Deep learning-based thunderstorm gale quantitative prediction method and device

Abstract

The application provides a thunderstorm gale quantitative prediction method and device based on deep learning, and relates to the technical field of weather prediction, and the method comprises the following steps: determining target multi-source meteorological data of a target area based on multiple meteorological data sources; inputting the target multi-source meteorological data into a thunderstorm gale prediction model to obtain multiple predicted wind speed values output by the thunderstorm gale prediction model; the thunderstorm gale prediction model is obtained by training based on sample multi-source meteorological data; comparing the multiple predicted wind speed values with a preset wind speed threshold value, and dividing the target area into an extreme thunderstorm gale area and a common thunderstorm gale area according to the comparison result. According to the technical scheme of the application, the thunderstorm gale prediction model can learn the storm evolution law in a long time sequence, improve the accuracy of thunderstorm gale area prediction, and accurately distinguish the common thunderstorm gale area from the extreme thunderstorm gale area according to the comparison between the multiple predicted wind speed values and the preset wind speed threshold value.

Description

Technical Field

[0001] This invention relates to the field of weather forecasting technology, and in particular to a method and apparatus for quantitative forecasting of thunderstorms and strong winds based on deep learning. Background Technology

[0002] Thunderstorm gales refer to gale-force winds with instantaneous wind speeds ≥ 17.2 m / s accompanied by thunderstorms, while extreme thunderstorm gales refer to severe convective weather with instantaneous wind speeds ≥ 25 m / s accompanied by lightning. These often cause disasters comparable to or exceeding the destructive power of tornadoes. Due to their sudden onset, high destructive force, short defense time, and difficulty in early warning, extreme thunderstorm gales can easily lead to severe casualties and huge economic losses. Therefore, short-term nowcasting and early warning technology for extreme thunderstorm gales has always been a key focus and challenge in severe convective weather research.

[0003] Current technologies for short-term forecasting of extreme thunderstorms and strong winds primarily rely on Doppler weather radar extrapolation. However, Doppler weather radar extrapolation is based solely on motion vectors from the most recent few frames, neglecting the ability to learn about storm evolution patterns over long time series, resulting in a high rate of missed predictions for rapidly intensifying extreme thunderstorms and strong winds. Furthermore, the forecasts output by current technologies do not specifically identify and classify extreme thunderstorms and strong winds (gusts ≥ 25 m / s), making it difficult to accurately distinguish between ordinary thunderstorms and strong winds and extreme thunderstorms and strong winds. Summary of the Invention

[0004] This invention provides a method and apparatus for quantitative forecasting of thunderstorms and gales based on deep learning. It addresses the shortcomings of existing technologies, such as neglecting the ability to learn the evolution patterns of storms over long time series, high underreporting rates for rapidly intensifying extreme thunderstorms and gales, and the lack of specific identification and tiered forecasting for extreme thunderstorms and gales, making it difficult to accurately distinguish between ordinary and extreme thunderstorms and gales. The technical solution of this invention enables the learning of storm evolution patterns over long time series through a thunderstorm and gale forecasting model, improving the accuracy of thunderstorm and gale area prediction. Furthermore, it can accurately distinguish between ordinary and extreme thunderstorm and gale areas by comparing multiple forecast wind speed values ​​with preset wind speed thresholds.

[0005] This invention provides a method for quantitative forecasting of thunderstorms and strong winds based on deep learning, comprising the following steps.

[0006] Target multi-source meteorological data for the target area is determined based on multiple meteorological data sources; The target multi-source meteorological data is input into the thunderstorm and gale forecasting model to obtain multiple forecast wind speed values ​​output by the thunderstorm and gale forecasting model; the thunderstorm and gale forecasting model is trained based on sample multi-source meteorological data; By comparing the multiple forecast wind speed values ​​with preset wind speed thresholds, the target area is divided into extreme thunderstorm wind areas and ordinary thunderstorm wind areas based on the comparison results.

[0007] According to the present invention, a deep learning-based quantitative forecasting method for thunderstorms and strong winds is provided, the method further comprising: Multiple wind speed increments are determined based on the multiple predicted wind speed values; For each of the aforementioned wind speed increments, if the wind speed increment exceeds a preset increment threshold, an early warning signal is generated.

[0008] According to the present invention, a deep learning-based quantitative forecasting method for thunderstorms and strong winds is provided, the method further comprising: The display device displays one or more of the following: the multiple forecast wind speed values, the extreme thunderstorm wind area, the ordinary thunderstorm wind area, and the warning signal.

[0009] According to the present invention, a deep learning-based quantitative forecasting method for thunderstorms and strong winds is provided, wherein determining target multi-source meteorological data for a target area based on multiple meteorological data sources includes: First multi-source meteorological data of the target area is obtained from the multiple meteorological data sources; Spatial coordinate matching and time step interpolation are performed on the first multi-source meteorological data to obtain the second multi-source meteorological data; The second multi-source meteorological data is normalized to obtain the third multi-source meteorological data; The third multi-source meteorological data is stacked and recombined to obtain the target multi-source meteorological data.

[0010] According to the deep learning-based quantitative forecasting method for thunderstorms and strong winds provided by the present invention, the step of stacking and recombining the third multi-source meteorological data to obtain the target multi-source meteorological data includes: The multi-source meteorological data is extracted using a sliding window sampler to obtain a multi-source meteorological sequence; The multi-source meteorological sequence is recombined in three dimensions—sample, time step, and spatial features—using a tensor stacker to obtain the target multi-source meteorological data.

