Sea fog visibility forecast method and system based on deep convolutional neural network model

Abstract

The present invention discloses a sea fog visibility forecasting method and system based on a deep convolutional neural network model, which relates to the field of sea fog visibility forecasting. The method divides a day into preset time periods to obtain forecast time periods, establishes a preset number of deep convolutional neural network models in each forecast time period, and trains the deep convolutional neural network model through a sample data set to obtain a sea fog visibility forecast model corresponding to each forecast time effectiveness in each forecast time period. The method constructs a sea fog visibility forecast model corresponding to each forecast time effectiveness in each forecast time period of each day, so that when actually making a forecast, the corresponding sea fog visibility forecast model can be selected for prediction according to the forecast time period corresponding to the current time and the required forecast time effectiveness. The method effectively improves the accuracy of the model prediction and solves the problem that the current conventional forecast products have low accuracy and cannot well meet the meteorological support service needs of ship scheduling and berth operations in ports.

Description

Technical Field

[0001] The present invention relates to the field of sea fog visibility forecasting, and in particular to a sea fog visibility forecasting method and system based on a deep convolutional neural network model. Background Art

[0002] Currently, most ports serve as cargo transportation hubs, placing high demands on weather forecast services. However, sufficient observation data is currently lacking at sea, and the complex terrain near ports makes it prone to both systemic, large-scale fog and small, clustered fog along shipping lanes. Effective research and forecasting methods for these conditions are still relatively limited. Conventional forecast products are primarily qualitative, with low accuracy and temporal and spatial resolution. These products cannot adequately meet the meteorological support needs for ship scheduling and berthing operations in ports, severely impacting port operational efficiency and navigation safety.

[0003] Currently, artificial intelligence technology is booming, especially machine learning technology represented by deep learning technology, which has been widely studied and applied in many fields such as image recognition, speech recognition, and autonomous driving. Deep learning technology has also begun to be applied in weather forecasting. For example, the forecast of short-term rainfall is transformed into a video prediction problem. Compared with traditional methods, some progress has also been made in the automatic identification of weather systems, satellite cloud image recognition, and cloud cover calculation. In response to the problem that the current conventional forecast products have weak sea fog forecasting capabilities (low forecast precision), the present invention proposes a sea fog visibility forecasting method and system based on a deep convolutional neural network model based on deep learning technology. Summary of the Invention

[0004] In order to improve the forecast accuracy of sea fog visibility and to address the problem that conventional forecast products currently have low accuracy, temporal and spatial resolution, and cannot well meet the meteorological support service requirements for ship scheduling and berth operations in ports, the present invention proposes a sea fog visibility forecast method based on a deep convolutional neural network model, comprising the following steps:

[0005] S1: Obtain observation data from each monitoring station within a preset time period, convert it into grid point data within the port area through interpolation, and include it in the sample dataset; obtain forecast data corresponding to each forecast time within the preset time period through forecast products, convert it into grid point data through downscaling and spatiotemporal interpolation, and include it in the sample dataset;

[0006] S2: Divide a day into preset time periods to obtain each prediction period, and establish a preset number of deep convolutional neural network models in each prediction period;

[0007] S3: training a deep convolutional neural network model using a sample data set to obtain a sea fog visibility forecast model corresponding to each forecast time validity in each forecast period; the number of sea fog visibility forecast models in each forecast period is equal to a preset number; the forecast time validity represents the time difference between the model start time point and the forecast time, and the forecast time is the time validity cutoff point of the forecast time validity;

[0008] S4: According to the current forecast period and the required forecast time, select the corresponding sea fog visibility forecast model, and output the visibility forecast value through the sea fog visibility forecast model.

[0009] Furthermore, before step S03, the following steps are further included:

[0010] The sample data set is filtered using preset filtering rules, and the data that meets the preset filtering rules is normalized to obtain a target data set; in step S03, a deep convolutional neural network model is trained using the target data set.

[0011] Furthermore, the deep convolutional neural network model is trained using the target dataset, specifically including the following steps:

[0012] S31: Based on the start and end time points of each forecast period and the forecast time limits corresponding to the forecast period, the grid point data corresponding to the observation data in the target data set are combined with the grid point data corresponding to the forecast data to obtain a forecast factor combination, and the combination is divided into target data subsets corresponding to the forecast time limits under each forecast period;

[0013] S32: The forecast factor combination in the target data subset is used as the input of the model, and the visibility data in the observation data of the target data subset corresponding to the forecast time is used as the output of the model. The deep convolutional neural network model corresponding to the forecast time is trained to obtain the sea fog visibility forecast model corresponding to each forecast time in each forecast period.

