Meteorological risk intelligent early warning method and system based on multi-source data analysis

Abstract

The invention relates to the technical field of risk early warning, and discloses a meteorological risk intelligent early warning method and system based on multi-source data analysis, and the method comprises the following steps: obtaining multi-source meteorological data, historical meteorological risk event data and environment change related data; carrying out normalization, feature extraction and coding processing on the data; fusing the processed multi-source data by adopting a deep neural network; constructing a deep reinforcement learning model; network parameters of the deep reinforcement learning model are updated through minimizing the loss function, and an optimized early warning strategy is obtained; according to the method, feature extraction and fusion are carried out by adopting a deep learning method after multi-source heterogeneous data are fused, meteorological risk influence factors can be comprehensively analyzed, the early warning accuracy is remarkably improved, and an early warning strategy can be dynamically adjusted based on a deep reinforcement learning model according to real-time data and historical experience, so that early warning better meets actual requirements.

Description

Technical Field

[0001] The present invention relates to the technical field of risk warning, and more specifically, to a meteorological risk intelligent warning method and system based on multi-source data analysis. Background Art

[0002] As global climate change intensifies, the impact of frequent meteorological disasters on human society is becoming increasingly significant. Traditional meteorological early warning systems mainly rely on a single data source (such as meteorological station observation data) and judgment methods based on empirical thresholds, which have problems such as low early warning accuracy and poor timeliness; The existing early warning system often takes a long time to process data and make decisions. In particular, when faced with sudden extreme weather events, it is unable to issue accurate warnings in a timely manner, which affects the effectiveness of disaster prevention and mitigation. At the same time, due to the failure to fully consider the influence of factors such as geographical environment characteristics and historical risk events, the existing early warning system has a high false alarm rate and missed alarm rate, which affects the credibility and practicality of the warning. Summary of the invention

[0003] The purpose of the present invention is to provide a meteorological risk intelligent early warning method and system based on multi-source data analysis in order to solve the above problems.

[0004] The present invention provides a meteorological risk intelligent early warning method based on multi-source data analysis, comprising the following steps: Step 100, obtaining multi-source meteorological data, historical meteorological risk event data and environmental change related data, and normalizing, feature extracting and encoding the data; Step 200, using a deep neural network to fuse the processed multi-source data and output a fused feature vector; Step 300, constructing a deep reinforcement learning model, taking the fused feature vector as the state space, taking the early warning strategy as the action space, and performing model training by setting an exploration strategy and a priority experience playback mechanism; Step 400, based on the data samples in the experience replay pool, the network parameters of the deep reinforcement learning model are updated by minimizing the loss function to obtain an optimized early warning strategy.

[0005] Furthermore, the multi-source meteorological data includes at least one of temperature data, humidity data, wind speed data, and air pressure data.

[0006] Furthermore, the reward function in the deep reinforcement learning model The calculation method is: ; in, , , is the adjustment coefficient, which is used to adjust the relative importance of different reward items in the total reward. is the improvement of early warning accuracy, which is calculated as follows: ; in, is the number of true positive warnings, that is, the number of correct warnings of meteorological risk events; is the number of true negative warnings, that is, the number of times that the meteorological risk event was correctly warned but did not occur; is the number of false alarms, that is, the number of times that meteorological risk events are mistakenly warned; is the number of missed reports, that is, the number of actual meteorological risk events that were not warned; is the benchmark accuracy, this formula represents the current warning strategy The difference between the accuracy under the prediction and the benchmark accuracy is used to measure the improvement effect of the early warning strategy on the accuracy; is the change in false alarm rate, which is calculated as follows: ; in, is the baseline false alarm rate; is the warning timeliness gain, which is calculated as follows: ; in The warning release time is is the actual time when the meteorological event occurred, is the time-effect attenuation coefficient; when three consecutive false alarms occur, the penalty multiplication factor is triggered , the false positive item is corrected to: ; in is the penalty coefficient, that is, when there are three consecutive false alarms, the change in the false alarm rate will be doubled to strengthen the control of false alarms.

[0007] Furthermore, the method further comprises the following steps: Step 501: Obtain forecast data from a weather model ,in Represents different forecast time periods and variables. Perform feature extraction and normalization; Step 502: Based on step 200, the meteorological model prediction data The feature vectors of are input into the deep neural network for fusion, and the fused feature vector is output; Step 503, input the fused feature vector obtained in step 502, and output the trend prediction result of the meteorological risk in the future period of time.