[0011] According to the present invention, a deep learning-based quantitative forecasting method for thunderstorms and strong winds is provided. The multiple meteorological data sources include one or more of Doppler weather radar, regional automatic weather station network, wind profiler radar, and satellite data receiving station. The first multi-source meteorological data includes one or more of gust wind speed, precipitation over a preset period, historical weather, current weather, sea level pressure, vertical velocity, low cloud cover, middle cloud cover, middle cloud type, and low cloud type.

[0012] According to the present invention, a deep learning-based quantitative forecasting method for thunderstorms and strong winds is provided. The thunderstorm and strong wind forecasting model includes an input layer, multiple bidirectional long short-term memory network layers, multiple random deactivation layers, and an output layer. The multiple bidirectional long short-term memory network layers and the multiple random deactivation layers are alternately connected in series. The bidirectional long short-term memory network layers are configured with input gates, forget gates, output gates, and peak retention unit state update branches.

[0013] According to the present invention, a deep learning-based quantitative forecasting method for thunderstorms and strong winds is provided, wherein the thunderstorm and strong wind forecasting model is trained based on the following steps: The sample multi-source meteorological data is input into the initial thunderstorm and gale forecast model. Based on the root mean square error and mean absolute error between the training forecast wind speed value and the actual wind speed value output by the initial thunderstorm and gale forecast model, as well as the wind force level score of the training forecast wind speed value, the initial thunderstorm and gale forecast model is iteratively trained to obtain the thunderstorm and gale forecast model. The network structure of the thunderstorm and gale forecast model is the same as that of the initial thunderstorm and gale forecast model.

[0014] According to the present invention, a deep learning-based quantitative forecasting method for thunderstorms and strong winds is provided, wherein the plurality of bidirectional long short-term memory network layers include a first bidirectional long short-term memory network layer, a second bidirectional long short-term memory network layer, a third bidirectional long short-term memory network layer, and a fourth bidirectional long short-term memory network layer; the random inactivation layer includes a first random inactivation layer, a second random inactivation layer, a third random inactivation layer, and a fourth random inactivation layer; The step involves inputting the target multi-source meteorological data into the thunderstorm and gale forecasting model to obtain multiple forecast wind speed values ​​output by the model, including: The target multi-source meteorological data is input into the input layer to obtain the first feature output by the input layer; The first feature is input into the first bidirectional long short-term memory network layer to obtain the second feature output by the first bidirectional long short-term memory network layer. The second feature is input into the first random deactivation layer to obtain the third feature output by the first random deactivation layer; The third feature is input into the second bidirectional long short-term memory network layer to obtain the fourth feature output by the second bidirectional long short-term memory network layer; The fourth feature is input into the second random deactivation layer to obtain the fifth feature output by the second random deactivation layer; The fifth feature is input into the third bidirectional long short-term memory network layer to obtain the sixth feature output by the third bidirectional long short-term memory network layer. The sixth feature is input into the third random deactivation layer to obtain the seventh feature output by the third random deactivation layer; The seventh feature is input into the fourth bidirectional long short-term memory network layer to obtain the eighth feature output by the fourth bidirectional long short-term memory network layer. The eighth feature is input into the fourth random deactivation layer to obtain the ninth feature output by the fourth random deactivation layer; The ninth feature is input into the output layer to obtain the plurality of forecast wind speed values ​​output by the output layer.

[0015] This invention also provides a deep learning-based quantitative forecasting device for thunderstorms and strong winds, comprising the following modules: The determination module is used to determine the target multi-source meteorological data for a target area based on multiple meteorological data sources; The forecasting module is used to input the target multi-source meteorological data into the thunderstorm and gale forecasting model to obtain multiple forecast wind speed values ​​output by the thunderstorm and gale forecasting model; the thunderstorm and gale forecasting model is trained based on sample multi-source meteorological data; The segmentation module is used to compare the multiple forecast wind speed values ​​with preset wind speed thresholds, and to segment the target area into extreme thunderstorm wind areas and ordinary thunderstorm wind areas based on the comparison results.

[0016] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the deep learning-based quantitative forecasting method for thunderstorms and gales as described above.

[0017] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the deep learning-based quantitative forecasting method for thunderstorms and gales as described above.

[0018] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the deep learning-based quantitative forecasting method for thunderstorms and strong winds as described above.

[0019] This invention provides a deep learning-based quantitative forecasting method and apparatus for thunderstorms and strong winds. It determines target multi-source meteorological data for a target area based on multiple meteorological data sources; inputs the target multi-source meteorological data into a thunderstorm and strong wind forecasting model to obtain multiple forecast wind speed values ​​output by the model; the thunderstorm and strong wind forecasting model is trained based on sample multi-source meteorological data; compares the multiple forecast wind speed values ​​with a preset wind speed threshold, and divides the target area into extreme thunderstorm and strong wind areas and ordinary thunderstorm and strong wind areas based on the comparison results. This invention's technical solution enables the learning of storm evolution patterns over long time series through the thunderstorm and strong wind forecasting model, improving the accuracy of thunderstorm and strong wind area prediction, and accurately distinguishing between ordinary thunderstorm and strong wind areas and extreme thunderstorm and strong wind areas based on the comparison of multiple forecast wind speed values ​​with preset wind speed thresholds. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the deep learning-based quantitative forecasting method for thunderstorms and strong winds provided by the present invention.

[0022] Figure 2 This is a schematic diagram of the structure of the thunderstorm and gale forecasting model provided by the present invention.

[0023] Figure 3 This is a schematic diagram of the structure of the deep learning-based quantitative forecasting device for thunderstorms and strong winds provided by the present invention.

