Extreme weather event prediction method and system coupled with pulse neural network

Abstract

The invention relates to the technical field of weather prediction methods, in particular to a spiking neural network coupled extreme weather event prediction method and a spiking neural network coupled extreme weather event prediction system. By constructing a multi-scale meteorological element time sequence database and utilizing a WRF climate mode to obtain high-resolution data, GPU cluster parallel processing is adopted, and a biological neuron and recurrent neuron hybrid model SNN-LSTM is combined, so that an extreme weather event prediction result is obtained. The model comprises two layers of 20 GRU cells, and is used for resampling and time sequence prediction of meteorological data in a one-hour scale to identify extreme weather events, the method extracts and updates internal states in different time scales, and finally outputs a northwest Pacific Ocean Typhoon trajectory and forms a probability prediction model. According to the method, the accuracy, efficiency and interpretability of extreme weather prediction are effectively improved, a new development direction is provided for the weather prediction technology, and the method has remarkable theoretical and practical values.

Description

Technical Field

[0001] The present invention relates to the technical field of weather forecasting methods, and in particular to a pulse neural network coupled extreme weather event forecasting method and system thereof. Background Art

[0002] In recent years, with the intensification of global climate change, the frequency and intensity of extreme weather events have continued to increase, significantly impacting the social economy and people's lives. Typhoons, one of the most destructive extreme weather events, have a particularly significant impact on the Northwest Pacific region. Therefore, accurately predicting extreme weather events, particularly the formation, development, and path of typhoons, has become a top priority in meteorological research.

[0003] Traditional extreme weather prediction methods primarily rely on numerical weather forecasting models, such as the widely used WRF (Weather Research and Forecasting) model. These models simulate atmospheric motion by solving a set of atmospheric dynamics equations. While they can predict weather changes to a certain extent, they still have many limitations when dealing with nonlinear, multi-scale extreme weather events. First, numerical models are extremely sensitive to initial and boundary conditions, and small errors can lead to significant deviations in prediction results. Second, the models are computationally complex and require a large amount of computing resources, making it difficult to achieve real-time, efficient predictions. Furthermore, traditional models are also inadequate in processing long-term data series and capturing the complex relationships between multi-scale meteorological elements.

[0004] In recent years, with the rapid development of artificial intelligence technology, researchers have begun to attempt to apply machine learning methods to extreme weather prediction. Among them, deep learning models such as long short-term memory networks (LSTMs) have performed well in time series prediction tasks and are widely used in weather forecasting. However, these models still face challenges in processing highly nonlinear and multi-scale meteorological data. For example, although the LSTM model can capture long-term dependencies, it is not very effective in dealing with extreme events with sudden, short-term changes. In addition, traditional deep learning models have poor interpretability and it is difficult to provide a physical explanation for the prediction results, which is a significant flaw in key areas such as weather forecasting.

[0005] Meanwhile, spiking neural networks (SNNs), a new neural network model that mimics biological neural systems, have recently demonstrated great potential in spatiotemporal data processing. SNNs can effectively process time-varying data and offer advantages such as low power consumption and high parallelism. However, simple SNN models still struggle to process long-term data series and are unable to effectively capture the long-term dependencies found in meteorological systems.

[0006] Given the shortcomings of existing technologies, new extreme weather forecasting methods are urgently needed that can leverage the strengths of numerical models, deep learning, and spiking neural networks. Such methods should be able to effectively process multi-scale, highly nonlinear meteorological data, accurately capture both short-term mutations and long-term trends, and possess high computational efficiency and interpretability. Summary of the Invention

[0007] The proposed method and system for extreme weather event prediction using a spiking neural network coupled with the proposed method are designed to address these technical issues. By innovatively integrating the SNN and LSTM models and coupling them with the traditional WRF numerical model, this method achieves efficient and accurate prediction of extreme weather events.