[0008] Furthermore, the method further comprises the following steps: Step 601, obtaining multi-source meteorological data, historical meteorological risk event data, real-time environmental change related data and inter-regional geographical relationship data of multiple adjacent regions; performing the same processing as step 100 on these data of each region; Step 602 constructs a multi-layer fusion network. First, the normalized meteorological data feature vector, the historical meteorological risk event feature vector, and the environmental change feature representation in each region are fused to obtain a regional fusion feature vector. Then, the regional fusion feature vectors of all regions and the geographic relationship data between regions are input into a higher-layer fusion network to output a multi-region collaborative fusion feature vector. Step 603, building a deep reinforcement learning model; the same as step 300, but the state space is updated to a multi-region collaborative fusion feature vector; Step 604: Model training and updating; same as step 400.

[0009] Furthermore, the method further comprises the following steps: Step 701, quickly screen out key subset data that can be quickly obtained at the early stage of an extreme event from multi-source meteorological data, historical meteorological risk event data, and real-time environmental change related data, and perform normalization processing on these key subset data in the same manner as step 100; Step 702, designing a lightweight neural network, inputting feature vectors corresponding to the preprocessed key subset data, and outputting fast fusion feature vectors; Step 703, using a fast decision model based on Bayesian reasoning, inputting the fast fusion feature vector obtained in step 702, and outputting a warning decision for extreme weather events; Step 704, using historical data marked with extreme weather events to train the Bayesian model, optimize the estimation of prior probability and likelihood probability, so as to improve the accuracy of the model in warning extreme weather events under limited data.

[0010] Furthermore, the method further comprises the following steps: Step 801, in addition to meteorological data, also obtains wave height data, wave period data, current speed data, current direction data, ocean-related data, and historical risk event data of marine navigation and operation activities ; Normalize the ocean-related data, and similarly normalize the wave period, current speed, and current direction data; Extract key features, construct feature vectors, and perform the same processing as step 100 on meteorological data and other data; Step 802, using a deep neural network to fuse multi-source data, inputting normalized meteorological data feature vectors, ocean-related data feature vectors, and maritime historical risk event feature vectors, and outputting a fused feature vector; Step 803 is the same as step 300, but the state space Updated to the fused feature vector, action space Adjust the early warning strategy to suit maritime activities; Step 804, model training and updating, is the same as step 400.

[0011] Furthermore, the method further comprises the following steps: Step 901, in addition to the data in step 100, high-precision terrain data of mountainous areas is also obtained , historical meteorological risk event data in mountainous areas ;right Perform feature extraction and encoding to obtain the corresponding feature representation. Extract key features, construct feature vectors, and perform the same processing as step 100 on meteorological data and other data; Step 902, using a deep neural network to fuse multi-source data, inputting normalized meteorological data feature vectors, mountain terrain feature representations, and mountain historical risk event feature vectors, and outputting a fused feature vector; Step 903, building a deep reinforcement learning model, is the same as step 300, but the state space Updated to the fused feature vector, action space Adjust the early warning strategy to suit mountain areas; Step 904, model training and updating; same as step 400.

[0012] Furthermore, the method further comprises the following steps: Step 1001, in addition to the data in step 100, also obtain the distribution data of each functional area of ​​the city , historical meteorological risk event data for each functional area , and real-time dynamic data related to the functional area ,right Encode the corresponding feature representation. Extract key features, construct feature vectors, and Perform normalization processing, and perform the same processing as step 100 on the meteorological data and other data; Step 1002, using a deep neural network to fuse multi-source data, inputting normalized meteorological data feature vectors, functional area distribution feature representations, functional area historical risk event feature vectors, and functional area real-time dynamic data feature vectors, and outputting a fused feature vector; Step 1003, building a deep reinforcement learning model, is the same as step 300, but the state space Updated to the fused feature vector, action space Adjust to early warning strategies for different functional areas; Step 1004, model training and updating, is the same as step 400.

[0013] The present invention provides a meteorological risk intelligent early warning system based on multi-source data analysis, which is used to store computer-readable instructions. When the computer-readable instructions are read, the aforementioned meteorological risk intelligent early warning method based on multi-source data analysis can be executed.