[0014] Furthermore, the observation data of each detection site includes: actual visibility, relative humidity, and temperature corresponding to each hour; the forecast data corresponding to each forecast time includes: relative humidity, temperature, and predicted visibility corresponding to the forecast time.

[0015] Furthermore, the preset screening rules are:

[0016] The actual visibility in the observed data is less than the preset value and the relative humidity is greater than the preset percentage, or the relative humidity in the forecast data is greater than the preset percentage.

[0017] Furthermore, in step S32, the visibility data in the observation data is:

[0018] Actual visibility or actual visibility difference;

[0019] The actual visibility difference is the difference between the actual visibility corresponding to the forecast time and the actual visibility at the current time; the current time represents the starting time point of the model.

[0020] Furthermore, the formula for normalizing the data that meets the preset screening rules is:

[0021]

[0022] Where x represents the actual visibility; x max Represents the maximum visibility value in the filtered sample data set; x min represents the minimum visibility value in the filtered sample data set; y represents the dimensionless actual visibility obtained after normalization of the actual visibility.

[0023] Furthermore, the actual visibility difference is normalized before being input into the model, and the formula is:

[0024]

[0025] Where y represents the actual visibility difference; y max The maximum absolute value of the difference between the actual visibility at each forecast time in the filtered sample data set and the actual visibility at the current time; y out It represents the dimensionless actual visibility difference obtained after normalization of the actual visibility difference.

[0026] Furthermore, the observation data of each detection site also includes: port area control records.

[0027] The present invention also proposes a sea fog visibility forecasting system based on a deep convolutional neural network model, comprising:

[0028] The dataset acquisition module is used to obtain the observation data of each detection station and the port area control records within the preset time period, convert them into grid point data within the port area through interpolation, and include them in the sample dataset; obtain the forecast data corresponding to each forecast time within the preset time period through forecast products, convert them into grid point data through downscaling spatiotemporal interpolation, and include them in the sample dataset;

[0029] A model building module is used to divide a day into preset time periods to obtain each prediction period, and to build a preset number of deep convolutional neural network models in each prediction period;

[0030] A training module is configured to train a deep convolutional neural network model using a sample data set to obtain a sea fog visibility forecast model corresponding to each forecast time validity period in each forecast period; the number of sea fog visibility forecast models in each forecast period is equal to a preset number; the forecast time validity period represents the time difference between the model start time point and the forecast time number, and the forecast time number is the time validity cutoff point of the forecast time validity period;

[0031] The prediction module is used to select the corresponding sea fog visibility forecast model according to the current forecast period and the required forecast timeliness, and output the visibility forecast value through the sea fog visibility forecast model.

[0032] Compared with the prior art, the present invention has at least the following beneficial effects:

[0033] (1) The present invention divides a day into preset time periods to obtain a forecast period, establishes a preset number of deep convolutional neural network models in each forecast period, and trains the deep convolutional neural network model through a sample data set to obtain a sea fog visibility forecast model corresponding to each forecast time in each forecast period. It constructs a sea fog visibility forecast model corresponding to each forecast time (each forecast time is not equal) in each forecast period of each day, so that when actually making a forecast, the corresponding sea fog visibility forecast model can be selected according to the forecast period corresponding to the current time and the required forecast time for prediction. It effectively improves the accuracy of the model prediction and solves the problem that the current conventional forecast products have low accuracy and cannot well meet the meteorological support service needs of ship scheduling and berth operations in ports;

[0034] (2) The present invention obtains observation data from each detection site within a preset time period and forecast data corresponding to each forecast time within the preset time period, and converts the observation data and forecast data into grid point data to obtain a sample data set for model training. This solves the problem of low temporal and spatial resolution of current conventional forecast products.

[0035] (3) The present invention filters the sample data set by preset screening rules, and normalizes the data that meets the preset screening rules to obtain a target data set, and trains a deep convolutional neural network model based on the target data set. The preset screening rules are used to increase the proportion of foggy samples in the sample data set, and the accuracy of the data in the sample data set is improved by normalizing the data that meets the preset screening rules, thereby further improving the forecast accuracy of the sea fog visibility forecast model.