[0008] The present invention proposes a method for predicting extreme weather events using a pulse neural network coupling method, comprising:

[0009] The acquisition steps include:

[0010] Construct a multi-scale meteorological element time series database, obtain the element initialization and downscaling calculation results of specific regions in the climate model WRF, and obtain fine-resolution meteorological element data as a data source;

[0011] Loading the data source into the GPU cluster through parallel processing;

[0012] Processing steps include:

[0013] Constructing a hybrid model of biological neurons and recurrent neurons, the SNN-LSTM model includes two layers of neurons, each layer includes 20 GRU cells, and each GRU cell includes three parts: an activation function f(), a reset gate r(), and an update gate z();

[0014] Meteorological data is resampled on a 1-hour scale, and time windows are constructed for meteorological elements based on bionic neurons. Meteorological elements are defined as biological neurons, which are then input into the hybrid model of biological neurons and recurrent neurons, SNN-LSTM.

[0015] Extract and update internal states at different time scales to achieve time series forecasting of meteorological elements and predict extreme weather events in meteorological elements;

[0016] Output steps include:

[0017] Based on the time series prediction results, the northwest Pacific typhoon track prediction and typhoon formation probability prediction models are constructed respectively.

[0018] Preferably, the specific area in the obtaining step refers to the northwest Pacific Ocean.

[0019] Preferably, constructing the biological neuron and recurrent neuron hybrid model SNN-LSTM in the processing step specifically includes:

[0020] Use sigmoid function as activation function f();

[0021] Introduce bionic neuron parameters in the reset gate r() and update gate z();

[0022] The bionic neuron introduces two variables: the time-dependent function τ(t j ,t) and transient pulse rate ω(t j ,t); The improved STDP learning rule is used to update the weights of each bionic neuron.

[0023] Preferably, the improved STDP learning rule includes:

[0024] Neurons i and j are activated simultaneously;

[0025] Calculate the spatial coupling strength between related neuron pairs within the time window;

[0026] Adjust the memory unit of neuron i by neuron j.

[0027] Preferably, in the processing step, the time window constructed according to the bionic neuron is: [t j ,t), where t j is the delay from the last activation to the current activation, t is the current time, τ(t j ,t) is the time correlation function, ω(t j ,t) is the transient pulse rate.

[0028] Preferably, the STDF learning rule in the bionic neuron corresponding to the time window is as follows:

[0029] δw=η(r⊙Δz+z⊙Δr),

[0030] Among them, r and z are the output values of the reset gate and update gate of the bionic neuron respectively, w is the internal weight of the bionic neuron, η is the learning rate, Δz and Δr are the state changes of the update gate and reset gate respectively, and p(t j ,t) is the memory item of the bionic neuron.

[0031] Preferably, the processing step further comprises:

[0032] During the extraction and updating process of meteorological data, within the time window, the outputs of the reset gate and update gate of the bionic neuron are connected to the fully connected layer of the neural network. The time window is updated by adjusting the weights of the neuron and the neurons in the fully connected layer.

[0033] Preferably, the processing step further comprises:

[0034] A weather prediction network that integrates the SNN-LSTM hybrid model is used. The coarse-resolution numerical weather forecast is used as input data. The error between the SNN-LSTM hybrid model and the fine-resolution numerical weather forecast is fed back multiple times, and the error is further input into the neural network. At the same time, the gradient descent method of training error estimation is used to fine-tune all model parameters.

[0035] Preferably, the typhoon track prediction model in the output step specifically includes:

[0036] The typhoon track forecast results predicted by the SNN-LSTM hybrid model are combined with the initial position coordinates;

[0037] Based on historical data on typhoon occurrence, intensification, and dissipation, an adaptive threshold is used to forecast typhoon levels;

[0038] Typhoon levels are classified based on the predicted wind speed at 10m height at the typhoon eye and sea level pressure greater than or equal to 255,000 hPa or sea level pressure near the center less than or equal to 998 hPa.

[0039] The extreme weather event prediction system coupled with a pulse neural network includes:

[0040] The meteorological element time series database module is used to build a multi-scale meteorological element time series database, initialize elements for specific regions in the climate model WRF, and obtain fine-resolution meteorological element data as a data source; at the same time, the data source is loaded into the GPU cluster through parallel processing;

[0041] The SNN-LSTM model module is used to construct the hybrid model of biological neurons and recurrent neurons (SNN-LSTM). The hybrid model of biological neurons and recurrent neurons (SNN-LSTM) contains two layers of neurons, each layer contains 20 GRU cells, and each GRU cell consists of three parts: an activation function f(), a reset gate r(), and an update gate z().