[0024] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0026] To address the aforementioned problems in existing technologies, this invention provides a deep learning-based quantitative forecasting method for thunderstorms and strong winds. Figure 1 This is a flowchart illustrating the deep learning-based quantitative forecasting method for thunderstorms and strong winds provided by this invention. Figure 1As shown, the method includes the following steps 110 to 130.

[0027] Step 110: Determine the target multi-source meteorological data for the target area based on multiple meteorological data sources.

[0028] Specifically, multiple standardized data interfaces can be pre-configured to connect to multiple meteorological data sources, and the interface protocols of each standardized data interface support decoding of meteorological-specific data formats. This allows for the determination of target multi-source meteorological data for a target area based on multiple meteorological data sources.

[0029] In one embodiment, determining the target multi-source meteorological data for the target area based on multiple meteorological data sources includes: First multi-source meteorological data of the target area is obtained from the multiple meteorological data sources; Spatial coordinate matching and time step interpolation are performed on the first multi-source meteorological data to obtain the second multi-source meteorological data; The second multi-source meteorological data is normalized to obtain the third multi-source meteorological data; The third multi-source meteorological data is stacked and recombined to obtain the target multi-source meteorological data.

[0030] Specifically, first multi-source meteorological data for the target area can be obtained from multiple meteorological data sources. Then, second multi-source meteorological data can be obtained by performing spatial coordinate matching and time-step interpolation on the first multi-source meteorological data. Spatial coordinate matching involves matching the spatial coordinates of data from different data sources using a grid. Time-step interpolation is used to complete the interpolation of asynchronous observation data. Furthermore, the second multi-source meteorological data can be normalized to obtain third multi-source meteorological data; this normalization process can be a minimum-maximum normalization operation.

[0031] Furthermore, the third multi-source meteorological data can be stacked and recombined to construct tensors and obtain the target multi-source meteorological data.

[0032] In the above embodiments, spatial coordinate matching and time step interpolation ensure that multi-source meteorological data correspond to the same region in the spatial dimension and maintain a consistent step size in the time dimension, while normalization processing compresses the values ​​into a unified range, ensuring data consistency.

[0033] In one embodiment, stacking and recombining the third multi-source meteorological data to obtain the target multi-source meteorological data includes: The multi-source meteorological data is extracted using a sliding window sampler to obtain a multi-source meteorological sequence; The multi-source meteorological sequence is recombined in three dimensions—sample, time step, and spatial features—using a tensor stacker to obtain the target multi-source meteorological data.

[0034] Specifically, a sliding window sampler can be used to extract data from the third-party multi-source meteorological data to obtain a multi-source meteorological sequence. For example, a sliding window sampler can be used to extract data from the third-party multi-source meteorological data along the time axis with a step size of 1 hour and a window length of 24 hours to obtain a multi-source meteorological sequence.

[0035] Furthermore, a tensor stacker can be used to reconstruct multi-source meteorological sequences in three dimensions—samples, time steps, and spatial features—to obtain target multi-source meteorological data. It is easy to understand that the target multi-source meteorological data is tensor data.

[0036] In the above embodiments, the data accuracy was further improved by sliding window sampling and three-dimensional reconstruction, laying the foundation for subsequent prediction.

[0037] In one embodiment, the plurality of meteorological data sources include one or more of Doppler weather radar, regional automatic weather station network, wind profiler radar, and satellite data receiving station; the first multi-source meteorological data includes one or more of gust wind speed, precipitation over a preset period, historical weather, current weather, sea level pressure, vertical velocity, low cloud cover, middle cloud cover, middle cloud type, and low cloud type.

[0038] Specifically, the multiple meteorological data sources include one or more of Doppler weather radar, regional automatic weather station network, wind profiler radar and satellite data receiving station. The first multi-source meteorological data includes one or more of gust wind speed, precipitation for a preset period, historical weather, current weather, sea level pressure, vertical velocity, low cloud cover, middle cloud cover, middle cloud type and low cloud type.

[0039] In the above embodiments, the diversity of meteorological data sources and meteorological data further enhances information diversity, thereby improving forecast accuracy.

[0040] Step 120: Input the target multi-source meteorological data into the thunderstorm and gale forecasting model to obtain multiple forecast wind speed values ​​output by the thunderstorm and gale forecasting model; the thunderstorm and gale forecasting model is trained based on sample multi-source meteorological data.

[0041] Specifically, target multi-source meteorological data can be input into the thunderstorm and gale forecasting model to obtain multiple forecast wind speed values ​​output by the model. The thunderstorm and gale forecasting model is trained based on sample multi-source meteorological data. Among them, the multiple forecast wind speed values ​​can be forecast wind speed values ​​at multiple forecast times and at various observation stations.

[0042] Alternatively, the forecast wind speed field can be determined based on multiple forecast wind speed values.

[0043] In one embodiment, the thunderstorm and gale forecasting model includes an input layer, multiple bidirectional long short-term memory network layers, multiple random deactivation layers, and an output layer; the multiple bidirectional long short-term memory network layers and the multiple random deactivation layers are alternately connected in series; the bidirectional long short-term memory network layers are configured with input gates, forget gates, output gates, and peak retention unit state update branches.