[0014] The beneficial effects of the present invention are as follows: by fusing multi-source heterogeneous data and then adopting a deep learning method to extract and fuse features, the present invention can comprehensively analyze the factors affecting meteorological risks, significantly improve the accuracy of early warnings, and based on a deep reinforcement learning model, can dynamically adjust early warning strategies according to real-time data and historical experience, so that early warnings are more in line with actual needs.

[0015] By adopting a lightweight fast fusion network and a fast decision-making model based on Bayesian reasoning, the system can make early warning decisions quickly in the early stages of extreme weather events.

[0016] For different application scenarios (such as at sea, in mountains, in cities, etc.), the system integrates corresponding professional data, builds specific early warning models, and achieves more targeted early warnings. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of a meteorological risk intelligent early warning method based on multi-source data analysis of the present invention; Figure 2 is an example of raw data obtained by the present invention; Figure 3 is a feature vector sample after data preprocessing of the present invention; Figure 4 is an example of a fused feature vector of the present invention; Figure 5 is an example of the early warning result given by the model of the present invention; Figure 6 This is an example of actual situation feedback data of the present invention. DETAILED DESCRIPTION

[0018] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the contents of this specification. Each example may omit, replace or add various processes or components as needed. In addition, the features described relative to some examples may also be combined in other examples.

[0019] At least one embodiment of the present invention discloses a meteorological risk intelligent early warning method based on multi-source data analysis, such as Figure 1 As shown, the following steps are included: Step 100: Data preprocessing. Obtain multi-source meteorological data, including temperature data , humidity data , wind speed data , air pressure data etc., among which Represents different time points or monitoring locations. These data are normalized. For temperature data, the original temperature data is , normalized temperature data .here Indicates the minimum value of all temperature data. Represents the maximum value of all temperature data. This formula is used to map the temperature data to This normalization method can make data of different magnitudes on the same comparable scale, which is helpful for subsequent model training. Similarly, humidity, wind speed, air pressure and other data are normalized. Obtain historical meteorological risk event data ,in Represent different historical events, extract key features from these data, and construct feature vectors. Obtain data related to real-time environmental changes , such as terrain changes, impacts of human activities, etc. Represent different environmental factors and encode them to obtain corresponding feature representations.

[0020] Step 200: Multi-source data fusion. A deep neural network is used to fuse multi-source data. Suppose the fusion network is , whose structure is a multi-layer perceptron (MLP), and the input is the normalized meteorological data feature vector , historical meteorological risk event feature vectors and environmental change feature representation, output fusion feature vector ,here It represents the function operation of the multi-layer perceptron network to perform fusion calculation on the input data, that is, the input data is transformed and combined through the weights and activation functions inside the network to obtain the fused feature vector.

[0021] Step 300: Build a deep reinforcement learning model. Define the state space To fuse feature vectors, action space Different early warning strategies (such as adjustment of warning level, adjustment of warning release time, etc.). Specifically include: Network architecture design: Construct a dual-channel deep Q network (DQN), a reinforcement learning architecture consisting of a main network and a target network. In DQN, the main network parameters By gradient descent update, the target network parameters Periodic synchronization Value. Both the main network and the target network adopt multi-layer perceptron structure. The main network contains an input layer (the dimension is the same as the fusion feature vector), two hidden layers (128 neurons in each layer, ReLU activation) and an output layer (the dimension is the same as the size of the action space); the target network structure is mirror-symmetric to the main network, and the parameters are updated synchronously every 100 steps, which helps to improve the stability and convergence of training.

[0022] Explore policy settings: Adoption -greedy strategy, initial exploration rate , which decays linearly with the number of training steps to . In each decision, the probability Randomly select an action to explore, with probability Select the currently estimated optimal action to use. Set the priority experience replay mechanism based on the TD error Dynamically adjust the sample sampling probability, where is the target Q value, Is the current state Take action The TD error reflects the difference between the current estimated Q value and the target Q value. The priority experience replay mechanism enables the model to pay more attention to samples with larger TD errors, thereby improving learning efficiency.

