KR102116917B1 - Prediction system of ocean environmental data and required power of ships using deep learning and big data - Google Patents

Abstract

The present invention relates to a system for predicting marine weather and required power using big data and deep learning, which can perform deep learning based on marine weather data in accordance with existing water temperature, wave heights, wave periods, wave directions, wind speeds, wind directions, flow speeds and flow directions and required power data with respect to several ships (for example, 72 ships with 13,000 TEU), and predict the marine weather and the required power with respect to a predicted course of the specific ship based on performed deep learning.

Description

Prediction system of ocean environmental data and required power of ships using deep learning and big data}

The present invention performs deep learning on the basis of the existing water temperature, wave height, wave period, wave direction, wind speed, wind direction, flow rate and flow direction, and the required horsepower data for 72 ships of 13,000TEU class, and performs deep learning. It relates to a system for predicting the sea and the required horsepower using big data and deep learning, which can predict the sea and the required horsepower for the expected route of a specific vessel based on running.

Recently, research using big data in the shipbuilding and offshore sector has attracted attention.

In particular, the Automatic Identification System (AIS) records a lot of information related to the operation of the ship, and the need for research to acquire and accumulate it is increasing.

Indeed, not only the classification but also related companies and academies are starting such research to find a way to analyze and utilize AIS data.

This AIS is public data that does not require measurement, and can be used to derive meaningful results by securing it with marine environment data and ship data.

In the future, if additional measurement data, engine operation data, etc. are to be secured, it will be used as a variety of engineering application analysis data such as CBM (Condition-Based Monitoring & Maintenance) and Sea Margin estimation.

By linking these AIS data with sea weather data, it is possible to perform a multi-faceted analysis of the ship's past operation records, acquire big data, and perform analysis based on it. It was difficult to develop jointly as a service.

In addition, because there is a delay time of about 2 to 3 months before sea data is processed and released, it has been somewhat difficult to utilize the latest data or analysis of the expected route of the ship.

Accordingly, the present applicant expects that by developing a predictive evaluation model for the predicted route of a specific ship, even if there is no recent data of sea weather data, it is possible to perform analysis for navigation and provide information.

As a related technology, an optimal navigation route providing system and method for predicting a real sea environment in published patent publication No. 10-2018-0045341 and an optimal method using big data related to coastal weather and ocean information in published patent publication No. 10-2018-0054410 There are technologies such as a route providing system.

However, the above-described techniques are related to a technique for providing an optimal route according to a marine situation, that is, a technique for proposing a route through which a ship passes, such as navigation.

In other words,  As in the present invention,  It is different from the technology that performs and predicts the analysis of the route the ship will navigate.

Published Patent Publication No. 10-2018-0045341 (2018.05.04.) Published Patent Publication No. 10-2018-0054410 (2019.05.22.) Registered Patent Publication No. 10-1850866 (Apr. 20, 2018) Registered Patent Publication No. 10-1322415 (2013.10.25. Announcement)

The object of the present invention is to perform deep learning based on the existing water temperature, wave height, wave period, wave direction, wind speed, wind direction, flow rate, and sea weather data according to flow rate and required horsepower data for 72 ships of 13,000TEU class, The present invention provides a system for predicting the sea and the required horsepower using big data and deep learning, which can predict the sea and required horsepower for the expected route of a specific ship based on the performed deep learning.

It was devised to achieve the above object, the sea and the required horsepower prediction system using big data and deep learning according to the present invention,

Deep learning learning based on water temperature, wave height, wave period, wave direction, wind speed, wind velocity, flow rate and flow direction, and sea horsepower data for 72 ships of 13,000TEU class Do it,

The main technical features are predicting the sea data on the predicted route to which the corresponding ship will sail, and predicting the horsepower required to navigate the predicted route according to the entered ship data based on the learned learning.

According to the sea weather and required horsepower prediction system using big data and deep learning according to the present invention, the existing sea temperature data according to the existing water temperature, wave height, wave period, wave direction, wind speed, wind direction, flow velocity and flow direction, and 13,000TEU class ship 72 Deep learning is performed based on required horsepower data for the chuck, and it has the effect of predicting the sea and required horsepower for the expected route of a specific vessel based on the performed deep learning.