[0036] (4) The present invention obtains a forecast factor combination according to the start and end time points of each forecast period and each forecast time validity period corresponding to the forecast period, and divides the forecast factor combination into a target data subset corresponding to each forecast time validity period under each forecast period, takes the forecast factor combination in the target data subset as the input of the model, and uses the visibility data in the observation data of the forecast time corresponding to the target data subset as the output of the model to train a deep convolutional neural network model corresponding to the forecast time validity period. It constructs a target data subset corresponding to a forecast time validity period in a forecast period to train a deep convolutional neural network model corresponding to the forecast time validity period (or forecast time), thereby obtaining a sea fog visibility forecast model corresponding to each forecast time validity period under each forecast period, which refines the sea fog visibility forecast model and greatly improves the prediction accuracy of the sea fog visibility in each forecast period;

[0037] (5) The present invention increases the proportion of fog samples in the sample data set by incorporating port control records into the sample data set;

[0038] (6) In the present invention, since the visibility data in the observation data is the actual visibility or the actual visibility difference, the model can be trained to output the sea fog visibility forecast model of the actual visibility or the actual visibility difference, which improves the flexibility of the model prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of a sea fog visibility forecasting method based on a deep convolutional neural network model;

[0040] Figure 2 This is a module diagram of a sea fog visibility forecast system based on a deep convolutional neural network model;

[0041] Figure 3 This is the structure diagram of the sea fog visibility forecast model;

[0042] Figure 4 This is the structure diagram of the residual convolution module. DETAILED DESCRIPTION

[0043] The following are specific embodiments of the present invention and the accompanying drawings to further describe the technical solutions of the present invention, but the present invention is not limited to these embodiments.

[0044] Example 1

[0045] In order to improve the forecast accuracy of sea fog visibility, and to solve the problem that the current conventional forecast products have low accuracy, time and space resolution, and cannot well meet the meteorological support service needs of ship scheduling and berth operations in ports, such as Figure 1 As shown, the present invention proposes a sea fog visibility forecasting method based on a deep convolutional neural network model, comprising the steps of:

[0046] S1: Obtain observation data from each monitoring station within a preset time period, convert it into grid point data within the port area through interpolation (specifically, downscaling and spatiotemporal interpolation of grid data), and include it in the sample dataset; obtain forecast data corresponding to each forecast time within the preset time period through forecast products, convert it into grid point data through downscaling and spatiotemporal interpolation, and include it in the sample dataset;

[0047] In this embodiment, the grid point data corresponding to the observation data and the grid point data corresponding to the forecast data have the same temporal and spatial resolution.

[0048] The present invention obtains the observation data of each detection site within a preset time period and the forecast data corresponding to each forecast time within the preset time period, and converts the observation data and forecast data into grid point data to obtain a sample data set for model training. It solves the problem that the current conventional forecast products have low time and space resolution.

[0049] The observation data of each detection site includes: actual visibility, relative humidity, and temperature corresponding to each hour; the forecast data corresponding to each forecast time includes: relative humidity, temperature, and predicted visibility corresponding to the forecast time.

[0050] Before step S03, the following steps are also included:

[0051] The sample data set is filtered using preset filtering rules, and the data that meets the preset filtering rules is normalized to obtain a target data set; in step S03, a deep convolutional neural network model is trained using the target data set.

[0052] The preset screening rules are:

[0053] The actual visibility in the observation data is less than the preset value (1 km) and the relative humidity is greater than the preset percentage (80%), or the relative humidity in the forecast data is greater than the preset percentage (90%).

[0054] It should be noted that since the occurrence of sea fog is a low-probability event and the frequency of sea fog occurrence has a certain regularity on the hourly scale of each day, setting preset filtering rules to filter the sample data set can increase the proportion of foggy samples in the sample data set.

[0055] The formula for normalizing the data that meets the preset screening rules is:

[0056]

[0057] Where x represents the actual visibility; x max Represents the maximum visibility value in the filtered sample data set; x minrepresents the minimum visibility value in the filtered sample data set; y represents the dimensionless actual visibility obtained after normalization of the actual visibility.

[0058] The present invention filters the sample data set through preset screening rules, normalizes the data that meets the preset screening rules to obtain a target data set, and trains a deep convolutional neural network model through the target data set. The preset screening rules are used to increase the proportion of foggy samples in the sample data set, and the accuracy of the data in the sample data set is improved by normalizing the data that meets the preset screening rules, thereby further improving the forecast accuracy of the sea fog visibility forecast model.

