is NN's historical meteorological data as the training set, and the predicted value y of the meteorological conditions is calculated. i，t , and the true value y of the training set i，label For comparison, calculate the loss function: Step 2-d2, through the back propagation algorithm, updates all weight parameters of the neural network model to minimize the loss function value, including the following steps: Step 2-d2-1, update the parameters of the LSTM layer of the long short-term memory network, including the forget gate f t , input gate i t and candidate memory cell states Parameters for the forget gate f t , first calculate the error gradient df of the forget gate t : where dc i,t Indicates the memory cell state c i,t The error gradient, dh i,t represents the error gradient of the hidden layer; Step 2-d2-2, based on the error gradient of the forget gate, calculate the gradient dW of the weight parameter of the forget gate respectively f and the gradient db of the bias parameter f , the formula is: dW f =df t ⊙(h′ i,t-1 ,he i ) Step 2-d2-3, use gradient descent method to update the forget gate f t The weight W at time t f，t and bias b f，t , where η is the learning rate, the formula is: IN f，t =In f，t-1 -ηdW f b f，t =b f，t-1 -ηdb f Step 2-d2-4, update the input gate i in sequence t , candidate memory cell states Memory cell state parameter c i，t , the hidden layer output value h′ at the current time t i，t ; Step 2-d2-5, based on the loss function L and the new parameters of the LSTM layer, the weight parameter W of the graph convolutional neural network GCN layer is updated using the standard back propagation algorithm x and W h ; Step 2-d2-6, based on the new parameters of the LSTM layer and the GCN layer, recalculate steps 2-c1 to 2-c2 to update the predicted value y i，t , and recalculate the loss function; when the loss function value is less than the set threshold ζ1, the training process is completed; Step 2-d2-7, use the trained long short-term memory network to predict the current weather conditions, and record the real-time detection data; when the deviation between the predicted data and the real-time detection data exceeds the set threshold ζ2, or the interval from the last training time is greater than t M Hours, repeat steps 2-d2-1 to 2-d2-7 to update the neural network training.

10. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores program codes, and when the program codes are executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 9.

Images