1 is a view to help explain the information processing of the sea and the required horsepower prediction system using big data and deep learning according to the present invention.

The terms or words used in the present specification and claims should not be construed as being limited to ordinary or lexical meanings, and the inventor can appropriately define the concept of terms in order to best describe his or her invention. Based on the principles, it should be interpreted as meanings and concepts consistent with the technical spirit of the present invention.

Therefore, the examples shown in the embodiments and the drawings described herein are only the most preferred embodiments of the present invention and do not represent all of the technical spirit of the present invention, and thus various equivalents can be substituted at the time of application. It should be understood that there may be water and variations.

Hereinafter, prior to the description with reference to the drawings, not necessary for revealing the gist of the present invention, that is, a well-known configuration that can be added by those skilled in the art will not be shown or not specifically described Well, reveal the notes.

The present invention performs deep learning based on the existing water temperature, wave height, wave period, wave direction, wind speed, wind direction, flow rate and flow rate, and the required horsepower data for a small vessel, and based on the performed deep learning. The present invention relates to a system for predicting the sea and the required horsepower using big data and deep learning, which can predict the sea and the required horsepower for the expected route of a specific ship.

Referring to Figure 1 of the accompanying drawings to describe the present invention.

1 is a view to help explain the information processing of the sea and the required horsepower prediction system using big data and deep learning according to the present invention.

That is, the expected route for a particular vessel; Weather data; After receiving data on ship data (ship, odd), and pre-processing the data as described below,

It is to provide predicted values through a predetermined program through a model for predicting the sea and required horsepower.

That is, as shown in FIG. 1, the present invention includes a big data processing platform, a deep learning platform, and a program for providing information.

In this case, the meaning of the platform may be composed of a separate hardware or server, or may be installed in a predetermined terminal in the form of software.

The big data processing platform, as well as data to be utilized as big data, the ship data including the predicted route, sea data and ship speed, odd number input to predict the sea and required horsepower for navigation of a specific ship It functions to process.

As an example, data collected for utilizing big data can be pre-processed and then learned through a deep learning platform described below.

As another example, ship data is collected and pre-processed so that learning-based results can be predicted through a deep learning platform.

The predicted results are provided by the program.

First, the big data processing platform functions to receive data and / or ship data, pre-process it, and convert and store the data in an integrated format.

At this time, in the case of ship data, in the format input by the administrator or the user, as described above, it is input and received to include the expected route, the sea data and the ship speed and odd numbers.

In addition, the data refers to the existing water temperature, wave height, wave period, wave direction, wind speed, wind direction, sea weather data for flow rate and flow direction, and the required horsepower data for 72 ships of 13,000TEU class,

At this time, data on water temperature, wave height, wave period, wave direction, wind speed and wind direction receive data for 20 years provided by ECMWF (above, longitude resolution of 0.75 degrees interval (about 80 km), time resolution of 6 hours interval). At this time, it is also possible to use data related to the eight types (water temperature, wave height, wave period, wave direction, wind speed, wind direction, flow velocity, flow direction) of the past sea level and data required by several ships in addition to the data provided by the ECMWF.

In addition, the data on flow velocity and frankincense receive data from HYCOM provided for 4 years and 4 months (above, longitude resolution 0.08 degree intervals (about 9 km), time resolution 3 hour intervals).

The received data or ship data is stored in a raw file format (* .NC) and converted to an HDPS file (H5) using an H5PY package to be stored.

In addition, data received before being converted, stored, and received as described above performs a function of removing data corresponding to an outlier of data.

At this time, in order to maintain the data trend, pre-processing is performed using a pre-trained model, not pre-processing by simple interpolation.

For example, in the case of pre-processing data on water temperature, as shown in [Table 1] below, the learning is performed in advance by using a part having no outlier, and when the temperature before and after the water temperature differs by 1.2 degrees or more, it is determined as an outlier, and the outlier is learned. Replace it with the predicted value from the model.

As another example, pre-processing data for wind direction, wind speed, direction, and flow rate causes the u and v components of the received data to be converted into speed and direction, respectively, and uses the second law of Cosine to determine the difference between the previous direction and the next direction. After calculating and setting it in the direction, compare the values to perform pre-processing and prediction for the direction.