[0042] The extreme weather prediction module is used to resample meteorological data on a one-hour scale, construct time windows for meteorological elements based on bionic neurons, define meteorological elements as biological neurons, input them into the SNN-LSTM hybrid model of biological neurons and recurrent neurons, and extract and update internal states at different time scales to achieve time series prediction of meteorological elements and predict extreme weather events among meteorological elements;

[0043] The Northwest Pacific typhoon track and typhoon formation prediction modules are used to construct the Northwest Pacific typhoon track prediction and typhoon formation probability prediction models respectively based on the time series prediction results in the extreme weather prediction module.

[0044] The method of the present invention has the following significant advantages and beneficial effects:

[0045] First, by constructing a multi-scale time-series database of meteorological elements, we integrated meteorological data from different sources and resolutions, providing comprehensive and rich input information for forecasting. Second, we innovatively designed a hybrid model of biological neurons and recurrent neurons, the SNN-LSTM, which effectively combines the advantages of SNNs in processing time-varying data with the ability of LSTMs to capture long-term dependencies, significantly improving the model's ability to represent complex meteorological systems.

[0046] Furthermore, the present invention introduces an improved STDP learning rule, enabling the model to dynamically adjust internal parameters based on the characteristics of the input data, improving its adaptability and robustness. By extracting and updating internal states at different time scales, this method can simultaneously capture both short-term mutations and long-term trends, providing more comprehensive and accurate information for the prediction of extreme weather events.

[0047] It is particularly noteworthy that the method proposed in this paper, through its innovative coupling with the WRF model, achieves complementary advantages between numerical models and machine learning methods. This not only improves the accuracy of predictions, but also enhances the interpretability of the model, providing a physical basis for the prediction results.

[0048] In practical applications, the method of this invention has demonstrated excellent performance in typhoon forecasting in the Northwest Pacific, accurately predicting the formation, development, and path of typhoons. This is of great significance for reducing the economic losses and casualties caused by typhoons. Furthermore, this method utilizes GPU parallel processing technology, significantly improving computational efficiency and enabling real-time, efficient extreme weather forecasting.

[0049] In summary, the method proposed in this invention effectively solves the challenges faced by existing technologies in extreme weather forecasting through a number of innovative designs, provides a new technical path for improving forecast accuracy, efficiency and interpretability, and has important theoretical significance and practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is the overall flow chart of the present invention.

[0051] Figure 2 This is a flow chart of the extreme weather prediction module 3 of the present invention.

[0052] Figure 3 This is a flow chart of the northwest Pacific typhoon track and typhoon formation prediction module 4 of the present invention. DETAILED DESCRIPTION

[0053] Please refer to the attached Figure 1-3 The present invention provides a method and system for predicting extreme weather events using a pulse neural network coupling method. The method mainly includes the following steps:

[0054] First, in the acquisition step, the present invention constructs a multi-scale time-series database of meteorological elements. Specifically, element initialization and downscaling calculations are performed on specific regions of the WRF climate model to obtain fine-resolution meteorological element data as a data source. Preferably, the present invention uses parallel processing technology to load this data source onto a GPU cluster to improve data processing efficiency.

[0055] Secondly, in the processing step, the present invention constructs a hybrid model of biological neurons and recurrent neurons, SNN-LSTM. The model contains two layers of neurons, each layer contains 20 GRU cells. Each GRU cell consists of three parts: an activation function f(), a reset gate r(), and an update gate z(). In one embodiment of the present invention, the activation function f() adopts a sigmoid function, whose mathematical expression is:

[0056]

[0057] Where x is the input value and f(x) is the output value.

[0058] The bionic neuron parameters are introduced into the reset gate r() and the update gate z(), which is an innovation of the present invention. Specifically, the bionic neuron introduces two key variables: the time-dependent function τ(t j ,t) and transient pulse rate ω(t j ,t). The introduction of these two variables enables the model to better capture the long-term and short-term dependencies in time series data.