[0044] Specifically, the thunderstorm and gale forecasting model includes an input layer, multiple bidirectional long short-term memory (LSTM) network layers, multiple random deactivation layers, and an output layer. The input layer receives target multi-source meteorological data. Multiple LSTM network layers and multiple random deactivation layers are alternately connected in series. For example, there are two LSTM network layers and two random deactivation layers. The input layer is followed by the first LSTM network layer, which is then followed by the first random deactivation layer, which is followed by the second LSTM network layer, which is followed by the second random deactivation layer, and finally the output layer. The LSTM network layers are configured with input gates, forget gates, output gates, and peak retention unit state update branches. The random deactivation layers are used to set the probability of randomly interrupting some neuron connections. The output layer adopts a temporally distributed wrapper layer structure, with multiple fully connected operation nodes arranged in parallel inside. Each node corresponds to a forecast time. Each node shares the deep composite features output by the feature extraction layer and independently maps the deep composite features to each future forecast time, realizing single forward operation and multi-step temporal output.

[0045] In the above embodiments, the alternating setting of bidirectional long short-term memory network layers and random deactivation layers avoids the error accumulation caused by traditional recursive prediction, and is computationally efficient, meeting the real-time requirements of meteorological operations for short-term forecasts. Furthermore, the design of multi-layer bidirectional long short-term memory network layers provides a strong modeling capability for the context of long-term series, and combined with the deep fusion of multi-source data, it significantly improves the accuracy of wind speed forecasts in the 0-24 hour short-term period, especially improving the forecasting capability for rapid changes in wind speed and turning points.

[0046] In one embodiment, the thunderstorm and gale forecasting model is trained based on the following steps: The sample multi-source meteorological data is input into the initial thunderstorm and gale forecast model. Based on the root mean square error and mean absolute error between the training forecast wind speed value and the actual wind speed value output by the initial thunderstorm and gale forecast model, as well as the wind force level score of the training forecast wind speed value, the initial thunderstorm and gale forecast model is iteratively trained to obtain the thunderstorm and gale forecast model. The network structure of the thunderstorm and gale forecast model is the same as that of the initial thunderstorm and gale forecast model.

[0047] Specifically, an initial thunderstorm and gale forecasting model can be pre-constructed, with the same network structure as the standard thunderstorm and gale forecasting model. Then, multi-source meteorological data can be input into the initial thunderstorm and gale forecasting model. Based on the root mean square error and mean absolute error between the training forecast wind speed values ​​and the actual wind speed values, as well as the wind force rating of the training forecast wind speed values, the initial thunderstorm and gale forecasting model is iteratively trained to obtain the final model. The wind force rating can be a TS score, which is calculated as: (Number of correct forecasts / (Number of correct forecasts + Number of false alarms + Number of missed forecasts)) used to evaluate the model's ability to capture thunderstorm and gale events of different intensities.

[0048] Optionally, the predicted wind speed field of typical historical extreme wind cases can be multiplied with the actual wind speed for hourly spatial sequence comparison to train and evaluate the thunderstorm wind forecasting model.

[0049] Optionally, the thunderstorm and gale forecast model can be fine-tuned during the implementation of the technical solution of this application.

[0050] In the above embodiments, the initial thunderstorm and gale forecast model is trained using root mean square error, mean absolute error, and wind force rating, which further enhances the model's adaptability to new regions and scenarios.

[0051] In one embodiment, the plurality of bidirectional long short-term memory network layers include a first bidirectional long short-term memory network layer, a second bidirectional long short-term memory network layer, a third bidirectional long short-term memory network layer, and a fourth bidirectional long short-term memory network layer; the random inactivation layer includes a first random inactivation layer, a second random inactivation layer, a third random inactivation layer, and a fourth random inactivation layer; The step involves inputting the target multi-source meteorological data into the thunderstorm and gale forecasting model to obtain multiple forecast wind speed values ​​output by the model, including: The target multi-source meteorological data is input into the input layer to obtain the first feature output by the input layer; The first feature is input into the first bidirectional long short-term memory network layer to obtain the second feature output by the first bidirectional long short-term memory network layer. The second feature is input into the first random deactivation layer to obtain the third feature output by the first random deactivation layer; The third feature is input into the second bidirectional long short-term memory network layer to obtain the fourth feature output by the second bidirectional long short-term memory network layer; The fourth feature is input into the second random deactivation layer to obtain the fifth feature output by the second random deactivation layer; The fifth feature is input into the third bidirectional long short-term memory network layer to obtain the sixth feature output by the third bidirectional long short-term memory network layer. The sixth feature is input into the third random deactivation layer to obtain the seventh feature output by the third random deactivation layer; The seventh feature is input into the fourth bidirectional long short-term memory network layer to obtain the eighth feature output by the fourth bidirectional long short-term memory network layer. The eighth feature is input into the fourth random deactivation layer to obtain the ninth feature output by the fourth random deactivation layer; The ninth feature is input into the output layer to obtain the plurality of forecast wind speed values ​​output by the output layer.

[0052] Specifically, the bidirectional long short-term memory network layer can consist of four layers, and the random deactivation layer can also consist of four layers. Target multi-source meteorological data can be input into the input layer to obtain the first feature output by the input layer. The input port dimension of the input layer matches the historical time series length and the number of fused features. The first feature can then be input into the first bidirectional long short-term memory (LSTM) network layer to obtain the second feature output by the first LSTM network layer. The second feature is then input into the first random deactivation layer to obtain the third feature output by the first random deactivation layer. The third feature is then input into the second LSTM network layer to obtain the fourth feature output by the second LSTM network layer. The fourth feature is then input into the second random deactivation layer to obtain the fifth feature output by the second random deactivation layer. The fifth feature is then input into the third LSTM network layer to obtain the sixth feature output by the third LSTM network layer. The sixth feature is then input into the third random deactivation layer to obtain the seventh feature output by the third random deactivation layer. The seventh feature is then input into the fourth LSTM network layer to obtain the eighth feature output by the fourth LSTM network layer. The eighth feature is then input into the fourth random deactivation layer to obtain the ninth feature output by the fourth random deactivation layer. The ninth feature is then input into the output layer to obtain multiple forecast wind speed values ​​output by the output layer. The alternation between the bidirectional LSTM network layer and the random deactivation layer progressively abstracts deep composite features containing information on long-term climate background and short-term convective abrupt changes. In this process, the peak-preserving unit state update branch enhances and preserves the gradient flow corresponding to extreme thunderstorm and strong wind data in the multi-source meteorological data of the sample, so as to avoid the dilution of extreme features by general samples in multi-layer transmission.