This can be referred to [Table 2].

As another example, data for 72 13,000TEU-class vessels (container ships) is collected by collecting approximately 240,000 data sets into one csv file to generate a datasheet as shown in [Table 3].

Based on this data sheet data, outliers can be determined and eliminated. For example, if the required horsepower is less than 63,910KW or greater than 10,000KW, it can be determined to be the outlier for the required horsepower and removed. have.

In addition, if the speed is greater than 30 knots, it can be determined as an outlier and removed.

In addition, if the heading (athlete) of the ship exceeds 360 degrees, it can be determined as an outlier and removed.

After the outliers of the data are removed in this way, the sea data contained in the data from which the outliers are removed is changed to a relative value for the ship.

Refer to [Table 4] for changing the sea data to a relative value for the ship.

[table   4]   Based on   Under   Mathematically   Velocity and   Direction   To calculate.

That is, based on [Table 4], velocity and direction are calculated based on the velocity and direction components of u and v.

When the relative values for the speed and direction of the calculated ship are compared with the ship operation data, which is the absolute value of [Table 3], and when the comparison result exceeds a predetermined error range, it is determined as an outlier and the removed area The review can be manually reviewed by the administrator. This may be expressed in the form of an alarm, etc., through which it is possible to improve the accuracy of the pre-treatment according to the removal of the outlier.

The deep learning platform performs deep learning learning based on the pre-processed data collected from the big data processing platform.

Deep learning learning performed in the deep learning platform is a basic artificial neural network structure, and is performed using a structure including an input layer, a hidden layer, and an output layer.

For example, based on the input data, the node values of the next layer are input using the node values and connection weights of each layer (input layer, hidden layer, output layer), the predicted value determined by each layer is calculated, and then calculated Calculate how much the predicted value differs from the actual value (see [Table 5]).

At this time, the deep learning learning unit continuously updates the connection weight to minimize the difference value (Loss value) (see [Table 5]).

Based on this deep learning learning, we use Convolutional Long Short-Term Memory (LSTM), a kind of Recurrent Neural Network (RNN), for sea weather prediction.

LSTM can refer to [Table 6].

The Convolutional LSTM of the present invention is a model (see [Table 7]) modified from the structure of the LSTM to receive a two-dimensional image with respect to the LSTM above, and LSTM for learning the temporal meaning and CNN for learning the spatial meaning. It is a model that can learn temporal meaning and spatial meaning by combining. These models are often used to identify patterns in a series of images. That is, since multiple images can be used as input data, learning considering the relationship between the weather data is possible.

In summary, upon receiving the sea data, deep learning learning is performed through the model of [Table 7] in the above-described big data processing platform. It predicts until 7 days later, and if exceeded, the predicted data can be predicted by learning the predicted data again).

That is, the deep learning platform performs deep learning learning based on the received data in connection with the big data processing platform, and then, when receiving ship data corresponding to a specific point and time, based on the previously learned big data, it is necessary Predictive data can be provided through programs.

At this time, the data and ship data are described in the present specification, and it is sufficient if the data is interpreted to mean existing data for deep learning learning, and the ship data is based on pre-trained data. It can be interpreted as input data to be provided. That is, the necessary prediction data described above is output data provided based on the ship data.

In addition, in case of predicting required horsepower, a Deep Feedforward Network (DFN) is used.

In the case of required horsepower, it is preferable to use DFN since there is no periodic pattern over time.

The input data consists of a total of 9 types of data, including flight status data (3 types) and sea weather data (6 types), where the 3 types of operation status data are data for draft, ship speed, and ship heading, The six types of sea weather data are data on water temperature, wind speed, wind direction, wave height, wave direction, and wave period.

This required horsepower prediction is also learned as shown in [Table 8] based on [Table 5].

At this time, deep learning learning for predicting required horsepower is to use the collected sea weather data as it is.

Applicant tested the required horsepower prediction using DFN using all sea weather data as it is in the first embodiment, and converted the wind speed and wind direction into relative values for the ship in the second embodiment, using DFN I tested horsepower prediction,

As shown in [Table 9], it was confirmed that the accuracy when using all the sea weather data according to the first embodiment without processing was higher.