[0059] S2: Divide a day into preset time periods to obtain each prediction period, and establish a preset number of deep convolutional neural network models in each prediction period;

[0060] S3: training a deep convolutional neural network model using a sample data set to obtain a sea fog visibility forecast model corresponding to each forecast time validity in each forecast period; the number of sea fog visibility forecast models in each forecast period is equal to a preset number; the forecast time validity represents the time difference between the model start time point and the forecast time, and the forecast time is the time validity cutoff point of the forecast time validity;

[0061] In this embodiment, a day is divided into 12 equal periods to obtain 12 forecast periods. Each forecast period includes 8 forecast time periods: 3, 6, 9, 12, 15, 18, 21, and 24 hours. Therefore, a total of 12*8=96 sea fog visibility forecast models are established. This refines the sea fog visibility forecast model and greatly improves the prediction accuracy of sea fog visibility within each forecast period.

[0062] Training a deep convolutional neural network model using the target dataset involves the following steps:

[0063] S31: Based on the start and end time points of each forecast period and the forecast time limits corresponding to the forecast period, the grid point data corresponding to the observation data in the target data set are combined with the grid point data corresponding to the forecast data to obtain a forecast factor combination, and the combination is divided into target data subsets corresponding to the forecast time limits under each forecast period;

[0064] S32: The forecast factor combination in the target data subset is used as the input of the model, and the visibility data in the observation data of the target data subset corresponding to the forecast time is used as the output of the model. The deep convolutional neural network model corresponding to the forecast time is trained to obtain the sea fog visibility forecast model corresponding to each forecast time in each forecast period.

[0065] Specifically, based on the start time of each forecast period, the forecast time limits corresponding to each forecast period, and the physical laws and correlation analysis of sea fog formation, the grid point data corresponding to the observation data in the target dataset are combined with the grid point data corresponding to the forecast data to construct a forecast factor combination, which is then assigned to the target data subset corresponding to each forecast time limit in each forecast period. In addition, the model can construct two outputs: one is the output of actual visibility or the difference between actual visibility and actual visibility. The following table illustrates the forecast factor combinations corresponding to each forecast time limit in a preset period (it should be noted in the table that: Current Time represents the model start time point; Visibility 3 hours ago represents the visibility 3 hours ago relative to the current time point; Visibility 6 hours ago represents the visibility 6 hours ago relative to the current time point; Forecast Time represents the time limit of the corresponding forecast time limit):

[0066]

[0067]

[0068] It should be noted that, since the forecast product of the present invention adopts the fine grid forecast product issued by the European Meteorological Center (EC), the forecast time interval is 3 hours. The EC fine grid forecast product is issued twice a day (the reporting time is 08:00 and 20:00). After deducting the time consumed by EC numerical model calculation, data distribution, decoding, etc., the forecast data of the EC product can be stably obtained on the system, which is delayed by about 8 hours compared to the reporting time. Therefore, after downscaling the forecast data of the EC product by spatiotemporal interpolation, the forecast data within the time period of 4 to 16 o'clock is set, and the forecast data of the corresponding time reported by the EC product at 20:00 the day before is adopted. The forecast data from 17:00 to 3:00 the next day adopts the forecast data of the EC product at 08:00.

[0069] After training each sea fog visibility prediction model, the present invention further comprises testing each trained model with independent samples, and obtaining a final sea fog visibility prediction model according to the test results.

[0070] The present invention obtains a forecast factor combination according to the start and end time points of each forecast period and each forecast time validity period corresponding to the forecast period, and divides the target data subset corresponding to each forecast time validity period under each forecast period, takes the forecast factor combination in the target data subset as the input of the model, and uses the visibility data in the observation data of the forecast time corresponding to the target data subset as the output of the model to train a deep convolutional neural network model corresponding to the forecast time validity period. The present invention constructs a target data subset corresponding to a forecast time validity period in a forecast period to train a deep convolutional neural network model corresponding to the forecast time validity period (or forecast time), thereby obtaining a sea fog visibility forecast model corresponding to each forecast time validity period under each forecast period, which refines the sea fog visibility forecast model and greatly improves the prediction accuracy of the sea fog visibility in each forecast period.