[0023] Defining the reward function :

[0024] in, , , is the adjustment coefficient ( and ), which is used to adjust the relative importance of different reward items in the total reward, is the improvement of early warning accuracy, which is calculated as follows:

[0025] in, is the number of true positive warnings, that is, the number of correct warnings of meteorological risk events; is the number of true negative warnings, that is, the number of times that the meteorological risk event was correctly warned but did not occur; is the number of false alarms, that is, the number of times that meteorological risk events are mistakenly warned; is the number of missed reports, that is, the number of actual meteorological risk events that were not warned; is the benchmark accuracy, this formula represents the current warning strategy The difference between the accuracy under the prediction and the benchmark accuracy is used to measure the improvement effect of the early warning strategy on the accuracy; is the change in false alarm rate, which is calculated as follows:

[0026] in, is the baseline false alarm rate, and the formula represents the current warning strategy The difference between the false alarm rate under the threshold and the baseline false alarm rate is used to evaluate the impact of the early warning strategy on the false alarm rate. is the warning timeliness gain, which is calculated as follows:

[0027] in The warning release time is is the actual time when the meteorological event occurred, is the time-effect attenuation coefficient; when three consecutive false alarms occur, the penalty multiplication factor is triggered , the false positive item is corrected to:

[0028] in is the penalty coefficient, that is, when there are three consecutive false alarms, the change in the false alarm rate will be doubled to strengthen the control of false alarms.

[0029] Experience replay optimization: set up experience replay pool The capacity is 10,000. It uses a combination of random sampling and importance sampling, and performs batch sampling every 10 training steps (batch_size = 64). The experience replay pool is used to store experience samples generated by the interaction between the agent and the environment. Random sampling can increase the diversity of samples, and importance sampling can sample according to the importance of samples. The combination of the two helps to improve training efficiency. Sliding window processing is used for time series data, and the window length is time steps, that is, processing 10 consecutive time steps of data each time. This processing method can capture the short-term dependencies in time series data.

[0030] Network training configuration: Use Adam optimizer, set the initial learning rate to 0.001, set the gradient clipping threshold to 1.0, and use gradient clipping to prevent gradient explosion. When the norm of the gradient exceeds the threshold, the gradient is clipped. Introduce the dual Q learning mechanism, and use the target Q value calculation method. ,in represents the target network, is in state Take action After the reward, is a discount factor used to measure the importance of future rewards, Indicates in status The double Q-learning mechanism can reduce the deviation of Q-value estimation and improve the stability of the algorithm.

[0031] Step 400: Model training and updating. From the experience replay pool Randomly sample a batch of data , calculate the target Q value ,in is the discount factor. By minimizing the loss function To update the network parameters of DQN, here Represents each sample in a batch of sampled data A summation operation is performed to minimize the loss function so that the Q value output by the main network is as close to the target Q value as possible, thereby optimizing the network parameters and enabling the model to better learn the optimal early warning strategy.

[0032] Through the above steps, we will eventually obtain a meteorological risk warning system that can adaptively adjust warning strategies based on real-time multi-source meteorological data and environmental changes. This system can improve the accuracy and timeliness of warnings, and provide strong support for meteorological disaster prevention with more accurate warning levels and more appropriate warning release times.

[0033] In one embodiment of the present invention, in order to further improve the foresight and scientific nature of the early warning system, a meteorological risk trend prediction function is added. By introducing meteorological model prediction data and combining it with the LSTM network, the system can not only provide real-time warnings, but also predict meteorological risk trends in the future, thereby providing data support for long-term disaster prevention and mitigation planning. Specifically, the following steps are included: Step 501: Data preprocessing. In addition to multi-source meteorological data, historical meteorological risk event data, and real-time environmental change related data, the forecast data of the meteorological model is also obtained. ,in Represents different forecast time periods and variables. Perform feature extraction and normalization.

[0034] Step 502: Multi-source data fusion. Based on step 200, the meteorological model prediction data The feature vector is input into the deep neural network Fusion is performed and the fused feature vector is output ,In this way, data from different sources can be integrated to provide more comprehensive information for subsequent models.

[0035] Step 503, construct a prediction model. The long short-term memory network (LSTM) is used. LSTM is a special recurrent neural network (RNN) that can effectively handle the long-term dependency problem in time series data. It controls the flow of information through a gating mechanism (input gate, forget gate, and output gate). The input is the fused feature vector obtained in step 502, and the output is the trend prediction result of meteorological risks in the future. Suppose the LSTM network is , then the prediction result .

[0036] Step 504: Model training and optimization. The LSTM network is trained using historical meteorological risk data and corresponding future trend data, using a mean square error loss function. ,in is the predicted value, is the true value, is the number of samples, and the mean square error loss function is used to measure the average error between the predicted value and the true value; Indicates the number of samples from the first to the The network parameters are updated by optimization algorithms such as gradient descent, so that the LSTM network can learn the patterns in the data and accurately predict the future meteorological risk trends.