However, in the case of [Table 8], since the pre-trained data occupies a significant number, it is necessary to be able to optimize the hyperparameter that can affect the required horsepower prediction result.

At this time, as the elements for optimization, the number of hidden layers, the number of nodes in the hidden layer, the learning rate, the dropout probability, and the type of optimizer are determined as elements.

에 대한 a의 값을 의미하고, Dropout 확률은 모델의 일반화를 위하여 학습에 참여하는 각층별 노드를 줄이는 확률을 의미하며, optimizer의 종류는 ADAM(ADAptive Moment estimation), RMSprop(Root Mean Square propagation), SGD(Stochastic Gradient Descent)를 포함한다.At this time, the learning rate is the degree to which the gradient changes,

The value of a for, and the dropout probability means the probability of reducing each node participating in learning for generalization of the model.The types of optimizers are ADAM (ADAptive Moment Estimation), RMSprop (Root Mean Square propagation), Includes SGD (Stochastic Gradient Descent).

As a result of evaluating the optimization of the hyperparameters for these elements, as shown in [Table 10], the number of hidden layers is 6; The number of nodes in the hidden layer is 16; Learning rate is 0.01;%; Dropout probability is 0%; For the type of optimizer, it was confirmed that using ADAM is the most advantageous for hyperparameter optimization, so the error is the lowest.

It is obvious that the above description using the drawings is only the main points of the present invention, and that the present invention is not limited to the configuration of the drawings, as various designs are possible within the technical scope.

Claims (8)

Based on the collected data, we perform deep learning learning that predicts the weather and required horsepower,
After receiving ship data including expected route, sea data and ship data, odd data, based on deep learning learning performed, BIC predicts the sea weather and required horsepower while the ship means the ship is sailing. In the sea and required horsepower prediction system using data and deep learning,
The system,
A data processing platform that receives and processes data and ship data,
Deep learning platform that performs learning based on the data received from the data processing platform and provides results for predicting the sea and required horsepower according to the ship data received from the data processing platform based on the previously performed learning. and,
Includes a program that outputs the results provided by the deep learning platform,
The above data for deep learning learning,
Data on water temperature, wave height, wave period, wave direction, wind speed and wind direction over 20 years provided by ECMWF,
Data on flow rate and frankincense for 4 years and 4 months provided by HYCOM,
Includes data with 240,000 data sets for 72 13,000TEU-class vessels (container ships),
The data processing platform,
Convert the received data or ship data into data having the same format and store it,
Before conversion, has a pre-processing function to remove data corresponding to the outliers of the data or ship data,
The pretreatment,
-In the case of water temperature, if it differs by more than 1.2 degrees from the previously received water temperature, it is determined as an outlier and removed, and replaced with the predicted value based on the pre-trained model.
-The pre-treatment,
In the case of wind direction, wind speed, direction, and flow rate, u and v components are converted into speed and direction on the data, and the difference between the previous direction and the next direction is calculated using the second law of Cosine, and the values are set in the direction and compared. To perform pre-processing and prediction for the direction,
-For the required horsepower data of the ship,
(a) If the required horsepower is less than 63,910KW or greater than 10,000KW, it is determined as an outlier for the required horsepower and removed.
(b) If the ship's speed is greater than 30 knots, it is judged as an outlier and removed.
(c) When the angle of a player exceeds 360 degrees, it is determined as an outlier and removed, and the system for predicting sea and required horsepower using big data and deep learning.
delete delete delete delete delete The method according to claim 1,
After the data for the outliers is removed through the pre-processing, velocity and direction are calculated by changing the sea weather data included in the data from which the outliers are removed to the relative values for the ship through the following equation. Characterized in that, sea and sea horsepower prediction system using big data and deep learning.

The method according to claim 1,
The deep learning platform,
Learning is performed using a structure including an input layer, a hidden layer, and an output layer,
Based on the input data, the node value of the next layer is input using the node value and connection weight of each layer, the final prediction value is predicted through the output layer, and the actual value and the prediction value are compared through the following equation. After determining the loss value (L), characterized in that the learning in the direction to continuously update the connection weight in the direction of minimizing the loss value, the sea and the required horsepower prediction system using big data and deep learning.

Images