[0053] For example, Figure 2 This is a schematic diagram of the structure of the thunderstorm and gale forecasting model provided by the present invention, as shown below. Figure 2As shown, the thunderstorm and gale forecasting model, in structural order, includes an input layer, a first bidirectional long short-term memory network layer, a first random inactivation layer, a second bidirectional long short-term memory network layer, a second random inactivation layer, a third bidirectional long short-term memory network layer, a third random inactivation layer, a fourth bidirectional long short-term memory network layer, and a fourth random inactivation layer. Following the fourth random inactivation layer are a first fully connected layer, a second fully connected layer, a third fully connected layer, and an output layer (the output layer is located in...). Figure 2 (Not shown in the image). Figure 2 The first "None" after the input / output indicates the batch size, meaning the model supports any number of batches; the second indicates the timestep / sequence length, which is set to 4 in this example; the third indicates the feature dimension of a single timestep, including 21, 64, 32, 16, and 1. For example, 64 indicates that this bidirectional long short-term memory network layer has 64 hidden neurons / hidden units, and each timestep is encoded into a 64-dimensional feature vector.

[0054] In the above embodiments, the ability to extract deep composite features of the model is further enhanced by alternating between four layers of bidirectional long short-term memory network and random deactivation layers.

[0055] Step 130: Compare the multiple forecast wind speed values ​​with the preset wind speed threshold, and divide the target area into extreme thunderstorm wind area and ordinary thunderstorm wind area according to the comparison results.

[0056] Specifically, multiple forecast wind speed values ​​can be compared with preset wind speed thresholds, and the target area can then be divided into extreme thunderstorm wind areas and ordinary thunderstorm wind areas based on the comparison results. It should be noted that each forecast wind speed value can correspond to a different preset wind speed threshold, or it can correspond to the same preset wind speed threshold. This preset wind speed threshold can be set in advance as needed. For example, if the preset wind speed threshold is set to 25 m / s, the area corresponding to a forecast wind speed greater than 25 m / s is an extreme thunderstorm wind area, and the area corresponding to a forecast wind speed less than or equal to 25 m / s is an ordinary thunderstorm wind area.

[0057] This invention provides a deep learning-based quantitative forecasting method for thunderstorms and strong winds. It determines target multi-source meteorological data for a target area based on multiple meteorological data sources; inputs the target multi-source meteorological data into a thunderstorm and strong wind forecasting model to obtain multiple forecast wind speed values ​​output by the model; the thunderstorm and strong wind forecasting model is trained based on sample multi-source meteorological data; compares the multiple forecast wind speed values ​​with a preset wind speed threshold, and divides the target area into extreme thunderstorm and strong wind areas and ordinary thunderstorm and strong wind areas based on the comparison results. This invention's technical solution enables the learning of storm evolution patterns over long time series through the thunderstorm and strong wind forecasting model, improving the accuracy of thunderstorm and strong wind area prediction, and accurately distinguishing between ordinary thunderstorm and strong wind areas and extreme thunderstorm and strong wind areas based on the comparison of multiple forecast wind speed values ​​with preset wind speed thresholds.

[0058] In one embodiment, the method further includes: Multiple wind speed increments are determined based on the multiple predicted wind speed values; For each of the aforementioned wind speed increments, if the wind speed increment exceeds a preset increment threshold, an early warning signal is generated.

[0059] Specifically, the wind speed increment of an observation station in a corresponding time period can be determined based on the forecast wind speed values ​​at the previous forecast time and the forecast wind speed values ​​at the next forecast time for the same observation station.

[0060] For each wind speed increment, an early warning signal can be generated if the wind speed increment exceeds a preset increment threshold. The preset increment threshold could be, for example, 10 meters per second per hour.

[0061] Existing technologies lack a corresponding anomaly warning mechanism, failing to promptly alert staff to verify forecasts when sudden changes occur, potentially leading to missed or misjudged warnings. In the above embodiment, abnormal wind speed changes are identified through wind speed increments, enabling accurate warnings to be sent to staff for verification.

[0062] In one embodiment, the method further includes: The display device displays one or more of the following: the multiple forecast wind speed values, the extreme thunderstorm wind area, the ordinary thunderstorm wind area, and the warning signal.

[0063] Specifically, the display device may include a display screen and an operation panel. The display device can show one or more of the following: forecast wind speed values, extreme thunderstorm wind areas, ordinary thunderstorm wind areas, and warning signals. The extreme thunderstorm wind areas and ordinary thunderstorm wind areas can be marked with different colors.

[0064] In the above embodiments, the display device displays one or more of the following: forecast wind speed values, extreme thunderstorm wind areas, ordinary thunderstorm wind areas, and warning signals, enabling staff to make more accurate judgments about thunderstorm winds in the target area.