[0071] In step S32, the visibility data in the observation data is:

[0072] Actual visibility or actual visibility difference;

[0073] The actual visibility difference is the difference between the actual visibility corresponding to the forecast time and the actual visibility at the current time; the current time represents the starting time point of the model.

[0074] The actual visibility difference is normalized before being input into the model. The formula is:

[0075]

[0076] Where y represents the actual visibility difference; y max The maximum absolute value of the difference between the actual visibility at each forecast time in the filtered sample data set and the actual visibility at the current time; y out It represents the dimensionless actual visibility difference obtained after normalization of the actual visibility difference.

[0077] S4: According to the current forecast period and the required forecast time, select the corresponding sea fog visibility forecast model, and output the visibility forecast value through the sea fog visibility forecast model.

[0078] In this embodiment, after selecting the corresponding sea fog visibility forecast model, the corresponding forecast factor combination is input according to the forecast period and the required forecast timeliness, so as to output the visibility forecast value (actual visibility or actual visibility difference) through the sea fog visibility forecast model.

[0079] The observation data of each detection site also includes: port area control records.

[0080] Port control records refer to records of maritime and port authorities initiating waterway traffic and terminal operation controls due to weather or other reasons. These records also include the reason for the control, the controlled area, and the start and end times.

[0081] The present invention increases the proportion of fog samples in the sample data set by incorporating port control records into the sample data set.

[0082] It should be noted that if Figure 3 As shown, the specific structure of the sea fog visibility forecast model in the present invention includes the following sequentially connected layers: a signal input layer, a convolution layer, a batch normalization layer, an activation layer, a convolution layer, a first residual convolution module (whose input size IN is 64; output size OUT is 64), a second residual convolution module (whose input size IN is 64; output size OUT is 64), a third residual convolution module (whose input size IN is 64; output size OUT is 128), a fourth residual convolution module (whose input size IN is 128; output size OUT is 128), a fifth residual convolution module (whose input size IN is 128; output size OUT is 256), a sixth residual convolution module (whose input size IN is 256; output size OUT is 256), an activation layer, and a fully connected layer output;

[0083] The specific structure of each residual convolution module is as follows Figure 4 Shown, including:

[0084] Input layer, judgment layer (i.e. Figure 4 IN==OUT in the judgment layer), when the condition of the judgment layer is not met, it enters the left channel, which includes the first channel and the second channel. The first channel includes a batch normalization layer and a convolutional layer (the last layer of the channel), and the second channel includes a batch normalization layer, an activation layer, a convolutional layer, a batch normalization layer, an activation layer, and a convolutional layer (the last layer of the channel) connected in sequence; the last layers of the first and second channels are connected to the first fusion layer and then output;

[0085] When the conditions of the judgment layer are met, the right channel is entered. The right channel includes the first channel and the second channel. The first channel includes a batch normalization layer, an activation layer, a convolutional layer, a batch normalization layer, an activation layer, and a convolutional layer (the last layer of the channel) connected in sequence; the second channel and the last layer of the first channel are connected to the second fusion layer and then output.

[0086] The present invention obtains prediction periods by dividing a day into preset periods, establishes a preset number of deep convolutional neural network models in each prediction period, and trains the deep convolutional neural network model through a sample data set to obtain a sea fog visibility forecast model corresponding to each forecast time validity in each prediction period. The present invention constructs a sea fog visibility forecast model corresponding to each forecast time validity (each forecast time validity is not equal) in each prediction period of each day, so that when actually making a prediction, the corresponding sea fog visibility forecast model can be selected for prediction according to the prediction period corresponding to the current time and the required forecast time validity, which effectively improves the accuracy of the model prediction and solves the problem that the current conventional forecast products have low accuracy and cannot well meet the meteorological support service needs of ship scheduling and berth operations in ports.