[0059] During the processing process, the present invention resamples meteorological data to a one-hour timescale. This timescale was chosen based on empirical values from meteorological research to capture short-term weather changes while minimizing the introduction of excessive noise. Subsequently, time windows are constructed for meteorological elements based on bionic neurons and defined as biological neurons. This processing approach enables the model to better simulate the behavior of neurons in nature, thereby improving forecast accuracy.

[0060] Next, the present invention feeds the processed data into the constructed SNN-LSTM hybrid model. Within the model, time-series forecasting of meteorological elements is achieved by extracting and updating internal states at different time scales. This multi-scale forecasting approach enables the model to simultaneously capture short-term, medium-term, and long-term weather trends, leading to more accurate predictions of extreme weather events.

[0061] Finally, in the output step, based on the above time series prediction results, the present invention constructs a typhoon trajectory prediction model for the Northwest Pacific and a typhoon formation probability prediction model. The construction of these two models fully utilizes the advantages of the SNN-LSTM hybrid model and can more accurately predict the formation and path of typhoons.

[0062] It is worth noting that the specific region in the method of the present invention is specifically the Northwest Pacific. This region was chosen as the research object because it is one of the regions with the most frequent typhoon activity in the world, and accurate prediction of this region is of great practical significance.

[0063] In the construction process of the SNN-LSTM hybrid model, the present invention adopts an improved STDP (Spike-Timing-Dependent Plasticity) learning rule to update the weights of each bionic neuron. This improved learning rule enables the model to better adapt to the characteristics of meteorological data and improves the accuracy of prediction. Specifically, for the i-th bionic neuron, its weight update can be expressed as:

[0064] δz i =net z,i -z i ,

[0065] Among them, δz i Indicates the weight update amount, net z,i represents the net input of neuron i, z i represents the current state of neuron i.

[0066] Time correlation function τ(t j ,t) and transient pulse rate ω(t j ,t) can be calculated by the following formula:

[0067]

[0068] Where τ0 and ω0 are time constants, t j The introduction of these two functions enables the model to more accurately capture the temporal dependencies in time series data.

[0069] Through the above steps, the method of the present invention can effectively predict extreme weather events, especially typhoons in the northwest Pacific region. This method not only improves prediction accuracy but also enhances computational efficiency through GPU parallel processing, thus having important practical application value.

[0070] In a preferred embodiment of the present invention, the improved STDP learning rule includes the following specific steps: First, neurons i and j are activated simultaneously. This simultaneous activation mechanism simulates the co-activation phenomenon in biological neural networks and helps capture the interactive relationship between multiple meteorological elements.

[0071] Next, this method calculates the spatial coupling strength between the relevant neuron pairs within the time window. The calculation formula for the spatial coupling strength is as follows:

[0072]

[0073] Among them, S ij represents the spatial coupling strength between neurons i and j, x i (t) and x j (t) represents the output of neurons i and j at time t, and t0 and t1 represent the start and end times of the time window, respectively. This calculation method takes into account the temporal correlation of neuron outputs and can better reflect the dynamic correlation between meteorological elements.

[0074] Finally, the method of the present invention adjusts the memory unit of neuron j to neuron i. The adjustment formula is as follows:

[0075] ΔM ij =ηS ij (x j (t)-M ij ),

[0076] Among them, ΔM ij Represents the adjustment amount of the memory unit, η is the learning rate, M ij Represents the current memory cell value. The learning rate η is typically set between 0.001 and 0.1, and the specific value needs to be tuned based on the actual data characteristics. This memory cell adjustment mechanism enables the model to better retain historical information and improve its ability to capture long-term meteorological trends.

[0077] In another embodiment of the present invention, STDP (Spike Time Dependent Plasticity) is used to adjust the temporal correlation function and the transient spike rate for each GRU cell. The STDP learning rule satisfies the mathematical expression of balancing spike time facilitation and inhibition. Specifically, τ0 and ω0 are time constants, R(t) represents the activity time of the neuron, and r(t j ,t) and z(t j ,t) are the hidden layer output results of the reset gate and update gate respectively, and E(t) is the pulse energy.