[0037] By introducing meteorological model prediction data and building an LSTM prediction model, we can effectively predict future meteorological risk trends, provide strong data support for long-term disaster prevention and mitigation planning, make planning more scientific and reasonable, prepare for response in advance, and reduce long-term losses caused by meteorological disasters.

[0038] In one embodiment of the present invention, considering that meteorological disasters are often cross-regional, the early warning of a single region may not be able to fully cope with complex meteorological risks, so a multi-region collaborative early warning method is proposed. By considering the geographical relationship between regions and the correlation of meteorological data, collaborative early warning of multiple adjacent regions is achieved, thereby improving the ability to prevent cross-regional meteorological disasters. Specifically, the following steps are included: Step 601: Data preprocessing. Obtain multi-source meteorological data for multiple adjacent areas. , , , , historical meteorological risk event data , real-time environmental change related data , and data on geographic relationships between regions ,in The same processing as step 100 is performed on these data of each region.

[0039] Step 602: Multi-region data fusion. Construct a multi-layer fusion network First, the normalized meteorological data feature vector in each region is , historical meteorological risk event feature vectors and environmental change feature representations are fused to obtain the fused feature vector in the region , The expression formula is: ; Then, the feature vectors of all regions and the geographic relationship data between regions are integrated into the The multi-region collaborative fusion feature vector is outputted together into a higher-level fusion network. , The expression formula is: .

[0040] Step 603: Build a deep reinforcement learning model. Same as step 300, but the state space Updated to multi-region collaborative fusion feature vector.

[0041] Step 604: Model training and updating. Same as step 400.

[0042] By integrating multi-regional data and inter-regional geographical relationships and building a reinforcement learning model based on this, multi-regional coordinated early warning can be achieved. When facing cross-regional meteorological disasters, each region can issue an early warning based on a collaboratively optimized early warning strategy, thereby improving the accuracy and effectiveness of the overall early warning and reducing the losses caused by cross-regional meteorological disasters.

[0043] In one embodiment of the present invention, a lightweight rapid warning method is proposed to address the need for rapid response when extreme weather events occur. By combining Bayesian reasoning, a relatively accurate warning decision can be made quickly even when data is limited. Specifically, the following steps are included: Step 701: Rapid data screening and preprocessing: Rapidly screen out key subset data that can be quickly obtained at the early stage of an extreme event from multi-source meteorological data, historical meteorological risk event data, and real-time environmental change related data, and perform normalization processing similar to step 100 on these key subset data.

[0044] Step 702: Build a lightweight fast fusion network. Design a lightweight neural network , the input is the feature vector corresponding to the preprocessed key subset data, and the output is the fast fusion feature vector Here is a lightweight neural network The input data is transformed and combined through the weights and activation functions inside the network to obtain a fused feature vector.

[0045] Step 703: Construct a fast decision model. Use a fast decision model based on Bayesian reasoning , the input is the fast fusion feature vector obtained in step 702, and the output is the warning decision for extreme weather events. Suppose the conditional probability formula of the Bayesian model is ,in Indicates the different states of extreme meteorological events (such as occurrence, non-occurrence, etc.). represents the evidence represented by the fast fusion feature vector. is the prior probability, which is an estimate of the probability of different states of extreme meteorological events occurring without additional information (i.e., based only on historical data); is the likelihood probability, which is the probability that an extreme meteorological event is in a certain state. When evidence emerges The probability of is the posterior probability, which is the probability given the known evidence. In the case of extreme weather events, The probability of estimating the prior probability through historical data and the likelihood , thereby calculating the posterior probability to make decisions.

[0046] Step 704: Model training and optimization: Use historical data marked with extreme weather events to train the Bayesian model, optimize the estimation of prior probability and likelihood probability, and improve the accuracy of the model in warning extreme weather events under limited data.

[0047] By quickly screening key data, building lightweight fusion networks and rapid decision-making models based on Bayesian reasoning, we can use limited data to quickly make accurate early warning decisions in the early stages of extreme weather events, thereby buying precious time to respond to extreme weather events and effectively reducing the losses caused by extreme weather events.