[0087] Example 2

[0088] like Figure 2 As shown, the present invention also proposes a sea fog visibility forecasting system based on a deep convolutional neural network model, comprising:

[0089] The dataset acquisition module is used to obtain the observation data of each detection station and the port area control records within the preset time period, convert them into grid point data within the port area through interpolation, and include them in the sample dataset; obtain the forecast data corresponding to each forecast time within the preset time period through forecast products, convert them into grid point data through downscaling spatiotemporal interpolation, and include them in the sample dataset;

[0090] A model building module is used to divide a day into preset time periods to obtain each prediction period, and to build a preset number of deep convolutional neural network models in each prediction period;

[0091] A training module is configured to train a deep convolutional neural network model using a sample data set to obtain a sea fog visibility forecast model corresponding to each forecast time validity period in each forecast period; the number of sea fog visibility forecast models in each forecast period is equal to a preset number; the forecast time validity period represents the time difference between the model start time point and the forecast time number, and the forecast time number is the time validity cutoff point of the forecast time validity period;

[0092] The prediction module is used to select the corresponding sea fog visibility forecast model according to the current forecast period and the required forecast timeliness, and output the visibility forecast value through the sea fog visibility forecast model.

[0093] In this embodiment, after selecting the corresponding sea fog visibility forecast model, the corresponding forecast factor combination is input according to the forecast period and the required forecast timeliness, so that the sea fog visibility forecast model outputs a visibility forecast value (actual visibility or actual visibility difference). It should be noted that after the model outputs the visibility forecast value, the system also includes:

[0094] The level classification module is used to interpolate the visibility prediction value to the hourly times in the grid point data within the port area, and divide the visibility into five levels to form a refined sea fog visibility forecast system with higher temporal and spatial resolution.

[0095] It should be noted that the visibility classification standards are based on actual business service needs, as well as national industry standards, "Fog Forecast Grades" (GB / T 27964-2011), and "Horizontal Visibility Grades" (GB / T33673-2017). Visibility is divided into five levels so that different service measures can be adopted according to different levels in actual meteorological support services. The specific visibility classification levels are shown in the table below:

[0096]

[0097]

[0098] It should be noted that in this system, if the actual visibility at a detection site is <1km at the current time, or the maximum relative humidity of the grid points within the port area at the corresponding forecast time in the latest EC product forecast data is >90%, the corresponding sea fog visibility forecast model will be activated for forecasting. Otherwise, the forecast data of the EC product will be used directly for forecasting.

[0099] The proposed sea fog visibility forecast system effectively addresses the current lack of objective sea fog forecast tools with high temporal and spatial resolution in port and shipping meteorological services. Compared to current EC fine-grid visibility forecast products, the system's sea fog visibility forecast model achieves a TS score exceeding 63% in sea fog trend forecast verification in actual operational applications, exceeding the TS score of EC fine-grid forecast products by 13% to 20%. This demonstrates operational usability and enhances the meteorological support capabilities of ports and shipping.

[0100] It should be noted that all directional indications in the embodiments of the present invention (such as up, down, left, right, front, back, etc.) are only used to explain the relative position relationship, movement status, etc. between the various components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0101] In addition, in the present invention, descriptions such as "first," "second," and "one" are for descriptive purposes only and should not be understood to indicate or imply their relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include at least one of such features. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

[0102] In the present invention, unless otherwise specified or limited, the terms "connection" and "fixation" should be understood in a broad sense. For example, "fixation" can mean fixed connection, detachable connection, or integration; mechanical connection or electrical connection; direct connection or indirect connection through an intermediate medium; internal communication between two elements or interaction between two elements, unless otherwise specified. Those skilled in the art will be able to understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0103] In addition, the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the fact that ordinary technicians in this field can implement it. When the combination of technical solutions is mutually contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

Claims

1. A sea fog visibility forecasting method based on a deep convolutional neural network model, characterized in that: Including steps: S1: Obtain observation data from each monitoring station within a preset time period, convert it into grid point data within the port area through interpolation, and include it in the sample dataset; obtain forecast data corresponding to each forecast time within the preset time period through forecast products, convert it into grid point data through downscaling and spatiotemporal interpolation, and include it in the sample dataset; S2: Divide a day into preset time periods to obtain each prediction period, and establish a preset number of deep convolutional neural network models in each prediction period; Filter the sample data set according to the preset filtering rules, and normalize the data that meets the preset filtering rules to obtain the target data set; S3: training a deep convolutional neural network model using the target data set to obtain sea fog visibility forecast models corresponding to each forecast time validity in each forecast period; the number of sea fog visibility forecast models in each forecast period is equal to a preset number; the forecast time validity represents the time difference between the model start time point and the forecast time, and the forecast time is the time validity cutoff point of the forecast time validity; Training a deep convolutional neural network model using the target dataset involves the following steps: S31: Based on the start and end time points of each forecast period and the forecast time limits corresponding to the forecast period, the grid point data corresponding to the observation data in the target data set are combined with the grid point data corresponding to the forecast data to obtain a forecast factor combination, and the combination is divided into target data subsets corresponding to the forecast time limits under each forecast period; S32: using the forecast factor combination in the target data subset as the input of the model, and the visibility data in the observation data corresponding to the forecast time in the target data subset as the output of the model, training the deep convolutional neural network model corresponding to the forecast time to obtain the sea fog visibility forecast model corresponding to each forecast time in each forecast period; S4: According to the current forecast period and the required forecast time, a corresponding sea fog visibility forecast model is selected, and the visibility forecast value is output through the sea fog visibility forecast model.