[0078] For positive energy pulses, the more frequent the neuron activity, the stronger the pulse time-dependent plasticity, and vice versa. This mechanism simulates the plasticity characteristics of biological neural systems, enabling the model to dynamically adjust its internal parameters according to the characteristics of the input data. In the SNN-LSTM hybrid model, the SNN model and the recurrent neuron The input and output methods and working principles of the Lange unit are as follows:

[0079] First, the STDP learning rule in each bionic neuron updates the internal weight w and bias b of the bionic neuron through the parameters in the GRU. According to the STDP rule, the pulse energy of the bionic neuron is calculated as follows:

[0080]

[0081] Among them, t f Indicates the time when the pulse occurs, τ E is the energy decay time constant. The larger the energy value, the more frequent the pulse activity of the bionic neuron.

[0082] Secondly, the time window is controlled by the reset gate and the update gate. The output of the reset gate r is the input of the neuron at the current moment and the cell state c at the current moment. t-1 The weighted sum of the update gate z output is the activation function sigmoid in the reset gate and cell state c t-1 The product of the cell state and the output value of the activation function sigmoid on the reset gate is used as the new state of the cell, and the output of the cell at the current moment is calculated based on the new state of the cell.

[0083] In a preferred embodiment of the present invention, the input of the bionic neuron includes recursive input, cyclic input, long-range input and pulse input. Specifically:

[0084] 1. Recursive input:

[0085] r t =χ(Dr t ),

[0086] Dr t =[r t ,τ(t j ,t)x t ],

[0087] Among them, Dr. t To reset the gate dynamics, r t is the reset gate output, χ(Dr t ) is the reset gate function.

[0088] 2. Loop input:

[0089] z t=v(Dz t ),

[0090] Among them, Dz t To update the door dynamics, z t is the update gate output, v(Dz t ) is the update gate function.

[0091] 3. Long-range input:

[0092]

[0093] in, To reset the gate and update the output energy of the gate.

[0094] 4. Pulse input:

[0095]

[0096] Among them, y(t j ,t) is the output of bionic neurons.

[0097] In GRU, the recurrent input can be viewed as a constant value for the reset gate and the update gate. Therefore, the activation function of the bionic neuron can be defined as follows:

[0098]

[0099] Among them, f, r t 、z t They are the activation function, reset gate, and update gate of the bionic neuron respectively. Represents element-wise multiplication.

[0100] Through the detailed model design and parameter adjustments described above, the method of the present invention can more accurately capture the spatiotemporal characteristics of meteorological data, thereby improving the accuracy of extreme weather event predictions. This method demonstrates significant advantages when processing long-term and multi-scale meteorological data.

[0101] In another embodiment of the present invention, the processing step further includes employing a weather prediction network that incorporates a SNN-LSTM hybrid model. This network architecture is designed to leverage the advantages of the SNN model in processing spatiotemporal data and the LSTM model's ability to capture long-term dependencies. Specifically, this method uses coarse-resolution numerical weather forecasts as input data and provides multiple feedback loops using the error between the SNN-LSTM hybrid model and the fine-resolution numerical weather forecast.

[0102] Preferably, the method of the present invention further inputs the above-mentioned error into the neural network. This error feedback mechanism can continuously optimize the model's prediction results and improve prediction accuracy. At the same time, this method uses the gradient descent method of training error estimation to fine-tune all model parameters. The learning rate of the gradient descent method is usually set between 0.001 and 0.01, and the specific value needs to be adjusted according to the actual data characteristics and the model convergence speed. This parameter fine-tuning strategy can make the model better adapt to different meteorological conditions and improve the model's generalization ability.

[0103] In a preferred embodiment of the present invention, the typhoon track prediction model in the output step has the following characteristics: First, the model combines the typhoon track forecast results predicted by the SNN-LSTM hybrid model with the initial position coordinates. This combination fully utilizes historical data and real-time observation data to more accurately predict the typhoon's movement trajectory.