[0048] In one embodiment of the present invention, considering the particularity of coastal areas, the marine meteorological risk warning is optimized. By integrating ocean-related data, the warning system can better serve marine activities. Specifically, the following steps are included: Step 801: Data preprocessing. In addition to meteorological data, sea wave height data is also obtained. , Wave cycle data , ocean current speed data , ocean current direction data Other ocean-related data, as well as historical risk event data of maritime navigation and operation activities . Normalize the ocean-related data. For example, for the wave height data, the original wave height data is , normalized wave height data .in, represents the raw wave height data, Represents the minimum value of all wave height data. Represents the maximum value of all wave height data. This formula is used to map the wave height data to the corresponding interval. Similarly, the wave period, current speed, current direction and other data are normalized. Extract key features and construct feature vectors. At the same time, perform the same processing as step 100 on the meteorological data and other data.

[0049] Step 802: Multi-source data fusion. A deep neural network is used to fuse multi-source data. The fusion network is assumed to be , the input is the normalized meteorological data feature vector , ocean-related data feature vector , feature vector of maritime historical risk events, output fusion feature vector , The expression formula is: .

[0050] Step 803: Build a deep reinforcement learning model. Same as step 300, but the state space Updated to the fused feature vector, action space Adjust to early warning strategies suitable for maritime activities (such as ship route adjustment recommendations, maritime operation suspension / continuation decisions, etc.).

[0051] Step 804: Model training and updating. Same as step 400.

[0052] By considering ocean factors such as waves and currents and integrating relevant data, a deep reinforcement learning model suitable for maritime activities is constructed. This can provide accurate meteorological risk warnings for maritime navigation and operational activities, give more targeted warning strategies, ensure the safety of maritime activities, and reduce maritime accidents caused by meteorological risks.

[0053] In one embodiment of the present invention, considering the impact of mountain terrain on meteorological risk, the meteorological risk warning in mountainous areas is optimized, and the early warning system can better serve mountainous activities by integrating mountain terrain data and related historical risk event data. Specifically, the following steps are included: Step 901: Data preprocessing. In addition to the data from step 100, high-precision terrain data of mountainous areas is also obtained. (such as altitude, slope, aspect, etc.), historical meteorological risk event data in mountainous areas .right Perform feature extraction and encoding to obtain the corresponding feature representation. Extract key features and construct feature vectors. At the same time, perform the same processing as step 100 on the meteorological data and other data.

[0054] Step 902: Multi-source data fusion. A deep neural network is used to fuse multi-source data. The fusion network is assumed to be , the input is the normalized meteorological data feature vector , mountain terrain feature representation, mountain historical risk event feature vector, output fusion feature vector .

[0055] Step 903: Build a deep reinforcement learning model. Same as step 300, but the state space Updated to the fused feature vector, action space Adjust to early warning strategies suitable for mountainous areas (such as mountain road closure / traffic decisions, mountain resident evacuation recommendations, etc.).

[0056] Step 904: Model training and updating. Same as step 400.

[0057] By integrating mountain terrain data and related historical risk event data and building a reinforcement learning model based on this, we can fully consider the impact of mountain terrain on meteorological risks, provide accurate meteorological risk warnings and warning strategies suitable for the characteristics of mountainous areas, effectively improve the ability of mountainous areas to respond to meteorological disasters, and ensure the safety of mountain residents and infrastructure.

[0058] In one embodiment of the present invention, in order to meet the needs of refined urban management, an early warning method for different functional areas of a city is proposed; by considering the distribution and characteristics of urban functional areas, a more targeted early warning is achieved. Specifically, the following steps are included: Step 1001: Data preprocessing. In addition to the data in step 100, the distribution data of each functional area in the city is also obtained. (including information on functional area type, area, location, etc.), historical meteorological risk event data for each functional area , and real-time dynamic data related to the functional area (Such as the flow of people in commercial areas, the intensity of production activities in industrial areas, etc.). Encode the corresponding feature representation. Extract key features, construct feature vectors, and Normalization is performed. At the same time, the meteorological data and other data are processed in the same manner as step 100.

[0059] Step 1002: Multi-source data fusion. A deep neural network is used to fuse multi-source data. Suppose the fusion network is , the input is the normalized meteorological data feature vector , functional area distribution feature representation, functional area historical risk event feature vector, functional area real-time dynamic data feature vector, output fusion feature vector .