[0065] The following is a complete example to illustrate this application: The system is configured with 8 standard serial interfaces and 2 gigabit Ethernet interfaces, connecting to the local Doppler weather radar, regional automatic weather station network, wind profiler radar, and satellite data receiving station, respectively. The spatiotemporal alignment processor is an embedded field-programmable gate array (FPGA) chip with a built-in bilinear interpolation algorithm core; normalization processing is performed on the parallel computing array of this FPGA. The sliding window sampler is configured with a window length register of 24 and a step size register of 1. The tensor stacker stacks 10 meteorological elements (gust speed, 6-hour precipitation, past weather, present weather, sea level pressure, vertical velocity, low cloud cover, medium cloud cover, medium cloud type, and low cloud type) from 259 observation stations into a three-dimensional tensor (i.e., target multi-source meteorological data) with dimensions "batch size, 24, 2590", where the batch size register is set to 64. The input dimension of the input layer in the thunderstorm and gale forecast model is (64, 24, 2590). It also includes four bidirectional long short-term memory (LSTM) network layers, each with 50 hidden units. The forward and reverse memory sublayers are each composed of independent LSM network kernels, with merged nodes performing vector concatenation operations. Random deactivation layers are placed between adjacent bidirectional LSM network layers. The output layer uses a temporally distributed wrapper layer, internally containing 24 fully connected computation nodes arranged in parallel. Each node corresponds to a future forecast time, with an output dimension of 259 (number of stations), resulting in a final output tensor dimension of (64, 24, 259). The pre-trained weights for loading and fine-tuning units are stored based on weights obtained from training on multi-source meteorological data from May to September 2012-2021. When deploying fine-tuning units in new regions, 32 local samples are used as fine-tuning batches. The initial learning rate is set to 0.001, decaying to 0.8 every 20 rounds. The early stopping strategy is set to terminate when the validation set loss does not decrease for 10 consecutive rounds. Wind speed ratings are categorized into levels 6–8, 8–10, 10–12, and 12 and above, with TS scores calculated based on three timeframes: the first 6 hours, 6–12 hours, and 12–24 hours. The warning signal output interface utilizes an RJ45 network port, supporting the transmission control protocol / internet protocol to push JSON-formatted warning signals to the meteorological operational platform. The display device can be configured with a screen and a capacitive touch control panel, supporting multi-window simultaneous display.

[0066] The following describes the deep learning-based quantitative forecasting device for thunderstorms and gales provided by the present invention. The deep learning-based quantitative forecasting device for thunderstorms and gales described below can be referred to in correspondence with the deep learning-based quantitative forecasting method for thunderstorms and gales described above.

[0067] Figure 3 This is a schematic diagram of the deep learning-based quantitative forecasting device for thunderstorms and strong winds provided by the present invention. Figure 3 As shown, the deep learning-based quantitative forecasting device for thunderstorms and strong winds 300 includes the following modules: Module 310 is used to determine the target multi-source meteorological data for the target area based on multiple meteorological data sources; The forecast module 320 is used to input the target multi-source meteorological data into the thunderstorm and gale forecast model to obtain multiple forecast wind speed values ​​output by the thunderstorm and gale forecast model; the thunderstorm and gale forecast model is trained based on sample multi-source meteorological data; The segmentation module 330 is used to compare the multiple forecast wind speed values ​​with preset wind speed thresholds, and to segment the target area into extreme thunderstorm wind areas and ordinary thunderstorm wind areas based on the comparison results.

[0068] In one embodiment, the deep learning-based quantitative forecasting device for thunderstorms and strong winds 300 further includes an early warning module, which is specifically used for: Multiple wind speed increments are determined based on the multiple predicted wind speed values; For each of the aforementioned wind speed increments, if the wind speed increment exceeds a preset increment threshold, an early warning signal is generated.

[0069] In one embodiment, the deep learning-based thunderstorm and gale quantitative forecasting device 300 further includes a display module, which is specifically used for: The display device displays one or more of the following: the multiple forecast wind speed values, the extreme thunderstorm wind area, the ordinary thunderstorm wind area, and the warning signal.

[0070] In one embodiment, the determining module 310 is specifically used for: First multi-source meteorological data of the target area is obtained from the multiple meteorological data sources; Spatial coordinate matching and time step interpolation are performed on the first multi-source meteorological data to obtain the second multi-source meteorological data; The second multi-source meteorological data is normalized to obtain the third multi-source meteorological data; The third multi-source meteorological data is stacked and recombined to obtain the target multi-source meteorological data.

[0071] In one embodiment, the determining module 310 is further configured to: The multi-source meteorological data is extracted using a sliding window sampler to obtain a multi-source meteorological sequence; The multi-source meteorological sequence is recombined in three dimensions—sample, time step, and spatial features—using a tensor stacker to obtain the target multi-source meteorological data.

[0072] In one embodiment, the plurality of meteorological data sources include one or more of Doppler weather radar, regional automatic weather station network, wind profiler radar, and satellite data receiving station; the first multi-source meteorological data includes one or more of gust wind speed, precipitation over a preset period, historical weather, current weather, sea level pressure, vertical velocity, low cloud cover, middle cloud cover, middle cloud type, and low cloud type.

[0073] In one embodiment, the thunderstorm and gale forecasting model includes an input layer, multiple bidirectional long short-term memory network layers, multiple random deactivation layers, and an output layer; the multiple bidirectional long short-term memory network layers and the multiple random deactivation layers are alternately connected in series; the bidirectional long short-term memory network layers are configured with input gates, forget gates, output gates, and peak retention unit state update branches.