[0104] Secondly, this method uses adaptive thresholds to predict typhoon levels based on historical data on typhoon occurrence, intensification, and dissipation. The use of adaptive thresholds enables the model to dynamically adjust the prediction criteria based on the characteristics of typhoons in different periods and regions, thereby improving the flexibility and accuracy of the prediction. Specifically, the adaptive threshold θ t It can be calculated by the following formula:

[0105] θ t =Δθ t-1 +(1-α)x t ,

[0106] Among them, α is the smoothing factor, which is usually between 0.8 and 0.95, and x t is the current observation value. This formula describes the exponential weighted average method used to calculate θ t , which combines the smoothed estimate θ from the previous time step t-1 and the observation value x at the current time step t .

[0107] Finally, the method of the present invention classifies typhoons based on the predicted 10-meter wind speed at the typhoon eye and the sea level pressure. The specific classification criteria are as follows: a typhoon is considered strong when the wind speed at 10 meters is greater than or equal to 25.5 m / s, or when the sea level pressure near the center is less than or equal to 998 hPa. These thresholds are based on empirical values from meteorological research and can accurately distinguish typhoons of varying intensities.

[0108] The present invention also provides an extreme weather event prediction system coupled with a pulse neural network corresponding to the above method. The system includes the following modules:

[0109] Meteorological Element Time Series Database Module 1: This module is used to build a multi-scale meteorological element time series database. Specifically, it initializes elements for specific regions in the WRF climate model to obtain fine-resolution meteorological element data as a data source. Preferably, this module loads the data source into the GPU cluster through parallel processing to improve data processing efficiency.

[0110] SNN-LSTM model module 2 is used to construct the SNN-LSTM hybrid model of biological neurons and recurrent neurons. In one embodiment of the present invention, this hybrid model includes two layers of neurons, each layer containing 20 GRU cells. Each GRU cell consists of three parts: an activation function f(), a reset gate r(), and an update gate z(). This structural design enables the model to simultaneously process short-term and long-term temporal dependencies, making it suitable for processing complex meteorological time series data.

[0111] Extreme Weather Forecast Module 3 resamples meteorological data to a one-hour scale, constructs time windows for meteorological elements based on bionic neurons, defines meteorological elements as biological neurons, and inputs these into a SNN-LSTM hybrid model. This module extracts and updates internal states at different time scales, enabling time series forecasting of meteorological elements and predicting extreme weather events within these elements. This multi-scale forecasting approach captures meteorological changes across different time scales, improving the comprehensiveness and accuracy of forecasts.

[0112] Northwest Pacific Typhoon Track and Formation Prediction Module 4: Based on the time-series forecast results from Extreme Weather Prediction Module 3, this module constructs a Northwest Pacific typhoon track prediction model and a typhoon formation probability prediction model. The combination of these two models enables the system to comprehensively predict the formation, development, and movement of typhoons, providing important decision-making support for disaster prevention and mitigation.

[0113] By combining these modules, the system can efficiently and accurately predict extreme weather events, particularly typhoons in the Northwest Pacific. This system not only improves forecast accuracy but also boosts computational efficiency through GPU parallel processing, demonstrating its significant practical application.

[0114] It should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A pulse neural network coupled extreme weather event prediction method, characterized in that: include: The acquisition steps include: Construct a multi-scale meteorological element time series database, obtain the element initialization and downscaling calculation results of specific regions in the climate model WRF, and obtain fine-resolution meteorological element data as a data source; Loading the data source into the GPU cluster through parallel processing; Processing steps include: Constructing a hybrid model of biological neurons and recurrent neurons, the SNN-LSTM model includes two layers of neurons, each layer includes 20 GRU cells, and each GRU cell includes three parts: an activation function f(), a reset gate r(), and an update gate z(); Meteorological data is resampled on a 1-hour scale, and time windows are constructed for meteorological elements based on bionic neurons. Meteorological elements are defined as biological neurons, which are then input into the hybrid model of biological neurons and recurrent neurons, SNN-LSTM. Extract and update internal states at different time scales to achieve time series forecasting of meteorological elements and predict extreme weather events in meteorological elements; Output steps include: Based on the time series prediction results, the northwest Pacific typhoon track prediction and typhoon formation probability prediction models are constructed respectively.

2. The extreme weather event prediction method coupled with a pulse neural network according to claim 1, characterized in that: The specific area in the obtaining step refers to the northwest Pacific Ocean.