[0060] Step 1003, build a deep reinforcement learning model. Same as step 300, but the state space Updated to the fused feature vector, action space Adjust to early warning strategies for different functional areas (such as issuing shopping mall closure / business adjustment notices in commercial areas, issuing resident prevention recommendations in residential areas, and issuing production activity adjustment recommendations in industrial areas, etc.).

[0061] Step 1004: Model training and updating. Same as step 400.

[0062] By integrating relevant data of urban functional areas and building a reinforcement learning model based on this, we can carry out refined meteorological risk warnings based on the characteristics of different functional areas in the city, provide more targeted warning strategies, improve the overall efficiency of the city's response to meteorological disasters, and ensure the normal operation of various functional areas in the city and the safety of life and property of residents.

[0063] The following is an early warning of the weather risk in coastal cities in the next 24 hours. The system obtains various data once an hour and gives an early warning for the next hour.

[0064] from Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6It is clear how the system acquires and integrates multi-source data and uses deep reinforcement learning models to give dynamically adjusted meteorological risk warnings. At the same time, the system is constantly learning and optimizing in order to give more accurate and effective warnings.

[0065] In at least one embodiment of the present invention, a meteorological risk intelligent warning system based on multi-source data analysis is provided, which is used to store computer-readable instructions, and when the computer-readable instructions are read, the aforementioned meteorological risk intelligent warning method based on multi-source data analysis can be executed.

[0066] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation mode. The above-mentioned specific implementation mode is merely illustrative and not restrictive. Under the guidance of this embodiment, ordinary technicians in this field can also make more forms of equivalent embodiments, all of which are protected by this embodiment.

Claims

1. A meteorological risk intelligent early warning method based on multi-source data analysis, characterized in that: The following steps are involved: Step 100, obtaining multi-source meteorological data, historical meteorological risk event data and environmental change related data, and normalizing, feature extracting and encoding the data; Step 200, using a deep neural network to fuse the processed multi-source data and output a fused feature vector; Step 300, constructing a deep reinforcement learning model, taking the fused feature vector as the state space, taking the early warning strategy as the action space, and performing model training by setting an exploration strategy and a priority experience playback mechanism; Step 400, based on the data samples in the experience replay pool, the network parameters of the deep reinforcement learning model are updated by minimizing the loss function to obtain an optimized early warning strategy.

2. The method for intelligent early warning of meteorological risks based on multi-source data analysis according to claim 1 is characterized in that: The multi-source meteorological data includes at least one of temperature data, humidity data, wind speed data, and air pressure data.

3. The meteorological risk intelligent early warning method based on multi-source data analysis according to claim 1 is characterized in that: Reward function in deep reinforcement learning model The calculation method is: ； in, , , is the adjustment coefficient, which is used to adjust the relative importance of different reward items in the total reward. is the improvement of early warning accuracy, which is calculated as follows: ； in, is the number of true positive warnings, that is, the number of correct warnings of meteorological risk events; is the number of true negative warnings, that is, the number of times that the meteorological risk event was correctly warned but did not occur; is the number of false alarms, that is, the number of times that meteorological risk events are mistakenly warned; is the number of missed reports, that is, the number of actual meteorological risk events that were not warned; is the benchmark accuracy, this formula represents the current warning strategy The difference between the accuracy under the prediction and the benchmark accuracy is used to measure the improvement effect of the early warning strategy on the accuracy; is the change in false alarm rate, which is calculated as follows: ； in, is the baseline false alarm rate; is the warning timeliness gain, which is calculated as follows: ； in The warning release time is is the actual time when the meteorological event occurred, is the time-effect attenuation coefficient; when three consecutive false alarms occur, the penalty multiplication factor is triggered , the false positive item is corrected to: ； in is the penalty coefficient, that is, when there are three consecutive false alarms, the change in the false alarm rate will be doubled to strengthen the control of false alarms.

4. The method for intelligent early warning of meteorological risks based on multi-source data analysis according to claim 1 is characterized in that: The following steps are also included: Step 501: Obtain forecast data from a weather model ,in Represents different forecast time periods and variables. Perform feature extraction and normalization; Step 502: Based on step 200, the meteorological model prediction data The feature vectors of are input into the deep neural network for fusion, and the fused feature vector is output; Step 503, input the fused feature vector obtained in step 502, and output the trend prediction result of the meteorological risk in the future period of time.