[0074] In one embodiment, the deep learning-based thunderstorm and gale quantitative forecasting device 300 further includes a training module, which is specifically used to train the thunderstorm and gale forecasting model based on the following steps: The sample multi-source meteorological data is input into the initial thunderstorm and gale forecast model. Based on the root mean square error and mean absolute error between the training forecast wind speed value and the actual wind speed value output by the initial thunderstorm and gale forecast model, as well as the wind force level score of the training forecast wind speed value, the initial thunderstorm and gale forecast model is iteratively trained to obtain the thunderstorm and gale forecast model. The network structure of the thunderstorm and gale forecast model is the same as that of the initial thunderstorm and gale forecast model.

[0075] In one embodiment, the plurality of bidirectional long short-term memory (LSTM) network layers include a first bidirectional LSM network layer, a second bidirectional LSM network layer, a third bidirectional LSM network layer, and a fourth bidirectional LSM network layer; the random inactivation layer includes a first random inactivation layer, a second random inactivation layer, a third random inactivation layer, and a fourth random inactivation layer; the prediction module 320 is specifically used for: The target multi-source meteorological data is input into the input layer to obtain the first feature output by the input layer; The first feature is input into the first bidirectional long short-term memory network layer to obtain the second feature output by the first bidirectional long short-term memory network layer. The second feature is input into the first random deactivation layer to obtain the third feature output by the first random deactivation layer; The third feature is input into the second bidirectional long short-term memory network layer to obtain the fourth feature output by the second bidirectional long short-term memory network layer; The fourth feature is input into the second random deactivation layer to obtain the fifth feature output by the second random deactivation layer; The fifth feature is input into the third bidirectional long short-term memory network layer to obtain the sixth feature output by the third bidirectional long short-term memory network layer. The sixth feature is input into the third random deactivation layer to obtain the seventh feature output by the third random deactivation layer; The seventh feature is input into the fourth bidirectional long short-term memory network layer to obtain the eighth feature output by the fourth bidirectional long short-term memory network layer. The eighth feature is input into the fourth random deactivation layer to obtain the ninth feature output by the fourth random deactivation layer; The ninth feature is input into the output layer to obtain the plurality of forecast wind speed values ​​output by the output layer.

[0076] This invention provides a deep learning-based quantitative forecasting device for thunderstorms and strong winds. It determines target multi-source meteorological data for a target area based on multiple meteorological data sources; inputs the target multi-source meteorological data into a thunderstorm and strong wind forecasting model to obtain multiple forecast wind speed values ​​output by the model; the thunderstorm and strong wind forecasting model is trained based on sample multi-source meteorological data; compares multiple forecast wind speed values ​​with preset wind speed thresholds, and divides the target area into extreme thunderstorm and strong wind areas and ordinary thunderstorm and strong wind areas based on the comparison results. This invention's technical solution, through the thunderstorm and strong wind forecasting model, can learn the evolution patterns of storms over long time series, improving the accuracy of thunderstorm and strong wind area prediction, and can accurately distinguish between ordinary thunderstorm and strong wind areas and extreme thunderstorm and strong wind areas based on the comparison of multiple forecast wind speed values ​​with preset wind speed thresholds.

[0077] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 4 As shown, the electronic device may include: a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other via the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a deep learning-based quantitative forecasting method for thunderstorms and strong winds, which includes: Target multi-source meteorological data for the target area is determined based on multiple meteorological data sources; The target multi-source meteorological data is input into the thunderstorm and gale forecasting model to obtain multiple forecast wind speed values ​​output by the thunderstorm and gale forecasting model; the thunderstorm and gale forecasting model is trained based on sample multi-source meteorological data; By comparing the multiple forecast wind speed values ​​with preset wind speed thresholds, the target area is divided into extreme thunderstorm wind areas and ordinary thunderstorm wind areas based on the comparison results.

[0078] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0079] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the deep learning-based quantitative forecasting method for thunderstorms and strong winds provided by the above methods. This method includes: Target multi-source meteorological data for the target area is determined based on multiple meteorological data sources; The target multi-source meteorological data is input into the thunderstorm and gale forecasting model to obtain multiple forecast wind speed values ​​output by the thunderstorm and gale forecasting model; the thunderstorm and gale forecasting model is trained based on sample multi-source meteorological data; By comparing the multiple forecast wind speed values ​​with preset wind speed thresholds, the target area is divided into extreme thunderstorm wind areas and ordinary thunderstorm wind areas based on the comparison results.

[0080] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the deep learning-based quantitative forecasting method for thunderstorms and strong winds provided by the methods described above, the method comprising: Target multi-source meteorological data for the target area is determined based on multiple meteorological data sources; The target multi-source meteorological data is input into the thunderstorm and gale forecasting model to obtain multiple forecast wind speed values ​​output by the thunderstorm and gale forecasting model; the thunderstorm and gale forecasting model is trained based on sample multi-source meteorological data; By comparing the multiple forecast wind speed values ​​with preset wind speed thresholds, the target area is divided into extreme thunderstorm wind areas and ordinary thunderstorm wind areas based on the comparison results.

[0081] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0082] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for quantitative forecasting of thunderstorms and strong winds based on deep learning, characterized in that, include: Target multi-source meteorological data for the target area is determined based on multiple meteorological data sources; The target multi-source meteorological data is input into the thunderstorm and gale forecasting model to obtain multiple forecast wind speed values ​​output by the thunderstorm and gale forecasting model; the thunderstorm and gale forecasting model is trained based on sample multi-source meteorological data; By comparing the multiple forecast wind speed values ​​with preset wind speed thresholds, the target area is divided into extreme thunderstorm wind areas and ordinary thunderstorm wind areas based on the comparison results.