3. The extreme weather event prediction method coupled with a pulse neural network according to claim 1, characterized in that: The processing steps of constructing the hybrid model of biological neurons and recurrent neurons SNN-LST M specifically include: Use sigmoid function as activation function f(); Introduce bionic neuron parameters in the reset gate r() and update gate z(); The bionic neuron introduces two variables: the time-dependent function τ(t j ,t) and transient pulse rate ω(t j ,t); The improved STDP learning rule is used to update the weights of each bionic neuron.

4. The extreme weather event prediction method coupled with a pulse neural network according to claim 3, characterized in that: The improved STDP learning rule includes: Neurons i and j are activated simultaneously; Calculate the spatial coupling strength between related neuron pairs within the time window; Adjust the memory unit of neuron i by neuron j.

5. The extreme weather event prediction method coupled with a pulse neural network according to claim 1, characterized in that: In the processing step, the time window constructed according to the bionic neuron is: [t j ,t), where t j is the delay from the last activation to the current activation, t is the current time, τ(t j ,t) is the time correlation function, ω(t j ,t) is the transient pulse rate.

6. The method for predicting extreme weather events using a pulse neural network coupling method according to claim 5, wherein: The STDF learning rule in the bionic neuron corresponding to the time window is as follows: δw=η(r⊙Δz+z⊙Δr), Among them, r and z are the output values of the reset gate and update gate of the bionic neuron respectively, w is the internal weight of the bionic neuron, η is the learning rate, Δz and Δr are the state changes of the update gate and reset gate respectively, and p(t j ,t) is the memory item of the bionic neuron.

7. The extreme weather event prediction method coupled with a pulse neural network according to claim 1, characterized in that: The processing steps further include: During the extraction and updating process of meteorological data, within the time window, the outputs of the reset gate and update gate of the bionic neuron are connected to the fully connected layer of the neural network. The time window is updated by adjusting the weights of the neuron and the neurons in the fully connected layer.

8. The extreme weather event prediction method coupled with a pulse neural network according to claim 1, characterized in that: The processing steps further include: A weather prediction network that integrates the SNN-LSTM hybrid model is used. The coarse-resolution numerical weather forecast is used as input data. The error between the SNN-LSTM hybrid model and the fine-resolution numerical weather forecast is fed back multiple times, and the error is further input into the neural network. At the same time, the gradient descent method of training error estimation is used to fine-tune all model parameters.

9. The extreme weather event prediction method coupled with a pulse neural network according to claim 1, characterized in that: The typhoon track prediction model in the output step specifically includes: The typhoon track forecast results predicted by the SNN-LSTM hybrid model are combined with the initial position coordinates; Based on historical data on typhoon occurrence, intensification, and dissipation, an adaptive threshold is used to forecast typhoon levels; Typhoon levels are classified based on the predicted wind speed at 10m height at the typhoon eye and sea level pressure greater than or equal to 255,000 hPa or sea level pressure near the center less than or equal to 998 hPa.

10. A pulse neural network coupled extreme weather event prediction system for executing the method according to any one of claims 1 to 9, characterized in that: include: The meteorological element time series database module is used to build a multi-scale meteorological element time series database, initialize elements for specific regions in the climate model WRF, and obtain fine-resolution meteorological element data as a data source; at the same time, the data source is loaded into the GPU cluster through parallel processing; The SNN-LSTM model module is used to construct the hybrid model of biological neurons and recurrent neurons (SNN-LSTM). The hybrid model of biological neurons and recurrent neurons (SNN-LSTM) contains two layers of neurons, each layer contains 20 GRU cells, and each GRU cell consists of three parts: an activation function f(), a reset gate r(), and an update gate z(). The extreme weather prediction module is used to resample meteorological data on a one-hour scale, construct time windows for meteorological elements based on bionic neurons, define meteorological elements as biological neurons, input them into the SNN-LSTM hybrid model of biological neurons and recurrent neurons, and extract and update internal states at different time scales to achieve time series prediction of meteorological elements and predict extreme weather events among meteorological elements; The Northwest Pacific typhoon track and typhoon formation prediction modules are used to construct the Northwest Pacific typhoon track prediction and typhoon formation probability prediction models respectively based on the time series prediction results in the extreme weather prediction module.

Images