5. The method for intelligent early warning of meteorological risks based on multi-source data analysis according to claim 1 is characterized in that: The following steps are also included: Step 601, obtaining multi-source meteorological data, historical meteorological risk event data, real-time environmental change related data and inter-regional geographical relationship data of multiple adjacent regions; performing the same processing as step 100 on these data of each region; Step 602 constructs a multi-layer fusion network. First, the normalized meteorological data feature vector, the historical meteorological risk event feature vector, and the environmental change feature representation in each region are fused to obtain a regional fusion feature vector. Then, the regional fusion feature vectors of all regions and the geographic relationship data between regions are input into a higher-layer fusion network to output a multi-region collaborative fusion feature vector. Step 603, building a deep reinforcement learning model; the same as step 300, but the state space is updated to a multi-region collaborative fusion feature vector; Step 604: Model training and updating; same as step 400.

6. The method for intelligent early warning of meteorological risks based on multi-source data analysis according to claim 1 is characterized in that: The following steps are also included: Step 701, quickly screen out key subset data that can be quickly obtained at the early stage of an extreme event from multi-source meteorological data, historical meteorological risk event data, and real-time environmental change related data, and perform normalization processing on these key subset data in the same manner as step 100; Step 702, designing a lightweight neural network, inputting feature vectors corresponding to the preprocessed key subset data, and outputting fast fusion feature vectors; Step 703, using a fast decision model based on Bayesian reasoning, inputting the fast fusion feature vector obtained in step 702, and outputting a warning decision for extreme weather events; Step 704, using historical data marked with extreme weather events to train the Bayesian model, optimize the estimation of prior probability and likelihood probability, so as to improve the accuracy of the model in warning extreme weather events under limited data.

7. The method for intelligent early warning of meteorological risks based on multi-source data analysis according to claim 1 is characterized in that: The following steps are also included: Step 801, in addition to meteorological data, also obtains wave height data, wave period data, current speed data, current direction data, ocean-related data, and historical risk event data of marine navigation and operation activities ; Normalize the ocean-related data, and similarly normalize the wave period, current speed, and current direction data; Extract key features, construct feature vectors, and perform the same processing as step 100 on meteorological data and other data; Step 802, using a deep neural network to fuse multi-source data, inputting normalized meteorological data feature vectors, ocean-related data feature vectors, and maritime historical risk event feature vectors, and outputting a fused feature vector; Step 803 is the same as step 300, but the state space Updated to the fused feature vector, action space Adjust the early warning strategy to suit maritime activities; Step 804, model training and updating, is the same as step 400.

8. The method for intelligent early warning of meteorological risks based on multi-source data analysis according to claim 1 is characterized in that: The following steps are also included: Step 901, in addition to the data in step 100, high-precision terrain data of mountainous areas is also obtained , historical meteorological risk event data in mountainous areas ;right Perform feature extraction and encoding to obtain the corresponding feature representation. Extract key features, construct feature vectors, and perform the same processing as step 100 on meteorological data and other data; Step 902, using a deep neural network to fuse multi-source data, inputting normalized meteorological data feature vectors, mountain terrain feature representations, and mountain historical risk event feature vectors, and outputting a fused feature vector; Step 903, building a deep reinforcement learning model, is the same as step 300, but the state space Updated to the fused feature vector, action space Adjust the early warning strategy to suit mountain areas; Step 904, model training and updating; same as step 400.

9. The method for intelligent early warning of meteorological risks based on multi-source data analysis according to claim 1 is characterized in that: The following steps are also included: Step 1001, in addition to the data in step 100, also obtain the distribution data of each functional area of ​​the city , historical meteorological risk event data for each functional area , and real-time dynamic data related to the functional area ,right Encode the corresponding feature representation. Extract key features, construct feature vectors, and Perform normalization processing, and perform the same processing as step 100 on the meteorological data and other data; Step 1002, using a deep neural network to fuse multi-source data, inputting normalized meteorological data feature vectors, functional area distribution feature representations, functional area historical risk event feature vectors, and functional area real-time dynamic data feature vectors, and outputting a fused feature vector; Step 1003, building a deep reinforcement learning model, is the same as step 300, but the state space Updated to the fused feature vector, action space Adjust to early warning strategies for different functional areas; Step 1004, model training and updating, is the same as step 400.

10. A meteorological risk intelligent early warning system based on multi-source data analysis, characterized in that: It is used to store computer-readable instructions, which, when read, can execute a meteorological risk intelligent early warning method based on multi-source data analysis as described in any one of claims 1-9.

Images