2. The method for predicting sea fog visibility based on a deep convolutional neural network model according to claim 1, wherein: The observation data of each detection site includes: actual visibility, relative humidity, and temperature corresponding to each hour; the forecast data corresponding to each forecast time includes: relative humidity, temperature, and predicted visibility corresponding to the forecast time.

3. The method for predicting sea fog visibility based on a deep convolutional neural network model according to claim 2, wherein: The preset screening rules are: The actual visibility in the observed data is less than the preset value and the relative humidity is greater than the preset percentage, or the relative humidity in the forecast data is greater than the preset percentage.

4. The method for predicting sea fog visibility based on a deep convolutional neural network model according to claim 2, wherein: In step S32, the visibility data in the observation data is: Actual visibility or actual visibility difference; The actual visibility difference is the difference between the actual visibility corresponding to the forecast time and the actual visibility at the current time; the current time represents the starting time point of the model.

5. The method for predicting sea fog visibility based on a deep convolutional neural network model according to claim 2, wherein: The formula for normalizing the data that meets the preset screening rules is: ； Where x represents the actual visibility; Represents the maximum visibility value in the filtered sample data set; Represents the minimum visibility value in the filtered sample data set; It represents the dimensionless actual visibility obtained by normalizing the actual visibility.

6. The method for predicting sea fog visibility based on a deep convolutional neural network model according to claim 4, wherein: The actual visibility difference is normalized before being input into the model. The formula is: ； Where, Indicates the actual visibility difference; It represents the maximum absolute value of the difference between the actual visibility at each forecast time in the filtered sample data set and the actual visibility at the current time; It represents the dimensionless actual visibility difference obtained after normalization of the actual visibility difference.

7. The method for predicting sea fog visibility based on a deep convolutional neural network model according to claim 1, wherein: The observation data of each detection site also includes: port area control records.

8. A sea fog visibility forecasting system based on a deep convolutional neural network model, characterized in that: include: The dataset acquisition module is used to obtain the observation data of each detection station and the port area control records within the preset time period, convert them into grid point data within the port area through interpolation, and include them in the sample dataset; obtain the forecast data corresponding to each forecast time within the preset time period through forecast products, convert them into grid point data through downscaling spatiotemporal interpolation, and include them in the sample dataset; A model building module is used to divide a day into preset time periods to obtain each prediction period, and to build a preset number of deep convolutional neural network models in each prediction period; Filter the sample data set according to the preset filtering rules, and normalize the data that meets the preset filtering rules to obtain the target data set; A training module is configured to train a deep convolutional neural network model using a target data set to obtain a sea fog visibility forecast model corresponding to each forecast time validity period in each forecast period; the number of sea fog visibility forecast models in each forecast period is equal to a preset number; the forecast time validity period represents the time difference between the model start time point and the forecast time number, and the forecast time number is the time validity cutoff point of the forecast time validity period; Train a deep convolutional neural network model using the target dataset, including: According to the start and end time points of each forecast period and the forecast time limits corresponding to the forecast period, the grid point data corresponding to the observation data in the target dataset are combined with the grid point data corresponding to the forecast data to obtain a forecast factor combination, and the combination is divided into the target data subset corresponding to each forecast time limit under each forecast period; The prediction factor combination in the target data subset is used as the input of the model, and the visibility data in the observation data of the target data subset corresponding to the forecast time is used as the output of the model. The deep convolutional neural network model corresponding to the forecast time is trained to obtain the sea fog visibility forecast model corresponding to each forecast time in each forecast period. The prediction module is used to select the corresponding sea fog visibility forecast model according to the current forecast period and the required forecast timeliness, and output the visibility forecast value through the sea fog visibility forecast model.

Images