2. The method for quantitative forecasting of thunderstorms and strong winds based on deep learning according to claim 1, characterized in that, The method further includes: Multiple wind speed increments are determined based on the multiple predicted wind speed values; For each of the aforementioned wind speed increments, if the wind speed increment exceeds a preset increment threshold, an early warning signal is generated.

3. The method for quantitative forecasting of thunderstorms and strong winds based on deep learning according to claim 2, characterized in that, The method further includes: The display device displays one or more of the following: the multiple forecast wind speed values, the extreme thunderstorm wind area, the ordinary thunderstorm wind area, and the warning signal.

4. The method for quantitative forecasting of thunderstorms and strong winds based on deep learning according to claim 1, characterized in that, The target multi-source meteorological data for determining the target area based on multiple meteorological data sources includes: First multi-source meteorological data of the target area is obtained from the multiple meteorological data sources; Spatial coordinate matching and time step interpolation are performed on the first multi-source meteorological data to obtain the second multi-source meteorological data; The second multi-source meteorological data is normalized to obtain the third multi-source meteorological data; The third multi-source meteorological data is stacked and recombined to obtain the target multi-source meteorological data.

5. The method for quantitative forecasting of thunderstorms and strong winds based on deep learning according to claim 4, characterized in that, The process of stacking and recombining the third multi-source meteorological data to obtain the target multi-source meteorological data includes: The multi-source meteorological data is extracted using a sliding window sampler to obtain a multi-source meteorological sequence; The multi-source meteorological sequence is recombined in three dimensions—sample, time step, and spatial features—using a tensor stacker to obtain the target multi-source meteorological data.

6. The method for quantitative forecasting of thunderstorms and strong winds based on deep learning according to claim 4, characterized in that, The multiple meteorological data sources include one or more of Doppler weather radar, regional automatic weather station network, wind profiler radar and satellite data receiving station; the first multi-source meteorological data includes one or more of gust wind speed, precipitation during a preset period, historical weather, current weather, sea level pressure, vertical velocity, low cloud cover, middle cloud cover, middle cloud type and low cloud type.

7. The method for quantitative forecasting of thunderstorms and strong winds based on deep learning according to any one of claims 1 to 6, characterized in that, The thunderstorm and gale forecasting model includes an input layer, multiple bidirectional long short-term memory network layers, multiple random deactivation layers, and an output layer; the multiple bidirectional long short-term memory network layers and the multiple random deactivation layers are alternately connected in series; the bidirectional long short-term memory network layers are configured with input gates, forget gates, output gates, and peak retention unit state update branches.

8. The method for quantitative forecasting of thunderstorms and strong winds based on deep learning according to claim 6, characterized in that, The thunderstorm and gale forecasting model was trained based on the following steps: The sample multi-source meteorological data is input into the initial thunderstorm and gale forecast model. Based on the root mean square error and mean absolute error between the training forecast wind speed value and the actual wind speed value output by the initial thunderstorm and gale forecast model, as well as the wind force level score of the training forecast wind speed value, the initial thunderstorm and gale forecast model is iteratively trained to obtain the thunderstorm and gale forecast model. The network structure of the thunderstorm and gale forecast model is the same as that of the initial thunderstorm and gale forecast model.

9. The method for quantitative forecasting of thunderstorms and strong winds based on deep learning according to claim 7, characterized in that, The plurality of bidirectional long short-term memory network layers include a first bidirectional long short-term memory network layer, a second bidirectional long short-term memory network layer, a third bidirectional long short-term memory network layer, and a fourth bidirectional long short-term memory network layer; the random inactivation layer includes a first random inactivation layer, a second random inactivation layer, a third random inactivation layer, and a fourth random inactivation layer; The step involves inputting the target multi-source meteorological data into the thunderstorm and gale forecasting model to obtain multiple forecast wind speed values ​​output by the model, including: The target multi-source meteorological data is input into the input layer to obtain the first feature output by the input layer; The first feature is input into the first bidirectional long short-term memory network layer to obtain the second feature output by the first bidirectional long short-term memory network layer. The second feature is input into the first random deactivation layer to obtain the third feature output by the first random deactivation layer; The third feature is input into the second bidirectional long short-term memory network layer to obtain the fourth feature output by the second bidirectional long short-term memory network layer; The fourth feature is input into the second random deactivation layer to obtain the fifth feature output by the second random deactivation layer; The fifth feature is input into the third bidirectional long short-term memory network layer to obtain the sixth feature output by the third bidirectional long short-term memory network layer. The sixth feature is input into the third random deactivation layer to obtain the seventh feature output by the third random deactivation layer; The seventh feature is input into the fourth bidirectional long short-term memory network layer to obtain the eighth feature output by the fourth bidirectional long short-term memory network layer. The eighth feature is input into the fourth random deactivation layer to obtain the ninth feature output by the fourth random deactivation layer; The ninth feature is input into the output layer to obtain the plurality of forecast wind speed values ​​output by the output layer.

10. A deep learning-based quantitative forecasting device for thunderstorms and strong winds, characterized in that, include: The determination module is used to determine the target multi-source meteorological data for a target area based on multiple meteorological data sources; The forecasting module is used to input the target multi-source meteorological data into the thunderstorm and gale forecasting model to obtain multiple forecast wind speed values ​​output by the thunderstorm and gale forecasting model. The thunderstorm and gale forecasting model was trained based on multi-source meteorological data. The segmentation module is used to compare the multiple forecast wind speed values ​​with preset wind speed thresholds, and to segment the target area into extreme thunderstorm wind areas and ordinary thunderstorm wind areas based on the comparison results.

Images